Home

mass flow measurement of multi-phase mixtures by - MWEB

1. SYSTEM GRAVEL VOLUME FRACTION 10 8 6 volume fraction 96 8 6 time s BOTTOM SYSTEM GRAVEL VOLUME FRACTION volume fraction 96 10 8 6 1 2 time s 3 FIGURE 7 19 biased correlation biased correlation 0 8 0 6 0 4 02 time s GRAVEL CROSS CORRELATION 0 8 0 6 0 45 0 2 PLOTS OF THE AIR AND GRAVEL VOLUME FRACTION PROFILES AT THE TOP AND BOTTOM MEASUREMENT PLANES WHERE THE VELOCITY OF THE 0 04 VOLUME FRACTION AIR BUBBLE WAS 0 5m s AND THE VELOCITY OF THE 0 04 VOLUME FRACTION GRAVEL BUBBLE WAS 0 2m s INCLUDED ARE THE CROSS CORRELATION FUNCTIONS OF THE AIR AND GRAVEL VOLUME FRACTION PROFILES 105 Firstly each cross correlation function contains two peaks corresponding to the partial separation of the air and gravel bubble In the ideal case where the reconstruction algorithm completely separates the detection of the air and gravel phases only one peak will exist in each cross correlation function In this particular example the reconstruction algorithm has achieved a satisfactory separation Consequently the larger peak of the air cross correlation function correctly predicts the velocity of the air phase as 0 48m s and the larger peak of the gravel cross correlation function correctly predicts the velocity of the gravel phase as 0 20m s However for tests where a successful separation between the t
2. add markers to the chart to indicate what the algorithm determined to be the peaks of the different plots if airtopdesiremax lt gt 1 then begin TopAirPlot Series 2 AddXY airtopmax 1 10 clBlue TopAirPlot Series 2 AddXY airtopdesiremax 1 10 clRed end if graveltopdesiremax lt gt 1 then begin TopGravelPlot Series 2 AddXY graveltopmax 1 10 clBlue TopGravelPlot Series 2 AddXY graveltopdesiremax 1 10 clRed end if airbotdesiremax lt gt 1 then begin BotAirPlot Series 2 AddXxY airbotmax 1 10 clBlue BotAirPlot Series 2 AddXY airbotdesiremax 1 10 clRed end if gravelbotdesiremax lt gt 1 then begin BotGravelPlot Series 2 AddXY gravelbotmax 1 10 clBlue BotGravelPlot Series 2 AddXY gravelbotdesiremax 1 10 clRed end end ACTUAL VELOCITY CALCULATION PROCEDURE since the stepper motors only provide a relatively low acceleration rate the bubble does not always reach constant speed before entering the measuring rings The cross correlation algorithm calculates an average velocity between the two measuring rings so if the bubble has not reached constant velocity before entering the measuring section the cross correlation determined velocity will not equal the desired velocity This algorithm takes that into consideration and adjusts the desired velocity using the same princples as the algorithm to calculate the position of the bubble at a s
3. This unit provides a means of verifying the dynamic results captured during the real time testing It operates purely in terms of volume fraction Using a defined starting position for the bubble and knowledge of the bubble velocity and accelaration it is able to predict when the bubble should be passing through a measurement section since the position of the measurement electrodes is specified according to the bubble starting position Since it can predict fairly accurately the position of a bubble at a certain time it is therefore also able to predict the desired volume fractions of the different phases at that time These desired volume fraction profiles are then compared to those captured from the real time testing and an error calculation is performed To compensate for slippage between the belt and pulley of the bubble drives the real time volume fraction profile is shifted to coincide with the desired volume fraction profile unit dynunit interface uses Windows Messages SysUtils Classes Graphics Controls Forms Dialogs StdCtrls ExtCtrls Spin Buttons TeEngine Series TeeProcs Chart Registry ath TeeFunci type TDynVerify class TForm Paneli TPanel Panel3 TPanel Label1 TLabel AirPhaseConfig TGroupBox GravPhaseConfig TGroupBox PlateConfig TGroupBox Label2 TLabel Label3 TLabel Label4 TLabel Label5 TLabel LowerSystemBottom TEdit LowerSystemTop TEdit UpperSystemBottom TEd
4. B gate voltage IRF9530 A B gate voltage PREVENTION i IRF9530 CIRCUITRY PUSH PULL POWER MOSFET OUTPUT STAGE The operation of the cross conduction prevention circuitry can be explained using the accompanying plots of the gate voltages for the two power MOSFETS At point A The IRF520 turns off quickly because the diode on the gate of the IRF520 conducts thus rapidly discharging its gate capacitance The IRF9530 turns on slowly because the diode on the gate of the IRF9530 is reverse biased so the gate capacitance of the IRF9530 discharges slowly At point B The IRF520 turns on slowly because the diode on the gate of the IRF520 is reverse biased so the gate capacitance of the IRF520 charges slowly The IRF9530 turns off quickly because the diode on the gate of the IRF9530 conducts thus rapidly charging its gate capacitance Subsequently a dead time is introduced into the output transitions to minimise the cross conduction in the power MOSFETs FIGURE 5 9 OPERATION OF CROSS CONDUCTION PREVENTION CI RCUITRY IN PUSH PULL POWER MOSFET OUTPUT STAGE 42 The cross conduction prevention circuitry creates an RC time delay using the gate capacitance of the power MOSFETSs Schottky diodes are used to vary this time delay depending on whether the corresponding MOSFET is switching off or on The operation of the cross conduction circuitry can be explained as follows when the TC4427 output goes low at point A t
5. implementation uses loadunit mainunit R DFM tM M DATABASE VIEWING PROCEDURES 5 move to the next test case in the database by incrementing currentindex and calling DisplayData if at the last element in the database then disable the button procedure TRecViewMain RecViewNextClick Sender TObject begin inc currentindex RecViewPrevious Enabled true if currentindex RecLoadDBase GetTotalTest RecLoadDBase GetTotalTrain 1 then begin RecViewNext Enabled false UseContinouos Checked false end DisplayData end move to the previous test case in the database by decrementing currentindex and calling DisplayData if first element in the database then disable the button procedure TRecViewMain RecViewPreviousClick Sender TObject begin dec currentindex if currentindex 0 then begin UseContinouos Checked false RecViewPrevious Enabled False end RecViewNext Enabled true DisplayData end display the desired network output for the current test case in the database procedure TRecViewMain DisplayData var contair contgrav Boolean index row col top bottom integer tempout TOutput begin RecViewVolLabel Caption Volume fractions for test inttostr currentindex 1 if RecLoadDBase RecDesiredRecon ItemIndex 0 then begin 163 for the image reconstruction use a TTomo component to display the desired network output RecViewlm
6. 0 to temphidden 1 do for rowindex 0 to 128 do ReadLn tempfile HiddenWeight rowindex colindex end load in the weights for the output layer neurons for colindex 0 to tempoutputs 1 do for rowindex 0 to upperindex do ReadLn tempfile OutputWeight rowindex colindex CloseFile tempfile Starttraining1 Enabled true Validatedatabase1 Enabled true Validatecurrentsamples1 Enabled true end end PROGRAM LISTING OF LOADUNIT PAS This unit loads the data from the specified databases into the matrices which will be used for network training This unit also performs pre processing on the loaded data standardisation is performed with the option of calibrating the system according to test cases where the rig is filled only with water unit loadunit interface uses Windows Messages SysUtils Classes Graphics Controls Forms Dialogs ExtCtrls StdCtrls Spin ComCtrls Buttons DBTables Math Registry type TRecLoadDBase class TForm Paneli TPanel Panel2 TPanel Panel3 TPanel RecTrainStart TSpinEdit RecTrainStop TSpinEdit RecTrainName TComboBox Bevel1 TBevel Label TLabel Label2 TLabel Label3 TLabel Label4 TLabel Bevel2 TBevel Label5 TLabel Label6 TLabel Label7 TLabel Label8 TLabel RecTestName TComboBox RecTestStart TSpinEdit RecTestStop TSpinEdit RecLoadData TSpeedButton Label9 TLabel Label10 TLabel RecPreProcess TSpeedButton Rec
7. Cryarxa cos w at sin w Sin w gt sin w At cos wgt sin w which can be rewritten using trigonometric identities as Vauucra t PARA i cos 2w t 89 sin gw t G C ES TxBRxA cos TxBRxA wat w t cos wgt wat sin wgt w sin wgt wat Gtyarxa V gmucta t ERE if wipe lt min 2w 4 wg wal 4 8 Consequently the multiplier output is proportional to the conductance between transmitter A and receiver A provided the cutoff frequency of the low pass filter is less than the minimum of 2wa Wg Wa Wg Wa To determine the conductance between transmitter B and receiver A the received signal is multiplied by the reference for transmitter B and the output is as follows G TxARxA cos w Wet cos w t wgt sin Wat sin w At wgt 0 4 9 di cos 2wst SPERA sin 2w pt GTxBRxA t imt m if wipe lt min 2wg w A ws 4 10 The same conditions apply for the low pass filter cutoff frequency as before To determine the capacitance between transmitter A and receiver A the received signal is multiplied by a 90 phase shifted version of the reference for transmitter A as follows t Grana Sin w At cos w cos w 4t cos w t 4 11 Gryerxa Sin wgt cos w qt cos wat cos w at ie which can be rewrit
8. II if the tables do not exist if not InpTable Exists and OutTable Exists then begin setup the fields for the tables InpTable FieldDefs Add TestNo ftFloat InpTable FieldDefs Add System ftInteger OutTable FieldDefs Add TestNo ftFloat II fields for 128 input voltages obtained from the system for i 1 to 128 do begin S Voltage inttostr i InpTable FieldDefs Add s ftFloat end II fields for the desired volume fractions for specific I test cases OutTable FieldDefs Add WaterVolume ftFloat OutTable FieldDefs Add GravelVolume ftFloat OutTable FieldDefs Add AirVolume ftFloat 11 fields for output pixel values for specific test cases for i 1 to 100 do begin s Pixel inttostr i OutTable FieldDefs Add s ftSmalllnt end for all tables the testno will be used as the index InpTable IndexDefs Add TestNo ixPrimary InpTable CreateTable OutTable IndexDefs Add TestNo ixPrimary OutTable CreateTable add an empty record to the table to initialise the table for further use npTable Open OutTable Open npTable Edit OutTable Edit npTable Append npTable FieldByName TestNo Value 0 OutTable Append OutTable FieldByName TestNo Value 0 npTable Post OutTable Post end else begin else simply open the tables and allow editing InpTable Open OutTable Open InpTable Edit OutTable Edit end tM GENERAL PROCEDURES 5
9. procedure TRecTrainNetwork TrainDoubleNetwork var outputnodeindex inputnodeindex currsample i integer weightsum deltasum deltahidden newderiv real delta array of real hiddenderiv array of real weightstore array 1 2 of double begin initialise the important variables currepoch 1 currentfails 0 SetLength delta numoutputs SetLength hiddenderiv numhidden 1 stop when the maximum number of epochs has been exceeded or when early stopping enabled and the validation performance has continued to 11 deteriorate for a certain minimum number of epochs set by maxfails while currepoch lt maxepochs and not currentfails gt maxfails and earlystop do begin for all training points for currsample 0 to totaltrain 1 do begin calculate outputs for hidden layer where the hidden layer simply consists of tanh activation functions for outputnodeindex 0 to numhidden do begin weightsum 0 if outputnodeindex 0 then begin bias node in hidden layer of neurons OutputHidden currsample outputnodeindex 1 hiddenderiv outputnodeindex 0 end else begin calculate the weighted summation the input 169 to the hidden layer nodes for inputnodeindex 0 to 128 do weightsum weightsum InputData currsample inputnodeindex HiddenWeight inputnodeindex outputnodeindex 1 prevent arithmetic overflow if weightsum 200 then weightsum 200 else if weightsum
10. unit mainunit var RecMain TRecMain iterface SampHardware TSamp provide interface to the uses sampling hardware InputData array of array of real Il input voltage readings Windows Messages SysUtils Classes Graphics Controls Forms Dialogs OutputData array of array of real JI desired network outputs ExtCtrls Menus Registry StdCtrls Spin Buttons sampunit type TRecMain class TForm RecMenu TMainMenu File1 TMenultem Loaddatabase1 TMenultem N1 TMenultem Loadnetworkweights TMenultem Savenetworkweights TMenultem N2 TMenultem Savenetworkperformancet TMenultem N3 TMenultem Exit1 TMenultem Networktraining1 TMenultem Settestnumber1 TMenultem Initialisenetworkweights TMenultem Setnetworktrainingparameters1 TMenultem Starttraining1 TMenultem Networktesting1 TMenultem Validatedatabase1 TMenultem Validatecurrentsamples1 TMenultem Paneli TPanel Viewdatabase1 TMenultem Writedatatoatextfile1 TMenultem RecSetTestNumber TPanel Label1 TLabel RecNewTestNumber TSpinEdit RecSetTestNumberOK TBitBtn RecSetTestNumberCan TBitBtn RecSetWeightFactor TPanel Label2 TLabel RecWeightFactor TEdit RecWeightOK TBitBtn RecWeightCan TBitBtn Realtimetesting1 TMenultem Savenetworkpreprocessing1 TMenultem Savecalibrationdata1 TMenultem SaveCalibDialog TSaveDialog procedure Exitt Click Sender TObject procedure Loaddatabase1 Click Se
11. 200 then weightsum 200 calculate hidden layer outputs OutputHidden currsample outputnodeindex tanh weightsum hiddenderiv outputnodeindex CalcDeriv OutputHidden currsample outputnodeindex end end calculate network outputs and adjustments to output weights for outputnodeindex 0 to numoutputs 1 do begin weightsum 0 calculate the weighted summation of the hidden layer outputs to the output node weights for inputnodeindex 0 to numhidden do weightsum weightsum OutputHidden currsample inputnodeindex OutputWeight inputnodeindex outputnodeindex prevent arithmetic overflow if weightsum gt 200 then weightsum 200 else if weightsum lt 200 then weightsum 200 Il if use a 1 of C encoding then the three outputs corresponding to a single pixel can only be calculated once the weighted sums for all three nodes have been calculated since a soft max Il activation function is used if numoutputs 264 then begin if outputnodeindex 1 mod 3 lt gt 0 then weightstore outputnodeindex 1 mod 3 weightsum else begin we can now calculate the outputs from the I softmax function since all three weighted sums are known for i 2 downto 0 do begin OutputNetwork currsample outputnodeindex i SoftMax 3 i weightstore 1 weightstore 2 weightsum calculate the adjustments to the weights of the output layer nodes delta outputnodeindex i OutputNetwork currsam
12. APPENDIX F APPENDIX G APPENDIX H APPENDIX I APPENDIX J APPENDIX K APPENDIX L APPENDIX M APPENDIX N APPENDIX O TABLE OF THE MANUFACTURING COSTS FOR 16 ELECTRODE FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM CIRCUIT DIAGRAMS FOR 8 ELECTRODE FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM CIRCUIT DIAGRAMS FOR 16 ELECTRODE FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM TABLES OF THE CAPACITANCE AND CONDUCTANCE READINGS FOR 16 ELECTRODE FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM MICROCONTROLLER PROGRAM CODE FOR 8 ELECTRODE FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM MICROCONTROLLER PROGRAM CODE FOR 16 ELECTRODE FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM FUNDAMENTAL FREQUENCIES AND HARMONICS FOR 8 ELECTRODE FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM FUNDAMENTAL FREQUENCIES AND HARMONICS FOR 16 ELECTRODE FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM AUTOMATIC CODE GENERATOR PROGRAM CODE PBIL OPTIMISATION ALGORITHM AND SIMULATOR PROGRAM CODE SAMPLER PROGRAM CODE NEURAL NETWORK PROGRAM CODE SCREEN CAPTURES OF SAMPLER PROGRAM SCREEN CAPTURES OF NEURAL NETWORK PROGRAM PAPER PUBLISHED IN MEASUREMENT SCIENCE AND TECHNOLOGY JOURNAL ON AUTHOR S PREVIOUS RESEARCH 78 82 82 85 90 93 93 96 98 101 103 103 105 106 107 110 111 112 116 117 122 129 130 134 135 136 137
13. AirBubblePosition ItemIndex Reg ReadInteger AirBubblePosition GravelBubblePosition ItemIndex Reg ReadInteger GravelBubblePosition FlowDirSelect ItemIndex Reg ReadInteger FlowDirection AirDrive Value Reg ReadInteger AirDrive GravelDrive Value Reg Readinteger GravelDrive AirAccel Text Reg ReadString AirAccel GravelAccel Text Reg ReadString GravelAccel finally Reg Free end end tM INITIALISATION PROCEDURES 9 download the AT6400 operating system procedure TStepperConfig StepDownloadClick Sender TObject begin RealMain Stepper DownloadOS end initialise the drives using the standard settings found in init txt textfile that must be located in the same directory as this program procedure TStepperConfig SteplnitialiseClick Sender TObject begin RealMain Stepper Initialise 768 init txt RealMain Stepper SetZeroPos end reset the stepper drive positions to zero procedure TStepperConfig StepResetClick Sender TObject begin RealMain Stepper ResetLimits RealMain Stepper ResetDrives end tM BUBBLE MOVEMENT PROCEDURES jog the air bubble if left click then bubble moves upwards and if you right click then the bubble moves upwards NOTE the bubble continues to move while button is pressed and since no hardware limits set user must know where bubble is positioned before starting a move proced
14. HiddenWeight rowindex colindex essentially w t 1 w t LR dE dw MOM w t w t 1 Il is the update rule HiddenWeight rowindex colindex HiddenWeight rowindex colindex learnrate DiffHiddenWeight rowindex colindex momentum HiddenWeight rowindex colindex OldHiddenWeight rowindex colindex OldHiddenWeight rowindex colindex tempweight DiffHiddenWeight rowindex colindex 0 end end end update the output layer weights for rowindex 0 to upperindex do begin for colindex 0 to numoutputs 1 do begin tempweight OutputWeight rowindex colindex essentially w t 1 w t LR dE dw MOM w t w t 1 IL is the update rule OutputWeight rowindex colindex OutputWeight rowindex colindex learnrate DiffOutputWeight rowindex colindex momentum OutputWeight rowindex colindex OldOutputWeight rowindex colindex OldOutputWeight rowindex colindex tempweight DiffOutputWeight rowindex colindex 0 end end end MM NETWORK VALIDATION PROCEDURES 5 calculate the network outputs for the test data by propagating the voltages forward through the network At the same time calculate the threshold and sum void error for the testing database procedure TRecTrainNetwork CalculateTestOutput start stop integer var count rowindex colindex i upperindex totalerrors integer tempvalue weightsum real weightstore array 1 2 of real begin initialise imp
15. MOVLW B 01010001 MOVWF PORTB OP OP MOVLW B 00000000 MOVWF PORTB OP OP MOVLW 00001000 MOVWF PORTB OP OP MOVLW 10001000 MOVWF PORTB OP OP MOVLW B 10101000 MOVWF PORTB OP OP MOVLW B 11101010 MOVWF PORTB OP OP MOVLW 11101010 MOVWF PORTB OP OP MOVLW 01111110 MOVWF PORTB OP OP MOVLW B 01111110 N MOVWF PORTB OP 131 NOP MOVLW B 00111111 MOVWF PORTB NOP NOP MOVLW B 00011111 MOVWF PORTB NOP NOP MOVLW B 10011111 MOVWF PORTB NOP NOP MOVLW B 10010111 MOVWF PORTB NOP NOP MOVLW B 11000101 MOVWF PORTB NOP NOP MOVLW B 11000101 MOVWF PORTB NOP NOP MOVLW B 01000101 MOVWF PORTB NOP NOP MOVLW B 01100101 MOVWF PORTB NOP NOP MOVLW B 00100000 MOVWF PORTB NOP NOP MOVLW B 00100000 MOVWF PORTB NOP NOP MOVLW B 10110000 MOVWF PORTB NOP NOP MOVLW B 10110000 MOVWF PORTB NOP NOP MOVLW B 11110010 MOVWF PORTB NOP NOP MOVLW B 11011010 MOVWF PORTB NOP NOP MOVLW 01011010 MOVWF PORTB NOP NOP MOVLW 01011010 MOVWF PORTB NOP NOP MOVLW 00001011 MOVWF PORTB NOP NOP MOVLW 00001011 MOVWF PORTB NOP NOP MOVLW B 10001111 MOVWF PORTB NOP NOP MOVLW B 10101111 MOVWF PORTB NOP NOP MOVLW 11101101 MOVWF PORTB NOP NOP MOVLW 11101101 MOVWF PORTB NOP NOP MOVLW B 01111101 MOVWF PORTB NOP NOP MOVLW B 01110101 MOVWF PORTB NOP NOP MOVLW 001101
16. being viewed starttime TDateTime time test started stoptime TDateTime time test finished timediff TDateTime time difference numtophidden integer numbottomhidden integer busyreconstruction Boolean indicate whether still busy performing a reconstruction busysampling Boolean indicate whether still busy sampling upperindex integer airdelay integer air transit time graveldelay integer gravel transit time topairthreshold real threshold values topgravthreshold real botairthreshold real botgravthreshold real correlationthreshold real drive1steps integer drive2steps integer drive3steps integer driverps real drive2rps real drive3rps real procedure ProcessResults stepper motor parameters procedure PerformCorrelation perform cross correlation procedure RealReconstruction perform reconstruction procedure Stepperlnit linitialise stepper procedure PlotTestOutput testposition integer display output public Public declarations Stepper TStep object used to control and 1 access AT6400 stepper motor controller frameinterval real estimated interval between I frames airvelocity real gravelvelocity real airactualvel real gravelactualvel real flowmeterfactor real systemseparation real procedure WndProc var TheMsg TMessage override end TTomolnputData record wh
17. bestair airerror bestgravel gravelerror bestthresh currthresh RecModelBestPerf Caption floattostrf bestsum ffGeneral 7 5 RecModelNumBest Caption inttostr numbest StoreBestWeights end end else begin if currthresh bestthresh then begin numbest numhidden bestthresh currthresh bestsum currsum bestwater watererror bestair airerror bestgravel gravelerror RecModelBestPerf Caption floattostrf bestthresh ffGeneral 7 5 RecModelNumBest Caption inttostr numbest StoreBestWeights end end increment the number of hidden layer neurons numhidden numhidden RecSetParam RecModellnc Value if not at the maximum number of hidden layer neurons then restart the process if numhidden lt RecSetParam RecModelStop Value then begin ModelSearchTimer Enabled true PauseTrain Enabled true RecMain Savenetworkperformancet Click RecMain Savenetworkpreprocessing1 Click RecMain Savenetworkweights Click end else PauseTrain Visible false end keep a record of the weights corresponding to the best network performance procedure TRecTrainNetwork StoreBestWeights var colindex rowindex integer begin OutputWeightBest nil HiddenWeightBest nil SetLength OutputWeightBest numhidden 1 numoutputs SetLength HiddenWeightBest 129 numhidden for colindex 0 to numhidden 1 do for rowindex 0 to 128 do HiddenWeightBest rowindex colindex HiddenW
18. end else begin airtopposition airtopposition airstartpos gravtopposition gravtopposition gravelstartpos end based on the positions of the bubbles calculate the corresponding volume fractions CalculateVolume airtopposition gravtopposition upperairvol uppergravvol lowerairvol lowergravvol if StepperConfig AddPhaseSelect ItemIndex 0 then begin desiredvolume testposition 1 100 upperairvoltuppergravvol lupperplatevolume desiredvolume testposition 2 0 desiredvolume testposition 3 100 lowerairvol lowergrawvol lowerplatevolume desiredvolume testposition 4 0 end else begin desiredvolume testposition 1 100 upperairvol upperplatevolume desiredvolume testposition 2 100 uppergravvol upperplatevolume desiredvolume testposition 3 100 lowerairvol lowerplatevolume desiredvolume testposition 4 100 lowergravvol lowerplatevolume end TopAirPlot Series 0 AddY desiredvolume testposition 1 TopGravelPlot Series 0 AddY desiredvolume testposition 2 BotAirPlot Series 0 AddY desiredvolume testposition 3 BotGravelPlot Series 0 Add Y desiredvolumeftestposition 4 end end calculate the volume of the bubble within the measuring regions based on the position of the bubble Note that this algorithm makes various assumptions regarding the shape of the bubble namely that the bubble is composed of a cylinder between two hemispheres and the length of the cylindri
19. maxvalue then stopmax testcase end else begin if volumestore testcase 2 gt maxvalue then begin maxvalue volumestore testcase 2 startmax testcase stopmax 0 end else if volumestore testcase 2 maxvalue then stopmax testcase end end TopGravPeak Caption floattostrf maxvalue ffFixed 4 3 if stopmax lt gt 0 then graveltopmax floor startmax 0 5 stopmax startmax else graveltopmax startmax maxvalue 0 startmax 0 stopmax 0 for testcase 1 to RealMain DesiredFrames Value do begin if desiredvolume testcase 2 gt maxvalue then begin maxvalue desiredvolume testcase 2 startmax testcase Stopmax 0 end else if desiredvolume testcase 2 maxvalue then stopmax testcase if desiredvolume testcase 2 gt 0 and graveltopstart 0 then graveltopstart testcase else if desiredvolume testcase 2 gt 0 and graveltopstart lt gt 0 then graveltopstop testcase end if maxvalue 0 then graveltopdesiremax 1 else if stopmax lt gt 0 then graveltopdesiremax floor startmax 0 5 stopmax startmax else graveltopdesiremax startmax topairerror 0 airtopoffset airtopmax airtopdesiremax graveltopoffset graveltopmax graveltopdesiremax shift the data and calculate the mean absolute error between the 11 desired and predicted volume fraction data over the width of the peak where MAE mean abs error if airtopdesiremax lt gt 1
20. power 0 5 airdiam 2 end else if airtopposition gt uppersystop 0 5 airdiam and airtopposition airleng 0 5 airdiam lt uppersysbot then begin upperairvol pi uppersystop uppersysbot power 0 5 airdiam 2 end else if airtopposition airleng 0 5 airdiam gt uppersysbot and airtopposition airleng 0 5 airdiam lt uppersysbot 0 5 airdiam then begin tempheight airtopposition airleng 0 5 airdiam uppersysbot upperairvol pi tempheight power 0 5 airdiam 2 1 3 pi power tempheight 3 pi uppersystop uppersysbot tempheight power 0 5 airdiam 2 end else if airtopposition airleng 0 5 airdiam gt uppersysbot 0 5 airdiam and airtopposition airleng 0 5 airdiam lt uppersystop then begin tempheight uppersystop airtopposition airleng 0 5 airdiam upperairvol airhemivol pi tempheight power 0 5 airdiam 2 end else if airtopposition airleng 0 5 airdiam gt uppersystop and airtopposition airleng lt uppersystop then begin tempheight airtopposition airleng 0 5 airdiam uppersystop upperairvol airhemivol pi tempheight power 0 5 airdiam 2 1 3 pi power tempheight 3 end else upperairvol 0 end else begin if airtopposition gt uppersysbot and airtopposition airleng lt uppersysbot then upperairvol pi power 0 5 airdiam 2 airtopposition uppersysbot else if airtopposition airleng gt uppersysbot and airtopposition lt uppersystop then upperairvol pi power 0 5 airdiam 2
21. 100 FGravelVolume 0 FAirVolume 0 ResetDesired Self OnMouseDown MouseDown Self OnMouseMove MouseMove end the phase of each pixel is set by clicking on the pixel as follows l left mouse button air central button water l right mouse button gravel 1 however only change the pixel phase if the cursor is inside the pipe I and if editing of the pipeline contents is enabled procedure TTomo MouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Integer var row col newX newY integer begin row X div 20 col Y div 20 if FDesired row col lt gt outpipe and FAllowEdit then begin Il air pixel if ssLeft in Shift then begin Canvas Pen Color clWhite Canvas Brush Color clWhite SetDesired row col air end gravel pixel else if ssRight in Shift then 150 begin Canvas Pen Color clGray Canvas Brush Color clGray SetDesired row col gravel end water pixel else begin Canvas Pen Color clBlue Canvas Brush Color clBlue SetDesired row col water end newX row 20 newY col 20 Canvas Rectangle newX newY newX 20 newY 20 end end the pipeline cross section is painted as a sequence of sqaures where the colour of each square is specified by the Desired property procedure TTomo Paint var row col X Y integer begin for row 0 to 9 do begin for col 0 to 9 do begin X row 20 Y col 20 if
22. Boolean begin Reg TRegistry Create try KeyGood Reg OpenKey SoftwareWeural False l if the key exists then read from the key else it is the first time the program is running so do not attempt I to read from the registry if KeyGood then testnumber Reg ReadInteger TestNumber weightfactor Reg ReadFloat WeightFactor finally Reg Free end RecNewTestNumber Value testnumber Randomize end when close form clear the memory allocated for the matrices and write the testnumber and weightfactor to the registry procedure TRecMain FormClose Sender TObject var Action TCloseAction var Reg TRegistry begin Reg TRegistry Create try Reg OpenKey Software Neural True Reg Writelnteger TestNumber testnumber Reg WriteFloat WeightFactor weightfactor finally Reg Free end free dynamically allocated memory InputData nil OutputData nil VolumeData nil OutputNetwork nil OutputHidden nil OutputWeight nil DiffOutputWeight nil OldOutputWeight nil HiddenWeight nil DiffHiddenWeight nil OldHiddenWeight nil OutputDeltaWeight nil HiddenDeltaWeight nil OldDiffHiddenWeight nil OldDiffOutputWeight nil OutputWeightBest nil HiddenWeightBest nil free the resources allocated for the sampling hardware SampHardware ReleaseSample SampHardware Free end set the test number procedure TRecMain RecSetTestNumberOK
23. MOVWF PORTB NOP NOP MOVLW B 01110101 MOVWF PORTB NOP NOP MOVLW B 01110101 MOVWF PORTB NOP NOP MOVLW 00110000 MOVWF PORTB NOP NOP MOVLW 00010000 MOVWF PORTB NOP NOP MOVLW B 10010000 MOVWF PORTB NOP NOP MOVLW B 10010000 MOVWF PORTB NOP NOP MOVLW B 11000010 MOVWF PORTB NOP NOP MOVLW B 11001010 MOVWF PORTB NOP NOP MOVLW 01001010 MOVWF PORTB NOP NOP MOVLW 01101010 MOVWF PORTB NOP NOP MOVLW B 00101011 MOVWF PORTB NOP NOP MOVLW B 00101011 MOVWF PORTB NOP NOP MOVLW B 10111111 MOVWF PORTB NOP NOP MOVLW B 10111111 MOVWF PORTB NOP NOP MOVLW B 11111101 MOVWF PORTB NOP NOP MOVLW B 11011101 MOVWF PORTB NOP NOP MOVLW B 01011101 MOVWF PORTB NOP NOP MOVLW B 01010101 MOVWF PORTB NOP NOP MOVLW B 00000100 MOVWF PORTB NOP NOP MOVLW 00000100 MOVWF PORTB NOP NOP MOVLW B 10000100 MOVWF PORTB NOP NOP MOVLW B 10100100 MOVWF PORTB NOP NOP MOVLW B 11100010 MOVWF PORTB NOP NOP MOVLW B 11100010 MOVWF PORTB NOP NOP MOVLW 01110010 MOVWF PORTB NOP NOP MOVLW 01110010 MOVWF PORTB NOP NOP MOVLW B 00110011 MOVWF PORTB NOP NOP MOVLW B 00011011 MOVWF PORTB NOP NOP MOVLW B 10011011 MOVWF PORTB NOP NOP MOVLW B 10011011 MOVWF PORTB NOP NOP MOVLW 11001001 MOVWF PORTB NOP NOP MOVLW 11001001 MOVWF PORTB NOP NOP MOVLW 01001101 MOVWF PORTB NOP NOP MOVLW B 011
24. RESISTANCE READINGS OF THE DIFFERENT TRANSMITTER RECEIVER ELECTRODE PAIRS AT THE BOTTOM MEASUREMENT PLANE FOR THE LABORATORY SCALE RIG FILLED COMPLETELY WITH POTABLE WATER RxB RxC RxD RxE RxF RxG 3 55k 3 54k 3 57k 3 57k 3 57k CAPACITANCE READINGS OF THE DIFFERENT TRANSMITTER RECEIVER ELECTRODE PAIRS AT THE BOTTOM MEASUREMENT PLANE FOR THE LABORATORY SCALE RIG FILLED COMPLETELY WITH POTABLE WATER RxB RxC RxD RxE RxF RxG 42 2p 43 3p 42 0p 38 3p 36 0p 40 1p 37 5p 41 0p 422 sa 420 3383 4010 375 36 0p 33 4p 33 0p 37 5p 36 1p 42 5p CAPACITANCE READINGS OF THE DIFFERENT TRANSMITTER RECEIVER ELECTRODE PAIRS AT THE BOTTOM MEASUREMENT PLANE FOR THE LABORATORY SCALE RIG FILLED COMPLETELY WITH AIR 129 APPENDIX E MI CROCONTROLLER PROGRAM CODE FOR 8 ELECTRODE FREQUENCY DI VISI ON MULTI PLEXED MPEDANCE TOMOGRAPHY SYSTEM INCLUDE lt PIC1684 SRC gt START BSF STATUS RPO select page 1 CLRF TRISB set PORTB to all outputs BCF STATUS RPO select page 0 CLRF PORTB clear all outputs on PORTB LOOP MOVLW B 00000000 MOVWF PORTB NOP NOP MOVLW B 00000000 MOVWF PORTB NOP NOP MOVLW B 10000000 MOVWF PORTB NOP NOP MOVLW B 10100000 MOVWF PORTB NOP NOP MOVLW B 11100010 MOVWF PORTB NOP NOP MOVLW B 11101010 MOVWF PORTB NOP NOP MOVLW 01111010 MOVWF PORTB NOP NOP MOVLW 01111010 MOVWF PORTB NOP NOP MOVLW B 00111011 MOVWF PORTB NOP NOP MOVLW B 00011011 MOVWF PORTB NOP N
25. The user can specify whether to sample from the upper or lower system It also provides the feature to automatically save frames of data to a text file at a specified rate so that the drift of the voltages can be analysed unit curunit interface uses Windows Messages SysUtils Classes Graphics Controls Forms Dialogs StdCtrls ExtCtrls Grids Buttons TeEngine Series TeeProcs Chart Spin type TCurMain class TForm Paneli TPanel Panel2 TPanel Label5 TLabel Label3 TLabel CurCapTable TStringGrid CurResTable TStringGrid Label4 TLabel CurSystemSelect TRadioGroup CurTimer TTimer CurExit TSpeedButton DriftTimer TTimer Calibrate TSpeedButton StartMonitor TSpeedButton DriftSave TSaveDialog Label TLabel RecordRate TSpinEdit SampleNumber TLabel procedure CurExitClick Sender TObject procedure FormShow Sender TObject procedure CurTimerTimer Sender TObject procedure DriftTimerTimer Sender TObject procedure CalibrateClick Sender TObject procedure StartMonitorClick Sender TObject private Private declarations I file to where frames of data are saved for analysis driftfile textfile driftfilename string sampnumber integer tempreadings array 1 128 of double to temporarily store the readings obtained from the sampler 1 process the results and 1 display in table procedure ProcessResults public Public declarations pr
26. Value WriteLn tempfile Weight factor floattostr weightfactor save the network performance results to the file WriteLn tempfile Threshold error 4floattostr RecTrainNetwork GetThresholdError WriteLn tempfile Sum void error floattostr RecTrainNetwork GetSumVoidError WriteLn tempfile Water error floattostr RecTrainNetwork GetWaterError WriteLn tempfile Gravel error floattostr RecTrainNetwork GetGravelError WriteLn tempfile Air error floattostr RecTrainNetwork GetAirError WriteLn tempfile Number of hidden for best inttostr RecTrainNetwork GetNumBest CloseFile tempfile end save the data used to pre process the training data to a text file so that future tests can be done without having to load the training database real time captured data has to be pre processed according to the same rules as the training data that was used to produce the neural network being used for the real time test procedure TRecMain Savenetworkpreprocessing1Click Sender TObject var tempfile textfile filename string tempair tempgrav Boolean i integer begin filename c TestData preproc inttostr testnumber txt AssignFile tempfile filename ReWrite tempfile keep a record of the mean of each input voltage for i 1 to 128 do WriteLn tempfile meanrec i keep a record of the standard deviation of each input voltage for i 1 to 128 do WriteLn tempfile stddevrec i
27. and so on By using the interleaving data collection protocol a smaller separation between the electrode planes can be achieved 10 As mentioned previously this will improve the performance of the cross correlator especially for flows that are highly skewed 10 The disadvantage of this technique is that the data acquisition rate is effectively half 12 Various factors determine the selection of the spacing between the two systems If the systems are close then the cross correlation has a well defined and large correlation peak enabling accurate measurement of the transit delay In addition the cross covariance between the two planes will be higher since the properties of the turbulence would have remained fairly consistent from the upstream to the downstream sensor 50 However in systems where the sensors are not point electrodes such as capacitance tomography systems it is difficult to define what the sensor 90 separation should be since the length of the electrodes are comparable to the system separation Consequently a large separation is required in these situations to reduce this uncertainty Between these two extremes an optimal value for system separation exists and has been shown to be between three to four pipe diameters 50 For the current application it was deemed impractical to separate the two planes by three to four pipe diameters the maximum possible separation was 1 4 pipe diameters Therefore alternate techniques
28. begin if StepperConfig AirBubblePosition ItemIndex 0 then tempsteps StepperConfig airsteps else tempsteps StepperConfig airsteps end case StepperConfig AirDrive Value of 1 begin drivetrps StepperConfig airrps drive1steps tempsteps end 2 begin drive2rps StepperConfig airrps drive2steps tempsteps end 3 begin drive3rps StepperConfig airrps drive3steps tempsteps end end end if StepperConfig gravelrps 0 then begin if StepperConfig FlowDirSelect lItemIndex 0 then begin if StepperConfig GravelBubblePosition ItemIndex 0 then tempsteps StepperConfig gravelsteps else tempsteps StepperConfig gravelsteps end else begin if StepperConfig GravelBubblePosition ItemIndex 0 then tempsteps StepperConfig gravelsteps else tempsteps StepperConfig gravelsteps end case StepperConfig GravelDrive Value of 1 begin drive1rps StepperConfig gravelrps drive1steps tempsteps end 2 begin drive2rps StepperConfig gravelrps drive2steps tempsteps end 3 begin drive3rps StepperConfig gravelrps drive3steps tempsteps end end end Stepper MoveFixedDist drive1rps drive2rps drive3rps drive1 steps drive2steps drive3steps end initialise the stepper motor variables procedure TRealMain Stepperlnit begin drivetrps 0 drive2rps 0 drive3rps 0 drive1steps 0 drive2steps 0 drive3steps 0 es RESULT VERIFICAT
29. begin outstring TxF end 7 begin outstring TxG end 8 begin outstring TxH end end while col lt 8 do begin outstring outstring floattostrf temprestable row col ffGeneral 5 4 inc col end WriteLn tempfile outstring end CloseFile tempfile end end rM WEIGHT INITIALISATION PROCEDURES allow the user to set the weight factor procedure TRecMain Initialisenetworkweights 1 Click Sender TObject begin RecSetWeightFactor Visible true RecWeightFactor Text floattostr weightfactor end user decides not to set the weight factor procedure TRecMain RecWeightCanClick Sender TObject begin RecSetWeightFactor Visible false end randomly initialise the weights in the range from 0 5 to 0 5 multiplied by the weight factor procedure TRecMain RecWeightOKClick Sender TObject var row col upperindex integer begin weightfactor strtofloat RecWeightFactor Text upperindex 128 if RecSetParam GetNumHidden gt 0 then begin I if there is a hidden layer of neurons then initialise the weights for the hidden layer upperindex RecSetParam GetNumHidden for row 0 to 128 do begin for col 0 to RecSetParam GetNumHidden 1 do begin HiddenWeight row col weightfactor Random 0 5 DiffHiddenWeight row col 0 depending on the type of network training also II initialise whatever additional weight matrices a
30. changeplus maxdelta OutputWeight rowindex colindex OutputWeight rowindex colindex Sign DiffOutputWeight rowindex colindex OutputDeltaWeight rowindex colindex OldDiffOutputWeight rowindex colindex DiffOutputWeight rowindex colindex end else if OldDiffOutputWeight rowindex colindex DiffOutputWeight rowindex colindex 0 then begin if opposite sign then crossed valley so reduce the amount by which the weights are changes and return without changing weights OutputDeltaWeight rowindex colindex Max OutputDeltaWeight rowindex colindex changeminus mindelta OldDiffOutputWeight rowindex colindex 0 end else begin I previously crossed valley so can now make cahnges to I weights OutputWeight rowindex colindex OutputWeight rowindex colindex 171 Sign DiffOutputWeight rowindex colindex OutputDeltaWeight rowindex colindex OldDiffOutputWeight rowindex colindex DiffOutputWeight rowindex colindex end DiffOutputWeight rowindex colindex 0 end end end adjust the network weights using gradient descent procedure TRecTrainNetwork AdjustWeights var rowindex colindex upperindex integer tempweight double begin upperindex 128 if numhidden gt 0 then begin upperindex numhidden I for a double layer network update the hidden layer weights for rowindex 0 to 128 do begin for colindex 0 to numhidden 1 do begin tempweight
31. col lt 255 do begin tempreadings index result col inc row inc index if row 8 then begin row 0 inc col 9 end else inc col end 11 display the results in table taking care to sort the data according to the multiplexor inputs on the synchronous detector cards index 1 for overall 1 to 16 do begin case overall of 1 begin row 4 tresult capacitance end 2 begin 4 tresult resistance 3 begin 3 tresult capacitance end 4 begin row 3 tresult resistance end 5 begin row 2 tresult capacitance end 6 begin row 2 tresult resistance end 7 begin TOW 1 tresult capacitance end 8 begin row 1 tresult resistance end 9 begin row 5 tresult resistance end 10 begin row 5 tresult capacitance end 11 begin row 6 tresult resistance end 12 begin row 6 tresult capacitance end 13 begin row 7 tresult resistance end 14 begin row 7 tresult capacitance end 15 begin row 8 tresult resistance end 16 begin row 8 tresult capacitance end end for col 1 to 8 do begin S Format 2 3f double tempreadings index if tresult capacitance then CurCapTable Cells col row 5 else CurResTable Cells col row s inc index end end end NI VOLTAGE MONITORING PROCEDURES each tim
32. else airtopposition 0 25 airacceleration power mintottime 2 end else begin during acceleration if time lt airfinaccel then airtopposition 0 5 airacceleration power time 2 during constant velocity else if time lt airfinaccel airtimeflat then airtopposition 0 5 airfinaccel tempairrps time airfinaccel tempairrps during deceleration else if time lt airtottime then airtopposition airfinaccel tempairrps airtimeflat tempairrps 0 5 airacceleration power airtottime time 2 else airtopposition airfinaccel tempairrps airtimeflat tempairrps end convert revolutions into metres by multiplying by the circumference of the pulley airtopposition airtopposition aircircumference end else airtopposition 0 calculate the position of the gravel bubble if tempgravrps gt 0 then begin same calculations are performed as for the air bubble except the gravel bubble parameters are used gravfinaccel tempgravrps gravacceleration gravtimeflat gravelrevs tempgravrps gravfinaccel gravtottime gravfinaccel 2 gravtimeflat mintottime 2 sqrt gravelrevs gravacceleration if gravfinaccel gt 0 5 mintottime then begin if time lt 0 5 mintottime then gravtopposition 0 5 gravacceleration power time 2 else if time lt mintottime then gravtopposition 0 25 gravacceleration power mintottime 2 0 5 gravacceleration power mintottime time 2 else gravtopposition 0 25 gra
33. end I load the water only endpoints to provide calibration Il for the end of the training data for traincalib 1 to numtrainregions do begin for trainversions 1 to trainnumvrs do begin index trainstopindex traincalib trainversions 1000 TrainInpTable FindKey index for col 1 to 128 do begin 51 Voltage 4inttostr col trainstopvalues traincalib col trainstopvalues traincalib col TrainInpTable FieldByName s1 Value end end end I load the water only testpoints to provide a calibration for the start of the test data for testcalib 1 to numtestregions do begin for testversions 1 to testnumvrs do begin index teststartindex testcalib testversions 1000 TestinpTable FindKey index for col 1 to 128 do begin 51 Voltage inttostr col teststartvalues testcalib col teststartvalues testcalib col TestInpTable FieldByName s1 Value end end end I load the water only testpoints to provide a calibration for the end of the test data for testcalib 1 to numtestregions do begin for testversions 1 to testnumvrs do begin index teststopindex testcalib testversions 1000 TestInpTable FindKey index for col 1 to 128 do begin 51 Voltage 4inttostr col teststopvalues testcalib col teststopvalues testcalib col TestInpTable FieldByName s1 Value end end end for row 1 to 5 do for col 1 to 128 do begin trainstartvalues row col trai
34. pulley resulted in the correct positioning of the bubbles e Two systems were designed to operate in the laboratory scale rig at the same time This feature is used to simulate two different bubbles moving at different velocities through the measurement sections at the same time 82 3mm ANODISED ALUMINIUM PLATING 140mm DIAMETER POLYACETAL PLASTIC PULLEY STEPPER MOTOR DETAIL 3mm O RING DRIVE BELT STEPPER 3mm ANODISED ALUMINIUM PLATING 20mm PVC PIPE AIR BUBBLE POLYACETAL PLASTIC MOTOR MOUNT 20mm STEEL SQUARE TUBE 20mm PVC PIPE POSITIONING BASE DETAIL when the pvc pipe is in one of the standard holes then the position of the bubble corresponds ge ra PLASC to one of the standard PULLEY positions 6mm PERSPEX PULLEY POSITIONING BASE FIGURE 7 2 DIAGRAM OF THE FLOW SIMULATION APPARATUS WITH INCLUDED DETAIL OF THE MOTOR MOUNT AND POSITIONING BASE 83 FIGURE 7 3 PHOTOGRAPH OF THE FLOW SIMULATION APPARATUS STANDING NEXT TO THE LABORATORY SCALE RIG FIGURE 7 4 PHOTOGRAPH OF THE FLOW SIMULATION APPARATUS FOR THE SPECIFIC TEST OF A 0 04 VOLUME FRACTION GRAVEL AND A 0 04 VOLUME FRACTION AIR BUBBLE PASSING THROUGH THE PIPELINE AT THE SAME TIME Figure 7 3 is a photograph of the flow simulation apparatus standing next to the rig and figure 7 4 is a photograph of both flow simulation systems before an air gravel water test The stepper motors are control
35. the gravel bubble is correctly detected as gravel when it first enters the measurement volume In terms of the neural network reconstruction the changes introduced by a gravel bubble are not as defined as those introduced by an air bubble This is because gravel has a slightly higher dielectric constant than air Subsequently as the air bubble enters the measurement volume the minor changes introduced are probably comparable with those of a gravel bubble and it is detected as gravel However once the air bubble is completely within the measurement volume the changes introduced are greater than those of a gravel bubble and it is correctly detected as air 104 Secondly tests revealed that the reconstruction algorithm was unable to consistently differentiate between the gravel and air phases Although this is not evident in figure 7 18 on page 104 air bubbles would sometimes be detected only as gravel and gravel bubbles would be detected only as air A possible explanation for this is that the capacitance readings are not accurate enough to distinguish between the air and gravel phases since this separation is based primarily on the difference in their dielectric constants Further the gravel volume fraction measured by the volume fraction predictor consists of a continuous output offset Superimposed on this offset are the changes resulting from the introduction of a bubble This volume fraction also contains a frequency component By performing
36. 04 volume fraction air bubble is 3 6m s In contrast the weight of the 0 04 volume fraction gravel bubble limits the maximum upward flow velocity to 1 4m s In addition the drive belt starts to slip on the pulley for the higher velocity tests This slippage obviously limits the accuracy of a dynamic verification algorithm since the position of the bubble is not guaranteed This is examined in greater detail in the following section It should be noted that the current system is unable to measure the velocity of the water component in a three phase air gravel water mixture Clearly this could be a problem for certain applications of this technology where it is important to know the individual velocities of all the components in a multi phase mixture However in terms of the intended application of this system it is only really the velocity of the gravel component that is critical Subsequently this limitation was largely ignored although future research should address this problem in order to apply the technology to a wider range of applications Because of this limitation no attempt was made to simulate the movement of water through the measurement sections 7 2 2 Verifying dynamically acquired data The neural network reconstruction software includes a Real time sampling form which allows one to test the dynamic performance of the dual plane impedance tomography system The commented program code for this unit can be found in Appendix L o
37. 1 of C output encoding This ensures that the sum of the three dummy output variables always equals 1 and are therefore valid posterior probabilities function TRecTrainNetwork SoftMax numout integer wat double grav double air double double var sum double begin sum exp wat exp grav exp air if numout 1 then SoftMax exp wat sum else if numout 2 then SoftMax exp grav sum else SoftMax exp air sum end calculate the derivative of the hyperbolic tangent at a specific point function TRecTrainNetwork CalcDeriv a extended extended begin CalcDeriv 1 intpower a 2 end return the sign of the argument function TRecTrainNetwork Sign arg extended extended begin if arg gt 0 then Sign 1 else if arg lt 0 then Sign 1 else Sign 0 tM WEIGHT ADJUSTMENT PROCEDURES adjust the network weights using Resilient back propagation procedure TRecTrainNetwork AdjustRPRop var colindex rowindex indexend integer begin indexend 128 if numhidden gt 0 then begin indexend numhidden 1 adjustments to hidden layer for colindex 0 to numhidden 1 do begin for rowindex 0 to 128 do begin if OldDiffHiddenWeight rowindex colindex DiffHiddenWeight rowindex colindex gt 0 then begin Il if same sign then increase amount by which change the weights since going down hill HiddenDeltaWeight rowindex colindex Min HiddenDeltaWe
38. 13 on page 48 provides an example of a 4 electrode system illustrating the problem of the square wave harmonics mixing in the multiplication stage of the synchronous detector Specifically the frequency of transmitter A is 1kHz and the frequency of transmitter B is 1 81kHz The signal at the output of receiver B consists of the frequency harmonics of both transmitters as shown When multiplied by the reference for transmitter B the sum and difference frequency components are generated The frequency components below 1kHz are shown in figure 5 13 Specifically the 50Hz component is the difference between the 9 and 5 harmonic of transmitter A and transmitter B respectively the 430Hz component is the difference between the 5 and 3 harmonic of transmitter and transmitter B respectively and the 810Hz component is the difference between the fundamental frequencies of both transmitters Since the 50Hz component is within the passband of the system it will not be filtered and will appear at the synchronous detector output as ripple Although this is a manufactured example it does illustrate the complication of using square waves in a FDM impedance tomography system It is also shown in figure 5 13 that if the bandwidth of the received signal was limited to approximately 8kHz then there would be no output ripple if an ideal low pass filter was used 47 TRANSMITTER A OUTPUT FREQUENCY SPECTRUM TxB kHz 1 3 5 7 9 TRANSMITTER B OUTPUT
39. 138 142 152 202 204 211 viii LI ST OF ILLUSTRATIONS FIGURES 2 1 2 2 2 3 2 4 3 1 3 2 3 3 4 1 4 2 4 3 44 4 5 4 6 5 1 5 2 5 3 5 4 5 5 5 6 5 7 Diagram illustrating the first two stages of the standard capacitance tomography data collection protocol showing for each stage a cross section of the pipeline and the capacitance plates mounted around the external perimeter of the pipeline Diagram of the 28 capacitance measurements in an 8 electrode standard capacitance tomography system Block diagram of the transducer circuitry for a standard AC based capacitance tomography System where each electrode is connected to its own transducer circuitry Diagram illustrating the four electrode adjacent pair measurement protocol for resistance tomography on a pipeline cross section where the electrodes are embedded in the pipeline wall Diagram illustrating the use of a single layer feed forward neural network to perform a two phase air seawater image reconstruction of the pipeline contents Diagram illustrating the use of a single layer feed forward neural network with 1 of C output encoding to perform a three phase air gravel seawater image reconstruction Diagram illustrating the use of a double layer feed forward neural network to perform a three phase air gravel seawater volume fraction prediction Diagram of a pipeline cross section where the linkages between the electrodes de
40. 16 ELECTRODE IMPEDANCE TOMOGRAPHY SYSTEM 6 1 Analysis of Capacitance and Conductance Measurement Drift 6 2 Training and Test Database Generation 6 3 Neural Network Reconstruction Results 6 3 4 Two phase air water reconstruction results 6 3 2 Three phase air gravel water reconstruction results 66 67 69 73 73 75 vii T DYNAMIC SITUATION RESULTS FOR THE DUAL PLANE 16 ELECTRODE IMPEDANCE TOMOGRAPHY CROSS CORRELATION SYSTEM 7 1 7 2 7 3 7 4 Cross correlation Flowmeter Fundamentals Real time Sampling and Reconstruction Software Description 7 2 4 Design of a flow simulation apparatus 7 2 2 Verifying dynamically acquired data 7 2 8 Cross plane interference problem Two phase Air Water Reconstruction Results 7 3 4 Real time volume fraction prediction and image reconstruction results 7 3 2 Reconstruction results for an air plug passing through the measurement section 7 3 8 Reconstruction results for a homogenous bubble flow 7 3 4 Cross correlation velocity prediction results Three phase Air Gravel Water Reconstruction Results 7 4 4 Real time volume fraction prediction and image reconstruction results 7 4 2 Cross correlation velocity prediction results 7 4 8 Limitations of three phase reconstruction algorithms 8 CONCLUSIONS 9 RECOMMENDATIONS FOR FUTURE DEVELOPMENT 10 REMAINING SYSTEM ISSUES THAT STILL NEED TO BE ADDRESSED LIST OF REFERENCES APPENDIX A APPENDIX B APPENDIX C APPENDIX D APPENDIX E
41. 256 MAXWORDS 128 Genral purpose constants MAXLINE 76 BASEP 0300 BACKSPACE 48 ENTER 13 LINEFEED 10 SPACE 32 27 DELETE 126 CTRLK 11 TERMINATOR 0 CharBuffer array 0 257 of char WordBuffer array 0 128 of word TDirection CW CCW specify motor direction TRequestData MotorPos Velocity specify what data requesting from motor TStep class private Address word base address Command CharBuffer command to AT6400 Error integer error integer code word_high word used for status sampling word_low word fast_status longint send command to AT6400 function SendAT6400Block Address word Command CharBuffer word receive command from AT6400 function RecvAT6400Block Address word var Response CharBuffer word convert string command into appropriate data format procedure ProcessCommand temp string Il initiate an update status procedure RequestStatus get status from AT6400 procedure ReadStatus var status high status low word var status longint set what specific information want to retrieve procedure SetPointer status_offset integer public load an initialisation file which contains a sequence of commands to initialise the stepper motors procedure Initialise NewAddress word readfilename string general purpose procedure which sets the velocity and direction of both drives simultaneously p
42. 8 begin top 1 bottom 8 end 9 begin top 2 bottom 7 end end for row top to bottom do begin tempphasepred water tempphasedes water if two phase reconstruction if RecLoadDBase GetNumOutputs 88 then begin RecLoadDBase GetPhases containsair containsgravel if network output less than 0 then pixel is water else pixel is other phase if testoutput index gt 0 then begin if containsair then tempphasepred air else tempphasepred gravel 176 end if OutputData testindex index 1 then begin if containsair then tempphasedes air else tempphasedes gravel end inc index end else three phase prediction else begin if testoutput index 2 gt testoutput index 1 and testoutput index 2 gt testoutput index then pixel is air tempphasepred air else if testoutput index gt testoutput index 1 and testoutput index gt testoutput index 2 then pixel is water tempphasepred water else tempphasepred gravel if OutputData testindex index 1 then tempphasedes water else if OutputData testindex index 1 1 then tempphasedes gravel else tempphasedes air inc index 3 end tempoutputdes col row tempphasedes tempoutputpred col row tempphasepred end end RecTestDesired Desired tempoutputdes RecTestNetwork Desired tempoutputpred force a repaint to show changes if not usecurrentsamples then RecTestDesired Repaint RecTestNetw
43. 9 Rz z 2 5 E a 9 Iz aa SCREEN CAPTURE OF LOADUNIT PAS 204 S 5 c 9 3 5 c 9 wn SCREEN CAPTURE OF PARAMUNIT PAS 205 Train the neural network Train 4 neural network specified network training Network training progress rk training validation arch results SCREEN CAPTURE OF NETWORKUNIT PAS 206 Test the network yolume fraction prediction Volume fraction predictions For t 652 Air volume prediction Gravel volume prediction Water volume prediction SCREEN CAPTURE OF TESTUNIT PAS 207 Real time impedance tomography reconstruction Reconstruction results Configuration 2 Verification SCREEN CAPTURE OF REALUNIT PAS 208 Stepper motor configuration Overall controller configuration Individual motor configuration Air bubble motor Additional bubble motor E m SCREEN CAPTURE OF STEPCONFIG PAS 209 Dynamic system verification onfiguration Volume Fraction plots For top and bottom tem Test case results SCREEN CAPTURE OF DYNUNIT PAS 210 APPENDI X O PAPER PUBLI SHED MEASUREMENT SCI AND TECHNOLOGY JOURNAL ON AUTHOR S PREVI OUS RESEARCH 211
44. 96192 202052 091992 056242 66942 952672 SeCpGC 08 92 21262 122452 560957 202042 202052 ANAS 52060 282 LISHE 546142 vlt pz 52060 65942 95262 GOSS 28492 8925 02 892582 2 9328932 928992 0042 2119820 529832 819542 886902 ANGERS 85482 88092 S3G 6pc 29 S 29 2 2522040 Ipee 28 2 949280 881092 S68tvc sep cep ow e wo 122142 22192 4 9442 256452 9688 02 9306292 3906542 261242 18552 159 JUBANI 34 10 51014221 eunjo 222692 958942 316892 958942 986092 469292 21462 MrrEST 954662 vct Gc 8 09842 480480 v603pPC tcO6PC BOE RZ 20040 509 2 6SG pz2 3G bpC 522092 89260 ScOSpC 8 2 004 2 602192 243090 4940 954880 809802 22092 PIEZ S0202 5689152 LISA 396860 9189002 6996070 209682 80952 POS 250800 022802 21462 62287 26 822 919822 sorer z 9992527 964622 220922 62 axuepaduir sbuipea abeyo waysAs woog 4 do 6 andano au asegejep 0 pajdwes ppy Jr 202 SCREEN CAPTURE OF ADDUNIT PAS 0 supe a4 abeyo wajsdis dydeibowo ejep ajdwes djsnonunu 203 SCREEN CAPTURE OF CURUNIT PAS APPENDI X N SCREEN CAPTURES OF NEURAL NETWORK PROGRAM x Testing databa Training database v D F 5
45. B 10111100 MOVWF PORTB NOP NOP MOVLW B 10111100 MOVWF PORTB NOP NOP MOVLW B 11111110 MOVWF PORTB NOP NOP N N N N N N N N N N N N N N N N N N N N N N MOVLW B 11011110 MOVWF PORTB NOP NOP MOVLW B 01011110 MOVWF PORTB NOP NOP MOVLW 01010110 MOVWF PORTB NOP NOP MOVLW B 000001 11 MOVWF PORTB NOP NOP MOVLW B 000001 11 MOVWF PORTB NOP NOP MOVLW B 100001 11 MOVWF PORTB NOP NOP MOVLW B 10100111 MOVWF PORTB NOP NOP MOVLW 11100001 MOVWF PORTB NOP NOP MOVLW 11100001 MOVWF PORTB OP OP MOVLW B 01110001 MOVWF PORTB OP OP MOVLW B 01110001 MOVWF PORTB OP OP MOVLW 00110000 MOVWF PORTB OP OP MOVLW B 0001 1000 MOVWF PORTB OP OP MOVLW 10011000 MOVWF PORTB OP OP MOVLW 10011000 MOVWF PORTB OP OP MOVLW 11001010 MOVWF PORTB OP OP MOVLW 11001010 MOVWF PORTB OP OP MOVLW B 01001110 MOVWF PORTB OP OP MOVLW B 01101110 MOVWF PORTB OP OP MOVLW B 00101111 N MOVWF PORTB OP 132 NOP MOVLW B 00101111 MOVWF PORTB NOP NOP MOVLW B 10111111 MOVWF PORTB NOP NOP MOVLW B 10110111 MOVWF PORTB NOP NOP MOVLW B 11110101 MOVWF PORTB NOP NOP MOVLW 11010101 MOVWF PORTB NOP NOP MOVLW B 01010101 MOVWF PORTB NOP NOP MOVLW B 01010101 MOVWF PORTB GOTO LOOP 133 APPENDI X F MI CROCONTROLLER PROGRAM CODE FOR 16 E
46. DC voltage is produced whose magnitude is proportional to the capacitance between receiver A and transmitter A The third synchronous detector multiplies the received signal by the reference for transmitter B After low pass filtering a DC voltage is produced proportional to the conductance between receiver A and transmitter B This process is replicated for the remaining transmitters to measure both the capacitance and conductance between receiver A and each of the four transmitters All this electronic circuitry is then replicated for receiver B C and D 20 Voltages proportional to the capacitance and conductance of all the different transmitter receiver electrode pairs are continuously available on the outputs of the low pass filters and no switching is required In addition since each transmitter receiver electrode pair has its own detection path each signal can be amplified individually for optimum dynamic range and offset This is in contrast to standard tomography systems where one set of electronics is used to perform all the measurements and consequently must be tuned to some general setting that is a compromise for all electrode pairs Since the readings are continuously available the frame rate of this system is only limited by the cutoff frequency of the low pass filters Subsequently this technique can achieve far higher frame rates than even the most advanced TDM impedance tomography systems One of the drawbacks of the FDM impedance
47. DETECTOR FOR A 4 ELECTRODE SYSTEM NOTE THAT THE HORIZONTAL SCALE OF THE MULTIPLIER OUTPUT FREQUENCY SPECTRUM IS DIFFERENT TO THAT OF THE TRANSMITTER AND RECEIVER FREQUENCY SPECTRUMS 48 It was decided that a better understanding of these frequency mixing principles could be obtained through the simulation of the system operation in software Using the specified eight transmitter frequencies the simulator calculates the multiplier output frequencies of each synchronous detector in a channel Since the output frequency components are independent of the particular receiver the simulation is only performed for a single receiver Initially the magnitudes of the frequency components within the system were not incorporated Instead frequencies higher than some maximum threshold were simply ignored Although the initial results of this simulation were promising it would fail to predict the generation of certain frequency components resulting from the mixing of those higher frequency harmonics that had been ignored It was therefore necessary to include the magnitude of the frequency components in the simulation to ensure that all possible output frequency components are correctly predicted for the specified set of transmitter frequencies The program code for the modified simulator is included in Appendix J on page 138 and operates as follows 1 Calculate the magnitude of the received frequency components All even and odd harmonics up to and including
48. DURING TESTING ONLY EVERY 50 FRAME IS INCLUDED AS BEFORE THE BLACK IS THE WATER PHASE THE WHITE IS THE AIR PHASE AND THE LIGHT GREY REPRESENTS THE REGION OUTSIDE THE PIPELINE 95 190mm bubble 90mm bubble 60mm bubble desired 9 z o E 3 gt 5 o 5 a c o E 0 005 0 01 0 015 0 02 0 025 0 03 0 035 0 04 0 045 0 05 0 055 0 06 0 065 0 07 0 075 actual volume fraction FIGURE 7 10 EFFECT OF THE BUBBLE LENGTH ON THE VOLUME FRACTION MEASUREMENT ACCURACY Figure 7 10 compares the actual volume fractions to the measured volume fractions for each of the three bubble sizes As can be seen the system accurately measures the volume fractions of the 190mm bubbles However as the length of the bubble decreases so the accuracy of the corresponding volume fraction measurement decreases and the readings deviate further from the desired plot The accuracy of these results could possibly be improved through the inclusion of shorter bubbles in the training database although this assumption has not been verified Table 7 2 compares the dynamic volume fraction error for each of the three bubble sizes at the top and bottom measurement planes TABLE 7 2 COMPARISON BETWEEN THE AVERAGE DYNAMIC VOLUME FRACTION ERRORS FOR THE 60mm 90mm AND 190mm BUBBLE RESPECTIVELY ALL RESULTS REPRESENT THE ABSOLUTE ERROR AND ARE ROUNDED OFF TO TWO SIGNIFICANT DIGITS DYNAMIC VOLUME FRACTION ERROR FOR SPECIFIC BUBBLE LE
49. FREQUENCY SPECTRUM Voltage proportional to conductance between transmitter B and receiver B The following is a sequence of frequency spectrum plots for the simplified case of a 4 electrode system kHz 5 43 1 The plots depict the problem with using square waves as frequency generators RECEIVER B OUTPUT FREQUENCY SPECTRUM the harmonics mix in the multiplier to i i i i i produce low frequency ripple in this case i i H i 50Hz at the low pass filter output Since this i i ripple is within the desired bandwidth of the system 200frames s it cannot be filtered out Consequently the accuracy with which the desired DC component is measured at the multiplier output is limited by this ripple This ripple can be reduced by including a bandpass filter in the receiver This is used to block any DC drift component and attenuate the high frequency harmonics that mix to form the low frequency ripple at the multiplier output For example if the cutoff was set to approximately 8 2 in the receiver filter then there would be no 50Hz component Y at the multiplier output for an ideal low pass i filter while at the same time ensuring reasonable transmission of the square waves 4 050 200 430 810 low pass filter response MULTIPLIER OUTPUT FREQUENCY SPECTRUM FIGURE 5 13 EXAMPLE OF THE SQUARE WAVE HARMONICS MIXING IN THE MULTIPLICATION STAGE OF THE SYNCHRONOUS
50. Label8 TLabel FiltCap2 TEdit FiltRes2 TEdit Label9 TLabel Label10 TLabel FiltRes1 TEdit FiltCap1 TEdit MagThresh TEdit Label11 TLabel Bevel1 TBevel procedure LoadFregsClick Sender TObject procedure FormShow Sender TObject procedure TestMultClick Sender TObject procedure PBILStartClick Sender TObject procedure MultiSearchPBILClick Sender TObject private Private declarations xtal array 1 23 of double standard crystal frequencies AvailFreq array 1 2000 of double available frequencies RecvFreq array 1 2000 of double received frequencies MultOutFregs array 1 200000 of double multiplier output frequencies resrecord array 1 9 1 100 of double record of results FreqMag array 1 801 1 2 of extended record of simulated frequency harmonics and corresponding magnitudes ntrials integer maxgen integer besteverfit double resi double 1 component parameters of the receiver filters double res2 double cap2 double TC1 extended TC2 extended gaint extended gain2 extended temp1 temp2 temp3 temp4 temp5 extended diffrec array 0 7 of double RefFreq clock double RefHarm array 1 51 of double threshold double magnitude threshold all recieved frequency components whose magnitudes are less than the threshold are ignored besteverreal array 1 8 of double function MultOut inputs a
51. OUTPUT LAYER NEURONS VOLTAGE 5 2s 525 aa o9 FIGURE 3 2 DIAGRAM ILLUSTRATING THE USE OF A SINGLE LAYER FEED FORWARD NEURAL NETWORK WITH 1 OF C OUTPUT ENCODING TO PERFORM A THREE PHASE AIR GRAVEL SEAWATER IMAGE RECONSTRUCTION THE STRUCTURE OF THIS NEURAL NETWORK IS VERY SIMILAR TO THE ONE EXAMINED IN FIGURE 3 1 HOWEVER SINCE EACH PIXEL CAN BE ONE OF THREE PHASES INSTEAD OF JUST TWO A 1 OF C OUTPUT LAYER ENCODING IS REQUIRED THIS IS ACHIEVED BY INTRODUCING ADDITIONAL DUMMY OUTPUT NEURONS FOR EACH PIXEL SUBSEQUENTLY THE NUMBER OF OUTPUT LAYER NEURONS IS INCREASED TO 264 SINCE THREE NEURONS ARE REQUIRED FOR EVERY PIXEL EACH OUTPUT NEURON CALCULATES THE WEIGHTED SUM OF THE INPUT LAYER NEURON OUTPUTS AS BEFORE A SOFTMAX ACTIVATION FUNCTION IS THEN APPLIED TO THESE WEIGHTED SUMS AS SPECIFIED IN EQUATION 3 7 THE THREE OUTPUT NEURONS CORRESPONDING TO A PARTICULAR PIXEL ARE COMPARED AND THE PHASE OF THE PIXEL IS SET ACCORDING TO WHICH OF THE THREE OUTPUT NEURONS HAS THE LARGEST OUTPUT IN THIS WAY A THREE PHASE IMAGE OF THE VESSEL CROSS SECTION IS OBTAINED Most tomography systems employ standard reconstruction algorithms to perform image reconstruction of the vessel contents This image is then analysed to extract the required information for the interpretation or control of the corresponding process However improved results can typically be achieved by modifying the reconstruction algorithm to instead perform a parameter
52. OutputData count colindex 2 lt gt 1 then inc totalerrors end else begin 11 predicts pixel is gravel totalgravelpixels totalgravelpixels 1 Il but if not gravel then increment the number of 11 errors if OutputData count colindex 1 lt gt 1 then inc totalerrors end end end two phase prediction else if numoutputs 88 then begin OutputNetwork count colindex tanh weightsum II if network output is less than 0 then predicts that the pixel is water if OutputNetwork count colindex gt 0 then begin if OutputData count colindex lt gt 1 then inc totalerrors if containsair then totalairpixels totalairpixels 1 else totalgravelpixels totalgravelpixels 1 end else begin if OutputData count colindex lt gt 1 then inc totalerrors end end volume fraction prediction else begin if weightsum gt 0 then OutputNetwork count colindex weightsum else OutputNetwork count colindex 0 if colindex 0 then totalairpixels OutputNetwork count colindex else totalgravelpixels OutputNetwork count colindex end end totalwaterpixels 100 totalairpixels totalgravelpixels calculate the errors in the volume fraction predictions for this test case and add to the previous error percentages watererror watererror Abs VolumeData count 2 totalwaterpixels gravelerror gravelerror Abs VolumeData count 1 totalgravelpixels airerror airerror Abs Volume
53. SampHardware StartSample 50 busyreconstruction true RealReconstruction end end else begin do the reconstruction for the final set of 50 frames RealReconstruction RealStart Caption Start testing StepperStop Visible false stoptime Now timediff stoptime starttime calculate the frame rate DecodeTime timediff hours mins secs msecs permin Caption inttostr mins persec Caption inttostr secs permsec Caption inttostr msecs totframes Caption inttostr totalfrms 1 frameinterval mins 60 secs msecs 1000 totalfrms 1 frmrate Caption floattostri 1 frameinterval ffFixed 10 6 Update perform a correlation on this captured data PerformCorrelation end end end this procedure implements the double layer feedforward neural network reconstruction on the real time captured data and places the results in outputstore very similar to the reconstruction algorithm for the test unit procedure TRealMain RealReconstruction var rowindex colindex i integer tempvalue weightsum real weightstore array 1 2 of real numrecon integer begin for numrecon 1 to 50 do begin first the top rig upperindex 128 if there are hidden layer neurons then calculate the output from the hidden layer first if numtophidden gt 0 then begin upperindex numtophidden for colindex 0 to numtophidden do begin if colindex 0 then tophiddenoutputs colind
54. Sender TObject procedure RecLoadReturnClick Sender TObject procedure FormClose Sender TObject var Action TCloseAction procedure FormShow Sender TObject private Private declarations maxepochs integer numhidden integer maximum number of epochs number of hidden layer neurons deltainit real initial delta values for RPROP deltamax real maximum delta values for RPROP learnrate real I learning rate for gradient descent momentum real numfails integer momentum constant for gradient descent maximum number of fails before stop training if using early stopping public Public declarations function GetMaxEpochs integer data access procedures function GetNumHidden integer function GetDeltalnit real function GetDeltaMax real function GetLearnRate real function GetMomentum real function GetNumFails integer function GetUpdateFreq integer 165 end var RecSetParam TRecSetParam implementation uses loadunit mainunit R DFM DATA ACCESS PROCEDURES function TRecSetParam GetMaxEpochs integer begin GetMaxEpochs maxepochs end function TRecSetParam GetDeltalnit real begin GetDeltalnit deltainit end function TRecSetParam GetDeltaMax real begin GetDeltaMax deltamax end function TRecSetParam GetLearnRate real begin GetLearnRate learnrate en
55. Series 0 Clear TopGravelPlot Series 0 Clear BotGravelPlot Series 0 Clear obtain the air and gravel bubble parameters airdiam strtofloat AirDiameter Text airleng strtofloat AirLength Text gravdiam strtofloat GravelDiameter Text gravleng strtofloat GravelLength Text uppersystop strtofloat UpperSystemTop Text uppersysbot strtofloat UpperSystemBottom Text lowersystop strtofloat LowerSystemTop Text lowersysbot strtofloat LowerSystemBottom Text airstartpos strtofloat AirBubbleStartPosition Text gravelstartpos strtofloat GravelBubbleStartPosition Text calculate the volumes of the two measuring regions given that the pipe diameter is 350mm and the electrode length is given 11 by the difference between uppersystop and uppersysbot upperplatevolume pi power 0 175 2 uppersystop uppersysbot lowerplatevolume pi power 0 175 2 lowersystop lowersysbot for testposition 1 to RealMain DesiredFrames Value do begin calculate the estimated time of a particular frame as the frame number divided by the frame rate or multiplied I by the frame interval time testposition RealMain frameinterval 1 calculate the position of the bubbles at this time CalculateDistance time airtopposition gravtopposition if StepperConfig FlowDirSelect ItemIndex 0 then begin airtopposition airstartpos airtopposition gravtopposition gravelstartpos gravtopposition
56. Text floattostr AddInputTomo GravelVolume end AddStatusBar Panels 3 Text Current test floattostr currenttest AddStatusBar Panels 4 Text Total number of samples inttostr InpTable RecordCount 1 if currenttest 1 then AddPrevious Enabled false else AddPrevious Enabled true if AddInputTomo AllowEdit false then begin if currenttest InpTable RecordCount 1 div numframes then AddNext Enabled false else AddNext Enabled true end else AddNext Enabled true if SampHardware GetSampleStatus then AddStatusBar Panels 2 Text SAMPLING else AddStatusBar Panels 2 Text READY end override the message handling procedure to determine whether the sampling hardware has completed If it is the stop notification then stop the hardware 145 and process the results else handle the message in the default way procedure TAddToDatabase WndProc var TheMsg TMessage begin if TheMsg Msg notifvalue then begin Cursor crDefault SampHardware StopSample result ProcessResults end else inherited WndProc TheMsg end when open the form move to the last test case in the database i e the last test taken before the form was closed and display the voltage readings in the table Also display what the desired output for that particular test case was procedure TAddToDatabase FormShow Sender TObject begin InpTable Last OutTable Last get the total number
57. a specific location in the second phase using the bubble placement system A set of readings is then taken corresponding to the capacitance and conductance measurements for this particular configuration The bubble is then moved to a different position and a new set of readings is again taken and recorded This process is repeated until the superposition of all the individual bubble positions effectively covers the vessel cross section 21 30 31 Drawn on the clear acrylic lid is a 10 by 10 matrix representing the pixels used to specify the desired network output The user can identify which pixels are air gravel or seawater simply by looking through the lid from above and setting the corresponding pixel values in the database generation software running on the reconstruction computer This software also records the capacitance and conductance readings corresponding to this particular bubble location and this database is used for training the neural network Neural networks are capable of generalisation for previously unseen bubble configurations Generalisation is described as the ability to predict the correct outputs for new inputs even though they do not form part of the training database 29 To test the generalisation performance of the neural network a validation database was generated This consists of combinations of air bubbles combinations of gravel bubbles and combinations of both air and gravel bubbles which do not form part of the tra
58. airleng else if airtopposition gt uppersystop and airtopposition airleng lt uppersystop then upperairvol pi power 0 5 airdiam 2 uppersystop airtopposition airleng else upperairvol 0 end gravel bubble for bottom measuring ring if gravleng gt lowersystop lowersysbot then begin if gravtopposition gt lowersysbot and gravtopposition lt lowersysbot 0 5 gravdiam then begin tempheight lowersysbot gravtopposition 0 5 gravdiam owergravvol gravhemivol pi tempheight power 0 5 gravdiam 2 1 3 pi power tempheight 3 end else if gravtopposition gt lowersysbot 0 5 gravdiam and gravtopposition lt lowersystop then begin tempheight gravtopposition 0 5 gravdiam lowersysbot owergravvol gravhemivol pi tempheight power 0 5 gravdiam 2 end else if gravtopposition gt lowersystop and gravtopposition lt lowersystop 0 5 gravdiam then begin tempheight lowersystop gravtopposition 0 5 gravdiam owergravvol pi tempheight power 0 5 gravdiam 2 1 3 pi power tempheight 3 pi lowersystop lowersysbot tempheight power 0 5 gravdiam 2 nd se if gravtopposition gt lowersystop 0 5 gravdiam and gravtopposition gravleng 0 5 gravdiam lt lowersysbot then begin owergravvol pi lowersystop lowersysbot power 0 5 gravdiam 2 nd se if gravtopposition gravleng 0 5 gravdiam gt lowersysbot nd gravtopposition gravieng 0 5 gravdiam lt lowersysbot 0 5 gravdi
59. and increment newindex if OutputData origindex 0 lt gt 0 and RecDesiredRecon ItemIndex 0 or OutputData origindex 0 lt gt 1 and RecDesiredRecon ItemIndex 1 then begin for colindex 0 to 128 do InputData newindex colindex InputData origindex colindex for colindex 0 to numoutputs 1 do OutputData newindex colindex OutputData origindex colindex for colindex 0 to 2 do VolumeData newindex colindex VolumeData origindex colindex inc newindex if origindex traintotal then inc temptraintotal else inc temptesttotal end else simply increment origindex RecLoadProgressBar Steplt end InputData Copy InputData 0 newindex OutputData Copy OutputData 0 newindex OutputNetwork Copy OutputNetwork 0 newindex VolumeData Copy VolumeData 0 newindex traintotal temptraintotal testtotal temptesttotal close the form procedure TRecLoadDBase RecLoadReturnClick Sender TObject begin RecLoadStatus Visible false RecMain Savenetworkpreprocessing1 Enabled true Close end when show form load all the possible database names into the combo boxes a user then chooses which databases contain the training and testing data The previous settings used are also loaded from the registry procedure TRecLoadDBase FormShow Sender TObject var templist TStringList i integer Reg TRegistry KeyGood Boolean begin custload false templist TStringList Create
60. be associated with the propagation velocity of these disturbances Clearly an electrical impedance tomography system does not strictly satisty assumption one Since the electrode length is generally comparable to the plane separation questions are raised as to how exactly the plane separation should be defined 5 For a cross correlation flowmeter to work correctly the signals detected by the top and bottom systems should be a random series These signals should represent a random modulation effect of the fluid in the sensing volume of the sensor In a bubbly gas liquid flow this modulation is achieved through the random positioning of the bubbles In contrast a stratified or annular flow produces no such random effect assuming that the interface is not wavy and consequently the cross correlation flowmeter will not work 12 As an example of the problems associated with assumption three consider the case of a pneumatic conveying System where pellets of a certain material are conveyed along a pipeline using air Here the pellets move along the pipe in a wavelike manner forming slugs It is these slugs that are detected and correlated by the tomography System and not the individual pellets However the velocity of the individual pellets cannot be directly linked to the propagation velocity of the slugs and so the cross correlation algorithm fails to determine the true material velocity 5 9 80 In the current application the volume frac
61. both the top and bottom measurement planes TABLE 7 3 COMPARISON BETWEEN THE AVERAGE DYNAMIC VOLUME FRACTION ERRORS OF THE DIFFERENT NEURAL NETWORK RECONSTRUCTION TECHNIQUES CONSIDERED FOR A THREE PHASE AIR GRAVEL WATER RECONSTRUCTION ALL RESULTS REPRESENT THE ABSOLUTE ERROR AND ARE ROUNDED OFF TO TWO SIGNIFICANT DIGITS TOP PLANE BOTTOM PLANE RECONSTRUCTION AIR VOLUME GRAVEL VOLUME AIR VOLUME GRAVEL VOLUME TECHNIGUE FRACTION ERROR FRACTION ERROR FRACTION ERROR FRACTION ERROR IMAGE RECONSTRUCTION 0 72 1 5 0 68 1 4 SINGLE LAYER VOLUME FRACTION PREDICTOR 0 62 1 5 0 67 1 5 DOUBLE LAYER VOLUME FRACTION PREDICTOR 10 1 5 0 89 1 6 As expected the dynamic volume fraction error of the gravel phase is poorer than that of the air phase 103 na m m Sesa ee 2 e om im o som om 2 om u mw om 4 FIGURE 7 18 SCREEN CAPTURES OF A SIMULATED 0 2m s BUBBLE FLOW WHERE THE 0 04 VOLUME FRACTION GRAVEL BUBBLE IS SITUATED NEAR THE LEFT HAND EDGE OF THE RIG AND THE 0 04 VOLUME FRACTION AIR BUBBLE IS SITUATED NEAR THE RIGHT HAND EDGE OF THE RIG THESE IMAGE RECONSTRUCTIONS WERE PERFORMED ON LINE AND THEIR RESULTS DISPLAYED IN REAL TIME FOR EACH SCREEN CAPTURE THE TOP FRAME REPRESENTS THE RECONSTRUCTION RESULTS OF THE TOP MEASUREMENT PLANE AND THE BOTTOM FRAME REPRESENTS THE RECONSTRUCTION RESULTS OF THE BOTTOM MEASUREMENT PLANE THE SEQUENCE OF SCREEN CA
