Home

Applied to and Demonstrated on Cupola

1. Validate Sensor P2 conf comparator sdata 8 gt Sub start top 8 0 8 E 0 8 gt gt 2 qstd gt Sub start base gt 8b 0 8 n lt lt au 8 gt end top y gt gt l 6 CAE gt gt hstd 8 70 5 v 0 h tstd 5 8 1 255 8b ac EET 8 545 std MER gt 8 8 65025 195678 0 0 8 pop d 137 trap 1 start top trap 0 start base trap 2 _lop Q a 3 end base height l slope 544 std VALIDATE SENSOR 3 138 8 gt sdata Ldvs0 gt E t 1 trapl 0 P Lg D tlregO gt E t 2 trap2 _ r0 ra D t2reg0 a a gt tart t start toprO mee 8 gt D streg0 g gt gt E end base end _ baser0 D 8 gt D ebreg0 gt gt 0 0 0 E g D tOreg0 4 gt t 3 3 0 _ gt D Brego A gt gt E start base start baserO 7 P sbreg0 g gt gt E height height g D hregO a gt end to end 0 E oer ane A gt E slope slope 0 8 gt D sreg0 A dy E 544 stdr0 sq4_ std sq4reg0 B gt
2. RF R5 OMF Sug Mem M17 M18 M19 21 22 2 3 1 2 3 CompFirSt Bxx ArrayOp vo 1 V2 V3 V4 V5 V6 V8 V9 V10 V11 Eval1 gt 0 gt 1 22 3 gt 4 gt 5 gt gt 7 gt 8 gt 9 gt 10 gt 11 wx Eval3 Eval4 1 2 3 4 5 6 7 8 9 10 11 ZA ZB Eval5 Out 1 3 4 5 Count SelRule 8 0 1 2 3 4 5 6 7 8 9 10 11 12 0 2 3 4 5 SelRuleOut 0 1 2 3 4 5 6 7 8 9 10 11 103 A 6 MSF Hardware Block Diagrams 104 16 16 larea 16 16 larea 6 1 16 larea gt rarea 0 91042 0 rarea larea 2 8 0 M4 rarea larea 8 noofsensors 8 cen pos A 20 lt lt 2 4 LdH 105 PE 106 107 AREA Lee 0 larea 2 aga 70 8 1 E Ee aee _1 8 2 comp ct sb amp amp ct lt st ee TMP Area Lye i 8 Logic 8 3 Leer amp g a8 a Cc 5 larea 3 ct gt st amp amp ct lt et comp 107 y larea 4 ay MULTI Z gt 75 multiplier Unit
3. VALIDATE SENSOR 4 139 trap2 rl gt end baserl A 3 rl gt height rl slope rl Ldvsl gt trap 1 trapl_rl tregi A gt E trap 2 gt D t2regl tart start top 5 e E s streg Jg gt E end _ base 10 ebregl 0 rl trap r 0 T S Jd tOregl gt gt E 3 WP t3regl start base start baserl sbregl E height g 12 hregl gt end toprl end top ea 5 gt D erregl Es E slope D sregl 8 gt E 544 std 44 He gt P clk 544 stdrl Note same diagram can be repeated with appropriate modifications for the rest of the sensors 140 REFERENCES 1 Nagrath LJ and Gopal M Control Systems Engineering Second Edition New Age International P Ltd Publishers 1995 2 Maciejowski J M Multivariable Feedback Design Addison Wesley Publishers Ltd 1990 3 Richard R Brooks and S S Iyengar Multi Sensor Fusion Fundamentals and Applications with Software Prentice Hall Inc New Jersey 1998 4 Ren C Luo and Michael G Kay Multiple Integration and Fusion in Intelligent Systems JEEE Transactions on Systems Man and Cybernetics vol 19
4. 132 5 3 Results and Analysis of Demo Runs ee eeeeee esee eren ee eren 134 B osa pesce S abr 148 APPENDIX SA 4 iiv dud RAM M MANI 149 CHAPTER u D 155 6 1 SUMMARY AND CONCLUSIONS eene 155 REFERENCES M MAMMA MM 161 APPENDIX A USER MANUAL eene 166 List of Figures Figure 1 1 Graphical Representation of Project Organization 21 Figure 1 2 Detailed representation of Project Organization 21 Figure 1 3 Overall System Vision for I3PSC applied to Cupola Furnaces 23 Figure 2 1 Schematic Diagram of a Feedback Control 33 Figure 2 2 Feedback Control System with Multiple Sensor Fusion 35 Figure 2 3 Block Diagram of Proposed System sse 37 Figure 2 4 Membership Funetionsaaiscecod ara Pep o tei ae a enis tud 44 Figure 2 5 Block Diagram of the Self Validation Technique esses 44 Figure 3 1 Individual Gaussian Functions and the Cumulative PDF 54 Figure 3 2 Estimation of the Measurand 56 Figure 3 3 Comparison of Estimate with the Centroid
5. nomi 4 charts 0 File Di Input Variables Input Variables 2 Input Variables 3 Input Variables 4 RESO coke in charge v blast temperature coke size 7 normalized mass of air blast 2400 0 2800 0 2400 0 Excel Files E Read New File 2500 2500 0 2350 0 Experiment xls 2300 0 NNData xls Close Window 2250 0 2400 0 2300 0 nominal xls 3 Real fs xls 2200 0 2200 0 22500 RealSensor xls 2150 0 SimRef xls 2100 0 2000 0 2200 0 Testmv xls Output Variables 20500 7 n 1 T 2150 0 7 0 1 Testnmy xls metal temperature 1 00 2500 5000 7500 5000 10000 1500 0 00 i S 20 Output Variables 2 melt rate Output Variables 3 combustion efficiency ratio w 10 04 554 5 0 4 5 40 80 50 2500 5000 7500 5000 10000 15004 5000 10000 15004 Output Variables 4 Offgas 2500 5000 7500 5000 10000 1500 Figure A 61 View N M Correlation Graphs View Nominal Graphs vi 215 216 A 3 Online Analysis Online Menu vi Online Setup Simulate Data Collection Analyze Collected Data View Results Exit Figure A 62 Online Analysis Menu Online Menu vi The online menu is used to interrogate the models in a real time situation The 4 D array data structure is setup and populated by the online analysis func
6. OL UR NN RN NEN EI DLE ME 12 1 4 Analysis and Validation of 5 eee 13 1 4 1 Self Validation Fuzzy Logic 13 1 4 2 Self Validation Preprocessing aci eae etia ire 15 1 4 3 Self Validation Execution 1 1 1 7 4 2 4 444 1 40 16 1 4 3 1 Determination of Timing with Hardware 16 1 4 3 2 Determination of Theoretical 17 1 5 Self Validation Software Implementation ccssccssssccsssessseseees 18 1 5 1 Self Validation Fuzzy Logic Code 18 1 5 2 Self Validation Preprocessing Code sese 20 1 6 Communication Software Development the CPU to Host Interface 21 17 SMITA ria ec 22 2 YEAR 2 ACCOMPLISHMENTS 23 2 1 decla 23 2 2 Communication Protocols 4 eee eee 1 ee eene 24 2 2 1 Develop low level communication protocol sss 24 2 2 2 Write and test the low level communication code for initialization ond Eee lye 2 odes ose tb ER a Ue 24 2 2 3 Define specifications for high level communication protocol details for MS SV program for User ntetfabe iie desea trt tasa t ea reete t rented 25 2 2 4 Outline method for
7. virtual Sensor Kalman MR r Plant Manual Melt Rate irav combustion efficiency ratio Fusion melt rate value Modality Property raw value Plant x Figure 47 Setup Variable to Monitor Select the Y nodes in the Y data list box Then click on ADD button to add it into the selected variables list If you want to delete the selected variables please select the variable in the selected variables list box then click on REMOVE button After you select variables need to be monitored for this window click on OK button to return back the Cupola Operation Monitor Use the same procedure to select the variables you want to monitor in other windows There are three display modes namely whole data shifting window and fixed segment Whole data is the default mode that will display the complete data of the running Shifting window only displays the latest data with the data length defined in Window L1 Fixed segment display the fixed length data statically The length of the fixed segment is defined by the start point and window length The Start Point controller only displayed while the fixed segment display mode is selected 201 202 Cupola Operation Monitor Main Screen new win 6 18 02 vi File Edit Operate Tools Browse Window Help E3 m 130 Application Font 69
8. 00 A00 i MFval01 A01 j MFval02 A02 i IMFval10 A10 IMFval11 A11 P4 Output Registers fisComputelnput fisTrapezoid LdIMFvalOO gt E Input D Value00 Reg LdIMFval01 E Input D Value01 Reg LdIMFval02 E Input D Value02 Reg LdIMFval10 E Input D Value10 Reg LdIMFval11 mE Input D Value11 Reg LdIMFval12 gt E Input D Value12 gt Reg MFval12 A12 comparator trap1 trap2 LdIMFval20 gt Input Value20 Reg MFval20 A20 Input Value21 Reg Fval21 A21 LdIMFval22 E Input D Value22 Reg 22 22 89 PA MFParameters fisComputelnputMf fisTrapezoid MDin 15 0 LdIMfPOa E 15 8 Para 0a D 8 5 Reg E 770 Para 1a Reg He N LdIMfP2a E 15 8 D Para 2a 8 Reg E 7 0 D Para 3a 8 5 Reg ALdIMfPOb E 15 8 D Para Ob 8 7 0 Para 1b 8 Reg He LdIMfP2b gt E 15 8 Para 2b 8 Reg E 7 0 3 8 Reg gt SellMfpb reg 8 W 2 8 Da 1 b paraOb reg 8 Para2 c
9. 17 Figure 3 5 Measurand Estimate with High Confidence 29 0 4 lt Integration Limits gt 0 3 1 0 2 PDF 1 0 1 S Seen s 1 T 0 wa 6 8 10 12 14 16 18 20 Figure 3 6 Measurand Estimate with Low Confidence 3 2 Considering Self Confidence in Redundant Sensor Fusion The self confidence of a sensor data obtained from the self validation technique explained in section 2 2 is a measure of how much this data agrees with the expected characteristics of the sensor as estimated from historical data Thus a change in the sensor noise level or if the sensor data or its rate of change exceeds the expected limits the self confidence value decreases Integration of this self confidence into the redundant sensor fusion is necessary to decrease the effect of the failed sensors on the estimated value In Section 3 1 the Parzen like procedure for estimating the PDF which was then used to get a best estimate for the measurand value was presented The function used in Parzen estimation was a Gaussian function with a standard deviation that depends upon the sensor noise This Gaussian function can be thought of as a representation of the 60 probability in finding the true value of the measurand data around the sensor reading As the self confidence decreases the probability of finding the true value of the measurand in the neighborhood of the sensor measurement