62. col water then dec FWaterVolume else if FDesired row col gravel then dec FGravelVolume else dec FAirVolume if value water then inc FWaterVolume else if value gravel then inc FGravelVolume else inc FAirVolume FDesired row col value end end end I pixels can also be drawn while the mouse is moving over the pipeline I depending on what mouse button is held down while the mouse moves procedure TTomo MouseMove Sender TObject Shift TShiftState X Y Integer var row col newX newY integer begin row X div 20 col Y div 20 FXPosition row 1 FYPosition col 1 if SsLeft in Shift or ssRight in Shift ssMiddle in Shift then begin if FDesired row col lt gt outpipe and FAllowEdit then begin Il air phase if ssLeft in Shift then begin Canvas Pen Color clWhite Canvas Brush Color clWhite SetDesired row col air end gravel phase else if ssRight in Shift then begin Canvas Pen Color clGray Canvas Brush Color clGray SetDesired row col gravel end water phase else begin Canvas Pen Color clBlue Canvas Brush Color clBlue SetDesired row col water end newX row 20 newY col 20 Canvas Rectangle newX newY newX 20 newY 20 end end end end 151 APPENDIX L NEURAL NETWORK PROGRAM CODE procedure RecWeightOKClick Sender TObject procedure Starttraining1 Click Sender TObject procedure Savenetworkperformanc
63. condition v Perform early stopping to prevent over fitting Sum total of volume fraction errors v Conduct model search Network specific parameters Model search parameters C Use gradient descent 1 Minimum number of hidden layer nodes 100 Specify the initial delta value Maximum number of hidden layer nodes Specify the maximum delta value Increment F Specify the performace refresh rate M v View the training database performance measures Return FIGURE 5 20 SCREEN CAPTURE OF THE NEURAL NETWORK PARAMETER FORM THIS SHOWS A TYPICAL CONFIGURATION FOR A DOUBLE LAYER FEED FORWARD NEURAL NETWORK TRAINED USING RESILIENT BACK PROPAGATION TO PERFORM A VOLUME FRACTION PREDICTION A MODEL SEARCH WILL BE CONDUCTED TO DETERMINE THE OPTIMAL NUMBER OF HIDDEN LAYER NEURONS THE MINIMUM NUMBER OF HIDDEN LAYER NEURONS IS SET AT 5 AND THE MAXIMUM NUMBER OF HIDDEN LAYER NEURONS IS SET AT 100 EARLY STOPPING IS ALSO PERFORMED TO PREVENT OVER FITTING Set the network training parameters opens a form where the desired training parameters are specified These include the type of training algorithm to use whether early stopping should be performed the number of hidden layer neurons for a double layer neural network and so on A model search is simply an iterative search to determine the optimal number of hidden layer neurons in a double layer feed forward neural network and
64. conductance probes are invasive they are designed to be non intrusive so as not to influence the flow parameters 59 Figure 5 5 is a photograph of the conductance probes from above As can be seen in figure 5 4 on page 37 the conductance probes for the transmitters are in electrical contact with the capacitance plates For the receiver a grounded metal ring surrounds the conductance probe In the previous System there existed a leakage current between the conductance probe and the capacitance plate Specifically if there were any moisture in the region of the conductance probe then a leakage current would flow from the conductance probe to the capacitance plate A potential difference exists between the conductance probe and the capacitance plate and therefore a current will flow Since the current through the capacitance is very small any leakage current seriously distorts the results By placing a grounded metal ring around the conductance probe any leakage currents will flow to ground and the problem should be solved 38 OUTER EARTHED SHIELD TO PREVENT EXTERNAL i INTERFERENCE PROJECTED RADIAL SHIELD PASSING BETWEEN THE CAPACITANCE PLATES TO MINIMISE INTER ELECTRODE CAPACITANCES EXTERNAL TO THE VESSEL FIGURE 5 6 DIAGRAM ILLUSTRATING THE OUTER EARTHED SCREEN AND THE PROJECTED RADIAL GUARDS PASSING BETWEEN THE CAPACITANCE PLATES AS A CROSS SECTION OF THE MEASUREMENT SECTION An earthed screen m
65. conductance measurements it is expected that the readings should decrease since air has both a lower permittivity and conductivity than seawater This phenomenon has been explained for the case of multiple transmitters operating in parallel for conductance tomography by Hervieu and Seleghim as follows 42 the presence of a non conducting object in the region of an electrode results in a current deficiency at that particular electrode and a current excess at its neighbours For larger non conductive objects this current deficiency extends to multiple electrodes and although the current excess is still present it is compensated by an overall current reduction resulting from the global reduction in conductivity For the case of capacitance tomography Xie et al explain this situation as follows 47 the introduction of a different phase into the vessel will redistribute the electric flux lines within the vessel Consequently certain detector electrodes may absorb more electric flux lines than before resulting in over shooting effects Subsequently it became clear that the voltage levels in the system could not be tuned with only seawater in the rig as the maximum expected capacitance and conductance readings since the introduction of a bubble could result in the saturation of certain op amp outputs Saturation of the op amp outputs would introduce a further non linearity into the readings making the reconstruction task even more complex To ensure tha
66. current test case if TestOutTable FieldByName AirVolume Value gt 0 then tempair true if TestOutTable Field ByName GravelVolume Value gt 0 then tempgravel true Il if the phases are the correct phases for the desired reconstruction then load the data normally if RecDesirePhase ItemIndex 1 and tempair and not tempgravel Jor RecDesirePhase ItemIndex 2 and not tempair and tempgravel or not tempair and not tempgravel or RecDesirePhase lItemIndex 0 then begin OutputData row 0 TestOutTable FieldByName AirVolume Value OutputData row 1 TestOutTable FieldByName GravelVolume Value end else set a flag so that teh test case can be removed Il at later stage else if RecDesirePhase ItemIndex 1 and tempgravel or RecDesirePhase ItemIndex 2 and tempair then begin OutputData row 0 OutputData row 1 end end 4 4 VolumeData row 0 TestOutTable FieldByName AirVolume Value VolumeData row 1 TestOutTable FieldByName GravelVolume Value VolumeData row 2 TestOutTable FieldByName WaterVolume Value inc row RecLoadProgressBar Steplt Application ProcessMessages end end RecLoadProgress Visible false opened true 11 display a summary of the loading results if TomoSystem 0 then S1 upper else S1 lower RecLoadStatus Caption Loaded inttostr traintotal testtotal datapoints from the s1 system s2 R
67. developed Due to the extremely high parallelism in FDM impedance tomography the component count increases exponentially as the number of electrodes increases Subsequently the current FDM impedance tomography system is based on square wave transmitter outputs instead of sine wave since it is cheaper to perform the multiplication in the synchronous detectors for a square wave system Although the square wave approach achieves good results tests revealed that it is limited in terms of the resolution and accuracy of the reconstruction results specifically for three phase mixtures This is because of the ripple on the outputs of the low pass filters resulting from the harmonics of the square waves mixing in the multipliers A sine wave system should address some of these problems since it is a simple task to select transmitter frequencies for a sine wave system to ensure that no frequency components are generated within the passband of the system vi TABLE OF CONTENTS Acknowledgments Terms of Reference Synopsis List of Illustrations 1 INTRODUCTION 2 TIME DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY 2 1 Standard Capacitance Tomography Measurement Systems 2 2 Standard Conductance Tomography Measurement Systems 3 IMPEDANCE TOMOGRAPHY RECONSTRUCTION ALGORITHMS 3 1 Standard Impedance Tomography Reconstruction Algorithms 3 2 Neural Network Reconstruction of Impedance Tomography Systems 3 2 1 Multi layer perceptron fundamentals 3 2 2 Training multi
68. different transmitters As before the received signal is buffered and inverted These signals are then applied to four DG303 CMOS analogue switches in parallel The first DG303 has Ao and Ago as control inputs to produce voltages proportional to the conductance and capacitance between transmitter A and receiver A respectively The second DG303 has Bo and Bs as control inputs to produce voltages proportional to the conductance and capacitance between transmitter B and receiver A respectively The third DG303 has and as control inputs to produce voltages proportional to the conductance and capacitance between transmitter C and receiver A respectively and finally the fourth DG303 has Do and Ds as control inputs to produce output voltages proportional to the conductance and capacitance between transmitter D and receiver A respectively All this circuitry is duplicated for receiver B C and D so as to produce voltages proportional to all the different transmitter receiver electrode pairs in parallel Selection of the low pass filter cutoff frequency is an important design parameter If it is too low then the settling time will be too long thus limiting the effective frame rate of the system Since a frame rate of 200frames s is required the cutoff frequency of the low pass filter must be greater than 200Hz If it is too high then the passband of the low pass filter may overlap with the undesired harmonics in the DG303 output resulting in a poo
69. distance at a certain speed the diameter of the pulley must be specified The user stipulates whether an upward or downward flow is desired and whether the bubble is on the inside or the outside of the pulley relative to the stand and this unit calculates in which direction the motor must step If the AT6400 stepper controller has been switched off since the previous test then it is necessary to download the operating system otherwise the stepper motors simply need to be re initialised unit stepconfig interface uses Windows Messages SysUtils Classes Graphics Controls Forms Dialogs Buttons ExtCtrls StdCtrls Registry StepUnit Spin Math type TStepperConfig class TForm Paneli TPanel Label1 TLabel Panel2 TPanel StepClose TSpeedButton Label2 TLabel Bevel1 TBevel StepDownload TSpeedButton Steplnitialise TSpeedButton StepReset TSpeedButton Label3 TLabel Label4 TLabel Label5 TLabel Label6 TLabel Label7 TLabel AirDiameter TEdit AirVelocity TEdit AirDistance TEdit Label9 TLabel GravelDiameter TEdit GravelVelocity TEdit Label10 TLabel Label11 TLabel GravelDistance TEdit AirBubblePosition TRadioGroup GravelBubblePosition TRadioGroup FlowDirSelect TRadioGroup AirDrive TSpinEdit GravelDrive TSpinEdit Label13 TLabel Label14 TLabel AirAccel TEdit Label8 TLabel Label12 TLabel GravelAccel TEdit AirBubbleJog TPanel GravelBubbleJog TPanel La
70. do sum sum InputData i colnum MeanColumn sum traintotal end calibration works as follows before start generating a training database a set of frames is taken for the rig filled only with water During database generation drift takes place in the readings as a result of the properties of the water changing temperature changes etc NOTE training database generation typically takes place over a couple of days At the end of training database generation another set of frames is captured for the rig filled only with water These provide the two endpoints and the calibration for points in between are calculated as a linear interpolation between these two points The process is then repeated for the testing database procedure TRecLoadDBase PerformCalibrate var rowindex colindex region numrepeat integer driftcomp real begin rowindex 0 region 1 numrepeat 0 while rowindex lt traintotal do begin for colindex 1 to 128 do begin driftcomp trainstopvalues region colindex trainstartvalues region colindex rowindex trainstartindex region numrepeat RecTrainStop Value 1 l trainstopindex region trainstartindex region 1 4 trainstartvalues region colindex InputData rowindex colindex InputData rowindex colindex driftcomp end inc rowindex RecLoadProgressBar Steplt if rowindex trainstopindex region numrepeat RecTrainStop Value then begin inc region if region numtrai
71. end end interpret the network outputs and display the results procedure TRecTest PlotTestOutput var tempoutputpred tempoutputdes TOutput col row index top bottom integer tempphasepred tempphasedes TBubblePhase containsair containsgravel Boolean begin if RecTestVolume Visible then begin display the volume fractions in bar charts where the desired volume faction is the bar on the right and the network prediction is the bar on the left RecAirChart Series 0 Clear RecAirChart Series 0 Addy testoutput 0 cITeeColor if not usecurrentsamples then RecAirChart Series 0 Addy OutputData testindex 0 clTeeColor RecGravChart Series 0 Clear RecGravChart Series 0 Addy testoutput 1 clTeeColor if not usecurrentsamples then RecGravChart Series 0 Addy OutputData testindex 1 clTeeColor RecWaterChart Series 0 Clear RecWaterChart Series 0 Addy 100 testoutput 0 testoutput 1 clTeeColor if not usecurrentsamples then RecWaterChart Series 0 Addy 100 OutputData testindex 0 OutputDataltestindex 1 clTeeColor end else begin display an image of the vessel cross section index 0 for col 0 to 9 do begin for row 0 to 9 do begin tempoutputdes col row outpipe tempoutputpred col row outpipe end end for col 0 to 9 do begin case col of 0 begin top 2 bottom 7 end 1 begin top 1 bottom 8 end 2 7 begin top 0 bottom 9 end
72. filename c TestData weight inttostr BottomTestNum Value txt AssignFile tempfile filename Reset tempfile ReadLn tempfile tempoutputs ReadLn tempfile temphidden if temphidden gt 0 then begin upperindex temphidden numbottomhidden temphidden SetLength bottomhiddenweights 129 temphidden SetLength bottomhiddenoutputs temphidden 1 load in the weights for the hidden layer neurons for colindex 0 to temphidden 1 do for rowindex 0 to 128 do ReadLn tempfile bottomhiddenweights rowindex colindex end SetLength bottomweights upperindex 1 totaloutputs load in the weights for the output layer neurons for colindex 0 to tempoutputs 1 do for rowindex 0 to upperindex do ReadLn tempfile bottomweights rowindex colindex CloseFile tempfile load the pre processing data for the top system filename c TestData preproc inttostr Top TestNum Value txt AssignFile tempfile filename Reset tempfile for index 1 to 128 do ReadLn tempfile topmean index for index 1 to 128 do ReadLn tempfile topstddev index if totaloutputs 88 then begin ReadLn tempfile tempphase if tempphase 0 then tempair true else tempgrav true RecLoadDBase SetPhases tempair tempgrav end CloseFile tempfile load the pre processing data for the bottom system filename c TestData preproc inttostr BottomTestNum Value txt AssignFile tempfile filename Rese
73. find the mean of the different inputs over the 50 frames for index 1 to 128 do begin topcalibvalue index topcalibvalue index 50 bottomcalibvalue index bottomcalibvalue index 50 end end else real time captured data else begin topindex 1 bottomindex 1 topsyst true cnt 0 I pre process the data by performing calibration and standardisation and place the data inputstore for index 0 to 12799 do begin if topsyst then begin inputstore storeposition readings topindex result index topcalibvalue topindex topmean topindex topstddev topindex inc topindex if topindex 129 then topindex 1 end else begin inputstore storeposition readings bottomindex 128 result index bottomcalibvalue bottomindex bottommean bottomindex bottomstddev bottomindex inc bottomindex if bottomindex 129 then begin bottomindex 1 inputstore storeposition sampletime Now starttime inc storeposition if storeposition 6000 then storeposition 1 inc totalfrms end end inc cnt if cnt 8 then begin topsyst not topsyst cnt 0 end end busysampling false II if still more frames to capture if totalfrms lt DesiredFrames Value then begin II if finished with the previous reconstruction then start the capture of another set of frames and perform the reconstruction of this set if not busyreconstruction then begin busysampling true
74. for row 0 to 9 do begin s Pixel tinttostr index OutTable FieldByName s Value AddinputTomo Desired col row 144 inc index end extract from the sampled data the readings specific to each particular system and store the results index 1 col vers 1 256 row 0 if AddSystemSelect ItemIndex 1 then col col 8 while col lt vers 256 1 do begin S Voltage inttostr index InpTable FieldByName s Value result col inc row inc index if row 2 8 then begin row 0 inc col 9 end else inc col end store the volume fractions OutTable FieldByName WaterVolume Value AddinputTomo WaterVolume OutTable FieldByName GravelVolume Value AddinputTomo GravelVolume OutTable FieldByName AirVolume Value AddlnputTomo Air Volume force the database to write the changes to the tables InpTable Post OutTable Post move to the next record in the database InpTable Next OutTable Next end 11 display the current samples in the tables InpTable FindKey currenttest 0 001 OutTable FindKey currenttest 0 001 UpdateTables end start the sampling as a background process specifying the number of frames of data to capture procedure TAddToDatabase AddSampleStartClick Sender TObject begin Cursor crHourGlass SampHardware StartSample numframes end tM GENERAL PROCEDURES disable the statu
75. frames By finding the peaks of the cross correlation functions the algorithm is able to determine the transit time between the two measurment systems and hence calculate the individual component velocities unit realunit interface uses Windows Messages SysUtils Classes Graphics Controls Forms Dialogs Buttons StdCtrls ExtCtrls Tomo Spin math TeEngine Series TeeProcs Chart Registry StepUnit type TRealMain class TForm Panel2 TPanel Panel3 TPanel Label1 TLabel TopTomo TTomo BottomTomo TTomo Label2 TLabel Label3 TLabel Label4 TLabel Label5 TLabel Label6 TLabel RealCalibrate TSpeedButton RealStart TSpeedButton Label9 TLabel Label10 TLabel persec TLabel permsec TLabel totframes TLabel Label11 TLabel frmrate TLabel RealOnline TRadioGroup RealExit TSpeedButton RealWeights TSpeedButton Label7 TLabel Label8 TLabel BottomTestNum TSpinEdit TopTestNum TSpinEdit RealView TSpeedButton RealTimer TTimer RealFrameRate TSpinEdit Label12 TLabel Label13 TLabel frmnumber TLabel permin TLabel TopVolume TPanel BottomVolume TPanel TopAirChart TChart Series TBarSeries TopGravChart TChart Series2 TBarSeries TopWaterChart TChart BarSeries1 TBarSeries 178 BottomAirChart TChart BarSeries2 TBarSeries BottomGravChart TChart BarSeries3 TBarSeries BottomWaterChart TChart BarSeries4 TBarSeries Label14 TLabe
76. hardware to indicate completion of sampling curcalibrate Boolean procedure CalculateTestCase calculate network output for test case procedure PlotTestOutput II display network output procedure ProcessResults process sampled rsults public Public declarations procedure WndProc var TheMsg TMessage override override the default Windows message handling procedure to capture message II indicating sampling completion end var RecTest TRecTest usecurrentsamples Boolean use samples taken from the DAQ result array 0 51200 of double array to store sampling results calibdata array 1 25600 of double calibval array 1 128 of double calibration reference point implementation uses loadunit mainunit networkunit paramunit realunit R DFM tM MM TESTING PROCEDURES each time the timer overflows start a new test For the testing database this is simply copying a new test case from the testing database and performing a reconstruction For the sampling instruct the sampling hardware to collect a new frame of data procedure TRecTest RecTestTimerTimer Sender TObject var colindex integer begin RecTestTimer Interval floor 1000 1 RecTestFrameRate Value if not usecurrentsamples then begin I copy the test case from the testing database for colindex 0 to 128 do testdata colindex InputData testindex colindex perform a
77. hmenste 3200 nee fee ts DX rt me Further figure 7 15 shows a sequence of image reconstructions performed on line during the test For each screen capture the top frame is the reconstruction of the top measurement plane and the bottom frame is the reconstruction of the bottom plane The order of the screen captures is left to right and only every 50 frame is captured No means was provided to verify these results other than visual inspection of the rig during a test Subsequently it cannot be said that the system is detecting individual bubbles although it is certainly true that the System is detecting groups of bubbles Although the results of this test are promising better reconstruction results will be achieved with a neural network if some partial separation exists between the multiple phases in the pipeline under investigation Whether this can be achieved or not depends on the physical limitations and requirements of the intended application 100 7 3 4 Cross correlation velocity prediction results In order to assess the accuracy of the cross correlation velocity prediction a number of tests were conducted using the 0 01 volume fraction air bubble at different simulated velocities A comparison was made between the simulated air phase velocity and the velocity measured by the cross correlation algorithm and the results are plotted in figure 7 16 Included in figure 7 16 is the mean absolute err
78. i 1 to j 1 do begin if AvailFreq i lt gt AvailFreq i 1 then begin TempFreq k AvailFreq i inc k end end for i 1 to 1 do begin AvailFreq i TempFreq i end end setup array of all standard crystal frequencies procedure TPBILMain FormShow Sender TObject begin xtal 1 166 xtal 2 1 638466 xtal 3 1 843266 xtal 4 266 xtal 5 2 457666 xtal 6 3e6 xtal 7 3 276866 xtal 8 3 579545 6 xtal 9 3 686466 xtal 10 4e6 xtal 11 4 433619 6 xtal 12 586 xtal 13 5 0688e6 xtal 14 666 xtal 15 6 14466 xtal 16 866 xtal 17 8 867238e6 xtal 18 1066 xtal 19 11 059266 xtal 20 1266 xtal 21 1666 xtal 22 16 3866 xtal 23 2066 end calculate the minimum frequency component on the LPF outputs by simulating the mixing in the synchronous detectors and aliasing in the switched capacitor l filters for a set of transmitter frequencies The magnitudes are the frequency harmonics are simulated so only those frequencies with a magnitude greater than the threshold are included function TPBILMain MultOut inputs array of double double var a b c d e upper numfact integer begin Qsort inputs 0 7 include a DC component with a magnitude of 0 5 FreqMag 1 1 0 FreqMag 1 2 0 5 b 2 calculate up to the 100th harmonic of the transmitter but only II include those frequency
79. ij where n is the learning rate and controls the speed of convergence and u is the momentum parameter that adds inertia to the movement of the algorithm through the weight space 29 Resilient back propagation is a local adaptive learning algorithm that uses local gradient information to modify the weights of the network directly 34 35 36 Unlike gradient descent Resilient back propagation simply makes use of the sign of the derivative and not the magnitude to update the network weights according to the following rule 36 OE A t if gt 0 y t ow OE ij 0 otherwise where Aj t is the weight update value for weight w t at epoch t These weight update values are modified during the training process to give optimal convergence on the minimum of the error function The network training process can be summarised as follows Initially the weights of the network are randomly set The training data inputs are propagated through the network and the network outputs are calculated using these initial weights The network predictions are then compared to the desired network outputs and the derivative of the error function is calculated using back propagation This gradient information is then used to adjust the network weights using either gradient descent or Resilient back propagation and the process is repeated Through multiple iterations the error in the network predictions steadily decreases as the neural network lea
80. in parallel Since the transmitters now have a much lower output impedance good quality square waves are generated even when all eight transmitters are operating simultaneously Figure 5 8 on the following page is a block diagram of a transmitter 41 LEVEL SHIFTING FROM FREQUENCY GENERATOR BOARD HIGH SPEED POWER MOSFET DRIVER PUSH PULL POWER MOSFET OUTPUT STAGE PROBE FIGURE 5 8 BLOCK DIAGRAM OF THE OUTPUT TRANSMITTER WITH PUSH PULL POWER MOSFET OUTPUT STAGE TO CAPACITANCE PLATE AND CONDUCTANCE A drawback of the push pull power MOSFET output stage is shoot through resulting from both MOSFETs being on simultaneously since a single MOSFET driver is being used to switch both power MOSFETS in parallel This shoot through resulted in ringing on the outputs serious degradation of the power supply quality and heat dissipation in the power MOSFETS Standard push pull power MOSFET configurations use two independent drive signals phased so that one MOSFET is always switched off before the other is switched on Since time and cost limitations meant that it was impractical to redesign the transmitter for a break before make driver some other technique had to be used in order to generate these two independently phased drive signals from one source The solution to this problem is illustrated in figure 5 9 CROSS CONDUCTION LEVEL SHIFTING TC4427 DUAL HIGH SPEED POWER MOSFET DRIVER TC4427 output A B
81. into the population 61 By iterating the above process over a number of generations the populations of trial solutions produced move steadily towards regions of high performance although it is not guaranteed that the optimum solution will be found The PBIL algorithm attempts to create a single probability vector from which trial solutions are drawn to produce the next generation 61 Each trial solution is encoded as a binary sequence of 1 or 0 The probability vector represents the probability of each bit position containing 1 61 At the start of the algorithm the values of the probability vector are initialised to 0 5 representing an equal probability of generating a 1 or O in that particular bit position Consequently the trial solutions generated are a random sequence of 1 and 0 For each generation the PBIL algorithm actively updates this probability vector to represent high evaluation trial solutions This is achieved through gradually shifting the values in the probability vector away from 0 5 and towards the bit values of the best trial solution for each generation The probability update rule is described below 61 PV PV jx 1 LR LR x TS 5 3 where is the probability vector and PV is the probability that bit position j contains a 1 LR is the learning rate and TS is the trial solution towards which the probability vector is being shifted Consequently the values of th
82. is a fairly time consuming process The other training parameters are fairly self explanatory and figure 5 20 is a screen capture of this form A full scale version of this screen capture is included in Appendix N on page 205 This shows a typical configuration of a double layer feed forward neural network trained using Resilient back propagation to perform a volume fraction prediction Before network training commences the network weights must be randomly initialised and this is performed by selecting Randomly initialise network weights Start training opens a form that allows the user to monitor the network training process and in particular the generalisation performance of the neural network on the test database A screen capture of this form can be found in Appendix N on page 206 Network training can be stopped temporarily by clicking Stop training if it is necessary to perform some other task on the computer during network training The training will resume on clicking Start training Having trained a neural network the network weights performance results and pre processing data need to be saved in order for the network to be reused at a later date Save network weights Save network performance and Save network pre processing in the file menu perform these functions The network performance can then be tested using the various options provided in the network testing menu 64 Validate database
83. layer perceptrons 4 FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY OF AN 8 ELECTRODE SYSTEM 4 1 Frequency Division Multiplexed Impedance Tomography Concept 4 2 Design of an 8 electrode System to Verify the Frequency Division Multiplexed Impedance Tomography Concept 4 2 1 Generation of transmitter frequencies 4 2 2 Output transmitter details 4 2 3 Conductance and capacitance receiver details 4 2 4 Synchronous detection of receiver signals 4 2 5 Capture of capacitance and conductance readings and the reconstruction results achieved 4 2 6 Limitations of the current system 5 FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY OF A 16 ELECTRODE SYSTEM 5 1 Construction of a Laboratory Scale Rig to Simulate a Pipeline Section 5 1 1 Reasoning behind certain rig dimensions 5 1 2 Design of transmitter and receiver electrodes and guarding 5 2 Electronic Design of a 16 electrode Frequency Division Multiplexed Impedance Tomography System 5 21 Output transmitter details 5 22 Conductance and capacitance receiver details 5 23 Synchronous detection of receiver signals 5 2 4 Population based incremental learning fundamentals 5 25 Generation of transmitter frequencies 5 3 Software Design of a 16 electrode Frequency Division Multiplexed Impedance Tomography system 5 3 1 Data capture software description 5 3 2 Training and test database generation software description 5 3 8 Neural network reconstruction software description 6 STATIC SITUATION RESULTS FOR THE
84. namely 51 1 grounded rings above and below the central measurement electrodes 2 driven segmented axial guards above and below the central measurement electrodes where the guard electrodes are driven to the same potential as the corresponding measurement electrode 3 asingle ring of measurement electrodes without guarding Since the electric field of a capacitance tomography sensor is three dimensional and the reconstructed images represent a two dimensional approximation of that field it is important to ensure the generation of a homogenous electric field In addition the measurement space must be focussed The particular guard electrode design employed has a significant influence on these specifications 51 Consequently numerous simulations have been performed analysing the merits of each of these different approaches These simulations typically use finite element techniques to model the three dimensional electric field of the above three sensor configurations However contrasting results have been achieved and so the following section will briefly compare the major results of a few of these simulations SA ji Although there is general agreement that the DRIVEN GROUNDED GUARD B use of driven axial guard electrodes both capacitance and resistance tomography improve the axial evenness of the electric field or current distribution different opinions VIRTUAL EARTH exist regard
85. of the air and gravel phases at the top and bottom systems were correlated to determine the individual component velocities Specifically tests were conducted for the simpler two phase air water reconstruction Using a 0 01 volume fraction air bubble the mean absolute error of the air phase velocity measurement was 0 13m s at the maximum flow simulation apparatus velocity of 4 2m s This corresponds to a velocity discrimination of 3 196 which is comparable to the theoretical limit calculated as 1 596 for the current System In terms of the three phase air gravel water reconstruction tests revealed that the system was able to calculate the individual velocities of the air and gravel phases provided the reconstruction algorithm performed a satisfactory separation between these two phases This research concluded that frequency division multiplexed impedance tomography provides a means of performing on line real time reconstructions at a far higher frame rate than standard time division multiplexed impedance tomography systems Subsequently it enables tomography to be used in industrial applications where previously the limited frame rates of tomography systems were a hindrance Whether it is applicable to this particular application or not depends on the results of further research and testing Various modifications have been recommended which should improve the performance of the system As an example an 8 electrode sine wave system is currently being
86. of the synchronous detector board multiplexing are provided in Appendix C on page 125 Initial tests revealed the existence of ripple on the low pass filter output voltages It was determined using a spectrum analyser that the frequency of this ripple was typically within the 200Hz passband of the system Further it was observed that these ripple frequencies were transmitter dependent For example if the output of the synchronous detector for the measurement of the conductance between transmitter A and receiver A contains a 75Hz component then this 75Hz component will also appear at the output of the synchronous detector for the measurement of the conductance between transmitter A and the other receivers In addition this 75 2 component will also appear at the output of the synchronous detector for the measurement of the capacitance between transmitter A and receiver A and so on It was also noted that this output ripple did not exist when only one transmitter was operating but would appear when two or more transmitters were operating simultaneously Finally it was shown that these frequency components would change if the transmitter frequencies were changed thus proving that they were dependent on the transmitter frequencies Further tests revealed that these frequency components were the result of the transmitter output square wave harmonics mixing in the multiplication stage of the synchronous detectors Figure 5 13 on page 48 provides an example o
87. of the transmitter and receiver electrode are detailed in figure 5 4 A capacitance plate length of 120mm was chosen When this electrode array was designed the intended DRIVEN GROUNDED AXIAL AXIAL Hears application was for flow velocities up to 12m s This has since been increased to 20m s but the choice of the electrode length will be explained in terms of the initial specification An object moving at 12m s will pass FIGURE 5 4 DIMENSIONS OF THE TRANSMITTER AND RECEIVER ELECTRODES AND GUARDS ALL DIMENSIONS ARE IN MILLIMETRES through each measurement section in a time of 10ms Provided the impedance tomography system performs on line reconstructions at a speed greater than 100frames s this object is guaranteed to be captured at least once In terms of the spatial filtering effect of the electrode it was assumed that the measurement volume is confined to the measurement plate namely 120mm For a maximum velocity of 12m s the highest frequency components generated will be 100Hz as calculated using equation 5 1 above To satisfy the Nyquist sampling theorem the impedance tomography system is required to capture data and perform reconstructions at 200frames s 37 Hence the minimum required frame rate is 200frames s for this particular application The widths of the plates were fixed by the circumference of the pipeline for a given number of electrodes FIGURE 5 5 PHOTOGRAPH OF THE PIPELINE FROM ABOV
88. operating correctly The addition of the brown out detector solved this problem The circuit details of the frequency generator board are provided in Appendix C on page 124 Having designed the frequency generator board it was necessary to select eight frequencies that maximise the frequency separation of the harmonics As mentioned previously the PBIL optimisation algorithm will be used to perform this task The implementation of the PBIL algorithm to this specific application will be discussed briefly in the following section Firstly the algorithm requires a set of possible transmitter frequencies from which it can select a subset of eight frequencies as a trial solution By combining the set of standard crystal frequencies available with all the possible division factors that the PIC16F84A microcontroller can implement a set of 256 different frequencies were generated between 17 6kHz and 80kHz The index of each frequency is coded as an 8 bit value Subsequently a trial solution is a 64 bit binary sequence where each byte indexes into this array to select the corresponding transmitter frequency Secondly the simulation algorithm is modified to find the minimum ripple frequency on the output of the synchronous detectors for a given set of eight transmitter frequencies This corresponds to the fitness of a trial solution If this minimum ripple frequency is maximised then the output ripple will be minimised In this way the algorithm selects a set
89. pipeline diameter and is represented by a single pixel in the 10 by 10 image reconstruction In total the bubble sizes simulated correspond to 10 20 and 30 of the pipeline diameter and represent volume fractions of 0 01 0 04 and 0 09 when situated within the measurement volume In terms of the lengths of the bubbles all bubbles simulated are 190mm long thus ensuring the bubble extends beyond the sensitivity region of the measurement electrodes Figure 6 4 is a photograph of the simulated air and gravel bubbles used in the training and test database generation The diameter of the 1096 bubble is 35mm FIGURE 6 4 PHOTOGRAPH OF THE THREE SIZES OF AIR AND GRAVEL BUBBLES CORRESPONDING TO 30 20 AND 10 OF THE PIPELINE DIAMETER RESPECTIVELY Although the use of potable water dramatically reduced the drift of the capacitance and conductance readings it was observed that the readings would still drift It is believed that this drift is the result of temperature variations Since the task of generating the training and test database takes place over a couple of days any drift of the readings will have a detrimental effect on the networks trained using these databases In addition it was observed that the properties of the water would vary as a result of the impurities introduced by the gravel bubbles Subsequently the capacitance and conductance readings would drift during database generation It was therefore decided that some form of software c
90. quad op amp drives eight DG303 switches in parallel a capacitive load of 56pF is created resulting op amp instability Replacing the standard quad op amps with LT1356 quad op amps from Linear Technology solved this problem These op amps are unity gain stable and are capable of driving capacitive loads making them ideal for this particular application Multiplexing was required to capture the output voltages since the PC30G card only has 16 analogue input channels and 128 voltages must be captured from two FDM impedance tomography systems Unlike standard tomography systems the multiplexing for FDM impedance tomography takes place after the measurements are performed Consequently it does not affect the readings and the system does not have to wait between each sample for the voltages to stabilise In FDM impedance tomography the multiplexing is purely for sampling purposes and it could be avoided by using a 128 input data acquisition card In the current system this was not a limiting factor and the multiplexing did not reduce the potential frame rate of the FDM impedance tomography system Consequently only a 16 input data acquisition card was used A single DG506A 16 input CMOS analogue Switch is used to multiplex the eight capacitance and eight conductance readings for each channel The address lines of these multiplexers are driven by the sample controller board which will be discussed in greater detail at a later stage The circuit details
91. real var airfinaccel gravfinaccel airtimeflat gravtimeflat real airtottime gravtottime tempairrps tempgravrps mintottime real begin if VelSource ItemIndex 0 then begin tempairrps StepperConfig airrps tempgravrps StepperConfig gravelrps end else begin if RealMain airvelocity lt gt 1 then tempairrps RealMain airvelocity StepperConfig aircircumference else tempairrps 0 if RealMain gravelvelocity lt gt 1 then tempgravrps RealMain gravelvelocity StepperConfig gravelcircumference else tempgravrps 0 end the position of the bubble is obtained by segmenting the path into three separate sections namely acceleration constant velocity and deceleration with StepperConfig do begin calculate the position of the air bubble if tempairrps 0 then begin time taken for bubble to accelerate to desired velocity airfinaccel tempairrps airacceleration time during which bubble moves at constant velocity airtimeflat airrevs tempairrps airfinaccel total time airtottime airfinaccel 2 airtimeflat mintottime 2 sqrt airrevs airacceleration if airfinaccel gt 0 5 mintottime then begin if acceleration not high enough to reach constant velocity if time lt 0 5 mintottime then airtopposition 0 5 airacceleration power time 2 else if time lt mintottime then airtopposition 0 25 airacceleration power mintottime 2 0 5 airacceleration power mintottime time 2
92. reconstructions for the previous set of 50 frames is performed 5 busyreconstruction and busysampling provide flags indicating when it is OK to start the capture of the next set of 50 frames or perform the reconstructions for the previous set of frames 6 if the application is still busy with the reconstruction for the previous set of frames then the sampling stops until it is complete otherwise overflow In this way the sampling and reconstruction tasks are overlapped thus achieving maximum possible frame rate for the system procedure TRealMain ProcessResults var index topindex bottomindex cnt integer topsyst Boolean begin pre process for a calibration if curcalibrate true then begin curcalibrate false topindex 1 bottomindex 1 topsyst true cnt 0 separate the data between the two systems for index 0 to 12799 do begin if topsyst then begin 180 topcalib topindex result index inc topindex end else begin bottomcalib bottomindex result index inc bottomindex end inc cnt if cnt 8 then begin topsyst not topsyst cnt 0 end end for index 1 to 128 do begin topcalibvalue index 0 bottomcalibvalue index 0 end for index 1 to 6400 do begin cnt index mod 128 if cnt 0 then cnt 128 topcalibvalue cnt topcalibvalue cnt topcalib index bottomcalibvalue cnt bottomcalibvalue cnt bottomcalib index end
93. results panel and start the model search timer numhidden RecSetParam RecModelStart Value RecModelResults Visible true bestsum 300 bestthresh 100 ModelSearchTimer Enabled true PauseTrain Visible true end else begin 1 else call the appropriate training procedure if RecSetParam GetNumHidden gt 0 then TrainDoubleNetwork else TrainSingleNetwork end performingtrain false RecNetworkStart Caption Start training end else begin 168 currepoch maxepochs performingtrain false RecNetworkStart Caption Start training end end procedure to train a single layer feed forward neural network where the network weights can be modified using gradient descent or RPROP Early stopping is performed to prevent overfitting the training data thus ensuring good generalisation performance for the test data procedure TRecTrainNetwork TrainSingleNetwork var outputnodeindex inputnodeindex currsample integer weightsum newnodeoutput newnodederiv delta real weightstore array 1 2 of real i integer begin initialise important variables currepoch 1 currentfails 0 currthresh 0 currsum 0 11 stop if the number of epochs exceeded or user specifies early stopping the number of fails has been exceeded i e the validation error has continued to increase for a certain minimum number of iterations while currepoch lt maxepochs and not currentf
94. sections are obtained using an implementation of the point by point cross correlation algorithm Since the cross correlation is often performed on sampled data the sampled data form of the standard cross correlation function is as follows 50 oq ix 4 Ry A nai 1 042 7 4 n 1 where At is the sample interval N is the number of samples and x is sample number n from sensor x This equation is calculated using either point by point calculation or evolutionary calculation Point by point cross correlation expands the above equation into the set of individual equations shown below 50 R yy 0 Xoyo XNYN Ry 1 x X vy 1 1 2 Xoya NY NH 7 5 R yy J t Xoy 24J XNY Na Since the aim is to detect the peak of the function it is not necessary to divide by N Evolutionary cross correlation provides a means of obtaining a quick approximation of the position of the peak without having to do the complete cross correlation calculation as is done in the point by point technique The method of evolutionary calculation is described by rewriting the above equations as follows 50 yo YN 72 7 6 Youu YN J uu Ey 1 2 Clearly if all the evolutions are included in the calculation then it is effectively the same as the point by point method and no speed advantages are achieved However a quick approximation of the cross correlation function can be obtained b
95. sensing field is more sensitive for a thinner pipe wall 48 Ultimately the pipe wall material and thickness is determined by the practical working conditions Such as pressure corrosion abrasion temperature and so on 48 In this application the pipe wall thickness was Set by the strength required to support the large volume of water contained within the rig Specifically the rig will hold approximately 160 litres of seawater when full Consequently the rig was constructed using 19mm rolled acrylic Cross correlation flow measurement depends on the successful detection of disturbances at both the top and bottom measurement planes In practice these disturbances will gradually change on moving downstream If the position of the downstream sensor were reasonably close to that of the upstream sensor then the patterns or signals generated would be sufficiently similar to be recognised by the cross correlator 12 Generally the selection of the plane separation is a compromise between the similarity of the flow patterns and the resolution with which the transit time can be determined 12 Interference between the electric fields of the two systems requires a large plane separation 43 Various simulation studies have produced varying results regarding the desired plane separation for dual plane cross correlation flowmeters Loh et a show that a separation of at least one pipe diameter is required to reduce the cross talk to approximately 19
96. such a multi phase flow mixture To obtain an initial answer to this question it was decided that the system should be tested on a homogenous bubble flow since the bubbles in such a flow are smaller than the system resolution It is important to remember that although the tomography System may not be able to resolve the individual air bubbles within a particular area it should provide a ratio of the phases within that area 65 The specific flow pattern achieved in a bubble column depends on the superficial velocity of the gas properties of the liquid phase sparger design column diameter and height of dispersion 66 Simply by increasing the 98 FIGURE 7 13 PHOTOGRAPH OF THE SPARGER USED TO GENERATE A HOMOGENOUS BUBBLE FLOW IN THE LABORATORY SCALE RIG superficial velocity of the gas phase the observed pattern changes from a homogenous bubble flow to a transition regime and then a turbulent churning flow 66 67 A sparger was designed for the simulation of a homogenous bubble flow It consists of two disks separated by a ring spacer and the top disk is drilled with a sequence of 1mm holes A quick coupler is used to connect the sparger to an air hose This air hose connects to the regulator of an air compressor Simply by varying the regulator pressure different flow patterns can be generated Figure 7 13 is a photograph of the sparger designed for the generation of a homogenous bubble flow The real time volume fractio
97. the smallest bubble with which the network has been trained this represents the maximum achievable resolution for the current system However it is possible that greater resolution could be achieved through training the network with smaller bubbles e By using the Test current samples feature of the reconstruction program it was determined that these results could be replicated for on line reconstruction results where the readings are captured directly from the system and a reconstruction is performed Further the reconstruction results were unaffected by minor variations in the properties of the water thus proving the correct operation of the calibration feature 73 FIGURE 6 5 SCREEN CAPTURES OF TEST CASES FOR THE TWO PHASE AIR WATER IMAGE RECONSTRUCTION OF THE TOP FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM WHERE THE DESIRED OUTPUTS ARE ON THE RIGHT AND THE NETWORK PREDICTIONS ARE ON THE LEFT IN EACH SCREEN CAPTURE IN EACH FRAME THE BLACK IS THE WATER PHASE THE WHITE IS THE AIR PHASE AND THE LIGHT GREY REPRESENTS THE AREA OUTSIDE THE PIPELINE Previous research has shown that greater volume fraction prediction accuracy can be achieved by training a neural network to predict the volume fractions directly 3 4 In addition the results are obtained faster than if an image reconstruction is first performed Multiple single layer feed forward neural networks were trained to perform the two phase air water volume fr
98. the reconstruction process is stopped Otherwise the resistivity distribution p is updated according to the following equation Pki 7 ox o T Px Vo 3 4 and the procedure is iterated until the convergence criterion has been met and a suitably accurate prediction obtained 19 By incorporating prior knowledge of the structure of the matrices involved in this process various modifications to the basic algorithm can be performed that reduce the total reconstruction time 28 As mentioned previously neural networks can be trained to perform image reconstructions of three phase air gravel seawater mixtures 3 4 Although the training process may require a considerable amount of time to perform for the class of neural networks implemented the network execution once the training is complete is fast since no iterations are required Further it was shown that a neural network could be used to perform on line real time reconstructions of an impedance tomography system and that the major source of delay in such an on line process would be the data acquisition system and not the reconstruction stage 3 Since a high frame rate is required for the intended application neural network reconstruction techniques were employed 3 2 Neural Network Reconstruction of Impedance Tomography Systems A neural network is a highly parameterised mathematical function that offers a general framework for representing non linear functional mappings fr
99. the volume fractions for the current test case unit addunit interface uses Windows Messages SysUtils Classes Graphics Controls Forms Dialogs StdCtrls Buttons ExtCtrls ComCtrls Grids TeEngine Series TeeProcs Chart Db DBGrids Tomo type TResultTable capacitance resistance TAddToDatabase class TForm AddReadings TPanel AddStatusBar TStatusBar AddCapTable TStringGrid AddResTable TStringGrid Label3 TLabel Label4 TLabel StatusUpdate TTimer Paneli TPanel Panel2 TPanel Label1 TLabel Label2 TLabel Label5 TLabel AddVolume TChart Series1 TBarSeries Series2 TBarSeries Series3 TBarSeries AddSystemSelect TRadioGroup AddReset TSpeedButton AddPrevious TSpeedButton AddSampleStart TSpeedButton AddNext TSpeedButton AddExit TSpeedButton AddSetAir TEdit AddSetGravel TEdit AddSetWater TEdit AddUserSetVolume TCheckBox AddInputTomo TTomo procedure AddExitClick Sender TObject procedure FormCreate Sender TObject procedure StatusUpdateTimer Sender TObject procedure FormShow Sender TObject procedure AddResetClick Sender TObject procedure AddSampleStartClick Sender TObject procedure AddPreviousClick Sender TObject procedure AddNextClick Sender TObject private Private declarations currenttest integer current test number procedure ProcessResults process the results from the sampling hardware
100. then begin calculate the time delay between the desired volume fraction peak and the real time peak if there was no slippage for example a synchronous belt was used then this would give l a good indication of the time response of the system electronics TopAirDelay Caption floattostrf RealMain frameinterval airtopoffset ffFixed 4 3 for testcase airtopstart to airtopstop do begin if totaloutputs 2 then topairerror topairerror abs desiredvolume testcase 1 outputstore testcase airtopoffset prediction 1 else topairerror topairerror abs desiredvolume testcase 1 volumestore testcase airtopoffset 1 end topairerror topairerror airtopstop airtopstart 1 end if the algorithm is unable to determine a peak then simply calculate the mean absolute error over all the width of the opposite phase else begin TopAirDelay Caption unknown for testcase graveltopstart to graveltopstop do begin if totaloutputs 2 then topairerror topairerror abs 0 outputstore testcase graveltopoffset prediction 1 else topairerror topairerror abs 0 volumestore testcase graveltopoffset 1 end topairerror topairerror graveltopstop graveltopstart 1 end TopAirVolError Caption floattostrf topairerror ffFixed 4 3 topgravelerror 0 if graveltopdesiremax lt gt 1 then begin TopGravDelay Caption floattostrf RealMain frameinterval graveltopoffset ffFixed 4 3 for testcase gra
101. then begin if StepperConfig FlowDirSelect ItemIndex 0 then begin if StepperConfig GravelBubblePosition ItemIndex 0 then tempsteps StepperConfig gravelsteps else tempsteps StepperConfig gravelsteps end else begin if StepperConfig GravelBubblePosition ItemIndex 0 then tempsteps StepperConfig gravelsteps else II initialise charts and other test variables AirPlot Series 0 Clear AirPlot Series 1 Clear GravelPlot Series 0 Clear GravelPlot Series 1 Clear AirCorrPlot Series 0 Clear GravelCorrPlot Series 0 Clear topairthreshold strtofloat TopAirThresh Text topgravthreshold strtofloat TopGravThresh Text botairthreshold strtofloat BotAirThresh Text botgravthreshold strtofloat BotGravThresh Text systemseparation strtofloat SystemSeplInput Text flowmeterfactor strtofloat FlowFactorlnput Text correlationthreshold strtofloat CorrThreshInp Text fori 1 to 6000 do begin volumestore i 1 0 volumestore i 2 0 volumestore i 3 0 volumestore i 4 0 end StepperStop Visible true VerifyResults Visible true busyreconstruction false busysampling true RealStart Caption Stop testing totalfrms 1 storeposition 1 reconposition 1 curcalibrate false determine the desired direction and speed of the stepper motors for both the air and gravel phases Stepperlnit if StepperConfig airrps gt 0 then begin if StepperConfig Flow
102. then read from the key else it is the first time the program is running so do not attempt to read from the registry if KeyGood then begin numframes Reg ReadInteger NumFrames SampNumFramesValue Value numframes dbasename Reg ReadString Database automatically load the last database to be used CreateTables dbasename end finally Reg Free end end when close program release the resources allocated for sampling and write to the registry the current configuration procedure TSampMain FormClose Sender TObject var Action TCloseAction var Reg TRegistry begin SampHardware ReleaseSample SampHardware Free Reg TRegistry Create try Reg OpenKey Software Sampler True Reg Writelnteger NumFrames numframes Reg WriteString Database dbasename finally Reg Free end end 143 show AddToDatabase modally to view the data currently in the database procedure TSampMain Database1 Click Sender TObject begin AddToDatabase Caption View database data 11 stop editing of data purely for viewing sampled data AddToDatabase AddInputTomo AllowEdit false AddToDatabase AddSampleStart Visible false AddToDatabase AddReset Visible false AddToDatabase AddUserSetVolume Visible false AddToDatabase AddSetGravel Enabled false AddToDatabase AddSetAir Enabled false AddToDatabase ShowModal end to continuously sample the system initialise the sam
103. time taken to capture the next set of 50 frames Hence the computer waits for the sampling to complete and immediately instructs the capture of the next set of 50 frames once the sampling is complete Consequently both measurement planes are continuously sampled In contrast the time taken to perform the more advanced three phase reconstruction algorithms is longer than the sampling time Hence the sampling system has to wait for the computer to finish the reconstruction of the previous set of 50 frames before it starts the capture of the next set of 50 frames A dead time is therefore introduced in the sampling of the measurement planes Consequently if a bubble were to pass through a particular measurement plane during this dead time then it would not be detected since the system was not being sampled at that time Further since the frames captured do not correspond to a constant sampling interval the transit time predicted by the cross correlation algorithm is erroneous This problem can be solved through the use of a faster reconstruction computer or a reduction in the sampling rate So that the timing of the process is once again determined by the sampling and not the reconstruction 106 CHAPTER 8 CONCLUSI ONS Based on the findings of this research the following conclusions have been drawn 1 Frequency division multiplexed impedance tomography provides a means of performing impedance tomography at a far higher frame rate than st
104. to the inverse of the sensitivity matrix can be found then better reconstruction results can be achieved An example of an improved approximation technique is the single step Tikhonov method 22 The application of this method will be illustrated for the case of resistance tomography but is equally applicable to capacitance tomography For electrical resistance tomography the linearised forward problem is expressed using the following equation 22 V V Jo p po 3 1 where Jo is the Jacobian matrix that maps the change in resistivity values p to the change in voltages V Since the change in voltages is actually what is observed and the change in resistivity is the desired output a regularised solution to the inverse of the above equation is obtained as follows 1 aA 5 V Vo 3 2 where a is the regularisation parameter A is typically the identity matrix and T indicates the transpose operation 22 The matrix 7 0 aA JI be pre calculated Hence approximate tomograms can be achieved very quickly simply through a matrix multiplication The regularisation parameter a has a significant effect on the reconstructed images If the regularisation is too large then the images will be blurred and important features could be missed If the regularisation is too small then the images are dominated by measurement noise and become nonsensical 22 Various techniques exist to determine an optimal regularisation value Al
105. using the same Software as the 8 electrode prototype Simply by using two different crystal frequencies eight different output frequencies are generated This is discussed in greater detail in section 5 2 4 As mentioned in section 4 2 1 on page 25 the periods of the square waves generated by each microcontroller are 8 12 16 and 20 cycles respectively Consequently the second harmonic of the output whose period is 16 cycles equals the fundamental frequency of the output whose period is 8 cycles This would not be a problem if perfect square waves were generated and propagated throughout the system However frequency components are generated at the even harmonics of the fundamental frequency because of imperfections in the transmitters and the limited slew rate of the receiver op amps Since all four frequencies from each microcontroller are related by the crystal frequency of that microcontroller it is impossible to generate eight frequencies whose harmonics do not coincide using this technique Subsequently it was decided that each transmitter should be driven by its own independent frequency generator source In addition since each frequency is generated independently of the other frequencies it should be possible to determine a set of eight frequencies that would result in minimal output ripple Specifically an optimisation search algorithm would be used in conjunction with the simulator to determine the set of eight output frequencies that produc
106. versa Clearly the separation between the gravel and air phases represents the most complex requirement of the reconstruction algorithm As mentioned previously the conductance of the gravel and air phases can be considered the same Subsequently any distinction between these two phases is based solely on their differences in dielectric constants In addition these results are for the ideal situation of static bubbles that are contiguous masses of gravel chips or polystyrene foam cylinders It is not known how the neural network would perform for the situation of a moving distributed three phase mixture This issue will be discussed further at a later stage 75 uu FIGURE 6 6 SCREEN CAPTURES OF TEST CASES FOR A THREE PHASE AIR GRAVEL WATER IMAGE RECONSTRUCTION OF THE TOP FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM WHERE THE DESIRED OUTPUTS ARE ON THE RIGHT AND THE NETWORK PREDICTIONS ARE ON THE LEFT OF EACH SCREEN CAPTURE AS BEFORE THE BLACK IS THE WATER PHASE THE WHITE IS THE AIR PHASE THE DARK GREY IS THE GRAVEL PHASE AND THE LIGHT GREY REPRESENTS THE AREA OUTSIDE THE PIPELINE 76 Multiple single layer feed forward neural networks were trained to perfor
107. volume fraction error TopAirVolError Caption Write tempfile desiredvolume row co 8 4 WriteLn tempfile top air bubble delay TopAirDelay Caption Write tempfile aircorr row 8 4 WriteLn tempfile top gravel volume fraction error TopGravVolError Caption Write tempfile gravelcorr row 8 4 WriteLn tempfile top gravel bubble delay TopGravDelay Caption WriteLn tempfile WriteLn tempfile bottom air volume fraction error BotAirVolError Caption end WriteLn tempfile bottom air bubble delay BotAirDelay Caption with RealMain do WriteLn tempfile bottom gravel volume fraction error begin BotGravVolError Caption WriteLn tempfile desired frames inttostr DesiredFrames Value WriteLn tempfile bottom gravel bubble delay BotGravDelay Caption WriteLn tempfile desired output inttostr totaloutputs CloseFile tempfile WriteLn tempfile view results RealOnline Items RealOnline ItemIndex end WriteLn tempfile top system test number inttostr TopTestNum Value WriteLn tempfile bottom system test number inttostr BottomTestNum Value end WriteLn tempfile system separation SystemSepInput Text 201 SCREEN CAPTURES OF SAMPLER PROGRAM APPENDI X M G 2ZE se duies jo saquinu e30 1621 3591 juaung AQv3d 0 umay lt Bujdues 135 SUORE AWNJOA 495 5 89256 2 954262 3289 27 682857 122192 000052 0 0 pz BL EARS 629392 906292 829572 958942 829572 091992 264252
108. wait until is AT6400 is ready while Port Address STATUS and IB IS EMPTY lt gt 0 do begin end send command to AT6400 1 50 while WordBuffer Cmd I lt gt 0 and 1 lt MAXWORDS do begin PortW Address WordBuffer Cma I Inc I end tell AT6400 we re done Port Address 4 Status CMD READY return the number of words successfully sent SendAT6400Block end receive a block of commands from the AT6400 function TStep RecvAT6400Block Address word var Response CharBuffer word var begin 0 while Port Address STATUS and OB HAS DATA lt gt 0 and I lt MAXWORDS do begin WordBuffer Response l PortW Address Inc end WordBuffer Response l 0 Inc I return the number of words received RecvAT6400Block end request the AT6400 to update its fast status area in memory procedure TStep RequestStatus begin portfaddress STATUS REQ_STATUS request fast status update 11 wait for fast status information to be updated while port address STATUS and CS UPDATED 0 do begin end end read the status of the drives from the fast status area procedure TStep ReadStatus var status_high status_low word var status longint begin status high portw address FASTSTATUS status low portw address FASTSTATUS extract the upper and lower portions from the data status_high status_high shl 8 status_high shr 8 stat
109. weightsum OutputWeight rowindex colindex tempvalue end prevent arithmetic overflow if weightsum gt 200 then weightsum 200 else if weightsum lt 200 then weightsum 200 IL if using 1 of C then three outputs corredponding to a pixel can IL only be calculated once all three weighted inputs for those three outputs have been calculated since uses soft max activation function if numoutputs 264 then begin if Colindex 1 mod 3 lt gt 0 then weightstore colindex 1 mod 3 weightsum else begin calculate the dummy variable outputs for i 2 downto 0 do OutputNetwork count colindex i SoftMax 3 i weightstore 1 weightstore 2 weightsum interpret the network outputs to determine the pixel phase and whether this prediction is correct or not 1 Also keep a record of the total number of air and gravel pixels if OutputNetwork count colindex OutputNetwork count colindex 1 and OutputNetwork count colindex OutputNetwork count colindex 2 then begin predicts pixel is air totalairpixels totalairpixels 1 I but if not air then increment the number of 11 errors if OutputData count colindex lt gt 1 then 172 inc totalerrors end else if OutputNetwork count colindex 2 OutputNetwork count colindex 1 and OutputNetwork count colindex 2 OutputNetwork count colindex then begin 11 predicts pixel is water but if not water then increment the number of 11 errors if
110. were tested in an attempt to reduce the cross plane interference Tests revealed that the top and bottom measurement systems interfere when both are operating simultaneously The effect of this interference can be explained as follows both planes are using the same set of eight transmitter frequencies However each plane has its own frequency generator board Consequently the frequencies generated are not exactly the same As an example receiver A of the top system receives a frequency component from transmitter A of the top system at frequency f and a component from transmitter A of the bottom system at frequency f x where x is typically small Subsequently the difference frequency x is produced at the output of the multiplication stage of the synchronous detector and passes through the low pass filters to appear on the output as a substantial ripple component Figure 7 7 compares the capacitance voltage output for this transmitter A receiver A electrode pair for the top system when only the top system is operating and when both the top and bottom Systems are operating at the same time ONLY TOP SYSTEM OPERATING BOTH TOP AND BOTTOM SYSTEM OPERATING 3 3 2 5 1 2 5 1 2 1 2 1 gt gt 8 8 8 1 5 4 8 1 5 4 o o c 9 8 8 1 1 1 1 0 5 1 0 5 1 0 1 0 1 0 10 20 30 40 0 10 20 30 40 time s time s FIGURE 7 7 COMPARISON BETWEEN THE CAPACITANCE OUTPUT VOLTAGES FOR THE TRANSMITTER A RECEIVER A ELE
111. work could be started on constructing a laboratory scale impedance tomography flowmeter it was necessary to verify whether the FDM concept would actually work In addition a comparison to the performance of a TDM impedance tomography system was required in order to assess the loss of resolution resulting from the reduction in available readings This was achieved through modification of the rig used in the previous project research 1 2 3 4 The rig is a polyester pipe of inner diameter 220mm and height 300mm One end of the pipe is sealed so that it retains water when the axis is vertical The capacitance electrode array is mounted on the external periphery of the vessel and consists of eight galvanised steel plates At the centre of each plate is a 6mm stainless steel machine screw which is threaded through the pipe wall and ground flush with the inside of the pipe An earthed outer shield is also provided to prevent external electromagnetic interference 23 Figure 4 3 is a block diagram of the major components in a FDM impedance tomography system although only one complete channel is shown The following sections will examine the implementation of each of these components in greater detail PIC16F84 FREQUENCY GENERATOR CAPACITANCE CONDUCTANCE LOW PASS i TRANSMITTER RECEIVER i FILTER RECONSTRUCTION COMPUTER SYNCHRONOUS DETECTION BOARD FIGURE 4 3 BLOCK DIAGRAM SHOWING THE INTERACTIONS BETWEEN THE TRANSM
112. 00 MOVWF PORTB NOP NOP MOVLW B 00010100 MOVWF PORTB NOP NOP MOVLW B 10010100 MOVWF PORTB NOP NOP MOVLW B 10010100 MOVWF PORTB NOP NOP MOVLW B 11000010 MOVWF PORTB NOP NOP MOVLW B 11000010 MOVWF PORTB NOP NOP MOVLW B 01000010 MOVWF PORTB NOP NOP MOVLW B 01100010 MOVWF PORTB NOP NOP MOVLW 00100011 MOVWF PORTB NOP NOP MOVLW B 00101011 MOVWF PORTB NOP NOP MOVLW B 10111011 MOVWF PORTB NOP NOP MOVLW B 10111011 MOVWF PORTB NOP NOP MOVLW B 11111001 MOVWF PORTB NOP NOP MOVLW B 11011001 MOVWF PORTB NOP NOP MOVLW B 01011101 MOVWF PORTB NOP NOP MOVLW B 01011101 MOVWF PORTB NOP NOP MOVLW 00001100 MOVWF PORTB NOP NOP MOVLW B 00001 100 MOVWF PORTB NOP NOP MOVLW 10001100 MOVWF PORTB NOP NOP MOVLW B 10100100 MOVWF PORTB NOP NOP MOVLW B 11100110 MOVWF PORTB NOP NOP MOVLW B 11100110 MOVWF PORTB NOP NOP MOVLW B 01110110 MOVWF PORTB NOP NOP MOVLW B 01110110 MOVWF PORTB NOP NOP MOVLW 00110011 MOVWF PORTB NOP NOP MOVLW 00010011 MOVWF PORTB NOP NOP MOVLW B 1001001 1 MOVWF PORTB NOP NOP MOVLW B 1001001 1 MOVWF PORTB NOP NOP MOVLW 11000001 MOVWF PORTB NOP NOP MOVLW 11001001 MOVWF PORTB NOP NOP MOVLW B 01001001 MOVWF PORTB NOP NOP MOVLW B 01101001 MOVWF PORTB NOP NOP MOVLW B 00101000 MOVWF PORTB NOP NOP MOVLW B 00101000 MOVWF PORTB NOP NOP MOVLW
113. 01101 MOVWF PORTB NOP NOP MOVLW B 00101100 MOVWF PORTB NOP NOP MOVLW B 00101100 MOVWF PORTB NOP NOP MOVLW 10111100 MOVWF PORTB NOP NOP MOVLW B 10110100 MOVWF PORTB NOP NOP MOVLW B 11110110 MOVWF PORTB NOP NOP MOVLW B 11010110 MOVWF PORTB NOP NOP MOVLW B 01010110 MOVWF PORTB NOP NOP MOVLW B 01010110 MOVWF PORTB NOP NOP MOVLW 00000011 MOVWF PORTB NOP NOP MOVLW 00000011 MOVWF PORTB NOP NOP MOVLW 10000011 MOVWF PORTB NOP NOP MOVLW B 1010001 1 MOVWF PORTB NOP NOP MOVLW 11100001 MOVWF PORTB NOP NOP MOVLW B 11101001 MOVWF PORTB NOP NOP MOVLW B 01111001 MOVWF PORTB NOP NOP MOVLW B 01111001 MOVWF PORTB NOP NOP MOVLW B 001 11000 MOVWF PORTB NOP NOP MOVLW B 0001 1000 MOVWF PORTB NOP NOP MOVLW B 10011100 MOVWF PORTB NOP NOP MOVLW B 10011100 MOVWF PORTB NOP NOP N N N N N N N N N N N N N N N N N N N N N N MOVLW 11001110 MOVWF PORTB NOP NOP MOVLW 11001110 MOVWF PORTB NOP NOP MOVLW B 01001110 MOVWF PORTB NOP NOP MOVLW B 01100110 MOVWF PORTB NOP NOP MOVLW B 001001 11 MOVWF PORTB NOP NOP MOVLW 00100111 MOVWF PORTB NOP NOP MOVLW B 10110111 MOVWF PORTB NOP NOP MOVLW B 10110111 MOVWF PORTB NOP NOP MOVLW B 11110001 MOVWF PORTB OP OP MOVLW 11010001 MOVWF PORTB OP OP MOVLW B 01010001 MOVWF PORTB OP OP
114. 1 to 128 do Write tempfile InputData row col WriteLn tempfile end WriteLn tempfile endcol write the desired output for the testing data to the text file for row RecLoadDBase GetTotalTrain to RecLoadDBase GetTotalTrain 4 RecLoadDBase GetTotalTest 1 do begin for col 0 to endcol 1 do Write tempfile OutputData row col WriteLn tempfile end CloseFile tempfile end save all the network parameters describing the current test as well as the network training results to a text file so that they can be viewed at a later stage procedure TRecMain Savenetworkperformance1 Click Sender TObject var tempfile textfile filename s string begin the test number is used to specify the filename filename cATestData perf inttostr testnumber txt AssignFile tempfile filename ReWrite tempfile write the training parameters to the file this will allow repeating the test at a later stage since all the required parameters are saved WriteLn tempfile Test number inttostr testnumber WriteLn tempfile Number of outputs inttostr RecLoadDBase GetNumOutputs WriteLn tempfile Training database RecLoadDBase RecTrainName Text WriteLn tempfile Test database RecLoadDBase RecTestName Text WriteLn tempfile Train start inttostr RecLoadDBase RecTrainStart Value WriteLn tempfile Train stop inttostr RecLoadDBase RecTrainStop Value WriteLn tempfile Test start inttostr RecLoadDB
115. 138 3ONVIOnQNOO 39NVIOnQNOO BONVLINAGNOD FONVLONGNOD BONVLONGNOD pe 3114 SSvd MO1 M3O0807HIsnod 31113 SSvd MwO1 43040 HL4NO3 31115 SSvd MO1 N3Q3O0 HIsnO03 31113 SSvd MOT1 H3Q0NO HINnO3 31113 SSvd MO1 30N0 H1Nn043 314 SSvd MO1 M3QONO HlINn03 SSvd MO1 43040 H14NO3 31113 SSvd MO1 MH3QMO H1sn04 NHOIVN3N39 AON3no3W4 Wows v OMVO8 N3AI393M WON 39Nv1I2n0NO3 FIGURE 5 12 Ob od wii 2O 5 uuzumg rcoarc zotr25 lt 2 15 ZZEE ZIVS touts Ia amp r aOo Low to aH 2200 2 lt 00 oantal ors 2 08 2 905 JEEE O ao 25 2 20359 23202 orug Iorot corti aqra 122551 95500 OHSU gt 25094 FHF O Cagar Og LO ud qu yt lt obkazs8 554232 oozou 2 5 2 oOooomr te our rosQu Zona gt wW u S EZ trtan56 oxoo uw uiu Soo nac 20237 cc CC o a OI 49225 2aczoo x0 oa Oros E EE ETE 45 As before DG303 dual CMOS analogue switches were used to emulate the multiplication phase of the synchronous detectors This created a problem for the quad op amps used to generate an inverted and buffered version of the input signal to each DG303 Specifically each DG303 has a typical input capacitance of 7pF Since each
116. 2002 In terms of support a technical assistant from the Cape Technikon was allocated to this project to investigate a particular module of the system namely the transmitter module This will be discussed in greater detail in section 5 2 1 on page 41 Further the services of an external electronic design company were employed for the printed circuit board design and manufacture Finally DebTech was involved in the design and construction of the laboratory scale test rig used to assess the system s performance SYNOPSI S This thesis investigates the use of a dual plane impedance tomography system to calculate the individual mass flow rates of the components in an air gravel seawater mixture The long term goal of this research is to develop a multi phase flowmeter for the on line monitoring of an airlift used in an offshore mining application This requires the measurement of both the individual component volume fractions and their velocities Tomography provides a convenient non intrusive technique to obtain this information Capacitance tomography is used to reconstruct the dielectric distribution of the material within a pipeline It is based on the concept that the capacitance of a pair of electrodes depends on the dielectric distribution of the material between the electrodes By mounting a number of electrodes around the periphery of the pipeline and measuring the capacitances of the different electrode combinations it is possible to recons
117. 3 Bi 84 EACH INDIVIDUAL Sm 2m zm 2m OUTPUT FREQUENCY 92 522 922 IS COMBINATION o 69 n OF THE CRYSTAL FREQUENCY AND CRYSTAL CRYSTAL CRYSTAL CRYSTAL THE DIVISION FACTOR PROGRAMMED IN FREQB FREQD FREQF FREQH THE MICROCONTROLLER B 0 D O z 0 0 z 9 z HIO Dy 8 90 Dy 0 90 By 90 2c 90 95 95 95 4 82 o 4 82 o 4 82 o 4 ga 29 mzo gt 2 mzo 29 mzo 29 929 AO 3o AO zo 46 2 gt Fao ze Fao Ze Bar 25 mox On 1433 JSS ON ARS OV 095 r ozz oz lt r ozz c 455 E 90 990 99 E 056 m rz m Tx m rs m 292 a O75 a maj a od od oa od zm zm zm 94 on 5 Cm T FIGURE 5 15 BLOCK DIAGRAM OF FREQUENCY GENERATOR CIRCUIT BOARD SHOWING THE EIGHT PIC16F84A MICROCONTROLLERS USED TO GENERATE THE EIGHT INDEPENDENT TRANSMITTER FREQUENCIES AND THEIR QUADRATURE WAVEFORMS A BROWN OUT DETECTOR WAS ADDED TO ENSURE THAT THE MICROCONTROLLERS WOULD START UP CORRECTLY Figure 5 15 is a block diagram of the frequency generator board A brown out detector was included to ensure that all eight microcontrollers start up correctly Initial tests revealed that certain transmitters were not operating correctly after switch on Further tests revealed that a brown out condition was occurring during start up Subsequently certain microcontrollers would enter an undefined state resulting in the corresponding transmitters not
118. 30G data acquisition card is sufficient There is an occasional loss of synchronisation between the sample controller board and the data acquisition card This loss of synchronisation is the result of false triggering of the data acquisition card by noise on the external trigger line and the effects are temporary Replacing the ribbon cable connecting the sample controller board to the data acquisition card with a multi core shielded cable should solve this problem In addition the electronics should be modified to improve the noise separation between the digital and analogue ground planes The current flow simulation apparatus is extremely limited in terms of the test cases that it is able to simulate Slippage between the drive belt and pulley means that it is impossible to predict when exactly a bubble should be passing through a measurement section Consequently it cannot be used to verify the time delay response of the measurement electronics In addition it is limited in terms of the size of the bubble and the maximum velocity that it can simulate The o ring drive belt and pulley arrangement should be replaced by a toothed belt and sprocket arrangement Further higher torque stepper motors should be used The MOSFET drivers in the transmitters should be replaced with break before make MOSFET drivers Although the cross conduction circuitry discussed previously greatly reduced the cross conduction in the push pull power MOSFET output stage t
119. 5 output is held high for a period of 180us for this sampling to complete Specifically since the maximum sampling rate of the card is 100ksamples s the minimum time taken to capture 16 samples is 160us In total the time taken to configure the multiplexers and capture 16 voltage readings is 200us corresponding to a potential sampling rate of 80ksamples s Since the required sampling rate is only 51 2ksamples s it is clear that the addition of a sample controller board achieves the desired data capture rate The circuit details of this sample controller board are included in Appendix C on page 128 In summary the operation of the sample controller board is as follows 1 oO DN digital output PAO is held low to keep the 555 in the reset state digital output PA1 is held high to reset the 4 bit counter to 0000 digital output PA1 returns to the low state sampling is started by outputting a high on PAO thus releasing the 555 from the reset state since the multiplexers are already configured for address zero the positive transition of the 555 output results in the voltages on input 0 of all 16 multiplexers being sampled by the PC30G card on the negative transition the counter increments to 0001 and the multiplexers connect the voltages on input 1 to the analogue input channels of the PC30G card on the positive transition these 16 voltages are sampled by the PC30G card this process repeats until the counter o
120. 6 for a resistance tomography system 49 In contrast Beck and Plaskowski state that a separation of three to four pipe diameters is necessary 50 Ultimately it was the physical size and cost that 35 limited the laboratory scale rig to a separation of 500mm which only represents 1 4 pipe diameters separation This will be examined in greater detail when the problem of cross plane interference is addressed at a later stage 5 1 2 Design of transmitter and receiver electrodes and guarding The range and accuracy of the capacitance measurement circuitry determines the number of electrodes used Standing capacitances refer to the capacitances between electrodes when the vessel is filled only with the low permittivity material Clearly the standing capacitances of the diametrically opposite electrodes represent the smallest capacitance that must be reliably measured by the electronics Only by increasing the electrode length can more electrodes be incorporated and the image resolution increased 11 Typically capacitance tomography Systems use 12 electrodes giving 66 independent measurements Since the FDM protocol decreases the number of measurements available for a given number of electrodes the number of capacitance plates was increased to 16 Consequently 64 independent capacitance measurements are obtained for a 16 electrode FDM tomography System Three types of electrode configurations are typically used for capacitance tomography systems