10. arable Monitor F1 Variable Monitor2 F2 Trend Monitor F4 Variable 1 graph trend Real Albany sensor imelt rate raw value 00 Trend Monitors 1 y 1 0000 Variables in Monitor Description 1 PositiveMediumFast PositiveMediumFast Variable Name 2 graph trend 2 Real Albany sensor metal temperature 1 0 00 Confidence 2 Press Shift to make multiple selections Max 4 0 7000 Description 2 egativeMediumFast Alarm setup 1 IConstantMedium elect parameters NegativeMediumFast Important Case num of par lt 4 Alarm setup 2 Variable Name 3 graph trend 3 Real Albany sensor sensor2 raw value 0 00 Confidence 3 num of par lt 4 A 1 0000 Description 3 PositiveMediumFast PositiveMediumFast PositiveMediumSlow Figure A 48 Trend Monitor On the trend monitor tab Figure A 48 you can select the variable whose trend will be monitored At most four trend monitor windows are arranged on the right side of the window Also two situation diagnostics can be set up in the same tab Click on the blue Select Parameters button The dialog to setup the diagnostics in Figure A 49 is shown The results of the diagnostics will be displayed by the LED on the Cupola Monitor window The bright LED indicates the alert situation is happening On this wi
11. Soo scimatGoF BR acus BR acus BR total oxygen percent oxyo 227p 227p essensdiWozRe me imecorespd 720 238 79038 AM cet eat blast rate et E Ee i ofgas amount of caf i2746 moex 1 combustone amountorcd s28p amp 528 space Kalman amountofio i9 S4t omehr 21495 omw l Siam MR e Kaman metal temper 201990716 31773280 Datacast Tenfmetal temper 2020 907 K 3179 6326 F Bain Temp Bath Temp Spout Tempe metal temper 201950716 _3177 8326 6 SPoutremp SpoutTemp Spoutremp Spout Temp Spout Temp Spout Temp Final Carbon amount of caf 3eop 56 c e o ec Final Carbon Melt Rate 2 amountofir 19541 _21 4951 tonfhr Meit Rate 2 Melt Rate 2_ Melt Rate 2 Rate 2 MeltRate2 Melt Rate metal temper metal temper 201990716 317782006 Temp Temp Temp tap temp metal temper Final Carbon amount of caf seo sep gt _ 1 FisiCarbon radar tevel evel feor
12. module The dotted lines show the buses going outside the module start _topr0 start toprl start topr2 stmux Start _topr start _baser0 start baserl start baser2 8 sbmux start baser3 8 n vus end topr end toprl z end topr2 etmux end topr3 8 8 end _ baser0 end baserl end baser2 8 ebmux end _ baser3 8 8 height rO height rl 8 height r2 8 htmux height r3 8 544 std 0 544 std rl rx 123 INPUT MUX sel sen no Ldinp gt E j t st sig st_reg 5 8 E sb sig 5 8 Dsb reg 8 LUE pet reg 4 E eb sig b 8 Deb reg E ht sig h 8 gt D ht reg ib E m_ sig j m_reg 8 H gt 4 std 16 Dsq4 reg 4 20 544 std r2 544 std r3 Multipliers for MSF sel mul8 8 sstd mull mul in mull fus std min iui sstd mul2 slope mul2 fus Ld8m8 8 8 21 7877 124 mul8 8 716 22 Inputs from RON val sen Area Trap Height Fused Conf Span 8 8 102 16 Ldl6m8 16 8 mull6 8out 24 125 126
13. nin Q3 2 4 n 4 15 2 PBK k 101 A min 0 where is the radius of a Ball c D Then the state of the system is given by 4 lt lt AQ xc 4 17 where 3 15 a class KL function In other words the steady state of the system lies within the ball B shown in Figure 4 4 for all time gt t Proof The system equation can be separated into the feedback stabilized part and the uncertain term as x A BK x BKg x 4 18 Figure 4 4 Region of Stability 102 derivative of the Lyapunov function is V P A BK x x Px x PBKg x Since the last two terms are scalar they can be combined and using the proof of the Theorem 4 1 the derivative of the Lyapunov function is V x Q x 2x PBKg x 4 19 Using Holder s Inequality gives V lt x Q x 2 x PBk lec The uncertainty g x is given by the expression 4 14 Using this relation in the expression gives lt Ap JPBK 214 lt 2 2 2 20 lt Quas 2 8 2 5 The term A Q 2 is always positive as is bounded as given 4 16 This gives an expression for defined in 4 13 2 PBK zy 2 PBK 103 Substituting the bound on and 6 satisfies the relation BERT The co
14. reg 8 8 para2b reg 8 Para3 d para3b reg 1 b Input val reg x 0 a PBD 7 0 90 16 LDIV Input D Value 2 1 fisComputelnputMf fisTrapezoid 8 Sel0 8 Comparat para 1 gt pu cM RN Sel Comparator Z parag Sel2 Q Comparator Sel3 Para0 NOT EQUAL Para1 255 0 0 1 8 Input val reg 0 J rap1 0 gt 3 Multiply Subtractor X 256 Subtractor Shift by 8 16b 8 16 iv out Trap1 8 Para1 0 Divider Subtractor Pet peo TRAP2 fisComputelnputMf fisTrapezoid Para2 c d 8 ara3 d Input val reg Comparator Comparator Comparator Subtractor para 2 NOT EQUAL para 3 para3 Input val reg gt Para Input val reg para3 Input val reg Sel1 Sel3 Sel2 X 256 Shift by 8 Subtractor rap 2 0 gt gt gt 3 8 8 2 Subtractor 16 b 16 8 91 92 firingStrength V 0 11 Comparator Eval1 FisEvaluate1 LdFSN p E NOT EQUAL FSNotO et Reg lt L CLK 127 CIrFSN
15. 0 6 f 04 0 2 0 10 20 30 40 50 60 70 80 90 100 Figure 3 13 Confidence of the Estimate Value Using PDF 790 TC4 J TC5 780 Estimate 770 760 750 Li 1 1 1 1 1 1 50 52 54 56 58 60 62 64 Figure 3 14 A Closeup that Shows Effect of Not Considering Self Confidence 3 3 2 Results of the Sensor Fusion Methodology Considering Self Confidence The test is repeated using the same data presented in Section 3 3 1 However this time the methodology presented in Section 3 2 is used self confidences of the sensors over the considered time period is shown in Figure 3 12 The estimates of the 66 measurand value are shown in Figure 3 15 close up of Figure 3 15 is shown in Figure 3 16 A Comparison between Figure 3 14 and Figure 3 16 shows the advantages of including the self confidence in the sensor fusion methodology When the high noise level was injected the self confidence of the sensor affected by the noise is reduced See Figure 3 12 and in turn its effect over the PDF function is reduced This leads the estimate of measurand value to depend more on the other two sensors with higher self confidence parameters Moreover the overall confidence of the estimate increases See Figure 3 17 This is because the energy around the third sensor is decreased by the inclusion of the self confidence 8
16. 2 0 0 Go Busy 0002 PBRF PBW Busy RFA 4 SR Write Busy to Status Stat To PBD Reg gt lt WaitEx0 Y 1000 PBRF PBR Busy LdCmd RFA 0 CR NotBusy 2000 PBRF PBW Busy RFA 4 SR Stat To PBD Read Cmd Reg Write NOT Busy to Status Reg 5 SV FPGA Controller Comp Input MF Values 1 ImFinitO 0004 PMR PMCe LdlmFp0a IncMAC PBRF PBR LdlV RFA 1 In0 ImFinit1 0008 PMR LdlmFp2a IncMAC 97 2 Read Mem 0 IMF paramo 1 Read RF1 Input0 Read Mem 1 IMF param2 3 ImF 1 0010 PMR PMCe IncMAC Lost cycle Read Mem Even Which param Reg Set to Load A or B Set A Load Sugeno Count 0 1 1 0 Set LdSugO LdimfPOb LdimfP0a gt ImF2 0020 PMR IncMAC Count 0 1 Set A Read Mem odd 0 Set B LdimfP2b LdimfP2a ASM Chart SV FPGA Controller Comp Input MF Values 2 S IMF3 0040 IncCnt P 3 Inc Cycle for next Which Param set to Read next SetPB 98 SetPA Count 8 LdinfVal22 Circ LdimfVal00 LdimfvVal01 Ldimfval02 PBRF PBR RFA 2 in1 LdimfVal10 LdimfVal11 LdimfVal12 PBRF PBR RFA 3 in 2 LdimfVal20 LdimfVal21 Load Correct O
17. see 57 Figure 3 4 Comparison of Estimate with Peak 5 57 Figure 3 5 Measurand Estimate with High Confidence 58 Figure 3 6 Measurand Estimate with Low Confidence eee 59 Figure 3 7 Estimation of Measurand without Considering Self Confidence 60 Figure 3 8 Estimation of Measurand Considering Self Confidence 61 Figure 3 9 Block Diagram of Multiple Sensor Fusion eee 61 Figure 3 10 Estimated Value from PDF without Considering Self Confidence 63 Figure 3 11 Estimated Measurand Value Using Average Method 64 Figure 3 12 Self Confidence of the Three Sensors sse 64 Figure 3 13 Confidence of the Estimate Value Using PDF sse 65 Figure 3 14 A Closeup that Shows Effect of Not Considering Self Confidence 65 Figure 3 15 Estimated Value using PDF Considering Self Confidence 66 Figure 3 16 Close Up of Figure detta SU pe e Dale eo MM e 67 Figure 3 17 Confidence of the Estimate from PDF including the Self Confidence 67 Figure 3 18 Multiple Sensor Fusion without trend information sss 70 10 Figure 319 Confidence Plot 71 Figure 3 20 Sources of Trend Information 71 Figure 3 21 General Method
18. 1 Modify Modality OK Back To Main Menu Path of that executes Modality 9 C MyDocuments vipin CupolaProj LabVIEW_new Figure A 20 Delete Modify Modality The Groups of Modality list box on the left of the screen lists all the groups of the modality created earlier Selecting one of them displays the details of that modality group in the Modality Details indicator The user is provided with an option of either deleting the group as such Delete Modality or modifying the contents of the group Modify Modality If click Delete Modality an alert confirmation dialog is popped 185 186 up to request the user to confirm whether he wants to delete the group The alert dialog is as shown in Figure A 21 IE Delete Confirm vi Figure A 21 Delete Confirmation Alert Modify Modality option allows the user to change the details of the modality groups The dialog of Update Group is as shown in Figure A 22 gt Update Group vi Update Group Change Variable List Change Parameters Done Updating Figure A 22 Update Modality Group Details The Change Variable List option allows the user to change the variable list of the group The interface is shown in Figure A 23 Double clicking on any of the variables in the list opens an interface as shown in Figure A 24 which allows the user to change the variable iables of Modality coke ratio Controller raw