121. 96 251953 255615 251221 2 51221 242188 242676 248291 2 39746 250732 244629 249756 250488 248535 249512 254883 253418 248535 245117 250244 246826 246826 246094 249268 2 49268 245605 249756 254639 249023 247314 25024 251953 244873 243023 247070 250732 250732 2 54395 251221 247803 254395 249756 254639 252930 254150 250732 252197 254150 256348 256836 256348 253906 254639 247070 250000 251221 2 58789 246826 249756 2 49268 Start sampling gt Return x 10 Y 9 READY Current test 1291 Total number of samples 32275 FIGURE 5 19 SCREEN CAPTURE OF THE DATABASE GENERATION FORM AT THE TOP LEFT IS THE GRAPHICAL COMPONENT USED TO SPECIFY THE DESIRED NETWORK OUTPUT WHERE BLUE REPRESENTS SEAWATER WHITE REPRESENTS AIR AND BROWN REPRESENTS GRAVEL THIS WAS A 0 09 VOLUME FRACTION AIR BUBBLE SITUATED NEAR THE LEFT HAND EDGE OF THE RIG AND A 0 04 VOLUME FRACTION GRAVEL BUBBLE SITUATED NEAR THE RIGHT HAND EDGE OF THE RIG BELOW THIS COMPONENT IS A BAR GRAPH OF THE DESIRED VOLUME FRACTIONS TO THE RIGHT OF THESE COMPONENTS ARE TWO TABLES OF THE VOLTAGES CORRESPONDING TO THE CAPACITANCE AND RESISTANCE READINGS RESPECTIVELY BELOW THESE TABLES ARE THE TWO ARROWS USED TO NAVIGATE THROUGH THE DATABASE 62 The data capture program provides an additional feature for the monitoring of the system electronics Since 128 voltage readings are available it is impractical to monitor the performance of the system in terms of individual volt
122. A It will also receive a signal from TxD RxA transmitter B except that this signal will be modulated by the material between transmitter B and receiver A fee It will receive a signal from transmitter C modulated x IxB by the material between transmitter C and receiver A Finally it will receive a signal from transmitter D modulated by the material between transmitter D and RxB j TxC receiver A This configuration achieves maximum parallelism since all measurements can be made imultan ly without havin witch between TATA simultaneously without having to switch betwee DIAGRAM OF A PIPELINE CROSS SECTION WHERE THE different electrode configurations The only remaining LINKAGES BETWEEN THE ELECTRODES DEMONSTRATE THE GENERATION OF THE DIFFERENT MEASUREMENT issue that needs to be addressed is how to separate PERMUTATIONS USING MULTIPLE TRANSMITTERS AND RECEIVERS OPERATING IN PARALLEL WITH FREQUENCY the signals from the different transmitters which are DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY received simultaneously at a receiver In contrast to TDM impedance tomography systems FDM impedance tomography separates the different signals in the frequency domain Specifically each transmitter operates at a different frequency in figure 4 1 above The signal at a receiver output is therefore a superposition of these frequency components where the amplitude and phase of a particular component has been modulated by the con
123. ACTION PROFILES AT THE TOP AND BOTTOM MEASUREMENT SYSTEMS WHICH ARE PLOTTED ONCE A TEST IS COMPLETE THE BLUE PLOT REPRESENTS THE BOTTOM MEASUREMENT SYSTEM AND THE RED PLOT REPRESENTS THE TOP MEASUREMENT SYSTEM BELOW THESE VOLUME FRACTION PROFILES ARE THE AIR AND GRAVEL CROSS CORRELATION FUNCTIONS WHICH ARE ALSO PLOTTED ONCE A TEST IS COMPLETE Cross correlation flowmeters require a zero flow detector because cross correlators are incapable of measuring flow below a certain minimum value 50 If there is no flow then the sensor signals will be predominantly electrical noise The cross correlator will always attempt to perform a correlation Consequently spurious velocity measurements will result from the correlation of this random electrical noise Since the laboratory scale rig does not have a zero flow detector it was necessary to code thresholds into the cross correlation software Firstly the volume fraction thresholds are used to remove a constant output offset from the volume fraction profile if such an offset exists Secondly a threshold is performed on the peak of the cross correlation function If the peak of the cross correlation is not larger than the threshold then it is determined that the peak of the cross correlation is probably the result of noise Consequently the phase velocity of the corresponding component is predicted as Zero 86 The network weights and data pre processing are loaded for the top and bottom networks
124. ATED NEAR THE LEFT HAND EDGE OF THE RIG THE DYNAMIC VOLUME FRACTION ERROR IS CALCULATED AS THE MEAN ABSOLUTE ERROR BETWEEN THE DESIRED AND MEASURED VOLUME FRACTION PROFILES To calculate the desired volume fraction profiles the dimensions of the bubbles used in the test must be specified as well as the positions of the top and bottom measurement electrodes relative to the reference marker discussed earlier Using this information the software is able to calculate the desired volume fractions as follows since the frame rate is known the sample time corresponding to a particular frame number can be calculated Then since the acceleration and velocity of the bubble is known the position of the bubble can be calculated for the corresponding frames Since the position and size of the bubble is known and the position of the measurement electrodes has been specified the desired volume fraction for a particular frame can be calculated In particular the volume fractions are calculated as the fraction of the bubble within the cylindrical volume of the measurement electrodes These desired volume fraction profiles are plotted in red and the measured volume fraction profiles in blue The volume fraction error for the dynamic performance of the system can now be calculated 89 The dynamic volume fraction error is the mean absolute error between the desired volume fraction profile and the measured volume fraction profile Since the desired volume fractio
125. B and van den Bleek C M 1999 Image Reconstruction of an Electrical Capacitance Tomography System Using an Artificial Neural Network Proc 1 World Congr on Industrial Process Tomography Buxton pp 174 180 West R M Jia X and Williams 1999 Parametric Modelling in Industrial Process Tomography Proc 1 World Congr on Industrial Process Tomography Buxton pp 444 450 Isaksen 1996 A review of reconstruction techniques for capacitance tomography Meas Sci Technol 7 325 337 Dickin F and Wang M 1996 Electrical resistance tomography for process applications Meas Sci Technol 7 247 260 Loh W W and Dickin F J 1996 Improved modified Newton Raphson algorithm for electrical impedance tomography Electron Lett 32 206 207 Bishop C M 1995 Neural Networks for Pattern Recognition New York Oxford University Press Nooralahiyan A Y Hoyle B S and Bailey N J 1994 Neural network for pattern association in electrical capacitance tomography Proc IEE 141 517 521 Nooralahiyan A Y and Hoyle B S 1997 Three component tomographic flow imaging using artificial neural network reconstruction Chem Eng Sci 52 2139 2148 Ratajewicz Mikolajczak E Shirkoohi G H and Sikora J 1998 Two ANN Reconstruction Methods for Electrical Impedance Tomography IEEE Trans Magn 34 2964 2967 Adler A and Guardo R 1994 A Neural Network Image Reconstruction Technique for Electrical Impedance Tomography EEE Trans Med Imag 13 594 560 Riedmiller M and Braun H 1993 A Direct Adapt
126. CTRODE PAIR OF THE TOP SYSTEM WHEN ONLY THE TOP SYSTEM IS OPERATING AND WHEN BOTH THE TOP AND BOTTOM MEASUREMENT PLANES ARE OPERATING AT THE SAME TIME THE RIGHT HAND GRAPH SHOWS THE EFFECT OF THE BEAT FREQUENCY BETWEEN THE TWO SIMILAR TRANSMITTER A OUTPUT FREQUENCIES AT THE TOP AND BOTTOM MEASUREMENT PLANES 91 A grounded band between the two planes of measurement electrodes has been used to reduce cross plane interference in capacitance tomography systems 14 The same has also been used in resistance tomography Systems A galvanised steel disk was inserted between two rubber insertion gaskets at the flange joining the top and bottom measurement planes The disk protruded slightly into the pipeline and hence made electrical contact with the water It was grounded and the interference analysis test was repeated By introducing a grounded band between the two measurement planes the ripple on the output voltages was substantially reduced To reduce the interference even further the top portion of the laboratory scale rig was inserted between the top and bottom measurement sections as a spacer to increase the plane separation to 2 4 pipe diameters In addition a grounded disk was inserted at the flange joining the top measurement section to the spacer and at the flange joining the spacer to the bottom measurement section The effect of these modifications was to reduce the interference even further Table 7 1 compares the average output voltage rip
127. Click Sender TObject begin testnumber RecNewTestNumber Value RecSetTestNumber Visible false end user decides not to set the test number procedure TRecMain RecSetTestNumberCanClick Sender TObject begin RecSetTestNumber Visible false end allow the user to set the test number procedure TRecMain Settestnumber1 Click Sender TObject begin RecSetTestNumber Visible true end show the form to set the network training parameters procedure TRecMain Setnetworktrainingparameters1Click Sender TObject begin RecSetParam ShowModal end show the form to start the network training procedure TRecMain StarttrainingClick Sender TObject begin RecTrainNetwork ShowModal end show the form to test the network performance on the testing database procedure TRecMain Validatedatabase1 Click Sender TObject begin usecurrentsamples false RecTest ShowModal end show the form to test the network performance on the current samples procedure TRecMain Validatecurrentsamples1Click Sender TObject begin usecurrentsamples true RecTest ShowModal end create the object to access the sampling hardware procedure TRecMain FormCreate Sender TObject begin SampHardware TSamp Create end show real time testing form this is where data is captured from both systems simultaneously and reconstructions are performed in real time with cross correlation algo
128. D CurResTable Cells 5 0 RxE CurResTable Cells 6 0 RxF CurResTable Cells 7 0 RxG CurResTable Cells 8 0 RxH CurCapTable Cells 0 1 TxA CurCapTable Cells 0 2 CurCapTable Cells 0 3 TxC CurCapTable Cells 0 4 TxD CurCapTable Cells 0 5 CurCapTable Cells 0 6 CurCapTable Cells 0 7 TxG 147 CurCapTable Cells 0 8 TxH CurCapTable Cells 1 0 RxA CurCapTable Cells 2 0 RxB CurCapTable Cells 3 0 RxC CurCapTable Cells 4 0 RxD CurCapTable Cells 5 0 RxE CurCapTable Cells 6 0 RxF CurCapTable Cells 7 0 RxG CurCapTable Cells 8 0 RxH CurTimer Enabled true end tM SAMPLING PROCEDURES each time the timer overflows initiate data capture of 1 frame of data procedure TCurMain CurTimerTimer Sender TObject begin SampHardware StartSample 1 end sort the results from the sampunit depending on whether we are sampling from the upper or lower system and display the results in a table procedure TCurMain ProcessResults var row col index overall integer tresult TResultTable S string begin index 1 col 0 row 0 data from the bottom system is offset by 8 with the order being 8 from top system 8 from bottom system 8 from top system and so on if CurSystemSelect ItemIndex 1 then col col 8 while
129. DURES access the network training results if a model search has been conducted then return the results corresponding to the best network function TRecTrainNetwork GetAirError real begin if RecSetParam RecPerformModel Checked then GetAirError bestair else GetAirError airerror end function TRecTrainNetwork GetGravelError real begin if RecSetParam RecPerformModel Checked then GetGravelError bestgravel else GetGravelError gravelerror end function TRecTrainNetwork GetSumVoidError real begin if RecSetParam RecPerformModel Checked then GetSumVoidError bestsum else GetSumVoidError currsum end function TRecTrainNetwork GetThresholdError real begin if RecSetParam RecPerformModel Checked then GetThresholdError bestthresh else GetThresholdError currthresh end function TRecTrainNetwork GetWaterError real begin if RecSetParam RecPerformModel Checked then GetWaterError bestwater else GetWaterError watererror end function TRecTrainNetwork GetNumBest integer begin GetNumBest numbest end end PROGRAM LISTING OF TESTUNIT PAS This unit is used for testing the trained networks the networks can be tested using either data in the test database or on samples obtained from the data aquisition system The reconstructions take place continously at a rate which can be specified by the user For the test database option the program cycles thro
130. Data count 0 totalairpixels var rowindex colindex totalerrors integer begin II initialise the important variables currtrainthresh 0 currtrainsum 0 totalerrors watererror 0 gravelerror 0 airerror 0 for rowindex start to stop do begin totaltraingravelpixels 0 totaltrainairpixels 0 colindex 0 while colindex numoutputs do begin three phase prediction if numoutputs 264 then begin if OutputNetwork rowindex colindex 2 gt OutputNetwork rowindex colindex 1 and OutputNetwork rowindex colindex 2 OutputNetwork rowindex colindex then begin predicts pixel is air totaltrainairpixels totaltrainairpixels 1 but if not air then increment number of errors if QutputData rowindex colindex 2 lt gt 1 then inc totalerrors end else if OutputNetwork rowindex colindex OutputNetwork rowindex colindex 1 and OutputNetwork rowindex colindex gt OutputNetwork rowindex colindex 2 then begin predicts pixel is water but if not water then increment the number of errors if OutputData rowindex colindex lt gt 1 then inc totalerrors end else begin predicts pixel is gravel totaltraingravelpixels totaltraingravelpixels 1 but if not gravel the increment the number of errors if OutputData rowindex colindex 1 lt gt 1 then inc totalerrors end inc colindex 3 end two phase prediction else if numoutputs 88 then begin i
131. DirSelect lItemIndex 0 then begin if StepperConfig AirBubblePosition ItemIndex 0 then tempsteps StepperConfig airsteps else tempsteps StepperConfig airsteps end else begin if StepperConfig AirBubblePosition ItemIndex 0 then tempsteps StepperConfig airsteps else tempsteps StepperConfig airsteps end case StepperConfig AirDrive Value of 1 begin drive1rps StepperConfig airrps drive1steps tempsteps end 2 begin drive2rps StepperConfig airrps drive2steps tempsteps tempsteps StepperConfig gravelsteps end case StepperConfig GravelDrive Value of 1 begin driveirps StepperConfig gravelrps drive1steps tempsteps end 2 begin drive2rps StepperConfig gravelrps drive2steps tempsteps end 3 begin drive3rps StepperConfig gravelrps drive3steps tempsteps end end end start the stepper motion Stepper MoveFixedDist drive1 rps drive2rps drive3rps drive1 steps drive2steps drive3steps keep a record of when started sampling starttime Now II initiate the capture of 50 frames of data SampHardware StartSample 50 end the real time unit operates as follows 1 data is captured in sets of 50 frames and is then separated and pre processed for the top and bottom system 2 initially 50 frames of data is captured 3 once the sampling is complete the system is instructed to capture another set of 50 frames 4 while the sampling is in process the
132. E WITH THE 16 STAINLESS STEEL CONDUCTANCE PROBES VISIBLE AT THE TOP AND BOTTOM MEASUREMENT SECTIONS Stainless steel 5mm machine screws were used for the conductance probes to ensure a high electrical conductivity in comparison to the seawater 17 In addition stainless steel has the advantages of low cost ease of fabrication and resistance to corrosion and abrasion 42 Although standard resistance tomography systems typically use plate electrodes whose dimensions are chosen to optimise the compromise between current injection and voltage detection 58 results can be achieved for point electrodes In general the current injection electrodes of a resistance tomography system should be as large as possible to ensure an even current distribution within the vessel 27 In contrast the voltage measurement electrodes should be as small as possible to avoid averaging over several equipotential values 27 A cause for concern was the fact that a potential difference will exist between the conductance probe and the capacitance plate The voltage on the conductance probe will vary with the received signal whereas the capacitance plate will be held at virtual earth by the transimpedance amplifier Hence it was necessary to keep the conductance probe as small as possible so as to minimise any interference between the capacitance plates and the conductance probes since both measurements are performed simultaneously Although the
133. ECTRODE FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM AND THE 8 ELECTRODE PROTOTYPE FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM FOR A THREE PHASE AIR GRAVEL WATER VOLUME FRACTION PREDICTION USING A DOUBLE LAYER FEED FORWARD NEURAL NETWORK ALL RESULTS REPRESENT THE ABSOLUTE ERROR AND ARE ROUNDED OFF TO TWO SIGNIFICANT DIGITS PERFORMANCE MEASURE 16 ELECTRODE LABORATORY SCALE FDM 8 ELECTRODE PROTOTYPE IMPEDANCE TOMOGRAPHY SYSTEM FDM IMPEDANCE TOP SYSTEM BOTTOM SYSTEM TOMOGRAPHY SYSTEM AIR VOLUME FRACTION ERROR GRAVEL VOLUME FRACTION ERROR WATER VOLUME FRACTION ERROR SUM VOLUME FRACTION ERROR Although an improvement in the accuracy of the volume fraction predictions was achieved through the addition of a hidden layer of neurons the results were still comparable with those of the image reconstruction It is evident from the results of the previous sections that a consistent difference exists between the performance of the top measurement system and the bottom measurement system This difference is probably the result of each System having different amplification settings due to component tolerances and minor variations in the measurement electrodes Further it is the nature of the neural network reconstruction algorithm that this performance difference could be the result of minor variations in the training database generation for the top and bottom systems However the effect of these variati
134. ERROR SUM VOLUME FRACTION ERROR 32 4 2 6 Limitations of the current system Although the above results for the FDM impedance tomography system are promising they do illustrate certain limitations of the prototype system Firstly having to set the cutoff frequency of the low pass filter to only 3Hz limits the frame rate of the system to only 3frames s and consequently negates any possible speed advantages achieved through the use of parallelism Clearly a low pass filter with a steeper rolloff would be required in the next system Alternate low pass filtering techniques were tested on the prototype rig before being implemented on the laboratory scale rig Specifically the LMF100 dual second order switched capacitor filter was tested Secondly the transmitter outputs contain additional frequency harmonics due to the distortions introduced by the power op amp These additional frequency harmonics result in the generation of unwanted frequency components on the DG303 outputs and impose greater limitations on the low pass filtering stage to ensure minimal ripple on the output voltages Research was done at the Cape Technikon to assess the performances of different power op amps 45 Since only a square wave is required various MOSFET drivers were also tested This research concluded that MOSFET drivers should be used in the laboratory scale rig The operation of these drivers will be examined at a later stage Finally the sampl
135. ETWEEN THE CAPACITANCE PLATES INDICATE THE DIFFERENT MEASUREMENT PATHS IN STAGE 2 ELECTRODE B IS CONNECTED TO THE ALTERNATING VOLTAGE SOURCE AND ALL THE REMAINING ELECTRODES EXCEPT ELECTRODE A ARE CONNECTED TO THE TRANSIMPEDANCE AMPLIFIER ONE AT A TIME ELECTRODE AIS NOT CONNECTED TO THE TRANSIMPEDANCE AMPLIFIER SINCE THE CAPACITANCE OF THIS ELECTRODE PAIR HAS ALREADY BEEN MEASURED Alternate data collection protocols do exist For example protocol 2 refers to grouping electrodes and exciting them in pairs An advantage of these more complex data collection protocols is that a larger number of independent measurements are generated for the same number of electrodes hence improving the resolution of the reconstructed images 11 FIGURE 2 2 DIAGRAM OF THE 28 CAPACITANCE MEASUREMENTS IN AN 8 ELECTRODE STANDARD CAPACITANCE TOMOGRAPHY SYSTEM IT SHOWS A CROSS SECTION OF THE PIPELINE AND THE CAPACITANCE PLATES MOUNTED AROUND THE EXTERNAL PERIMETER OF THE PIPELINE THE LINES BETWEEN CAPACITANCE PLATES INDICATE THE DIFFERENT MEASUREMENT PATHS The capacitance of each transmitter receiver electrode pair is measured using one of the following capacitance measurement techniques e Charge discharge capacitance transducer AC based capacitance transducer It has been shown that the AC based transducer has a high signal to noise ratio and good stray immunity compared to the charge discharge technique 6 Stray immunity is the
136. EURAL NETWORK WITH A 1 OF C OUTPUT ENCODING ALL RESULTS REPRESENT THE ABSOLUTE ERROR AND ARE ROUNDED OFF TO TWO SIGNIFICANT DIGITS PERFORMANCE MEASURE TIME DIVISION FREGUENCY DIVISION MULTIPLEXED IMPEDANCE MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM TOMOGRAPHY SYSTEM THRESHOLD ERROR 20 13 AIR VOLUME FRACTION ERROR 9 1 2 9 GRAVEL VOLUME FRACTION ERROR WATER VOLUME FRACTION ERROR SUM VOLUME FRACTION ERROR An additional performance assessment criterion is included labelled sum volume fraction error This is simply the sum of the seawater gravel and air volume fraction errors Since the intended application of the system is the accurate measurement of the volume fractions this measure provides a simple means of comparing the volume fraction accuracies of two networks without having to compare the individual volume fraction error percentages for each of the three phases As can be seen from table 4 5 the FDM impedance tomography system is capable of distinguishing between the gravel and air phases and actually produces more accurate results than the TDM impedance tomography system examined previously In particular the accuracy of the seawater and air volume fraction predictions has improved considerably It is believed that this is a result of isolating the conductance and capacitance measurements on completely separate measurement channels It has been shown that improved volume fraction predictions can
137. Electrical tomography can be classified as electrical resistance tomography electrical capacitance tomography or electro magnetic tomography The term electrical impedance tomography has been used to describe systems where only the resistive component is measured as well as systems where both the capacitance and conductance components are measured thus providing the complex impedance of the vessel contents In this thesis electrical impedance tomography will be used to describe the latter configuration Electrical capacitance tomography attempts to reconstruct the dielectric distribution within the cross section of a vessel based on the capacitances measured between plates mounted on the wall of the vessel Capacitance tomography is used for visualising the internal structure of various industrial processes involving non conducting substances 6 It operates on the principle that the capacitance of a pair of electrodes depends on the dielectric distribution of the material between the electrodes By mounting several of these electrodes on the periphery of the vessel it is possible to reconstruct an image of the vessel contents Electrical resistance tomography attempts to reconstruct the resistivity distribution within the cross section of a vessel based on conductance measurements between electrodes mounted in the wall of the vessel It is intended for the imaging of multi phase conductive mixtures In order to discriminate between more than two p
138. FDesired row col water then begin Canvas Brush Color clBlue Canvas Pen Color clBlue end else if FDesired row col gravel then begin Canvas Brush Color clGray Canvas Pen Color clGray end else if FDesired row col air then begin Canvas Brush Color clWhite Canvas Pen Color clWhite end else begin Canvas Brush Color clBtnFace Canvas Pen Color clBtnFace end Canvas Rectangle X Y X 20 Y 20 end end end 11 reset all the pixels in the pipeline to water procedure TTomo ResetDesired var row col bottom top integer begin FWaterVolume 100 FGravelVolume 0 FAirVolume 0 Canvas Pen Color clBlue Canvas Pen Width 1 Canvas Brush Color clBlue for row 0 to 9 do begin for col 0 to 9 do begin FDesired row col outpipe end end for row 0 to 9 do begin case row of 0 begin top 2 bottom 7 end 1 begin top 1 bottom 8 end 8 begin top 1 bottom 8 end 9 begin top 2 bottom 7 end 2 7 begin top 0 bottom 9 end end for col top to bottom do begin FDesired row col water end end end adjust the volume fractions of the different phases according to the change in pixel at position row col to phase value procedure TTomo SetDesired row col integer value TBubblePhase begin if row lt 9 and col lt 9 then begin if FDesired row col lt gt outpipe then begin if FDesired row
139. ING OF INVERTERS AND BUFFERS FOR THE SYNCHRONOUS DETECTION OF THE RECEIVED SIGNALS 118 CIRCUIT DIAGRAM SHOWING SYNCHRONOUS DETECTION OF THE CAPACITANCE AND CONDUCTANCE SIGNALS FROM RECEIVER A AND RECEIVER B 119 CIRCUIT DIAGRAM SHOWING SYNCHRONOUS DETECTION OF THE CAPACITANCE AND CONDUCTANCE SIGNALS FROM RECEIVER C AND RECEIVER D 120 CIRCUIT DIAGRAM OF MULTIPLEXING STAGE FOR CAPTURE OF THE 16 CAPACITANCE AND 16 CONDUCTANCE READINGS USING A PC30G DATA ACQUISITION CARD 121 APPENDI X C CI RCUI T DI AGRAMS FOR 16 ELECTRODE FREQUENCY DIVI SI ON MULTI PLEXED I MPEDANCE TOMOGRAPHY SYSTEM CIRCUIT DIAGRAM OF THE DIFFERENT TRANSMITTERS WHERE THE HIGH CURRENT TRANSMITTER HAS AN ADDITIONAL PUSH PULL POWER MOSFET OUTPUT STAGE 122 CIRCUIT DIAGRAM OF CAPACITANCE AND CONDUCTANCE RECEIVER SHOWING ADDITIONAL LOW PASS AND BANDPASS FILTER STAGES 123 CIRCUIT DIAGRAM OF FREQUENCY GENERATOR BOARD 124 CIRCUIT DIAGRAM OF INPUT STAGE AND MULTIPLEXING OF THE RECEIVER A AND B SYNCHRONOUS DETECTOR BOARD 125 CIRCUIT DIAGRAM SHOWING THE SYNCHRONOUS DETECTION OF THE CAPACITANCE AND CONDUCTANCE SIGNAL FROM RECEIVER A 126 CIRCUIT DIAGRAM SHOWING THE SYNCHRONOUS DETECTION OF THE CAPACITANCE AND CONDUCTANCE SIGNAL FROM RECEIVER B 127 CIRCUIT DIAGRAM OF THE SAMPLE CONTROLLER BOARD 128 APPENDIX D TABLES OF THE CAPACI TANCE AND CONDUCTANCE READI NGS FOR 16 ELECTRODE FREQUENCY DI VI SI ON MULTI PLEXED MPEDANCE TOMOGRAPHY SYSTEM
140. ION PROCEDURES to verify the dynamic performane of a test show the dynverify form procedure TRealMain VerifyResultsClick Sender TObject begin DynVerify ShowModal end to view real time results after a test enable the realtimer procedure TRealMain RealViewClick Sender TObject begin if RealTimer Enabled then begin RealTimer Enabled false RealView Caption Continue end else begin RealTimer Enabled true RealView Caption tests end end each time the timer overflows show the network prediction for that particular frame show the sample time of that particular frame and move viewposition to the next frame in the matrix of results procedure TRealMain RealTimerTimer Sender TObject begin RealTimer Interval floor 1000 1 RealFrameRate Value PlotTestOutput viewposition frmnumber Caption inttostr viewposition DecodeTime outputstore viewposition sampletime hours mins secs msecs smins Caption inttostr mins min ssecs Caption inttostr secs smsecs Caption inttostr msecs ms inc viewposition ViewInc Value I if have reached the end of the results then loop back to the start if viewposition gt DesiredFrames Value then viewposition ViewInc Value rM M FILE HANDLING PROCEDURES load the neural network weights from the specified files procedure TRealMain RealWeightsClick Sender TO
141. ITTERS RECEIVERS AND MAIN SYNCHRONOUS DETECTION BOARD IN THE 8 ELECTRODE FREGUENCY DIVISION MULTIPLEKED IMPEDANCE TOMOGRAPHY SYSTEM ONLY ONE COMPLETE MEASUREMENT PATH IS INCLUDED IN THIS DIAGRAM WHEREAS A TYPICAL FREGUENCY DIVISION MULTIPLEKED IMPEDANCE TOMOGRAPHY SYSTEM CONSISTS OF MULTIPLE MEASUREMENT PATHS EACH MEASUREMENT PATH CONSISTS OF THE FREGUENCY GENERATOR THE TRANSMITTER THE VESSEL ITSELF THE CONDUCTANCE AND CAPACITANCE RECEIVERS THE SYNCHRONOUS DETECTORS THE DATA ACQUISITION CARD AND THE RECONSTRUCTION COMPUTER THE OPERATION OF EACH OF THESE MODULES WILL BE EXAMINED IN THE FOLLOWING SECTIONS 4 2 41 Generation of transmitter frequencies Although the previous analysis was for sine wave excitation cost constraints limited the application of the FDM technique to square waves The use of square wave excitation is still theoretically correct However instead of a single frequency component for each transmitter the square wave outputs consist of a number of additional odd harmonics of the fundamental frequency as is well known in Fourier analysis 41 pp 9 45 The implications of this will be examined at a later stage Since the conductance measurements are based on a two electrode measurement protocol the contact impedance between the electrode and the substance being imaged becomes an important parameter As mentioned previously this contact impedance is the result of electrochemical reactions taking place between the me
142. LATOR CONSISTING OF AN ANALOGUE MULTIPLIER AND A LOW PASS FILTER TO PRODUCE A VOLTAGE PROPORTIONAL TO THE CAPACITANCE BETWEEN THE CORRESPONDING TRANSMITTER RECEIVER ELECTRODE PAIR THE T SWITCH CMOS NETWORK IS USED TO AVOID UNWANTED COUPLING BETWEEN THE TRANSMITTER AND RECEIVER CHANNELS WITHIN THE SAME TRANSDUCER WHEN IT IS USED AS A RECEIVER SOURCE Gamio et al 6 The frame rates of standard capacitance tomography data acquisition systems are fairly limited Jaworski and Dyakowski made use of a dual plane capacitance tomography system to monitor a pneumatic conveying system The data acquisition rate of the 8 electrode capacitance tomography system was 100frames s from both planes simultaneously 5 Yang developed an AC based 12 electrode capacitance tomography system using phase sensitive demodulation The frame rate achieved was 140frames s 16 2 2 Standard Conductance Tomography Measurement Systems Resistance tomography systems reconstruct the resistivity distribution of the pipeline cross section To achieve this a number of conductance probes are mounted in the pipeline wall around the measurement volume The standard data collection protocol for resistance tomography is the four electrode adjacent pair measurement protocol It operates by injecting an alternating current from one electrode to another for an adjacent pair of electrodes The voltages between all remaining adjacent electrode pairs are then measured 17 The current
143. LE 1 1 REQUIRED SPECIFICATIONS OF THE FLOWMETER TO BE DEVELOPED FOR THE ON LINE MONITORING OF AN AIR GRAVEL SEAWATER MIXTURE IN AN OFFSHORE MINING APPLICATION DESCRIPTION SPECIFICATION PIPELINE DIAMETER 350mm MIKTURE VELOCITY up to 20m s SOLIDS CONTENT 5 to 10 AIR CONTENT approximately 30 CARRIER MEDIUM seawater Initially two commercial impedance tomography systems were to be purchased These commercial systems would then have been used to conduct the above tests However a literature study revealed that the commercially available tomography systems were unable to operate at the frame rate necessary for the intended application In addition cost was a prohibitive factor Consequently the objectives of this thesis were expanded to include the development of a low cost high frame rate impedance tomography system Appendix A on page 116 contains a table of the approximate costs involved in manufacturing the system discussed in this thesis In comparison to commercially available systems this new system clearly meets the requirement of being a low cost system Typical electrical tomography systems consist of the sensors data acquisition system and reconstruction computer 15 As mentioned previously research has been done by the author on the use of neural networks as an alternate reconstruction algorithm for an 8 electrode impedance tomography system 3 4 Although the results of this thesis rely heavily on the reconstructi
144. LECTRODE FREQUENCY DI VISI ON MULTI PLEXED MPEDANCE TOMOGRAPHY SYSTEM LOOP INCLUDE lt PIC1684 SRC gt BSF STATUS RPO CLRF TRISB STATUS RPO CLRF PORTB MOVLW B 00000010 MOVWF PORTB NOP NOP MOVL W B 0000001 1 MOVWF PORTB NOP NOP MOVL W B 00000001 MOVWF PORTB NOP NOP MOVL W B 00000000 MOVWF PORTB GOTO LOOP 134 APPENDI X G FUNDAMENTAL FREQUENCI ES AND HARMONI CS FOR 8 ELECTRODE FREQUENCY DI VI SI ON MULTI PLEXED I MPEDANCE TOMOGRAPHY SYSTEM HARMONIC TxA TxB TxC TxD APPENDI X H FUNDAMENTAL FREQUENCI ES AND HARMONI CS FOR 16 ELECTRODE FREQUENCY DI VI SI ON MULTI PLEXED I MPEDANCE TOMOGRAPHY SYSTEM TxB T T T HARMONIC TxA TxH TxD xE xF 136 APPENDIX I AUTOMATI C CODE GENERATOR PROGRAM CODE PROGRAM LISTING OF AUTONEW PAS program automatically generates the program code for the PIC16F84 to achieve a specific division factor The code is saved to a temporary textfile compiled using MPASM and downloaded using the software provided with the PIC programmer The output square waves are generated using the following sequence of bit sets 1 1 0 PORTB11 1 0 0 unit autonew interface uses Windows Messages SysUtils Classes Graphics Controls Forms Dialogs StdCtrls Spin type TAutoMain class TForm Label1 TLabel DivSelect TSpinEdit Generate TButton Compile TButton Downl
145. LoadReturn TSpeedButton RecLoadProgress TPanel Label11 TLabel RecLoadProgressBar TProgressBar Session TSession RecDesiredRecon TRadioGroup RecLoadStatus TLabel RecLoadCalibrate TCheckBox Label12 TLabel Label13 TLabel RecTrainVers TSpinEdit RecTestVers TSpinEdit RecDesirePhase TRadioGroup procedure RecLoadReturnClick Sender TObject procedure FormShow Sender TObject procedure RecLoadDataClick Sender TObject procedure FormClose Sender TObject var Action TCloseAction procedure RecPreProcessClick Sender TObject private Private declarations trainnumcases integer testnumcases integer traintotal integer testtotal integer trainnumvrs integer total number of training data samples total number of testing data samples I total training samples loaded total testing samples loaded number of versions for each training sample number of versions for each testing sample containsgravel Boolean testnumvrs integer loaded data contains gravel 156 containsair Boolean 1 loaded data cotains air threephase Boolean performing a three phase reconstruction tomosystem integer II specify which system loaded data is from opened Boolean opened databases numoutputs integer number of network outputs TrainInpTable TTable table to load training input data from training database TrainOutTable TTable table to loa
146. MASS FLOW MEASUREMENT OF MULTI PHASE MI XTURES BY MEANS OF TOMOGRAPHI C TECHNI QUES Gavin Teague Thesis Presented for the Degree of DOCTOR OF PHILOSOPHY in the Department of Electrical Engineering at the UNIVERSITY OF CAPE TOWN September 2002 ACKNOWLEDGEMENTS The author wishes to thank the following people for the assistance they provided over the duration of this thesis Prof J Tapson for his knowledgeable supervision and guidance over the past two years DebTech and specifically A Roebl for the financial backing of the project and belief in its potential V Ramphal for the development of the MOSFET driver circuitry at the Cape Technikon B Teague for his excellent advice and assistance in the design and construction of the flow simulation apparatus and other test equipment M Teague for the endless supply of nylon stockings used to construct the gravel masses TERMS OF REFERENCE This project was initiated by A Roebl of DebTech Control and Instrumentation on 1 January 2001 The long term goal of this research is to develop an instrument for the monitoring of a multi phase airlift in an offshore mining application Previous research had been done to assess how applicable impedance tomography is to the monitoring of an air gravel seawater mixture However the previous work has considered only static pipeline configurations This research was initiated to assess the performance of impedance tomography for the on line monitoring of a
147. MPLING COMPLETES SEND A WINDOWS NOTIFICATION MESSAGE TO THE CALLING APPLICATION FIGURE 5 18 DIAGRAM ILLUSTRATING THE INTERACTIONS BETWEEN THE CALLING APPLICATION SAMPLE UNIT AND EDR DEVICE DRIVER WHERE THE ARROWS INDICATE THE PROGRAM EXECUTION REQUIRED TO CAPTURE A SET OF FRAMES USING STREAMING Calling StartSample initiates the data capture process The sampling unit uses a buffer to store the results from the streaming It instructs the EDR device driver to begin sampling as a background process and specifies the number of samples to be captured It also provides the device driver with the address in memory of the buffer as the desired location for the samples Sampling starts by releasing the 555 from the reset state as discussed before and the program execution returns to the calling application From this point on the sampling is completely controlled by the EDR device driver as a background process Consequently the calling application can proceed with some completely unrelated task while the device driver captures the next set of frames As will be seen when real time on line reconstructions are examined this provides a convenient time for the image reconstructions to be performed for the previous set of captured frames Once the required number of frames has been captured the EDR device driver posts a Windows notification message to the calling application To ensure that this message is captured by the calling application it is
148. NGTH 60mm 90mm 190mm TOP MEASUREMENT 0 36 PLANE 4 j BOTTOM 0 31 MEASUREMENT PLANE E i 7 3 2 Reconstruction results for an air plug passing through the measurement section As discussed previously the bubbles used in training database generation were not allowed to come into contact with the conductance probes Although this has increased the resolution at the centre of the pipeline it was not known how the system would react to having one or all 16 conductance probes temporarily immersed in air In order to simulate the presence of this air plug it was decided that the water level in the rig should be lowered past the level of the conductance probes while performing an on line real time reconstruction Although these tests were only conducted for the top measurement plane it is believed that similar results would be achieved if the test were repeated for the bottom measurement plane 96 This test should also provide a useful assessment of the axial guard electrodes situated above and below the central plane of measurement electrodes In theory the capacitance measurements should only be dependent on the material within the volume enclosed by the measurement electrodes Subsequently there should be no change in the capacitance readings as the water level in the pipeline drops until it reaches the top of the measurement electrodes If the capacitance readings change while the water level is within the region of the guard elec
149. OLUME FRACTION ERROR 14 1 4 5 0 WATER VOLUME FRACTION ERROR 1 4 1 4 5 0 Figure 6 5 on page 74 is a sequence of screen captures illustrating the neural network reconstruction results for certain test cases from the top measurement system For each screen capture the desired output is the right hand frame and the network prediction is the left hand frame Certain conclusions can be made regarding the 16 electrode FDM impedance tomography system based on the screen captures in figure 6 5 e extending the training database to include examples of the masking effect the neural network reconstruction accuracy for these configurations is greatly improved For example the test cases in figure 6 5 reveal that when a bubble is situated towards the edge and another bubble is situated at the centre of the pipeline the bubble at the centre is also detected Previously only the bubble towards the edge would have been detected e resolution of the system has improved to where the 10 bubble can be detected at the centre of the pipeline Previously the 8 electrode prototype was unable to detect the 2096 bubble at the centre of the pipeline By increasing the number of electrodes to 16 and by not allowing the training database bubbles to come into contact with the conductance probes the resolution at the centre of the pipeline has increased to 1096 of the pipeline diameter corresponding to a 0 01 volume fraction Since the 1096 bubble is
150. ONDUCTANCE COMPONENTS THE RECEIVED SIGNAL IS MULTIPLIED DIRECTLY BY THE CORRESPONDING REFERENCE WAVEFORM THE OUTPUT OF THE LOW PASS FILTER IS A VOLTAGE PROPORTIONAL TO THE CONDUCTANCE BETWEEN THE CORRESPONDING TRANSMITTER RECEIVER ELECTRODE PAIR FOR THE CAPACITANCE COMPONENTS THE RECEIVED SIGNAL IS MULTIPLIED BY A 90 PHASE SHIFTED VERSION OF THE CORRESPONDING REFERENCE WAVEFORM THE 90 BLOCK REPRESENTS THE INTRODUCTION OF THIS 90 PHASE SHIFT IN THE REFERENCE WAVEFORM IN THE ABOVE DIAGRAM THE OUTPUT OF THE LOW PASS FILTER IS A VOLTAGE PROPORTIONAL TO THE CAPACITANCE BETWEEN THE CORRESPONDING TRANSMITTER RECEIVER ELECTRODE PAIR Figure 4 2 illustrates the application of this concept to an 8 electrode system Specifically only the signal from receiver A will be examined although it is a simple extension to perform the same analysis for the other receivers As mentioned previously the output from receiver A consists of a superposition of the signals from transmitter A B C and D This signal is connected to eight synchronous detectors in parallel For example the first synchronous detector multiplies the received signal by the reference for transmitter A After low pass filtering a DC voltage is produced whose magnitude is proportional to the conductance between receiver A and transmitter A The second synchronous detector multiplies the received signal by a 90 phase shifted version of the reference for transmitter A After low pass filtering a
151. ONDUCTANCE RECEIVER FOR THE 8 ELECTRODE FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM THE RECEIVER ELECTRODE CONSISTS OF A CAPACITANCE PLATE AND A CONDUCTANCE PROBE THAT ARE NOT IN ELECTRICAL CONTACT THE CAPACITANCE RECEIVER IS A TRANSIMPEDANCE AMPLIFIER THE CONDUCTANCE RECEIVER CONSISTS OF AN INPUT RESISTOR TO GROUND FOLLOWED BY A BUFFER The capacitance receiver is a transimpedance amplifier In the ideal case of capacitor input and feedback elements the output voltage Vo is given by the following equation relating the feedback capacitance to the unknown capacitance for a transmitter voltage Ve 6 Vo mc 4 16 Ideally Ce should be chosen as small as possible For capacitance feedback to dominate the feedback resistor and capacitor component values must be chosen so that lt lt Re 16 The resistive feedback is required to ensure that the output does not saturate as a result of the DC bias current charging the feedback capacitor 6 27 Unlike standard tomography systems the receiver outputs for FDM impedance tomography consist of a superposition of waveforms corresponding to the signals from the different transmitters In addition square waves are used instead of sine waves Although it may have been possible to calculate theoretically ideal resistor and capacitor values for the receiver the application of these values to this specific application would not have been practi
152. OP MOVLW B 10011111 MOVWF PORTB NOP NOP MOVLW B 10011111 MOVWF PORTB NOP NOP MOVLW B 11001101 MOVWF PORTB NOP NOP MOVLW B 11001101 MOVWF PORTB NOP NOP MOVLW B 01001101 MOVWF PORTB NOP NOP MOVLW B 01100101 MOVWF PORTB NOP NOP MOVLW 00100100 MOVWF PORTB NOP NOP MOVLW 00100100 MOVWF PORTB NOP NOP MOVLW B 10110100 MOVWF PORTB NOP NOP MOVLW B 10110100 MOVWF PORTB NOP NOP MOVLW B 11110010 MOVWF PORTB NOP NOP MOVLW B 11010010 MOVWF PORTB NOP NOP MOVLW B 01010010 MOVWF PORTB NOP NOP MOVLW B 01010010 MOVWF PORTB NOP NOP MOVLW B 0000001 1 MOVWF PORTB NOP NOP MOVLW 00001011 MOVWF PORTB NOP NOP MOVLW B 10001011 MOVWF PORTB NOP NOP MOVLW B 10101011 MOVWF PORTB NOP NOP MOVLW B 11101001 MOVWF PORTB NOP NOP MOVLW B 11101001 MOVWF PORTB NOP NOP MOVLW B 01111101 MOVWF PORTB NOP NOP MOVLW B 01111101 MOVWF PORTB NOP NOP MOVLW B 00111100 MOVWF PORTB NOP NOP MOVLW B 00011100 MOVWF PORTB NOP NOP MOVLW B 10011100 MOVWF PORTB NOP NOP MOVLW B 10010100 MOVWF PORTB NOP NOP MOVLW B 11000110 MOVWF PORTB NOP NOP MOVLW B 11000110 MOVWF PORTB NOP NOP MOVLW 01000110 MOVWF PORTB NOP NOP MOVLW B 01100110 MOVWF PORTB NOP NOP MOVLW 00100011 MOVWF PORTB NOP NOP MOVLW 00100011 MOVWF PORTB NOP NOP MOVLW B 10110011 MOVWF PORTB NOP NOP MOVLW 10110011 MOVWF PORTB NOP NO
153. P MOVLW B 11110001 MOVWF PORTB NOP NOP MOVLW 11011001 MOVWF PORTB NOP NOP MOVLW B 01011001 MOVWF PORTB NOP NOP MOVLW B 01011001 MOVWF PORTB NOP NOP MOVLW 00001000 MOVWF PORTB NOP NOP MOVLW 00001000 MOVWF PORTB NOP NOP MOVLW 10001100 MOVWF PORTB NOP NOP MOVLW B 10101100 MOVWF PORTB NOP NOP MOVLW B 11101110 MOVWF PORTB NOP NOP MOVLW B 11101110 MOVWF PORTB NOP NOP MOVLW 01111110 MOVWF PORTB NOP NOP MOVLW B 01110110 MOVWF PORTB NOP NOP MOVLW B 00110111 MOVWF PORTB NOP NOP MOVLW B 000101 11 MOVWF PORTB NOP NOP MOVLW B 10010111 MOVWF PORTB NOP NOP MOVLW B 10010111 MOVWF PORTB NOP NOP MOVLW 11000001 MOVWF PORTB NOP NOP MOVLW 11000001 MOVWF PORTB NOP NOP MOVLW B 01000001 MOVWF PORTB NOP NOP MOVLW B 01100001 MOVWF PORTB NOP NOP MOVLW B 00100000 MOVWF PORTB NOP NOP MOVLW B 00101000 MOVWF PORTB NOP NOP MOVLW 10111000 MOVWF PORTB NOP NOP MOVLW 10111000 MOVWF PORTB NOP NOP MOVLW B 11111010 MOVWF PORTB NOP NOP MOVLW 11011010 MOVWF PORTB NOP NOP MOVLW B 01011110 MOVWF PORTB NOP 130 NOP MOVLW B 01011110 MOVWF PORTB NOP NOP MOVLW B 00001 111 MOVWF PORTB NOP NOP MOVLW B 00001 111 MOVWF PORTB NOP NOP MOVLW B 10001111 MOVWF PORTB NOP NOP MOVLW B 10100111 MOVWF PORTB NOP NOP MOVLW B 11100101 MOVWF PORTB NOP NOP MOVLW B 11100101
154. PTURES IS LEFT TO RIGHT AND ONLY EVERY 50 FRAME IS CAPTURED AS BEFORE THE BLACK IS THE WATER PHASE THE WHITE IS THE AIR PHASE THE DARK GREY IS THE GRAVEL PHASE AND THE LIGHT GREY REPRESENTS THE REGION OUTSIDE THE PIPELINE t ntm ma A 2 Figure 7 18 is a sequence of screen captures corresponding to the specific configuration of a 0 04 volume fraction air bubble situated near the right hand edge of the rig and a 0 04 volume fraction gravel bubble situated near the left hand edge of the rig For each screen capture the top frame represents the reconstruction results of the top measurement plane and the bottom frame represents the reconstruction results of the bottom measurement plane The sequence of the screen captures is left to right and only every 50 frame is captured Various characteristics common to the three phase reconstruction results will be analysed with reference to this sequence of screen captures Firstly the air bubble is detected as gravel as it enters the measurement volume It is then correctly detected as air As it exits the measurement volume so the air bubble is once again detected as gravel These gravel transition regions are similar to the two phase air water reconstruction results where the air bubble was first detected as a small air bubble as it entered the measurement volume It can also be seen in figure 7 18 that no transition regions exist for the gravel bubble In other words
155. Plot Series 1 Clear BotAirPlot Series 2 Clear BotGravelPlot Series 0 Clear BotGravelPlot Series 1 Clear BotGravelPlot Series 2 Clear for testindex 1 to RealMain DesiredFrames Value do begin if totaloutputs 2 then begin volume fraction data TopAirPlot Series 1 AddY outputstore testindex prediction 1 TopGravelPlot Series 1 AddY outputstore testindex prediction 2 BotAirPlot Series 1 AddY outputstore testindex prediction 3 BotGravelPlot Series 1 AddY outputstore testindex prediction 4 end else begin I image data TopAirPlot Series 1 AddY volumestore testindex 1 TopGravelPlot Series 1 AddY volumestore testindex 2 BotAirPlot Series 1 AddY volumestore testindex 3 BotGravelPlot Series 1 Add Y volumestore testindex 4 end end VelSource SetFocus end save the results to a textfile so that they may be loaded in MATLAB for analysis at a later stage procedure TDynVerify RealSaveResultsClick Sender TObject var tempfile textfile filename string tempstring string row col totalmsecs integer hours mins secs msecs word topairposition topgravelposition real begin order for output file is sample time interpolated time air bubble position gravel position top air volume fraction top gravel volume fraction bottom air volume fraction bottom gravel fraction desired top air fraction desired top gravel fraction desired bottom air fraction desired bott
156. RPROP VolumeData array of array of real volume fractions OutputWeightBest array of array of real best weights for output layer in model search HiddenWeightBest array of array of real best weights for hidden layer in model search testnumber integer overall test number which is automatically appended to all files saved datacompare array 1 128 of real the dimensions of the matrices is as follows OutputData OutputNetwork traintotal testtotal X numoutputs InputData traintotal testtotal X 129 OutputHidden traintotal testtotal X numhidden 1 HiddenWeight DiffHiddenWeight OldHiddenWeight HiddenDelta Weight 129 Xnumhidden OutputWeight OldOutputWeight DiffOutputWeight OutputDeltaWeight numhidden 1 OR 129 X numoutputs implementation uses loadunit viewunit paramunit networkunit testunit realunit R DFM tM GENERAL PROCEDURES exit the program procedure TRecMain Exit1 Click Sender TObject begin Close end show the load databases form procedure TRecMain Loaddatabase1 Click Sender TObject 152 begin RecLoadDBase ShowModal end show the view databases form procedure TRecMain Viewdatabase1 Click Sender TObject begin RecViewMain ShowModal end when show form load the previous test number and weight factor from the registry procedure TRecMain FormShow Sender TObject var Reg TRegistry KeyGood
157. S OUTPUT LAYER NEURONS PIXEL FIGURE 3 1 DIAGRAM ILLUSTRATING THE USE OF A SINGLE LAYER FEED FORWARD NEURAL NETWORK TO PERFORM A TWO PHASE AIR SEAWATER IMAGE RECONSTRUCTION OF THE PIPELINE CONTENTS THE NEURAL NETWORK CONSISTS OF AN INPUT LAYER OF NEURONS THAT SIMPLY DISTRIBUTES THE 28 CAPACITANCE AND 28 CONDUCTANCE READINGS OBTAINED FROM THE IMPEDANCE TOMOGRAPHY DATA ACQUISITION SYSTEM TO THE OUTPUT LAYER OF NEURONS INCLUDED IN THE INPUT LAYER IS AN OFFSET NEURON WHOSE OUTPUT EQUALS UNITY THE 57 INPUT NEURONS ARE FULLY CONNECTED TO THE 88 OUTPUT NEURONS EACH CONNECTION HAS A WEIGHT VALUE THAT CONTROLS THE STRENGTH OF THE CONNECTION FROM THE CORRESPONDING INPUT NEURON TO THE CORRESPONDING OUTPUT NEURON EACH OUTPUT LAYER NEURON FORMS THE WEIGHTED SUM OF THE INPUT LAYER NEURON OUTPUTS AND THE OUTPUT IS THE HYPERBOLIC TANGENT OF THIS WEIGHTED SUM IF THE OUTPUT IS GREATER THAN OR EQUAL TO ZERO THEN THE PHASE OF THE CORRESPONDING PIXEL IN THE RECONSTRUCTED IMAGE IS SET AS AIR OTHERWISE IT IS SET AS SEAWATER IN THIS WAY AN IMAGE OF THE VESSEL CROSS SECTION IS OBTAINED Figure 3 2 on page 13 illustrates a three phase image reconstruction using a single layer feed forward neural network The structure of this network is very similar to the two phase image reconstruction neural network considered previously However since each pixel can be one of three phases instead of just two a different output layer encoding is required The standard multi class
158. S PREDICTING THE VOLUME FRACTIONS OF THE AIR AND GRAVEL PHASES DIRECTLY EACH OUTPUT LAYER NEURON SIMPLY FORMS THE WEIGHTED SUM OF THE HIDDEN LAYER NEURON OUTPUTS THE SEAWATER VOLUME FRACTION IS CALCULATED AS 1 LESS THE AIR AND GRAVEL VOLUME FRACTIONS IN THIS WAY THE THREE PHASE VOLUME FRACTIONS IN A VESSEL CROSS SECTION ARE OBTAINED 3 2 2 Training multi layer perceptrons The adjustable parameters or weights of the network control the functional mapping and are obtained through network training Training of a neural network is achieved through the presentation of known input output pairs to the network and adjustment of the network weights until the error is satisfactorily small To train a neural network requires a large quantity of representative data Unlike previous research into neural network reconstruction techniques the training data in this study was obtained from an existing impedance tomography rig The generation of this large database required a method of systematically varying the vessel contents A bubble placement system was developed for this task consisting of a bed of nails and clear acrylic lid Figure 6 3 on page 70 is a photograph of a bubble placement system It is used to locate a simulated air or gravel bubble at a 14 defined position within the vessel The bubble placement system plays an important role in the training of a neural network If the position of a bubble within the training database is not precisel
159. SULT OF SYSTEM IMPERFECTIONS THAT ARE NOT RELATED TO THE FREQUENCY MIXING IN THE MULTIPLICATION STAGE OF THE SYNCHRONOUS DETECTORS As can be seen in figure 5 16 the simulation software is able to predict a large percentage of the actual frequency components measured The simulation software has shown that it is impossible to obtain a set of eight transmitter frequencies for which there is no output ripple in a square wave FDM impedance tomography system In a pure sine wave system it is a simple task to choose eight output frequencies for which there will be no output ripple since there are no frequency harmonics to consider A sine wave FDM impedance tomography system is currently being developed at the University of Cape Town for the 8 electrode prototype rig to verify this assumption 62 In theory the sine wave system is expected to produce more accurate results at a higher image resolution However a drawback of the sine wave system is cost since analogue multipliers which are considerably more expensive than the DG303 CMOS switches used in the current system are required to perform the multiplication It is also more complex and expensive to generate a sine wave and its quadrature waveform All tests in the following sections are conducted with the 16 electrode square wave FDM impedance tomography system 56 5 3 Software Design of a 16 electrode Frequency Division Multiplexed Impedance Tomography System The following sections examin
160. TWO SIGNIFICANT DIGITS PERFORMANCE MEASURE TIME DIVISION FREGUENCY DIVISION MULTIPLEXED IMPEDANCE MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM TOMOGRAPHY SYSTEM THRESHOLD ERROR AIR VOLUME FRACTION ERROR WATER VOLUME FRACTION ERROR In a surprising result the FDM impedance tomography system outperforms the TDM impedance tomography system on all the performance criteria considered Since only 32 readings are available for the new system instead of 56 it was expected that the resolution and consequently the accuracy of the reconstructed images would have deteriorated One possible explanation for this improvement in performance is that the FDM impedance tomography system is using separate capacitance and conductance receivers Previously a single receiver was used for both the capacitance and conductance signals Since there is a large conductance component from the seawater the gain on this receiver was fairly low to ensure the op amp output did not saturate Consequently the received capacitance component was very small and was affected by the conductance signal By using separate receivers each receiver can be tuned to give optimal signal detection for that particular measurement hence a much larger capacitance gain was achieved In addition the interference between the two measurements was reduced These results were confirmed when a single layer feed forward neural network was trained using gradient descent to perfor
161. a Fourier transform on the gravel volume fraction profile it was determined that this ripple was at the same frequency as the ripple on the low pass filter outputs of the synchronous detectors This proved that the System was sensitive to this output ripple as expected Greater reconstruction accuracy should therefore be achieved if this output ripple were eliminated through the use of a sine wave excitation system 7 4 2 Cross correlation velocity prediction results Tests were conducted to determine whether the cross correlation algorithm was able to calculate the velocities of individual components in a multi phase flow These tests consisted of a 0 04 volume fraction air bubble and a 0 04 volume fraction gravel bubble moving through the laboratory scale rig at different velocities Figure 7 19 is an example of such a test where the air bubble was set to move at 0 5m s and the gravel bubble was set to move at 0 2m s Included in the figure are the air and gravel volume fraction profiles at the top and bottom measurement planes as well as the air and gravel cross correlation functions The dashed lines represent the desired volume fraction profiles according to the bubble size and flow simulation apparatus configuration An analysis of figure 7 19 reveals certain results TOP SYSTEM AIR VOLUME FRACTION BOTTOM SYSTEM AIR VOLUME FRACTION 10 10 1 AIR CROSS CORRELATION volume fraction 96 volume fraction Yo o
162. a InputData currsample inputnodeindex end end end else begin IL if wish to perform a regression then the output Il activation function is linear if numoutputs 2 then begin if weightsum gt 0 then newnodeoutput weightsum else newnodeoutput 0 newnodederiv 1 end else just two phases so simple tanh output else begin newnodeoutput tanh weightsum newnodederiv CalcDeriv newnodeoutput end calculate the adjustments to the output layer weights OutputNetwork currsample outputnodeindex newnodeoutput delta newnodeoutput OutputData currsample outputnodeindex newnodederiv for inputnodeindex 0 to 128 do begin DiffOutputWeight inputnodeindex outputnodeindex DiffOutputWeight inputnodeindex outputnodeindex delta InputData currsample inputnodeindex end end end end after all the training data has been processed update the weights where the method of update depends on the training technique being I used if useRPROP then AdjustRPROP else AdjustWeights every updatefreq epochs update the training display to indicate the current performance of the network if currepoch mod updatefreq 0 then UpdateNetworkDisplay inc currepoch end RecNetworkProgressBar Position RecNetworkProgressBar Max end procedure to train a double layer feed forward neural network where the hidden layer neurons consist of simple hyperbolic tangent activation functions
163. abases and the number of versions and range of training and test data points to load Once this information has been specified the required data will be loaded by clicking Load databases If Calibrate is selected then additional calibration data will also be required This will be discussed at a later stage once the operation of the software calibration feature has been explained After the data has been loaded it must be pre processed Standardisation of the training data is performed when Pre process data is clicked On returning to the main menu the user has the option of loading a previous set of network weights thus allowing one to perform tests without having to retrain a neural network Since this is a discussion on training and testing a new neural network the next stage requires the setting of the neural network parameters Each network trained has a different test number which can be specified by selecting Set test number in the network training menu All network training results network weights and other data are saved according to this test number Therefore if one wished to load the network weights for a specific network then it would be necessary to specify the corresponding test number beforehand 63 Set the network training parameters X General parameters Early stopping parameters Maximum number of epochs 2000 Number of Failures before training stops E Number of hidden layer neurons mi Stopping
164. ability to measure small inter electrode capacitances while ignoring the large stray capacitances to ground Gamio et a describe a standard AC based capacitance tomography system as follows 6 a sinusoidal signal is generated by a direct digital synthesiser and buffered before being applied to the transmitter electrode through a CMOS switch network The charge on the receiver electrode is measured by an op amp with capacitive and resistive feedback temporarily connected to the receiver electrode through a CMOS switch A programmable gain amplifier amplifies the op amp output before it is demodulated using a phase sensitive demodulator consisting of an analogue multiplier and low pass filter Figure 2 3 is a block diagram of this system T SWITCH NETWORK CAPACITANCE PLATE DDS EXCITATION SOURCE DDS REFERENCE SIGNAL LOW PASS FILTER PHASE SENSITIVE DEMODULATOR PROGRAMMABLE GAIN AMPLIFIER FIGURE 2 3 BLOCK DIAGRAM OF THE TRANSDUCER CIRCUITRY FOR A STANDARD AC BASED CAPACITANCE TOMOGRAPHY SYSTEM WHERE EACH ELECTRODE IS CONNECTED TO ITS OWN TRANSDUCER CIRCUITRY AS A TRANSMITTER ELECTRODE THE T SWITCH CMOS NETWORK CONNECTS THE DIRECT DIGITAL SYNTHESISER DDS TO THE CAPACITANCE PLATE THROUGH A BUFFER AS A RECEIVER ELECTRODE A SINGLE CMOS SWITCH CONNECTS THE CAPACITANCE PLATE TO THE TRANSIMPEDANCE AMPLIFIER THE OUTPUT IS AMPLIFIED AND DEMODULATED USING A PHASE SENSITIVE DEMODU
165. aborate way of generating a square wave and its quadrature but the microcontroller implementation has certain advantages compared to the logic gate based designs namely e Component count since each output square wave and its quadrature are generated using a single 18 pin integrated circuit e Versatility since the output frequency generated can easily be modified by reprogramming the microcontroller The program code implemented on the microcontroller is similar to that used for the 8 electrode prototype However only one square wave is generated and therefore a sequence of only four port set instructions is required to generate the output square wave and its quadrature Between each of the four port set instructions are a number of nop instructions The first three port set instructions are followed by a minimum of two nop instructions The goto instruction after the fourth port set instruction requires two instruction cycles to execute hence the nop instructions ensure the correct duty cycle and phase shift Additional nop instructions can also be added equally after all four port set instructions to decrease the output frequency for a given crystal frequency The program code is essentially implementing a PIC16F84 20 division of the crystal frequency where the io 000000 yA MO MOWE PORTB minimum division factor is 64 corresponding to NOP TNT TERT no additional nop instructions The maxim
166. according to the specified test number when Load network data is clicked A set of calibration data is captured from the rig when Calibrate system is clicked This provides a reference point against which all future tests are compared Specifically these calibration voltages are subtracted from all the voltages captured during testing so that only the differences are input to the neural network reconstruction algorithm An advantage of this approach is that it is possible to calibrate the effect of the flow simulation apparatus in the rig Consequently when a test is conducted the flow simulation apparatus is ignored and only the bubble is detected Clicking on Stepper motor setup displays a form that is used for the configuration of the flow simulator apparatus A screen capture of this form can be found in Appendix N on page 209 This form initialises the stepper controller and allows the user to specify the desired direction and velocity of the simulated bubble flow The desired flow velocity is specified in metres per second Using the pulley diameter and the drive resolution this form converts the velocity and distance parameters into step units that are then used to configure the stepper controller before a test To calculate the required pulley direction it is necessary to indicate whether the bubble is situated on the inside portion of the drive belt or the outside For example figure 7 3 on page 84 is a photograph of the bubble on t
167. action prediction and the averaged results are detailed in table 6 3 As is evident from table 6 3 an improvement in the accuracy of the volume fraction prediction was achieved for both the top and bottom system compared to the image reconstruction results considered in table 6 2 The following section examines the reconstruction results for static three phase air gravel water configurations TABLE 6 3 PERFORMANCE OF THE 16 ELECTRODE FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM FOR A TWO PHASE AIR WATER VOLUME FRACTION PREDICTION USING A SINGLE LAYER FEED FORWARD NEURAL NETWORK ALL RESULTS REPRESENT THE ABSOLUTE ERROR AND ARE ROUNDED OFF TO TWO SIGNIFICANT DIGITS PERFORMANCE MEASURE 16 ELECTRODE LABORATORY SCALE FDM IMPEDANCE TOMOGRAPHY SYSTEM TOP SYSTEM BOTTOM SYSTEM AIR VOLUME FRACTION ERROR 0 20 WATER VOLUME FRACTION ERROR 0 20 74 6 3 2 Three phase air gravel water reconstruction results Multiple single layer feed forward neural networks were trained to perform the three phase air gravel water image reconstruction and the averaged results are detailed in table 6 4 which includes a comparison to the 8 electrode prototype system The networks employ a 1 of C output encoding and early stopping is used to prevent over fitting As is evident from table 6 4 the 16 electrode FDM impedance tomography system outperforms the 8 electrode prototype on all the performance assessment criteria considered This is consi
168. ages By selecting Current samples under the view menu a form opens showing all 128 voltages in a tabular format that is continuously updated A screen capture of this form can be found in Appendix M on page 203 A further feature is provided whereby frames of data can be saved to a single text file at a specified rate This is used for the assessment of the system drift since this text file can be loaded in a spreadsheet and the voltages plotted Having generated a training and test database the neural network program is used to train a neural network to perform reconstruction and this will be examined in the following section 5 3 3 Neural network reconstruction software description As before the underlying structure of the neural network program is very similar to the software developed as part of the author s previous research The commented program code is provided in Appendix L on page 152 The following is not an attempt to analyse the implementation of the neural networks Instead it describes briefly how one would train and verify a neural network Firstly the training and test databases must be loaded Selecting Load database in the file menu opens a form where the specific details of the database are entered A screen capture of this form is included in Appendix N on page 204 In particular the user is required to specify whether image reconstruction or volume fraction prediction is required the names of the training and test dat
169. ails gt maxfails and earlystop do begin for each training point for currsample 0 to totaltrain 1 do begin for each output node calculate output of node for outputnodeindex 0 to numoutputs 1 do begin weightsum 0 calculate the weighted summation to that node for inputnodeindex 0 to 128 do begin weightsum weightsum InputData currsample inputnodeindex OutputWeight inputnodeindex outputnodeindex end prevent arithmetic overflow if weightsum gt 200 then weightsum 200 else if weightsum lt 200 then weightsum 200 IL if using 1 of C then outputs can only be computed once the weights to all three output nodes corresponding to a single pixel have been found if numoutputs 264 then begin if outputnodeindex 1 mod 3 lt gt 0 then weightstore outputnodeindex 1 mod 3 weightsum else begin we can now calculate the outputs of the SoftMax function since we have all three inputs to the 1 nodes corresponding to water gravel and air II respectively for i 2 downto 0 do begin OutputNetwork currsample outputnodeindex i SoftMax 3 i weightstore 1 weightstore 2 weightsum 11 calculate the adjustments to the output layer 11 weights delta OutputNetwork currsample outputnodeindex i OutputData currsample outputnodeindex i for inputnodeindex 0 to 128 do DiffOutputWeight inputnodeindex outputnodeindex i DiffOutputWeight inputnodeindex outputnodeindex i delt
170. ainNetwork implementation uses paramunit loadunit mainunit R DFM tM NETWORK TRAINING PROCEDURES A when start the network training initialise all the important variables and call the appropriate training procedure procedure TRecTrainNetwork RecNetworkStartClick Sender TObject begin if not performingtrain then begin II initialise important variables performingtrain true RecNetworkStart Caption Stop training RecMain Savenetworkperformancet Enabled true RecMain Validatedatabase1 Enabled true RecMain Validatecurrentsamples1 Enabled true RecNetworkProgressBar Position 0 numoutputs RecLoadDBase GetNumOutputs totaltrain RecLoadDBase GetTotalTrain totaltest RecLoadDBase GetTotalTest maxepochs RecSetParam GetMaxEpochs maxfails RecSetParam GetNumFails updatefreq RecSetParam GetUpdateFreq earlystop RecSetParam RecPerformEarlyStop Checked useRPROP RecSetParam RecUseRPROP Checked numhidden RecSetParam RecSetNumHidden Value RecLoadDBase GetPhases containsair containsgravel RecNetworkProgressBar Max RecSetParam GetMaxEpochs div updatefreq if useRProp then begin maxdelta RecSetParam GetDeltaMax mindelta 0 end else begin learnrate RecSetParam GetLearnRate momentum RecSetParam GetMomentum end if RecSetParam RecPerformModel Checked then begin II if wish to perform model search then show the model search
171. aining parameters specific to RPROP procedure TRecSetParam RecUseRPROPClick Sender TObject begin if RecUseRProp Checked then RecRPropParam Visible true else RecRPropParam Visible false end close form procedure TRecSetParam RecLoadReturnClick Sender TObject begin Close end when close form save the current network parameters to the registry and dynamically allocate the memory for the different weight matrices procedure TRecSetParam FormClose Sender TObject var Action TCloseAction var Reg TRegistry begin save the current parameters to the registry Reg TRegistry Create try Reg OpenKey Software Neural True Reg Writelnteger MaxEpochs strtoint RecSetMaxEpochs Text Reg Writelnteger NumHidden RecSetNumHidden Value Reg WriteBool EarlyStop RecPerformEarlyStop Checked Reg WriteBool Model RecPerformModel Checked Reg WriteBool UseGradient RecUseGradDescent Checked Reg WriteFloat Deltalnit strtofloat RecSetDeltalnit Text Reg WriteFloat DeltaMax strtofloat RecSetDeltaMax Text Reg WriteFloat LearnRate strtofloat RecSetLearn Text Reg WriteFloat Momentum strtofloat RecSetMomentum Text Reg Writelnteger NumFails RecNumFails Value Reg WriteBool StopVolume RecStopVolume Checked Reg Writelnteger ModelStart RecModelStart Value Reg Writelnteger ModelStop RecModelStop Value Reg Writelnteger Modellnc RecModellnc Value Re
172. airposition 1 UpdateDisplay end procedure TDynVerify AirShape2MouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Integer begin airposition 2 UpdateDisplay end procedure TDynVerify AirShape3MouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Integer begin airposition 3 UpdateDisplay end procedure TDynVerify AirShape4MouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Integer begin airposition 4 UpdateDisplay end procedure TDynVerify AirShape5MouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Integer begin airposition 5 UpdateDisplay end procedure TDynVerify GravShape1 MouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Integer begin gravelposition 1 UpdateDisplay end procedure TDynVerify GravShape2MouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Integer 199 begin gravelposition 2 UpdateDisplay end procedure TDynVerify GravShape3MouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Integer begin gravelposition 3 UpdateDisplay end procedure TDynVerify GravShape4MouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Integer begin gravelposition 4 UpdateDisplay end procedure TDynVerify GravShape5MouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Inte
173. alibration feature should be provided to compensate for this remaining drift The software calibration feature calculates the changes resulting from the introduction of a bubble At the start of the training database a frame of voltages is captured with the measurement section filled only with water This will provide a starting reference All training data points are then captured normally as before On completion of the training database generation a further frame of voltages is once again captured with the measurement section filled only with water thus providing an end reference This process is repeated for the generation of a test database 71 When the data is loaded into the neural network reconstruction software the following correction is applied to each voltage using the start and stop points as references Vik Vk v Vik 6 1 where k is the index of a particular voltage in the 128 voltages corresponding to a complete frame j is the index of this particular frame in the database is the total number of frames in the database and V4 and Vyk represent the water only tests conducted at the beginning and the end of the database generation respectively In this way the neural network reconstructions are based on the differences resulting from the introduction of a bubble instead of the absolute voltage readings Subsequently any minor drift during database generation is corrected for Before an on line real time test is
174. alue then stopmax testcase end end BotGravPeak Caption floattostrf maxvalue ffFixed 4 3 if stopmax lt gt 0 then gravelbotmax floor startmax 0 5 stopmax startmax else gravelbotmax startmax maxvalue 0 startmax 0 stopmax 0 for testcase 1 to RealMain DesiredFrames Value do begin if desiredvolume testcase 4 gt maxvalue then begin maxvalue desiredvolumeftestcase 4 startmax testcase stopmax 0 end else if desiredvolume testcase 4 maxvalue then stopmax testcase if desiredvolume testcase 4 gt 0 and gravelbotstart 0 then gravelbotstart testcase else if desiredvolume testcase 4 gt 0 and gravelbotstart lt gt 0 then gravelbotstop testcase end if maxvalue 0 then gravelbotdesiremax 1 else if stopmax lt gt 0 then gravelbotdesiremax floor startmax 0 5 stopmax startmax else gravelbotdesiremax startmax botairerror 0 airbotoffset airbotmax airbotdesiremax gravelbotoffset gravelbotmax gravelbotdesiremax if airbotdesiremax lt gt 1 then begin BotAirDelay Caption floattostrf RealMain frameinterval airbotoffset ffFixed 4 3 for testcase airbotstart to airbotstop do begin if totaloutputs 2 then botairerror botairerror abs desiredvolume testcase 3 outputstore testcase airbotoffset prediction 3 else botairerror botairerror abs desiredvolume testcase 3 volumestore testcase airbotoffset 3 end botairerror b
175. am then egin tempheight gravtopposition gravieng 0 5 gravdiam lowersysbot owergravvol pi tempheight power 0 5 gravdiam 2 1 3 pi power tempheight 3 pi lowersystop lowersysbot tempheight power 0 5 gravdiam 2 nd se if gravtopposition gravleng 0 5 gravdiam gt lowersysbot 0 5 gravdiam and gravtopposition gravleng 0 5 gravdiam lt lowersystop then begin tempheight lowersystop gravtopposition gravieng 0 5 gravdiam owergravvol gravhemivol pi tempheight power 0 5 gravdiam 2 end se if gravtopposition gravieng 0 5 gravdiam gt lowersystop and gravtopposition gravleng lt lowersystop then begin tempheight gravtopposition gravieng 0 5 gravdiam lowersystop owergravvol gravhemivol pi tempheight power 0 5 gravdiam 2 1 3 pi power tempheight 3 oo D D D end else owergravvol 0 end else begin if gravtopposition lowersysbot and gravtopposition gravleng lowersysbot then lowergravvol pi power 0 5 gravdiam 2 gravtopposition lowersysbot else if gravtopposition gravleng gt lowersysbot and gravtopposition lt lowersystop then lowergravvol pi power 0 5 gravdiam 2 gravleng 195 else if gravtopposition lowersystop and gravtopposition gravleng lowersystop then lowergrawvol pi power 0 5 gravdiam 2 lowersystop gravtopposition gravleng else lowergravvol 0 end gravel bubble for top measuring ring if grav
176. ance plate configuration Bubble diameter m 07 Bubble diameter m 07 Upper system top m o 738 Bubble length m g B Bubble length m Jo 19 Upper system bottom m p 618 Bubble start position m Bubble start position m Jo Lower system top m 0 235 Lower system bottom m Jo 115 Bubble position Bubble position Volume fraction plots For top and bottom system Test case results Air volume fraction for top system Gravel volume fraction For top system r Source of bubble velocity jl desired cross correlation Calculate desired volume fractions Calculate actual velocity Calculate volume fraction error Air volume error delay s 500 1 000 0 200 400 600 800 1 000 1 200 0 502 4 028 0 004 Air volume fraction for bottom system Gravel volume fraction for bottom system Bottom 0 426 4 064 0 011 Gravel volume error 9 5 Top 0 000 0 000 unknown Bottom 0 000 0 000 unknown Custom filename Save network results Close 0 200 400 600 800 1 000 1 200 FIGURE 7 6 SCREEN CAPTURE OF THE DYNAMIC VERIFICATION FORM THE VOLUME FRACTION PROFILES OF THE AIR AND GRAVEL PHASES ARE PLOTTED AT THE TOP AND BOTTOM SYSTEM IN BLUE USING THE SPECIFIED BUBBLE DIMENSIONS THE DESIRED VOLUME FRACTION PROFILES OF THE AIR AND GRAVEL PHASES AT THE TOP AND BOTTOM SYSTEMS CAN BE CALCULATED AND PLOTTED IN RED THE ABOVE SCREEN CAPTURE IS OF A 0 04 VOLUME FRACTION AIR BUBBLE SITU
177. ance tomography the sensitivity matrix or Jacobian maps the changes in the resistivity distribution to the changes in the voltages measured on the boundary of the region 22 These matrices can be calculated numerically using the finite element modelling technique or they can be measured experimentally 23 It is actually the inverse of the sensitivity matrix that is required and the inverse problem that needs to be solved For capacitance tomography the inverse problem can be described as follows given the electrode potentials applied to the vessel and set of measured capacitances C calculate the permittivity distribution 24 Since there are generally fewer readings than pixels in the reconstructed images this inverse problem is ill posed and some form of regularisation is required 25 Since the sensitivity matrix is not square an approximation technique is necessary to calculate the inverse The standard linear back projection algorithm simply uses the transpose of the sensitivity matrix 11 The linearisation of the reconstruction imposes the following restrictions 25 1 there are no large changes in the conductivity or permittivity field 2 the spatial variation is low Results produced using linear back projection are typically blurry and are sub optimal for a given set of capacitance or conductance readings In addition the second restriction results in images with very low spatial resolution 25 If a better approximation
178. andard time division multiplexed impedance tomography systems with minimal loss in resolution A dual plane 16 electrode FDM impedance tomography system has been developed that operates at an on line real time frame rate of approximately 280frames s for a two phase image reconstruction or three phase volume fraction prediction Further the resolution of the tomography system was shown to be 1096 of the pipeline diameter at the centre of the pipeline which is considered standard for electrical tomography systems The only major drawback of the FDM impedance tomography technique is that the conductance readings are obtained using a two electrode collection protocol Consequently the conductance readings are sensitive to variations in the contact impedance between the conductance probes and the material being imaged and therefore drift considerably over time These contact impedance variations have been shown to be primarily the result of electrochemical reactions taking place between the metal conductance probes and the seawater In the current system this problem was circumvented by using potable water which has a far lower conductance than seawater However a feasibility study is currently being performed on implementing a frequency division multiplexed four electrode adjacent pair measurement protocol for the conductance measurements This would eliminate the problems associated with two electrode conductance measurements Neural networks trained w
179. ard air bubble length of 190mm For each test the peak of the measured volume fraction profile was determined and the results averaged over a number of tests 94 rt I D vee ma wa te I i i t i i i 7 Hi oem a vee i I wa I I f ew fA Ar Tert i eS om a 4 Li i rec e i om t om te en eee me m aom ie ooa im i a I H t c B D I I i f gt Fu 168 i i A mensim uM UND A A we n i i 3 Se u 00 See t M m u Mm FIGURE 7 9 SCREEN CAPTURES OF A 0 04 VOLUME FRACTION AIR BUBBLE NEAR THE LEFT HAND EDGE OF THE RIG AND A 0 01 VOLUME FRACTION AIR BUBBLE NEAR THE RIGHT HAND EDGE OF THE RIG MOVING THROUGH THE MEASUREMENT PLANES AT THE SAME TIME THESE IMAGE RECONSTRUCTIONS WERE PERFORMED ON LINE AND THEIR RESULTS DISPLAYED IN REAL TIME IN EACH SCREEN CAPTURE THE TOP FRAME REPRESENTS THE TOP MEASUREMENT PLANE AND THE BOTTOM FRAME REPRESENTS THE BOTTOM MEASUREMENT PLANE THE ORDER OF THE FRAMES IS FROM LEFT TO RIGHT AND THEN TOP TO BOTTOM DUE TO THE LARGE NUMBER OF FRAMES CAPTURED
180. artmax testcase stopmax 0 end else if volumestore testcase 1 maxvalue then stopmax testcase end end TopAirPeak Caption floattostrf maxvalue ffFixed 4 3 if stopmax lt gt 0 then airtopmax floor startmax 0 5 stopmax startmax else airtopmax startmax then calculate the desired width of the volume fraction peak and the position of the desired volume fraction peak maxvalue 0 startmax 0 stopmax 0 for testcase 1 to RealMain DesiredFrames Value do begin if desiredvolume testcase 1 gt maxvalue then begin maxvalue desiredvolume testcase 1 startmax testcase Stopmax 0 end else if desiredvolume testcase 1 maxvalue then stopmax testcase 196 if desiredvolume testcase 1 0 and airtopstart 0 then airtopstart testcase else if desiredvolume testcase 1 gt 0 and airtopstart lt gt 0 then airtopstop testcase end if maxvalue 0 then airtopdesiremax 1 else if stopmax lt gt 0 then airtopdesiremax floor startmax 0 5 stopmax startmax else airtopdesiremax startmax gravel volume for top measuring ring startmax 0 stopmax 0 maxvalue 0 for testcase 1 to RealMain DesiredFrames Value do begin if totaloutputs 2 then begin if outputstore testcase prediction 2 gt maxvalue then begin maxvalue outputstore testcase prediction 2 startmax testcase stopmax 0 end else if outputstore testcase prediction 2
181. ase RecTestStart Value WriteLn tempfile Test stop inttostr RecLoadDBase RecTestStop Value WriteLn tempfile Max epochs inttostr RecSetParam GetMaxEPochs WriteLn tempfile Number of hidden nodes tinttostr RecSetParam GetNumHidden if RecSetParam RecPerformEarlyStop Checked then S true else S false WriteLn tempfile Perform early stopping s if RecSetParam RecPerformModel Checked then S true else S false WriteLn tempfile Conduct model search s if RecSetParam RecUseGradDescent Checked then S Gradient descent else s RPROP WriteLn tempfile Training algorithm 5 WriteLn tempfile Learning rate floattostr RecSetParam GetLearnRate WriteLn tempfile Momentum 4floattostr RecSetParam GetMomentum WriteLn tempfile Initial delta floattostr RecSetParam GetDeltalnit WriteLn tempfile Maximum delta floattostr RecSetParam GetDeltaMax WriteLn tempfile Update frequency floattostr RecSetParam GetUpdateFreq WriteLn tempfile Number of fails floattostr RecSetParam GetNumFails if RecSetParam RecStopVolume Checked then S Sum volume else S Threshold WriteLn tempfile Stopping condition s WriteLn tempfile Minimum hidden for model inttostr RecSetParam RecModelStart Value WriteLn tempfile Maximum hidden for model inttostr RecSetParam RecModelStop Value WriteLn tempfile Increment for model inttostr RecSetParam RecModellnc
182. asonable transmission of all eight transmitter frequencies FIRST ORDER LOW PASS FILTER FIRST ORDER BANDPASS FILTER TO SYNCHRONOUS DETECTOR BOARD FROM CAPACITANCE Fion 1kHz Freon 340kHz PLATE 340kHz N TRANSIMPEDANCE AMPLIFIER FIGURE 5 10 BLOCK DIAGRAM OF CAPACITANCE RECEIVER As before the conductance receiver consists of a potential divider formed by an input resistor to ground and a voltage buffer Figure 5 11 is a block diagram of the conductance receiver A first order bandpass filter was included to attenuate the high frequency components of the received signal and to block any DC electrochemical voltages Electrochemical reactions between the electrodes and the conductive solution result in the generation of DC offset voltages These DC voltages have a magnitude of up to hundreds of millivolts A bandpass filter has been shown to reduce these offset voltages 12 The circuit details of the conductance and capacitance receiver are provided in Appendix C on page 123 FIRST ORDER BANDPASS FILTER FROM CONDUCTANCE 1kHz F TO SYNCHRONOUS PROBE DETECTOR BOARD VOLTAGE BUFFER WITH GAIN POTENTIAL DIVIDER FIGURE 5 11 BLOCK DIAGRAM OF CONDUCTANCE RECEIVER 43 5 2 3 Synchronous detection of receiver signals Each synchronous detection board performs the synchronous detection of two complete channels where a channel incorporates both the conductance and capacitance
183. asurement of the Gas Phase Velocity in Gas Liquid Flows Using a Dual Plane ERT System Proc 2 World Cong on Industrial Process Tomography Hannover pp 669 676 Georgakopoulos D and Yang W Q 2001 Key Issues in the Design of a Capacitance Transducer for Tomography Proc 2 World Cong on Industrial Process Tomography Hannover pp 586 594 Ramphal V 2002 Implementation of MOSFETs in an impedance tomography system BTech thesis Cape Technikon Yang W Q 1996 Calibration of capacitance tomography systems a new method for setting system measurement range Meas Sci Technol 7 863 867 Xie C G et al 1992 Electrical capacitance tomography for flow imaging system model for development of image reconstruction algorithms and design of primary sensors Proc 139 89 98 Xie C G Stott A L Plaskowski A and Beck M S 1990 Design of capacitance electrodes for concentration measurement of two phase flow Meas Sci Technol 1 65 78 Loh W W Dickin F J Waterfall R C and Pinheiro P A T 1998 Alternative inter plane data collection protocol for electrical resistance tomography in flow monitoring applications Electron Lett 34 1487 1488 Beck M S and Plaskowski A 1987 Cross Correlation Flowmeters their Design and Application Bristol Adam Hilger Yan H Shao and Wang S 1999 Simulation Study of Capacitance Tomography Sensors Proc 1 World Cong on Industrial Process Tomography Buxton pp 388 394 Xu H Yang G and Wang S 1999 Effect of Axial Guard Electrode
184. asurement system must be tuned as a compromise between all the electrode combinations Having verified the operation of the FDM impedance tomography data capture technique it was necessary to construct a laboratory scale rig for the assessment of the impedance tomography system on dynamic flow simulations The diameter of the pipeline is 350mm The laboratory scale rig stands vertically and consists of a top section two impedance tomography measurement sections and a bottom section A flow simulation apparatus was designed for the rig and is used to simulate the movement of air or gravel masses in the water at specified velocities under computer control Each measurement section consists of a central plane of 16 capacitance measurement plates mounted on the outside of the pipeline Driven axial guard electrodes are provided above and below this central plane of measurement electrodes At the centre of each capacitance plate is a conductance probe mounted through the pipeline wall and in electrical contact with the water Grounded radial shields pass between the measurement electrodes and an outer grounded shield prevents external interference Various modifications were made to the data acquisition system of the laboratory scale rig based on the results of the 8 electrode system in an attempt to improve the volume fraction prediction accuracy To assess the results of these modifications the laboratory scale measurement sections were tested on static conf
185. ate the capacitance and conductance components It was therefore decided that the square waves would be generated using a different software approach D ii L T L Loop movw Boooooooo MOVWF PORTB DELLI bF LI NOP Co Li MOVLW B 00000000 MOVWF PORTB wap NOP JE NOP Bo ii MOVLW 10000000 MOVWF PORTB Boo bd ee eee NOP zE NOP CR oT a E MOVLW B 01010101 Ay or os MOVWF PORTB vvv GOTO LOOP 123 FIGURE 4 4 OUTPUT FREQUENCY GENERATION FOR THE 8 ELECTRODE FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM USING A PIC16F84 MICROCONTROLLER INCLUDED IS THE MICROCONTROLLER PROGRAM CODE USED TO GENERATE THE FIRST THREE PORT SET INSTRUCTIONS HIGHLIGHTED IN THE TIMING DIAGRAM SIMPLY BY COMBINING THE CORRECT SEQUENCE OF PORT SET INSTRUCTIONS THE FOUR OUTPUT SQUARE WAVES AND THEIR QUADRATURE WAVEFORMS ARE GENERATED A GOTO INSTRUCTION AT THE END OF THE PROGRAM ENSURES THAT THE SEQUENCE OF PORT SET INSTRUCTIONS REPEAT The output square waves are generated by hard coding the microcontroller port set instructions Figure 4 4 illustrates this method Included in figure 4 4 is the microcontroller program code for the first three port set instructions as well as the last port set instruction in the sequence The full program code can be found in Appendix E on page 130 Simply by combining the correct sequence of port set instructions the four output square waves and their
186. ate the individual volume fractions of the seawater gravel and air phases as they pass through it In addition it will be required to monitor the velocities of the individual components so that the mass flow rate of each component can be calculated This information is invaluable for the on line monitoring and control of such a process Due to the potential advantages that such an instrument would possess this project has progressed through a number of stages over roughly a four year period Firstly research was done by Quinton Smit to determine whether capacitance tomography is a feasible technique for the monitoring of an air gravel seawater mixture 1 2 Secondly research was done by the author on neural networks as a means of improving the accuracy of the reconstructed images 3 4 Work to date has considered only static situations This thesis will address the issue of monitoring a flow and in particular the measurement of the individual component flow rates Turbulent multi phase flows are characteristic of many industrial processes Although experimental observations and measurements of such processes are extremely complex tomography has developed over the last decade as a reliable method of investigating such systems 5 Process tomography techniques provide a novel means to visualise the internal dynamics of industrial processes such as multi component flows in pneumatic conveyors and the process of mixing or separation in plant vessels
187. bbles based on the time procedure CalculateDistance time real var airtopposition real var gravtopposition real procedure ResetColours procedure UpdateDisplay calculate the desired volume fraction for the two phases based on the positions procedure Calculate Volume airtopposition real gravtopposition real var upperairvol real var uppergravvol real var lowerairvol real var lowergrawvol real public Public declarations airposition integer position of air bubble gravelposition integer position of gravel bubble end airdiam real II position of electrodes for bottom system var DynVerify TDynVerify implementation uses realunit stepconfig R DFM DESIRED VOLUME FRACTION ESTIMATION PROCEDURES the following procedure calculates the position of the two bubbles at a certain time using the velocity and accelerations stipulated in the stepper configuration form The position of the bubble is defined to be the top of the bubble Near the base of the pipeline there is a horizontal band defined to be position zero At the start of a test the top of the bubble must conincide with this band As the bubble moves up the pipe so the position stipulated in metres from the start band increases The positions of the measurement electrodes are also provided with respect to this start band procedure TDynVerify CalculateDistance time real var airtopposition gravtopposition
188. be obtained by training a neural network to predict the volume fractions of the different phases directly without having to first perform an image reconstruction 3 4 A double layer feed forward neural network was trained using gradient descent to predict the volume fractions of a three phase air gravel seawater mixture directly The operation of this neural network was examined in figure 3 3 on page 14 The number of hidden layer neurons was chosen to be 25 based on the results achieved for the TDM impedance tomography system Early stopping was employed to prevent over fitting and table 4 6 provides a comparison between the performances of the FDM impedance tomography system and the TDM impedance tomography system As is evident from table 4 6 the performance of the volume fraction predictor is also better for the FDM impedance tomography system TABLE 4 6 COMPARISON BETWEEN THE PERFORMANCE OF THE FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM AND A STANDARD TIME DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM FOR A THREE PHASE AIR GRAVEL SEAWATER VOLUME FRACTION PREDICTION USING A DOUBLE LAYER FEED FORWARD NEURAL NETWORK ALL RESULTS REPRESENT THE ABSOLUTE ERROR AND ARE ROUNDED OFF TO TWO SIGNIFICANT DIGITS PERFORMANCE MEASURE TIME DIVISION FREGUENCY DIVISION MULTIPLEXED IMPEDANCE MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM TOMOGRAPHY SYSTEM AIR VOLUME FRACTION ERROR 90 GRAVEL VOLUME FRACTION ERROR WATER VOLUME FRACTION
189. bel TLabel RecNetworkStart TSpeedButton Label2 TLabel RecNetworkEpoch TLabel RecNetworkProgressBar TProgressBar Label3 TLabel RecTrainResults TPanel Label4 TLabel Label5 TLabel RecTrainThreshErr TLabel RecTrainVoidErr TLabel RecTestThreshErr TLabel RecTestVoidErr TLabel RecThreshLabel TLabel RecSumLabel TLabel RecModelResults TPanel Label6 TLabel Label7 TLabel Label8 TLabel Label9 TLabel RecModelBestPerf TLabel 167 RecModelNumBest TLabel RecModelCurrent TLabel ModelSearchTimer TTimer RecNetworkReturn TSpeedButton PauseTrain TSpeedButton procedure RecNetworkReturnClick Sender TObject procedure FormCreate Sender TObject procedure RecNetworkStartClick Sender TObject procedure FormShow Sender TObject procedure ModelSearchTimerTimer Sender TObject procedure PauseTrainClick Sender TObject private Private declarations changeplus real multiplication factor for delta values in the RPROP algorithm increase changeminus real multiplication factor for delta values in the RPROP algorithm decrease numoutputs integer number of network outputs numhidden integer number of hidden layer neurons learnrate real learning rate for gradient descent momentum real momentum constant for gradient descent totalwaterpixels real number of test water pixels totalgravelpixels real number of test gravel pixels t
190. bel15 TLabel AddPhaseSelect TComboBox procedure StepCloseClick Sender TObject procedure FormShow Sender TObject procedure StepDownloadClick Sender TObject procedure SteplnitialiseClick Sender TObject procedure StepResetClick Sender TObject procedure AirBubbleJogMouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Integer procedure AirBubbleJogMouseUp Sender TObject Button TMouseButton Shift TShiftState X Y Integer procedure GravelBubbleJogMouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Integer procedure GravelBubbleJogMouseUp Sender TObject Button TMouseButton Shift TShiftState X Y Integer private Private declarations public Public declarations airsteps integer airrps real airrevs real gravelsteps integer gravelrps real gravelrevs real aircircumference real gravelcircumference real airacceleration integer gravacceleration integer end number of steps of air drive revolutions per second of air drive revolutions of air drive number of steps of gravel drive revolutions per second of gravel drive revolutions of gravel drive air pulley circumference gravel pulley circumference var StepperConfig TStepperContig implementation uses realunit dynunit R DFM tM GENERAL PROCEDURES when close form calculate what the require
191. bject var tempfile textfile filename string tempair tempgrav Boolean index temphidden tempoutputs colindex rowindex tempphase integer begin upperindex 128 numtophidden 0 numbottomhidden 0 tempair false 185 tempgrav false filename cATestData Weight inttostr TopTestNum Value txt AssignFile tempfile filename Reset tempfile ReadLn tempfile tempoutputs totaloutputs tempoutputs load the number of outputs from the files and set the neural network output display appropriately if totaloutputs 2 then begin TopVolume Visible true BottomVolume Visible true end else begin TopVolume Visible false BottomVolume Visible false end load the weights for the top system ReadLn tempfile temphidden if temphidden gt 0 then begin upperindex temphidden numtophidden temphidden SetLength tophiddenweights 129 temphidden SetLength tophiddenoutputs temphidden 1 load in the weights for the hidden layer neurons for colindex 0 to temphidden 1 do for rowindex 0 to 128 do ReadLn tempfile tophiddenweights rowindex colindex end SetLength topweights upperindex 1 totaloutputs load in the weights for the output layer neurons for colindex 0 to tempoutputs 1 do for rowindex 0 to upperindex do ReadLn tempfile topweights rowindex colindex CloseFile tempfile load the weights for the bottom system upperindex 128
192. c Measurements of Solids Mass Flow in Dense Pneumatic Conveying What We Need to Know About the Flow Physics Proc 2 World Congr on Industrial Process Tomography Hannover pp 353 361 Loh W W Waterfall R C Cory J and Lucas G P 1999 Using ERT for Multi Phase Flow Monitoring Proc 17 World Congr on Industrial Process Tomography Buxton pp 47 53 Byars M 2001 Developments in Electrical Capacitance Tomography Proc 2 World Congr on Industrial Process Tomography Hannover pp 542 549 Deng X Dong F Xu L J Liu X P and Xu L A 2001 The design of a dual plane ERT system for cross correlation measurement of bubbly gas liquid pipe flow Meas Sci Technol 12 1024 1031 Thorn R Johansen G A and Hammer E A 1999 Three Phase Flow Measurement in the Offshore Oil Industry Is There a Place for Process Tomography Proc 17 World Congr on Industrial Process Tomography Buxton pp 228 235 Arko A Waterfall R C and Beck M S 1999 Development of Electrical Capacitance Tomography for Solids Mass Flow Measurement and Control of Pneumatic Conveying Systems Proc 1 World Congr on Industrial Process Tomography Buxton pp 140 146 Dickin F J et al 1992 Tomographic imaging of industrial process equipment techniques and applications Proc IEE 139 72 81 Yang W Q 2001 Advance in AC based capacitance tomography system Proc 2 World Congr on Industrial Process Tomography Hannover pp 557 564 Dickin F J Williams R A and Beck M S 1993 Determination of compositio
193. cViewCapTable Cells 0 3 RecViewCapTable Cells 0 4 TxD RecViewCapTable Cells 0 5 RecViewCapTable Cells 0 6 TxF RecViewCapTable Cells 0 7 TxG RecViewCapTable Cells 0 8 TxH RecViewCapTable Cells 1 0 RxA RecViewCapTable Cells 2 0 RxB RecViewCapTable Cells 3 0 RxC RecViewCapTable Cells 4 0 RxD RecViewCapTable Cells 5 0 RxE RecViewCapTable Cells 6 0 RxF RecViewCapTable Cells 7 0 RxG RecViewCapTable Cells 8 0 RxH RecViewVolume SeriesList 0 Add 100 RecViewVolume SeriesList 1 Add 0 RecViewVolume SeriesList 2 Add 0 end when show form display either an image of the vessel cross section or a chart of the volume fractions depending on what the desired network output is procedure TRecViewMain FormShow Sender TObject begin currentindex 0 RecViewPrevious Enabled false RecViewNext Enabled true if RecLoadDBase RecDesiredRecon ItemIndex 0 then begin RecViewlmgLabel Visible true RecViewInputTomo Visible true end else begin RecViewlmgLabel Visible false RecViewInputTomo Visible false end DisplayData end scroll through all the test cases procedure TRecViewMain ScrollTimerTimer Sender TObject begin if RecViewNext Down and UseContinouos Checked then RecViewNextClick sender if RecViewPrevious Down and UseContinouos Checked then RecViewPreviousClick sende
194. cal Hence the feedback capacitor and resistor values were chosen as follows with only the lowest frequency transmitter operating the resistor and capacitor value were chosen so that the output from the square wave transition decayed to zero at roughly the same time as the next square wave transition This ensured that a satisfactory signal level was received for all transmitters Then with all transmitters operating simultaneously this RC combination remained fixed while the feedback resistor value was reduced to ensure that the output of the amplifier did not saturate The circuit details of the conductance and capacitance receiver are given in Appendix B on page 117 4 2 44 Synchronous detection of receiver signals Phase sensitive detection can be achieved using either multiplication or switching techniques Results have shown that the multiplying phase sensitive detector produces more accurate results since the switching technique suffers from the following problems 43 1 the charge injection from the CMOS switches 2 the introduction of odd harmonics In addition it has been shown that the multiplication demodulator is 1 2 times faster than the switch based phase sensitive demodulator for the same level of accuracy 43 However since the FDM impedance tomography system uses large numbers of multipliers in parallel cost constraints limited the electronics to CMOS switch based techniques In particular a DG303 dual single pole double thro
195. cal section is greater than or equal to the length of the measuring section electrodes The equations below are simply standard volume calculations obtained from tables and will not be discussed in any major detail procedure TDynVerify CalculateVolume airtopposition gravtopposition real var upperairvol uppergravvol lowerairvol lowergravvol real var tempheight airhemivol gravhemivol real begin airhemivol 2 3 pi power 0 5 airdiam 3 gravhemivol 2 3 pi power 0 5 gravdiam 3 air bubble volume for bottom measuring ring as round at top of bubble enters measuring section if airleng gt lowersystop lowersysbot then begin if airtopposition gt lowersysbot and airtopposition lt lowersysbot 0 5 airdiam then begin tempheight lowersysbot airtopposition 0 5 airdiam oweraitvol airhemivol pi tempheight power 0 5 airdiam 2 1 3 pi power tempheight 3 nd as cylinder section of bubble enters measuring ring se if airtopposition gt lowersysbot 0 5 airdiam nd airtopposition lt lowersystop then tempheight airtopposition 0 5 airdiam lowersysbot owerairvol airhemivol pi tempheight power 0 5 airdiam 2 nd as round at top of bubble exits measuring section se if airtopposition gt lowersystop nd airtopposition lt lowersystop 0 5 airdiam then 2 5 tempheight lowersystop airtopposition 0 5 airdiam owerairvol pi tempheight pow
196. ce tomography system is to be installed The major rig dimensions are provided in figure 5 1 The rig consists of four distinct sections namely the two measurement planes and a top and a bottom as can be seen in the accompanying figure The top and bottom sections are provided for the flow simulation tests Specifically they provide a region before and after the measurement planes for the bubbles to accelerate and decelerate Ideally the bubble should be travelling at constant velocity while moving through the measurement sections since cross correlation measures the average velocity between the two measurement planes Included in the bottom section are two alignment strips used for the exact positioning of test cases The application of these alignment strips will be illustrated at a later stage An advantage of constructing different sections is the flexibility of the design As an example the top section can be placed between the two measurement sections to increase the measurement plane separation Also each measurement plane can be tuned and tested separately simply by constructing a blanking off disk The measurement section is then simply a larger version of the prototype rig thus permitting a consistent comparison between the performance of the 8 electrode prototype and the 16 electrode laboratory scale rig The sections are joined using stainless steel flanges and are sealed with 3mm rubber insertion gaskets The entire rig is suppo
197. classification scheme is 1 of C encoding A 1 of C output encoding is achieved by introducing additional dummy output variables for each pixel Subsequently the number of output layer neurons is increased to 264 since three output neurons are required for every pixel Each dummy output is given a target of zero except for that particular output corresponding to the desired classification which has a target value of unity In addition each output neuron employs a softmax activation function 29 This ensures that the sum of the three dummy output neurons always equals unity and so these outputs can be interpreted as valid posterior probabilities 29 The pixel classification is the dummy output neuron with the largest output and consequently the largest posterior probability 12 The outputs are calculated using the following softmax activation function formula 29 e 3 1 E 8 7 where a is the weighted input for each dummy output neuron Although this neural network is able to perform three phase air gravel seawater image reconstructions the increase in the number of output neurons substantially increases the computation time of the reconstruction algorithm This will be examined more closely when the issue of real time on line reconstructions is addressed NS 1 OF C ENCODING se 2 lt Ji OSV O READINGS eO Wa PROPORTIONAL gi AND 4 CONDUCTANCE Y INPUT LAYER NEURONS
198. components on the outputs of the low pass filters were monitored using a spectrum analyser As an example figure 5 16 on page 56 contains the frequency spectrum captured on the output of the low pass filter for the measurement of the conductance between transmitter B and receiver G Included in figure 5 16 is a table of the frequency components predicted for this particular transmitter receiver electrode pair using the simulation software The highlighted entries indicate those frequency components observed with the spectrum analyser The predicted frequency components are displayed on the frequency spectrum as square markers 55 30 gt 8 40 o 50 60 70 80 90 0 20 40 60 80 100 140 180 frequency Hz 10 4625Hz 20 925Hz 31 3875 2 41 85 2 52 3125 2 62 775 2 73 2375 2 83 7 2 94 1625 2 104 625 2 115 0875 2 125 55Hz 136 0125Hz 146 475Hz 156 9375Hz 167 4Hz 177 8625Hz 188 325Hz 198 825Hz 209 25Hz FIGURE 5 16 COMPARISON BETWEEN THE FREQUENCY SPECTRUM MEASURED USING A SPECTRUM ANALYSER AND THE SPECTRAL PEAKS PREDICTED BY THE SIMULATION SOFTWARE WHICH ARE PLOTTED AS SQUARE MARKERS FOR THE TRANSMITTER B RECEIVER G ELECTRODE PAIR THE TABLE LISTS THE FREQUENCY COMPONENTS PREDICTED BY THE SIMULATION SOFTWARE FOR TRANSMITTER B AND THE HIGHLIGHTED ENTRIES INDICATE THOSE SPECTRAL PEAKS THAT COINCIDE WITH THE MEASURED FREQUENCY SPECTRUM THE SPECTRAL PEAKS NOT PREDICTED ARE THE RE
199. conductance of that electrode pair is measured The system consists of a single set of measurement electronics and the electrodes for a specific electrode pair are temporarily connected to the measurement electronics through multiplexed switches These measurements are performed serially and therefore the time taken to capture a complete frame of data is the number of electrode combinations multiplied by the time taken to perform an individual measurement Further the process of switching can introduce impulses in the received signal which increase the measurement time 37 Subsequently the time taken to collect a complete frame of data can become significant thus limiting the on line frame rate of the system In contrast the impedance tomography data acquisition technique developed for this application uses frequency division multiplexing FDM where the measurements are separated in the frequency domain The operation of this technique will be explained for the simple configuration of a 4 electrode system although it is completely scaleable and has been successfully implemented on a 16 electrode system FDM impedance tomography systems specify electrodes as either dedicated transmitter or receiver electrodes For the 4 electrode system the order of these electrodes around the pipeline perimeter would be transmitter A receiver A transmitter B receiver B Each transmitter outputs a sine wave at a different frequency and all transmitters operate at t
200. d function TRecSetParam GetMomentum real begin GetMomentum momentum end function TRecSetParam GetNumFails integer begin GetNumFails numfails end function TRecSetParam GetNumHidden integer begin GetNumHidden numhidden end function TRecSetParam GetUpdateF req integer begin GetUpdateFreq RecSetUpdateFreq Value end tM GENERAL PROCEDURES if user selects to perform early stopping then enable user to set the early stopping parameters procedure TRecSetParam RecPerformEarlyStopClick Sender TObject begin if RecPerformEarlyStop Checked then RecSetEarlyParam Enabled true else RecSetEarlyParam Enabled false end if user selects to conduct a model search to determine the optimal number of hidden layer neurons then enable the user to set the model search parameters procedure TRecSetParam RecPerformModelClick Sender TObject begin if RecPerformModel Checked then RecSetModelParam Enabled true else RecSetModelParam Enabled false end if user selects to use gradient descent then enable the user to set the training parameters specific to gradient descent procedure TRecSetParam RecUseGradDescentClick Sender TObject begin if RecUseGradDescent Checked then RecRPropParam Visible false else RecRPropParam Visible true end if user selects to use Resilient back propagation then enable the user to set the tr
201. d frames The on line data capture and reconstruction rate of a two phase air water image reconstruction was 280frames s For the more advanced three phase air gravel water reconstruction the on line data capture and reconstruction rate reduced to 140frames s for image reconstruction However since it is only the individual volume fractions that are required the frame rate can be increased to 180frames s using the volume fraction predictor neural network Since the positions of the simulated bubbles are precisely controlled it was possible to determine what the volume fraction of a specific phase should be at a certain time Subsequently a dynamic volume fraction error was calculated based on the mean absolute error between the measured volume fraction profile and the desired volume fraction profile for a specific test The average dynamic volume fraction error for the air phase was 1 0 at the top System and 0 89 at the bottom system using a double layer volume fraction predictor neural network The average dynamic volume fraction error for the gravel phase was 1 596 at the top system and 1 696 at the bottom System Since the water was static no attempt was made to assess the dynamic volume fraction error of the water phase These results are comparable with those of the static tests thus proving that neural networks trained on static configurations can be used to perform reconstructions of dynamic flow simulations The volume fraction profiles
202. d steps and steps second for each of the drives based on the specified test parameters procedure TStepperConfig StepCloseClick Sender TObject var Reg TRegistry tempaccel1 tempaccel2 tempaccel3 integer begin write the current settings to the registry Reg TRegistry Create try Reg OpenKey Software Neural True Reg WriteString AirPulley AirDiameter Text Reg WriteString GravelPulley GravelDiameter Text Reg WriteString AirVelocity AirVelocity Text 187 Reg WriteString GravelVelocity GravelVelocity Text Reg WriteString AirDistance AirDistance Text Reg WriteString GravelDistance GravelDistance Text Reg WriteInteger AirBubblePosition AirBubblePosition ItemIndex Reg Writelnteger GravelBubblePosition GravelBubblePosition ItemIndex Reg Writelnteger FlowDirection FlowDirSelect ItemIndex Reg Writelnteger AirDrive AirDrive Value Reg Writelnteger GravelDrive GravelDrive Value Reg WriteString AirAccel AirAccel Text Reg WriteString GravelAccel GravelAccel Text finally Reg Free end calculate the number of revolutions for each drive based on the desired distance and the circumference of the pulley aircircumference pi strtofloat AirDiameter Text gravelcircumference pi strtofloat GravelDiameter Text airrps strtofloat AirVelocity Text aircircumference gravelrps strtofloat GravelVelocity Text gravelcircumference assuming 400 st
203. d training output data from training database TestInpTable TTable table to load testing input data from testing database TestOutTable TTable I table to load testing output data from testing database trainstartvalues array 1 5 1 128 of real trainstopvalues array 1 5 1 128 of real teststartvalues array 1 5 1 128 of real teststopvalues array 1 5 1 128 of real records of voltages for rig filled only with water this will be used to calibrate the system procedure PerformCalibrate I offset the data according to the calibration data so that only the changes resulting from inserting a bubble are stored procedure ReOrganiseData in the current implementation all the data is loaded however if only want to train for air water reconstruction then the other data is removed using this procedure procedure StandardiseData standardise loaded data function StdDevColumn colnum integer mean double double find the standard deviatioon for a column function MeanColumn colnum integer double find the mean of a column public Public declarations function GetTomoSystem integer get which system using function GetTotalTrain integer get total training points function GetTotalTest integer get total test points function GetNumOutputs integer get number of outputs procedure GetPhases var contair Boolean var contgrav Boolean procedure SetPhas
204. dInteger GravelPosition DynVerify UpperSystemTop Text Reg ReadString UpperSystTop DynVerify UpperSystemBottom Text Reg ReadString UpperSystBot DynVerify LowerSystemTop Text Reg ReadString LowerSystTop DynVerify LowerSystemBottom Text Reg ReadString LowerSystBot DynVerify VelSource ItemIndex Reg Readlnteger VelSource finally Reg Free end VerifyResults Visible false StepperStop Visible false viewposition Viewlnc Value notifvalue SampHardware InitSample 1 Handle curcalibrate false end to perform a calibration initiate a capture of 50 frames of data for the rig filled only with water procedure TRealMain RealCalibrateClick Sender TObject begin curcalibrate true SampHardware StartSample 50 end override the default Windows message handling procedure to capture the sampling complete message when sampling is complete call StopSampling and process the results procedure TRealMain WndProc var TheMsg TMessage begin if TheMsg Msg notifvalue then begin SampHardware StopSample result ProcessResults end else inherited WndProc TheMsg end end PROGRAM LISTING OF STEPCONFIG PAS This unit is used to set up the stepper motors for a dynamic test Jog features allow the user to carefully position the bubble at one of the defined starting positions before a test To calculate how many steps are required to move the bubble a certain
205. dProgressBar Max traintotal testtotal RecLoadProgressBar Position 0 row 0 II load the training data II for the two phase predictions load the input data but set the output to zero for incorrect phase inputs to serve as a flag so that this data can be overwritten at a later stage for trainversions 1 to RecTrainVers Value do begin for trainnumber RecTrainStart Value to RecTrainStop Value do begin tempair false tempgravel false index trainnumber trainversions 1000 TrainOutTable FindKey index TrainInpTable FindKey index TomoSystem TrainInpTable FieldByName System Value InputData row 0 1 II load the input data for col 1 to 128 do begin 51 Voltage 4inttostr col InputData row col TrainInpTable FieldByName s1 Value end if RecDesiredRecon lItemIndex 0 then begin if desired output is image reconstruction calculate what phases are in this training case for col 1 to 100 do begin 51 Pixel inttostr col tempvalue TrainOutTable FieldByName s1 Value if tempvalue 0 then tempair true 158 else if tempvalue 2 then tempgravel true end imgcol 0 for col 1 to 100 do begin if the phases for the current test point are the correct phases for the desired reconstruction then load the data normally if RecDesirePhase ItemIndex 1 and tempair and not tempgravel Jor RecDesirePhase ItemIndex 2 and not tempair and temp
206. deltas for the output nodes connected to this hidden node by the weight of the connection for outputnodeindex 0 to numoutputs 1 do deltasum deltasum OutputWeight inputnodeindex 1 outputnodeindex delta outputnodeindex deltahidden hiddenderiv inputnodeindex 1 deltasum calculate adjustments to hidden weights for outputnodeindex 0 to 128 do DiffHiddenWeight outputnodeindex inputnodeindex DiffHiddenWeight outputnodeindex inputnodeindex deltahidden InputData currsample outputnodeindex where the derivative of the error function with II respect to the weights is given by the calculated delta for the hidden node multiplied by the II input to that node end end once gone through entire training set make the adjustments to the weights where the type of adjustment depends on the training algorithm being used if UseRProp then AdjustRPROP else AdjustWeights Il every updatefreq epochs update the network display to indicate the current training results if currepoch mod updatefreq 0 then UpdateNetworkDisplay inc currepoch end RecNetworkProgressBar Position RecNetworkProgressBar Max end update the network display to show current network training results This procedure also tests the validation performance of the network and whether the validation performance is deteriorating as a result of overfitting the training data procedure TRecTrainNetwork UpdateNetworkDisplay var tem
207. der the case of the 4 electrode system with transmitter A operating at 3kHz and transmitter B operating at 5kHz In this example the 5 harmonic of transmitter A equals the 3 harmonic of transmitter B and duplicate frequencies would be received To calculate for example the conductance between transmitter A and receiver A the received signal is multiplied by reference A at 3kHz However transmitter B contains a harmonic whose frequency equals a harmonic of transmitter A and will therefore pass through the low pass filter stage of the synchronous detector Consequently the output of the low pass filter will not only represent the conductance between transmitter A and receiver A but will also incorporate the conductance between transmitter B and receiver A which is clearly an undesirable situation In terms of output ripple it also undesirable to have duplicate frequency components Due to component tolerances the duplicate frequency components will not be exactly equal in a practical implementation and instead a low frequency difference will exist This difference frequency will typically be within the passband of the system and will appear as ripple on the low pass filter outputs 49 3 Calculate the harmonics of the reference square waves for all eight transmitter frequencies Since the DG303 uses TTL compatible input levels the reference waveforms are assumed to be perfect square waves and hence only odd harmonics are included All odd harmonic
208. derably more prohibitive and the circuit board area increased by a factor of 8 Appendix A on page 116 includes a table of the costs involved in manufacturing a dual plane 16 electrode FDM impedance tomography system From a theoretical standpoint the operation of the FDM impedance tomography concept can be considered for the simple case of a 4 electrode system with sine wave excitation Given that the two transmitter outputs are described by the following equations Vqxa t sin w at 4 2 4 sin wgt 4 3 then the receiver output consists of a superposition of these two waveforms where each waveform is scaled and phase shifted by the conductance and capacitance of that particular transmitter receiver electrode pair The output waveform for receiver A is therefore t K raga 5 5 1 4 4 Each received signal can be resolved into two components namely a conductance component that is in phase with the original transmitter output and a capacitance component that is phase shifted by 90 38 pp 30 35 21 Consequently the above equation can be rewritten as follows t Gryanxa Sin w Crxarxa COS W Grygera Sin wgt cos wat 4 5 To determine the conductance between transmitter A and receiver A this signal is then multiplied by the reference for transmitter A as follows t Gryarxa Sin w At sin w
209. dex 1 Il offset the data according to which system wish to perform JI calibration for if RecLoadDBase GetTomoSystem 1 then col col 8 while col lt 12800 do begin calibdata index result col inc row inc index if row 8 then begin row 0 inc col 9 end else inc col end for index 1 to 128 do calibval index 0 for index 1 to 6400 do begin row index mod 128 if row 0 then row 128 calibval row calibval row calibdata index end for index 1 to 128 do calibval index calibval index 50 end end MM GENERAL PROCEDURES override the default Windows message handling procedure to capture the sampling complete message when sampling is complete call StopSampling and process the results procedure TRecTest WndProc var TheMsg TMessage begin if TheMsg Msg notifvalue then begin SampHardware StopSample result ProcessResults end else inherited WndProc TheMsg end close the form procedure TRecTest RecTestReturnClick Sender TObject begin 177 RecTestTimer Enabled false Close end provide the option of stopping the testing by disabling the timer which triggers the start of a new test procedure TRecTest RecTestStopClick Sender TObject begin if RecTestTimer Enabled then begin RecTestTimer Enabled false RecTestStop Caption Continue end else begin RecTestTimer Enabled
210. dex 1 totaltraingravelpixels airerror airerror Abs VolumeData rowindex 0 totaltrainairpixels end calculate the threshold and sum void error percentages for the complete training database currtrainthresh 100 totalerrors 88 stop start 1 currtrainsum watererror gravelerror airerror stop start 1 end tm MODEL SEARCH PROCEDURES conduct a model search to determine the optimal number of hidden layer neurons in a double layer network This is achieved by continually training networks where the number of hidden layer neurons is incremented after each network has been trained The current network performance is then compared to the best performance so far and the records updated appropriately procedure TRecTrainNetwork ModelSearchTimerTimer Sender TObject begin ModelSearchTimer Enabled false PauseTrain Enabled false RecModelCurrent Caption inttostr numhidden RecNetworkProgressBar Position 0 RecMain RecWeightOKClick Sender train the network for the current number of hidden layer neurons TrainDoubleNetwork compare the current performance to the best achieved performance and update the record of the best performance appropriately keep a 11 record of the weights for the best network by calling StoreBestWeights if RecSetParam RecStopVolume Checked then begin if currsum lt bestsum then begin numbest numhidden bestsum currsum bestwater watererror
211. display the results inc currenttest InpTable FindKey currenttest 0 001 OutTable FindKey currenttest 0 001 DisplayData UpdateTables end else begin I if at the end then create space for numframes of records and set the TestNo fields Also clear the tables to illustrate that sampling must still take place inc currenttest for vers 1 to numframes do begin InpTable Append OutTable Append InpTable FieldByName TestNo Value currenttest vers 1000 OutTable FieldByName TestNo Value currenttest vers 1000 for col 1 to 8 do for row 1 to 8 do begin AddCapTable Cells col row 0 AddResTable Cells col row 0 end end end end update the tables with the values in the database procedure TAddToDatabase UpdateTables var row col index overall integer tresult TResultTable 51 82 string begin index 1 for overall 1 to 16 do begin case overall of 1 begin row 4 tresult capacitance end 2 begin row 4 tresult resistance end 3 begin row 3 tresult capacitance end 4 begin row 3 tresult resistance end 5 begin row 2 tresult capacitance end 6 begin row 2 tresult resistance end 7 begin row 1 tresult capacitance end 8 begin row 1 tresult resistance end 9 begin row 5 tresult resistance end 10 begin row 5 tresult capacitance end 11 begin row 6 t
212. ds passing between the capacitance plates as a cross section of the measurement section Block diagram illustrating the interactions between the different circuit boards in a 16 electrode PAGE N 12 13 14 19 20 24 25 27 29 34 35 36 37 38 39 5 8 5 9 5 10 5 11 5 12 5 16 5 17 5 18 5 19 5 20 6 1 6 2 6 3 6 4 6 5 6 6 7 1 7 2 7 3 frequency division multiplexed impedance tomography system where the number next to each line indicates the number of signals 40 Block diagram of the output transmitter with push pull power MOSFET output stage 42 Operation of cross conduction prevention circuitry in push pull power MOSFET output stage 42 Block diagram of capacitance receiver 43 Block diagram of conductance receiver 43 Block diagram of the synchronous detector circuit board showing only one complete channel whereas each synchronous detector circuit board actually supports two complete channels 45 Example of the square wave harmonics mixing in the multiplication stage of the synchronous detector for a 4 electrode system 48 Output square wave generation with 0 and 90 references for the synchronous detection of the conductance and capacitance signals respectively 53 Block diagram of frequency generator circuit board showing the eight PIC16F84A microcontrollers used to generate the eight independent transmitter frequencies and their quadrature waveforms 54 Compari
213. ductance and capacitance of the material between that particular transmitter receiver electrode pair Synchronous detection is then performed on the receiver output to separate this signal into the components from the different transmitters Synchronous detection is also called phase sensitive detection lock in detection and coherent demodulation The following section will examine the practical application of this concept to an 8 electrode system This concept will then be proven from a theoretical standpoint 19 TRANSMITTER Vout proportional to Gaa Vour proportional to Caa Vour proportional to Gas Vout proportional to Cas Vour proportional to Gac NOTATION Gas is the conductance between receiver A and transmitter B Vour proportional to Cac Cas is the capacitance between receiver A and transmitter B Vour proportional to Gao Vour proportional to Cap FIGURE 4 2 BLOCK DIAGRAM ILLUSTRATING THE CONCEPT OF FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY OF AN 8 ELECTRODE SYSTEM ONLY THE SIGNALS FOR RECEIVER A ARE INCLUDED SPECIFICALLY THE SOLID LINE REPRESENTS THE PATH OF THE RECEIVED SIGNAL AND THE DASHED LINES REPRESENT THE PATHS OF THE REFERENCE WAVEFORMS IN THE SYSTEM THE SIGNAL FROM RECEIVER A IS APPLIED TO EIGHT SYNCHRONOUS DETECTORS OPERATING IN PARALLEL EACH SYNCHRONOUS DETECTOR CONSISTS OF A MULTIPLIER FOLLOWED BY A LOW PASS FILTER THE LPF BLOCK REPRESENTS THE LOW PASS FILTERS IN THE ABOVE DIAGRAM FOR THE C
214. e 9 WriteLn tempfile 9 WriteLn tempfile 9 WriteLn tempfile 9 for i 1 to numextra WriteLn tempfile WriteLn tempfile 9 WriteLn tempfile 9 WriteLn tempfile 9 WriteLn tempfile 9 for i 1 to numextra do WriteLn tempfile WriteLn tempfile 9 WriteLn tempfile 9 WriteLn tempfile 9 CLRF OVWF PORTB OP OP do OVLW B 0000001 1 OVWF OP OP do 9 OVLW B 00000001 OVWF OP OP 9 OVLW B 00000000 OVWF PORTB for i 1 to numextra do WriteLn tempfile 9 NOP WriteLn tempfile 9 GOTO LOOP WriteLn tempfile 9 END CloseFile tempfile end end compile the text file procedure TAutoMain CompileClick Sender TObject var tempstring string begin tempstring MPASM SaveDialog FileName pPIC16F84 alNHX8M WinExec PChar tempstring SW_SHOWMAXIMIZED end download the text file procedure TAutoMain DownloadClick Sender TObject begin WinExec PP SW_SHOWMAXIMIZED end end 137 APPENDI X J PBI L OPTIMISATION ALGORITHM AND SIMULATOR PROGRAM CODE PROGRAM LISTING OF PBILFORM PAS following application performs PBIL to determine the set of optimal frequencies for an 8 electrode FDM impedance tomography system The program uses all the standard crystal frequencies and all division factors that can be prog
215. e can cause iterative image reconstruction algorithms to diverge 23 In addition multiple local optima are often encountered 25 The algorithm can then become stuck at these local optima 17 Reconstruction time is also greatly increased as a result of the iterations making it unacceptable for industrial processes that have rapid dynamics 25 According to Byars 11 iterative reconstruction techniques produce images of good resolution that are close to the theoretical limit determined by the measurement protocol and number of electrodes An example of an iterative reconstruction algorithm for resistance tomography is the modified Newton Raphson reconstruction algorithm The Newton Raphson iteration is applied to the standard non linear reconstruction algorithm which minimises the mean square difference between the measured and estimated voltage responses as D p 1 2 f p Vo fp Vo 3 3 where Vo is the measured voltage and f p is the estimated voltage for a resistivity distribution p 19 Initially a resistivity distribution p equal to the vessel filled with the background phase is assumed The forward problem is then solved using a finite element model and a set of voltages is calculated for this resistivity distribution These voltages are then compared to the measured voltages and the least squares error is calculated between the two 10 voltages sets as described above 27 If itis less than a predefined value then
216. e probability vector are steadily shifted away from 0 5 and towards either 1 0 or 0 0 thus defining a single high evaluation trial solution A mutation operator is also provided to ensure that the algorithm does not converge too quickly on a sub optimal solution Mutation in PBIL can either be performed on the trial solutions themselves or on the probability vector 61 The bits to be mutated in the probability vector are chosen randomly according to the mutation probability value and are updated according to the following equation 61 PV PV x 1 random 0 or 1 x SHIFT 5 4 where SHIFT controls the amount of mutation to apply to the probability vector 51 Overall the operation of the PBIL optimisation algorithm can be described using the following pseudo code for all j do initialise PV to 0 5 for generations do for population do generate trial solution using PV evaluate fitness of trial solution if fitness better than best trial solution TS fitness then update record of best trial solution TS for all j do adjust PV towards TS using the probability update rule PV Pv jx LR LR x TS if mutation is to be applied to bit j then mutate probability vector according to the mutation update rule PV PVi x 1 SHIFT random 0 or 1 SHIFT As is evident above the PBIL algorithm requires the setting of four parameters namely the population size the learning rate the mutation probability and the mutation s
217. e reference for the multipliers is also a square wave a large number of frequency components are generated at the multiplier output since the sum and difference of all the frequency harmonics in the received signal and the reference waveform are generated In an initial attempt to improve the situation it was decided that the bandwidth of the receiver signals should be limited This motivated the filtering of the high frequency components in the capacitance and conductance receivers as discussed in the previous section Although attenuation of these upper frequency components was achieved they were not completely removed and subsequently ripple was still present on the synchronous detector outputs The output ripple limits the accuracy of the capacitance and conductance measurement systems Specifically as proven in Chapter 4 the information regarding the capacitance or conductance between a particular transmitter receiver electrode pair is incorporated in the DC component on the low pass filter output Any ripple superimposed on this DC component limits the accuracy with which the capacitance or conductance can be determined if only a single measurement is taken Subsequently the accuracy and more importantly the resolution of the impedance tomography system is limited Although averaging could improve the measurement accuracy the required frame rate creates a time constraint resulting in only single measurements being taken Figure 5
218. e PC30G card must be configured as follows A D configuration single ended 5V to 5V range slave trigger DMA single DMA channel 5 base address 1000H unit sampunit interface uses EDR32 const rst555 byte 00 sequence to reset 555 on sample control rstclk byte 02 sequence to reset counter on sample control card 510555 byte 01 sequence for run mode datamask word 0FFF mask to extract A D reading from raw binary data type TSamp class private boardnum integer numframes integer busy_sampling Boolean number of frames to capture Boolean indicating whether the system is currently sampling buffer array 0 65535 of word buffer store for sampled data public II initialise the card and get Windows notification value function InitSample bdnum integer Handle cardinal cardinal check the status of the sampling function GetSampleStatus Boolean release the resources allocated to sampling procedure ReleaseSample start sampling where the total number of frames to be captured is nmfrmes procedure StartSample nmfrmes integer once sampling is complete process sampled data procedure StopSample var result array of double end implementation intialise the sampling by getting the notification message value setting digital IO port A to output loading channel list with channel 0 to 15 setting the burst length to 16 setting the trans
219. e electrode configurations considered Wang 53 confirmed this result for the case of resistance tomography In contrast Ma et al 54 55 illustrate the compression of the measurement volume through the use of driven axial guards for resistance tomography For the laboratory scale rig being developed guard electrodes were provided for the capacitance plates but no guarding was provided for the conductance probes This ensured that reasonable standing capacitances were measured and that a near parallel electric field improved the accuracy of the reconstructed images The axial length of the measurement volume plays an important role in the cross correlation algorithm In particular it affects the spatial filtering and hence the bandwidth of the signals to be correlated 50 The effective cutoff frequency of capacitance electrodes of DRIVEN GROUNDED AXIAL AXIAL GUARD GUARD length at a flow velocity u is 50 pan 5 1 Small capacitance electrodes are desirable in order to CAPACITANCE RECEIVER obtain wide spatial bandwidth signals 56 as particles pass into and out the sensing volume 57 However short electrodes result in small capacitance readings Os CONDUCTANCE PROBE CONDUCTANCE PROBE LEAKAGE CURRENT SHIELD and the measurement volume must be large relative to the size of the component particles to ensure an accurate volume fraction measurement 57 Hence a compromise is made The dimensions
220. e fraction at the top plane This represents the region where the bubbles are situated between the two measurement planes This change is probably a consequence of signals from the top plane being detected at the bottom plane and vice versa Further it is observed that a good correlation exists between the measured volume fraction profile and the desired volume fraction profile As is evident in figure 7 8 on page 93 a slight offset exists between the measured and desired volume fraction profiles at the top measurement plane As explained earlier this is a result of the slippage between the drive belt and the pulley of the flow simulation apparatus Since the effect is cumulative as the bubbles move further up the pipeline so the difference in position at the top measurement plane is more noticeable than at the bottom measurement plane The average dynamic volume fraction error was calculated at 0 6896 for the top measurement plane and 0 5896 for the bottom measurement plane using a single layer feed forward volume fraction predictor neural network These results represent the absolute error and are rounded off to two significant digits Figure 7 9 on page 95 is a sequence of screen captures of the image reconstruction results for this test It is important to note that these reconstructions were performed in real time This is in contrast to standard TDM impedance tomography systems where the required number of frames are captured first and the reconstruct
221. e lItemIndex 0 then begin OutputData row 0 TrainOutTable FieldByName AirVolume Value OutputData row 1 TrainOutTable FieldByName GravelVolume Value end else set a flag so that this test case can be removed Il at a later stage else if RecDesirePhase ItemIndex 1 and tempgravel or RecDesirePhase ItemIndex 2 and tempair then begin OutputData row 0 OutputData row 1 end end 4 4 II load the volume fractions into volume data to speed up the calculation of the sum void error VolumeData row 0 TrainOutTable FieldByName AirVolume Value VolumeData row 1 TrainOutTable FieldByName GravelVolume Value VolumeData row 2 TrainOutTable FieldByName WaterVolume Value inc row RecLoadProgressBar Steplt Application ProcessMessages end end load the test data for testversions 1 to RecTestVers Value do begin for testnumber RecTestStart Value to RecTestStop Value do begin tempair false tempgravel false index testnumber testversions 1000 TestOutTable FindKey index TestInpTable FindKey index InputData row 0 1 load the input voltages for col 1 to 128 do begin 51 Voltage 4inttostr col InputData row col TestInpTable FieldByName s1 Value end if the desired output is an image of the vessel cross section if RecDesiredRecon ItemIndex 0 then begin determine the phases in this test case for col 1 to 100 do begin
222. e more complex task of a three phase air gravel water reconstruction 7 4 Three phase Air Gravel Water Reconstruction Results Although the intended application requires accurate volume fraction predictions of the air gravel and water phases image reconstructions will be performed as a means of visually assessing the accuracy of the system Screen captures of these image reconstruction results will be included in the following sections where appropriate 7 41 Real time volume fraction prediction and image reconstruction results In order to perform three phase reconstructions the appropriate neural network reconstruction algorithms were loaded before the commencement of the real time testing A consequence of the increased processing required to perform a three phase reconstruction was a reduction in the on line frame rate Specifically the frame rate of the three phase image reconstruction decreased to approximately 140frames s and the frame rate of the three phase double layer volume fraction predictor decreased to approximately 180frames s The frame rate of the single layer volume fraction predictor remained approximately 280frames s The consequences of this reduction in frame rate will be discussed at a later stage Tests were conducted using both air and gravel bubbles and their dynamic volume fraction errors were recorded Table 7 3 compares the average dynamic volume fraction errors for the different reconstruction algorithms considered at
223. e near the right hand edge of the rig moving through the measurement planes at the same time 95 7 10 Effect of the bubble length on the volume fraction measurement accuracy 96 7 11 Plots of capacitance and conductance variations for receiver A as the water level is lowered past the top measurement plane 97 7 12 Screen captures of the top measurement system image reconstruction as the water level is lowered past the conductance probes 98 7 18 Photograph of the sparger used to generate a homogenous bubble flow in the laboratory scale rig 99 7 14 Plot of the air volume fraction measurements at the top and bottom measurement planes for a homogenous bubble flow 99 7 15 Screen captures of a homogenous bubble flow where for each screen capture the top frame is the top measurement plane and the bottom frame is the bottom measurement plane 100 7 16 Cross correlation velocity measurement accuracy for the 0 01 volume fraction air bubble 101 7 17 Plots of the volume fraction profiles at the top and bottom measurement planes for two 0 04 volume fraction air bubbles moving at 1m s and 0 2m s respectively 102 7 18 Screen captures of a simulated 0 2m s bubble flow where the 0 04 volume fraction gravel bubble is situated near the left hand edge of the rig and the 0 04 volume fraction air bubble is situated near the right hand edge of the rig 104 7 19 Plots of the air and gravel volume fraction profiles at the top and bottom measurement planes wher
224. e operation of the on line real time reconstruction software can be explained using the following pseudo code Start the capture of the first set of 50 frames Set busysampling to true WAIT until sampling completes Start the capture of the next set of 50 frames Set busysampling to true LOOP WHILE busysampling is true DO Start reconstructions for previous set of 50 frames Set busyreconstruction to true IF busysampling is false but busyreconstruction still true THEN Complete reconstructions for previous set of 50 frames ELSE IF busysampling is false and busyreconstruction is false THEN Start the capture of the next set of 50 frames Set busysampling to true ELSE IF busysampling still true THEN WAIT until the sampling completes Start the capture of the next set of 50 frames Set busysampling to true REPEAT LOOP while there are still more frames to be captured The use of these flags ensures that a consistent timing relationship is maintained between the sampling process and the reconstruction process For both image reconstruction and volume fraction prediction neural networks the correlation is performed on the volume fraction profiles of the air and gravel phases at the top and bottom systems A standard point by point cross correlation is performed For the cross correlation calculation it is necessary to include values of one of the signals beyond the measurement period T The standard solution to this problem is to extend the appropriate data r
225. e optimal frequency separation and hence minimal output ripple 50 5 2 4 Population based incremental learning fundamentals Population based incremental learning PBIL is a stochastic search technique that combines the mechanisms of genetic algorithms with competitive learning to produce an optimisation tool that is simpler than genetic algorithms and which out performs genetic algorithms both in terms of speed and accuracy for a large set of optimisation problems 61 Genetic algorithms are biologically motivated adaptive systems that are based on the principles of natural selection and genetic recombination 61 The operation of a genetic algorithm to solve an optimisation problem can be explained as follows genetic algorithms maintain a population of trial solutions to a particular optimisation problem These trial solutions are initially randomly generated For each generation these trial solutions are evaluated as to their fitness which is a measure of how well a trial solution optimises the objective function 61 Trial solutions that produce good results are then combined using various recombination operators to generate a new generation of potential solutions In addition genetic diversity must be maintained to ensure that the algorithm performs a thorough search of the feature space and does not converge too quickly on a sub optimal solution This is achieved through mutation Mutation helps preserve diversity by introducing random changes
226. e physical properties of the process itself since smaller bubbles would give rise to the generation of signals with higher frequency components as the bubbles move through the measurement space 2 the flow velocity since there are more transitions per second for a faster flow 3 the spatial filtering effect of the sensor The spatial filtering effect often determines the maximum frequency component of the correlated signals Further if a digital cross correlator is used it is important to satisfy the Nyquist sampling theorem to ensure erroneous results are not generated through aliasing 50 It can be shown that the variance of the measured time delay for bandwidth limited white noise is given by the following equation 50 0 038 ey where B is the bandwidth of the system T is the time duration over which the correlation is performed and 0 defines the mean square signal to noise ratio From this it is evident that a slight decrease in the nn system bandwidth B will result in a sharp increase in the variance of time delay measurement thus indicating the importance of high bandwidth signals for cross correlation 50 Various assumptions are made when measuring the velocity field by cross correlating tomographic information 5 9 1 the size of the sensor is small relative to the separation between planes 2 there exist measurable disturbances in the flow field being investigated 3 the velocity field can
227. e replaced with type 316 marine grade stainless steel In addition carbon probes were tested since carbon is considered both durable and chemically resistant To provide a reference the measurement section was filled with potable water and a drift analysis was performed Since potable water has a far lower conductance than seawater the drift resulting from electrochemical reactions should be greatly reduced Table 6 1 compares the average drift of both the capacitance and conductance readings for the different conductance probe materials tested over a 15 hour period It also includes the drift measured when the measurement section is filled with potable water TABLE 6 1 COMPARISON BETWEEN THE DRIFT OF THE CAPACITANCE AND CONDUCTANCE READINGS OVER A 15 HOUR PERIOD USING DIFFERENT CONDUCTANCE PROBE MATERIALS STANDARD MARINE GRADE CARBON PROBE POTABLE WATER STAINLESS STEEL STAINLESS STEEL REFERENCE CAPACITANCE 121mV 235mV 159mV 28mV CONDUCTANCE 167mV 269mV 264mV 34mV Since insufficient funds were available to investigate more advanced conductance probe materials it was decided that all future tests should be done with potable water This required an increase of the conductance receiver gain A larger input resistor was used to increase the conductance receiver output and specifically an 8200 resistor was used whereas a 15Q resistor had been used for seawater tests This is only a temporary solution and clearly more research must be d
228. e software performs an on line reconstruction at a rate of 280frames s As mentioned previously the data is captured in sets of 50 frames and is controlled by the EDR device driver as a background process This creates an interval during which the reconstructions for the previous set of 50 frames can be performed and their results displayed To ensure a balance is maintained between the two processes a busysampling and busyreconstruction flag are used These flags are defined as follows busyreconstruction is true while the program is performing the reconstructions for the previous set of 50 frames and busysampling is true while the next set of 50 frames is being captured It is important to note that although the data is captured in sets of 50 frames a constant sampling rate is still maintained This is an important requirement if the system is to be used for on line control purposes Specifically the time taken to perform the reconstructions for the previous set of 50 frames is typically shorter than the time taken to capture the next set of 50 frames Subsequently the sampling process can start the capture of the next set of 50 frames as soon as it finishes the capture of the current set of 50 frames since it does not have to wait for the reconstruction process to complete In this way both 87 tomography systems are sampled at a constant frame rate This will be examined in greater detail in section 7 4 3 on page 106 Overall th
229. e task of the low pass filter to extract the DC component from this signal It is evident in figure 4 6 that an inverted DG303 output is equivalent to the multiplication of the receiver signal by the reference waveform The situation for capacitance detection is very similar to the above except that the switches are controlled by Ago instead of Ap so as to produce an output proportional to the capacitance component of the received signal Since the DG303 is dual single pole double throw CMOS analogue switch the capacitance demodulation for a particular transmitter receiver electrode pair is performed using the same integrated circuit as the conductance demodulation detailed above 28 REFERENCE WAVEFORM FROM PIC TRANSMITTER OUTPUT RECEIVED SIGNAL TO LOW PASS FILTER BUFFERED RECEIVER OUTPUT INVERTED RECEIVER REFERENCE WAVEFORM OUTPUT REFERENCE SWITCH 1 SWITCH 3 OPEN CLOSED OUTPUT FILTER FIGURE 4 6 SYNCHRONOUS DETECTION OF THE RECEIVER SIGNAL USING A DG303 CMOS ANALOGUE SWITCH TO EMULATE MULTIPLICATION BY A REFERENCE WAVEFORM NOTE THAT SOME OF THE DELAYS GENERATED BY NON IDEAL CIRCUIT RESPONSES SUCH AS SLEW RATE ARE SHOWN IN THE WAVEFORMS FOR THE SAKE OF REALISM INCLUDED IS A TABLE SPECIFYING THE STATES OF SWITCHES 1 AND 3 IN THE DG303 BASED ON THE REFERENCE WAVEFORM When all transmitters are operating simultaneously the received signal consists of a superposition of components from the
230. e the drift timer overflows open the text file add all 128 voltages as a single line to the file separated by spaces and close the file again also increment sampnumber to give visual indication of the number of frames that have been saved to this file procedure TCurMain DriftTimerTimer Sender TObject var value string TOW col integer begin value for col 1 to 8 do for row 1 to 8 do value value CurCapTable Cells col row for col 1 to 8 do for row 1 to 8 do value value CurResTable Cells col row AssignFile driftfile driftfilename Append driftfile WriteLn driftfile value CloseFile driftfile SampleNumber Caption inttostr sampnumber inc sampnumber end procedure to write the contents of tables to Excel spreadsheet for analysis l at a later stage procedure TCurMain CalibrateClick Sender TObject var tempfile textfile TOW col integer begin if DriftSave Execute then begin AssignFile tempfile DriftSave FileName ReWrite tempfile if CurSystemSelect ItemIndex 0 then WriteLn tempfile system else WriteLn tempfile Bottom system WriteLn tempfile DateToStr Date WriteLn tempfile Capacitance for row 0 to 8 do begin 148 for col 0 to 8 do Write tempfile CurCapTable Cells col row WriteLn tempfile end WriteLn tempfile Resistance for row 0 to 8 do begin for col 0 to 8 do W
231. e the software developed for the dual plane 16 electrode FDM impedance tomography system 5 3 1 Data capture software description As mentioned previously the polled sampling technique used for the 8 electrode impedance tomography system is not fast enough to achieve the desired frame rate In addition the maximum data capture rate can only be achieved if DMA is used Specifically since 128 voltages must be captured from two systems at a rate of 200frames s the desired sampling rate is 51 2ksamples s Therefore it was necessary to modify the data capture software to use streaming so that the maximum frame rate could be achieved Streaming is similar to single channel DMA sampling in that it only requires a single DMA channel However streaming allows for unlimited amounts of data to be transferred through the use of a circular buffer 63 All the register level instructions for programming the Eagle data acquisition cards are incorporated in the EDR software development kit provided with the card Specifically this functionality is accessed by linking the associated dynamic link library DLL device driver into a program at run time These high level procedures then perform the register level operations to achieve a specific task An additional complication of the current system is that the analogue multiplexers have to be configured for each set of 16 voltage readings Initially the address lines of these multiplexers were to be driven by the digi
232. e the velocity of the 0 04 volume fraction air bubble was 0 5m s and the velocity of the 0 04 volume fraction gravel bubble was 0 2m s 105 TABLES 1 1 Required specifications of the flowmeter to be developed for the on line monitoring of an air gravel seawater mixture in an offshore mining application 3 3 1 Dielectric constants of the materials used during testing 15 4 1 Output frequencies generated with a 4MHz crystal for the 8 electrode system 26 4 2 LM675 specifications appropriate to this application 26 4 3 Comparison between the performance of the freguency division multipleked impedance tomography system and a standard time division multiplexed impedance tomography system for a two phase air seawater image reconstruction using a single layer feed forward neural network 31 44 4 5 4 6 5 1 5 2 6 1 6 2 6 3 6 4 6 5 6 6 7 1 7 2 7 3 Comparison between the performance of the frequency division multiplexed impedance tomography system and a standard time division multiplexed impedance tomography system for a two phase gravel seawater image reconstruction using a single layer feed forward neural network Comparison between the performance of the frequency division multiplexed impedance tomography system and a standard time division multiplexed impedance tomography system for a three phase air gravel seawater image reconstruction using a single layer feed forward neural network with a 1 of C output enc
233. ecLoadStatus Visible true RecMain Viewdatabase1 Enabled true RecMain Writedatatoatextfile1 Enabled true end end t DATA ACCESS PROCEDURES function TRecLoadDBase GetTotalTrain integer begin GetTotalTrain traintotal end function TRecLoadDBase GetTotalTest integer begin GetTotalTest testtotal end procedure TRecLoadDBase GetPhases var contair contgrav Boolean begin contair containsair contgrav containsgravel end function TRecLoadDBase GetNumOutputs integer begin GetNumOutputs numoutputs end function TRecLoadDBase GetTomoSystem integer begin GetTomoSystem tomosystem end procedure TRecLoadDBase SetPhases contair contgrav Boolean begin containsair contair containsgravel contgrav end tm PRE PROCESSING PROCEDURES pre process the loaded data by optionally performing calibration and then standardise procedure TRecLoadDBase RecPreProcessClick Sender TObject begin RecLoadProgressBar Position 0 RecLoadProgressBar Max traintotal testtotal RecLoadProgress Visible true RecLoadProgress Update if RecLoadCalibrate Checked then PerformCalibrate if only desire two phase reconstruction then need to remove those test cases relating to the other phase as well as three phase mixtures if RecDesirePhase ItemIndex lt gt 0 then ReOrganiseData RecLoadProgressBar Positio
234. ecified distance This procedure is used for flow simulation purposes where a bubble is moved over a fixed distance from the base of the rig to the top of the rig at a fixed velocity StopDrive brings all the drives to a controlled stop using the defined deceleration ramp e GetMotorData accesses the fast data area of the stepper controller to determine the motor position and velocity Since an encoder is not used these readings are meaningless if the stepper motor fails 84 Tests were conducted to determine the maximum acceleration of the stepper motors under load Clearly it is desirable to use the maximum possible acceleration so as to permit the simulation of higher flow velocities within the limited distance of the flow simulation apparatus However due to the large diameter of the pulley the torque of the stepper motors limited the acceleration to approximately 50 52 Consequently the maximum velocity that could be simulated is 4 7m s at the centre of the rig and the maximum average velocity between the two measurement planes is 4 2m s Once the bubbles are attached to the drive belt and the apparatus is inserted into the water of the laboratory scale rig then these maximum velocities are reduced However the buoyancy of the air bubble partly compensates for the friction of the water in a simulated upward flow Specifically the maximum velocity of the 0 01 volume fraction air bubble is 4 2m s and the maximum velocity of the 0
235. ecord for a further period of T seconds where x t 0 for t T 64 The peak of each of the cross correlations is calculated and this corresponds to the number of frames delay between the top and bottom systems This is then converted into a transit time using the estimated frame rate In particular a system timer keeps track of the total time taken to perform the specified number of frame captures and reconstructions and hence is able to estimate the on line frame rate Using these transit times and the specified plane separation the software proceeds to calculate the velocities of the air and gravel phases A comparison is also made to the desired phase velocities as specified when the stepper controller was configured Since the positions of the bubbles are exactly controlled by the stepper controller it is possible to perform a comparison between the measured dynamic response of the dual plane impedance tomography system and the desired response Clicking on Verify results displays a form that is used to perform this dynamic comparison and a Screen capture of this form is included in figure 7 6 on page 89 A full scale version of this screen capture is included in Appendix N on page 210 The plots of the different volume fraction profiles of the air and gravel phases for both the top and bottom system are displayed 88 Dynamic system verification Test case configuration r Air bubble configuration Gravel bubble configuration Capacit
236. ed Since this technique permits no means of verifying the synchronisation between the hardware and the card it is necessary to ensure that such a loss of synchronisation does not occur Tests have shown that sufficiently reliable results can be achieved if the maximum number of frames captured without software intervention is limited to 50 Consequently all further data capture software will capture readings in sets of 50 frames so as to maximise the efficiency introduced through the use of streaming while at the same time minimise the possibility of a loss of synchronisation The operation of the sampling unit can be explained with reference to figure 5 18 on page 60 When the program starts the sampling unit and the PC30G card are initialised by calling InitSample This procedure configures the sample controller board by resetting the 4 bit counter and holding the 555 in the reset state It then proceeds to configure the PC30G data acquisition card as follows e the channel list represents the sequence of analogue input channels that must be sampled when a trigger is received InitSample loads channel 0 to 15 into the channel list and sets the gain of each channel to unity e burst mode is a feature provided by the data acquisition card where for every trigger received a specific number of channels are sampled at the maximum sampling rate of the card instead of just a single conversion InitSample configures the PC30G card to use b
237. ed that the neural network would perform better reconstructions of these situations if the training database included examples of this masking effect Consequently it was decided that the training database for the 16 electrode FDM impedance tomography system should be extended to include these configurations as well as combinations of both air and gravel bubbles Tests were conducted to compare the performances of neural networks trained on the extended training databases and those trained on the original databases Results showed that the neural networks trained on the extended databases consistently outperformed those trained on the original databases in terms of the detection of bubbles at the centre of the pipeline and the accurate reconstruction of configurations exhibiting the masking effect 70 Finally smaller bubbles were used in the generation of the training database in an attempt to extract the maximum possible resolution from the 16 electrode FDM impedance tomography system Previously the smallest bubble used in the training database generation corresponded to 20 of the pipeline diameter thus setting the maximum achievable resolution at 2096 This bubble size corresponds to four pixels in the 10 by 10 image reconstruction In tests the previous system was only able to detect the 2096 bubble towards the edges of the pipeline For the 16 electrode FDM impedance tomography system the smallest bubble simulated corresponds to 10 of the