19. Dex zi 8 PA gt et amp amp ct lt eb larea 5 rea P2 108 6 EL 8 Logic larea 6 8 109 larea larea larea WW N NO A n 6 4 2 110 Area P3 ldarea_ f gt E larea sig larea BP 14 16 clear clr area gt E Ift point clear clr area clk 111 112 4 TMP AREA Logic MULTI Via tmp out multiplier ae mm 24 102 16 l Unit 9 16b 8b 24 MULTI muliplier Unit 2 al al 1 5 a2 ht 2 P 0 0 gt 4 A a2 8 2 0 8 8 6 2 d 8b 8b 6 ht 8 2 0 A 8 Note the multiplier amp divider units are implemented outside the Area module 16 8 16b 8b 16b gt cen add _ sig Cen D add 113 Cen reg Center pug CENTROID Ldcen add Ldcen E d cen ad gt 16 reg 16b 8b cen 518 EA noofsensors Clr cen 114 115 Dividers for MSF sel
20. sss 135 Figure 5 3 Metal Stream Changes suggested by I3PSC for control of C for Run 1 135 Figure 5 4 Individual Measurements and Fused Melt Rate for Run 1 136 Figure 5 5 Confidence of Fused 136 Figure 5 6 Individual Measurements and Fused Temperature 137 Figure 5 7 Confidence of Fused Temperature seen 137 Figure 5 8 Oxygen Enrichment for Temperature Control for Run 1 138 Figure 5 9 Blast Rate for Melt Rate Control for Run 1 138 Figure 5 10 Control of C during Run 2 perpe dte 139 Figure 5 11 Changes in MR during Run 2 cscccssesscossssssenssssceorseseduacesatsoncesacencessessers 140 Figure 5 12 Changes in MR during cedes 140 Figure 5 13 Metal Stream Changes control of for Run 2 142 Figure 5 14 Forward Change in CMR to Achieve Large Change MR Run 3 143 Figure 5 15 Changes in Metal Stream to compensate for Change in CMR 143 Figure 5 16 Changes in Oxygen Enrichment SCFM during Run 3 144 Figure 5 17 Changes in Blast Rate SCFM during Run 23 eese 144 Figure 5 18 Control of Melt Rate during Run 3 145 Figure 5 19 Changes in Iron Temperature deg F during Run 3 145 12 Figure 5 20 Changes in Carb
21. gt Set Standard Grammar vi E x File Path Read File Let Find 16 Standard Grammar 2 C ISpsc FILES Master List cj 6 23 txt p Tee Figure A 7 Set Standard Grammar File If the standard grammar file is stored in the default path shown in the file path controller click Read File button will load the file into your setup Otherwise you need click Let Me Find It button to search and load the file 175 Plant i podeis eR coke in charglweight aer coke ratio oe foero coke rati cupola diame diameter oft zm Od gt cupola well di diameter oft volume 0 338 So0 scimatSOF blast rate biast rate BlastRateR Blower Freq volume orar O 1339 w amp s toc soo scmateoF o BlowerFrea pressure drogpressure dro 0 1338 106 So0 sctmateoF esre oxygen additi volume ofan 0003125 7 scimat amp oF O2 Emich O2 Enrich o2FlowRatdvolumeofax 0 003125 7 scmateoF OzFbwRad blast temperdiemperature 699826 ko last fraction fraction of biel actua BR actual volumd 0 1339
22. o A Bet roptcavon Fone melt rate 3i 1 000000 melt temp Startup Time Duration min 3 1600 0000C Seo Carbon 000000 Figure A 38 Add startup routine The steady burn period is setup in the dialog showm in Figure A 39 Figure A 40 is the dialog to setup the transition In Figure A 41 user can setup the shut down time Add Steady Burn Routine vi lof x File Edit Operate Tools Browse Window Help on Define Start and Stop Times for Steady State Burn Start Time March 0 1 em a Figure A 39 Set up steady burn routine 195 196 Add Transition vi Start Time Start Value 1 End Value 1 melt rate 0 000000 Start Value 2 End Value 2 End Time Pyrometer Temperature 0 000000 Jo 000000 Start Value 3 End 2nd Pyrometer 10 000000 Figure A 40 Add transition It Add Shut Down Routine vi x Start Time Start Value 1 End Value 1 melt rate 0 000000 0 000000 Start Value 2 End Value 2 End Time Pyrometer Temperature 10 000000 Start Value End Value 3 2nd Pyrometer 0 000000 0 000000 Figure A 41 Add shutdown time 4 1 1 6 2 3 4 Controller setup Controller setup is used to setup the parameters of the controller variables which are listed in the menu box in Figure A 42 Highlight the contro
23. parallel processing are an integral part of this operating system The relatively random time periods consumed by these processes could not be separated from the actual code timing The final timing results from multiple runs had a variation of up to 70 of the average value A time consuming solution would have to be developed to eliminate these problems We decided that the complexity of the solution was not worthwhile So the Hardware Team decided to drop this hardware timing method and try the following theoretical timing method instead 1 4 3 2 Determination of Theoretical Timing Theoretical timing which measures the number of clock cycles required by the system to process the given code is determined by looking up every instruction s execution time in the microprocessor data book The main goal of performing this theoretical timing analysis is to compute the average time required in clock cycles for processing the SV code In order to do this the code was split into three types of blocks that perform computations This would help in calculating the average number of clock cycles required for processing the code for a given set of inputs So the code was divided into straight line blocks conditional blocks or procedure blocks A straight line block consists of statements through which the control flows without any branching to other blocks For instance assignment statements equations and initializations form part of strai
24. REARRANGE tarr2 0 y Output a gt Regd tarr2 1 Output tra SC gt Regl lp EVEN SORT tarr2 11 Output tra asc Reel A 2 trap 2 __ seltrap trap_Ord a ee 0 arr mux arr gt Reo0 J 3 trap Orl D ir mes tari3 1 bio Q5 A i jd 8 ODD f SORT trap r3 Z 8 11 gt 8 7 Regll 8 ct Ag 12 14 Regl2 8 tarr F Q e tarr reg l tarr reg H tarr reg 2 Ldspan 3 MET mul sinc 8 8 L clk NOTE 8 8 multiplier is used to calculate 3 stdmin for span module The multiplier is implemented outside the Rearrange module Span reg 16 Span E tarr _ tarr tarr tarr _ tarr _ tarr _ tarr _ tarr _ tarr _ arr _ _ tarr _ tarr tarr _ Even Sort U E U E U E U E U E U E SORT Odd Sort S _0 tar3 0 LR tarr2 1 fe L tarr2 2 gt gt gt 3 1 2 tarr2 3 A U E yas D 8 8 8 tarr2 4 gt gt fp tar3 3 P tarr2 5 gt UE nm 8 8 8 tarr2 6 gt gt fr tar3 5 P ta
25. Ty TKR 10 st Q FT 07 TT TTT TN NNN NNN Figure 5 16 Changes in Oxygen Enrichment SCFM during Run 3 Blast Rate 360 340 320 300 280 260 240 220 200 e e o e N o m 144 157 N wo om e 209 222 235 248 274 287 300 313 Q Figure 5 17 Changes in Blast Rate SCFM during Run 3 145 4000 MR 2 Value MR rv 3500 Fused MR v Manual MR rv 3000 2500 4 y fj 2000 4 4 1500 1000 14 E 40 O QN R 0 0 c q dq Figure 5 18 Control of Melt Rate during Run 3 olten ron 2800 1 V I Desire 5 ca at ln ea ah eee 2700 N A Vi IU Ld AW ll an T i desee EE 2 CIC A 2600 2500 Spout rv Datacast rv pyro2 rv enon Pyro rv Fused T v 2300 ea G 10 Or DF Dr oO F amp F O c0 0 Pr ec NOT st A t DTN Tr TT TTF QN NNN Figu
26. 0 051 0 149 0 000 0 129 0 030 0 119 0 055 3 460 4 498 14 187 35 640 51 211 60 554 67 474 76 817 83 045 90 311 79 239 76 125 112 111 143 253 129 412 101 384 49 135 37 024 27 336 8 997 3 114 2 422 5 882 5 190 6 228 6 920 10 035 15 225 14 187 8 651 9 689 20 415 19 377 17 993 16 609 20 069 24 567 23 183 40 138 45 329 61 592 51 211 43 253 52 941 31 142 47 405 73 709 412 699 412 695 882 688 235 681 765 680 588 680 588 680 588 682 941 684 118 692 941 694 706 698 235 699 412 709 412 717 059 720 588 727 647 737 647 739 412 748 823 759 412 763 529 2 3 Self confidence output file sv4_outx txt 0 501960814 0 501960814 0 501960814 0 501960814 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 0 009 0 159 0 060 0 119 0 110 0 011 0 000 0 000 0 039 0 018 0 138 0 030 0 055 0 018 0 161 0 131 0 055 0 122 0 159 0 030 0 149 0 184 0 064 59 862 69 896 90 657 65 052 52 941 14 187 3 460 5 536 33 218 40 484 41 522 29 758 33 218 90 657 81 661 98 616 95 502 79 931 103 806 125 952 108 304 156 055 136 332 74 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 0 501960814 1 000000000 1 000000000 1
27. 1 when confidence is 1 a 0 when confidence is 0 4 3 and 0 lt lt 1 for all other confidence between 0 and 1 4 3 Stability Analysis The controller designed as discussed above will be helpful only if the closed loop system with the controller is stable This application is developed for a cupola furnace plant in Albany research center which has a linear model Hence a linear model is considered in the stability analysis The stability conditions for the closed loop system are not trivial since the system is time variant as the controller parameters change with the time A theorem is stated and proved in this section The application of the theorem discusses the conditions on the stability of the closed loop system 97 Theorem 4 1 Consider a linear time varying system x t Ax t Bu t u t 2 K t y t 4 4 y t x t where xeR yeR uceR are the state output and input variables respectively A B and K are the matrices of appropriate dimensions The system will be asymptotically stable if 1 The controller parameter K is given by expression 4 1 2 There exists a Lyapunov function of the form given in 4 5 where P is a positive definite matrix that proves the stability of the system for both matrices K and V x Px 4 5 Proof The system equation after combining all the expressions is x A BK t x It is given that the closed loop system with the time invariant controller para
28. 5 September 1989 5 R C Luo M Lin and R S Scherp Dynamic multi sensor data fusion system for intelligent robots JEEE Journal Robotics and Automation vol RA 4 no 4 pp 385 396 1988 6 Keith E Holbert A Sharif Heger and Nahrul K Alang Rashid Redundant Sensor Validation by Using Fuzzy Logic Nuclear Science and Engineering vol 118 pp 54 64 1994 141 7 Asok Ray and Rogelio Luck An Introduction to sensor Signal Validation Redundant Measurement Systems JEEE Control Systems Magazine vol 11 no 2 pp 43 Feb 01 1991 8 Marcello R Napolitano Charles Neppach Van Casdorph Steve Naylor Mario Innocenti and Giovanni Silvestri Neural Network Based Scheme for Sensor Failure Detection Identification and Accomodation Journal of Guidance Control and Dynamics vol 18 no 6 Dec 1995 9 Mohamed Abdelrahman and Senthil Subramaniam An Intelligent Signal Validation System for Cupola Furnace Part 1 and Part 2 American Control Conference San Diego 1999 10 Janice C Yang and David Clarke A Self Validating Thermocouple IEEE Transactions on Control Systems Technology vol 5 no 2 March 1997 11 and D W Clarke The Self Validating sensor Rationale definitions and examples Control Eng Practice vol 1 no 4 pp 585 610 1993 12 T M Tsai and H P Chou Sensor fault detection with the single sensor parity relation Nuclear Science and Eng
29. EST gt bese 7 Comp p ED 9 EE fus 9 div 25 _ gt 119 120 Fused_Conf 3 conf t en 8 10 conf en_t 153 5 pis _ 10 8 conf st t en t Gimp en conf2st t amp lt t 11 8 ui f en co st t s _ 8 102 8 8 A 8 fus confll ie conf M st t 8 en conf 51 t 12 102 3 fus conf12 8 en conf Step st t M i en conf2st amp amp en conf st t 13 en conf st b _ 5 8 2 div in ap no conf13 8 121 Fused_Conf 4 0 1 fus conf par 3 d fus confO Ldconf fus confl 8 8 7 eonf acc fus conf 8 D gt fusconf _ fin he 16 16 8 M 7g 5 noofsensors fus confl3 8 clk clr fuscon Note 16 8 divider is implemented outside fus_conf module 122 Fused_Conf 5 0 lg c5 Ars 1 8 8 16 8 n 24 16 102 A tstd in 73 16 2 c3 5 mul amp 8 8 ht EF c5 NOTE All multipliers and dividers are implemented outside the
30. In this dialog 1 Desired Sample Interval is the sample interval of the system The default value is 60 seconds 2 Start will active the run of all modalities 3 If you want to monitor the system on another computer you can type that machine s IP address in the GUI Machine Name 4 Press Write Data to File will record all the running data into an Excel file and this file will be opened after you Stop the running 5 In the Modality File Name box the user can check all the setup modalities that are executing 6 While the system is running the user can setup the modality using the procedure mentioned in section A 1 1 6 and then press Update button to update the modality setup For example using Update the user can add a new variable into a fusion group even the system is still running 200 7 The user can select the variables to be written the Excel using the whose path is C ipsc Datastructure classes Data Structure Export to Excel New vi The variables in the Selected Parameters array will be written Data Structure Export to Excel New vi File Edit Operate Tools Browse Window 24pt Dialog Font reference duplicate reference Data Structure Data Structure Selected Parameters value value confidence value confidence Figure 45 Excel File Setup While running PPSC s the Cupola Operation Monitor window will be popped up
31. Modalities Modalities Modalities Figure A 2 Data Strucuture Model 170 171 Start d Shaded part is the predefined procedure Run Cupola Interface vi 4 4 A Set up a new What to do Quit this run Cupola applicaiton Run the application which has already been setup Set Up Application Y Quit Run What to do Sam Update the Stop without Write data to a Is run modality setup saving data file then stop while running Stop without Write Stop with Write Update Data to File Data to File Y1 Data Setup the variables to display on the graph Select Parameters for diagnostics Figure A 3 Top level Procedure 171 y Define Stardard Grammar y Create Modify Standard Grammar Grammar Select Modalities amp Variable amp Interface Need new modalities amp y Select Variable Properties eed to setup new ariable properties y Save Setup Information Save new setup Modality Setup Modify the modality s setup Modality Setup is a predefined procedure to setup a modality The flow cha
32. cree 4 4 eee eene eren netta see 33 2 10 Develop the MSF fixed point code 34 211 Hardware implementation of the MSF code ees 35 212 35 3 YEAR ACCOMPLISHMENTS 22 1 36 3 1 OPV CE VIG W E 36 3 2 SV Implementation Pr 36 3 3 SV 37 34 37 3 5 MSF Block 44 3 6 Virtex FPGA Board APS 240 45 MEE GU riu ec E 46 3 8 Comnmunicallona e 47 3 9 29050 47 REMEMBER 47 3 10 1 Work COMBI EEA send co 47 3 10 2 Future 49 REFERENCES 50 APPENDICES 22 22 503 Year 1 Accomplishments 11 Overview During the year 1999 the PPSC Hardware Team completed much of the background work needed to begin the implementation of the Signal Processing System Hardware portion of the project We first examined the literature available that was pertinent to our work which included lea
33. sensor2 sensor3 oxygen addition rate trend confidence value confidence melt rate melt rate SP metal temperature 1 metal temperature 1 SP Final Carbon Sio Relative Instant Add Variable Add Group Variables in Group coke ratio Controller raw value at oxygen addition rate Controller trend at 0 sensor Monitor confidence at 0 P S Double Click on any of the elements in the list above to modify it Back to Main Menu Figure A 17 Select Variables in Group NOTE Whenever you add a node with value property a node which has trend property is added automatically If you only need value property node you can delete the trend property node The procedure of deleting a node is double clicking Delete Update Modalities in dialog shown in Figure A 16 Then click the button on each popped up dialogs in the following order Modify Modality as shown in Figure A 22 Change Variable List as shown in Figure A 23 Double clicking on the variable you can Delete it on the dialog in Figure A 24 Clicking on the Add Group button in Figure A 17 dialog will open another dialog as shown in Figure A 18 This interface allows the user to split all the variables selected in the modality into input and output variables IE Split In Op_vi Input Nodes gt gt lt lt Output Nodes HE OER lt lt Figure A 18 Split Variables
34. 0 0591 0 0212 64 0739 40 4590 40 4093 60 9554 57 7605 61 0663 40 7508 27 7856 25 8352 7 9057 0 3 7637 0 4 8436 0 0301 0 0125 0 0806 0 1181 0 0883 0 0602 0 1500 0 1033 0 1114 0 0951 0 0984 0 1032 0 2165 0 1134 0 0655 0 0678 0 0943 0 0155 0 0055 0 0304 0 0336 0 0078 14 6033 35 5197 49 8494 59 2496 67 8309 78 6293 84 8290 90 1782 76 5131 73 3023 111 8839 145 3136 133 4396 103 4643 49 0632 37 4642 27 3302 9 5115 3 6494 2 9445 6 4174 5 6818 0 6 7649 0 0391 0 0053 0 0110 0 0891 0 0196 0 0043 0 0428 0 0629 0 0696 0 0147 0 0253 7 6432 11 1846 15 4671 13 1891 9 5560 9 9518 20 0147 19 4620 17 1302 17 1748 21 2207 64 748 0688 746 5332 743 2291 740 0242 730 4389 730 3578 722 5549 720 7940 713 5910 710 0772 709 1510 699 3033 695 9776 688 4638 681 8918 680 8008 680 8008 680 8008 683 0898 684 0997 693 0463 694 8339 698 3150 699 6347 709 5662 716 8192 720 3288 727 5724 737 6488 739 3268 748 6264 759 2078 763 4113 0 0985 0 0269 0 0524 0 0562 0 1521 0 0014 0 1239 0 0309 0 1221 0 0558 0 0160 0 1563 0 0583 0 1193 0 1133 0 0124 24 5630 23 1204 42 0565 47 5914 63 5263 49 1593 41 2988 52 5718 32 8590 48 8266 58 2177 68 3216 87 8930 65 2264 54 3268 15 0831 0 4 3865 0 6 3251 0 0388 0 0168 0 1420 0 0314 0 0561 0 0228 0 1602 0 1251 0 0566 0 1271 0 1599 0 0294 0 1476 0 1856 0 0667 3
35. 133 134 4 h sum mx ind sub h sum mx ind 1 16 h sum mx ind E lower h_sum mx_ind_1 a 16 upper lt lower 6 COMP 16 ind Ldtap 0 0 Ldfusval trap 0 8 conf 1 A 8 D gt gt M 1 fus clk 50 8 ind conf 0 See gt gt 1 tap A x ttl ME 7 trap 8 D 8 NOTE Trapht2 and max all run parallel Trapht2 max fus_val are combined into a single module trap2max vhd 135 Unit Element UE X yl 5 comp 229 8 0 amp d Validate Sensor MDin 16 E 7 0 conf D conf TET Ea conf Br gt 27 E 7 sstD A sdata Ar 5 E E gt E 15 8 p 8 sdata Br gt 15 8 DsdatA B 8 2 gt sstd 8 7 0 sstd 8 D 5840 IM i 3 8 0 136 sstd