238. edure TStep GetMotorData want TRequestData var axis1 double var axis2 double var axis3 double begin instruct the AT6400 to update the fast status area RequestStatus if want Velocity then begin get the drive velocity SetPointer AXIS1_VELOCITY ReadStatus word_high word_low fast_status axis1 fast_status 400 ReadStatus word high word low fast status axis2 fast status 400 ReadStatus word high word low fast status axis3 fast status 400 end else if want MotorPos then begin get the drive position SetPointer AXIS1_MOTOR ReadStatus word_high word_low fast_status axis1 fast_status ReadStatus word_high word_low fast_status axis2 fast_status ReadStatus word_high word_low fast_status axis3 fast status end end download the operating system procedure TStep DownloadOS begin WinExec AT6400 EXE SW SHOWMINIMIZED end reset the drive limits procedure TStep ResetLimits var temp string begin temp LHO0 191 ProcessCommand temp temp LH3 ProcessCommand temp end set the acceleration of the drives procedure TStep SetAcceleration axis1 axis2 axis3 integer var temp string begin temp A inttostr axis1 inttostr axis2 inttostr axis3 10 ProcessCommand temp temp AD inttostr axis1 inttostr axis2 inttostr axis3 10 ProcessCommand temp end end PROGRAM LISTING OF DYNUNIT PAS
239. egistry Create try Reg OpenKey Software Neural True Reg Writelnteger OutputDesired RecDesiredRecon ItemIndex Reg WriteString TrainDatabase RecTrainName Text Reg WriteString TestDatabase RecTestName Text Reg Writelnteger TrainStart RecTrainStart Value Reg Writelnteger TrainStop RecTrainStop Value Reg Writelnteger TestStart RecTestStart Value Reg Writelnteger TestStop RecTestSTop Value Reg Writelnteger TrainVers RecTrainVers Value Reg Writelnteger TestVers RecTestVers Value Reg Writelnteger DesiredLoadPhase RecDesirePhase ItemIndex Reg WriteBool Calibrate RecLoadCalibrate Checked finally Reg Free end if the data has been loaded then close the tables and free the 1 allocated resources if opened then begin TrainInpTable Close TrainOutTable Close TestlnpTable Close TestOutTable Close TrainInpTable Free TrainOutTable Free TestInpTable Free TestOutTable Free end end end PROGRAM LISTING OF CALIBSET PAS This unit simply provides an interface for the user to stipulate the endpoints for the calibration feature these endpoints represent test cases where the rig was filled only with water and provide a means of compensating for the drift that takes place over the days of generating training and test databases unit calibset interface uses Windows Messages SysUtils Classes Graphics Controls Forms Dialogs StdCtrl
240. eight rowindex colindex for colindex 0 to numoutputs 1 do for rowindex 0 to numhidden do OutputWeightBest rowindex colindex OutputWeight rowindex colindex end pause the model search procedure TRecTrainNetwork PauseTrainClick Sender TObject begin if ModelSearchTimer Enabled true then begin ModelSearchTimer Enabled false PauseTrain Caption Resume Training end else begin ModelSearchTimer Enabled true PauseTrain Caption Pause training end close the form procedure TRecTrainNetwork RecNetworkReturnClick Sender TObject begin Close end when create the form initialise the mulitplication factors for RPROP These factors control the rate at which RPROP converges on a solution procedure TRecTrainNetwork FormCreate Sender TObject begin changeplus 1 2 changeminus 0 5 end when show form update the display depending on the type of training to be performed procedure TRecTrainNetwork FormShow Sender TObject begin RecTrainResults Visible RecSetParam RecShowTrainPerf Checked PauseTrain Enabled true PauseTrain Visible false if RecLoadDBase GetNumOutputs 2 then begin RecTestThreshErr Visible false RecTrainThreshErr Visible false RecThreshLabel Visible false end else begin RecTestThreshErr Visible true RecTrainThreshErr Visible true RecThreshLabel Visible true end 174 tM DATA ACCESS PROCE
241. eightsum weightsum HiddenWeight rowindex colindex 1 testdata rowindex prevent arithmetic overflow if weightsum gt 200 then weightsum 200 else if weightsum lt 200 then weightsum 200 tanh activation function testhiddenoutput colindex tanh weightsum end end end calculate the network output for colindex 0 to numoutputs 1 do begin weightsum 0 for rowindex 0 to upperindex do begin if numhidden gt 0 then tempvalue testhiddenoutput rowindex else tempvalue testdata rowindex weightsum weightsum OutputWeight rowindex colindex tempvalue end prevent arithmetic overflow if weightsum gt 200 then weightsum 200 else if weightsum lt 200 then weightsum 200 I if using 1 of C then three outputs corredponding to a pixel can only be calculated once all three weighted inputs for those three outputs have been calculated since uses soft max activation function if numoutputs 264 then begin if colindex 1 mod 3 lt gt 0 then weightstore colindex 1 mod 3 weightsum else begin for i 2 downto 0 do testoutput colindex i RecTrainNetwork SoftMax 3 i weightstore 1 weightstore 2 weightsum end end two phase image reconstruction else if numoutputs 88 then testoutput colindex tanh weightsum else volume fraction prediction begin if weightsum 0 then testoutput colindex weightsum else testoutput colindex 0 end
242. el Once again a clear acrylic lid was constructed with the matrix of 10 by 10 pixels clearly marked Sequences of holes were drilled in this lid and the gravel bubbles were suspended from hooks In this way it is possible to simulate various static configurations of both air and gravel masses Figure 6 3 on page 70 is a photograph of the bubble placement system Each pixel on the clear acrylic lid is numbered and this numbering corresponds to the graphical user interface of the data capture software used to specify the desired network output 69 FIGURE 6 3 PHOTOGRAPH OF THE BED OF HOOKS BASE PLATE AND CLEAR ACRYLIC LID USED AS THE BUBBLE PLACEMENT SYSTEM FOR THE GENERATION OF STATIC TESTS Various changes were implemented in the generation of the training and test database Firstly the bubbles in both the training and test databases were not allowed to come into contact with the conductance probes When a bubble comes into contact with a conductance probe the measurements change dramatically In contrast the changes introduced by moving a bubble at the centre of the pipeline are much smaller By not allowing the bubbles to come into contact with the conductance probes the measurement range is significantly reduced Subsequently it is easier to train a neural network to detect the minor changes resulting from the movement of bubbles at the centre of the pipeline In contrast the previous neural networks were dominated by the bubbles at
243. en sample data keep a record readings array 1 256 of real of the system time as well for sampletime TDateTime comparison end TTomoOutputData record prediction array 1 528 of real when output data copy the sample sampletime TDateTime time across end var RealMain TRealMain result array 0 12800 of double temporary buffer for samples inputstore array 1 6000 of TTomolnputData once the sampled data has 1 been processed it is placed in the inputstore the reconstruction algorithm then I takes points from this store and places the reconstructed 179 outputs in the outputstore outputstore array 1 6000 of TTomoOutputData volumestore array 1 6000 1 4 of real gravelcorr array 1 6001 of real gravel and air correlation aircorr array 1 6001 of real I results reconposition integer II position in matrix where data IL is being taken from for reconstruction totaloutputs integer implementation uses mainunit networkunit stepconfig dynunit loadunit R DFM tM SAMPLING AND RECONSTRUCTION PROCEDURES 5 initialise variables for test start stepper motion and initiate the capture of 50 frames of data procedure TRealMain RealStartClick Sender TObject var i tempsteps integer begin end 3 begin drive3rps StepperConfig airrps drive3steps tempsteps end end end if StepperConfig gravelrps 0
244. ence square wave since Il itis a perfect sqaure wave TTL thresholding include only odd harmonics RefFreq inputs a RefHarm 1 0 139 6 81 for b 2 to 51 do begin RefHarm b RefFreq c inc c 2 end simulte the multiplier output frequencies as a sum and JI difference of the received and reference frequencies d 1 for c 1 to 51 do for b 1 to upper 1 do begin MultOutFreqs d RefHarm c RecvFreq b inc d MultOutFreqs d abs RefHarm c RecvFrea b inc d end Qsort MultOutFreqs 0 d 2 I clock frequency is generally the lowest transmitter frequency clock inputs 0 simulate the aliasing in the switched capacitor filters for c 1 to d 1 do begin if MultOutFreqs c gt clock then begin numfact floor MultOutFreqs c clock MultOutFreqs c MultOutFreqs c numfact clock end if MultOutFreqs c gt clock 2 then MultOutFreqs c clock MultOutFreqs c end ter while MultOutFreqs c 0 1 do inc c diffrec a MultOutFregs c end Qsort diffrec 0 7 MultOut diffrec 0 end I check the performance of a set of transmitter frequencies procedure TPBILMain TestMultClick Sender TObject var input array 1 8 of double res double begin threshold strtofloat MagThresh Text resi strtofloat FiltRes1 Text strtofloat FiltCap1 Text res2 strtofloat FiltRes2 Text cap strtofloat FiltCap2 Te
245. end BotAirPeak Caption floattostrf maxvalue ffFixed 4 3 if stopmax lt gt 0 then airbotmax floor startmax 0 5 stopmax startmax else airbotmax startmax maxvalue 0 startmax 0 stopmax 0 for testcase 1 to RealMain DesiredFrames Value do begin if desiredvolume testcase 3 gt maxvalue then begin maxvalue desiredvolume testcase 3 startmax testcase stopmax 0 end else if desiredvolume testcase 3 maxvalue then stopmax testcase if desiredvolume testcase 3 gt 0 and airbotstart 0 then airbotstart testcase else if desiredvolume testcase 3 gt 0 and airbotstart lt gt 0 then airbotstop testcase end if maxvalue 0 then airbotdesiremax 1 else if stopmax lt gt 0 then airbotdesiremax floor startmax 0 5 stopmax startmax else airbotdesiremax startmax 11 gravel volume fraction for bottom measuring ring startmax 0 stopmax 0 maxvalue 0 for testcase 1 to RealMain DesiredFrames Value do begin if totaloutputs 2 then begin if outputstore testcase prediction 4 gt maxvalue then begin maxvalue outputstore testcase prediction 4 startmax testcase stopmax 0 end else if outputstore testcase prediction 4 maxvalue then stopmax testcase end else begin if volumestore testcase 4 gt maxvalue then begin maxvalue volumestore testcase 4 startmax testcase stopmax 0 end else if volumestore testcase 4 maxv
246. ender TObject Shift TShiftState X Y Integer virtual procedure Paint override public Public declarations constructor Create AOwner TComponent override create tomo component procedure ResetDesired reset all the pixels to water property Desired TOutput read FDesired write FDesired published Published declarations property specifying whether the data can be changed or not property AllowEdit Boolean read FAllowEdit write FAllowEdit default true property specifying the percentage water in the pipeline property WaterVolume Integer read FWaterVolume write FWaterVolume default 100 property specifying the percentage gravel in the pipeline property GravelVolume Integer read FGravelVolume write FGravelVolume default 0 property specifying the percentage air in the pipeline property AirVolume Integer read FAirVolume write FAirVolume default 0 current X and Y positions of the cursor property XPosition Integer read FXPosition property YPosition Integer read FYPosition property OnPaint end procedure Register implementation register the component and place the icon on the samples palette procedure Register begin RegisterComponents Samples TTomo end TTomo I create a paintbox of specified size and initialise the volume fractions for 100 water constructor TTomo Create AOwner TComponent begin inherited Width 200 Height 200 FWaterVolume
247. ent descent SetLength OldOutputWeight numhidden 1 RecLoadDBase GetNumOutputs SetLength OldHiddenWeight 129 numhidden end end else begin SetLength OutputWeight 129 RecLoadDBase GetNumOutputs SetLength DiffOutputWeight 129 RecLoadDBase GetNumOutputs if RecUseRPRop Checked then begin allocate the memory for the matrices required by RPROP SetLength OutputDeltaWeight 129 RecLoadDBase GetNumOutputs SetLength OldDiffOutputWeight 129 RecLoadDBase GetNumOutputs end else begin allocate the memory for the matrix required by gradient descent SetLength OldOutputWeight 129 RecLoadDBase GetNumOutputs end end RecMain Loadnetworkweights Enabled true RecMain Savenetworkweights Enabled true RecMain Initialisenetworkweights1 Enabled true end when show form load the previously used network training parameters from the registry and initialise the important variables procedure TRecSetParam FormShow Sender TObject var Reg TRegistry KeyGood Boolean begin load the previously used parameters fom the registry Reg TRegistry Create try KeyGood Reg OpenKey SoftwareWeural False if the key exists then read from the key else it is the first time the program is running so do not attempt I to read from the registry if KeyGood then begin RecSetMaxEpochs Text inttostr Reg ReadInteger MaxEpochs RecSetNumHidden Value Reg ReadInteger NumHidden RecPerformEarlyS
248. eps per revolution drive resolution convert the number of revolutions into the number of steps airrevs strtofloat AirDistance Text aircircumference gravelrevs strtofloat GravelDistance Text gravelcircumference airsteps round 400 airrevs gravelsteps round 400 gravelrevs tempaccel 50 tempaccel2 50 tempaccel3 50 airacceleration strtoint AirAccel Text gravacceleration strtoint GravelAccel Text case AirDrive Value of 1 begin tempaccel1 airacceleration end 2 begin tempaccel2 airacceleration end 3 begin tempaccel3 airacceleration end end case GravelDrive Value of 1 begin tempaccel1 gravacceleration end 2 begin tempaccel2 gravacceleration end 3 begin tempaccel3 gravacceleration end end RealMain Stepper SetAcceleration tempaccel1 tempaccel2 tempaccel3 Close end when show form load the previous settings from the registry procedure TStepperConfig FormShow Sender TObject var Reg TRegistry KeyGood Boolean begin Reg TRegistry Create try KeyGood Reg OpenKey SoftwareWeural False if KeyGood then AirDiameter Text Reg ReadString AirPulley GravelDiameter Text Reg ReadString GravelPulley AirVelocity Text Reg ReadString AirVelocity GravelVelocity Text Reg ReadString GravelVelocity AirDistance Text Reg ReadString AirDistance GravelDistance Text Reg ReadString GravelDistance
249. er PBILStartClick Sender if besteverfit gt overbestfreq then begin overbestfreq besteverfit OverFreq1 Caption Freq1 Caption OverFreq2 Caption Freq2 Caption OverFreq3 Caption Freq3 Caption OverFreq4 Caption Freq4 Caption OverFreq5 Caption Freq5 Caption OverFreq6 Caption Freq6 Caption OverFreq7 Caption Freq7 Caption OverFreq8 Caption Freq8 Caption OverBestFreqFit Caption BestFreqFit Caption end resrecord 1 currentiter besteverfit for i 2 to 9 do resrecord i currentiter besteverreal i 1 filestring c multi inttostr currentiter txt AssignFile multirecord filestring ReWrite multirecord WriteLn multirecord Best fit floattostr resrecord 1 currentiter WriteLn multirecord Frequencies for best fit for i 2 to 9 do WriteLn multirecord floattostr resrecord i currentiter CloseFile multirecord end if SaveMultiRecord Execute then begin AssignFile multirecord SaveMultiRecord FileName ReWrite multirecord for currentiter 1 to Numlterations Value do begin WriteLn multirecord Best fit floattostr resrecord 1 currentiter WriteLn multirecord Frequencies for best fit for i 2 to 9 do WriteLn multirecord floattostr resrecordli currentiter WriteLn multirecord end CloseFile multirecord end 141 APPENDIX SAMPLER PROGRAM CODE PROGRAM LISTING OF MAINUNIT PAS This program is used to capture t
250. er 0 5 airdiam 2 1 3 pi power tempheight 3 pi lowersystop lowersysbot tempheight power 0 5 airdiam 2 end as only the cylindrical portion of the bubble remains in the measuring ring else if airtopposition gt lowersystop 0 5 airdiam and airtopposition airleng 0 5 airdiam lt lowersysbot then begin owerairvol pi lowersystop lowersysbot power 0 5 airdiam 2 nd as round at bottom of bubble enters measuring section se if airtopposition airleng 0 5 airdiam gt lowersysbot nd airtopposition airleng 0 5 airdiam lt lowersysbot 0 5 airdiam then egin tempheight airtopposition airleng 0 5 airdiam lowersysbot oweraitvol pi tempheight power 0 5 airdiam 2 1 3 pi power tempheight 3 pi lowersystop lowersysbot tempheight power 0 5 airdiam 2 end as round at bottom of the bubble is fully inside the measuring section else if airtopposition airleng 0 5 airdiam gt lowersysbot 0 5 airdiam and airtopposition airleng 0 5 airdiam lt lowersystop then o on 194 begin tempheight lowersystop airtopposition airleng 0 5 airdiam owerairvol airhemivol pi tempheight power 0 5 airdiam 2 nd as round at bottom of the bubble exits the measuring section se if airtopposition airleng 0 5 airdiam gt lowersystop nd airtopposition airleng lt lowersystop then egin tempheight airtopposition airleng 0 5 airdiam lowersystop owerairvol airhemivol pi tempheight power 0 5 a
251. er increments and sets the multiplexer On the positive edge the PC30G card starts the next conversion 16 muliplexer outputs SYNCHRONOUS DETECTOR BOARDS DG506 MUX8 FIGURE 5 17 BLOCK DIAGRAM OF SAMPLE CONTROLLER BOARD A SINGLE SAMPLE CONTROLLER BOARD IS USED TO CAPTURE DATA FROM TWO 16 ELECTRODE FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEMS A 4 BIT BINARY COUNTER IS USED TO SET THE MULTIPLEXER ADDRESS LINES AND THE TIMING IS CONTROLLED BY A 555 OSCILLATOR Central to the sample controller board is a 4 bit binary counter that is used to drive the address lines of all the multiplexers in parallel A CMOS 555 oscillator provides the clock signal to both the 4 bit counter and the external trigger input of the PC30G card A timing diagram of the 555 output is also included in figure 5 17 and the operation can be explained as follows on the negative transition of the 555 output the 4 bit binary counter is incremented and the multiplexers connect the next voltage inputs in the sequence to the analogue input channels of the PC30G card based on the address lines The 555 output is low for a period of 20us while the multiplexer outputs stabilise Specifically the maximum switching time of a DG506A CMOS analogue multiplexer is 1us On the positive transition of the 555 output the analogue to digital converter on the PC30G card is triggered and the process of sampling the 16 analogue inputs is started The 55
252. er of frames 1400 0 20 400 600 800 1000 1200 separation os Gravel volume fraction 0 0 Flowmeter Factor Air Gravel Water Top air threshold Bottom system Top gravel threshold Bottom air threshold 100 Bottom gravel threshold Correlation threshold 001 Til 0 200 400 600 800 1 000 1 200 Air correlation result m Load network data Calibrate system 1 Stepper motor setup Start motors Start testing ud uos 0 200 400 600 s00 1 000 1 200 Frame number 1400 3 Sampling period 0 E 50 ravel correlation result Sample time Number of frames captured 1400 0 min 5s Oms Calculated frame rate 277 227723 Set the desired frame rate Velocity predictions For offline viewing fi Desired air velocity m s 0 200 ia Desired gravel velocity m s 0 dee Verify results Predicted air velocity m s 0 207 i i 7 Cl Predicted gravel velocity m s unknown 0 200 400 600 800 1000 1200 ose FIGURE 7 5 SCREEN CAPTURE OF THE REAL TIME SAMPLING FORM THE FORM CONSISTS OF THREE SECTIONS ON THE FAR RIGHT IS THE SECTION WHERE THE SOFTWARE IS CONFIGURED BEFORE A TEST THIS CONSISTS OF SPECIFYING THE NEURAL NETWORKS TO PERFORM THE RECONSTRUCTION THE SYSTEM SEPARATION THE FLOWMETER FACTOR AND THE THRESHOLD VALUES ON THE FAR LEFT IS WHERE THE ON LINE MEASUREMENTS OF THE VOLUME FRACTIONS AT THE TOP AND BOTTOM MEASUREMENT SYSTEMS ARE DISPLAYED IN THE CENTRE ARE THE AIR AND GRAVEL VOLUME FR
253. erStopClick Sender TObject procedure RealStartMotorsClick Sender TObject procedure VerifyResultsClick Sender TObject private Private declarations notifvalue cardinal II Windows notification message value that is issued once streaming is complete curcalibrate Boolean topcalib array 1 6400 of real calibration data for top bottomcalib array 1 6400 of real and bottom systems topcalibvalue array 1 128 of real bottomcalibvalue array 1 128 of real bottommean array 1 128 of real pre processing data loaded bottomstddev array 1 128 of real from a file for the top topmean array 1 128 of real and bottom neural network topstddev array 1 128 of real topweights array of array of real matrices containing neural tophiddenweights array of array of real network weights for the top tophiddenoutputs array of real and bottom system both hidden bottomhiddenoutputs array of real layer and output layer bottomweights array of array of real bottomhiddenweights array of array of real totalfrms integer total number of frames to be captured hours mins secs msecs word used to calculate the total time taken to capture and reconstruct the specified number of frames storeposition integer position in matrix where data must be placed viewposition integer II position in matrix where data
254. eriod A drift analysis of the FDM impedance tomography system revealed considerable drift of both the capacitance and conductance readings Figure 6 1 illustrates the drift of both capacitance and conductance voltages over a 15 hour period Tests revealed that the behaviour of the drift was independent of the particular receiver analysed As is evident the drift of the capacitance between receiver A and transmitter A is very similar to the drift of the conductance between receiver A and transmitter A This also extends to completely different receivers For example the drift of the capacitance between receiver A and transmitter A is very similar to the drift of the capacitance between receiver B and transmitter A From this it was concluded that the drift was related to the transmitter outputs and not the receivers The average capacitance voltage drift was 121mV and the average conductance voltage drift was 167mV over a 15 hour period TRANSMITTER A TRANSMITTER C capacitance V capacitance V 5 10 time hours time hours TRANSMITTER A TRANSMITTER conductance V conductance V 5 10 5 10 time hours time hours FIGURE 6 1 DRIFT ANALYSIS OF BOTH CAPACITANCE AND CONDUCTANCE READINGS FOR TRANSMITTER A RECEIVER A AND TRANSMITTER C RECEIVER A ELECTRODE PAIRS OVER A 15 HOUR PERIOD 67 Standard conductance measurement systems use the four electrode adjacent pair measurement pro
255. es contair Boolean contgrav Boolean get and set phases end var RecLoadDBase TRecLoadDBase meanrec array 1 128 of real store the means of each column so that current samples be pre processed according to the same rule used for the training data stddevrec array 1 128 of real store the standard deviations of each column custload Boolean trainstartindex array 1 5 of integer trainstopindex array 1 5 of integer teststartindex array 1 5 of integer teststopindex array 1 5 of integer numtrainregions integer numtestregions integer implementation uses mainunit calibset R DFM MM DATA LOADING PROCEDURES this procedure loads the data in the specified databases into the matrices that will be used for network training and network testing To access the data temporary tables are created for the databases and opened NOTE the data is loaded as follows FOR THREE PHASE IMAGE RECONSTRUCTION O0 water O1 gravel 02 air FOR VOLUME PREDICTION O0 air 01 gravel FOR VOLUME DATA O0 air 01 gravel 02 water FOR TWO PHASE IMAGE RECONSTRUCTION 1 water target 1 other phase Specify the table sizes as if all the data were to be loaded Then depending on whether a test case contains air gravel or mixture load the data into the table This will allow the standard calibration feature to correctly calculate the offset for two phase networks w
256. estsum real sum void error corresponding to the best network bestwater real water volume fraction corresponding bestgravel real to the best network bestair real performingtrain Boolean procedure TrainDoubleNetwork train double layer network procedure TrainSingleNetwork train single layer network procedure AdjustRPRop adjust network weights using RPROP procedure AdjustWeights adjust network weights using gradient decsent procedure Calculate TestOutput start integer stop integer calculate the network outputs for the test data as well as the validation error percentages procedure Calculate TrainError start integer stop integer calculate the training error percentages function Sign arg extended extended return sign of arg procedure UpdateNetworkDisplay update display during training procedure StoreBestWeights keep record of best network weights II so far public Public declarations function SoftMax numout integer wat double grav double air double double calculate the soft max activation function output function CalcDeriv a extended extended calculate the derivative of the hyperbolic tangent function GetAirError real data access functions function GetWaterError real function GetGravelError real function GetSumVoidError real function GetThresholdError real function GetNumBest integer end var RecTrainNetwork TRecTr
257. et Click Sender TObject procedure SavenetworkweightsClick Sender TObject procedure LoadnetworkweightsClick Sender TObject procedure Validatedatabase1Click Sender TObject procedure Validatecurrentsamples1Click Sender TObject procedure FormCreate Sender TObject procedure Realtimetesting1Click Sender TObject procedure Savenetworkpreprocessing1 Click Sender TObject procedure Savecalibrationdata1 Click Sender TObject PROGRAM LISTING OF MAINUNIT PAS This program will be used to train and test neural networks for the image reconstruction and volume fraction prediction of a static vessel situation The neural networks provided are the single layer feed forward multi layer perceptron and the double layer feed forward multi layer perceptron These networks can be trained using gradient descent or Resilient back propagation these network training algorithms have been taken from the previous network training program private AUR Private declarations The main unit provides the general structure to the program through a menu It A also performs functions such as loading and saving the network weights saving weightfactor real I initi i a pM y d md the network performance and initialising the network weights to random values 7 Inmate MUNDSI All data matrices are allocated dynamically once the network training parameters Bb have been specified this saves memory st d ic declarations
258. eter Text Reg WriteString AirLength AirLength Text Reg WriteString AirBubbleStartPosition AirBubbleStartPosition Text Reg WriteString GravelDiameter GravelDiameter Text Reg WriteString GravelLength GravelLength Text Reg WriteString GravelBubbleStartPosition GravelBubbleStartPosition Text Reg Writelnteger AirPosition airposition Reg Writelnteger GravelPosition gravelposition Reg WriteString UpperSystTop UpperSystemTop Text Reg WriteString UpperSystBot UpperSystemBottom Text Reg WriteString LowerSystTop LowerSystemTop Text Reg WriteString LowerSystBot LowerSystemBottom Text Reg Writelnteger VelSource VelSource ItemIndex finally Reg Free end Close end when form opens initialise variables and load real time volume fraction data into charts procedure TDynVerify FormShow Sender TObject var testindex integer begin if StepperConfig airrps gt 0 then AirPhaseConfig Enabled true else begin airposition 0 AirPhaseConfig Enabled false end if StepperConfig gravelrps gt 0 then GravPhaseConfig Enabled true else begin gravelposition 0 GravPhaseConfig Enabled false end ErrorResults Visible false UpdateDisplay TopAirPlot Series 0 Clear TopAirPlot Series 1 Clear TopAirPlot Series 2 Clear TopGravelPlot Series 0 Clear TopGravelPlot Series 1 Clear TopGravelPlot Series 2 Clear BotAirPlot Series 0 Clear BotAir
259. eural networks trained on static data Consequently the bubble placement system was replicated for the dynamic simulations It consists primarily of a positioning base which is included as a detail in figure 7 2 The positioning base is placed in the bottom section of the laboratory scale rig and the alignment strips in the bottom section are used to align the positioning base with the measurement sections 20mm holes are drilled at the corners of each pixel as can be seen in the figure detail Before a test is conducted the 20mm PVC pipe used to support the pulleys is inserted into one of these holes Subsequently the position of the bubble on the drive belt corresponds to one of the standard bubble positions used during network training e Ideally the apparatus is required to simulate flow velocities up to 20m s Since the maximum stepper motor speed is 50rev s as set by the stepper driver this places a constraint on the minimum pulley diameter In addition the pulley diameter is required to be a multiple of 35mm in order to ensure that the position of the bubble corresponds to a standard bubble position Hence a pulley diameter of 140mm was chosen corresponding to a maximum flow velocity of 22m s A 3mm o ring drive belt was used and the pulleys were machined out of polyacetal plastic In addition the pulley mounting blocks were machined from polyacetal plastic and were specifically designed to ensure that the spacing between the PVC pipe and the
260. evious research The concept of the new data acquisition system will be explained in Chapter 4 In addition the static tomography rig used in the previous project research will be modified to incorporate this new measurement technique A comparison will be made between the results of this modified system and the original system Chapter 5 will discuss the construction of the laboratory scale dual plane impedance tomography flowmeter Static reconstruction results will be provided in Chapter 6 to verify the operation of the new system Chapter 7 will examine the performance of this laboratory scale rig in simulated flow situations Finally conclusions will be drawn and recommendations made regarding future project development Chapter 10 will complete the thesis by briefly examining the remaining issues of the current system that would need to be addressed before an industrial prototype is developed CHAPTER 2 TI ME DIVISION MULTI PLEXED MPEDANCE TOMOGRAPHY This chapter is a brief description of conventional capacitance and conductance tomography data acquisition Systems It will serve as a reference when the operation of the new data acquisition technique is explored in Chapter 4 21 Standard Capacitance Tomography Measurement Systems Capacitance tomography systems reconstruct the dielectric distribution of the pipeline cross section To achieve this a number of capacitance plates are mounted around the measurement volume and the capacitances of a
261. ewVolume SeriesList 2 YValue 0 VolumeData currentindex 1 end else begin for volume fraction prediction use a chart to display the desired volume fractions if OutputData currentindex 0 lt gt 1 then begin RecViewVolume SeriesList 0 Y Value 0 100 OutputData currentindex 0 OutputData currentindex 1 RecViewVolume SeriesList 1 YValue 0 OutputData currentindex 0 RecViewVolume SeriesList 2 YValue 0 OutputData currentindex 1 end else begin RecViewVolume SeriesList 0 Y Value 0 0 RecViewVolume SeriesList 1 YValue 0 0 RecViewVolume SeriesList 2 Y Value 0 0 end end UpdateTables end display the input voltages in a tabular format procedure TRecViewMain UpdateTables var row col index overall integer tresult TResultTable 51 52 string begin index 1 for overall 1 to 16 do begin case overall of 1 begin row 4 tresult capacitance end 2 begin row 4 tresult resistance end 3 begin row 3 tresult capacitance end 4 begin row 3 tresult resistance end 5 begin row 2 tresult capacitance end 6 begin row 2 tresult resistance end 7 begin row 1 tresult capacitance end 8 begin row 1 tresult resistance end 9 begin row 5 tresult resistance end 10 begin row 5 tresult capacitance end 11 begin row 6 tresult resistance end 12 begin r
262. ex 1 else begin weightsum 0 for rowindex 0 to 128 do begin if rowindex 0 then tempvalue 1 else tempvalue inputstore reconposition readings rowindex weightsum weightsum tophiddenweights rowindex colindex 1 tempvalue end prevent arithmetic overflow if weightsum gt 200 then weightsum 200 else if weightsum lt 200 then weightsum 200 181 tanh activation function tophiddenoutputs colindex tanh weightsum end end end once calculated hidden layer outputs calculate final network 11 prediction for colindex 0 to totaloutputs 1 do begin weightsum 0 for rowindex 0 to upperindex do begin if numtophidden 0 then begin if rowindex 0 then tempvalue 1 else tempvalue inputstore reconposition readings rowindex end else tempvalue tophiddenoutputs rowindex weightsum weightsum topweights rowindex colindex tempvalue end if weightsum gt 200 then weightsum 200 else if weightsum lt 200 then weightsum 200 I three phase image reconstruction if totaloutputs 264 then begin if colindex 1 mod 3 lt gt 0 then weightstore colindex 1 mod 3 weightsum else begin for i 2 downto 0 do outputstore reconposition prediction colindex i 1 RecTrainNetwork SoftMax 3 i weightstore 1 weightstore 2 weightsum end end two phase image reconstruction else if totaloutputs 88 then begin outputstore reconposit
263. exit the program procedure TSampMain Exitt Click Sender TObject begin Close end if the user changes his mind procedure TSampMain SampAliasNewCancClick Sender TObject begin SampAliasNewName Text SampAliasNew Visible false end procedure TSampMain SampAliasOpenCancClick Sender TObject begin SampAliasOpen Visible false end to start sampling initialise the sample hardware and display AddToDatabase modally procedure TSampMain Startsampling1 Click Sender TObject begin store the notification value for the sample unit so that can capture Windows notification message indicating that sampling is complete notifvalue SampHardware InitSample 1 AddToDatabase Handle AddToDatabase Caption Add sampled data to database allow editing of data AddToDatabase AddInputTomo AllowEdit true AddToDatabase AddSampleStart Visible true AddToDatabase AddReset Visible true AddToDatabase AddUserSetVolume Visible true AddToDatabase ShowModal end when start the program read from the registry to determine which was the last database to be used and how many frames were taken for each test case Load the last database to be used automatically procedure TSampMain FormCreate Sender TObject var Reg TRegistry KeyGood Boolean begin SampHardware TSamp Create Reg TRegistry Create try KeyGood Reg OpenKey Software Sampler False II if the key exists
264. f network output less than 0 then predicts that pixel is water if OutputNetwork rowindex colindex gt 0 then begin if OutputData rowindex colindex lt gt 1 then inc totalerrors if containsair then totaltrainairpixels totaltrainairpixels 1 end calculate the threshold and sum void errors for the entire testing database currthresh 100 totalerrors 88 stop start 1 currsum watererror gravelerror airerror stop start 1 watererror watererror stop start 1 gravelerror gravelerror stop start 1 airerror airerror stop start 1 end calculate the threshold and sum void error percentages for the training database procedure TRecTrainNetwork CalculateTrainError start stop integer else totaltraingravelpixels totaltraingravelpixels 1 end else begin if OutputData rowindex colindex lt gt 1 then inc totalerrors end inc colindex end volume fraction prediction 173 else begin if colindex 0 then totaltrainairpixels OutputNetwork rowindex colindex else totaltraingravelpixels OutputNetwork rowindex colindex inc colindex end end calculate the volume fraction predictions for this test and add to the previous error percentages totaltrainwaterpixels 100 totaltrainairpixels totaltraingravelpixels watererror watererror Abs VolumeData rowindex 2 totaltrainwaterpixels gravelerror gravelerror Abs VolumeData rowin
265. f the square wave harmonics mixing in the multiplication stage of the synchronous detector Unlike the 8 electrode prototype the frequency components generated for the 16 electrode system are within the passband of the system and therefore cannot be removed simply through the use of additional low pass filtering It was therefore necessary to perform a thorough analysis of these frequency components in an attempt to eliminate their generation instead of simply filtering them 46 The theoretical analysis of the FDM impedance tomography system in Chapter 4 was for sine wave signals Each transmitter only outputs a single frequency component The output of the multiplication stage would then contain the sum and difference of these fundamental components and consequently it is a trivial task to select appropriate transmitter frequencies to ensure that no output ripple exists within the passband of the system However this is not the case for a square wave FDM impedance tomography system Specifically each transmitter output consists of a fundamental frequency component with a number of frequency components at the odd harmonics of the fundamental frequency In addition if the output square waves are not perfect then minor frequency components also exist at the even harmonics of the fundamental frequency Subsequently the receiver signals consist of a considerably larger number of frequency components in a square wave system than a sine wave system Since th
266. fer mode to streaming I and the minimum block size to 1024 function TSamp InitSample bdnum integer Handle cardinal cardinal var message_val cardinal i integer frequency integer begin frequency 6000 boardnum bdnum get the message number that is sent to the notification window message val EDR_GetNotificationMsg set the window that will receive the notification messages for the board SetNotificationWindow boardnum Handle numframes 0 configure port A as output port in simple DIO mode EDR_DlOConfigurePort boardnum 0 EDR_DIO_SIMPLE 0 reset 555 and counter on sample control card EDR DlOPortOutput boardnum 0 rst555 EDR DlOPortOutput boardnum O rstclk load the channel list with channel 0 to 15 EDR SetADChanListLen boardnum 0 for i 0 to 15 do begin AddToADChanList boardnum i EDR SetADInGain boardnum i 1 end SetADClockmilliHz boardnum frequency 1000 set the transfer mode of the AD data as streaming SetADTransferMode boardnum EDR STREAM EDR SetADBurstLen boardnum EDR ADBURSTNUMCHANS set the size of the block written out by the streamer SetStreamBlockSize boardnum 1024 InitSample message val busy sampling false end release the resources allocated for sampling procedure TSamp ReleaseSample begin EDR DlOPortOutput boardnum 0 0 EDR SetNotificationWindow boardnum 0 StopBackgroundADIn b
267. flow and in particular the calculation of the individual component mass flow rates within that flow Consequently it was necessary to assess the volume fraction prediction accuracy of the impedance tomography System as well as the accuracy of the component velocities measured using the cross correlation of a dual plane System The specific instructions were to 1 Verify the performance of an impedance tomography system on a laboratory scale rig for static configurations of a multi phase air gravel seawater mixture Construct two complete impedance tomography systems and interface them for flow measurement Develop the neural network software that accompanies the laboratory scale rig for the volume fraction prediction of the phases within the mixture Included in the software must be the calculation of the individual component velocities so that the mass flow rates of the phases within the mixture can be determined 4 Perform a thorough characterisation of the dual plane impedance tomography system Demonstrate the operation of the flow measurement system Draw conclusions regarding the application of a dual plane impedance tomography system to the on line monitoring of an air gravel seawater mixture 7 Make recommendations for future project development and to discuss any remaining issues that would need to be addressed before an industrial prototype was developed 8 Submit the results of the research and all supporting documentation by 30 June
268. for that particular pixel namely where pixel e fair gravel seawater 3 6 By combining these neural network outputs in parallel an image of the contents of the vessel is obtained If the algorithm proceeds to sum the air gravel and seawater pixels in the image then an estimate of the individual component volume fractions within the vessel cross section can be obtained 11 Figure 3 1 illustrates a two phase image reconstruction using a single layer feed forward neural network The network consists of 56 input neurons corresponding to the 28 capacitance and 28 conductance readings obtained from the data acquisition system respectively An additional input neuron is provided labelled offset whose output is equal to unity These input neurons are fully connected to the output layer of neurons Although the image contains 100 pixels only 88 of these are situated within the cross section of the pipeline hence there are 88 output neurons Hyperbolic tangent activation functions are employed in the output layer If the neuron output is less than zero then the corresponding pixel contains seawater else it contains the second phase which is air for the example illustrated OFFSET IF OUTPUT lt 0 THEN SEAWATER Wn ELSE SECOND PHASE AIR OR GRAVEL ACTIVATION FUNCTION VOLTAGE READINGS PROPORTIONAL 2 AND IGHT CONDUCTANCE WEIGHTS z Os ti aa go INPUT LAYER NEURON
269. frequency of the low pass filter to the clock frequency in the current configuration is 60 f 1 1 Y fipp CLOCK 14 1 1 LPF 100 xs xs 5 2 where Q is the quality factor of the second order filter and equals unity in the current configuration A multi turn potentiometer on the input to the LMF100 controls the gain of the switched capacitor filter and provides a means of achieving the maximum measurement range for each individual transmitter receiver electrode pair One drawback of switched capacitor filters is that they are effectively a sampled data system where the sampling rate is set by the clock input Subsequently the Nyquist sampling theorem must be observed and frequency components greater than one half the clock frequency must be removed before the LMF100 To achieve this the first order op amp low pass filters were retained The cutoff frequencies of these low pass filters were set at 194Hz These op amps also provided an additional gain control stage The gains for the op amps in the capacitance paths were set at 2 5 The gains for the op amps in the conductance paths were also set at 2 5 except for those corresponding to adjacent transmitter receiver electrode pairs where the gain was reduced to 1 2 Another drawback of switched capacitor filters is that their outputs can be noisy In particular clock feedthrough results in the clock appearing as a small ripple on the output voltages of the switched capacitor fil
270. g WriteBool ShowTrainPerf RecShowTrainPerf Checked Reg Writelnteger UpdateFreq RecSetUpdateFreq Value finally Reg Free end get the required values from the form deltainit strtofloat RecSetDeltalnit Text deltamax strtofloat RecSetDeltaMax Text learnrate strtofloat RecSetLearn Text momentum strtofloat RecSetMomentum Text numfails RecNumFails Value if RecPerformModel Checked then numhidden RecModelStop Value else numhidden RecSetNumHidden Value maxepochs strtoint RecSetMaxEpochs Text dynamically allocate the memory for the different weight matrices now that the number of hidden nodes and the network training procedure has been specified if numhidden gt 0 then begin SetLength OutputWeight numhidden 1 RecLoadDBase GetNumOutputs SetLength HiddenWeight 129 numhidden SetLength OutputHidden RecLoadDBase GetTotalTrain RecLoadDBase GetTotalTest numhidden 1 SetLength DiffHiddenWeight 129 numhidden SetLength DiffOutputWeight numhidden 1 RecLoadDBase GetNumOutputs if RecUseRProp Checked then begin allocate the memory for the matrices required by RPROP SetLength OutputDeltaWeight numhidden 1 RecLoadDBase GetNumOutputs SetLength OldDiffOutputWeight numhidden 1 RecLoadDBase GetNumOutputs SetLength HiddenDeltaWeight 129 numhidden SetLength OldDiffHiddenWeight 129 numhidden 166 end else begin allocate the memory for the matrices required by gradi
271. gLabel Caption Desired output for test inttostr currentindex 1 RecLoadDBase GetPhases contair contgrav RecViewInputTomo ResetDesired for col 0 to 9 do begin for row 0 to 9 do begin tempout col row outpipe end end index 0 for col 0 to 9 do begin case col of 0 begin top 2 bottom 7 end 1 begin top 1 bottom 8 end 2 7 begin top 0 bottom 9 end 8 begin top 1 bottom 8 end 9 begin top 2 bottom 7 end end for row top to bottom do begin if contair and contgrav then begin if three phase then decode 1 of C encoding employed if OutputData currentindex index 1 then tempout col row water else if OutputData currentindex index 1 1 then tempout col row gravel else tempout col row air inc index 3 end else begin else simply two phase if OutputData currentindex index 1 then begin if contair then tempout col row air else tempout col row gravel end else if OutputData currentindex index 1 then tempout col row water else tempout col row outpipe inc index end end end set the TTomo component and force a repaint RecViewInputTomo Desired tempout RecViewInputTomo Repaint display the volume fraction data using the VolumeData matrix RecViewVolume SeriesList 0 YValue 0 VolumeData currentindex 2 RecViewVolume SeriesList 1 YValue 0 VolumeData currentindex 0 RecVi
272. ger begin gravelposition 5 UpdateDisplay end reset the graphical interface procedure TDynVerify ResetColours begin AirShapet Brush Color clBlack AirShape2 Brush Color clBlack AirShape3 Brush Color clBlack AirShape4 Brush Color clBlack AirShape5 Brush Color clBlack GravShapet Brush Color clBlack GravShape2 Brush Color clBlack GravShape3 Brush Color clBlack GravShape4 Brush Color clBlack GravShape5 Brush Color clBlack end update the graphical interface indicating bubble position procedure TDynVerify UpdateDisplay begin ResetColours case airposition of 0 begin end 1 begin AirShape1 Brush Color clWhite 2 begin AirShape2 Brush Color clWhite end 3 begin AirShape3 Brush Color clWhite end 4 begin AirShape4 Brush Color clWhite 5 begin AirShape5 Brush Color clWhite end end case gravelposition of 0 begin end 1 begin GravShape1 Brush Color clGray 2 begin GravShape2 Brush Color clGray end 3 begin GravShape3 Brush Color clGray end 4 begin GravShape4 Brush Color clGray end begin GravShape5 Brush Color clGray end end end when close form save the current settings to the registry procedure TDynVerify RealExitClick Sender TObject var Reg TRegistry begin Reg TRegistry Create try Reg OpenKey Software Neural True Reg WriteString AirDiameter AirDiam
273. get a list of all available database names Session GetAliasNames templist 161 for i 0 to templist Count 1 do begin RecTrainName ltems Add templist i RecTestName ltems Add templist i end templist Free opened false I load the previously used settings from the registry Reg TRegistry Create try KeyGood Reg OpenKey SoftwareWeural False II if the key exists then read from the key else it is the first time the program is running so do not attempt I to read from the registry if KeyGood then begin RecDesiredRecon lItemIndex Reg Readinteger OutputDesired RecTrainName Text Reg ReadString TrainDatabase RecTestName Text Reg ReadString TestDatabase RecTrainStart Value Reg Readinteger TrainStart RecTrainStop Value Reg ReadInteger TrainStop RecTestStart Value Reg ReadInteger TestStart RecTestStop Value Reg Readlnteger TestStop RecLoadCalibrate Checked Reg ReadBool Calibrate RecTrainVers Value Reg ReadInteger TrainVers RecTestVers Value Reg ReadInteger TestVers RecDesirePhase temIndex Reg ReadInteger DesiredLoadPhase end finally Reg Free end end when close form write the current parameters to the registry If the data has been loaded then free the resources allocated to the temporary tables procedure TRecLoadDBase FormClose Sender TObject var Action TCloseAction var Reg TRegistry begin Reg TR
274. gle layer feed forward neural network Performance of the 16 electrode frequency division multiplexed impedance tomography system for a three phase air gravel water volume fraction prediction using a single layer feed forward neural network Comparison between the performance of the 16 electrode frequency division multiplexed impedance tomography system and the 8 electrode prototype frequency division multiplexed impedance tomography system for a three phase air gravel water volume fraction prediction using a double layer feed forward neural network Comparison between the average peak to peak output ripple of the capacitance and conductance readings for the different configurations tested Comparison between the average dynamic volume fraction errors for the 60mm 90mm and 190mm bubble respectively Comparison between the average dynamic volume fraction errors of the different neural network reconstruction techniques considered for a three phase air gravel water reconstruction 31 32 32 52 55 69 73 74 75 77 77 92 96 103 Xii CHAPTER 1 I NTRODUCTI ON This thesis investigates the use of a dual plane impedance tomography system to calculate the mass flow rate of an air gravel seawater mixture The long term goal of this project is to develop an instrument for the monitoring of a multi phase airlift used in an offshore mining application This instrument is to be placed in line and will be required to calcul
275. gravel inc index 3 end else begin if outputstore testposition prediction index 88 gt 0 then begin if containsair then bottomphasepred air else bottomphasepred gravel end else bottomphasepred water inc index 1 end bottomoutputpred col row bottomphasepred if bottomphasepred air then volumestore testposition 3 volumestore testposition 3 1 else if bottomphasepred gravel then volumestore testposition 4 volumestore testposition 4 1 end end if volumestore testposition 1 lt topairthreshold then volumestore testposition 1 0 if volumestore testposition 2 lt topgravthreshold then volumestore testposition 2 0 if volumestore testposition 3 lt botairthreshold then volumestore testposition 3 0 if volumestore testposition 4 lt botgravthreshold then volumestore testposition 4 0 if testposition mod 5 0 then begin TopTomo Desired topoutputpred BottomTomo Desired bottomoutputpred TopTomo Repaint BottomTomo Repaint frmnumber Caption inttostr reconposition DecodeTime outputstore reconposition sampletime hours mins secs msecs smins Caption inttostr mins min ssecs Caption inttostr secs smsecs Caption inttostr msecs ms Update end end else begin if testposition mod 5 0 then begin TopAirChart Series 0 Clear 183 TopAirChart Series 0 Addy outputstore testposition prediction 1 clTeeColor TopG
276. gravel or not tempair and not tempgravel or 0 then begin s1 Pixel tinttostr col tempvalue TrainOutTable FieldByName s1 Value if tempvalue 3 then begin if threephase then begin OutputData row imgcol 0 OutputData row imgcol 1 0 OutputData row imgcol 2 0 if tempvalue 0 then OutputData row imgcol 2 1 else if tempvalue 1 then OutputData row imgcol 1 else OutputData row imgcol 1 1 inc imgcol 3 end else begin if tempvalue 1 then OutputData row imgcol 1 else OutputData row imgcol 1 inc imgcol end end end else just set a flag so that this test case can be removed at a later stage else if RecDesirePhase ItemIndex 1 and tempgravel or RecDesirePhase ItemIndex 2 and tempair then begin OutputData row imgcol 0 inc imgcol if imgcol 88 then imgcol 0 end end end else begin if desired output is volume fractions if TrainOutTable FieldByName AirVolume Value gt 0 then tempair true if TrainOutTable FieldByName GravelVolume Value gt 0 then tempgravel true if the phases for the current test case are the correct 11 phases for the desired reconstruction then load the data normally if RecDesirePhase ItemIndex 1 and tempair and not tempgravel Jor RecDesirePhase ItemIndex 2 and not tempair and tempgravel or not tempair and not tempgravel or RecDesirePhas
277. harmonics whose magnitudes after the filtering in the receiver are greater than the minimum for 1 to 100 do for a 0 to 7 do begin calculate the frequency by multiplying a tranmsitter I frequency by an integer factor FreqMag b 1 inputs a c temp1 inputs a c temp2 TC1 temp1 temp3 TC2 temp1 temp4 sqr temp2 temp5 sqr temp3 calculate the attenuation due to filter 1 gaint 1 sqrt 1 temp4 calculate the attenuation due to filter 2 gain2 1 sqrt 1 temp5 the magnitude of a square wave harmonic is given by 11 1 c pi where c is the order of the harmonic the magnitude of the received harmonic is then this magnitude multiplied by the gains of the two filter stages FreqMag b 2 gain1 gain2 1 c pi inc b end e 1 for c 1 to 800 do begin only frequencies larger than the threshold are kept if FreqMag c 2 gt threshold then begin FreqMag c 1 inc e end end sort the received frequencies and check for duplicates if there are any duplicate frequencies then one transmitter is some multiple of another and therefore the minimum frequency component is 0 2 upper e QSort RecvFreq 0 e 2 2 while RecvFreq a lt gt 0 do begin if RecvFreq a RecvFreq a 1 then begin MultOut 0 exit end inc a end for all 8 transmitter frequencies fora 0 to 7 do begin calculate the harmonics of the refer
278. hases a dual modality tomography system is required 7 Dual modality tomography makes use of separate measurement modalities to produce two tomographic images representing the same interrogated area or object 8 Since different modalities are used the resulting complementary images represent two different object property variations over the interrogated area In particular to image a three phase flow requires more than one modality 9 An electrical impedance tomography system will be used to provide both capacitance and conductance information to discriminate between the air gravel and seawater phases The conductance distribution will give a good separation between seawater and the other phases Since the conductance of the gravel and air phases is effectively the same capacitance tomography will provide the distinction between these two phases based on their different dielectric constants Various flowmeters exist using different measurement techniques These systems are typically flow regime dependent Hence assumptions must be made beforehand regarding the flow regime of the system to be monitored since the instruments themselves are not capable of determining local flow information 10 If these assumptions are inaccurate then the calculated flow rates are also inaccurate Tomography provides a way of extracting this local flow information such as volume fraction and component velocity 10 In practice the frame rates of current tomogra
279. he outside portion of the drive belt In order to verify the dynamic performance of the system all bubble simulations must start from a defined reference point The reference starting point for an upward flow is a point marked 115mm below the bottom measurement section This defines position zero and the positions of all other points are calculated in terms of the axial distance from this reference mark Before a test each bubble must be positioned so that the top of the bubble corresponds to the reference mark The jog buttons allow the user to perform this positioning Returning to the Real time sampling form the test is started by clicking Start testing This initiates the stepper motor movement and also the data capture and reconstruction for the dual plane impedance tomography system The on line reconstruction results for the top and bottom system are displayed in the left hand portion of the reconstruction results section Once the test is complete the volume fraction profiles for the air and gravel phases are plotted and the cross correlation performed In these charts the red series is the volume fraction profile of the top system while the blue series is the volume fraction profile of the bottom system As can be seen in figure 7 5 on page 86 the air bubble was first detected at the bottom system before being detected at the top system as would be expected for an upward flow The following is a brief discussion as to how th
280. he diode on the gate of the IRF520 conducts thus rapidly discharging its gate capacitance Subsequently the IRF520 turns off quickly At the same time the diode on the gate of the IRF9530 is reverse biased so the gate capacitance of the IRF9530 discharges slowly Subsequently the IRF9530 turns on slowly thus introducing a dead time between the turning off of the one MOSFET and the turning on of the second MOSFET When the output of the TC4427 goes high the above process is reversed and the IRF9530 turns off quickly while the IRF520 only turns on after a dead time Consequently shoot through is minimised The circuit details of the cross conduction circuitry as well as the push pull power MOSFET output stage are given in Appendix C on page 122 5 2 2 Conductance and capacitance receiver details Figure 5 10 is a block diagram of the capacitance receiver It consists of a transimpedance amplifier followed by various filtering stages The selection of the RC combination in the transimpedance amplifier follows the same design principles as discussed in the design of the 8 electrode prototype As will become evident later it was determined that the high frequency components of the received signals should be attenuated in order to achieve better separation at the synchronous detection stage A first order low pass and bandpass filter with an upper cutoff frequency of 340kHz attenuates these high frequency components This frequency was chosen to ensure re
281. he effect of this distortion is evident when the output waveforms are monitored using a spectrum analyser An ideal square wave only has frequency components at the odd harmonics of the fundamental frequency However these distortions resulted in significant components at the even harmonics of the fundamental frequency as well Since the system is attempting to isolate individual frequency components the generation of additional frequency components is a highly undesirable situation and will be examined in greater detail once the operation of the synchronous detector electronics has been explained It was observed that serious corrosion of the conductance probes would take place when only one transmitter was operating In addition bubbles would form on the conductance probes However the situation improved greatly once all transmitters were operating simultaneously and minimal bubble formation was observed The consequences of these electrochemical reactions between the seawater and the conductance probes will be examined at a later stage when the drift of the capacitance and conductance readings is analysed 26 4 2 3 Conductance and capacitance receiver details Initially the conductance probe and the capacitance plate of the receivers were in electrical contact and a single transimpedance amplifier was used to generate an output waveform whose amplitude and phase were a function of the complex impedance of the material within the rig The indi
282. he output square waves produced are still not ideal and contain minor irregularities In addition switching noise is still evident on the power supply lines 111 LI ST OF REFERENCES 10 11 12 13 14 15 16 17 18 Smit Q 2000 Material phase detection using capacitance tomography MSc Thesis University of Cape Town Smit Q Tapson J and Mortimer B J P 2000 Material phase detection system using capacitance tomography Proc IEEE IMTC Conf Baltimore WA Teague G 2000 Neural network reconstruction for electrical capacitance tomography system BSc Thesis University of Cape Town Teague G Tapson J and Smit Q 2001 Neural network reconstruction for tomography of a gravel air seawater mixture Meas Sci Technol 12 1102 1108 Jaworski A J and Dyakowski T 2001 Application of electrical capacitance tomography for measurement of gas solids flow characteristics in a pneumatic conveying system Meas Sci Technol 12 1109 1119 Gamio J C Yang W Q and Stott A L 2001 Analysis of non ideal characteristics of an ac based capacitance transducer for tomography Meas Sci Technol 12 1076 1082 Reinecke N and Mewes D 1996 Recent developments and industrial research applications of capacitance tomography Meas Sci Technol 7 233 246 Basarab Horwath Daniels A T and Green 2001 Image analysis dual modality tomography for material classification Meas Sci Technol 12 1153 1156 Jaworski A J and Dyakowski T 2001 Tomographi
283. he same time Hence receiver A will detect a signal from transmitter A modulated by the material between receiver A and transmitter A Specifically the conductance and capacitance of the material between this particular transmitter receiver electrode pair modulates the magnitude and phase of this component Further receiver A will detect a signal from transmitter B except that the material between receiver A and transmitter B will modulate this component Since each transmitter operates at a different frequency the received signal can be separated into the individual components using synchronous detection Synchronous detection is also called phase sensitive detection lock in detection and coherent demodulation In this way the capacitance and conductance of all the different electrode combinations are continuously available at the outputs of the synchronous detectors and no multiplexing is required Subsequently the potential frame rate of the system is only limited by the bandwidth of the electronics The only drawback of the FDM impedance tomography system is a reduction in the number of electrode 2 combinations for a given number of electrodes compared to TDM systems Specifically only transmitter receiver electrode pairs are obtained for an N electrode system To assess the consequences of this reduction in the number of electrode combinations an 8 electrode TDM impedance tomography system was modified into a FDM impedance tomography sys
284. here the data is not continuous procedure TRecLoadDBase RecLoadDataClick Sender TObject var row col tempvalue imgcol integer 51 52 string trainversions trainnumber testnumber testversions traincalib testcalib integer index real tempair tempgravel Boolean begin II if wish to perform calibration then show calibration setup form so that the user can specify which training data points represent the rig being filled only with water if RecLoadCalibrate Checked then CalibSetup ShowModal if CalibSetup ModalResult lt gt mrCancel then begin Update create the tables for the databases so that the data can be accessed TrainInpTable TTable Create Self TrainOutTable TTable Create Self TestInpTable TTable Create Self TestOutTable TTable Create Self set the database names to those specified by the user TrainInpTable DatabaseName RecTrainName Text TrainOutTable DatabaseName RecTrainName Text TestlnpTable DatabaseName RecTestName Text TestOutTable DatabaseName RecTestName Text TrainInpTable TableName InpTrain db TrainOutTable TableName OutTrain db TestInpTable TableName InpTrain db TestOutTable TableName OutTrain db the tables TrainInpTable Open TrainOutTable Open TestlnpTable Open TestOutTable Open InputData nil OutputData nil VolumeData nil calculate the number of training and testing poins to load traint
285. hift 61 Since the probability vector is used to generate the next population of trial solutions the learning rate has a direct affect on the regions of the function space that will be explored 61 If the learning rate is too high then the initial populations generated largely determine the focus of the search and a thorough exploration of the function space is not achieved 61 However if the learning rate is too low then it could take an enormous number of generations to shift the probability vector towards the high evaluation regions and escape from initial oscillation 61 In terms of the intended application the PBIL algorithm is used to determine the set of eight transmitter frequencies that maximise the frequency separation between the harmonics of these eight frequencies The accompanying program code can be found in Appendix J on page 140 The results of the PBIL optimisation algorithm will be discussed in the next section once the generation of the transmitter frequencies has been examined in greater detail TABLE 5 1 5 2 5 Generation of transmitter frequencies TRANSMITTER FREQUENCIES GENERATED USING TWO As mentioned previously the generation of the eight PIC16F84 MICROCONTROLLERS WITH 4 43361MHz AND NS 5MHz CRYSTALS RESPECTIVELY transmitter frequencies was initially achieved using two 5 MHz CRYSTAL PIC16F84 microcontrollers programmed using the TRANSMITTER OUTPUT FREQUENCY 78 125kHz 52 0833kHz for the 8 electrode system Sim
286. hnology 1999 EDR Software developers kit for Eagle Technology boards User Manual For Dos and Windows 114 64 65 66 67 Bendat J S and Piersol A G 1980 Engineering applications of correlation and spectral analysis Toronto John Wiley and Sons Neuffer D Alvarez A Owens D H Ostrowski K L Luke S P and Williams R A 1999 Control of Pneumatic Conveying using ECT Proc 1 World Cong on Industrial Process Tomography Buxton pp 71 76 Fransolet E Crine M L Homme G Toye D and Marchot P 2001 Analysis of electrical resistance tomography measurements obtained on a bubble column Meas Sci Technol 12 1055 1060 Bennett M A Luke S P Jia X West R M and Williams R A 1999 Analysis and Flow Regime Identification of Bubble Column Dynamics Proc 1 World Cong on Industrial Process Tomography pp 54 61 115 APPENDIX A TABLE OF THE MANUFACTURI NG COSTS FOR 16 ELECTRODE FREQUENCY DI VI SI ON MULTI PLEXED I MPEDANCE TOMOGRAPHY SYSTEM DESCRIPTION COST IN US DOLLARS COMPUTER EQUIPMENT PRINTED CIRCUIT BOARD MANUFACTURE AND DESIGN ELECTRONIC COMPONENTS MANUFACTURE OF LABORATORY SCALE RIG OTHER MATERIAL COSTS 116 APPENDI X B CI RCUI T DI AGRAMS FOR 8 ELECTRODE FREQUENCY DI VI SI ON MULTI PLEXED I MPEDANCE TOMOGRAPHY SYSTEM CIRCUIT DIAGRAM OF TRANSMITTER CAPACITANCE RECEIVER AND CONDUCTANCE RECEIVER 117 CIRCUIT DIAGRAM SHOWING PIC16F84 FREQUENCY GENERATOR AND INPUT STAGE CONSIST
287. ialise the tables and create the tables if they do not already exist public Public declarations end var SampMain TSampMain InpTable TTable OutTable TTable SampHardware TSamp notifvalue cardinal table to store input data table to store desired output object used to control sampling Windows notification value for streaming complete result array 0 65535 of double temporary array to store streaming results numframes integer number of frames of data to capture implementation uses addunit curunit R DFM tM DATABASE PROCEDURES get a list of all the available databases the user then selects one procedure TSampMain Opendatabase1 Click Sender TObject var templist TStringList i integer begin templist TStringList Create Session GetAliasNames templist for i 0 to templist Count 1 do SampAliasList Items Add templist i SampAliasOpen Visible true templist Free end when the user selects a specific database initialise the tables procedure TSampMain SampAliasListClick Sender TObject begin SampAliasOpen Visible false if SampAliasList Text lt gt then CreateTables SampAliasList Text end close the databases procedure TSampMain Closedatabase1 Click Sender TObject begin InpTable Close OutTable Close end allow the user to create a new database procedure TSam
288. if RecLoadDBase GetNumOutputs 88 then begin RecLoadDBase GetPhases tempair tempgrav if tempair then WriteLn tempfile 0 else WriteLn tempfile 1 end CloseFile tempfile end save the calibration data to a comma separated textfile for analysis at a later stage procedure TRecMain Savecalibrationdata1 Click Sender TObject var tempfile textfile index overall col row integer outstring string tresult TResultTable tempcaptable array 1 8 1 8 of double temprestable array 1 8 1 8 of double begin row 0 tresult capacitance if SaveCalibDialog Execute then begin AssignFile tempfile SaveCalibDialog Filename ReWrite tempfile to save the data in a representative format create a temporary table and then sort the data accordingly index 1 for overall 1 to 16 do begin case overall of 1 begin row 4 end 2 begin row 4 end 3 begin row 3 end 4 begin tresult capacitance tresult resistance tresult capacitance 154 row 3 tresult resistance end 5 begin row 2 tresult capacitance end 6 begin 2 tresult resistance 7 begin row 1 tresult capacitance end 8 begin row 1 tresult resistance end 9 begin row 5 tresult resistance end 10 begin 5 tresult capacitance end 11 begin 6 tresult resistance end 12 begin row 6 tres
289. ight rowindex colindex changeplus maxdelta HiddenWeight rowindex colindex HiddenWeight rowindex colindex Sign DiffHiddenWeight rowindex colindex HiddenDeltaWeight rowindex colindex OldDiffHiddenWeight rowindex colindex DiffHiddenWeight rowindex colindex end else if OldDiffHiddenWeight rowindex colindex DiffHiddenWeight rowindex colindex 0 then begin Il if opposite sign then crossed valley so reduce the amount by which change the weights and return without making adjustment to weights HiddenDeltaWeight rowindex colindex Max HiddenDeltaWeight rowindex colindex changeminus mindelta OldDiffHiddenWeight rowindex colindex 0 end else begin previous step crossed a valley so can now adjust weights HiddenWeight rowindex colindex HiddenWeight rowindex colindex Sign DiffHiddenWeight rowindex colindex HiddenDeltaWeight rowindex colindex OldDiffHiddenWeight rowindex colindex DiffHiddenWeight rowindex colindex end DiffHiddenWeight rowindex colindex 0 end end end adjustments to output layer for colindex 0 to numoutputs 1 do begin for rowindex 0 to indexend do begin if OldDiffOutputWeight rowindex colindex DiffOutputWeight rowindex colindex gt 0 then begin same sign so increase amount by which change the weights since going down hill OutputDeltaWeight rowindex colindex Min OutputDeltaWeight rowindex colindex
290. igurations of air gravel water mixtures A neural network was again used to perform the volume fraction prediction task and the volume fraction error for each of the three phases was assessed using a database of test cases Specifically the air volume fraction error was 1 2 for the top system and 1 4 for the bottom system the gravel volume fraction error was 1 5 for the top system and 1 7 for the bottom system and the water volume fraction error was 0 86 for the top system and 1 0 for the bottom system Clearly the volume fraction errors of the 16 electrode impedance tomography systems were considerably smaller than the 8 electrode system as expected Further tests revealed that the resolution of the system had improved to where an air bubble corresponding to 1096 of the pipeline diameter could be detected at the centre of the pipeline This bubble corresponds to an air phase volume fraction of 0 01 Although these results were promising it is the performance of the FDM impedance tomography system for real time on line monitoring of dynamic flow simulations that sets it apart from TDM impedance tomography systems Specifically tests revealed that the on line data capture rate of the 16 electrode dual plane impedance tomography System was 280frames s from both measurement planes simultaneously which clearly meets the required specification of 200frames s Further the neural network software permitted the on line reconstruction of these capture
291. in if GravelBubblePosition ItemIndex 1 then tempdir CW else tempdir CCW end else begin if GravelBubblePosition ItemIndex 1 then tempdir CCW else tempdir CW end RealMain Stepper JogMove GravelDrive Value tempdir end downward movement else if ssRight in Shift then begin if GravelDrive Value 1 or GravelDrive Value 3 then begin if GravelBubblePosition ItemIndex 1 then tempdir CCW else tempdir CW end else begin if GravelBubblePosition ItemIndex 1 then tempdir CW else tempdir CCW end RealMain Stepper JogMove GravelDrive Value tempdir end end once the mouse button is released stop the bubble movement procedure TStepperConfig GravelBubbleJogMouseUp Sender TObject Button TMouseButton Shift TShiftState X Y Integer begin RealMain Stepper StopDrive end end PROGRAM LISTING OF STEPUNIT PAS This unit it used to send and receive commands from the AT6400 4 axis controller and in so doing control the motion of the two stepper motors connected for generating test cases The SendCommand and ReceiveCommand procedures were ported from a Pascal device driver provided by Compumotor Since Delphi does not support the PORT and PORTW commands directly a special component called simport was installed to perform this task To download the AT6400 operating system the AT6400 EXE program provided by Compumotor is executed unit StepUnit interface uses sys
292. in FormShow Sender TObject var Reg TRegistry KeyGood Boolean begin Stepper TStep Create Reg TRegistry Create try KeyGood Reg OpenKey SoftwareWeural False if KeyGood then RealOnline ItemIndex Reg ReadInteger RealOnline TopTestNum Value Reg Readlnteger TopTestNumber BottomTestNum Value Reg Readinteger BottomTestNumber RealFrameRate Value Reg Readinteger RealFrameRate DesiredFrames Value Reg ReadInteger DesiredFrames SystemSepInput Text Reg ReadString Separation FlowFactorlnput Text Reg ReadString FlowFactor TopAirThresh Text Reg ReadString TopAirThresh TopGravThresh Text Reg ReadString TopGravThresh BotAirThresh Text Reg ReadString BotAirThresh BotGravThresh Text Reg ReadString BotGravThresh CorrThreshinp Text Reg ReadString CorrThresh Viewlnc Value Reg Readinteger ViewInc DynVerify AirDiameter Text Reg ReadString AirDiameter DynVerify AirLength Text Reg ReadString AirLength DynVerify AirBubbleStartPosition Text Reg ReadString AirBubbleStartPosition 186 DynVerify GravelDiameter Text Reg ReadString GravelDiameter DynVerify GravelLength Text Reg ReadString GravelLength DynVerify GravelBubbleStartPosition Text Reg ReadString GravelBubbleStartPosition DynVerify airposition Reg ReadInteger AirPosition DynVerify gravelposition Reg Rea
293. ing speed of the data acquisition card would need to be improved According to the PC30G manual the maximum sampling rate that can be achieved using single voltage captures under software control is typically 3ksamples s Although the card is rated at 100ksamples s this is only achieved through the use of Direct Memory Access DMA Consequently the data acquisition software for the laboratory scale rig would need to make use of DMA in order to achieve the high frame rate required for the intended application 33 CHAPTER 5 FREQUENCY DI VI SI ON MULTI PLEXED I MPEDANCE TOMOGRAPHY OF A 16 ELECTRODE SYSTEM The following chapter details the development of a dual plane 16 electrode FDM impedance tomography system 5 1 Construction of a Laboratory Scale Rig to Simulate a Pipeline Section TOP SECTION MEASUREMENT SECTION 26 zu o 500 260 STAINLESS STEEL SHIELD STAINLESS STEEL BASE WITH DRAIN VALVE FIGURE 5 1 OVERALL RIG DIMENSIONS FOR LABORATORY SCALE MULTI PHASE PIPELINE SECTION ALL DIMENSIONS ARE IN MILLIMETRES A laboratory scale rig was constructed to assess the performance of the FDM impedance tomography system on simulated flows The following sections examine the selection of certain parameters during the design of the rig 5 1 1 Reasoning behind certain rig dimensions The rig diameter was chosen to be equal to the pipeline diameter of the industrial airlift section where the impedan
294. ing whether driven axial guards RECEIVER GROUNDED increase or decrease the measurement WO m m wp We m 9 mW CS o WO o a AXIAL A Hr volume Figure 5 3 illustrates the generation DRIVEN 9 LU ia a GUARD of a homogenous electric field for TRANSMITTER FIPE WALL capacitance tomography using driven axial FIGURE 5 3 guard electrodes As is evident in figure 5 3 DIAGRAM ILLUSTRATING THE GENERATION OF A HOMOGENOUS ELECTRIC FIELD BETWEEN THE CAPACITANCE MEASUREMENT the electric field generated between the ELECTRODES USING DRIVEN AXIAL GUARD ELECTRODES ON THE TRANSMITTER SIDE THE GUARD ELECTRODES ARE SIMPLY measurement electrodes approximates a CONNECTED TO THE TRANSMITTER OUTPUT ON THE RECEIVER SIDE THE GUARD ELECTRODES TIED TO GROUND SINCE THE two dimensional field thus improving the TRANSIMPEDANCE AMPLIFIER HOLDS THE RECEIVER MEASUREMENT M ELECTRODE AT VIRTUAL EARTH validity of the reconstruction algorithm 36 In addition it has been shown that driven axial guards improve the axial resolution and measurement sensitivity 11 Further driven axial guards are normally essential for large diameter pipelines to ensure that the standing capacitances between diametrically opposite electrodes can be measured 11 In terms of the measurement volume Yan et al 51 and Xu et al 52 conclude that the measurement volume produced by driven axial guards is the largest of the three possibl
295. ining database The validation database is used to determine whether the neural network is able to generalise from a single bubble of only one phase in the training database to combinations of bubbles of different phases as would typically be achieved in a more realistic situation 15 The generalisation ability of the neural network is assessed using the following performance measures 1 Threshold error which is the percentage of pixels that are not predicted correctly For example if the network predicts that a particular pixel is gravel when it is actually air then this represents an error These errors are summed over the entire database and are divided by the total number of pixels in the database This performance measure applies only to the image reconstruction neural network and gives an indication of the positional correlation between the desired network output and the network prediction 2 Volume fraction error which is the percentage error of the volume fraction predictions for each of the three phases For example if the image reconstruction neural network predicts that there are 42 pixels of seawater when there are actually 44 then this represents an error of two pixels These errors in the seawater prediction are summed over the entire database and are divided by the total number of pixels in the database to give the seawater volume fraction error For the volume fraction predictor neural networks the absolute values of the diffe
296. injection is then moved to the next electrode pair and the process is repeated To avoid contact impedance problems the voltages of the electrodes adjacent to and including the injection pair are not measured The total number of independent measurements obtained is NN 3 where N is the number of electrodes 18 Figure 2 4 illustrates the four electrode adjacent pair measurement protocol EQUIPOTENTIAL CURRENT FLOW VOLTAGE MEASUREMENT ON REMAINING ELECTRODE PAIRS CURRENT INJECTION ELECTRODE PAIR ELECTRODE FIGURE 2 4 DIAGRAM ILLUSTRATING THE FOUR ELECTRODE ADJACENT PAIR MEASUREMENT PROTOCOL FOR RESISTANCE TOMOGRAPHY ON A PIPELINE CROSS SECTION WHERE THE ELECTRODES ARE EMBEDDED IN THE PIPELINE WALL AN ALTERNATING CURRENT IS INJECTED BETWEEN A PAIR OF ADJACENT ELECTRODES AND THE VOLTAGE DIFFERENCES BETWEEN THE REMAINING ELECTRODE PAIRS ARE MEASURED THE CURRENT INJECTION IS THEN MOVED TO THE NEXT ELECTRODE PAIR AND THE PROCESS IS REPEATED INCLUDED IN THE ABOVE FIGURE ARE LINES REPRESENTING THE CURRENT FLOW BETWEEN THE CURRENT INJECTION ELECTRODES AS WELL AS THE EQUIPOTENTIAL LINES FOR THE VOLTAGE MEASUREMENT ELECTRODE PAIR ILLUSTRATED The motivation for the adjacent measurement protocol is to ensure that any variation in the contact impedance between the electrodes and the substance being imaged does not influence the measured voltages 17 As a result of the forced current various reactions may occur at the injec
297. instruction is followed by the goto CRYSTAL FOR THE ELECTRODE SYSTEM instruction and consequently the correct duty cycle OUTPUT FREQUENCY and phase shift is maintained Using this technique the output frequencies generated with a 4MHz crystal AEZ are given in table 4 1 Further Appendix G on page 135 lists up to the 9 harmonic of these fundamental frequencies 4 2 2 Output transmitter details The load on a transmitter was measured using an RLC meter In particular for the rig filled only with seawater the average resistance of a transmitter receiver electrode pair is 50O Since each transmitter is required to drive four receiver electrodes in parallel the approximate load on a transmitter is 12 5 A power op amp was therefore required Specifically the LM675 power op amp was used for the output transmitter and the circuit details are provided in Appendix on page 117 Table 4 2 briefly lists the specifications of the LM675 that are appropriate to this application TABLE 4 2 LM675 SPECIFICATIONS APPROPRIATE TO THIS APPLICATION 25V Ta 25 C PARAMETER CONDITION TYPICAL OUTPUT POWER THD 1 fo 1kHz 80 OUTPUT VOLTAGE SWING 80 GAIN BANDWIDTH PRODUCT fo 20kHz Ave 1000 MAXIMUM SLEW RATE Although this op amp was capable of driving the load presented by the rig the output square waves are distorted as a result of the limited slew rate and the crossover distortion of the op amp T