36. 3 00 2 80 2 60 2 40 2 20 2 00 21 41 61 81 101 121 141 161 181 201 221 241 261 281 301 321 5 361 381 401 421 441 461 481 501 Figure 5 2 Control of Carbon Content Run 2 cast steel pig iron Figure 5 3 Metal Stream Changes suggested by I3PSC for control of for Run 1 136 4000 00 3500 00 3000 00 2000 00 1500 00 1000 00 500 00 0 00 2500 00 1 Melt Rate MR Fused MR v ManualMR rv 0 00 DORM DOK WO QN TT TT 433 451 469 487 505 Figure 5 4 Individual Measurements and Fused Melt Rate for Run 1 Fused MR Confidence Figure 5 5 Confidence of Fused MR 137 1 1 Tron Temperature 2900 00 2850 00 ooo UR Ww 7 2700 00 1 yu W A il juu i d TM ps in Pyro I Fused 2550 00 2500 00 I 116 139 162 185 208 231 254 277 300 323 346 369 392 415 438 461 484 507 Figure 5 6 Individual Measurements and Fused Temperature
37. A 1 4 Matlab Input Output File sv4_in out txt RLH 11 18 99 Execute sv4 m using sv4 fis and TpDtl wk1 raw input data ans 18 Nov 1999 Number of data points sz 1 ans 100 Raw data input time temp a a 1 0e 003 0 0 7792 0 0560 0 7796 0 1180 0 7791 0 1760 0 7764 0 2380 0 7718 0 2950 0 7691 0 3590 0 7658 0 4160 0 7611 0 4750 0 7596 0 5370 0 7553 0 5990 0 7488 0 6590 0 7469 0 7150 0 7470 0 7790 0 7442 0 8370 0 7404 0 8990 0 7335 0 9590 0 7299 1 0170 0 7313 1 0790 0 7241 1 1350 0 7175 1 1970 0 7128 1 2550 0 7095 1 3190 0 7064 1 3760 0 6987 1 4390 0 6925 1 4960 0 6902 1 5590 0 6843 1 6180 0 6806 1 6750 0 6794 1 7380 0 6759 1 7950 0 6782 1 8570 0 6813 1 9140 0 6820 1 9770 0 6871 2 0350 0 6937 2 0980 0 6992 2 1540 0 7030 2 2160 0 7114 2 2790 2 3350 2 3980 2 4540 2 5160 2 5790 2 6360 2 6940 2 7570 2 8140 2 8770 2 9350 2 9980 3 0550 3 1180 3 1750 3 2380 3 2960 3 3590 3 4160 3 4790 3 5360 3 5940 3 6570 3 7160 3 7780 3 8360 3 8980 3 9550 4 0180 4 0750 4 1380 4 1950 4 2580 4 3150 4 3780 4 4350 4 4940 4 5570 4 6150 4 6780 4 7350 4 7980 4 8560 4 9440 4 9750 0 7179 0 7241 0 7300 0 7362 0 7421 0 7546 0 7618 0 7655 0 7698 0 7753 0 7785 0 7765 0 7762 0 7803 0 7801 0 7743 0 7770 0 7746 0 7736 0 7674 0 7685 0 7671 0 7646 0 7607 0 7566 0 7557 0 7543 0 7481 0 7465 0 7432 0 7400 0 7304 0 7304 0 7226 0 7208 0 7136 0 7101 0 709
38. Interface In this option the user can select the modalities that need to be included in the IPSC system Also the variable and the interface related to the modalities can be selected The interface of Select Modalities amp Variable amp Interface is in Figure A 12 All modalities appearing in the Standard Grammar File appear in the Modalities window Once the modalities are selected they become available for the creation of modality groups The variables in the modality can also be selected If that variable exists in other modalities they are also automatically selected For each modality the modality interface VI path is also selected This interface VI associated with that modality will be called when IPSC starts running Click Continue button to close the window after selecting the modalities variables and interfaces 178 179 Modalities Variables amp Execution vi path Selection Modalities Controller Plant Virtual Sensor Planner Controller Plant Virtual Sensor INPUTS IMPUTS INPUTS coke ratio coke ratio actual BR blast rate blast rate total oxygen in blast Blower Freq OUTPUTS SCR pressure drop melt rate oxygen addition rate Kalman MR OUTPUTS O2 Flow Rate Pyrometer Temperature combustion efficiency ratio blast temperature Kalman PT melt rate actual BR 2nd Pyrometer Pyrometer Temperature total oxygen in blast Datacast Temperature Fusion Monitor Expert Controller interface path Plant interf
39. S C w orkNData amp nalysisND ata Run Name albany Setup Online a Files Figure A 63 Online Setup Online Setup vi A 3 2 Simulate Data Collection If the data 15 being collected from an existing data file this option 15 selected The current data file is displayed in the path field If that needs to be changed press the Declare Data File button and declare the text file with the sensor information and data The selected modalities are shown in a listbox select the modality that holds the sensor information that will map the text file data into the standard grammar Next press the Declare Run Name button Select the run name that was created for the current run Once this 15 done the Data button becomes enabled press it and the data is read from the text file into the data structure and written to the data structure file The counter file is also updated to indicate how many rows of data were collected While data is being collected all the buttons are disabled Once the process is complete they are 217 218 enabled You can collect data from another file or return to the previous menu at this point Simulate Online Data Collection vi 19 AFS model variable ANN Model variable Real Albany sensor BE 7 Figure A 64 Simulate Data Collection Simulate Online Data Collection vi A 3 3 Analyze Collected Data This function interrogates the models with the data
40. The controller parameters are adjusted directly only using plant input and output signals The identification and control functions are merged into one scheme Many adaptive control approaches have been developed Gain scheduling self tuning of the controller model reference adaptive control and variable structure adaptive control are some of the most commonly used approaches All of these approaches fall in one of the two categories mentioned above 16 Few approaches that involve both the direct and indirect method have also been developed Gain scheduling is the simplest type of the adaptive control In this approach the controller gains are made dependent on the parameters that can be measured or inferred from other measurements This approach is very conservative and poses many problems if the dependent parameter has high rate of variation 15 Parameter estimation forms the base for self tuning The required parameter is modeled and an observer is implemented to estimate the parameter The controller is designed as a dependent on the estimated parameter This method requires all the state 47 variables for parameter estimation which is not possible in all systems 16 Using state estimator may help but it results in a complex system Model reference adaptive control is based on a reference model for the plant The error between the actual output and the output from the reference model is used to change the controller parameters
41. The format is Path IP For example if the engine VI is on machine 149 149 0 1 and its path is C PPSC Fusion Modality Fusion MultipelSensor Fusion vi the path will be C PPSC Fusion Modality Fusion MultipelSensor Fusion vi 149 149 0 1 184 185 3 While defining the group in this modality the variables associated with this group will be given These variables are Threshold Trend Influence Factors Weight of Expected Value and Weight of Previous Fused Confidence Threshold means the number of points to ignore at the beginning of run Trend Influence Factor is a number of 0 1 that determines how much the trend effects the confidence calculation Weight of Expected Value is a weight factor that determines the effect of expected value calculated using previous fused value and trend on the fused results Weight of Previous Fused Confidence is a factor from 0 to 1 that determines the effect of previous fused confidence over the current confidence Once all the parameters are added click on OK button to add this group to the modality The dialog in Figure A 17 will show back for adding more groups 4 1 1 6 2 1 2 Delete Update Modalities Click Delete Update Modalities in Figure A 16 The dialog as shown in Figure 20 is popped up to delete or update the details in the existing modality groups Delete Update Group vi Modality Details Modality Name Group 2 Input Nodes Groups of Modality Delete Modality Group
42. The user can monitor the real running situation of the controlled cupola furnace by this monitor At the same time several other dialogs will be popped up depending on the modalities you set For example the Controller Options dialog is popped up if you set controller modality in the system A 1 2 1 Cupola Operation Monitor At the beginning of PPSC s running Cupola Operation Monitor is popped up You can monitor the real time running situation of the controlled cupola furnace by this interface The front panel of the monitor is shown in Figure A 46 Bile Edit Operae Tools Browse Window Help en Variable Monitort F warte Monitor2 F2 Trend Monitor 66 Figure A 46 Cupola Operation Monitor 200 201 There are two tabs for variable monitor and one tab for trend monitor each variable monitor tab four variable waveform windows are arranged You can select multiple variables in one window to monitor Select the X axis node first All the waveform on the same window must use the same X axis node Pull down the menu in Data list box to select X axis node The default X axis node 15 Time Plant Raw value node Then Click on the blue Y1 Data button A YI data setup window will popped up This window is shown in Figure A 47 i gt Setup G1 Yvariables vi File Edit Operate Tools Browse Window Help Y Data selected variables
43. This application allows the users to set a single set of inputs pick a model to process those inputs with and view the outputs The model is run only once with one set of inputs This is useful for seeing quick numerical results and testing to see that a model is reporting results correctly The values for the inputs can be changed at this point Pressing Return to Main Menu closes the window and returns control to the offline menu for this and all the other VIs described here 208 209 Single Run Interface vi AFS model variable ANN Model variable Real Albany sensor Figure A 55 Single Run Single Run Interface vi A 2 2 Single Variable Correlation This application is used to vary a single parameter while holding all other parameters constant It is useful for making a large number of iterations on a single variable The other applications can be configured to accomplish much the same tasks but it seems to be useful nonetheless The varied parameter is selected from the list the default value and unit are displayed for reference so the user knows an approximate value The slider in this and the other correlation applications are set to display 1 3 the default value as the lower bound and 3 times the default as the upper bound That does not mean that values outside of those bounds can not be set It also does not guarantee that the bounds make any physical sense For example in figure 7 the upper
44. a 88 CHAPTER d EDU mI UM DID M 92 4 INTEGRATION OF MULTIPLE SENSOR FUSION IN CONTROLLER DESIGN 92 4 1 93 4 2 Controller o 94 4 3 Stability 96 44 Tuzzy Controller 104 4 41 Controller d sigh 5 105 44 2 Smith Predictor oda cocto p neut editas actes 109 4 4 3 Integration of Sensor Fusion in Controller Design 110 444 SIMULATOR DESIGN tiet tic teinte 111 44 5 BASIC DAYOUT 111 4 46 NOISE DISTURBANCES AND VARYING PARAMETERS 112 qui RESULTS acumen d 113 4 4 8 Integration of Sensor Fusion In Controller 118 4 4 9 VARYING MODEL 8 2 121 4 4 10 Varying Pure Time Delay of the CMR see 122 4 4 11 COMBINING ALL NOISES AND DISTURBANCES 124 APPENDIX 4A nep uS 126 CHAPTERS cca tecto 129 5 DEMONSTRATION 8 1 22 129 5 1 arre mE 129 5 2 Setup IPSC for Demonstration Runs
45. however lead to a degree of variability in the cupola output This variability can be reduced by expert operation of the cupola by experienced personnel Reducing this variability is more important for some cast products than for others where iron temperature and composition are crucial as in the production of automotive parts holding furnaces sometimes hundreds of tons in size are used to pool the output of one or more cupolas and temperature and composition can be adjusted before the hot metal goes to the casting line 28 The economic and environmental costs of this variability can be substantial Iron that fails to meet specifications can cause substandard castings or even casting failure the material may be re melted but the energy spent melting it the first time is wasted The costs of installing maintaining and operating large holding furnaces to level out the variability is an additional cost of producing iron Materials such as coke breeze fines from the handling of coke that would cause poor operation if charged from above can also be injected through the tuyeres for added energy the incinerator like nature of the cupola incorporates these and even other hazardous wastes unrelated to cupola operation into the relatively benign cupola outputs cast iron CO CO2 and slag 15 1 2 Objectives and Scope of the Project The purpose of the project was to develop and demonstrate an intelligent integrated industrial process sen
46. our code were compared to the original Matlab results We found that the floating point code produced exactly the same results as Matlab while the fixed point code gave results within 1 of the Matlab results This accuracy is both reasonable and acceptable for our needs Finally the fuzzy logic source code was reorganized by splitting it into two files called setup c and exec c The setup file performs the entire house keeping activity All activities that are performed only once like reading the input data file and FIS file and 20 printing the results were grouped into setup c The exec consists of functions that perform the mathematical operations on the input data This code will be repeatedly executed during normal system operation Functions in the exec file are most likely to be implemented in FPGAs while all of the functions in the setup file will continue to be executed as code on the CPU board The current version of the code was tested using our preprocessed input data discussed in the next section The output of the fuzzy logic code see Appendix was compared with that of the Matlab Fuzzy Logic Tool Box see Appendix It was observed that most of the results were compatible with lt 2 error but some input sequences resulted in errors as high as 40 The cause of the small errors is due to the usage of fixed point numbers instead of the floating point numbers used in Matlab This error is acceptable for our purposes
47. the performance of the system when the multiple sensor fusion fails 94 4 2 Controller Design The controller designed with the assumption that the estimate from the sensor fusion is reliable may drive the system fast into the wrong direction if the estimate was wrong the other hand if the controller is designed considering the worst case scenario it will result in sluggish response One method that can be used to improve the performance of the system is to design the controller such that it has fast response when the confidence is high and a slow response when the confidence is low Hence the controller should be able to adapt itself and produce a controller that adapts its response depending upon the confidence in the estimate The required performance of the controller after the integration of the confidence can be summarized as 1 When the confidence is high the feedback signal is reliable So the controller should be fast enough to track the reference value 2 When the confidence is low the feedback signal is not reliable which implies that even if the controller has tracked the value fed to it the state that the system has reached may not be the correct reference value So the controller should not try to reach the reference value very fast One way to achieve this requirement on the controller is by designing two controllers a fast controller that will be active when the confidence is high and a slow controller which
48. 28 Fused Confidence Plot 81 3 4 6 Summary The section introduced the concept of incorporating trend in Sensor Fusion to deal with cases where reliable trend information sources such as virtual sensors or model are available The algorithm was tested using data obtained under various circumstances and the results are shown The results clearly indicate that the algorithm performed correctly under circumstances of sensors malfunctioning by incorporating trend information 3 5 Fusion of Linguistic Sources In this section the algorithm for multi modal sensor fusion is further strengthened using expert systems as additional sources of information An expert system provides linguistic information on the parameter which has to be converted to numerical form so that the sensor fusion algorithm can fuse it along with information from the other sources We start by fusing linguistic information on the trend of the measurand and then continue with fusing information on the value of the measurand itself 3 5 1 Linguistic Information on Trend Consider an expert system may be an operator who can predict the trend of the parameter This source of information would be in linguistic form and would be quite reliable This source of information need not be available at every instant of time It is 82 possible that the operator can intervene at certain instants of time when there is a sudden change in the operating conditions This section prop