298. ion var testindex integer upperindex integer i k maxindex integer tempsum maxval real denominator real begin perform direct correlation on the air volume fraction data from the top and bottom system upperindex DesiredFrames Value maxval 0 for i 0 to upperindex do begin tempsum 0 for k 1 to upperindex do begin if k i lt upperindex then begin if totaloutputs 2 then begin if StepperConfig FlowDirSelect ItemIndex 1 then tempsum tempsum outputstore k i prediction 1 outputstore k prediction 3 else tempsum tempsum outputstore k i prediction 3 outputstore k prediction 1 end else begin if StepperConfig FlowDirSelect ItemIndex 1 then tempsum tempsum volumestore k i 1 volumestore k 3 else tempsum tempsum volumestore k i 3 volumestore k 1 end end end aircorr i 1 tempsum upperindex keep a record of the maximum of the correlation function as well I as the corresponding frame number if aircorr i 1 gt maxval then begin maxval aircorr i 1 maxindex i 1 end end calculate the transit time for the air bubble airdelay maxindex 1 denominator airdelay frameinterval if denominator 0 or maxval lt correlationthreshold then begin airvelocity 1 AirVelPredict Caption unknown end else begin IL if non zero then calculate the average velocity of the air bubble airvelocity flowmete
299. ion For example if the system were to be used on a horizontal pipeline containing a stratified flow then the training database would have to be extended to include this possibility The cross correlation algorithm can accurately measure the velocities of individual components in a multi phase flow provided the reconstruction algorithm is able to distinguish between the different phases Specifically for a simulated air and gravel upward flow the cross correlation algorithm was able to measure the individual velocities of the air and gravel phases Further the cross correlation algorithm was able to measure the velocity of a 0 01 volume fraction air bubble with a velocity discrimination of only 3 196 at the maximum bubble velocity of 4 2m s which is comparable to the theoretical limit for this particular system It has also been shown that component volume fractions provide a computationally efficient means of calculating the average individual component velocities without having to perform a pixel by pixel cross correlation Square wave excitation can achieve good two phase reconstruction results at a reasonable cost However to achieve better three phase reconstruction results and specifically improved air gravel differentiation greater accuracy of the conductance and specifically the capacitance measurements are required Sine wave excitation systems provide an alternative to the square wave excitation system considered in this thesis which
300. ion and analysis is then performed off line afterwards In each screen capture the top frame represents the top measurement plane and the bottom frame represents the bottom measurement plane The order of the frames is from left to right and then top to bottom Due to the large number of frames captured during testing only every 50 frame is included in figure 7 9 As is evident in figure 7 9 the 0 04 volume fraction air bubble is generally detected first and last as the pair of bubbles move through a measurement section In addition when the 0 04 volume fraction air bubble is not completely within the measurement volume of the system it is detected as a smaller bubble As the bubble enters the measurement volume so the size of the bubble increases until the desired 0 04 volume fraction air bubble is predicted This raises questions regarding what effect the length of the bubble has on the volume fraction measurement accuracy Previously all training and test data considered only bubbles that extended beyond the axial length of the measurement electrodes For the laboratory scale rig the measurement plates are 120mm long Consequently all bubbles simulated were 190mm long Tests were conducted to determine what effect shorter bubbles would have on the accuracy of the volume fraction measurement Specifically the dynamic performance of the top measurement plane was tested with air bubbles 60mm and 90mm long Tests were also conducted for the stand
301. ion prediction colindex 1 tanh weightsum end I three phase volume fraction prediction else begin if colindex 1 1 then begin if weightsum gt topairthreshold then outputstore reconposition prediction 1 weightsum else outputstore reconposition prediction 1 0 end else begin if weightsum gt topgravthreshold then outputstore reconposition prediction 2 weightsum else outputstore reconposition prediction 2 0 end end end now the bottom rig upperindex 128 Il if there are hidden layer neurons then calculate the output I from the hidden layer first if numbottomhidden gt 0 then begin upperindex numbottomhidden for colindex 0 to numbottomhidden do begin if colindex 0 then bottomhiddenoutputs colindex 1 else begin weightsum 0 for rowindex 0 to 128 do begin if rowindex 0 then tempvalue 1 else tempvalue inputstore reconposition readings rowindex 128 weightsum weightsum bottomhiddenweights rowindex colindex 1 tempvalue end prevent arithmetic overflow if weightsum gt 200 then weightsum 200 else if weightsum lt 200 then weightsum 200 tanh activation function bottomhiddenoutputs colindex tanh weightsum end end end once calculated hidden layer outputs calculate the final network prediction for colindex 0 to totaloutputs 1 do begin weightsum 0 for rowindex 0 to upperindex do begin if
302. irdiam 2 1 3 pi power tempheight 3 o end else owerairvol 0 end else begin if airtopposition lowersysbot and airtopposition airleng lt lowersysbot then lowerairvol pi power 0 5 airdiam 2 airtopposition lowersysbot else if airtopposition airleng gt lowersysbot and airtopposition lt lowersystop then owerairvol pi power 0 5 airdiam 2 airleng else if airtopposition gt lowersystop and airtopposition airleng lt lowersystop then lowerairvol pi power 0 5 airdiam 2 lowersystop airtopposition airleng else oweraitvol 0 end II air bubble for top measuring ring if airleng gt uppersystop uppersysbot then begin if airtopposition gt uppersysbot and airtopposition lt uppersysbot 0 5 airdiam then begin tempheight uppersysbot airtopposition 0 5 airdiam upperairvol airhemivol pi tempheight power 0 5 airdiam 2 1 3 pi power tempheight 3 end else if airtopposition gt uppersysbot 0 5 airdiam and airtopposition lt uppersystop then begin tempheight airtopposition 0 5 airdiam uppersysbot upperairvol airhemivol pi tempheight power 0 5 airdiam 2 end else if airtopposition gt uppersystop and airtopposition lt uppersystop 0 5 airdiam then begin tempheight uppersystop airtopposition 0 5 airdiam upperairvol pi tempheight power 0 5 airdiam 2 1 3 pi power tempheight 3 pi uppersystop uppersysbot tempheight
303. irection dir2 TDirection dir3 TDirection var temp string dirtemp1 string dirtemp2 string dirtemps string begin continuous movement temp MC1111 ProcessCommand temp set the velocities of the drives temp V inttostr vel1 inttostr vel2 inttostr vel3 10 190 ProcessCommand temp set the directions of the drives if dirl CW then dirtemp1 else dirtemp1 if dir2 CW then dirtemp2 else dirtemp2 dir3 CW then dirtemp3 zs else dirtemp3 temp D dirtemp1 dirtemp2 dirtemp3 ProcessCommand temp initiate movement on drive if drive1 then dirtemp1 1 else dirtemp1 0 if drive2 then dirtemp2 1 else dirtemp2 0 if drive3 then dirtemp3 1 else dirtemp3 0 temp GO dirtemp1 dirtemp2 dirtemp3 0 ProcessCommand temp end issue a command to immediately stop the drive movement at a controlled deceleration ramp procedure TStep StopDrive var temp string begin temp s ProcessCommand temp end set the current drive position as the absolute zero procedure TStep SetZeroPos var temp string begin temp PSETO 0 0 0 ProcessCommand temp end disable the drives procedure TStep DisableDrives var temp string begin temp DRIVE0000 ProcessCommand temp temp 000000 ProcessCommand temp end reset the d
304. isation of the tomogram 25 The parameters identified during reconstruction are exactly those required for interpretation or control This has the advantage of more efficiently utilising the computational power available and reducing the reconstruction time 25 Since the number of parameters is typically small compared to the number of pixels in the reconstructed images the solution to the problem is also better posed than the standard image reconstruction task and hence better results can be achieved Tests revealed that a neural network could be trained to predict the volume fractions directly 3 4 In this case the network is required to model the functional mapping between the readings obtained from the data acquisition System and the continuous outputs representing the volume fractions of the air gravel and seawater phases This is achieved by reducing the number of output neurons to two thus giving the volume fraction of gravel and air phases respectively The volume fraction of seawater is found by subtracting the sum of air and gravel volume fractions from 1 Secondly linear activation functions are used for the two output neurons 13 Consequently the neural network performs a regression function namely lo raveleraction aleRACTION fREGRESSION 3 8 where aes 0 2 3 9 all FRACTION Figure 3 3 illustrates the operation of a double layer feed forward neural network performing a volume fraction predicti
305. it UpperSystemTop TEdit Label6 TLabel Label7 TLabel Label8 TLabel AirDiameter TEdit AirLength TEdit GravelDiameter TEdit GravelLength TEdit Label9 TLabel Label10 TLabel Label11 TLabel ShapePipeAir TShape AirShapet TShape AirShape3 TShape AirShape2 TShape RealExit TSpeedButton AirShape4 TShape AirShape5 TShape GravShape4 TShape GravShape3 TShape GravShape5 TShape GravShape2 TShape GravShape1 TShape ShapePipeGrav TShape Label12 TLabel Label13 TLabel TopAirPlot TChart Series3 TLineSeries Series4 TLineSeries BotAirPlot TChart LineSeries1 TLineSeries LineSeries2 TLineSeries TopGravelPlot TChart LineSeries3 TLineSeries LineSeries4 TLineSeries BotGravelPlot TChart LineSeries5 TLineSeries LineSeries6 TLineSeries CalcDesired TSpeedButton RealSaveResults TSpeedButton SaveResultsDialog TSaveDialog VelSource TRadioGroup CalcVolError TSpeedButton ErrorResults TPanel Label14 TLabel Label15 TLabel Label16 TLabel Label17 TLabel Label18 TLabel Label19 TLabel TopAirVolError TLabel BotAirVolError TLabel TopGravVolError TLabel BotGravVolError TLabel CalcActualVel TSpeedButton TopAirDelay TLabel BotAirDelay TLabel Label20 TLabel TopGravDelay TLabel BotGravDelay TLabel Series TBarSeries Series2 TBarSeries Series5 TBarSeries Series6 TBarSeries Label21 TLabel Label22 TLabel AirBubbleStartPositi
306. ith static bubble configurations provide a means of performing accurate on line real time reconstructions of dynamic flows A single layer feed forward neural network was trained to perform a two phase air water volume fraction prediction The volume fraction error of this network for static test cases was 0 20 for the top measurement plane and 0 18 for the bottom measurement plane For dynamic tests the volume fraction errors increased to 0 68 for the top measurement plane and 0 58 for the bottom measurement plane The average on line frame rate of this reconstruction algorithm is 280frames s A double layer feed forward neural network was trained to perform a three phase air gravel water volume fraction prediction The volume fraction errors of this network for static test cases were 1 2 1 5 and 0 86 for the air gravel and water phases at the top measurement plane respectively The volume fraction errors for the air gravel and water phases at the bottom measurement plane for static test cases were 1 4 1 7 and 1 0 respectively For dynamic tests the volume fraction errors measured at the top measurement plane were 1 096 and 1 596 for the air and gravel phases respectively and 0 8996 and 1 696 at the bottom measurement plane respectively The average on line frame rate of this reconstruction algorithm is 180frames s Although the initial results for the three phase reconstructions were promising the neural networks trained were unable to con
307. ive Method for Faster Backpropagation Learning The RPROP Algorithm Proc IEEE Int Conf on Neural Networks ICNN San Francisco ed H Ruspini pp 586 591 Riedmiller M 1994 Advanced Supervised Learning in Multi layer Perceptrons From Backpropagation to Adaptive Learning Algorithms nt J of Computers Standards and Interfaces 5 Riedmiller M 1994 Rprop Description and Implementation Details Technical Report University of Karlsruhe Szczepanik Z and Rucki 2 2001 Problems of the Data Acquisition Rate for Impedance Tomography Proc 274 World Cong on Industrial Process Tomography Hannover pp 511 516 Horowitz P and Hill W 1995 The art of electronics Cambridge Cambridge University Press Georgakopoulos D Waterfall R C and Yang W Q 2001 Towards the Development of a Multiple Frequency ECT System Proc 2 World Cong on Industrial Process Tomography Hannover pp 550 556 Brown B H 2001 Medical impedance tomography and process impedance tomography a brief review Meas Sci Technol 12 991 996 Morrison N 1994 ntroduction to Fourier Analysis Toronto John Wiley and Sons 113 42 43 44 45 46 4T 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 Hervieu E and Seleghim P Jr 1999 Direct Imaging of Two Phase Flows by Electrical Impedance Measurements Proc 1 World Cong on Industrial Process Tomography Buxton 62 69 Deng X Dong F Xu L J Liu X P and Xu L A 2001 Me
308. iver 45 To obtain a higher output current from the TC4427 the outputs were tied together and the inputs driven in parallel as can be seen in Appendix C on page 122 In addition a 10nF capacitor was placed from the output of the TC4427 to ground to minimise the output ringing and overshoot The functionality of this capacitor can be explained by examining the output stage of the TC4427 Specifically it consists of a push pull MOSFET stage and these MOSFETs have a fairly high on resistance of 10Q maximum This on resistance creates an RC combination with the capacitor thus slowing down the output rise and fall Subsequently the output ringing and overshoot is reduced In order to generate a 9V to 9V output square wave the ground line of the TC4427 was connected to 9V Since the required input is a TTL signal with respect to this ground line it was necessary to shift the input square wave from the PIC16F84 microcontroller The potential divider configuration on the input to the TC4427 shifts the zero to square wave from the microcontroller to a 9V to 4V square wave required for correct operation of the MOSFET driver Although good results were achieved when only a single transmitter was operating at any one time the output waveforms were seriously degraded when more than one transmitter was operating as would be the case for normal operation in an FDM impedance tomography system Specifically it was noted that superimposed on the
309. l Label15 TLabel Label16 TLabel Label17 TLabel Label18 TLabel Label19 TLabel Label20 TLabel DesiredFrames TSpinEdit Label21 TLabel smins TLabel ssecs TLabel smsecs TLabel MotorSetup TSpeedButton StepperStop TSpeedButton Motor1 Setting TLabel Motor2Setting TLabel AirVelActual TLabel GravelVelActual TLabel RealStartMotors TSpeedButton AirPlot TChart Series3 TLineSeries Series4 TLineSeries GravelPlot TChart LineSeries1 TLineSeries LineSeries2 TLineSeries AirCorrPlot TChart LineSeries3 TLineSeries GravelCorrPlot TChart LineSeries4 TLineSeries Label22 TLabel Label23 TLabel SystemSeplnput TEdit FlowFactorlnput TEdit Label24 TLabel Label25 TLabel Label26 TLabel AirVelPredict TLabel GravelVelPredict TLabel SetThreshold TPanel Label27 TLabel TopAirThresh TEdit Label28 TLabel Label29 TLabel TopGravThresh TEdit Label30 TLabel BotAirThresh TEdit Label31 TLabel BotGravThresh TEdit Label32 TLabel CorrThreshinp TEdit VerifyResults TSpeedButton Label33 TLabel ViewInc TSpinEdit procedure RealExitClick Sender TObject procedure FormShow Sender TObject procedure RealCalibrateClick Sender TObject procedure RealStartClick Sender TObject procedure RealWeightsClick Sender TObject procedure RealViewClick Sender TObject procedure RealTimerTimer Sender TObject procedure MotorSetupClick Sender TObject procedure Stepp
310. lect Items StepperConfig FlowDirSelect ItemIndex DecodeTime outputstore row sampletime hours mins secs msecs WriteLn tempfile air bubble diameter AirDiameter Text totalmsecs 1000 mins 60 secs msecs WriteLn tempfile air bubble length AirLength Text Write tempfile totalmsecs WriteLn tempfile air bubble position inttostr airposition Write tempfile row RealMain frameinterval 8 4 WriteLn tempfile air bubble start position AirBubbleStartPosition Text CalculateDistance row RealMain frameinterval topairposition WriteLn tempfile gravel bubble diameter GravelDiameter Text topgravelposition WriteLn tempfile gravel bubble length GravelLength Text Write tempfile topairposition 8 4 WriteLn tempfile gravel bubble position inttostr gravelposition Write tempfile topgravelposition 8 4 WriteLn tempfile gravel bubble start position for col 1 to 4 do GravelBubbleStartPosition Text begin WriteLn tempfile upper system top UpperSystemTop Text if totaloutputs 2 then WriteLn tempfile upper system bottom UpperSystemBottom Text Write tempfile outputstore row prediction col 8 4 WriteLn tempfile lower system top LowerSystemTop Text else WriteLn tempfile lower system bottom LowerSystemBottom Text Write tempfile volumestore row col 8 4 WriteLn tempfile source of bubble velocity end VelSource Items VelSource ItemIndex for col 1 to 4 do WriteLn tempfile top air
311. led using a stepper controller and driver system obtained from DebTech The stepper controller panel consists of an AT6400 4 axis stepper controller two OEM330 stepper drivers and one OEM530 stepper driver A controller card installed in the reconstruction computer controls the AT6400 4 axis stepper controller A Pascal device driver provided with the controller card was ported into Delphi to control the stepper motor for flow simulation purposes The complete functionality of the stepper controller was coded in a unit which can be found in Appendix L on page 189 Specifically this unit provides access to high level routines that can be called without having to generate the operating system syntax used by the stepper controller Examples of these high level procedures include e Initialise loads a text file of initialisation data such as the drive resolution and the stepper pulse width and configures the AT6400 stepper controller e DownloadOS the AT6400 stepper controller requires an operating system to run and this procedure downloads the operating system file to the stepper controller GoDrive initiates motion on the specified drives at the specified velocity and in the specified direction e JogMove jogs the specified drive in the specified direction and is used for the accurate positioning of the bubbles before a test e MoveFixedDist initiates motion on the specified drives at the specified velocity and for the sp
312. lements long to select frequency besteverfit 1 record of best ever result calculate the bit weights for g 0 to prec 1 do bw g 1 floor power 2 g initialise the probability vector to all 0 5 for g 1 to prec nvars do PV g 0 5 for g 1 to maxgen do begin PBILProgress Steplt bestfit 1 for trial 1 to ntrials do begin generate a random trial solution for i 1 to prec nvars do trialsol i short PV i gt Random convert the sequence of bits into integer values 1 for i 1 to nvars do begin 0 for j 1 to prec do begin R i R i trialsol k bw j inc k end end use these 8 integer values to index into the array of available transmitter frequencies for i 1 to nvars do inputs i AvailFreq R i 1 test current set of 8 transmitter frequencies fit MultOut inputs if the performance is the best so far for this generation then update the record if fit gt bestfit then begin bestfit fit for i 1 to nvars prec do begin bestsol i trialsol i if i lt nvars then bestreal i inputs i end end end res MultOut bestreal if res 0 1 then begin using the best solution adjust the probability vector towards the best solution for i 1 to prec nvars do 140 begin 0 9 PV i 0 1 bestsol i PV i PV i 0 005 PV i 0 5 end keep record of the best overall results if bestfi
313. leng gt uppersystop uppersysbot then begin if gravtopposition gt uppersysbot and gravtopposition lt uppersysbot 0 5 gravdiam then begin tempheight uppersysbot gravtopposition 0 5 gravdiam uppergravvol gravhemivol pi tempheight power 0 5 gravdiam 2 1 3 pi power tempheight 3 end else if gravtopposition gt uppersysbot 0 5 gravdiam and gravtopposition lt uppersystop then begin tempheight gravtopposition 0 5 gravdiam uppersysbot uppergravvol gravhemivol pi tempheight power 0 5 gravdiam 2 end else if gravtopposition gt uppersystop and gravtopposition lt uppersystop 0 5 gravdiam then begin tempheight uppersystop gravtopposition 0 5 gravdiam uppergravvol pi tempheight power 0 5 gravdiam 2 1 3 pi power tempheight 3 pi uppersystop uppersysbot tempheight power 0 5 gravdiam 2 end else if gravtopposition gt uppersystop 0 5 gravdiam and gravtopposition gravleng 0 5 gravdiam lt uppersysbot then begin uppergrawvol pi uppersystop uppersysbot power 0 5 gravdiam 2 end else if gravtopposition gravleng 0 5 gravdiam gt uppersysbot and gravtopposition gravieng 0 5 gravdiam lt uppersysbot 0 5 gravdiam then begin tempheight gravtopposition gravleng 0 5 gravdiam uppersysbot uppergravvol pi tempheight power 0 5 gravdiam 2 1 3 pi power tempheight 3 pi uppersystop uppersysbot tempheight power 0 5 gravdiam 2 end else if gravtopposition gravieng 0 5 gravdiam g
314. licability of the neural network reconstruction algorithms to the intended application should be assessed Various modifications to the generation of the neural network training database have been proposed throughout this thesis which should improve the overall performance of the system These include extending the database to incorporate bubbles that are shorter than the electrode length and depending on the intended application incorporating an all air training data point Although not discussed in this thesis it is also possible that improved air gravel separation could be achieved by training separate neural networks to predict the air and gravel phases individually In this way each neural network can be optimised for the detection of a particular phase Currently a single neural network is trained to perform both the air and gravel volume fraction predictions If an improvement in the air gravel separation were not achieved through these proposed modifications then it would be advisable to re examine the positioning of the tomography system in the intended application Specifically greater accuracy would be obtained if some form of partial separation of the phases in the flow had already been achieved In terms of the current application it has been proposed that the impedance tomography system is positioned before the introduction of the air phase in the airlift thus simplifying the measurement task to that of a two phase gravel seawater reconstr
315. ll the different electrode combinations are measured Various protocols define the order in which these capacitances are measured The standard capacitance tomography data collection technique is referred to as protocol 1 where the measurements are made only between single pairs of electrodes This measurement sequence involves applying an alternating voltage to the first electrode and then measuring the currents received on the remaining electrodes The second electrode is then connected to this alternating voltage source and the currents are again measured on the remaining electrodes The first two stages of this measurement protocol are illustrated in figure 2 1 Using this data collection protocol the number of independent measurements obtained is NN where N is the number of electrodes Figure 2 2 on page 6 illustrates the 28 linearly independent capacitance measurements possible for an 8 electrode system RECEIVER RxH RxH RxG RxG TRANSMITTER RxF RxF RxB TxB RxE RxE RxC RxC RxD RxD FIGURE 2 1 DIAGRAM ILLUSTRATING THE FIRST TWO STAGES OF THE STANDARD CAPACITANCE TOMOGRAPHY DATA COLLECTION PROTOCOL SHOWING FOR EACH STAGE A CROSS SECTION OF THE PIPELINE AND THE CAPACITANCE PLATES MOUNTED AROUND THE EXTERNAL PERIMETER OF THE PIPELINE IN THE FIRST STAGE ELECTRODE A IS CONNECTED TO THE ALTERNATING VOLTAGE SOURCE AND ALL THE REMAINING ELECTRODES ARE CONNECTED TO THE TRANSIMPEDANCE AMPLIFIER ONE AT A TIME THE LINES B
316. locity vector 10000000 1000000 100000 10000 E 9 5 2 2 o E 9 s 5 2 a E o 1000 100 6 minimum velocity m s FIGURE 7 1 COMPARISON BETWEEN THE COMPUTATIONS REQUIRED TO DETERMINE THE AVERAGE VELOCITY VELOCITY PROFILE AND VELOCITY VECTOR FOR A GIVEN FRAME RATE AND AT A SPECIFIED MINIMUM FLOW VELOCITY One possible reduction in the computational requirements can be achieved by noting that the signals of adjacent pixels are also themselves highly correlated 59 Consequently it may be possible to reduce the number of pixels in the images that are correlated Otherwise large amounts of redundant information may be obtained Using the velocity profile information obtained from the cross correlation function it is possible to determine the mass flow rate of a particular component n within a multi component flow as follows 50 M AUW p 7 9 where A is the cross sectional area of the pipe is the mean flow velocity W gives the fractional cross sectional area of the pipeline occupied by component n and p is the density of component n The volume fraction calculated by the neural network is effectively the same as W and uis calculated by the cross correlation of the volume fractions measured at the two planes This equation is based on the assumptions that the flow distribution is fixed and the components in the multi phase flow are moving at the same velocity 81 Fo
317. lts online then update the display if RealOnline ItemIndex 0 then begin PlotTestOutput reconposition end inc reconposition end if reconposition gt 6000 then reconposition 1 busyreconstruction false I if sampling was stopped to allow for the reconstruction to complete then restart the sampling and perform the reconstructions for the previous set of 50 frames if not busysampling and totalfrms lt DesiredFrames Value then begin busysampling true SampHardware StartSample 50 if reconposition storeposition 50 then RealReconstruction end end display the real time reconstruction results procedure TRealMain PlotTestOutput testposition integer var topoutputpred bottomoutputpred TOutput topphasepred bottomphasepred TBubblePhase index col row top bottom integer containsair containsgravel Boolean begin index 1 if totaloutputs 264 or totaloutputs 88 then begin for col 0 to 9 do begin for row 0 to 9 do begin topoutputpred col row outpipe bottomoutputpred col row outpipe end end for col 0 to 9 do begin case col of 0 begin top 2 bottom 7 end 1 begin top 1 bottom 8 end 2 7 begin top 0 bottom 9 end 8 begin top 1 bottom 8 end 9 begin top 2 bottom 7 end end for row top to bottom do begin if totaloutputs 264 then begin if outputstore testposition prediction i
318. m a two phase gravel seawater image reconstruction the results of which are given in table 4 4 below TABLE 4 4 COMPARISON BETWEEN THE PERFORMANCE OF THE FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM AND A STANDARD TIME DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM FOR A TWO PHASE GRAVEL SEAWATER IMAGE RECONSTRUCTION USING A SINGLE LAYER FEED FORWARD NEURAL NETWORK ALL RESULTS REPRESENT THE ABSOLUTE ERROR AND ARE ROUNDED OFF TO TWO SIGNIFICANT DIGITS PERFORMANCE MEASURE TIME DIVISION FREQUENCY DIVISION MULTIPLEXED IMPEDANCE MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM TOMOGRAPHY SYSTEM THRESHOLD ERROR 13 8 7 GRAVEL VOLUME FRACTION ERROR 4 5 4 4 WATER VOLUME FRACTION ERROR 4 5 4 4 31 A single layer feed forward neural network with a 1 of C output encoding was trained using Resilient back propagation to perform a three phase air gravel seawater image reconstruction The operation of this neural network was examined in figure 3 2 on page 13 The decision to use Resilient back propagation was based on the results of the previous research Once again three networks were trained and the average results are given in table 4 5 below TABLE 4 5 COMPARISON BETWEEN THE PERFORMANCE OF THE FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM AND A STANDARD TIME DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM FOR A THREE PHASE AIR GRAVEL SEAWATER IMAGE RECONSTRUCTION USING A SINGLE LAYER FEED FORWARD N
319. m the three phase air gravel water volume fraction prediction and the averaged results are detailed in table 6 5 TABLE 6 5 PERFORMANCE OF THE 16 ELECTRODE FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM FOR A THREE PHASE AIR GRAVEL WATER VOLUME FRACTION PREDICTION USING A SINGLE LAYER FEED FORWARD NEURAL NETWORK ALL RESULTS REPRESENT THE ABSOLUTE ERROR AND ARE ROUNDED OFF TO TWO SIGNIFICANT DIGITS PERFORMANCE MEASURE 16 ELECTRODE LABORATORY SCALE FDM IMPEDANCE TOMOGRAPHY SYSTEM TOP SYSTEM BOTTOM SYSTEM AIR VOLUME FRACTION ERROR GRAVEL VOLUME FRACTION ERROR WATER VOLUME FRACTION ERROR SUM VOLUME FRACTION ERROR As is evident from table 6 5 no major improvement in the accuracy of the volume fraction predictions was obtained compared to the image reconstruction results examined in table 6 4 Previous research has revealed that improved volume fraction predictions can be achieved through the use of a hidden layer of neurons 4 A model search was conducted to determine the optimal number of hidden layer neurons It was determined that optimal performance is achieved for the top system using 185 hidden layer neurons and 220 hidden layer neurons were used for the bottom system The reconstruction results of these double layer feed forward neural networks are detailed in table 6 6 which includes a comparison to the results of the 8 electrode prototype TABLE 6 6 COMPARISON BETWEEN THE PERFORMANCE OF THE 16 EL
320. m using sine wave excitation must be developed Tests should be conducted to assess whether an improvement in reconstruction accuracy and air gravel separation is achieved through the use of sine wave excitation If improved results are achieved then this system must be extended for use on the dual plane 16 electrode laboratory scale rig The various dynamic tests considered in this thesis can then be repeated and the results compared Research must be done into the drift of the conductance and capacitance readings resulting from the electrochemical reactions taking place within the rig These tests can be performed for the simpler configuration of a 2 electrode system consisting of a single transmitter and single receiver electrode situated diametrically opposite each other The University of Cape Town Control and Instrumentation department possess a 50mm 0 1 6m s flow rig that can be used to conduct these tests A special pipeline section must be constructed with these two electrodes Initially one end of the section can be sealed and placed vertically so that water is retained A drift analysis should be performed for this configuration in order to provide a reference point corresponding to static water The end seal can then be removed and the pipeline section placed within the flow loop of the flow rig The above drift analysis should then be repeated and a comparison made Various electrode materials could also be tested using this approach The app
321. monstrate the generation of the different measurement permutations using multiple transmitters and receivers operating in parallel with frequency division multiplexed impedance tomography Block diagram illustrating the concept of frequency division multiplexed impedance tomography of an 8 electrode system Block diagram showing the interactions between the transmitters receivers and main synchronous detection board in the 8 electrode frequency division multiplexed impedance tomography system Output frequency generation for the 8 electrode frequency division multiplexed impedance tomography system using a PIC16F84 microcontroller Diagram of the capacitance and conductance receiver for the 8 electrode frequency division multiplexed impedance tomography system Synchronous detection of the receiver signal using a DG303 CMOS analogue switch to emulate multiplication by a reference waveform Overall rig dimensions for laboratory scale multi phase pipeline section Photograph of laboratory scale multi phase pipeline section Diagram illustrating the generation of a homogenous electric field between the capacitance measurement electrodes using driven axial guard electrodes Dimensions of the transmitter and receiver electrodes and guards Photograph of the pipeline from above with the 16 stainless steel conductance probes visible at the top and bottom measurement sections Diagram illustrating the outer earthed screen and the projected radial guar
322. more detailed analysis of the dynamics of a flow In particular the cross spectral density function enables one to determine the attenuation and delay of the different frequency components in the propagation of turbulence from the first sensor to the second 50 Hence the cross spectral density function can be used to detect the existence of flow disturbances travelling at different speeds 50 This density function is defined over all frequencies both positive and negative In practice it is more convenient to work with non negative frequencies only and so the one sided spectral density function G f is defined as follows 64 To The Fourier transform and the inverse Fourier transform provide a computationally efficient means of calculating the cross correlation function as follows 50 firstly the Fourier transforms X k and Y k are calculated for the discrete functions x n and y n where X k FFT x n and Y k FFT y n The cross spectral density function is then calculated as G k X k Y where the denotes the conjugate Finally the cross correlation function is calculated using the inverse Fourier transform as R n IFFT G k 78 Although the above technique is more computationally efficient simplicity of solution has resulted in the cross correlation often being calculated using the direct or point by point method described below In particular the cross correlation results achieved in the following
323. mu II hae y 1 h 0 5 1 1 5 2 2 5 3 3 5 4 4 5 time s BOTTOM SYSTEM VOLUME FRACTION MEASUREMENT T T 1 1 1 1 1 0 5 1 1 5 2 2 5 3 3 5 4 4 5 time s AIR CROSS CORRELATION T T x 1 1 1 1 1 1 0 5 1 1 5 2 2 5 3 3 5 4 4 5 time s FIGURE 7 17 PLOTS OF THE VOLUME FRACTION PROFILES AT THE TOP AND BOTTOM MEASUREMENT PLANES FOR TWO 0 04 VOLUME FRACTION AIR BUBBLES MOVING AT 1m s AND 0 2m s RESPECTIVELY INCLUDED IS THE CROSS CORRELATION FUNCTION OF THE TWO PLOTS 102 However it is not that the cross correlation function responds only to the slower bubble but instead it responds to the bubble with the larger response To prove this theory the above test was repeated with a 0 01 volume fraction air bubble and a 0 04 volume fraction air bubble The 0 01 volume fraction air bubble was set to move at 0 2m s and the 0 04 volume fraction air bubble was set to move at 1m s Since the 0 04 volume fraction air bubble produced a larger response than the 0 01 volume fraction air bubble the cross correlation algorithm determined the average velocity to be 1m s This is a favourable result since it is preferable for the algorithm to respond to the large masses rather than individual bubbles in order to achieve a better assessment of the mass flow rate The following section will verify the performance of the dual plane 16 electrode FDM impedance tomography system on th
324. n 0 RecLoadProgressBar Max 128 StandardiseData RecLoadProgress Visible false end standardise the data by subtracting the mean of each column from the readings in that column and dividing by the standard deviation for that column procedure TRecLoadDBase StandardiseData var rowindex colindex integer mean stddev double begin for colindex 1 to 128 do begin find the mean of the column and keep a record for future tests mean MeanColumn colindex meanrec colindex mean find the standard deviation of the column and keep a record for future tests stddev StdDevColumn colindex mean stddevrec colindex stddev II pre process the data by subtracting the mean and dividing by the standard deviation for rowindex 0 to traintotal testtotal 1 do InputData rowindex colindex InputData rowindex colindex mean stddev RecLoadProgressBar Steplt end 160 end calculate the standard devitation of a column of training data function TRecLoadDBase StdDevColumn colnum integer mean double double var i integer sum double begin sum 0 for i 0 to traintotal 1 do sum sum IntPower InputData i colnum 2 StdDevColumn sqrt sum traintotal 1 end calculate the mean of a column of training data function TRecLoadDBase MeanColumn colnum integer double var sum double i integer begin sum 0 for i 0 to traintotal 1
325. n and motion of multicomponent mixtures in process vessels using electrical impedance tomography Principles and process engineering applications Chem Eng Sci 48 1883 1897 Yuen E L Mann R York T A and Grieve B D 2001 Electrical Resistance Tomography ERT Imaging of a Metal Walled Solid Liquid Filter Proc 2 World Congr on Industrial Process Tomography Hannover pp 183 190 112 19 20 21 22 23 24 25 26 2T 28 29 30 31 32 33 34 35 36 37 38 39 40 41 Cilliers J J Xie W Neethling S J Randall E W and Wilkinson A J 2001 Electrical resistance tomography using a bi directional current pulse technique Meas Sci Technol 12 997 1001 van Weereld J J A Collie D A L and Player M A 2001 A fast resistance measurement system for impedance tomography using a bipolar DC pulse method Meas Sci Technol 12 1002 1011 Hoyle B S Bailey N J and Nooralahiyan A Y 1995 Performance of neural network in capacitance based tomographic process measurement systems Meas Control 28 109 112 West R M Tapp H S Spink D M Bennett M A and Williams R A 2001 Application specific optimization of regularization for electrical impedance tomography Meas Sci Technol 12 1050 1054 Ge S M and Yang W Q 2001 Filtering in Image Reconstruction for Electrical Capacitance Tomography Proc 2 World Congr on Industrial Process Tomography Hannover pp 41 47 Sun T D Mudde R Schouten J C Scarlett
326. n is zero for a large proportion of the time the mean absolute error is only calculated over the region where the bubble is within the measurement section so as to get a more representative result An additional complication arose as a result of the slippage between the stepper motor pulley and the drive belt Consequently the position of the desired volume fraction profile did not coincide with the measured volume fraction profile as can be seen in figure 7 6 on page 89 To quantify the amount of slippage between the drive belt and pulley a 0 04 volume fraction air bubble was positioned at the top of the rig and a mark was made on the belt and the pulley at the same point The flow simulation apparatus was then instructed to move the bubble down for one metre and then raise the bubble by one metre so that it would return to its original position Observation of the marks on the belt and pulley after the movement showed that the belt had slipped by on average 150mm with respect to the pulley It was also noted that the pulley had returned to its original position proving that the slippage was not the result of stepper motor slip Unfortunately the slippage between the pulley and drive belt meant that no valid information could be obtained regarding the time delay dynamic response of the dual plane impedance tomography system A volume fraction error could still be calculated provided software compensation corrected for the slippage To achieve a more rep
327. n measurement was started at the same time as when the valve on the regulator was opened Initially no bubbles are generated However once the sparger has filled with air a large surge of air bubbles is generated in the rig This initial surge can be seen in figure 7 14 which shows the measured air volume fraction profiles during the test As expected the surge is detected at the bottom measurement plane before being detected at the top measurement plane TOP SYSTEM VOLUME FRACTION MEASUREMENT volume fraction 96 time s BOTTOM SYSTEM VOLUME FRACTION MEASUREMENT volume fraction 96 T T T time s FIGURE 7 14 PLOT OF THE AIR VOLUME FRACTION MEASUREMENTS AT THE TOP AND BOTTOM MEASUREMENT PLANES FOR A HOMOGENOUS BUBBLE FLOW 99 Sangu uma tt YA Bangi tam nee ion oe bm n FIGURE 7 15 SCREEN CAPTURES OF A HOMOGENOUS BUBBLE FLOW WHERE FOR EACH SCREEN CAPTURE THE TOP FRAME IS THE TOP MEASUREMENT PLANE AND THE BOTTOM FRAME IS THE BOTTOM MEASUREMENT PLANE THESE IMAGE RECONSTRUCTIONS WERE PERFORMED ON LINE AND THEIR RESULTS DISPLAYED IN REAL TIME THE ORDER OF THE FRAMES IS LEFT TO RIGHT AND ONLY EVERY 50 FRAME IS CAPTURED AS BEFORE THE BLACK IS THE WATER PHASE THE WHITE IS THE AIR PHASE AND THE LIGHT GREY REPRESENTS THE AREA OUTSIDE THE PIPELINE fre unte Ya te ts Cm m fee meo m 1e rom
328. n page 178 The following section will describe the functionality of this form as well as the accompanying dynamic verification form Figure 7 5 on page 86 is a screen capture of the Real time sampling form and a full scale version of this screen capture can be found in Appendix N on page 208 The form is divided into two separate sections namely the reconstruction results and the configuration section To perform a dynamic test it is necessary to configure the system beforehand This requires specifying the neural networks to perform the reconstructions of the top and bottom systems the plane separation and the flowmeter factor The flowmeter factor a relates the true mean flow velocity Ua to the velocity measured using the cross u correlation flow measurement as 50 In certain configurations this factor can be obtained through calibration In situations where calibration cannot be performed certain analytical methods exist as low cost alternative to calculate the flowmeter factor 50 Tests have shown that accurate velocity prediction results are obtained for the laboratory scale rig with a unity flowmeter factor 85 Real time impedance tomography reconstruction B x Reconstruction results Configuration Top system Cross correlation flowmeter View results Air volume fraction Online 1 100 3 C Offline Top system test number 101 m 122 Bottom system test number Desired numb
329. nd deltainit strtofloat RecSetDeltalnit Text deltamax strtofloat RecSetDeltaMax Text learnrate strtofloat RecSetLearn Text momentum strtofloat RecSetMomentum Text numfails RecNumFails Value numhidden RecSetNumHidden Value maxepochs strtoint RecSetMaxEpochs Text II initialise the dynamically allocated memory OutputWeight nil HiddenWeight nil DiffOutputWeight nil OldOutputWeight nil HiddenWeight nil DiffHiddenWeight nil OldHiddenWeight nil OutputDeltaWeight nil HiddenDeltaWeight nil OldDiffHiddenWeight nil OldDiffOutputWeight nil end end PROGRAM LISTING OF NETWORKUNIT PAS This unit is used for the training of the neural networks A single layer feed forward neural network and a double layer feed forward neural network can be trained using this unit The network weights can be adjusted using either gradient descent of Resilient back propagation and early stopping is provided to prevent the network from overfitting the training data which would result in poor generalisation performance A model search feature is also provided to determine the optimal number of hidden layer neurons in a double layer network unit networkunit interface uses Windows Messages SysUtils Classes Graphics Controls Forms Dialogs ExtCtrls ComCtrls StdCtrls Buttons Math type TRecTrainNetwork class TForm Paneli TPanel Panel2 TPanel La
330. nder TObject private 192 desiredvolume array 1 6000 1 4 of real 1 desired volume fractions for both phases at the top and bottom system diameter of air bubble gravdiam real diameter of gravel bubble topairerror real error in air prediction for top system topgravelerror real error in gravel prediction for top system botairerror real error in air prediction for bottom system botgravelerror real gravel prediction for bottom system airtopstart integer parameters specifying width of desired air airtopstop integer and gravel pulse airbotstart integer airbotstop integer graveltopstart integer graveltopstop integer gravelbotstart integer gravelbotstop integer airleng real length of air bubble gravleng real length of gravel bubble uppersystop real position of electrodes for top system uppersysbot real lowersystop real lowersysbot real airtopmax integer parameters used fro shifting real time airtopdesiremax integer data to coincide with desired data to graveltopmax integer compensate for slippage on pulley graveltopdesiremax integer airbotmax integer airbotdesiremax integer gravelbotmax integer gravelbotdesiremax integer airtopoffset integer airbotoffset integer graveltopoffset integer gravelbotoffset integer calculate the position of the two bu
331. nder TObject procedure Viewdatabase1 Click Sender TObject procedure Writedatatoatextfile 1 Click Sender TObject procedure FormShow Sender TObject procedure FormClose Sender TObject var Action TCloseAction procedure RecSetTestNumberOKClick Sender TObject procedure RecSetTestNumberCanClick Sender TObject procedure Settestnumber1 Click Sender TObject procedure Setnetworktrainingparameters1Click Sender TObject procedure Initialisenetworkweights1 Click Sender TObject procedure RecWeightCanClick Sender TObject OutputNetwork array of array of real network predictions OutputHidden array of array of real hidden layer neuron outputs OutputWeight array of array of real weights for output layer DiffOutputWeight array of array of real delta values for the output layer OldOutputWeight array of array of real old output layer weights HiddenWeight array of array of real weights for hidden layer DiffHiddenWeight array of array of real delta values for the hidden layer OldHiddenWeight array of array of real old hidden layer weights OutputDeltaWeight array of array of real output layer delta values for RPROP HiddenDeltaWeight array of array of real hidden layer delta values for RPROP OldDiffHiddenWeight array of array of real old output delta values for RPROP OldDiffOutputWeight array of array of real old hidden delta values for
332. ndex 2 gt outputstore testposition prediction index 1 and outputstore testposition prediction index 2 outputstore testposition prediction index then topphasepred air else if outputstore testposition prediction index outputstore testposition prediction index1 and outputstore testposition prediction index gt outputstore testposition prediction index 2 then topphasepred water else topphasepred gravel end else begin RecLoadDBase GetPhases containsair containsgravel if outputstore testposition prediction index gt 0 then begin if containsair then topphasepred air else topphasepred gravel end else topphasepred water end topoutputpred col row topphasepred if topphasepred air then volumestore testposition 1 volumestore testposition 1 1 else if topphasepred gravel then volumestore testposition 2 volumestore testposition 2 1 if totaloutputs 264 then begin if outputstore testposition prediction index 2 264 gt outputstore testposition prediction index 1 264 and outputstore testposition prediction index 2 264 outputstore testposition prediction index 264 then bottomphasepred air else if outputstore testposition prediction index 264 gt outputstore testposition prediction index 1 264 and outputstore testposition prediction index 264 gt outputstore testposition prediction index 2 264 then bottomphasepred water else bottomphasepred
333. necessary to override the default Windows message handling procedure WndProc Essentially if the message relates to some standard Windows functionality then the original WndProc procedure is called However if the message identification value matches that returned by InitSample then the sampling is complete and StopSample is called StopSample returns the 555 to the reset state and converts the raw binary data stored in the buffer into actual voltage readings in the range from 5V to 5V These voltage readings are then returned to the calling application This concludes the capture of the 50 frames When the program exits ReleaseSample must be called to de allocate the resources reserved for sampling Tests were conducted to determine the on line data capture rate of the sampling unit Although the sample controller board is designed to capture data at 80ksamples s corresponding to 312frames s from both systems the overhead of processing and storing of results reduced the on line frame rate to approximately 280frames s However this still meets the required specification of 200frames s from both systems 60 5 3 2 Training and test database generation software description As mentioned previously it is necessary to capture a large quantity of representative data in order to train a neural network A data capture application was developed for the task of generating this training database The complete program code for this applicati
334. nregions then begin region 1 inc numrepeat end end end region 1 numrepeat 0 while rowindex lt traintotal testtotal do begin for colindex 1 to 128 do begin driftcomp teststopvalues region colindex teststartvalues region colindex rowindex teststartindex region numrepeat RecTestStop Value traintotal 1 l teststopindex region teststartindex region 1 teststartvalues region colindex InputData rowindex colindex InputData rowindex colindex driftcomp end inc rowindex RecLoadProgressBar Steplt if rowindex teststopindex region numrepeat RecTestStop Value traintotal then begin inc region if region gt numtestregions then begin region 1 inc numrepeat end end end end not all the data that was loaded can be used for a two phase measurement therefore this routine effectively copies over the data that was flagged as incorrect phase during the data loading procedure procedure TRecLoadDBase ReOrganiseData var origindex newindex colindex temptraintotal temptesttotal integer begin newindex 0 temptraintotal 0 temptesttotal 0 RecLoadProgressBar Max traintotal testtotal RecLoadProgressBar Position 0 for origindex 0 to traintotal testtotal 1 do begin II if image reconstruction and valid test case then copy the data across and increment newindex II if volume fraction reconstruction and valid test case then copy the data across
335. nstartvalues row col trainnumvrs trainstopvalues row col trainstopvalues row col trainnumvrs teststartvalues row col teststartvalues row col testnumvrs teststopvalues row col teststopvalues row col testnumvrs end 11 set the sizes of the matrices according to the desired output if RecDesiredRecon ItemIndex 0 then begin I three phase measurement requires 1 of C encoding for Il each output pixel therefore 264 outputs if RecDesirePhase ItemIndex 0 then begin SetLength OutputData traintotal testtotal 264 SetLength OutputNetwork traintotal testtotal 264 numoutputs 264 threephase true containsair true containsgravel true S2 for a three phase measurement end two phase image reconstruction only requires a single output per pixel therefore 88 outputs else begin SetLength OutputData traintotal testtotal 88 SetLength OutputNetwork traintotal testtotal 88 numoutputs 88 if RecDesirePhase ItemIndex 1 then begin containsair true S2 for an air water two phase measurement end else begin containsgravel true S2 for a gravel water two phase measurement end end end volume fraction reconstruction requires two outputs one for air and one for gravel else begin SetLength OutputData traintotal testtotal 2 SetLength OutputNetwork traintotal testtotal 88 numoutputs 2 end RecLoadProgress Visible true RecLoadProgress Update RecLoa
336. ntroller It also provides three 8 bit bi directional digital input output ports The voltages are sampled using single ended input sampling Although this causes the sampled readings to be more sensitive to noise the lead lengths connecting the PC30G to the multiplexing board were kept short so as to minimise this effect In addition analogue ground was isolated from digital ground plane noise through an inductor Two DG506A 16 input CMOS analogue multiplexers were used to multiplex the 16 capacitance and 16 conductance voltages The address lines of the multiplexers were driven directly by digital output port A of the PC30G card In addition the enable lines of the analogue multiplexers were driven by a digital output of the card The two analogue outputs of the multiplexers were connected to analogue input channel 0 and 1 respectively A complete frame of voltages was captured as follows 1 Port A was set under software control so as to access address 0 on the multiplexers and the multiplexers were enabled 2 After a delay the voltages on channel 0 and channel 1 were sampled under software control and stored in the corresponding locations of the complete frame 3 Port A was then incremented to access the next address and the process was repeated until a complete frame of capacitance and conductance readings had been captured The circuit diagram of the multiplexing card can be found in Appendix B on page 121 A minimum bubble diameter c
337. nultem Currentsamples1 TMenultem Paneli TPanel Session TSession SampAliasOpen TPanel Label TLabel SampAliasList TComboBox SampAliasNew TPanel Label2 TLabel SampAliasNewName TEdit SampAliasNameOK TBitBtn SampAliasNewCanc TBitBtn SampAliasOpenCanc TBitBtn Setthenumberofframes1 TMenultem SampNumFrames TPanel Label3 TLabel SampNumFramesValue TSpinEdit SampNumFramesOK TBitBtn SampNumFramesCan TBitBtn procedure Exitt Click Sender TObject procedure Opendatabase1 Click Sender TObject procedure SampAliasListClick Sender TObject procedure Closedatabase1 Click Sender TObject procedure Createdatabase1 Click Sender TObject procedure SampAliasNameOKClick Sender TObject procedure SampAliasNewCancClick Sender TObject procedure SampAliasOpenCancClick Sender TObject procedure Startsampling1 Click Sender TObject procedure FormCreate Sender TObject procedure FormClose Sender TObject var Action TCloseAction procedure Database1 Click Sender TObject procedure Currentsamples1 Click Sender TObject procedure SampNumFramesOKClick Sender TObject procedure SampNumFramesCanClick Sender TObject procedure Setthenumberofframes1 Click Sender TObject private Private declarations dbasename string store the name of the current database so that it can be written to the registry on exit procedure CreateTables aliasname string II init
338. numbottomhidden 0 then begin if rowindex 0 then tempvalue 1 else tempvalue inputstore reconposition readings rowindex 128 end else tempvalue bottomhiddenoutputs rowindex weightsum weightsum bottomweights rowindex colindex tempvalue end if weightsum gt 200 then weightsum 200 else if weightsum lt 200 then weightsum 200 three phase image reconstruction if totaloutputs 264 then begin if colindex 1 mod 3 lt gt 0 then weightstore colindex 1 mod 3 weightsum else begin for i 2 downto 0 do outputstore reconposition prediction colindex i 265 RecTrainNetwork SoftMax 3 i weightstore 1 weightstore 2 weightsum end end two phase image reconstruction else if totaloutputs 88 then begin outputstore reconposition prediction colindex 89 tanh weightsum end three phase volume fraction prediction else begin if colindex 3 3 then begin if weightsum gt botairthreshold then outputstore reconposition prediction 3 weightsum else outputstore reconposition prediction 3 0 end else begin if weightsum gt botgravthreshold then outputstore reconposition prediction 4 weightsum else outputstore reconposition prediction 4 0 end end end copy the sample time across from the inputstore to keep as record for 182 analysis outputstore reconposition sampletime inputstore reconposition sampletime if wish to display the resu
339. oad TButton SaveDialog TSaveDialog procedure GenerateClick Sender TObject generate code sequence procedure CompileClick Sender TObject compile code sequence procedure DownloadClick Sender TObject download code sequence private Private declarations public Public declarations end var AutoMain TAutoMain implementation R DFM generate sequnce of instructions to achieve a specific division factor the number of included NOP instructions sets this factor The first three 1 instructions also require two NOP s to compensate for the instruction 1 cycles to loop back to the start after the fourth instruction this ensures that the duty cycle is 50 and 90 degrees of phase shift is I achieved procedure TAutoMain GenerateClick Sender TObject var tempfile textfile numextra i integer begin if SaveDialog Execute then begin calculate the number of extra NOP s required to achieve a division factor numextra DivSelect Value div 16 4 AssignFile tempfile SaveDialog FileName ReWrite tempfile configure PORTB as an output WriteLn tempfile 9 INCLUDE PIC1684 SRC WriteLn tempfile 9 BSF STATUS RPO WriteLn tempfile 9 CLRF TRISB WriteLn tempfile 9 BCF STATUS RPO WriteLn tempfile LOOP 9 MOVLW 00000010 WriteLn tempfile 9 WriteLn tempfile 9 WriteLn tempfile 9 for i 1 to numextra WriteLn tempfile WriteLn tempfil
340. oardnum EDR FreeBoardHandle boardnum end start sampling by reseting the 555 and counter instructing the EDR software to capture a specified number of samples and then putting the 555 in run mode procedure TSamp StartSample nmfrmes integer var i integer begin busy sampling true numframes nmfrmes reset 555 and couter EDR DlOPortOutput boardnum 0 rst555 EDR DlOPortOutput boardnum O rstclk EDR DlOPortOutput boardnum 0 rst555 acquire a block of AD data using streaming and place the data IL into buffer 149 EDR ADInBinBackground boardnum 256 numframes buffer 0 I create a delay while the Eagle card is configured fori 1 to 1000 do EDR DlOPortOutput boardnum 0 rst555 start the sampling by putting the 555 in run mode EDR DlOPortOutput boardnum 0 std555 end stop the sampling by calling EDR_StopBackroundADIn reseting the 555 and counter and process raw binary data into voltages in the range from 5N to 5V these voltages are then stored in the result array procedure TSamp StopSample var result array of double var temp double i integer begin EDR StopBackgroundADIn boardnum EDR DlOPortOutput boardnum 0 rst555 EDR_DIOPortOutput boardnum 0 rstclk EDR DlOPortOutput boardnum 0 rst555 for i 0 to 256 numframes 1 do begin temp buffer i AND datamask temp temp 2048 5 2048 result i temp end busy sampling false end get status of
341. ocedure WndProc var TheMsg TMessage override override the standard message handling procedure to catch the notification message 1 indicating completion of sampling end var CurMain TCurMain implementation uses mainunit addunit R DFM tM GENERAL PROCEDURES this procedure is called whenever the form receives a message from Windows If the message indicates that sampling is complete then stop the sampling and process the readings If not then handle the message in the default way procedure TCurMain WndProc var TheMsg TMessage begin if TheMsg Msg notifvalue then begin SampHardware StopSample result ProcessResults end else inherited WndProc TheMsg end disable the timer and exit procedure TCurMain CurExitClick Sender TObject begin DriftTimer Enabled false CurTimer Enabled false Close end setup the tables and enable the timer procedure TCurMain FormShow Sender TObject begin CurCapTable Col 1 CurCapTable Row 1 CurResTable Cells 0 1 TxA CurResTable Cells 0 2 TxB CurResTable Cells 0 3 CurResTable Cells 0 4 TxD CurResTable Cells 0 5 CurResTable Cells 0 6 TxF CurResTable Cells 0 7 TxG CurResTable Cells 0 8 TxH CurResTable Cells 1 0 RxA CurResTable Cells 2 0 RxB CurResTable Cells 3 0 RxC CurResTable Cells 4 0 Rx
342. oding Comparison between the performance of the frequency division multiplexed impedance tomography system and a standard time division multiplexed impedance tomography system for a three phase air gravel seawater volume fraction prediction using a double layer feed forward neural network Transmitter frequencies generated using two PIC16F84 microcontrollers with 4 43361MHz and 5MHz crystals respectively Transmitter frequencies selected using PBIL optimisation algorithm with the crystal frequencies and division factors required to generate these frequencies Comparison between the drift of the capacitance and conductance readings over a 15 hour period using different conductance probe materials Comparison between the performance of the 16 electrode frequency division multiplexed impedance tomography system and the 8 electrode prototype frequency division multiplexed impedance tomography system for a two phase air water image reconstruction using a single layer feed forward neural network Performance of the 16 electrode frequency division multiplexed impedance tomography system for a two phase air water volume fraction prediction using a single layer feed forward neural network Comparison between the performance of the 16 electrode frequency division multiplexed impedance tomography system and the 8 electrode prototype frequency division multiplexed impedance tomography system for a three phase air gravel water image reconstruction using a sin
343. odules The numbers next to each connection specify the number of signal lines between the corresponding modules 39 TO PC30G EAGLE DATA VOLTAGE TO SECOND ACQUISITION Pes A SAMPLE MUX MEASUREMENT CARD CONTROLLER OUTPUT SECTION VOLTAGES MUX OUTPUT VOLTAGES FREQUENCY GENERATOR REFERENCE SIGNALS CONDUCTANCE RxA CAPACITANCE RxA SYNCHRONOUS DETECTOR BOARD CONDUCTANCE RxB CAPACITANCE RxB Mea eT CONDUCTANCE RxC CAPACITANCE RxC SYNCHRONOUS DETECTOR BOARD CONDUCTANCE RxD CAPACITANCE RxD LIENS CONDUCTANCE RxE CAPACITANCE RxE SYNCHRONOUS DETECTOR BOARD CONDUCTANCE RxF CAPACITANCE RxF ee ie CONDUCTANCE RxG CAPACITANCE RxG SYNCHRONOUS DETECTOR BOARD CONDUCTANCE RxH CAPACITANCE RxH FIGURE 5 7 BLOCK DIAGRAM ILLUSTRATING THE INTERACTIONS BETWEEN THE DIFFERENT CIRCUIT BOARDS IN A 16 ELECTRODE FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM WHERE THE NUMBER NEXT TO EACH LINE INDICATES THE NUMBER OF SIGNALS 40 The operation and design of each of these modules will be explained in detail in the following sections 5 2 1 Output transmitter details As mentioned previously a LM675 power op amp was used as the output transmitter for the 8 electrode prototype Research was done at the Cape Technikon into the use of MOSFET drivers instead of power op amps and a design was proposed consisting of a TC4427 dual high speed power MOSFET dr
344. of eight transmitter frequencies that minimise the output ripple The accompanying program code can be found in Appendix J on page 138 54 The only required user input is the selection of the magnitude threshold that determines which frequency harmonics to consider Due to the simplicity of the simulation used to model the system behaviour this magnitude threshold cannot be equated to a measurable signal level and therefore does not have a unit Instead it is found through experimentation By observing the frequency components on the output of the low pass filters with a spectrum analyser it was possible to verify the simulation results Specifically tests revealed that for a threshold value of 0 0001 all frequency components observed with the spectrum analyser were successfully predicted by the simulation software However the best result achieved using the PBIL optimisation algorithm for this threshold value was a minimum frequency component at 35Hz on the low pass filter outputs Since this is within the passband of the system output ripple would be observed By increasing the threshold value to 0 01 the minimum frequency component increased but not all the frequency components were predicted Subsequently it was decided that a compromise should be reached by using a threshold value of 0 01 Specifically all the major frequency components will be detected and should be outside the passband of the low pass filters The drawback is that minor fre
345. of test cases by moving to the last element in the database and extracting the test number currenttest integer InpTable FieldByName TestNo Value AddInputTomo ResetDesired if currenttest lt gt 0 then begin InpTable FindKey currenttest 0 001 OutTable FindKey currenttest 0 001 display the desired output and update the tables DisplayData UpdateTables end StatusUpdate Enabled true end reset the InputTomo component procedure TAddToDatabase AddResetClick Sender TObject begin AddinputTomo ResetDesired AddInputTomo Repaint end when click the left arrow move to the previous test case in the database and display the results in the tables and the desired output using the InputTomo component procedure TAddToDatabase AddPreviousClick Sender TObject begin if currenttest gt 1 then begin dec currenttest InpTable FindKey currenttest 0 001 OutTable FindKey currenttest 0 001 DisplayData UpdateTables end end when click the right arrow either move to the next element in the database or create a complete new test case depending on the position of the cursor in the database procedure TAddToDatabase AddNextClick Sender TObject var vers col row integer begin if currenttest lt gt 0 currenttest lt gt InpTable RecordCount 1 div numframes then begin if cursor not at the end of the database then simply move to the next element in the database and
346. om gravel fraction air correlation gravel correlation if FileCustomName Checked then begin SaveResultsDialog Execute filename SaveResultsDialog FileName end else begin 200 if strtofloat StepperConfig AirVelocity Text 0 then WriteLn tempfile flowmeter factor FlowFactorlnput Text tempstring inttostr airposition WriteLn tempfile top air threshold TopAirThresh Text inttostr floor 10 strtofloat StepperConfig AirVelocity Text a WriteLn tempfile top gravel threshold TopGravThresh Text else WriteLn tempfile bottom air threshold BotAirThresh Text tempstring inttostr gravelposition WriteLn tempfile bottom gravel threshold BotGravThresh Text inttostr floor 10 strtofloat StepperConfig GravelVelocity Text g WriteLn tempfile correlation threshold CorrThreshinp Text WriteLn tempfile framerate frmrate Caption filename c TestData dyn tempstring txt WriteLn tempfile actual air velocity AirVelActual Caption end WriteLn tempfile actual gravel velocity GravelVelActual Caption AssignFile tempfile filename WriteLn tempfile predicted air velocity AirVelPredict Caption Rewrite tempfile WriteLn tempfile predicted gravel velocity GravelVelPredict Caption WriteLn tempfile RealMain DesiredFrames Value end for row 1 to RealMain DesiredFrames Value do WriteLn tempfile flow direction begin StepperConfig FlowDirSe
347. om several input variables to several output variables 29 Research into the use of neural networks for image reconstruction in capacitance tomography has been done 21 24 30 31 although the previous work has predominantly made use of simulated capacitance readings obtained using a finite element modelling technique The same is also true for resistance tomography systems 32 33 whereas the author s previous research concentrated more on the practical application of neural networks to an existing impedance tomography system 3 4 Multi layer perceptrons and radial basis function neural networks were analysed for the reconstruction task Tests revealed that the multi layer perceptron consistently out performed the radial basis function neural network because of the limited amount of training data available Hence only the multi layer perceptron was implemented in the new system software and will be discussed in the following section 3 2 1 Multi layer perceptron fundamentals A multi layer perceptron is a densely interconnected layered network of neurons 29 Each neuron consists of a combination function which forms the weighted sum of the neuron inputs and an activation function which introduces a non linearity into the network and confines the output to lie within a specific range To perform image reconstruction each pixel in the image can be considered the output of an individual neural network performing a classification function
348. on In contrast to the previous reconstruction neural networks examined an additional hidden layer of neurons is provided Specifically tests concluded that the optimal number of hidden layer neurons for this application is 25 An offset neuron is also provided for the hidden layer As before the neural network has a fully connected structure between each neuron in the input layer and each neuron in the hidden layer and between each neuron in the hidden layer and each of the two output neurons These output neurons predict the air and gravel volume fractions directly An additional advantage of predicting the volume fractions directly is a reduction in the computation time of the reconstruction DATA ACQUISITION SYSTEM OUTPUT LAYER NEURONS HIDDEN LAYER NEURONS INPUT LAYER NEURONS FIGURE 3 3 DIAGRAM ILLUSTRATING THE USE OF A DOUBLE LAYER FEED FORWARD NEURAL NETWORK TO PERFORM A THREE PHASE AIR GRAVEL SEAWATER VOLUME FRACTION PREDICTION IN CONTRAST TO THE PREVIOUS RECONSTRUCTION NEURAL NETWORKS EXAMINED AN ADDITIONAL HIDDEN LAYER OF NEURONS IS PROVIDED THESE HIDDEN LAYER NEURONS FORM THE WEIGHTED SUM OF THE INPUT LAYER NEURON OUTPUTS AND THE OUTPUT OF EACH NEURON IS THE HYPERBOLIC TANGENT OF THIS WEIGHTED SUM THE 57 INPUT LAYER NEURONS ARE FULLY CONNECTED TO THE 25 HIDDEN LAYER NEURONS INCLUDED IN THE HIDDEN LAYER IS ANOTHER OFFSET NEURON WHOSE OUTPUT EQUALS UNITY THESE HIDDEN LAYER NEURONS ARE FULLY CONNECTED TO THE TWO OUTPUT LAYER NEURON
349. on TEdit GravelBubbleStartPosition TEdit FileCustomName TCheckBox Label23 TLabel TopAirPeak TLabel BotAirPeak TLabel TopGravPeak TLabel BotGravPeak TLabel procedure AirShape1MouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Integer procedure AirShape2MouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Integer procedure RealExitClick Sender TObject procedure AirShape3MouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Integer procedure AirShape4MouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Integer procedure AirShape5MouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Integer procedure GravShapet MouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Integer procedure GravShape2MouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Integer procedure GravShape3MouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Integer procedure GravShape4MouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Integer procedure GravShape5MouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Integer procedure FormShow Sender TObject procedure RealSaveResultsClick Sender TObject procedure CalcDesiredClick Sender TObject procedure CalcVolErrorClick Sender TObject procedure CalcActualVelClick Se
350. on flow measurement is based on the principle of tagging disturbances in the flow generated through turbulence or suspended particles 12 If the disturbance detected by the upstream sensor reappears at the downstream sensor after a period of t seconds and the separation between the two sensors is known to be L then the velocity of the disturbance can be calculated as YA The time is calculated as the peak of the cross correlation function The present system is required to measure component velocities up to 20m s Based on this maximum velocity specification it was determined that an on line frame rate of 200frames s from both planes simultaneously is required for the intended application However a literature study revealed that none of the commercially available tomography systems could achieve this frame rate Further no results had been published to date of any research systems that could achieve a 200frames s frame rate for a dual plane 16 electrode impedance tomography system Consequently it was necessary to design a new impedance tomography data capture technique to address this problem To understand the potential speed advantages of the new data capture technique it is necessary to briefly examine standard tomography data acquisition systems Essentially standard tomography systems employ time division multiplexing TDM In other words each electrode pair has a set time during which the corresponding capacitance or
351. on is provided in Appendix K on page 142 with comments discussing the functionality of each of its procedures Although the fundamental concepts of this program are the same as those used for previous research various modifications were made in an attempt to reduce the time taken to generate a complete training database An example of such a modification was the extensive use of the registry to save the program configuration Although the time taken to configure the program for an individual test may not be that significant the advantage achieved through saving the program state is definitely observed when the same or similar test must be repeated The training and test data are stored in databases instead of individual text files This has the obvious advantage of encapsulating all the training or test data results into a minimum number of files In addition the use of databases simplifies the task of accessing a particular test point or number of test points and speeds up the loading of complete databases into program memory The software used to add entries to the database is also more flexible through the use of databases It is possible to edit entries overwrite entries or completely remove entries If a mistake was made during the training database generation process then that particular entry can be overwritten at a later stage If a similar mistake were made with the previous system s software then the training database generation process would ty
352. on stage the neural networks developed previously will be used In addition the sensors were constructed according to standard design principles Hence this thesis concentrates on the data acquisition stage and specifically the development of a high speed data acquisition system capable of operating at the frame rate necessary for the intended application A laboratory scale flowmeter is developed to prove the operation of the concept Consequently no effort has been made to ensure that the system can be used in an industrial environment although recommendations are made for the construction of such an industrial instrument In addition the dynamic performance of the rig will only be tested on simulated air or gravel masses Since the seawater will be static no attempt will be made to measure the mass flow rate of the seawater component A scaled down version of this system could be tested in a closed loop pump circuit at a later stage A high speed camera would permit verification of the measurement system results To gain a better understanding of how the new data acquisition system achieves a much higher frame rate than standard tomography systems it is necessary to first examine the operation of standard capacitance and resistance tomography systems and this is done in Chapter 2 Central to tomography systems is the reconstruction stage and Chapter 3 examines standard reconstruction techniques and highlights the major results of the author s pr
353. one on this topic before an industrial prototype is developed Further it is believed that in the intended application the constant material flow through the measurement section would assist in keeping the conductance probes clear of bubbles and clear away corrosion deposits Subsequent work by the author s colleague Mr A Giannopoulos in which the water is circulated in a static rig by means of a pump has shown this assumption to be correct 62 6 2 Training and Test Database Generation Training a neural network requires the generation of a representative training database As mentioned previously a bubble placement system developed as part of the author s previous research provided a convenient way to systematically vary the contents of the vessel The bubble placement system consists of a bed of nails placed at the base of the vessel and a clear acrylic lid Whereas in previous work long cylindrical bubbles were used that extended the full length of the vessel it was decided that shorter bubbles with rounded ends would be more realistic Consequently it was necessary to change the concept of a bed of nails into a bed of hooks so that the air bubbles could be anchored at specific positions within the measurement section remembering that an air bubble is actually a contiguous polystyrene foam mass The blanking off disk mentioned previously was used to provide this array of anchor points corresponding to specific positions within the vess
354. ons is not that significant since it is possible to perform a reconstruction of the top system using a neural network trained for the bottom system and vice versa although the reconstructed images may not be accurate The following chapter examines the results of using neural networks trained with static data to reconstruct dynamic flow situations 77 CHAPTER 7 DYNAMI C SI TUATI ON RESULTS FOR THE DUAL PLANE 16 ELECTRODE IMPEDANCE TOMOGRAPHY CROSS CORRELATI ON SYSTEM The theory of cross correlation flowmeters is briefly discussed in the following section before the dynamic performance of the dual plane 16 electrode FDM impedance tomography system is analysed 7 1 Cross correlation Flowmeter Fundamentals As mentioned previously cross correlation flow measurement is based on obtaining the cross correlation function of the signals obtained from two sensors spaced axially along the pipe The time delay corresponding to the maximum value of the cross correlation function gives an indication of the transit time between the two sensors The cross correlation function of the value of x t at time t at the upstream sensor and the value of y t at time t at the downstream sensor is calculated as follows 64 li T Ry 7 7 1 0 The cross spectral density function S f is the Fourier transform of the cross correlation function and is calculated as follows 64 S t Ryte Pr de 7 2 and is useful for providing a
355. or between the actual velocity and the measured velocity As expected the mean absolute error increases as the velocity of the bubble increases This is a result of the reduction in the transit time between the top and bottom measurement planes Since the sampling interval is fixed the accuracy with which this transit time can be determined is reduced as the number of frames delay determined by the cross correlation algorithm decreases In addition it was observed that the slippage between the drive belt and the stepper motor pulley increased as the flow velocity increased 0 2 average measured velocity desired velocity 0 18 mean absolute error 0 16 0 14 0 12 0 1 0 08 measured velocity m s mean absolute error m s 0 06 0 04 0 02 0 1 2 3 4 5 actual velocity m s FIGURE 7 16 CROSS CORRELATION VELOCITY MEASUREMENT ACCURACY FOR THE 0 01 VOLUME FRACTION AIR BUBBLE The theoretical velocity discrimination of this system can be calculated using equation 4 1 on page 18 in Chapter 4 as AV 7 12 Vax 201 AY sets 1 5 7 13 Vmax 2 280 frames s 0 50 m From figure 7 16 the velocity discrimination at the maximum simulated velocity of 4 2m s is 98 3 1 which is comparable to the theoretical limit as calculated using equation 7 12 Having determined the accuracy of the cross correlation algorithm for the simple case of a single bubble moving
356. ork Repaint end end completion of sampling process the results so that only the voltages from the desired system are obtained and then pre process the data according to the same rules applied to the training database Then calculate the network output and plot the network prediction procedure TRecTest ProcessResults var col row index integer begin data was captured as test case if not curcalibrate then begin col 0 row 0 index 1 I sort the voltages depending on which is the desired system if RecLoadDBase GetTomoSystem 1 then col col 8 while col lt 255 do begin testdata index result col inc row inc index if row 8 then begin row 0 inc col 9 end else inc col end testdata 0 1 remove the calibration offset if the network was trained using I calibrated data if RecLoadDBase RecLoadCalibrate Checked then for col 1 to 128 do begin testdata col testdata col calibval col end standardise the data by subtracting the mean and dividing by the standard deviation for col 1 to 128 do begin testdata col testdata col meanrec col stddevrec col end calculate the network output and plot the results CalculateTestCase PlotTestOutput end else capturing a set of data to provide a calibration point so that tests be offset according to this calibration else begin curcalibrate false col 0 row 0 in