49. 3 Richard R Brooks and S S Iyengar Multi Sensor Fusion Fundamentals and Applications with Software Prentice Hall Inc New Jersey 1998 4 Ren C Luo and Michael G Kay Multiple Integration and Fusion in Intelligent Systems JEEE Transactions on Systems Man and Cybernetics vol 19 no 5 September 1989 5 R C Luo M Lin and R S Scherp Dynamic multi sensor data fusion system for intelligent robots JEEE Journal Robotics and Automation vol RA 4 no 4 pp 385 396 1988 6 Keith E Holbert A Sharif Heger and Nahrul K Alang Rashid Redundant Sensor Validation by Using Fuzzy Logic Nuclear Science and Engineering vol 118 pp 54 64 1994 7 Asok Ray and Rogelio Luck An Introduction to sensor Signal Validation in Redundant Measurement Systems JEEE Control Systems Magazine vol 11 no 2 pp 43 Feb 01 1991 8 Marcello R Napolitano Charles Neppach Van Casdorph Steve Naylor Mario Innocenti and Giovanni Silvestri Neural Network Based Scheme for Sensor Failure Detection Identification and Accomodation Journal of Guidance Control and Dynamics vol 18 no 6 Dec 1995 9 Mohamed Abdelrahman and Senthil Subramaniam An Intelligent Signal Validation System for Cupola Furnace Part 1 and Part 2 American Control Conference San Diego 1999 10 Janice C Yang and David Clarke A Self Validating Thermocouple IEEE Transactions on Control Systems Technology vol 5 no 2 March 1997 11 M P H
50. A 1 1 Matlab M File SV4 m SV4 m Self Validation V4 RLH Modified to print out data for in out txt file Reading the three signal data from wk file a wklread TpDt1 Reading the FIS file to use it to find the self confidence of the sensor readings fis readfis sv4 fis Read the temperature and the time from the file and seperating them into individual data points and creating the three inputs required to the fuzzy system sz size a for j 1 sz 1 The Preprocessing required on the raw data to form the three input to the fuzzy System Getting Change in time ifj 1 ch in time a j 1 else ch in time a j 1 a j 1 1 end Finding the current median and previous median temperature to find the change in temperature value leading to the calculation of one of the input rate of change in Temperature if ch in time 0 temp 1 a j 2 temp 2 a j 2 temp 3 a j 2 temp 4 a j 2 temp 5 a j 2 temp 6 a j 2 else 55 temp 6 temp 5 temp 5 temp 4 temp 4 temp 3 temp 3 temp 2 temp 2 temp 1 temp 1 a j 2 end prev temp median temp 2 6 Previous Median Temperature curr temp median temp 1 5 oCurrent Median Temperature Setting the current median temperature as one of the input to the fuzzy system Median Temperature med temp j curr temp Finding the second input rate of change in temperature ch in temp curr temp prev tem
51. Clarke A Self Validating Thermocouple IEEE Transactions on Control Systems Technology vol 5 no 2 March 1997 11 and D W Clarke Self Validating sensor Rationale definitions and examples Control Eng Practice vol 1 no 4 pp 585 610 1993 12 T M Tsai and H P Chou Sensor fault detection with the single sensor parity relation Nuclear Science and Engineering vol 114 pp 141 1993 13 Mathieu Mercadal Sensor Failure detection using Generalized Parity relations for Flexible Structures Journal of Guidance Control and Dynamics vol 12 no 1 Feb 1989 14 Jeff Frolik C V PhaniShankar and Steve Orth Fuzzy Rules for Automated Sensor Self Validation and Confidence Measure n Proceedings of American Control Conference June 2000 15 Bernard Friedland Advanced Control System Design Prentice Hall Inc New Jersey 1996 163 16 B Wittenmark Adaptive Control Addison Wesley Publishing Co Reading MA 1989 17 Liu Hsu Aldayr D de Araujo Ramon Costa Analysis and design of I O based variable structure adaptive control input output variable structure model reference adaptive control systems JEEE Transactions on Automatic Control vol 39 no 1 4 Jan 1994 18 E Burdet A Codourey Evaluation of parametric and nonparametric nonlinear adaptive controllers Nonlinear controllers Robotica vol 16 no 1 1998 19 Judith Hoche
52. Computer Figure 5 1 Configuration for Interfacing I3PSC with ALRC DAQ for Demo Runs Three of the demonstration runs were focused on illustrating the integration of sensing and control of cupola parameters 52 Setup of IPSC for Demonstration Runs TPSC system was configured with the following modalities a Data Acquisition Modality A modality whose function is to collect and send raw data and IPSC control parameters from and to the ALRCDAQ 133 b Virtual Sensor Modality A modality using models to predict values of important d e cupola parameters Two virtual sensors were configured One for molten iron temperature and the other for the iron melt rate Sensor Fusion Modality For fusion of data collected from the cupola Three parameters were of interest in the demo runs namely iron temperature melt rate and Carbon content of the molten metal Monitoring modality This modality monitors trend of important variables and displays the current trend of such variables such as increasing decreasing constant etc This modality can also be setup to monitor for conditions such as bridging that would be reflected in changes in operational parameters such as cupola back pressure and exit temperature An example of such situation 1s shown later in this chapter Planner modality This is the modality which specifies the run plan in terms of the requirements on the variables of importance This was limited during the demo runs
53. External Advisory Board ect txt te aat oS rud td 19 1 4 3 Coordination of Teams Efforts 20 1 4 4 Overall System VISION e e pe Dens eund 22 1 5 Evaluation based on Proposed objectives eere 23 1 6 Summary and Report 0 26 APPENDIX RN 28 APPENDIX 1V c AE 31 THESES SUPPORTED BY THE PROJECT 31 CHAPTER 2 MER REUS 33 2 MOTIVATION AND 33 2 1 hits M osis 34 2 2 od ke bw eq 36 2 3 Multiple Sensor Fusion and Signal Validation 37 2 3 1 Multiple Sensor Fusion nean 37 253 2 Signal 40 2 3 3 Validattonos 42 24 45 2 5 Eddie 49 CHAPTER 5 RUP MD rer ID e D UP EID ES 52 3 MULTIPLE SENSOR 2 2 1 52 31 Parzen like Methodology for Redundant
54. Figure 3 11 The self confidences of the three sensors are presented in Figure 3 12 Figure 3 13 shows the total confidence in the estimate It is clear from 63 Figure 3 11 that the average method results in the estimated value to be affected by the sudden disturbances as well as by the noise introduced into one of the sensors In comparison Figure 3 10 shows that the effect of the disturbances were mitigated to some extent However a close up of the data in Figure 3 10 shows that the estimated value of the measurand is still affected by the readings of the sensor which was artificially injected with the high noise level This close up is shown in Figure 3 14 It should be noted however that the confidence in the estimates are lower in periods where the agreement between the considered sensors decreases as shown in Figure 3 13 800 750 700 650 TC4 TC5 Estimate 10 20 30 40 50 60 70 80 90 100 Figure 3 10 Estimated Value from PDF without Considering Self Confidence 800 750 700 650 0 8 0 6 0 4 0 2 Estimate 10 20 30 40 50 60 70 80 90 100 Figure 3 11 Estimated Measurand Value Using Average Method ConfTC3 ConfTC4 ConfTC5 10 20 30 4 50 60 70 80 90 100 Figure 3 12 Self Confidence of the Three Sensors 64 65 ost
55. MODEL INTERFACES 220 4 4 1 AFS Model Interface 220 A 4 1 1 AFS 221 168 A 4 1 1 1 Define AFS File Paths A 4 1 1 2 Charge Selection A 4 1 2 Metal Selection Option Menu A 4 1 2 1 Create Material Property Files 223 A 4 1 2 2 Material Selection 224 A 4 1 2 3 Set Metal 5 225 4 4 2 AFS Preprocessor 225 A 5 REAL SENSORS 226 168 169 PPSC Online System User Manual The user can start l PSC Intelligent Integrated Industrial Process Sensing and Control online system by running Cupola Interface vi path is C P psc Application Cupola Interface vi This VI s front panel is as shown in Figure A 1 gt Intelligent Cupola Sensing amp Control Systel Set up application Run Figure A 1 PSC Onl
56. P14 C32 133 C32 133 134 C32 133 134 C33 1 2 C32 133 134 C34 1 2 C32 133 S18 C32 133 134 C32 133 C33 C32 133 C34 C33 C33 135 P15 C33 135 P16 C34 aoa aaa aoa aa of times the function is called Setup functions called by exec c FUNCTION fisMin BLOCK NAME PROBABILTY PROBABILTY 1 0 5 0 5 0 5 0 165 0 165 0 165 0 5 0 5 0 5 0 5 0 5 0 165 0 165 0 165 0 5 0 5 0 5 0 5 WEIGHT ITERATIONS nnn nin nwid WEIGHT OF ITERATIONS Co Ce Cs Cl o gt Ce Ces Ce Co ss WEIGHT ITERATIONS Clock Cycles Weighted Clockcycles Col 1 2 3 11820 205 201 375 193 289 5 0 5 23 3 795 0 1 65 7 2 805 6 8 160 80 205 537 375 193 289 5 0 5 23 3 795 0 1 65 7 2 805 6 8 160 80 21 336 70 1120 Grand Total 12940 Clock Cycles Weighted Clockcycles Col 1 2 3 1 1 1731 05 1731 05 1872 75 12954 1 1 46 736 44 44 17 8 5 15 15 351 351 26 163 170 1792 50 800 7 3 5 29 14 5 3 3 48 768 33 33 0 0 Grand Total 15782 55 Clock Cycles Weighted Clockcycles Col 1 2 3 517 1 of times the function is called FUNCTION fisMax BLOCK NAME S17 1 of times the function is called 779 412 779 412 779 412 779 412 779 412 776 471 771 765 769 412 765 882 761 176 759 412 755 294 748 823 747 059 747 059 744 118 740 588 733 529 731 176 730 000 724 118 717 64