357. orresponding to 20 of the pipeline diameter was used in the author s previous research since capacitance tomography systems typically produce low resolution images and the readings were subject to measurement noise and drift To ensure a consistent comparison the training database for the FDM impedance tomography system was generated using the same bubble positions and minimum bubble sizes namely 20 40 and 50 of the pipeline diameter respectively for both air and gravel phases In addition the software developed previously was used to generate training and test databases and train neural networks to perform the reconstruction for the FDM impedance tomography system The only modifications included changing the data capture software to capture readings from the PC30G card instead of the serial port and modifying the neural networks to have 32 inputs instead of the previous 56 Since the program code is effectively the same as that developed previously it will not be examined here and has not been included in the appendices Instead the neural network reconstruction results will be examined and a comparison to the previous results achieved using a TDM impedance tomography system will be provided It is important to note that a consistent comparison was performed between the TDM impedance tomography system and the FDM impedance tomography system since the same pipeline section and electrode array were used for both sets of tests Further both sy
358. ortant variables upperindex 128 currthresh 0 currsum 0 totalerrors 0 watererror 0 gravelerror 0 airerror 0 if there is a hidden layer then propagate the inputs through the network calculating first the outputs from the hidden layer and then based on these outputs calculate the final network outputs for count start to stop do begin if numhidden gt 0 then begin calculate the outputs of the hidden layer neurons where the hidden layer neurons are simple hyperbolic tangents upperindex numhidden for colindex 0 to numhidden do begin if colindex 0 then OutputHidden count colindex 1 bias term else begin weightsum 0 calculate the weighted sum on the input to the neuron for rowindex 0 to 128 do weightsum weightsum HiddenWeight rowindex colindex 1 InputData count rowindex prevent arithmetic overflow if weightsum gt 200 then weightsum 200 else if weightsum lt 200 then weightsum 200 tanh activation function OutputHidden count colindex tanh weightsum end end end calculate network output totalgravelpixels 0 totalairpixels 0 for colindex 0 to numoutputs 1 do begin weightsum 0 calculate the weighted sum on the input to the output neuron for rowindex 0 to upperindex do begin if numhidden gt 0 then tempvalue OutputHidden count rowindex else tempvalue InputData count rowindex weightsum
359. otairerror airbotstop airbotstart 1 end else begin BotAirDelay Caption unknown for testcase gravelbotstart to gravelbotstop do begin if totaloutputs 2 then botairerror botairerror abs 0 outputstore testcase gravelbotoffset prediction 3 else botairerror botairerror abs 0 volumestore testcase gravelbotoffset 3 end botairerror botairerror gravelbotstop gravelbotstart 1 end BotAirVolError Caption floattostrf botairerror ffFixed 4 3 botgravelerror 0 if gravelbotdesiremax lt gt 1 then begin BotGravDelay Caption floattostrf RealMain frameinterval gravelbotoffset ffFixed 4 3 for testcase gravelbotstart to gravelbotstop do begin if totaloutputs 2 then botgravelerror botgravelerror abs desiredvolume testcase 4 outputstore testcase gravelbotoffset prediction 4 else botgravelerror botgravelerror abs desiredvolume testcase 4 volumestore testcase gravelbotoffset 4 end botgravelerror botgravelerror gravelbotstop gravelbotstart 1 end else begin BotGravDelay Caption unknown 198 for testcase airbotstart to airbotstop do begin if totaloutputs 2 then botgravelerror botgravelerror abs 0 outputstore testcase airbotoffset prediction 4 else botgravelerror botgravelerror abs 0 volumestore testcase airbotoffset 4 end botgravelerror botgravelerror airbotstop airbotstart 1 end BotGravVolError Caption floattostrf botgravelerror ffFixed 4 3
360. otal RecTrainStop Value RecTrainStart Value 1 RecTrainVers Value testtotal RecTestStop Value RecTestStart Value 1 RecTestVers Value II initialise matrices into which data will be loaded SetLength InputData traintotal testtotal 129 SetLength VolumeData traintotal testtotal 3 threephase false containsair false containsgravel false S2 for row 1 to 5 do begin for col 1 to 128 do begin trainstartvalues row col 0 trainstopvalues row col 0 teststartvalues row col 0 teststopvalues row col 0 end end calculate the maximum number of training and testing versions TrainInpTable Last trainnumcases integer TrainInpTable FieldByName TestNo Value trainnumvrs 1000 TrainInpTable FieldByName TestNo Value trainnumcases TestlnpTable Last testnumcases integer TestInpTable FieldByName TestNo Value testnumvrs 1000 TestInpTable FieldByName TestNo Value testnumcases II load the water only endpoints to provide calibration 157 I for the start of the training data for traincalib 1 to numtrainregions do begin for trainversions 1 to trainnumvrs do begin index trainstartindex traincalib trainversions 1000 TrainInpTable FindKey index for col 1 to 128 do begin 51 Voltage 4inttostr col trainstartvalues traincalib col trainstartvalues traincalib col TrainInpTable FieldByName s1 Value end end
361. otalairpixels real number of test air pixels totaltrainwaterpixels real number of train water pixels totaltraingravelpixels real number of train gravel pixels totaltrainairpixels real number of train air pixels containsgravel Boolean whether contains gravel containsair Boolean whether contains air totaltrain integer total number of training cases totaltest integer total number of test cases maxepochs integer maximum number of epochs maxfails integer maximum number of fails before stop network training if using early stopping updatefreq integer rate at which display updates earlystop Boolean 11 perform early stopping useRPROP Boolean train using RPROP currthresh real current test threshold error currsum real current test sum void error currtrainthresh real current train threshold error currtrainsum real current train sum void error maxdelta real maximum delta value mindelta real minimum delta value watererror real current test water volume error gravelerror real current test gravel volume error airerror real current test air volume error numbest integer number of hidden layer neurons for best performance currentfails integer current number of fails currepoch integer current training epoch bestthresh real threshold corresponding to the best network b
362. ounted on the outside of the measurement electrodes is used to reject external noise 51 Radial shields isolate the measurement electrodes to ensure that the inter electrode capacitances external to the vessel are minimal 46 They also have the advantage of reducing the standing capacitances of the adjacent electrode pairs thus reducing the required dynamic range of the capacitance sensing circuitry 47 56 By analysing the measurement combinations of a FDM impedance tomography protocol it can be seen that a significant proportion of the measurements are for adjacent transmitter receiver electrode pairs The radial shields ensure that these capacitance measurements do take place through the vessel Figure 5 6 illustrates the outer earthed screen as well as the projected radial shields passing between the capacitance plates Appendix D on page 129 contains tables of the capacitance and conductance readings of the different electrode combinations in the laboratory scale rig 5 2 Electronic Design of a 16 electrode Frequency Division Multiplexed Impedance Tomography System Due to the high level of parallelism in the FDM impedance tomography technique it was decided that the 16 electrode system should be broken down into separate fundamental modules where each module is then replicated according to the number of electrodes in the system Figure 5 7 on page 40 is a block diagram illustrating the interaction of these individual m
363. output nil end capture set of 50 frames to provide calibration point corresponding to the rig full of water procedure TRecTest StartCalibrateClick Sender TObject begin RecMain Savecalibrationdata1 Enabled true curcalibrate true SampHardware StartSample 50 end end PROGRAM LISTING OF REALUNIT PAS This unit implements the real time sampling reconstruction and cross correlation calculation that would be used in an on line flow monitoring application To begin a test the user specifies which network he wishes to use for the top system and which network he wishes to use for the bottom system The application then loads the network weights and pre processing data for those networks and based on the data in these files sets up the application for the appropriate reconstruction as well as hidden layers etc To calculate the velocity of the individual components the separation between the two measurment systems must be provided as well as the flowmeter factor Since the readings from the measurement systems contain ripple the volume fraction predictions also contain ripple this can be removed by specifying a threshold level larger than this noise level Once the stepper motors have been configured the test is started After the specified number of frames has been captured and the reconstructions calculated a cross correlation on the gravel and air volume fractions is performed over the duration of these
364. output waveform of a particular transmitter was a signal in phase with the output from the other transmitters that were operating at the same time This variation can also be explained with reference to the output stage of the TC4427 As mentioned previously the on resistance of the MOSFETs in the TC4427 is fairly high Even with the outputs paralleled the output impedance of the transmitter can be as high as 5Q For the situation where a single transmitter is operating the load on the transmitter is constant Subsequently the potential across this output impedance is constant and hence a perfect square wave is generated When multiple transmitters are operating simultaneously the load on a particular transmitter varies since the load on a transmitter also depends on the output states of the other transmitters in the system Subsequently the current drawn varies and the potential across this output impedance varies Hence the transmitter output contains components from the other transmitters and is no longer a perfect square wave The output impedance of the transmitter must be reduced to correct for this The author incorporated an additional push pull power MOSFET output stage for each transmitter in order to achieve a lower output impedance In particular an IRF520 N channel power MOSFET and an IRF9530 P channel power MOSFET were used These devices have a typical on resistance of only 0 20 The TC4427 was used to drive the gates of these two MOSFETs
365. ow 6 tresult capacitance end 13 begin row 7 tresult resistance end 14 begin row 7 tresult capacitance end 15 begin 8 tresult resistance end 16 begin row 8 tresult capacitance end end for col 1 to 8 do begin s1 Voltage 4inttostr index S2 Format 2 5f InputData currentindex index if tresult capacitance then RecViewCapTable Cells col row s2 else RecViewResTable Cells col row 52 inc index end end end 164 close the form procedure TRecViewMain RecViewReturnClick Sender TObject begin Close end when create the form initialise the tables and the chart representing the volume fractions procedure TRecViewMain FormCreate Sender TObject begin RecViewResTable Cells 0 1 TxA RecViewResTable Cells 0 2 TxB RecViewResTable Cells 0 3 RecViewResTable Cells 0 4 TxD RecViewResTable Cells 0 5 TxE RecViewResTable Cells 0 6 TxF RecViewResTable Cells 0 7 TxG RecViewResTable Cells 0 8 TxH RecViewResTable Cells 1 0 RxA RecViewResTable Cells 2 0 RxB RecViewResTable Cells 3 0 RxC RecViewResTable Cells 4 0 RxD RecViewResTable Cells 5 0 RxE RecViewResTable Cells 6 0 RxF RecViewResTable Cells 7 0 RxG RecViewResTable Cells 8 0 RxH RecViewCapTable Cells 0 1 TxA RecViewCapTable Cells 0 2 Re
366. owledge the first ever implementation of frequency division multiplexing to perform impedance tomography Further it has been shown that this technique provides a means of performing on line real time reconstructions at a far higher frame rate than standard time division multiplexed impedance tomography systems with minimal loss in resolution for a given number of electrodes Subsequently it enables tomography to be used in industrial applications where previously the limited frame rates of tomography systems were a hindrance such as on line multi phase flowmeters Although the frame rate of the current system is only 280frames s this measurement protocol has the potential to achieve frame rates of the order of thousands of frames per second simply by using sine wave excitation and high bandwidth electronic components Whether such a high frame rate is useful or not obviously depends on the intended application However DebTech have identified the potential of such a system and have since obtained a provisional patent in order to assess its marketability Further this system is comparatively cheap to implement compared to current commercially available systems Chapter 9 gives recommendations for future development 109 CHAPTER 9 RECOMMENDATI ONS FOR FUTURE DEVELOPMENT Based on the findings and conclusions of this research the following recommendations for future development can be made 1 An 8 electrode FDM impedance tomography syste
367. pMain Createdatabase1 Click Sender TObject begin SampAliasNew Visible true SampAliasNewName SetFocus end once the user has entered a name for their new database create a directory for the tables create a database alias and initialise the tables procedure TSampMain SampAliasNameOKClick Sender TObject var directory string begin directory c Courses Masters SampAliasNewName Text if Session IsAlias SampAliasNewName Text then I check to see whether the database already exists MessageDlg Database already exists mtError mbOK 0 142 else begin else add the database CreateDirectory PChar directory nil Session AddStandardAlias SampAliasNewName Text directory PARADOX Session SaveConfigFile CreateTables SampAliasNewName Text end SampAliasNew Visible false SampAliasNewName Text end create the tables and set them for the current database If the tables don t exist then specify the table fields the table indeces and add an empty record to initialise the table for use If the tables exist then simply open the tables and allow editing of the tables procedure TSampMain CreateTables aliasname string var i integer S string begin dbasename aliasname InpTable TTable Create self OutTable TTable Create self InpTable DatabaseName aliasname OutTable DatabaseName aliasname InpTable TableName InpTrain db OutTable TableName OutTrain db
368. pecific time procedure TDynVerify CalcActualVelClick Sender TObject var simtime airtoptime gravtoptime airbottime gravbottime real airtopposition gravtopposition airtransit gravtransit real begin airtoptime 0 gravtoptime 0 airbottime 0 gravbottime 0 simtime 0 to calculate the average velocity of the bubble between the two measuring rings find the time when the bubble passed the 11 centre of the first measuring ring the time when the bubble passed the centre of the second measuring ring and divide the 1 measuring ring separation by this transit time while simtime RealMain DesiredFrames Value RealMain frameinterval do begin CalculateDistance simtime airtopposition gravtopposition time when air bubble at centre of bottom measuring ring if airtopposition 0 5 airleng gt lowersysbot 0 5 lowersystop lowersysbot 0 005 and airtopposition 0 5 airleng lt lowersysbot 0 5 lowersystop lowersysbot 0 005 then airbottime simtime time when air bubble at centre of top measuring ring if airtopposition 0 5 airleng gt uppersysbot 0 5 uppersystop uppersysbot 0 005 and airtopposition 0 5 airleng lt uppersysbot 0 5 uppersystop uppersysbot 0 005 then airtoptime simtime time when gravel bubble at centre of bottom measuring ring if gravtopposition 0 5 gravleng lowersysbot 0 5 lowersystop lowersysbot 0 005 and gravtopposition 0 5 gravleng lo
369. performed a frame of voltages is captured with the measurement section filled only with water These reference voltages are then subtracted from the real time captured voltages during testing The advantage of this approach is that the neural networks trained are unaffected by changes to the properties of the water in the rig and it is not necessary to retune the synchronous detector boards each time the water in the rig is replaced Clearly this will only work to a point If the electrical properties have changed to such a degree that saturation is occurring in the measurement electronics then retuning would be necessary Using the techniques discussed above a training and test database were generated for both the top and bottom measurement sections These databases were of static configurations and consisted of both air and gravel bubbles as well as different combinations of air and gravel bubbles The author s undergraduate thesis compared the performances of the different neural networks for this reconstruction task and a paper detailing the major results of this research is included in Appendix O on page 211 Since the emphasis of this thesis is more on the dynamic performance of an FDM impedance tomography system the following section will simply tabulate the reconstruction results achieved and provide examples of results for the top measurement section 72 6 3 Neural Network Reconstruction Results Only limited accuracy had been achieved p
370. phy systems can be described as using time division multiplexing TDM where the readings are performed serially and each reading receives a portion of time within which to measure the capacitance or conductance of a particular electrode pair However it is this switching and the transients that result from the switching that increases the overall time of measurement 37 The measured signal passes through a multiplexing stage consisting of CMOS switches As the signal is switched from one electrode to another this Switching action may produce a step input to the next stage of processing This signal experiences a long transit time through the output stage low pass filter and consequently the system must be held in this configuration for a longer time to get a stable reading 12 Only once the output reading is stable can the system then proceed to measure the properties of the next electrode pair This greatly increases the total data acquisition time thus limiting the real time applicability of this technique Deng et a made use of a zero crossing switch to eliminate the negative effect of this impulse Using this technique a frame rate of 50frames s was achieved for a dual plane 12 electrode resistance tomography system 12 Within standard capacitance and resistance tomography systems there does exist the possibility of parallelism For example in resistance tomography systems the voltage differences for all the receiver electrode pairs corre
371. phy systems limit the application of this technique to situations where the fluid is moving relatively slowly 11 The velocity of individual components will be measured using the cross correlation of signals from the top and bottom impedance tomography systems Cross correlation flow measurement is based on the principle of tagging disturbances in the flow generated through turbulence or suspended particles 12 If the disturbance detected by the upstream sensor reappears at the downstream sensor after a period of seconds and L the separation between the two sensors is known to be L then the velocity V can be calculated as V Various T assumptions are made when using this technique of flow measurement and these will be detailed at a later stage The primary use of multi phase flow monitoring equipment is the quantification of the mass or volumetric flow rates of individual components within a flow 13 Since direct mass flow measurement techniques are rare for two phase flows and non existent for three phase flows an inferential method is required 13 Using the velocity distribution Vc of a specific component calculated with the cross correlation algorithm and the volume fraction of that component ac the volumetric flow rate Q of a particular component can be determined as Q acVcdA 1 1 A where the integration is performed over the cross sectional area of the vessel A 10 Combining the volumetric flow rate with the indi
372. pically have to be restarted A component was coded to provide a graphical user interface to set the desired network output In terms of the number of pixels in the image more pixels clearly produce a better quality image However by increasing the number of pixels the system sensitivity of each pixel will decrease In addition readings obtained from real tomography systems contain noise Consequently if the noise level is higher than the sensitivity level the reconstructed image results become vague 24 The minimum resolution of the system therefore determines the size of these pixels 17 It was decided that once again a 10 by 10 matrix of pixels would be used to display the neural network reconstruction results The details of this component and its operation are provided in Appendix K on page 150 The high data capture rate of the sampling unit also reduces the time taken to generate a training or test database It is often the case that more than one frame is to be stored for a particular configuration of air or gravel bubbles Since all these training data points have the same desired output but whose inputs differ slightly due to electrical noise the noise rejection performance of the reconstruction neural network is improved Consequently 25 frames were typically captured for every training or test configuration The following is a brief description of how one would use the software in order to generate a training or test database It i
373. ple of both the capacitance and conductance readings for the different configurations tested TABLE 7 1 COMPARISON BETWEEN THE AVERAGE PEAK TO PEAK OUTPUT RIPPLE OF THE CAPACITANCE AND CONDUCTANCE READINGS FOR THE DIFFERENT CONFIGURATIONS TESTED INCLUDED IS THE OUTPUT RIPPLE MEASURED WHEN SYNCHRONISED TRANSMITTERS ARE USED TWO GROUNDED DISKS AND SYNCHRONISED SPACER TRANSMITTERS BETWEEN PLANES CAPACITANCE 21mV pk to pk 1 809V pk to pk 354mV pk to pk 58mV pk to pk 31mV pk to pk CONDUCTANCE 14mV pk to pk 1 248V pk to pk 250 pk to pk 39mV pk to pk 28mV pk to pk ORIGINAL ORIGINAL CONFIGURATION CONFIGURATION GROUNDED WITH ONLY ONE WITH BOTH DISK BETWEEN PLANE PLANES PLANES OPERATING OPERATING Although these different pipeline configurations reduced the interference between the top and bottom measurement planes considerably the magnitude of the output peak to peak ripple was still twice as large as when only one measurement plane was operating as is evident in table 7 1 However further analysis revealed a simple solution to the problem of interference in a dual plane FDM impedance tomography system Essentially the ripple on the output voltages is a result of the small differences between the operating frequencies of the two planes Consequently if both systems were to use a common frequency generator board then this output ripple would not be produced This is the concept behind synchronised transmitters where one frequenc
374. ple outputnodeindex i OutputData currsample outputnodeindex i for inputnodeindex 0 to numhidden do DiffOutputWeight inputnodeindex outputnodeindex i DiffOutputWeight inputnodeindex outputnodeindex i delta outputnodeindex i OutputHidden currsample inputnodeindex end end end else begin if regression then simple linear output unit if numoutputs 2 then begin if weightsum 0 then OutputNetwork currsample outputnodeindex weightsum else OutputNetwork currsample outputnodeindex 0 newderiv 1 end else simply two phases so use tanh else begin OutputNetwork currsample outputnodeindex tanh weightsum newderiv CalcDeriv OutputNetwork currsample outputnodeindex end calculate delta for output units a record of the delta values is kept to allow for the calculation of the weight adjustment for the hidden I layers delta outputnodeindex OutputNetwork currsample outputnodeindex OutputData currsample outputnodeindex newderiv calculate weight adjustment for output units for inputnodeindex 0 to numhidden do DiffOutputWeight inputnodeindex outputnodeindex DiffOutputWeight inputnodeindex outputnodeindex delta outputnodeindex OutputHidden currsample inputnodeindex end end calculate delta for hidden layer nodes for inputnodeindex 0 to numhidden 1 do begin deltasum 0 calculate the delta for the hidden layer by adding the products of the
375. pling hardware and show CurMain modally procedure TSampMain Currentsamples1 Click Sender TObject begin notifvalue SampHardware InitSample 1 CurMain Handle CurMain ShowModal end set the number of frames procedure TSampMain SampNumFramesOKClick Sender TObject begin numframes SampNumFramesValue Value SampNumFrames Visible false end if the user changes his mind procedure TSampMain SampNumFramesCanClick Sender TObject begin SampNumFrames Visible false end procedure TSampMain Setthenumberofframes1 Click Sender TObject begin SampNumFrames Visible true end end PROGRAM LISTING OF ADDUNIT PAS This unit is used to capture samples from the impedance tomography system and store them in the database Also stored in the database is a cross sectional image of the pipe at the time of the test This image is set using a Tomo component which created specifically for this task Buttons are also provided which permit the user to look at previous test cases in the database To create new test case move to the end of the database and click the RIGHT arrow This will automatically create space in the database for the new test case The user then specifies what the desired output for that test case is and presses the Start sampling button to initiate the sampling as a background process The results of the current sample are displayed in a table format A bar chart displays
376. ply by using two 31 25kHz 39 0625kHz different crystal frequencies eight output square same square wave generation software as was used 4 43361 MHz CRYSTAL TRANSMITTER OUTPUT FREQUENCY waves and their quadrature waveforms were 69 2752kHz generated with the correct duty cycle and phase shift 46 1834kHz 27 7101kHz The crystal frequencies chosen were 4 43361MHz and 34 6376kHz 5MHz and the output frequencies generated are given in table 5 1 Transmitter G was chosen as the lowest output frequency since it was used as the clock input to the Switched capacitor low pass filters Subsequently the cutoff frequencies of the switched capacitor filters were 352Hz 52 It is impossible to generate eight independent output frequencies using this technique since all four frequencies for each microcontroller are related by the crystal frequency In order to generate independent frequencies a different frequency generation technique would be required Initially CMOS 555 oscillators were investigated However tests revealed that the output frequency would drift over time since the timing of the oscillator depended on an RC combination To achieve the required frequency stability it was necessary to employ a crystal based frequency generator Although a number of logic gate based designs were tested it was decided that a single PIC16F84A microcontroller would be used as each frequency generator This may seem an overly el
377. procedure UpdateTables update the tables with the new data procedure DisplayData display the desired output for a specific test case using the InputTomo component public Public declarations procedure WndProc var TheMsg TMessage override override the default Windows messaging procedure so as to catch the stop sampling I notification from the sampling hardware end var AddToDatabase TAddToDatabase implementation uses mainunit R DFM tM SAMPLING PROCEDURE process the results by extracting the data for each system either the upper or lower system and storing the voltages in InpTable Also store the volume fractions of the different components and whether the data is for the upper or lower system Finally store the picture of the desired network output in OutTable procedure TAddToDatabase ProcessResults var row col index vers integer S string begin InpTable FindKey currenttest 0 001 OutTable FindKey currenttest 0 001 for each test case capture numframes of data this is to improve the noise performance of the neural network for vers 1 to numframes do begin put tables in Edit mode InpTable Edit OutTable Edit store a record of which system we are currently using InpTable FieldByName System Value AddSystemSelect ItemIndex store the pixel image of the desired network output index 1 for col 0 to 9 do
378. pthresh tempsum real begin Il if user wishes to view performance of network for the training data if RecSetParam RecShowTrainPerf Checked then begin CalculateTrainError 0 totaltrain 1 RecTrainThreshErr Caption floattostrf currtrainthresh ffGeneral 7 5 RecTrainVoidErr Caption floattostrf currtrainsum ffGeneral 7 5 end keep a record of the current threshold and sum void errors so as to provide for a comparison when the new values are calculated tempthresh currthresh tempsum currsum calculate the network outputs for the test data as well as the new 170 threshold and sum void error percentages CalculateTestOutput totaltrain totaltrain totaltest 1 if numoutputs lt gt 2 then begin RecTestThreshErr Caption floattostrf currthresh ffGeneral 7 5 end RecTestVoidErr Caption floattostr currsum ffGeneral 7 5 if earlystop then begin if wish to perform early stopping then check whether the validation performance is deteriorating if RecSetParam RecStopThreshold Checked then begin if currthresh gt tempthresh then inc currentfails else currenttfails 0 end else begin if currsum gt tempsum then inc currentfails else currentfails 0 end end RecNetworkEpoch Caption Current epoch inttostr currepoch RecNetworkProgressBar Steplt Application ProcessMessages end tM SPECIAL FUNCTIONS softmax activation function for the
379. quadrature waveforms are generated A goto instruction at the end of the sequence ensures that the loop repeats and the square waves are repeated In order to generate complete waveforms a sequence of port set instructions corresponding to the lowest common multiple of all the square wave periods must be written This ensures that when the end of the sequence is reached all the square waves complete a cycle at the same time and no half cycles exist The periods of the above square waves are 8 12 16 and 20 cycles respectively The sequence of 240 port set instructions were calculated using a program written in Delphi They were written to a text file compiled using MPASM and downloaded using the download software provided with the microcontroller programmer One further subtlety of this software is the inclusion of the nop instructions To ensure that the output waveforms have the correct 5096 duty cycle and 90 phase shift it is necessary to analyse the timing of the instructions In particular it was noted that the goto instruction at the end of the sequence requires two instruction cycles to complete Consequently if this was not compensated for then the duty cycle and phase shift would be incorrect when the goto instruction was executed Two nop instructions are included after every port set instruction except 25 the last in the sequence to compensate for this The OUTPUT EBEGUENGIES GENGRATED WITH A 4MHz last port set
380. quency components may still exist on the low pass filter outputs that were not predicted by the simulation software Using this threshold the PBIL algorithm was able to determine a set of eight transmitter frequencies with a minimum frequency component at 829Hz which should be filtered out by the fourth order low pass filter stages It was observed that the performance of the PBIL algorithm varied on each run resulting from the convergence on local optima instead of an absolute optimal solution Subsequently the above tests were repeated on multiple occasions The eight output frequencies selected using the PBIL optimisation algorithm are detailed in table 5 2 which includes the corresponding crystal frequencies and division factors required to generate these frequencies Further Appendix on page 136 lists up to the 9 harmonic of these fundamental frequencies TABLE 5 2 TRANSMITTER FREQUENCIES SELECTED USING PBIL OPTIMISATION ALGORITHM WITH THE CRYSTAL FREQUENCIES AND DIVISION FACTORS REQUIRED TO GENERATE THESE FREQUENCIES TRANSMITTER OUTPUT FREQUENCY REQUIRED CRYSTAL REQUIRED PIC FREGUENCY DIVISION FACTOR 79 1718kHz 8 867238MHz 75kHz 12MHz 73 5294kHz 20MHz 66 6667kHz 16MHz 62 8364kHz 11 0592MHz 54 3478kHz 20MHz also LMF100 clock 23 4375kHz 12MHz 48kHz 6 144MHz A B D E F G H To assess the accuracy of the simulation the above transmitter frequencies were implemented and the frequency
381. r end end PROGRAM LISTING OF PARAMUNIT PAS This unit is used to set the desired network training parameters When the form is opened the previous parameters used are automatically loaded from the registry unit paramunit interface uses Windows Messages SysUtils Classes Graphics Controls Forms Dialogs StdCtrls Spin ExtCtrls Buttons Registry type TRecSetParam class TForm Paneli TPanel RecSetGenParam TGroupBox RecSetEarlyParam TGroupBox RecSetNetParam TGroupBox RecSetModelParam TGroupBox Label TLabel Label3 TLabel RecSetMaxEpochs TEdit RecSetNumHidden TSpinEdit RecPerformEarlyStop TCheckBox RecPerformModel TCheckBox RecUseGradDescent TRadioButton RecUseRPROP TRadioButton Label4 TLabel Label5 TLabel RecSetLearn TEdit RecSetMomentum TEdit RecRPropParam TPanel Label6 TLabel Label7 TLabel RecSetDeltalnit TEdit RecSetDeltaMax TEdit Label8 TLabel RecNumFails TSpinEdit Label9 TLabel RecStopVolume TRadioButton RecStopThreshold TRadioButton Label10 TLabel Label11 TLabel Label12 TLabel RecModelStart TSpinEdit RecModelStop TSpinEdit RecModellnc TSpinEdit RecLoadReturn TSpeedButton RecShowTrainPerf TCheckBox Label13 TLabel RecSetUpdateFreq TSpinEdit procedure RecPerformEarlyStopClick Sender TObject procedure RecPerformModelClick Sender TObject procedure RecUseGradDescentClick Sender TObject procedure RecUseRPROPClick
382. r signal to noise ratio 44 Since the minimum frequency separation of the four transmitter frequencies used is 6 25kHz the cutoff of the low pass filter must be less than 6 25kHz Consequently an initial cutoff frequency of 625Hz was chosen However only first order low pass filters were implemented and therefore a fairly substantial 6 25kHz ripple was still evident on the DC output of the low pass filters To reduce this ripple to a level where the accuracy of the reconstruction was not affected it was necessary to set the cutoff frequency of the low pass filters to 3Hz thus limiting the frame rate of this prototype system to 3frames s Clearly this limitation would need to be addressed in the final system in order to take full advantage of the FDM technique but it was deemed satisfactory in the short term since it was only necessary to prove that the concept actually worked The circuit details of the synchronous detection circuitry are given in Appendix B on page 119 29 4 2 5 Capture of capacitance and conductance readings and the reconstruction results achieved The voltages are captured using an Eagle PC30G data acquisition card plugged into the reconstruction computer as illustrated in figure 4 3 on page 24 This card supports 16 analogue input voltages in the range from 10V to 10V and can achieve a sampling rate of 100ksamples s provided the data is streamed directly to memory under the supervision of the Direct Memory Access DMA co
383. r the case of components travelling at different velocities the mass flow rate of component n is given by 50 2R 2R M f July DW x y 7 10 1 0 x 0 y 0 where R is the radius of the pipeline u x y t is the velocity of a particular pixel measured using the cross correlation of the pixel values and Wy X Yt pq if P X Yt Pp 7 11 W x y t 0 if p x y t is determined using the pipeline imaging system 7 2 Real time Sampling and Reconstruction Software Description The following sections detail the design of a flow simulation apparatus and the modifications made to the neural network reconstruction software to perform on line real time reconstructions 7 2 4 Design of a flow simulation apparatus In order to verify the dynamic reconstruction performance of the dual plane 16 electrode FDM impedance tomography system it was necessary to design some apparatus that could accurately control both the speed and position of simulated masses It was decided that stepper motors would provide this speed and position control Figure 7 2 on page 83 is a diagram of the flow simulation apparatus developed for this particular application The requirements of this flow simulation apparatus were as follows e The position of the bubble must coincide with one of the standard bubble positions defined by the bubble placement system This was necessary to ensure that optimal reconstruction results could be achieved using the n
384. raining data from the impedance tomography system in a static situation The desired network output for each test case is specified through an InputTomo component which allows the user to set the phase of each pixel in an image representation of the vessel cross section The voltages are captured from the system using the sampunit and the results are stored in a Delphi database A database consists of two tables namely InpTable for the 128 voltage readings from the rig and OutTable for the pixels specifying the desired network output and the volume fractions for the phases An option is provided to continuously sample data from the system and display the results in a tabular format this will be useful for de bugging the final system since the user can monitor all 128 voltages without having to measure at that particular point on the board using a multimeter unit mainunit interface uses Windows Messages SysUtils Classes Graphics Controls Forms Dialogs ExtCtrls Tomo Menus StdCtrls DBTables DB Buttons sampunit Spin Registry type TTomoSystem top bottom 1 dedicted type to differentiate between readings from the top and bottom system TSampMain class TForm SampMainMenu TMainMenu File1 TMenultem Opendatabase1 TMenultem Createdatabase1 TMenultem Closedatabase1 TMenultem N1 TMenultem Exit1 TMenultem View TMenultem Startsampling1 TMenultem View2 TMenultem Database TMe
385. ramed in the PIC16F84A to determine a set of possible frequencies between 17 6kHz and 80kHz trial solution is taken from this set and the process of mixing in the synchronous detectors is simulated as a sum and difference of the frequency harmonics The aliasing that occurs in the switched capacitor filters on the output is also simulated and the fitness of a trial solution is the minimum frequency component at the LPF outputs Clearly the objective is to maximise this frequency so that all frequency components are outside the passband of the LPF and can therefore be filtered out unit pbilform interface uses Windows Messages SysUtils Classes Graphics Controls Forms Dialogs StdCtrls Math ComCtrls Spin ExtCtrls type TPBILMain class TForm LoadFregs TButton TestMult TButton Label TLabel Label2 TLabel MaxGenerations TSpinEdit NumTrials TSpinEdit PBILStart TButton PBILProgress TProgressBar Freqt TLabel Freq2 TLabel Freq3 TLabel Freq4 TLabel Freq5 TLabel Freq6 TLabel Freq7 TLabel Freq8 TLabel Label3 TLabel BestFregFit TLabel MultiSearchPBIL TButton Label4 TLabel Numlterations TSpinEdit Label5 TLabel CurlterCaption TLabel Label6 TLabel OverBestFregFit TLabel OverFreq5 TLabel OverFreq6 TLabel OverFreq7 TLabel OverFreq8 TLabel OverFreq1 TLabel OverFreq2 TLabel OverFreq3 TLabel OverFreq4 TLabel SaveMultiRecord TSaveDialog Label7 TLabel
386. ravChart Series 0 Clear TopGravChart Series 0 Addy outputstore testposition prediction 2 clTeeColor TopWaterChart Series 0 Clear TopWaterChart Series 0 Addy 100 outputstore testposition prediction 1 outputstore testposition prediction 2 clTeeColor BottomAirChart Series 0 Clear BottomAirChart Series 0 Addy outputstore testposition prediction 3 clTeeColor BottomGravChart Series 0 Clear BottomGravChart Series 0 Addy outputstore testposition prediction 4 clTeeColor BottomWaterChart Series 0 Clear BottomWaterChart Series 0 Addy 100 outputstore testposition prediction 3 outputstore testposition prediction 4 clTeeColor frmnumber Caption inttostr reconposition DecodeTime outputstore reconposition sampletime hours mins secs msecs smins Caption inttostr mins min Ssecs Caption inttostr secs s smsecs Caption inttostr msecs ms Update end end end perform a point by point cross correlation on the volume fraction data obtained from top and bottom measurement systems the peak of this function gives the number of frames difference between the top and bottom systems By multiplying this number of frames by the frame interval the transit time of the gravel and sir bubbles is obtained Since the system separation is known this is then used to calculate the average velocity between the top and bottom system procedure TRealMain PerformCorrelat
387. re required if RecSetParam RecUseGradDescent Checked then OldHiddenWeight row col HiddenWeight row col else begin HiddenDeltaWeight row col RecSetParam GetDeltalnit OldDiffHiddenWeight row col 0 end end end end II initialise the weights for the output layer of neurons for row 0 to upperindex do begin for col 0 to RecLoadDBase GetNumOutputs 1 do begin OutputWeight row col weightfactor Random 0 5 DiffOutputWeight row col 0 if RecSetParam RecUseGradDescent Checked then OldOutputWeight row col OutputWeight row col else begin OutputDeltaWeight row col RecSetParam GetDeltalnit OldDiffOutputWeight row col 0 end end 155 end RecSetWeightFactor Visible false Starttraining Enabled true end save the network weights to a textfile if a model search has been done then save the weights for the best network procedure TRecMain SavenetworkweightsClick Sender TObject var tempfile textfile filename string rowindex colindex upperindex integer begin upperindex 128 filename cATestData Weight inttostr testnumber txt AssignFile tempfile filename ReWrite tempfile WriteLn tempfile RecLoadDBase GetNumOutputs if RecSetParam RecPerformModel Checked then begin I if a model search has been done then write the weights for the best network upperindex RecTrainNetwork GetNumBest WriteLn tempfile upperindex for colindex 0 to upperinde
388. reconstruction CalculateTestCase II display network prediction PlotTestOutput RecVoidLabel Caption Volume fraction predictions for test case inttostr testindex RecPredictLabel Caption Network prediction for test case inttostr testindex RecDesiredLabel Caption Desired output for test case inttostr testindex 175 inc testindex if testindex RecLoadDBase GetTotalTrain RecLoadDBase GetTotalTest then testindex RecLoadDBase GetTotalTrain end else I instruct hardware to capture a frame of data SampHardware StartSample 1 end calculate the network output for the test case by propagating the input voltages through the network procedure TRecTest CalculateTestCase var rowindex colindex i upperindex numoutputs numhidden integer tempvalue weightsum real weightstore array 1 2 of real begin upperindex 128 numoutputs RecLoadDBase GetNumOutputs numhidden RecSetParam GetNumHidden SetLength testoutput numoutputs I if there is a hidden layer then propagate the inputs through the network calculating first the outputs from the hidden layer and then based on these outputs calculate the final network outputs if numhidden gt 0 then begin upperindex numhidden SetLength testhiddenoutput numhidden 1 for colindex 0 to numhidden do begin if colindex 0 then testhiddenoutput colindex 1 bias term else begin weightsum 0 for rowindex 0 to 128 do w
389. rences between the desired and the predicted volume fractions are summed over the entire database and then divided by the total number of frames in the database It therefore represents the mean absolute error of the volume fraction predictor neural network for a specific phase This measure is independent of pixel position and operates simply in terms of volume It gives an indication of the ability of the network to predict volume fractions accurately The generalisation capability of the neural network is regularly monitored during network training Early stopping is used to prevent the neural network from fitting the training data exactly which would result in poor reconstruction results for unseen bubble configurations The network training is stopped as soon as the generalisation performance starts to consistently deteriorate The neural network training algorithms implemented in this system will be discussed briefly in the following section for completeness Training a multi layer perceptron is typically performed in two stages Firstly the derivative of the error function with respect to the network weights must be determined The standard sum of squares error function is as follows 29 C 2 E IY y t 3 10 2 k 1 where C is the total number of network outputs y is the network prediction for output and t is the desired prediction for output k The derivative of this error function is calculated in a computationally efficient way u
390. resentative dynamic volume fraction error the desired volume fraction profile was shifted so as to coincide with the measured volume fraction profile to allow for drive belt slippage The amount of shift was calculated as the number of frames between the peak of the desired volume fraction profile and the peak of the measured volume fraction profile 7 2 8 Cross plane interference problem Cross plane interference if not accurately modelled will result in a general loss of accuracy and can severely affect the reconstruction process 49 Further cross talk between sensors can distort the cross correlation function making accurate determination of the peak more difficult 57 The standard technique used to prevent cross plane interference is the method of interleaving This ensures that only one sensing array is activated at any time thus effectively eliminating cross talk For the adjacent measurement protocol in resistance tomography the operation of the interleaving data acquisition technique is as follows 43 1 current is injected between electrode 1 and 2 for system 1 and the potential differences are measured on the remaining electrodes of system 1 2 current is injected between electrode 1 and 2 for system 2 and the potential differences are measured on the remaining electrodes of system 2 3 current is injected between electrode 2 and 3 for system 1 and the potential differences are measured on the remaining electrodes of system 1
391. result resistance end 12 begin row 6 tresult capacitance end 13 begin row 7 tresult resistance end 14 begin fow 7 tresult capacitance end 15 begin row 8 tresult resistance end 16 begin row 8 tresult capacitance end end for col 1 to 8 do begin 51 Voltage 4inttostr index s2 Format 2 5f double Inp Table FieldByName s1 Value 146 if tresult capacitance then AddCapTable Cells col row 52 else AddResTable Cells col row 52 inc index end end end set the InputTomo component based on the database to reflect the desired output for a specific test case procedure TAddToDatabase DisplayData var col row index integer S string tempdesired TOutput begin index 1 for col 0 to 9 do for row 0 to 9 do begin S Pixel tinttostr index tempdesired col row OutTable FieldByName s Value inc index end AddInputTomo Desired tempdesired AddinputTomo WaterVolume OutTable FieldByName WaterVolume Value AddinputTomo GravelVolume OutTable FieldByName GravelVolume Value AddinputTomo AirVolume OutTable FieldByName AirVolume Value AddSystemSelect ItemIndex InpTable FieldByName System Value AddInputTomo Repaint end end PROGRAM LISTING OF CURUNIT PAS This unit continuously samples the impedance tomography system using streaming and displays the results in a tabular format
392. reviously with three phase reconstructions It was therefore decided that a better assessment of the 16 electrode FDM impedance tomography system performance could be achieved using the simpler situation of a two phase air water reconstruction 6 31 Two phase air water reconstruction results Multiple single layer feed forward neural networks were trained to perform the two phase air water image reconstruction and the averaged results are detailed in table 6 2 which includes a comparison to the results for the 8 electrode prototype As is evident in table 6 2 the 16 electrode system outperforms the 8 electrode system for all performance assessment criteria considered This was expected since more capacitance and conductance readings are obtained from the 16 electrode system than the 8 electrode system and therefore an increase in resolution is achieved TABLE 6 2 COMPARISON BETWEEN THE PERFORMANCE OF THE 16 ELECTRODE FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM AND THE 8 ELECTRODE PROTOTYPE FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM FOR A TWO PHASE AIR WATER IMAGE RECONSTRUCTION USING A SINGLE LAYER FEED FORWARD NEURAL NETWORK ALL RESULTS REPRESENT THE ABSOLUTE ERROR AND ARE ROUNDED OFF TO TWO SIGNIFICANT DIGITS PERFORMANCE MEASURE 16 ELECTRODE LABORATORY SCALE FDM 8 ELECTRODE PROTOTYPE IMPEDANCE TOMOGRAPHY SYSTEM FDM IMPEDANCE TOP SYSTEM BOTTOM SYSTEM TOMOGRAPHY SYSTEM THRESHOLD ERROR 96 2 1 2 2 9 1 AIR V
393. rfactor systemseparation denominator if airvelocity 0 001 then airvelocity 0 AirVelPredict Caption floattostrf airvelocity ffFixed 5 3 end repeat the process for the gravel bubble maxval 0 for i 0 to upperindex do begin tempsum 0 for k 1 to upperindex do begin if k i lt upperindex then begin if totaloutputs 2 then begin if StepperConfig FlowDirSelect ItemIndex 1 then tempsum tempsum outputstore k i prediction 2 outputstore k prediction 4 else tempsum tempsum outputstore k i prediction 4 outputstore k prediction 2 end else begin if StepperConfig FlowDirSelect ItemIndex 1 then tempsum tempsum volumestore k i 2 volumestore k 4 else tempsum tempsum volumestore k i 4 volumestore k 2 end end end gravelcorr i 1 tempsum upperindex if gravelcorr i 1 gt maxval then begin maxval gravelcorr i 1 maxindex i 1 end end graveldelay maxindex 1 denominator graveldelay frameinterval if denominator 0 or maxval lt correlationthreshold then begin gravelvelocity 1 GravelVelPredict Caption unknown end else begin gravelvelocity flowmeterfactor systemseparation denominator if gravelvelocity lt 0 001 then gravelvelocity 0 GravelVelPredict Caption floattostrf gravelvelocity ffFixed 5 3 end Iplot results for testindex 1 to upperindex do begin if totaloutputs 2 then begin AirPlo
394. riesList 1 Add AddInputTomo AirVolume AddVolume SeriesList 2 Add AddInputTomo GravelVolume end every 10ms update the status bar to indicate the current program situation specifically the X and Y position of the cursor on the InputTomo component the status of the sampling hardware the currenttest and totalsamples indicator and whether the left and right buttons are enabled to control the browsing through the database or create a new test sample Also update the chart of the volume fractions procedure TAddToDatabase StatusUpdateTimer Sender TObject begin AddStatusBar Panels 0 Text X inttostr AddInputTomo Xposition AddStatusBar Panels 1 Text Y inttostr AddInputTomo Y position if AddUserSetVolume Checked then begin AddSetGravel Enabled true AddSetAir Enabled true AddVolume SeriesList 1 YValue 0 strtofloat AddSetAir Text AddVolume SeriesList 2 YValue 0 strtofloat AddSetGravel Text AddVolume SeriesList 0 YValue 0 100 AddVolume SeriesList 1 YValue 0 AddVolume SeriesList 2 YValue 0 end else begin AddSetGravel Enabled false AddSetAir Enabled false AddVolume SeriesList 0 YValue 0 AddInputTomo WaterVolume AddVolume SeriesList 1 YValue 0 AddInputTomo AirVolume AddVolume SeriesList 2 YValue 0 AddinputTomo GravelVolume AddSetWater Text floattostr AddInputTomo WaterVolume AddSetAir Text floattostr AddInputTomo AirVolume AddSetGravel
395. rite tempfile CurResTable Cells col row WriteLn tempfile end CloseFile tempfile end end start drift ananlysis by specifying destination filename and desired capture rate procedure TCurMain StartMonitorClick Sender TObject begin if DriftTimer Enabled false then begin if DriftSave Execute then begin sampnumber 1 SampleNumber Visible true AssignFile driftfile DriftSave FileName driftfilename DriftSave FileName ReWrite driftfile StartMonitor Caption Stop recording DriftTimer Interval Round 3 6e6 RecordRate Value DriftTimer Enabled true CloseFile driftfile end end else begin StartMonitor Caption Start recording DriftTimer Enabled false SampleNumber Visible false end end end PROGRAM LISTING OF SAMPUNIT PAS This unit controls the sampling of the Eagle PC30G card The sampling is controlled by the sample control board which sets up the multiplexers and triggers the card to start sampling The card is configured for burst mode and samples channel 0 to 15 each time it receives a trigger on the external trigger line Streaming is used to capture the samples as a background process it is configured using the EDR32 unit provided with the card Once the sampling is complete a Windows notification message is issued by EDR32 This message is captured by the sampling software and the data is then processed For the sampling to work correctly th
396. rithm calculating the velocities of the individual components on completion procedure TRecMain Realtimetesting1Click Sender TObject begin RealMain ShowModal end tM FILE HANDLING PROCEDURES all file saves and loads are from a directory called TestData write the loaded data to a text file so that it can be used in Matlab procedure TRecMain Writedatatoatextfile1 Click Sender TObject var tempfile textfile filename string row col endcol integer begin endcol RecLoadDBase GetNumOutputs use the test number to specify the filename filename c TestData mtlbtst inttostr testnumber txt AssignFile tempfile filename Rewrite tempfile WriteLn tempfile RecLoadDBase GetTotalTrain write the input voltages for the training data to the text file for row 0 to RecLoadDBase GetTotalTrain 1 do begin for col 1 to 128 do Write tempfile InputData row col WriteLn tempfile end WriteLn tempfile endcol write the desired output for the training data to the text file for row 0 to RecLoadDBase GetTotalTrain 1 do begin for col 0 to endcol 1 do Write tempfile OutputData row col WriteLn tempfile end 153 WriteLn tempfile RecLoadDBase GetTotalTest write the input voltages for the testing data to the text file for row RecLoadDBase GetTotalTrain to RecLoadDBase GetTotalTrain RecLoadDBase GetTotalTest 1 do begin for col
397. rives procedure TStep ResetDrives var temp string begin temp RESET ProcessCommand temp end move the drive continuously at a fixed velocity in a specific direction procedure TStep JogMove drivenum integer dir TDirection var drive1 drive2 drive3 Boolean dirt dir2 dir3 TDirection velt vel2 vel3 integer begin drive1 FALSE drive2 FALSE drive3 FALSE vel 1 vel2 1 vel3 1 CW dir2 CW dir3 CW set the drive direction and enable drive case drivenum of 1 begin drivet TRUE dirt dir end 2 begin drive2 TRUE dir2 dir end 3 begin drive3 TRUE dir3 dir end end GoDrive drivet drive2 drive3 vel1 vel2 vel3 dir1 dir2 dir3 end procedure TStep MoveFixedDist velt real vel2 real vel3 real dist1 integer dist2 integer dist3 integer var temp string begin set motion to move a fixed distance and to work according to absolute co ordinates temp MC0000 ProcessCommand temp temp MA0000 ProcessCommand temp set the velocity temp V floattostrf vel1 ffGeneral 5 2 floattostrf vel2 ffGeneral 5 2 floattostrf vel3 ffGeneral 5 2 0 ProcessCommand temp set the distance temp D tinttostr dist1 inttostr dist2 inttostr dist3 0 ProcessCommand temp II initiate motion temp 001110 ProcessCommand temp end get drive status from AT6400 proc
398. rns to model the desired functional mapping inherent in the training database 17 CHAPTER 4 FREQUENCY DI VI SI ON MULTI PLEXED I MPEDANCE TOMOGRAPHY OF AN 8 ELECTRODE SYSTEM In order to determine the velocity of components by cross correlating the data from two planes the real time performance of the system must be guaranteed This incorporates the image reconstruction and data processing tasks 12 It can be shown that the frame rate v must meet the following requirement i 2L AV Vmax Vmax where L is the separation between the two planes Vmax is the maximum expected velocity and is the MAX 4 1 desired velocity discrimination 12 A frame rate of 200frames s is required for a velocity discrimination of 1096 with a maximum velocity of 20m s and a system separation of approximately 500mm No results had been published to date of a 16 electrode impedance tomography system capable of operating on line at this frame rate While the standard tomography measurement technique could possibly be pushed to a frame rate close to that required it is fundamentally limited by its sequential processing Consequently it was decided that the data acquisition problem should be examined from a different perspective to determine whether a different approach may yield higher frame rates more easily Standard tomography systems are based on the concept of switching between different electrode combinations Specifically standard tomogra
399. rocedure GoDrive drive1 Boolean drive2 Boolean 189 drive3 Boolean velt integer vel2 integer vel3 integer TDirection dir2 TDirection dir3 TDirection move a drive continously in a specific direction procedure JogMove drivenum integer dir TDirection move a drive a fixed distance at a certain velocity procedure MoveFixedDist velt real vel2 real vel3 real distt integer dist2 integer dist3 integer stop the motor drives procedure StopDrive disable the motor drives procedure DisableDrives reset the motor drives procedure ResetDrives get the status of the drives procedure GetMotorData want TRequestData var 1 double var axis2 double var axis3 double set the current drive position as the zero position for absolute movement procedure SetZeroPos download the operating system procedure DownloadOS procedure ResetLimits procedure SetAcceleration axis1 integer axis2 integer axis3 integer end implementation tM COMMAND CONTROL PROCEDURES send a block of commands to the AT6400 function TStep SendAT6400Block Address word Command CharBuffer word var Cmd CharBuffer begin clear out buffer for 0 to MAXBYTES do begin Cmd l TERMINATOR end 1 copy Command buffer to buffer 1 50 while Command l lt gt TERMINATOR do begin Cmd I Command I Inc I end
400. rray of double double public Public declarations end var PBILMain TPBILMain implementation R DFM Quick sort routine to sort an array in ascending order adapted from Borland s BP7 demo procedure QSort var X array of double Lbound Ubound Integer procedure Sort L R Integer var J integer U V double begin l L J R U X L R div 2 repeat while X I lt U do 1 1 while U lt X J do J J 1 lt J then begin V Xl X I KI V l 1 1 J J 1 end until gt J if L lt J then Sort L J R then Sort l R end begin Sort Lbound Ubound end E E 138 I frequency simulation procedures Il generate a list of all available transmitter frequencies procedure TPBILMain LoadFreqsClick Sender TObject var i j k integer factor real divfactor array 1 60 of double TempFreq array 1 2000 of double begin factor 64 i21 while factor lt 1008 do begin divfactor i 1 factor inc i factor factor 16 end for j 1 to 23 do begin for k 1 to 60 do begin TempFreq i xtal j divfactor k inc i end end 1 while lt gt 0 do begin if TempFreq i gt 17 6e3 and TempFreaq i lt 80e3 then begin AvailFreq j TempFreq i inc j end inc i end QSort AvailFreq 0 j 2 k 1 for
401. rted on a raised stainless steel base and a drain valve is provided for cleaning purposes All metal fittings were 34 constructed from stainless steel due to its corrosion resistance A photograph of the assembled laboratory scale rig can be seen in figure 5 2 Since the electrodes of a capacitance tomography System are typically mounted outside the vessel the measured inter electrode capacitances vary non linearly with respect to the permittivity of the enclosed material because of the capacitance of the insulating pipe wall 46 Ideally the permittivity of the pipe wall material should be small to minimise the field divergence of the capacitance tomography system 47 48 Clear acrylic was used for the construction of the vessel This was necessary to enable visual verification of the reconstruction results Further it is essential that training and test databases can be generated quickly and easily since neural networks are used to perform the reconstruction By using clear acrylic for the pipeline the task of aligning a specific test or training configuration was simplified FIGURE 5 2 PHOTOGRAPH OF LABORATORY SCALE MULTI PHASE PIPELINE SECTION The selection of the pipe wall thickness is an important design parameter If it is too thick then the capacitance measurements become highly non linear and the sensing fields of adjacent electrodes are almost completely confined to the pipe wall 47 In addition the capacitance
402. ruction of the top frequency division multiplexed impedance tomography system where the desired outputs are on the right and the network predictions are on the left of each screen capture 76 Comparison between the computations required to determine the average velocity velocity profile and velocity vector for a given frame rate and at a specified minimum flow velocity 81 Diagram of the flow simulation apparatus with included detail of the motor mount and positioning base 83 Photograph of the flow simulation apparatus standing next to the laboratory scale rig 84 7 4 Photograph of the flow simulation apparatus for the specific test of a 0 04 volume fraction gravel and a 0 04 volume fraction air bubble passing through the pipeline at the same time 84 7 5 Screen capture of the real time sampling form 86 7 6 Screen capture of the dynamic verification form 89 7 7 Comparison between the capacitance output voltages for the transmitter A receiver A electrode pair of the top system when only the top system is operating and when both the top and bottom measurement planes are operating at the same time 91 7 8 Plot of both desired and measured volume fraction profiles at the top and bottom measurement Systems for a simulated 0 2m s flow consisting of a 0 04 volume fraction air bubble and a 0 01 volume fraction air bubble 93 7 9 Screen captures of a 0 04 volume fraction air bubble near the left hand edge of the rig and a 0 01 volume fraction air bubbl
403. s ExtCtrls type TCalibSetup class TForm Panel TPanel CalibOK TButton CalibCancel TButton ainPanel TPanel ainPrompt TLabel ainEdit TEdit nputPanel TPanel nputPrompt TLabel Label1 TLabel Label2 TLabel nputStart TEdit nputStop TEdit procedure FormShow Sender TObject procedure CalibOKClick Sender TObject private Private declarations currentindex integer currentstage integer public Public declarations end var Cal libSetup TCalibSetup implementation uses loadunit R DFM RR GENERAL PROCEDURES initialise variables proc edure TCalibSetup FormShow Sender TObject begin end ainPanel Visible true nputPanel Visible false ainEdit Text 1 ainPrompt Caption Specify the number of training regions currentindex 1 currentstage 1 ainEdit SetFocus CalibOK ModalResult mrNone collect information regarding the calibration points for both the training and testing databases procedure TCalibSetup CalibOKClick Sender TObject begin if currentstage 1 then begin numtrainregions strtoint MainEdit Text currentstage 2 InputPanel Visible true InputPrompt Caption Specify the bounds for region inttostr currentindex InputStart Text 1 InputStart Enabled false InputStop Text 1 InputStop SetFocus end else if current
404. s up to and including the 100 harmonic of the reference frequency are included 4 Calculate the sum and difference frequencies at the multiplier outputs for each of the eight transmitter frequencies With each transmitter as a reference frequency the sum and difference frequency components between all the received frequency components and the harmonics of the reference square wave are calculated 5 Simulate the aliasing in the switched capacitor filters Since the LMF100 is a sampled data system any frequency components greater than one half the clock frequency result in aliasing In this application the clock frequency equals the frequency of transmitter G The effect of the aliasing is to shift frequency components from forocr X where X is the difference between the input frequency and foror down to eoo _ X Hence components in the region of the clock frequency are shifted down into the passband of the switched capacitor low pass filter and appear at the output as ripple For example if the clock frequency was 20kHz and the multiplier output contained a frequency component at 19 9kHz this frequency component would be shifted down to 100Hz and would appear on the switched capacitor filter output as a 100Hz ripple component Initial tests using this simulator showed that it is undesirable to have transmitter frequencies that are related Initially the eight transmitter frequencies were generated using two PIC16F84 microcontrollers
405. s not an attempt to explain the underlying implementation of the software which can be determined by reading the commented program code in Appendix K on page 142 When the program opens it automatically loads the most recently modified database Since a new database is desired selecting Close database in the file menu closes this old database A new database is created by selecting New database and the user is prompted to provide a name for this new database The software automatically configures the tables in this new database As mentioned previously it is typical to capture a number of versions of the same bubble configuration in order to improve the noise performance of the trained neural networks Set the number of frames under the edit menu 61 allows one to set the number of versions to capture Having configured the software one can now progress to capturing a database of readings Add to database in the edit menu opens a new window that is used for the capturing of database data Figure 5 19 is a screen capture of this form and a full scale version of this screen capture can be found in Appendix M on page 202 Firstly it is necessary to set the source of the data namely the top or bottom tomography system The right arrow is clicked to create space in the database for a new set of samples The desired network output is then drawn on the tomography component in the top left of the form Simply by clicking on a pixel the phase of tha
406. s on Sensing Field of Capacitance Tomographic Sensor Proc 1 World Cong on Industrial Process Tomography Buxton pp 348 352 Wang M 1999 Three dimensional Effects in Electrical Impedance Tomography Proc 1 World Cong Industrial Process Tomography Buxton pp 410 415 Ma Y Wang KULA and Jiang C 1997 Simulation study of the electrode array used an ERT system Chem Eng Sci 52 2197 2203 Ma Y Xu LA and Jiang C 1999 Experimental Study of the Guard Electrodes in an ERT System Proc 1 World Cong on Industrial Process Tomography Buxton pp 335 338 Yang W Q and York T A 1999 New AC based capacitance tomography system Proc 146 47 53 Lucas G P Cory J C and Waterfall R C 2000 A six electrode local probe for measuring solids velocity and volume fraction profiles in solids water flows Meas Sci Technol 11 1498 1509 Pinheiro P A T Loh W W and Dickin F J 1998 Optimal sized electrodes for electrical resistance tomography Electron Lett 34 69 Dyakowski T 1996 Process tomography applied to multi phase flow measurement Meas Sci Technol 7 343 353 National Semiconductor 1999 LMF100 High Performance Dual Switched Capacitor Filter datasheet Baluja S 1994 Population Based Incremental Learning A Method for Integrating Genetic Search Based Function Optimization and Competitive Learning Carnegie Mellon University Giannopoulos A 2003 Investigation of sine wave inputs for an FDM EIT system MSc Thesis University of Cape Town Eagle Tec
407. s update timer and exit procedure TAddToDatabase AddExitClick Sender TObject begin StatusUpdate Enabled false Close end when create form initialise the tables and the chart displaying the current volume fractions procedure TAddToDatabase FormCreate Sender TObject begin AddResTable Cells 0 1 TxA AddResTable Cells 0 2 TxB AddResTable Cells 0 3 TxC AddResTable Cells 0 4 TxD AddResTable Cells 0 5 AddResTable Cells 0 6 AddResTable Cells 0 7 TxG AddResTable Cells 0 8 TxH AddResTable Cells 1 0 RxA AddResTable Cells 2 0 RxB AddResTable Cells 3 0 RxC AddResTable Cells 4 0 RxD AddResTable Cells 5 0 RxE AddResTable Cells 6 0 AddResTable Cells 7 0 RxG AddResTable Cells 8 0 RxH AddCapTable Cells 0 1 TxA AddCapTable Cells 0 2 TxB AddCapTable Cells 0 3 TxC AddCapTable Cells 0 4 TxD AddCapTable Cells 0 5 AddCapTable Cells 0 6 TxF AddCapTable Cells 0 7 TxG AddCapTable Cells 0 8 AddCapTable Cells 1 0 RxA AddCapTable Cells 2 0 RxB AddCapTable Cells 3 0 RxC AddCapTable Cells 4 0 RxD AddCapTable Cells 5 0 RxE AddCapTable Cells 6 0 RxF AddCapTable Cells 7 0 RxG AddCapTable Cells 8 0 RxH AddVolume SeriesList 0 Add AddInputTomo WaterVolume AddVolume Se
408. should result in improved reconstruction accuracy since the output voltage readings would not contain the ripple inherent in a square wave excitation system A sine wave excitation system is currently being developed for the 8 electrode prototype rig to assess the validity of this statement Synchronised transmitters were used in the current system to minimise the cross plane interference when both systems operate simultaneously A drawback of this approach is that the received signals for a particular plane of electrodes also contain components resulting from the transmitters of the other plane of electrodes Subsequently a large disturbance passing through the first plane of electrodes will also introduce a change to the received signals at the second plane of electrodes Ideally an FDM impedance tomography system would use different sets of frequencies at the two measurement planes so that synchronised transmitters are not required and the synchronous detectors automatically reject any interference components Although this could not be implemented in the current square wave excitation system the sine wave system being developed has made provision for the generation of different sets of frequencies at the two measurement planes 108 Based on these conclusions it can be seen that good results have been achieved using the dual plane 16 electrode FDM impedance tomography system A literature study revealed that this is to the best of the author s kn
409. si Pixel tinttostr col tempvalue TestOutTable FieldByName s1 Value if tempvalue 0 then tempair true else if tempvalue 2 then tempgravel true end imgcol 0 if the current phases are the correct phases Il for the desired reconstruction then load the data normally for col 1 to 100 do begin if RecDesirePhase ItemIndex 1 and tempair and not tempgravel or RecDesirePhase temIndex 2 and not tempair and tempgravel or not tempair and not tempgravel or RecDesirePhase ItemIndex 0 then begin S1 Pixel inttostr col tempvalue TestOutTable FieldByName s1 Value if tempvalue lt gt 3 then begin if threephase then begin OutputData row imgcol 0 OutputData row imgcol 1 0 OutputData row imgcol 2 0 if tempvalue 0 then OutputData row imgcol 2 1 else if tempvalue 1 then OutputData row imgcol 1 else OutputData row imgcol 1 1 inc imgcol 3 end 159 else begin if tempvalue 1 then OutputData row imgcol 1 else OutputData row imgcol 1 inc imgcol end end end else set a flag so that the data can be removed at later stage else if RecDesirePhase ItemIndex 1 and tempgravel or RecDesirePhase ItemIndex 2 and tempair then begin OutputData row imgcol 0 inc imgcol if imgcol 88 then imgcol 0 end end end else desired output is volume fraction prediction else begin determine the phases in the
410. signal from a specific receiver as can be seen in figure 5 12 on page 45 As mentioned previously the low pass filtering in the 8 electrode system did not have a steep enough rolloff to prevent harmonics from the multiplication stage passing through the filter and appearing as ripple on the DC output voltages To correct this the 16 electrode system uses switched capacitor filters Specifically LMF100 dual Switched capacitor filters are employed The LMF100 can be configured to provide either a single fourth order low pass filter or two independent second order low pass filters Due to the high component count of the 16 electrode System it was decided that the LMF100 would be configured as two second order low pass filters and would perform the filtering of both the capacitance and conductance signals for each channel An advantage of switched capacitor filters is that their cutoff frequency is set by an external clock input Consequently the clock inputs of the Switched capacitor filters on the synchronous detector boards were tied together and were driven by the same frequency generator as transmitter G This was an important requirement since it was not known beforehand what low pass filter cutoff frequency would be required to achieve minimum output ripple If standard active filters had been used then a change in the cutoff frequency would have required replacing a considerable number of capacitive and resistive components The ratio of the cutoff
411. simply displays the neural network reconstruction results for each of the test data points sequentially It also displays the desired network output so that a comparison can be made A screen capture of this form is included in Appendix N on page 207 Validate current samples is an extension of this where instead of using data from the test database voltage readings are continuously captured from the impedance tomography rig and the reconstructions performed and displayed In this situation there is no desired output Instead a comparison is made between the network prediction and the bubble configuration within the rig at that point in time The desired frame rate can be set although the maximum achieved is typically only 20frames s due to the simplicity of the software implementation If the maximum frame rate of 280frames s is required then the Real time sampling form must be used The operation of this unit will be examined in greater detail when the dynamic performance of the FDM impedance tomography system is analysed in Chapter 7 65 CHAPTER 6 STATI C SI TUATI ON RESULTS FOR THE 16 ELECTRODE I MPEDANCE TOMOGRAPHY SYSTEM The following chapter examines the performance of the 16 electrode FDM impedance tomography system for static bubble configurations Further these tests are performed with only one of the measurement sections operating at any one time Performance assessments are generally made with respect to a reconstruction algorithm d
412. sing the back propagation algorithm 29 For the output layer of neurons the derivative of the error function with respect to the network weights is calculated as 29 ow 9 GO uz 3 11 where g a is the derivative of the output neuron activation function for combined input and z is the output of hidden layer neuron i with weight wy connecting to neuron the output layer For the hidden layer neurons the task of calculating the error derivative is complicated by the fact that there is no defined target value Although the accompanying software includes hidden layer capabilities these will not be discussed here since an explanation of the hidden layer calculations requires a thorough analysis of the back propagation technique For those wishing to read further Bishop provides an in depth analysis of the back propagation technique in 29 pp 140 148 16 Secondly these computed derivatives are used to adjust the network weights so as to minimise the error function Gradient descent and Resilient back propagation were implemented in this system to perform this particular task Gradient descent is a process whereby the gradient information is used to adjust the weights so as to move a small distance in the weight space in the direction in which the error function decreases most rapidly 29 This is achieved using the following weight update equation 29 ES u w t w t 1 3 12 Wi t 1 2 wi t n
413. sistently differentiate between the gravel and air phases This is believed to be a result of the limitations in the accuracy and resolution of the capacitance measurement circuitry resulting from the use of square wave excitation signals although this assumption has not yet been verified Neural networks were also trained to perform an image reconstruction of the pipeline contents although these were used primarily for verification purposes 107 Various concerns existed regarding the use of neural networks in applications where a distributed multi phase mixture is to be monitored instead of contiguous masses Initial tests conducted on a homogenous bubble flow illustrated that the neural networks developed were capable of imaging such flows However due to the simplicity of the test apparatus it was not possible to confirm whether the reconstruction results equated to the situation within the measurement volume The neural network reconstruction accuracy for situations involving the masking effect is improved by including examples of masking in the training database Further the reconstruction resolution is improved at the centre of the pipeline by not allowing any of the bubbles in the training database to come into contact with the conductance probes The drawback is that distorted reconstruction results are achieved if one or more conductance probes are immersed in air Whether this presents a problem or not depends on the specific applicat
414. son between the freguency spectrum measured using a spectrum analyser and the spectral peaks predicted by the simulation software which are plotted as sguare markers for the transmitter B receiver G electrode pair 56 Block diagram of sample controller board 58 Diagram illustrating the interactions between the calling application sample unit and EDR device driver where the arrows indicate the program execution required to capture a set of frames using streaming 60 Screen capture of the database generation form 62 Screen capture of the neural network parameter form 64 Drift analysis of both capacitance and conductance readings for transmitter A receiver A and transmitter C receiver A electrode pairs over a 15 hour period 67 Drift analysis of the capacitance between transmitter A and receiver A with the measurement section full of air 68 Photograph of the bed of hooks base plate and clear acrylic lid used as the bubble placement System for the generation of static tests 70 Photograph of the three sizes of air and gravel bubbles corresponding to 30 20 and 10 of the pipeline diameter respectively 71 Screen captures of test cases for the two phase air water image reconstruction of the top frequency division multiplexed impedance tomography system where the desired outputs are on the right and the network predictions are on the left in each screen capture 74 Screen captures of test cases for a three phase air gravel water image reconst
415. sponding to a particular current injection pair can be measured simultaneously simply by replicating the receiver circuitry 27 This reduces the data acquisition times and removes the phase shift associated with additional multiplexing stages 27 By using parallelism in the measurement it has been shown that higher frame rates can be achieved Frequency division multiplexing can be considered as an extension of the use of parallelism to achieve extremely high frame rates 18 41 Frequency Division Multiplexed Impedance Tomography Concept As previously mentioned the use of parallelism in impedance tomography greatly increases the potential frame rate of the system Frequency division multiplexed FDM impedance tomography makes use of this concept to achieve extremely high frame rates Each electrode is either a dedicated transmitter or receiver electrode Specifically every second electrode is a transmitter electrode with the electrodes in between being receiver electrodes This is illustrated in figure 4 1 for the case of an 8 electrode system Included in figure 4 1 are linkages showing the possible measurement permutations for this configuration All the transmitters operate simultaneously RECEIVER Consequently each receiver will receive signals from TxA all the transmitters simultaneously For example RxD TRANSMITTER receiver A will receive a signal from transmitter A modulated by the material between receiver A and transmitter
416. stage 2 then begin trainstartindex currentindex strtoint InputStart Text trainstopindex currentindex strtoint InputStop Text inc currentindex 162 if currentindex gt numtrainregions then begin InputPanel Visible false MainPrompt Caption Specify the number of testing regions currentindex 1 MainEdit Text 1 currentstage 3 MainEdit SetFocus end else begin InputPrompt Caption Specify the bounds for region inttostr currentindex InputStart Text inttostr strtoint InputStop Text 1 InputStart Enabled false InputStop Text InputStop SetFocus end end else if currentstage 3 then begin numtestregions strtoint MainEdit Text currentstage 4 nputPanel Visible true nputPrompt Caption Specify the bounds for region inttostr currentindex nputStart Text 1 nputStart Enabled false nputStop Text nputStop SetFocus end else if currentstage 4 then begin teststartindex currentindex strtoint InputStart Text teststopindex currentindex strtoint InputStop Text inc currentindex if currentindex gt numtestregions then begin CalibOK ModalResult mrOK currentstage 5 CalibOK Click end else begin InputPrompt Caption Specify the bounds for region inttostr currentindex InputStart Text inttostr strtoint InputStop Text 1 InputStart Enabled false InputStop Text InputStop SetFoc
417. stems were tested with the same air gravel seawater phase configurations and the same neural network topologies were used to perform the reconstructions of both systems Hence any difference in the reconstruction accuracy is related specifically to the measurement technique employed namely either TDM or FDM impedance tomography 30 Firstly to verify the operation of the system a single layer feed forward neural network was trained using gradient descent to perform a two phase air seawater image reconstruction The operation of this neural network was examined in figure 3 1 on page 12 Early stopping was used to prevent over fitting Since the neural network performance depends largely on the random initialisation of the network weights three networks were trained and their results averaged to give a more representative indication of system performance The neural network reconstruction results are detailed in table 4 3 below which also gives the corresponding performance results of the TDM impedance tomography system The white cells indicate which of the compared cell entries have the better performance TABLE 4 3 COMPARISON BETWEEN THE PERFORMANCE OF THE FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM AND A STANDARD TIME DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM FOR A TWO PHASE AIR SEAWATER IMAGE RECONSTRUCTION USING A SINGLE LAYER FEED FORWARD NEURAL NETWORK ALL RESULTS REPRESENT THE ABSOLUTE ERROR AND ARE ROUNDED OFF TO
418. stent with the results of the two phase air water image reconstruction TABLE 6 4 COMPARISON BETWEEN THE PERFORMANCE OF THE 16 ELECTRODE FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM AND THE 8 ELECTRODE PROTOTYPE FREQUENCY DIVISION MULTIPLEXED IMPEDANCE TOMOGRAPHY SYSTEM FOR A THREE PHASE AIR GRAVEL WATER IMAGE RECONSTRUCTION USING A SINGLE LAYER FEED FORWARD NEURAL NETWORK ALL RESULTS REPRESENT THE ABSOLUTE ERROR AND ARE ROUNDED OFF TO TWO SIGNIFICANT DIGITS PERFORMANCE MEASURE 16 ELECTRODE LABORATORY SCALE FDM 8 ELECTRODE PROTOTYPE IMPEDANCE TOMOGRAPHY SYSTEM FDM IMPEDANCE TOP SYSTEM BOTTOM SYSTEM TOMOGRAPHY SYSTEM THRESHOLD ERROR 36 13 AIR VOLUME FRACTION ERROR 29 GRAVEL VOLUME FRACTION ERROR 5 3 WATER VOLUME FRACTION ERROR 5 6 SUM VOLUME FRACTION ERROR 14 Figure 6 6 on page 76 includes screen captures of specific test cases from the top measurement system For each Screen capture the desired output is the right hand frame and the network prediction is the left hand frame These test cases confirm the conclusions of the two phase air water image reconstruction results However for the three phase reconstruction it is evident from figure 6 6 that the neural network is not always able to distinguish between the gravel and air phases In certain test cases the 0 04 volume fraction gravel bubble is occasionally detected as an air gravel combination or vice
419. t gt besteverfit then begin besteverfit bestfit res MultOut bestreal BestFreqFit Caption floattostrf res ffFixed 15 5 for i 1 to nvars prec do begin besteversol i bestsol i if i lt nvars then besteverreal i bestreal i end Qsort besteverreal 0 7 Freq Caption floattostrf Freq2 Caption floattostrf Freq3 Caption floattostrf Freq4 Caption floattostrf Freq5 Caption floattostrf Freq6 Caption floattostrf Freq7 Caption floattostrf Freq8 Caption floattostrf end end Application ProcessMessages end end besteverreal 1 ffFixed 15 4 besteverreal 2 ffFixed 15 4 besteverreal 3 ffFixed 15 4 besteverreal 4 ffFixed 15 4 besteverreal 5 ffFixed 15 4 besteverreal 6 ffFixed 15 4 besteverreal 7 ffFixed 15 4 besteverreal 8 ffFixed 15 4 E E since PBIL is inherently random in nature the first result achieved is not 1 necessarily the best therefore the process should be repeated multiple times the best results then used this procedure simply implements multiple PBIL searches keeping track of the best results procedure TPBILMain MultiSearchPBILClick Sender TObject var currentiter i integer overbestfreq double multirecord textfile filestring string begin LoadFreqsClick Sender overbestfreq 1 for currentiter 1 to Numlterations Value do begin CurlterCaption Caption inttostr currentit
420. t particular pixel changes and depending on which mouse button was pressed to either seawater gravel or air The bar chart simply displays the corresponding volume fractions for each of the three phases Having set the desired network output in software it is necessary to replicate this bubble configuration in the actual rig Once complete Start sampling is clicked and the software will capture a set of frames corresponding to this bubble configuration where the number of frames captured equals the desired number of versions To add another entry to the database the right arrow is once again clicked and the process is repeated The left and right arrows allow one to navigate through the database and if the end of the database is reached creates space for another set of samples The same process is followed for the generation of a test database Add sampled data to database x Set the desired network output ST Qi Top System Bottom System Voltage readings obtained From the impedance tomography system Im p 12 228027 229738 292666 243408 2 28516 229492 7 248779 249512 228770 228027 248047 242408 2 39502 5 2 36816 2 43896 245117 231445 2 31689 2 37305 243164 240723 2 36084 248535 251465 243652 251708 247070 246826 249268 240723 249756 247559 246094 250244 2 49023 2 46094 2 46338 2 48535 2 56348 257324 249756 249512 262695 260986 256836 263916 256836 259277 olume fractions For the current test case 2 53906 253906 2438
421. t uppersysbot 0 5 gravdiam and gravtopposition gravleng 0 5 gravdiam lt uppersystop then begin tempheight uppersystop gravtopposition gravleng 0 5 gravdiam uppergravvol gravhemivol pi tempheight power 0 5 gravdiam 2 end else if gravtopposition gravleng 0 5 gravdiam gt uppersystop and gravtopposition gravleng uppersystop then begin tempheight gravtopposition gravleng 0 5 gravdiam uppersystop uppergravvol gravhemivol pi tempheight power 0 5 gravdiam 2 1 3 pi power tempheight 3 end else uppergravvol 0 end else begin if gravtopposition gt uppersysbot and gravtopposition gravleng lt uppersysbot then uppergrawvol pi power 0 5 gravdiam 2 gravtopposition uppersysbot else if gravtopposition gravleng gt uppersysbot and gravtopposition lt uppersystop then uppergravvol pi power 0 5 gravdiam 2 gravieng else if gravtopposition uppersystop and gravtopposition gravleng uppersystop then uppergravvol pi power 0 5 gravdiam 2 uppersystop gravtopposition gravleng else uppergravvol 0 end end this procedure calculates the mean absolute error between the desired volume and the real time volume fraction data To compensate for slippage between the belt and pulley of the stepper motors the real time volume fraction data is shifted to coincide with the desired volume fraction data As the bubble passes through the measuring section a bell shaped volume frac
422. t Series 0 AddY outputstore testindex prediction 1 AirPlot Series 1 AddY outputstore testindex prediction 3 GravelPlot Series 0 AddY outputstore testindex prediction 2 GravelPlot Series 1 AddY outputstore testindex prediction 4 end else begin AirPlot Series 0 Add Y volumestore testindex 184 AirPlot Series 1 Add Y volumestore testindex 3 GravelPlot Series 0 Add Y volumestore testindex 2 GravelPlot Series 1 AddY volumestore testindex 4 end AirCorrPlot Series 0 AddY aircorr testindex GravelCorrPlot Series 0 AddY gravelcorr testindex end end tM STEPPER CONFIGURATION PROCEDURES configure the stepper motors for a test procedure TRealMain MotorSetupClick Sender TObject begin Stepperlnit StepperConfig ShowModal end stop stepper drives immediately procedure TRealMain StepperStopClick Sender TObject begin Stepper StopDrive end start stepper drives based on the current test configuration calculate the direction speed and distance for a fixed distance move procedure TRealMain RealStartMotorsClick Sender TObject var tempsteps integer begin StepperStop Visible true Stepperlnit if StepperConfig airrps gt 0 then begin if StepperConfig FlowDirSelect ItemIndex 0 then begin if StepperConfig AirBubblePosition ItemIndex 0 then tempsteps StepperConfig airsteps else tempsteps StepperConfig airsteps end else
423. t such a saturation would not occur the amplifications of all op amp stages within the system were reduced so as to provide a safety margin As an example the gains of the LMF100 switched capacitor filters were reduced so that the DC output voltage levels were at approximately 2 5V even though the LMF100 switched capacitor filters were on a 5V split supply and could produce output signals greater than 3 7V The output capacitance and conductance voltages were all tuned to be approximately 2 5V with the measurement section full of seawater As mentioned previously these voltages included a ripple component 66 resulting from the harmonic mixing in the multiplication stage of the synchronous detectors The average magnitude of this ripple was 21mV peak to peak for the capacitance readings and 14mV peak to peak for the conductance readings Having tuned the FDM impedance tomography system it was necessary to analyse the long term stability of these readings 6 1 Analysis of Capacitance and Conductance Measurement Drift To analyse the long term stability of the capacitance and conductance voltage readings the sampling software was modified to include a data logging feature This feature saves frames of data at a set rate to a comma separated text file which is then loaded into a spreadsheet for analysis Since only a drift analysis is required a logging rate of one frame every minute should give an indication of how the readings vary over a 24 hour p
424. t tempfile for index 1 to 128 do ReadLn tempfile bottommean index for index 1 to 128 do ReadLn tempfile bottomstddev index CloseFile tempfile end tM GENERAL PROCEDURES when exit deallocate the dynamic memory and stepper motor resources and save the settings to the registry procedure TRealMain RealExitClick Sender TObject var Reg TRegistry begin Reg TRegistry Create try Reg OpenkKey Software Neural True Reg Writelnteger RealOnline RealOnline ItemIndex Reg Writelnteger TopTestNumber TopTestNum Value Reg Writelnteger BottomTestNumber BottomTestNum Value Reg Writelnteger RealFrameRate RealFrameRate Value Reg Writelnteger DesiredFrames DesiredFrames Value Reg WriteString FlowFactor FlowFactorlnput Text Reg WriteString Separation SystemSeplInput Text Reg WriteString TopAirThresh TopAirThresh Text Reg WriteString TopGravThresh TopGravThresh Text Reg WriteString BotAirThresh BotAirThresh Text Reg WriteString BotGravThresh BotGravThresh Text Reg WriteString CorrThresh CorrThreshInp Text Reg Writelnteger ViewInc Viewlnc Value finally Reg Free end topweights nil tophiddenweights nil tophiddenoutputs nil bottomhiddenoutputs nil bottomweights nil bottomhiddenweights nil Stepper Free Close end when open form load previous settings from the registry procedure TRealMa
425. tal conductance probe and the ionic solution It has been shown that the contact impedance becomes negligible for the case where only one of the phases is conductive provided a sufficiently high excitation frequency is used 42 An excitation above 1kHz is required for potable water 42 In addition it must not be excessively high so as to avoid unwanted electromagnetic effects and external noise The output frequencies are generated using a PIC16F84A microcontroller with a 4MHz crystal It is necessary to generate four output square waves with 50 duty cycle as well as four 90 phase shifted versions for the synchronous detection of the capacitance components A microcontroller provides a convenient and simple solution to the generation of these quadrature waveforms Port B was used to output the four square waves and their quadrature waveforms In particular port BO was used to output Ao port B1 was used to output the quadrature waveform and so on The full details are provided in the circuit diagram in Appendix B on page 118 24 The output waveforms were initially generated by program loops Tests revealed that this technique was unable to generate frequencies high enough to extract a usefully measurable capacitance component It was deemed necessary to generate frequencies in the range from approximately 20kHz to 80kHz In addition this technique was unable to guarantee the 5096 duty cycle and 90 phase shift necessary to accurately separ
426. tal input output lines on the PC30G card as was done for the 8 electrode system However these port values must be set through software Consequently any advantages that may have been gained through the use of streaming are negated by the fact that every 16 samples the software has to increment the address lines of the multiplexers for the next set of 16 readings As expected the use of DMA only results in a performance increase if a large number of samples can be captured without software intervention Some other technique of setting the multiplexer address lines was therefore required It was decided that faster sampling could be achieved if the address lines of the multiplexers were driven by external hardware and figure 5 17 on page 58 is a block diagram of this sample controller board A single sample controller board is used to drive both impedance tomography systems and is configured using the digital input output lines of the PC30G card In contrast to the previous system these digital lines are only used to configure the sample controller board before and after a complete set of data has been captured and therefore do not affect the sampling rate of the system 57 SAMPLE CONTROLLER BOARD EXTERNAL TRIGGER PA1 PC30G CARD CHO RESET 555 OSCILLATOR 555 OUTPUT 20us 180us CD4516 4 bit COUNTER V CD4011 counter output CH15 On the negative edge the count
427. tance and conductance readings With adjacent transmitter receiver electrode pairs both the capacitance and conductance readings increase before decreasing as is evident in figure 7 11 This may be related to the phenomenon discussed earlier where the introduction of an air bubble actually results in an increase in certain capacitance and conductance readings even though the readings are expected to decrease In contrast the transmitter receiver electrode pairs that are spaced further apart behave more as expected For example the capacitance reading of the transmitter B receiver A electrode pair only starts to decrease when the water level drops below the top of the measurement electrode In addition there is no noticeable change to the capacitance reading as the water level 97 drops below the conductance probes In contrast the conductance reading decreases almost immediately to zero as the water level drops below the level of the conductance probes for the transmitter B receiver A electrode pair A sequence of screen captures were taken of the top measurement plane as the water level was lowered and these screen captures are displayed in figure 7 12 The sequence of screen captures is from left to right and only every 50 frame is captured As can be seen the drop in the water level is initially detected as an air bubble at the centre of the pipeline As the water level drops so the neural network image reconstruction results become more dis
428. tem Tests were conducted for static three phase configurations and the reconstructions were performed using a volume fraction predictor neural network The performance of the volume fraction predictor neural network was assessed on a database of test cases using the volume fraction error which is the mean absolute error in the volume fraction predictions for each of the three phases Image reconstructions were also performed but purely for verification purposes In a surprising result the FDM impedance tomography system outperformed the TDM impedance tomography System for the air gravel and water volume fraction predictions Specifically the air volume fraction error was reduced from 8 9 for the TDM impedance tomography system to 7 4 for the FDM impedance tomography system All results represent the absolute error and are rounded off to two significant digits Further the gravel volume fraction error dropped from 5 0 to 4 7 and the water volume fraction error dropped from 6 5 to 4 8 This improvement in performance is believed to be the result of the FDM impedance tomography system providing individual paths for the measurement of both the capacitance and conductance of the different electrode combinations Subsequently each synchronous detector can be tuned individually for the optimum measurement range and gain of the corresponding transmitter receiver electrode pair This is in contrast to the TDM impedance tomography system where a single me
429. ten using trigonometric identities as follows t STARA sinfgw gt Omama foos 2w gt 1 4 12 wat w At cos wgt Wat G C 4L ATXBRXA sin wet w sin wgt Wat VcMuLTA t xA ae if Wi pr lt min 2w wg W Wg wal 4 13 Consequently the multiplier output is proportional to the capacitance between receiver A and transmitter A provided the cutoff frequency of the low pass filter is less than the minimum of 2Wa Wg Wa OF Wg Wa 22 To determine the capacitance between transmitter B and receiver A the received signal is multiplied by a 90 phase shifted version of the reference for transmitter B and the multiplier output is G TxARxA Vcmutte t 5 sinw at Wet sin w At wgt cos w w gt cos w wgt 4 14 Sena fsinfow gt STEA foos 2wt 1 if Wipe lt minf2wg Wa 4 15 The above theoretical examination can also be performed for receiver B It is important to distinguish FDM impedance tomography from spectroscopic frequency tomography Dielectric spectroscopy is the technique of obtaining the permittivity of a material at different excitation frequencies 39 and has long been used for material characterisation As an example the system developed by Georgakopoulos et al is for a frequency range from 10kHz to 1MHz Various excitation techniques e
430. ters This was eliminated using a simple RC low pass filter on the output of the LMF100 The cutoff frequency of this RC combination was set at 234Hz It was decided that the convenience provided by the LMF100 to vary the cutoff frequency of all the filters in parallel far outweighed the above mentioned drawbacks Further since the design allowed for maximum signal range and gain the effects of the noise introduced by the switched capacitor filters could be minimised The circuit details of the low pass filters can be seen in Appendix C on page 126 44 v GVON N3AISOS8 TVNOIS 3ONVIIOVdYO Qquvoag gt 431113 55 331113 SSvd MO1 43040 HL4NO3J ES ONV HXL OXL N33M138 N33M138 BONVLIOVdYD 3ONVLIOVdVO 31113 55 43040 HL4n03 Yala 55 YJOYOHLYNOA X31 SSvd MO1 WIJOYOHLYNOA 331113 SSvd MO1 WIJOYOHLYNOA wala SSvd MO1 M3QSOH Dno ONY 3XL Namaa BONVLIOVdYS ONY 3X1 N33M138 gt OXL WXL N33A44138 SNI1dWvS 403 GVO NOIVN3N3O AON3nO383 O1 SOWLIOA 4nd4no aaxandicniw QNuvO8 HO1VN3N39 AONSNODAYS WON3 7IOHINOO N3xa laLnniw Waxandirnw Vx ONY HXL ONY OXL Wea ONY 3XL Yx ONY 3X1 VXM ONY VXL N33M138 N33M138 N33M138 8 N33M
431. the 100 harmonic of the fundamental frequency for all eight transmitters are incorporated but only those harmonics whose magnitudes are greater than some minimum threshold are included in the later stages of the simulation Consequently the attenuation of the low pass and bandpass filters in the capacitance and conductance receivers are incorporated in this calculation Specifically the initial magnitude of harmonic c is calculated using the Fourier transform of a unity magnitude square wave as L For a fundamental frequency the gain of the low pass filter for harmonic c is calculated as 1 The gain of vi ee Ci pg the bandpass filter is calculated in a similar way The final magnitude for harmonic c is calculated as gain pp x gain gpp TER and is only included in the later stages of the simulation if this magnitude is greater than nC the specified threshold For completeness a DC component is included in the received signal to compensate for any DC offsets resulting from electrochemical reactions component offsets and so on 2 Sortthe remaining received frequencies after thresholding and check for duplication Specifically if duplicate frequency components are generated then the minimum frequency separation can immediately be calculated as OHz Duplicate frequency components indicate that the frequencies of two or more transmitters are related and is an undesirable situation because of the following consi
432. the dual plane 16 electrode FDM impedance tomography System initial tests were conducted for the simpler configuration of a two phase air water reconstruction Although the emphasis of the research is the development of a mass flow meter which requires accurate volume fraction measurements image reconstructions were also performed at various stages during the testing as a means of verifying the correct operation of the tomography systems Where appropriate sequences of screen captures for specific tests will be included in the following sections 7 3 4 Real time volume fraction prediction and image reconstruction results The real time dynamic performance of the dual plane 16 electrode FDM impedance tomography system was tested using different air bubble configurations For all tests conducted the measured volume fraction profile was compared to the desired volume fraction profile and the dynamic volume fraction error was calculated Figure 7 8 is a plot of the results for such a test Specifically this test was for a 0 04 volume fraction air bubble positioned near the left hand edge of the rig and a 0 01 volume fraction air bubble positioned near the right hand edge of the rig Both bubbles were set to move upward at 0 2m s and both bubbles were set to start moving at the same time Hence the volume fraction of the air phase should peak at 0 05 when the bubbles were situated directly within the measurement volume In figure 7 8 the solid line is the
433. the edge of the pipeline and consequently the resolution at the centre was poor Tests were conducted to compare the performances of neural networks trained on data that included bubbles at the edges and data that did not Results showed that a greater resolution and accuracy was consistently achieved for the latter A drawback of this approach is that erroneous results are achieved when a bubble is allowed to come into contact with a conductance probe An example of this situation will be analysed in Chapter 7 Secondly the training database was extended to include combinations of air bubbles combinations of gravel bubbles and combinations of both air and gravel bubbles Previously the training database had consisted only of single air bubbles and single gravel bubbles In this situation the reconstruction of more complex configurations is dependent on the generalisation capability of the neural network Further it was shown that the neural network reconstruction algorithms consistently produced incorrect reconstructions for situations that contained the masking effect An example of the masking effect is the configuration consisting of a bubble near the edge and a bubble at the centre In this configuration the neural network was unable to detect the bubble at the centre and only detected the bubble near the edge since the reconstruction is dominated by the changes introduced by the bubble near the edge Although not verified at the time it was believ
434. the sampling function TSamp GetSampleStatus begin GetSampleStatus busy sampling end end PROGRAM LISTING OF TOMO PAS This unit implements a tomography component which provides a convenient way to display the contents of a pipeline as well as a way to specify the contents of a pipeline for training database generation The component segments the pipeline into a 10 by 10 matrix of pixels where the phase of each pixel can be either water gravel or air Of the 100 pixels only 88 fall within the area of the pipe and the rest are classified as outpipe unit Tomo interface uses Windows Messages SysUtils Classes Graphics Controls Forms Dialogs ExtCtrls type TBubblePhase air water gravel outpipe dedicated type to identify the phase of a pixel TOutput array 0 9 0 9 of TBubblePhase dedicated type to store the 10 by 10 matrix of pixels I where each pixel is of type TBubblePhase TTomo class TPaintBox 1 component inherits from the standard Delphi paintbox private Private declarations FAllowEdit Boolean FWaterVolume Integer FAirVolume Integer FGravelVolume Integer FXPosition Integer FYPosition Integer FDesired TOutput procedure SetDesired row integer col integer value TBubblePhase protected Protected declarations procedure MouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Integer virtual procedure MouseMove S
435. though simple to implement linear back projection is flawed and errors occur in the reconstruction resulting from the soft field nature of electrical sensors A soft field sensor is one for which the contents of the vessel distort the measurement field Consequently the sensitivity of the sensor depends on the component distribution within the region of interest 26 Various iterative reconstruction techniques exist as a solution to this problem Iterative reconstruction algorithms use the equation describing the forward problem to update the reconstructed images so that it converges on a more accurate solution The process of iterative reconstruction algorithms for the specific example of capacitance tomography is described as follows 26 1 an initial image is calculated using the linear back projection algorithm 2 the forward problem is solved for this approximate distribution using a finite element modelling technique and a set of capacitances are calculated for this approximate distribution 3 the error between the measured capacitances and these back calculated values is used to calculate an error distribution this error distribution is used to update the initial approximate distribution this process is iterated until a satisfactorily small error between the measured and back calculated capacitance readings is obtained It is then necessary to check that the distribution did converge on a solution 11 In particular measurement nois
436. through the rig it was decided that the performance of the system should be tested on the more realistic configuration of multiple bubbles moving at different velocities Initially these tests were only conducted with air bubbles although the next section will examine the results for an air and gravel bubble moving at different velocities 101 Tests were conducted with multiple air bubbles moving at different velocities through the laboratory scale rig For example figure 7 17 is a plot of the air volume fraction profiles at the top and bottom measurement planes for the configuration of two 0 04 volume fraction air bubbles Included in figure 7 17 is the cross correlation function of the volume fraction profiles One of the bubbles was set to move at 0 2m s while the other bubble was set to move at 1m s As expected these individual bubbles are detected at the top and bottom measurement planes with the response corresponding to the faster bubble being narrower than the response produced by the slower bubble In addition it is evident that the cross correlation function contains two peaks corresponding to the two different bubble velocities but it is the peak resulting from the slower bubble that is detected as the average velocity This is because of the wider response generated by the slower bubble volume fraction 96 volume fraction 96 biased crosscorrelation TOP SYSTEM VOLUME FRACTION MEASUREMENT T T
437. tion electrodes At the voltage levels typically used the predominant reaction is the electrolysis of water 19 This consists of the oxidation of one of the injection electrodes and the reduction of the other injection electrode 19 At the receiver the input amplifiers typically have high input impedances so minimal current flows between the ions in solution and the electrode Consequently no electrochemical reactions are present at the receiver electrodes 19 Since minimal current flows into the receiver electrodes the contact impedance between the electrode and solution as a result of electrode fouling only has a minimal effect The total net charge transported through the solution should be zero to ensure that its properties are not permanently changed 20 The bi directional current has the advantage of eliminating any long term polarisation effects at the electrodes 19 depending on the conductance probe material 20 The potential differences between the electrodes and solution that occur at an electrochemical interface may not all be the same resulting in voltage offsets These offsets may be the result of concentration fluctuations in the solution or variations in the electrode properties 20 The problems associated with these offset voltages will be re examined at a later stage The potential difference between adjacent electrodes is measured using a differential amplifier Input and output multiplexers temporarily connect the c
438. tion peak is created To improve the accuracy of the assessment the mean error is only calculated over the duration of this peak By detecting the peak of both the desired volume fraction data and the real time predicted data the data can be shifted so that the peaks conincide and any error as a result of slippage will be corrected procedure TDynVerify CalcVolErrorClick Sender TObject var testcase startmax stopmax integer maxvalue real begin II initialise variables topairerror 0 topgravelerror 0 botairerror 0 botgravelerror 0 airtopdesiremax 1 graveltopdesiremax 1 airbotdesiremax 1 gravelbotdesiremax 1 airtopstart 0 airtopstop 0 airbotstart 0 airbotstop 0 graveltopstart 0 graveltopstop 0 gravelbotstart 0 gravelbotstop 0 ErrorResults Visible true startmax 0 stopmax 0 maxvalue 0 air bubble at top measuring ring firstly find the peak of the real time predicted volume fraction data for testcase 1 to RealMain DesiredFrames Value do begin if totaloutputs 2 then begin if outputstore testcase prediction 1 gt maxvalue then begin maxvalue outputstore testcase prediction 1 startmax testcase stopmax 0 end else if outputstore testcase prediction 1 maxvalue then stopmax testcase end else begin if volumestore testcase 1 gt maxvalue then begin maxvalue volumestore testcase 1 st
439. tions of the air and gravel phases are correlated to determine the average velocity of each component This approach does not give any indication as to how the velocities of the individual components vary over the cross section of the pipeline However by cross correlating the individual pixels of two sequences of images the individual axial velocity vectors at each pixel can be obtained producing a velocity profile 14 For a two phase system 88 correlation algorithms would be performed instead of the single correlation based on component volume fraction Consequently the computation required would increase by a factor of 88 In certain applications the flow may take on spiralling forms Consequently every pixel in the first plane would need to be correlated with every other pixel in the second plane to give the average velocity vector of each pixel between the two planes 17 This approach will obviously increase the computation by a further factor of 88 Figure 7 1 is a plot showing the difference in computational requirements between each of these three techniques using a point by point cross correlator The computation is defined as the number of program instructions executed As the minimum velocity of the system decreases so the number of frames over which the correlation is performed increases for a given frame rate and hence the computation increases 1000000000 1tx average velocity 100000000 velocity profile tt ve
440. tocol to ensure that the readings are independent of any electrochemical reactions taking place within the vessel This was discussed in detail in Chapter 2 However the current FDM impedance tomography technique is limited to two electrode conductance measurements and is therefore sensitive to polarisation voltages as well as variations in the contact impedance between the conductance probe and the seawater Specifically corrosion results in the fouling of the conductance probe Since this contact impedance is included in the conductance measurement an increase in contact impedance results in a decrease in the measured conductance Further DC electrochemical potentials exist at the electrodes and these voltages differ for each electrode 37 Research is currently being done on the design of a FDM impedance tomography system to perform four electrode adjacent pair conductance measurements During a drift analysis it was observed that bubbles would form on the conductance probes The formation of these bubbles explains the almost oscillatory behaviour of the readings for transmitter C in figure 6 1 on page 67 As the bubble is formed so the contact impedance between the seawater and conductance probe increases as a result of the reduced contact area Consequently the conductance reading decreases steadily As soon as the bubble is large enough and breaks away the contact impedance drops and subsequently the conductance reading increases sharply F
441. tomography technique is that fewer measurements are available for a given number of electrodes compared to the standard capacitance and conductance tomography protocols 2 Specifically the number of transmitter receiver electrode pairs is where N is the number of electrodes Hence for an 8 electrode impedance tomography system only 16 capacitance and 16 conductance readings are available instead of 28 capacitance and 20 conductance readings Consequently a reduction in the resolution of the reconstructed images is expected To achieve the same level of resolution it is therefore necessary to increase the number of electrodes Further since separate transmitter and receiver electrodes are employed the sensitivity of the FDM system is effectively half that of the corresponding TDM system However as mentioned previously the use of separate transmitter and receiver electrodes ensures that no switching is required permitting extremely high frame rates to be achieved It is therefore a compromise between sensitivity and on line frame rate capability This compromise will have to be made at various other stages during project design as well Another drawback of the large parallelism associated with FDM impedance tomography is an exponential increase in cost as the number of electrodes increases For example it was fairly cost effective to develop the 8 electrode prototype system On increasing the number of electrodes to 16 the cost became consi
442. top Checked Reg ReadBool EarlyStop RecPerformModel Checked Reg ReadBool Model RecUseGradDescent Checked Reg ReadBool UseGradient RecSetDeltalnit Text floattostr Reg ReadFloat Deltalnit RecSetDeltaMax Text floattostr Reg ReadFloat DeltaMax RecSetLearn Text floattostr Reg ReadFloat LearnRate RecSetMomentum Text floattostr Reg ReadFloat Momentum RecNumFails Value Reg ReadInteger NumFails RecStopVolume Checked Reg ReadBool StopVolume RecModelStart Value Reg Readinteger ModelStart RecModelStop Value Reg Readinteger ModelStop RecModellnc Value Reg Readinteger Modellnc RecShowTrainPerf Checked Reg ReadBool ShowTrainPerf RecSetUpdateFreq Value Reg ReadInteger UpdateFreq end finally Reg Free end set the important variables if RecUseGradDescent Checked then RecRPropParam Visible false else RecRPropParam Visible true RecUseRProp Checked not RecUseGradDescent Checked RecStopThreshold Checked not RecStopVolume Checked if RecPerformEarlyStop Checked then RecSetEarlyParam Enabled true else RecSetEarlyParam Enabled false if RecPerformModel Checked then RecSetModelParam Enabled true else RecSetModelParam Enabled false if RecLoadDBase RecDesiredRecon ItemIndex 0 then RecStopThreshold Visible true else begin RecStopThreshold Visible false RecStopVolume Checked true e
443. torted until the water level has dropped below the conductance probes and the reconstructions stabilise Further it is observed that when the conductance probes are completely immersed in air the reconstructed images do not reflect this fact and instead retain a central disk of water This distortion in the reconstruction is probably the result of the training database not including an all air data point although this assumption has not been verified Further itis not known whether the inclusion of such a point would have a detrimental effect on the reconstruction resolution FIGURE 7 12 SCREEN CAPTURES OF THE TOP MEASUREMENT SYSTEM IMAGE RECONSTRUCTION AS THE WATER LEVEL IS LOWERED PAST THE CONDUCTANCE PROBES THESE IMAGE RECONSTRUCTIONS WERE PERFORMED ON LINE AND THEIR RESULTS DISPLAYED IN REAL TIME THE SEQUENCE OF THE FRAMES IS LEFT TO RIGHT AND ONLY EVERY 50 FRAME IS CAPTURED AS BEFORE THE BLACK IS THE WATER PHASE THE WHITE IS THE AIR PHASE AND THE LIGHT GREY REPRESENTS THE AREA OUTSIDE THE PIPELINE 7 3 8 Reconstruction results for a homogenous bubble flow The dual plane 16 electrode FDM impedance tomography system has only been trained and tested with contiguous polystyrene foam or gravel masses that are larger than the resolution of the system However the flow patterns in the intended application are likely to be distributed three phase mixtures and it is not known how a System trained on contiguous phases would respond to
444. trodes then the guard electrodes are not operating correctly In addition there should be no sudden changes resulting from the water level dropping below the ring of conductance probes In contrast the conductance readings should decrease rapidly to zero when the water level passes the ring of conductance probes Plots of the capacitance and conductance variations between transmitter A and receiver A and between transmitter B and receiver A are included in figure 7 11 Superimposed on these plots are the relative positions of the guard electrodes measurement electrode and conductance probe TRANSMITTERA TRANSMITTER B capacitance V capacitance V 100 200 100 200 time s time s TRANSMITTER A TRANSMITTER B conductance V conductance V 100 200 100 200 time s time s FIGURE 7 11 PLOTS OF CAPACITANCE AND CONDUCTANCE VARIATIONS FOR RECEIVER A AS THE WATER LEVEL IS LOWERED PAST THE TOP MEASUREMENT PLANE THE VERTICAL LINES INDICATE THE RELATIVE POSITIONS OF THE TOP GUARD PLATE THE TOP OF THE CAPACITANCE PLATE CONDUCTANCE PROBE THE BOTTOM OF THE CAPACITANCE PLATE AND THE BOTTOM GUARD PLATE RESPECTIVELY It is interesting to note that the response of the adjacent transmitter receiver electrode pairs are different to those of the electrode pairs that are spaced further apart This phenomenon is observed for both the capaci
445. truct the distribution of the phases within the pipeline provided the phases have different dielectric constants Resistance tomography is used to reconstruct the resistivity distribution within the cross section of the pipeline and operates in a similar way to capacitance tomography Impedance tomography can be described as a dual modal approach since both the capacitance and conductance of the different electrode combinations are measured to reconstruct the complex impedance of the material distribution Previous research has shown that impedance tomography can be used to reconstruct a three phase air gravel water mixture 3 4 In addition it has been shown that neural networks can be used to perform this reconstruction task 3 4 In particular a single layer feed forward neural network with a 1 of C output encoding can be trained to perform a three phase image reconstruction Further a double layer feed forward neural network can be trained to predict the volume fractions of the three phases within the flow directly based on the capacitance and conductance readings obtained from the data acquisition system However these tests were only for static configurations This thesis will readdress this problem from the dynamic viewpoint In addition the individual component velocities will be calculated using the cross correlation of the volume fraction predictions from two impedance tomography Systems spaced a certain distance apart Cross correlati
446. true RecTestStop Caption Stop tests end end when show form initialise the sampling hardware and start the timer procedure TRecTest FormShow Sender TObject begin curcalibrate false RecTestStop Caption Start tests SetLength testdata 129 set the form parameters according to the type of test being conducted if RecLoadDBase RecDesiredRecon ItemIndex 0 then begin Caption Test the network image reconstruction performance RecTestVolume Visible false end else begin Caption Test the network volume fraction prediction RecTestVolume Visible true end if usecurrentsamples then begin RecVoidLabel Caption Volume fraction predictions RecPredictLabel Caption Network prediction testindex 1 RecVoidNoteLabel Visible false RecDesiredLabel Visible false RecTestDesired Visible false StartCalibrate Visible true end else begin testindex RecLoadDBase GetTotalTrain RecVoidNoteLabel Visible true RecDesiredLabel Visible true RecTestDesired Visible true StartCalibrate Visible false end RecTestTimer Interval 10 I initialise the sampling hardware notifvalue SampHardware nitSample 1 Handle start the timer end when close form free memory allocated for testing procedure TRecTest FormClose Sender TObject var Action TCloseAction begin testdata nil testoutput nil testhidden
447. uction These tests should be conducted before any decision is made regarding whether the project should continue It should be noted that industrialising such an instrument would not be a trivial task In particular certain challenges would need to be overcome such as the selection of the pipeline material and the conductance probe material for the harsh environment of an offshore airlift In addition further testing should be done to assess how the system would perform on real flows that simulate the dynamics of the intended application more realistically Further recommendations specific to this particular system are discussed in Chapter 10 110 CHAPTER 10 REMAI NI NG SYSTEM ISSUES THAT STILL NEED TO BE ADDRESSED The following is a brief discussion of certain issues that must still be addressed in this particular system before further research is conducted The current system is based on the Eagle PC30G data acquisition card This card is intended for an ISA slot and is Soon to be obsolete since it is increasingly difficult to purchase computer motherboards that support the ISA standard It is recommended that a PCI based National Instruments card be purchased for future work The sampling software would need to be modified to use this new card However since the device specific calling procedures are encapsulated in a sampling class it would only be necessary to modify this sampling class The current data capture rate of the Eagle PC
448. ue to the large number of measurement variables in a tomography system Although the desired output is the volume fraction of each of the three phases and their corresponding component velocities image reconstructions will also be performed as a means of demonstrating the wide range of applications to which such a system is applicable The following sections will examine specifically the configuration and assessment of the top measurement section It should be noted that the same process was followed for the bottom measurement section due to the similarity of the results achieved this will not be discussed in detail Before training database generation could commence it was necessary to tune the system for optimal performance Firstly a blanking off disk was constructed to seal the top measurement section The flange clamps used to assemble the laboratory scale rig were used to secure the blanking off disk to the base of the measurement section with a rubber insertion gasket The measurement section was then filled with seawater and the system was switched on The sample controller board was connected to the PC30G card and the sampling software running on the reconstruction computer provided a means of monitoring all 128 voltages In this way it was possible to observe trends in the voltages as bubbles were introduced It was noted that the introduction of a bubble resulted in the increase of certain readings However for both the capacitance and
449. ugh all the test cases in the test database unit testunit interface uses Windows Messages SysUtils Classes Graphics Controls Forms Dialogs Buttons ExtCtrls TeEngine Series TeeProcs Chart Tomo StdCtrls Math Spin type TRecTest class TForm Paneli TPanel Panel2 TPanel RecTestReturn TSpeedButton RecDesiredLabel TLabel RecTestDesired TTomo RecPredictLabel TLabel RecTestNetwork TTomo RecTestVolume TPanel RecAirChart TChart Series1 TBarSeries RecGravChart TChart Series2 TBarSeries RecWaterChart TChart BarSeries1 TBarSeries RecVoidLabel TLabel Label4 TLabel Label5 TLabel Label6 TLabel RecVoidNoteLabel TLabel RecTestStop TSpeedButton RecTestTimer TTimer Label1 TLabel RecTestFrameRate TSpinEdit StartCalibrate TSpeedButton procedure RecTestReturnClick Sender TObject procedure RecTestStopClick Sender TObject procedure FormShow Sender TObject procedure RecTestTimerTimer Sender TObject procedure FormClose Sender TObject var Action TCloseAction procedure StartCalibrateClick Sender TObject private Private declarations testdata array of real input voltages for a specific test case testhiddenoutput array of real output of hidden layer neurons for I test case testoutput array of real network output for test case testindex integer index into testing database notifvalue cardinal II notification value for sampling
450. ult capacitance end 13 begin row 7 tresult resistance end 14 begin fow 7 tresult capacitance end 15 begin row 8 tresult resistance end 16 begin row 8 tresult capacitance end end for col 1 to 8 do begin if result capacitance then tempcaptable row col calibval index else temprestable row col calibval index inc index end end Il once have generated the resistance and capacitance tables write the tables out to the file WriteLn tempfile Capacitance calibration WriteLn tempfile RxA RxB RxC RxD RxE RxF RxG RxH for row 1 to 8 do begin col 1 case row of 1 begin outstring TxA end 2 begin outstring TxB end 3 begin outstring TxC end 4 begin outstring TxD end 5 begin outstring TxE end 6 begin outstring TXF end 7 begin outstring TxG end 8 begin outstring TxH end end while col lt 8 do begin outstring outstring floattostrf tempcaptable row col ffGeneral 5 4 inc col end WriteLn tempfile outstring end WriteLn tempfile Resistance calibration WriteLn tempfile RxA RxB RxC RxD RxE RxF RxG RxH for row 1 to 8 do begin col 1 case row of 1 begin outstring TxA end 2 begin outstring TxB end 3 begin outstring TxC end 4 begin outstring TxD end 5 begin outstring TxE end 6
451. um MOVWF PORTB NOP via division factor is limited to 4032 only by the X 00 J f m ine RE T div available program memory on the PIC16F84A NOP where Figure 5 14 illustrates these concepts A program standard MOVLW 00000000 crystal 7 1 2 lt xtal lt 20M Hz written in Delphi generates the assembler code xtal 64 lt lt 4032 used to achieve specific division factor PIC implementing division automatically The user specifies the desired factor div division factor and this program generates a text FIGURE 5 14 OUTPUT SQUARE WAVE GENERATION WITH 0 AND 902 file that is assembled using MPASM and then REFERENCES FOR THE SYNCHRONOUS DETECTION OF THE CONDUCTANCE AND CAPACITANCE SIGNALS RESPECTIVELY downloaded to the corresponding microcontroller The program code for this Delphi application can be found in Appendix on page 137 In addition the assembler code generated for a division factor of 64 is included in Appendix F on page 134 53 CRYSTAL CRYSTAL CRYSTAL CRYSTAL FREQA FREQC FREQE FREQG 0 0 0 GO z 0 z 0 0 0 o O O 2 90 Dy 90 Dy 90 Ay 90 8o 90 8o 8o 83 ora 83 Ora 83 ora ora n mao zm mzo zn mzo zn mao 42 30 2 Jo O DF Toa gt Baw zt Baa aS Aaa ON SES 932 9 lt gt 932 BROWN OUT DETECTOR 056 el 26 E BEFA m 292 m 292 m 29 29 i 8
452. ure TStepperConfig AirBubbleJogMouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Integer var tempdir TDirection begin upward movement if ssLeft in Shift then begin if AirDrive Value 1 or AirDrive Value 3 then begin if AirBubblePosition ItemIndex 1 then tempdir CW else tempdir CCW end else begin if AirBubblePosition ItemIndex 1 then tempdir CCW else tempdir CW end RealMain Stepper JogMove AirDrive Value tempdir end downward movement else if ssRight in Shift then begin if AirDrive Value 1 or AirDrive Value 3 then begin if AirBubblePosition ItemIndex 1 then tempdir CCW else tempdir CW end else begin if AirBubblePosition ItemIndex 1 then tempdir CW else tempdir CCW end RealMain Stepper JogMove AirDrive Value tempdir end end when the mouse button is released stop the bubble movement procedure TStepperConfig AirBubbleJogMouseUp Sender TObject Button TMouseButton Shift TShiftState X Y Integer begin RealMain Stepper StopDrive end jog the gravel bubble if left click then bubble moves upward else if right click then bubble moves downwards procedure TStepperConfig GravelBubbleJogMouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Integer var tempdir TDirection begin upward movement if ssLeft in Shift then begin if GravelDrive Value 1 or GravelDrive Value 3 then beg
453. urrent source to the injection electrodes and the differential amplifier to the detection electrodes A phase sensitive detector demodulates the differential amplifier output whose amplitude is modulated by the material within the vessel A low pass filter removes the switching harmonics to produce a DC voltage that is sampled by an analogue to digital converter 17 Chapter 3 examines the reconstruction algorithms for tomography systems in greater detail CHAPTER 3 IMPEDANCE TOMOGRAPHY RECONSTRUCTI ON ALGORI THMS It is the task of the reconstruction algorithm to determine an image of the contents of the vessel based on the limited capacitance and conductance measurements obtained from the data acquisition system 21 The following section briefly discusses conventional reconstruction algorithms before neural network reconstruction techniques are discussed 31 Standard Impedance Tomography Reconstruction Algorithms The linear back projection algorithm is common to both capacitance and resistance tomography systems Fundamental to this technique is the generation beforehand of a sensitivity matrix The image reconstruction is then performed in a single step through a matrix multiplication For the capacitance tomography reconstruction the sensitivity matrix describes how the capacitance readings for a particular electrode pair are affected by a change to the dielectric constant of a particular pixel within the vessel cross section 11 For resist
454. urst mode so that for every external trigger received all 16 analogue input channels are sampled at the maximum rate permitted e as mentioned previously streaming transfers the sampled data directly to memory InitSample instructs the PC30G card to use streaming and that the data should be transferred to memory in blocks of 1024 samples corresponding to four complete frames of data from both systems The larger the transfer block the more efficient the streaming A drawback is that samples can then only be captured in multiples of the transfer block size Consequently if only one frame were required from both systems an additional 768 samples would also be captured and then discarded The choice of the transfer block size is a compromise between these two factors e finally InitSample configures the EDR driver so that when the DMA transfer is complete a Windows notification message is sent by the driver to the application program InitSample then returns the message identification value for this notification message to the calling application 59 INITIALISE THE SAMPLE APPLICATION CONTROLLER CARD SAMPLE UNIT EDR DEVICE PROGRAM SOA DRIVER INITSAMPLE START THE CAPTURE OF STARTSAMPLE INITIATE THE SAMPLING A SET OF FRAMES AS A BACKGROUND USING STREAMING PROCESS DISABLE THE SAMPLE CONTROLLER CARD AND RETURN THE VOLTAGE SAMPLES STOPSAMPLE RELEASESAMPLE WNDPROC RELEASE THE RESOURCES ALLOCATED FOR SAMPLING WHEN THE SA
455. urther it was observed that the conductance probes would corrode over time Consequently it was decided that the drift of the capacitance and conductance readings was the result of the electrochemical reactions taking place within the rig and was not related to electronic component drift To prove this assumption correct a drift analysis was performed with the measurement section empty Subsequently no electrochemical reactions could take place and only electronic component drift would be e o iw ha o 7 c MALE HAN YH MI M time hours FIGURE 6 2 DRIFT ANALYSIS OF THE CAPACITANCE BETWEEN TRANSMITTER A AND RECEIVER A WITH THE MEASUREMENT SECTION FULL OF AIR monitored The resulting drift analysis is plotted in figure 6 2 which is the drift of the capacitance reading between transmitter A and receiver A As is evident from figure 6 2 minimal drift occurs when the measurement section does not contain seawater The remaining drift is probably the result of temperature variations From this result it was concluded that the capacitance and conductance voltage drift was the result of electrochemical reactions taking place between the conductance probes and the seawater 68 In order to minimise this drift various alternate materials were tested for the conductance probes Initially the conductance probes were 5mm stainless steel machine screws These wer
456. us end end end end PROGRAM LISTING OF VIEWUNIT PAS This unit displays the loaded data both the desired network output and the corresponding sampled voltages unit viewunit interface uses Windows Messages SysUtils Classes Graphics Controls Forms Dialogs Buttons ExtCtrls TeEngine Series TeeProcs Chart StdCtrls Grids DBGrids Db Tomo type TResultTable capacitance resistance TRecViewMain class TForm Panel1 TPanel RecViewCapTable TStringGrid Label3 TLabel Label5 TLabel Label4 TLabel RecViewResTable TStringGrid RecViewNext TSpeedButton RecViewPrevious TSpeedButton RecViewVolLabel TLabel RecViewVolume TChart Series1 TBarSeries Series2 TBarSeries Series3 TBarSeries RecViewReturn TSpeedButton RecViewlnputTomo TTomo RecViewlmgLabel TLabel ScrollTimer TTimer UseContinouos TCheckBox procedure RecViewReturnClick Sender TObject procedure FormCreate Sender TObject procedure FormShow Sender TObject procedure RecViewNextClick Sender TObject procedure RecViewPreviousClick Sender TObject procedure ScrollTimerTimer Sender TObject private Private declarations currentindex integer current index into database procedure DisplayData I display the desired network output procedure UpdateTables update the tables to show the corresponding input voltages public Public declarations end var RecViewMain TRecViewMain
457. us_low status_low shl 8 status_low shr 8 status status high status status shi 16 4 status low end set the fast status pointer to the status information that wish to access procedure TStep SetPointer status offsetinteger begin port address FASTSTATUS status offset end convert a string command into the appropriate format and send it to the AT6400 procedure TStep ProcessCommand temp string var CharPtr byte begin for CharPtr 0 to Length temp 1 do begin Command CharPtr temp CharPtr 1 end Command CharPtr ENTER Inc CharPtr Command CharPtr TERMINATOR Error SendAT6400Block Address Command CharPtr 0 Command CharPtr TERMINATOR end rM HIGH LEVEL GENERAL PURPOSE CONTROL PROCEDURES initialise the drives by reading a sequence of commands from an initialisation file procedure TStep Initialise NewAddress word readfilename string var initfile textfile temp string begin Address NewAddress AssignFile initfile readfilename Reset initfile while not SeekEOF initfile do begin read a command and then process it ReadLn initfile temp ProcessCommand temp end CloseFile initfile end general purpose procedure for initiating drive movement at a fixed velocity and in a certain direction procedure TStep GoDrive drive1 Boolean drive2 Boolean drive3 Boolean velt integer vel2 integer vel3 integer dir1 TD
458. utils simport windows const 11 AT6400 masks OB HAS DATA 01 AT6400 output buffer has data 5 EMPTY 02 AT6400 input buffer is empty GP INTERRUPT 04 AT6400 general purpose interrupt CS UPDATED 08 AT6400 card status update MASTER 10 AT6400 master interrupt enable KILL REQ 20 AT6400 kill has been requested OS LOADED 40 AT6400 operating system loaded LOADED 80 AT6400 Xilinx part loaded 1 AT6400 commands READY 42 tells AT6400 that input buffer has data CLEAR GPINT 44 tells AT6400 to clear general purpose interrupt REQ STATUS 48 tells AT6400 that you want a status update REQ 81 tells AT6400 that you want the kill command executed AT6400 address offsets FASTSTATUS 2 fast status offset STATUS 4 Il set clear status offset 1 AT6400 fast status AXIS1 MOTOR 00 AXIS2 MOTOR 02 AXIS3 MOTOR 04 AXIS4 MOTOR 06 AXIS1 ENCODER 08 AXIS2 ENCODER 0A AXIS3 ENCODER 0C AXIS4 ENCODER 0E AKISI VELOCITY 10 52 VELOCITY 12 AXIS3 VELOCITY 14 54 VELOCITY 16 AXIS1 STATUS 18 AXIS2 STATUS 14 AXIS3 STATUS 1C AXIS4 STATUS 1 INPUT STATUS 20 OUTPUT STATUS 22 LIMIT STATUS 24 INO STATUS 25 ANALOG STATUS 26 INT STATUS 28 SYSTEM STATUS 2A USER STATUS 2 TFRM STATUS 20 TIMER STATUS 2E Maximums MAXBYTES
459. vacceleration power mintottime 2 end 193 else begin if time gravfinaccel then gravtopposition 0 5 gravacceleration power time 2 else if time lt gravfinaccel gravtimeflat then gravtopposition 0 5 gravfinaccel tempgravrps time gravfinaccel tempgravrps else if time lt gravtottime then gravtopposition gravfinaccel tempgravrps gravtimeflat tempgravrps 0 5 gravacceleration power gravtottime time 2 else gravtopposition gravfinaccel tempgravrps gravtimeflat tempgravrps end gravtopposition gravtopposition gravelcircumference end else gravtopposition 0 end end calculate the desired volume fractions and add the traces to the charts the time is calculated using the estimated frame rate from the real time unit since the start of the sampling and the start of the stepper motors occurs at the same time The volume fractions are calculated as the volume of the bubble currently within the measuring region devided by the total volume of the measuring region This is not the same as the component fraction if the bubble is shorter than the electrode length procedure TDynVerify CalcDesiredClick Sender TObject var testposition integer airtopposition gravtopposition time real upperairvol uppergravvol lowerairvol lowergravvol upperplatevolume lowerplatevolume real airstartpos gravelstartpos real begin Il initialise the charts TopAirPlot Series 0 Clear BotAirPlot
460. veltopstart to graveltopstop do begin if totaloutputs 2 then topgravelerror topgravelerror abs desiredvolume testcase 2 outputstore testcase graveltopoffset prediction 2 else topgravelerror topgravelerror abs desiredvolume testcase 2 volumestore testcase graveltopoffset 2 end topgravelerror topgravelerror graveltopstop graveltopstart 1 end else begin TopGravDelay Caption unknown for testcase airtopstart to airtopstop do begin if totaloutputs 2 then topgravelerror topgravelerror abs 0 outputstore testcase airtopoffset prediction 2 else topgravelerror topgravelerror abs 0 volumestore testcase airtopoffset 2 end topgravelerror topgravelerror airtopstop airtopstart 1 end TopGravVolError Caption floattostrf topgravelerror ffFixed 4 3 air volume fraction for bottom measuring ring startmax 0 stopmax 0 maxvalue 0 for testcase 1 to RealMain DesiredFrames Value do begin if totaloutputs 2 then begin if outputstore testcase prediction 3 gt maxvalue then begin maxvalue outputstore testcase prediction 3 startmax testcase stopmax 0 end 197 else if outputstore testcase prediction 3 maxvalue then stopmax testcase end else begin if volumestore testcase 3 gt maxvalue then begin maxvalue volumestore testcase 3 startmax testcase stopmax 0 end else if volumestore testcase 3 maxvalue then stopmax testcase end
461. verflows to 0000 corresponding to the collection of a complete frame of 128 voltages from both impedance tomography systems sampling is disabled by outputting a low on PAO thus keeping the 555 in the reset state 58 Having designed a hardware system for automatically controlling the multiplexer address lines it was still necessary to implement a software unit that will provide high level access to the sampling functions This unit is included in Appendix K on page 149 It was determined that greater efficiency would be achieved by capturing the data in sets of frames instead of individual frames Although the sampling unit allows up to 200 frames to be captured at one time without software intervention tests have shown that readings for the later frames namely those above frame number 100 can be erroneous as a result of a temporary loss of synchronisation As is evident from the description of the sample controller board exact synchronisation is required between the hardware controlling the multiplexers and the PC30G card in order to know that for example sampled voltage number 1 corresponds to the capacitance between transmitter A and receiver A If noise on the external trigger line initiates a sample that is not related to the multiplexer settings then all samples captured from that point are out of synchronisation by the 16 additional samples captured Only by restarting the data capture process will synchronisation once again be achiev
462. vidual component density the mass flow rate can be calculated 13 This density information can be determined in advance for most applications 14 The specific objectives of this research were to 1 verify the performance of an impedance tomography system on a laboratory scale rig for static configurations of a multi phase air gravel seawater mixture 2 combine two of these measurement sections and determine whether any interference or cross coupling exists between the two systems when both are operating at the same time 3 examine the ability of a dual plane impedance tomography system to accurately measure the individual component mass flow rates in simulated flow situations Specifically the accuracy of component volume fraction measurements must be determined as well as the range and accuracy of the component velocity measurements 4 evaluate the ability of the system to differentiate between an air and gravel mass moving through the measurement section simultaneously and the corresponding effect on the component velocity measurement 5 draw conclusions regarding the application of a dual plane impedance tomography system to the on line monitoring of an air gravel seawater mixture 6 make recommendations for future project development and discuss any remaining issues that would need to be addressed before an industrial prototype is developed The required specifications for this multi phase flowmeter are detailed in table 1 1 TAB
463. vidual conductance and capacitance components were then separated at the synchronous detection stage A power op amp was required due to the large conductance component Specifically a LM6313 high speed high power op amp was used to perform this task The peak output current of the LM6313 is 300mA However initial tests revealed that although good results were achieved for the measurement of the conductance component the changes in capacitance were less detectable and were sensitive to conductance variations Hence it was decided that the conductance and capacitance receiver paths should be isolated A larger hole was drilled in the centre of the capacitance plates to ensure that there was no electrical contact between the conductance probes and the capacitance plates PIPE WALL The conductance receiver consists of an input resistor to ground followed by a buffer CAPACITANCE As the conductance of the material within PLAT E the rig changes so the potential across the input resistor changes In particular as the conductance increases so the TRANSIMPEDANCE zz AMPLIFIER amplitude of the output waveform increases The input resistor was chosen to CONDUCTANCE PROBE ensure that the buffer output did not saturate when all the transmitters were BUFFER operating simultaneously Figure 4 5 is a simplified diagram of the capacitance and conductance receiver FIGURE 4 5 DIAGRAM OF THE CAPACITANCE AND C
464. volume fraction profile measured by the neural network reconstruction algorithm and the dashed line is the desired volume fraction profile calculated according to the bubble sizes and flow simulation apparatus configuration TOP SYSTEM VOLUME FRACTION MEASUREMENT 10 T T T T T T T IR 1 9 6 4 o S 9 4L 2 9 2L 4 0 NI iy 0 0 5 1 1 5 2 2 5 3 3 5 4 4 5 5 time s BOTTOM SYSTEM VOLUME FRACTION MEASUREMENT 10 T T T T T T T gt 89r 1 E 9 6r o S EN o 4 4 E a S 2 4 baw pi 1 1 1 ya 1 1 1 0 0 5 1 1 5 2 2 5 3 3 5 4 4 5 5 time s FIGURE 7 8 PLOT OF BOTH DESIRED AND MEASURED VOLUME FRACTION PROFILES AT THE TOP AND BOTTOM MEASUREMENT SYSTEMS FOR A SIMULATED 0 2m s FLOW CONSISTING OF A 0 04 VOLUME FRACTION AIR BUBBLE AND A 0 01 VOLUME FRACTION AIR BUBBLE 93 As expected the bubbles are detected at the bottom measurement plane before they are detected at the top measurement plane In addition it is evident that when the bubbles pass through the bottom measurement plane they are not detected at the top measurement plane thus proving that minimal interference exists between the two measurement planes However the decay of the measured volume fraction at the bottom plane and the increase of the measured volume fraction at the top plane are slower than the increase of the measured volume fraction at the bottom plane or the decay of the measured volum
465. w CMOS analogue switch was used The following explanation of switch based synchronous detection will consider the received signals from receiver A with only transmitter A operating All appropriate waveforms are included in figure 4 6 on page 29 which shows the specific results for the received conductance signal Firstly the received signal is buffered and inverted to produce the original waveform and an inverted version with a 180 phase shift between the two These signals are connected to the inputs of two switches whose outputs are connected together The states of these two switches are opposite and are driven by Ao from the PIC16F84 which is also used to drive the transmitter The operation of these switches is given in the table within the figure For example when is high switch 1 closes and switch 3 opens and the inverted signal is connected to the low pass filter Then when A is low switch 3 closes and switch 1 opens and the buffered signal is connected to the low pass filter If there was no phase shift in the system and the transmitter outputs were perfect square waves then the output of the DG303 would be a pure DC voltage proportional to the conductance between transmitter A and receiver A However due to the phase shifts introduced by the different stages in the system and the limited slew rate of the output transmitters the DG303 output actually consists of a DC voltage with an AC component as can be seen in figure 4 6 It is th
466. wersysbot 0 5 lowersystop lowersysbot 0 005 then gravbottime simtime time when gravel bubble at centre of top measuring ring if gravtopposition 0 5 gravleng uppersysbot 0 5 uppersystop uppersysbot 0 005 and gravtopposition 0 5 gravleng uppersysbot 0 5 uppersystop uppersysbot 0 005 then gravtoptime simtime simtime simtime 0 001 end calculate transit times of two bubbles and then the average velocity airtransit abs airtoptime airbottime gravtransit abs gravtoptime gravbottime with RealMain do begin if airtransit 0 then begin airactualvel 0 AirVelActual Caption 0 end else begin airactualvel flowmeterfactor systemseparation airtransit if airactualvel lt 0 001 then airactualvel 0 AirVelActual Caption floattostrf airactualvel ffFixed 4 3 end if gravtransit 0 then begin gravelactualvel 0 GravelVelActual Caption 0 end else begin gravelactualvel flowmeterfactor systemseparation gravtransit if gravelactualvel lt 0 001 then gravelactualvel 0 GravelVelActual Caption floattostrf gravelactualvel ffFixed 4 3 end end end tM GENERAL PROCEDURES the following functions provide a simple graphical interface for specifying the position of the gravel and air bubble procedure TDynVerify AirShape1 MouseDown Sender TObject Button TMouseButton Shift TShiftState X Y Integer begin
467. wo phases was not achieved the predicted phase velocities were erroneous Secondly the problem of distinguishing between the air and gravel phases is evident Specifically the gravel transition regions of the air bubble at the top measurement plane can be seen Further the gravel bubble is partly detected as an air bubble at the bottom measurement plane Subsequently the accuracy with which the cross correlation algorithm calculates the velocities of the individual components will depend on the accuracy of the reconstruction algorithm to distinguish between the air and gravel phases Although these results are limited they do prove that a cross correlation algorithm is able to calculate the velocities of individual components within a multi phase flow provided a satisfactory separation between the different phases is achieved 7 4 3 Limitations of three phase reconstruction algorithms As mentioned previously the use of three phase reconstruction algorithms reduced the on line frame rate of the System Further the accuracy of the velocities calculated using the cross correlation algorithm decreased and occasionally bubbles would not be detected at certain measurement planes This behaviour can be explained as follows the sampling and reconstruction tasks are performed simultaneously and operate in sets of 50 frames For the simpler two phase reconstruction algorithms the time taken to perform the reconstruction of 50 frames is shorter than the
468. x 1 do for rowindex 0 to 128 do WriteLn tempfile HiddenWeightBest rowindex colindex for colindex 0 to RecLoadDBase GetNumOutputs 1 do for rowindex 0 to upperindex do WriteLn tempfile OutputWeightBest rowindex colindex end else begin WriteLn tempfile RecSetParam GetNumHidden if RecSetParam GetNumHidden gt 0 then begin save the weights for the hidden layer neurons upperindex RecSetParam GetNumHidden for colindex 0 to upperindex 1 do for rowindex 0 to 128 do WriteLn tempfile Hidden Weight rowindex colindex end save the weights for the output layer neurons for colindex 0 to RecLoadDBase GetNumOutputs 1 do for rowindex 0 to upperindex do WriteLn tempfile OutputWeight rowindex colindex end CloseFile tempfile end load the network weights from a text file where the filename depends on the test number procedure TRecMain LoadnetworkweightsClick Sender TObject var tempfile textfile filename string rowindex colindex upperindex tempoutputs temphidden integer begin upperindex 128 filename c TestData weight inttostr testnumber txt AssignFile tempfile filename Reset tempfile read in the number of outputs ReadLn tempfile tempoutputs read in the number of hidden layer neurons ReadLn tempfile temphidden if temphidden gt 0 then begin upperindex temphidden load in the weights for the hidden layer neurons for colindex
469. xist including 39 1 sequentially sweeping the frequency over the desired range 2 white noise generators 3 specific output functions such as the delta function or Sin The frequency components are then extracted either in hardware using phase sensitive demodulation or in the discrete time domain by means of the DTF transform 39 Indeed close similarities do exist between the author s research and research conducted at the University of Sheffield 40 where Electrical Impedance Tomographic Spectroscopy EITS is used in medical applications to investigate the bio impedance of certain tissues Systems developed by the University of Sheffield include a 16 electrode system that uses dedicated current injection and voltage detection electrodes and operates over eight frequencies from 9 6kHz to 1 2MHz at a frame rate of 67frames s Further a three dimensional system has been developed consisting of four rings of 16 electrodes to construct a three dimensional model of the interrogated region 40 However the fundamental measurement principle of the author s system is different to that of EITS in that the different frequencies are used as a means of separating the signals from the different transmitters operating in parallel and not as a means of identifying how the complex impedance of a substance changes with frequency 4 2 Design of an 8 electrode System to Verify the Frequency Division Multiplexed Impedance Tomography Concept Before
470. xt TC1 res1 cap1 2 pi TC2 res2 cap2 2 pi inpui 1 23437 275 input 2 48005 1375 input 3 54346 8875 input 4 62829 1 input 5 66692 8125 input 6 73528 525 input 7 74997 1875 input 8 79172 8125 res MultOut input BestFreqFit Caption floattostrf res ffFixed 5 4 l 7M PBIL procedures Il 1 perform PBIL to determine optimal set of transmitter frequencies for an 8 electrode FDM impedance tomography system procedure TPBILMain PBILStartClick Sender TObject var nvars prec 0 trial i j k integer bestfit fit double bw array 1 8 of integer PV array 1 64 of double array 1 8 of integer bestreal array 1 8 of double trialsol array 1 64 of short bestsol array 1 64 of short besteversol array 1 64 of short inputs array 1 8 of double res double begin resi strtofloat FiltRes1 Text strtofloat FiltCap1 Text res2 strtofloat FiltRes2 Text cap strtofloat FiltCap2 Text TC1 rest cap1 2 pi TC2 res2 cap2 2 pi threshold strtofloat MagThresh Text II initialise the random number generator with a random value Randomize ntrials NumTrials Value maxgen MaxGenerations Value PBILProgress Position 0 PBILProgress Max maxgen 1 nvars 8 number of variables prec 8 8 bit precision since want to index into I an array 256 e
471. y defined then the accuracy of the resulting neural network would be seriously degraded since the neural network would attempt to incorporate this outlier into the model of the system subsequently compromising the performance for true data Therefore great care must be taken during the generation of the training database to ensure that for each training case the bubbles are positioned correctly It is important to note that the term bubble does not refer to a bubble in the normal sense Instead it describes a contiguous mass of either gravel or polystyrene foam Air bubbles were simulated as polystyrene foam cylinders since polystyrene foam has a dielectric constant that is similar to air and is easier to work with Gravel bubbles were simulated as collections of gravel chips Figure 6 4 on page 71 is a photograph of the simulated air and gravel bubbles Further Table 3 1 lists the dielectric constants of the materials used during testing Although the polystyrene entry in the table has a dielectric constant that is closer to gravel than air it was polystyrene foam that was actually used which has a dielectric constant similar to air TABLE 3 1 DIELECTRIC CONSTANTS OF THE MATERIALS USED DURING TESTING DIELECTRIC CONSTANT 25 C AIR POLYSTYRENE POLYSTYRENE FOAM SAND GRAVEL WATER The training database is generated as follows essentially the smallest identifiable element of one phase or bubble is placed in
472. y generator board is used for both systems Using this technique it was possible to remove the spacers and grounded disks and still achieve an output ripple comparable to that when only a single plane was operating as is evident in table 7 1 Subsequently all further dynamic dual plane tests were performed using synchronised transmitters One drawback of this technique is that the receivers for a particular measurement plane are still receiving signals from the transmitters of the other measurement plane Hence if a large air bubble were to move through the bottom measurement plane a change would also be detected at the top measurement plane because of the change to the interference component from the bottom measurement plane Ideally an FDM impedance tomography system would use a different set of frequencies for the two measurement sections ensuring that even if signals from the second plane are received at the first plane the synchronous detectors of the first plane will reject them However because of the nature of square waves it was only possible to find a single set of eight transmitter frequencies that gave adequate performance as discussed earlier Consequently this cross plane interference technique could not be employed but it should certainly be considered for the dual plane sine wave excitation System currently being developed 62 92 7 Two phase Air Water Reconstruction Results To get a better assessment of the performance of
473. y only performing the calculation for the first few evolutions Evolution 1 and 2 are highlighted in the above equation As more evolutions are included in the calculation greater accuracy is achieved at the cost of increased processing time Since cross correlation flowmeters are only required to measure a time delay certain sensor properties are not critical For example the linearity and DC stability of the readings are unimportant 50 In contrast the phase delay of the transducers requires careful consideration Specifically the filters used in both transducers must be matched since a difference in the phase delay of the two systems would result in a constant bias to the velocity measurements These time delay estimation errors have been investigated and it can be shown that the error in time delay measurement is a function of the difference between the two time constants T and of the two planes Assuming the two measurement planes have a first order low pass filter response the error in the time delay estimation is approximately 50 1 Ita x X ri T 7 7 where x is the measured time delay and ta is the actual time delay 79 As mentioned previously the bandwidth of the signals plays an important role in the cross correlation function and specifically in terms of the accuracy with which the peak of the cross correlation function can be measured The bandwidth of a cross correlation system depends on three factors 50 1 th

mass flow measurement of multi-phase mixtures by - MWEB

Contents

Download Pdf Manuals

Related Search

Related Contents