57. SYSCLK CLK Application Self Validation Preprocessing Fuzzy Logic Top Level Hardware Architecture FPGA y 80 BMmode 16 0 PMA 17 1 47 16 0 gt 16 0 BMDout MDWe 15 0 PMDout 15 0 MDin 16 MD 15 0 n 15 0 BMR n BMW n BMCe n n BMDWe PMR_n gt MWE_n n MDWe PMW PMDWe 2 Port Memo Interface PC 104 Bus Interface 81 Base 300h B If usitc to RFad 104 BUS RF p gt MemAdHI Bus MemAdLo Address BAD Decoder Mem lt 7 10 Brd Sel BCS16_n BR_n BMR gt BW_n BMW v BDout eu 16 Memory BMA 16 0 Address Register 77 16 BMDout E 16 4 i 46 BDin 16 RF Address M c Register 9 Y 9 16 interface MDrd RF Bus Port gt 2 port Register BMmode BMmode File BMwrite 8 16 gt gt Proc Port la CLK BR_n BMW gt MS BW n Memory BMCe n Ifc Mem Decoder BMDWe 2000000 MM 16 MDrd 3 ri PBRF n PBw PBD DS to proc 82 83 Buslfc Register File Reg File 27 Bus RFAdReg RFe
58. Sensor Fusion and Control System Components 2 Algorithms for Intelligent Signal Preprocessing Multi Modal Sensor Fusion Model Fusion Sensor and Model Fusion 3 Re configurable Logic Implementation for Intelligent Signal Processing and Sensor Fusion Algorithms 4 Algorithms for Integration of Intelligent Sensor Fusion Data into the Controller 5 Prototype Implementation and Testing for ALRC Cupola 1 4 Project Organization Administration and Execution 1 4 1 Management Organization This project represented a model for collaboration between technical developers industry oversight and end users as represented in Figure 1 The technical expertise was provided by 1 Tennessee Technological University as the main contractor 19 2 Utah State University USU as subcontractor 3 Idaho National Environmental and Engineering Laboratory INEEL as a subcontractor and 4 Albany Research Center ALRC as a subcontractor The industry oversight was provided by American Foundry Society AFS and the end users represented by General Motors GM and US Pipe Detailed management organization of the project technical development team is shown in Figure 2 The main tasks of the project are listed in Table 1 along with the groups responsible for the completion of each task 1 4 2 External Advisory Board In a kickoff meeting in Detroit in January 1999 TTU USU INEEL GM and AFS agreed to create an exter
59. Temperature Fused Confidence Figure 5 7 Confidence of Fused Temperature 138 Figure 5 8 Oxygen Enrichment for Temperature Control for Run 1 Blast rate 320 300 280 260 240 220 200 A Oe NNNNN OM ono st st sb 0 Figure 5 9 Blast Rate for Melt Rate Control for Run 1 Figure 5 2 illustrates the change in the carbon content between the current carbon content and the desired content of 3 It should be noted that a stream of pig iron was added to the metal stream as a disturbance in place of part of the steel in the charge as illustrated in 139 Figure 5 3 The controller corrects for the disturbance by increasing the amount of steel in the charge Adjustments in oxygen enrichment and blast rate to maintain the temperature and melt rate are illustrated in Figures 5 8 and 9 The temperature and melt rate obtained from individual measurements are shown in Figure 5 4 and Figure 5 6 Changes in the confidence in the fused values reflecting agreement between the different measurements are shown in Figure 5 5 and Figure 5 7 Carbon Figure 5 10 Control of C during Run 2 140 Manual MR rv rv Fused MR v Figure 5 11 Changes in MR during Run 2 2900 2850 2800 2750 2700 2650 2600 2550 2500 245
60. This approach introduces a lot of nonlinearity through multipliers and additional error processing Hence determining the stability of the system is very difficult 16 Variable structure adaptive control from input and output variables has been discussed in the paper 17 Variable structure is similar to model reference adaptive control but instead of using parameter estimation it uses signal synreport discontinuous switching control function is designed to generate the sliding surface for the variable structure adaptive control The paper derives the stability of the adaptive control The disadvantage of the variable structure control is that it requires the knowledge of all state variables State estimator may be used but it results in a complex system In 18 Burdet and Codourey compares most of these adaptive control algorithms and have tested experimentally two of the best algorithms It was shown that the Adaptive FeedForward Controller AFFC is well suited for learning the parameters of the dynamic equation The resulting control performance is compared with the measured parameters for any trajectory in the workspace and was said to give better results The 48 paper also introduces an adaptive look up table memory and was shown to be simpler and better for tasks that requires repeating the same trajectory In 19 another type of adaptive control based on switching the controllers is developed In this paper the output of the
61. all the sensors when many sensors are used Efficiency and performance of the measured data are enhanced 3 Several techniques are available to fuse the values from the redundant sensors 4 The most obvious approach is to find the average of the sensor data In this case however the estimate will be affected by the invalid sensor data A simple improvement to this was to have a weighted average of the redundant information A weight is given to each sensor depending upon a threshold The threshold for the current decision is usually the previous estimate This helps in eliminating the spurious data The choice of threshold is important in this method If the process data has large variations between adjacent values the threshold technique may result in removing valid sensor data Kalman filtering technique is generally used for sensor fusion where Gaussian noise exists The performance of the Kalman filter technique depends upon the accuracy of the system model Chapter 12 3 It gives better results if there exists a linear model to the system and if both the system and the sensor noise can be modeled as Gaussian 39 noise Finding an accurate model for systems is not always possible in many cases and most of the real time systems are nonlinear A method developed by Luo and Lin 5 finds the estimate from the multiple sensor fusion of only consensus sensors The method first eliminates those sensors data that are likely to be erroneou
62. confidence level to adapt the speed of response of the controller This is achieved by scaling the output of the fuzzy inference in the fuzzy controller by a function related to the confidence In this paper the confidence is raised to a power and this is multiplied into the output of the fuzzy inference system Assume for example that the confidence level in a signal is 9095 it could be raised to the fifth power which results in reducing the change in the input to 6096 of the amount requested by the fuzzy inference system If the confidence is 50 the change in input 15 reduced 3 of the amount requested by the fuzzy engine Thus the effect of using the confidence in the measurement to scale the fuzzy engine output is to adapt the speed of response of the controller based on our confidence in the measurements This can be effective in mitigating the effects of failed sensor or external disturbances over the performance of the closed loop system 111 444 SIMULATOR DESIGN The model was implemented in Simulink for the transfer matrix first order responses and the time delay of the CMR The model requires loading the data of the steady state variables time constants CMR time delay operating point and input boundaries 4 4 5 BASIC LAYOUT The simulation is to represent a change in the outputs and inputs from a normal operating point The output of the controller reflects the desired rate of change of the system inputs The controller output
63. double click on the Save Setup Information on Setup Menu Figure A 6 to save the setup information A 1 1 6 Modality Setup Once all the modalities needed for system are selected the Modality Setup option Figure A 6 allows you to set up various properties of these modalities like defining groups of the modality e g multiple fusion groups defining properties of variables in the modality etc The Modality Setup window is shown in Figure A 14 IE Model Specific Setup vi x Declare Model Setup Files Return to Menu Modalities Controller Plant Virtual Sensor gun Setup Fusion y Figure A 14 Modality Specific Setup In this window the Modalities indicator shows all the modalities which you ve already selected previously The following steps are taken to setup the modalities 1 Click on Define button to Declare Model Setup File section A 1 1 6 1 180 181 2 Select the modality setup VI and then click Setup VI section A 1 1 6 2 to run the modality setup VI For each modality several setup VIs need to be run Table 1 shows the Setup VIs that need to be run for each modality 3 The Return to Menu button will exit modality setup and close the dialog in Figure A 14 Table 1 Setup VI s of Modalities Setup VIs Planner Controller Plant Virtual Fusion Monitor Modality Modality Modality Sensor Modality Modality Mod
64. environment as a cupola this noise could be considerable 113 Disturbances Disturbances in the form of a square wave were added to each of the integrator signals This represents not knowing exactly the system inputs An example would be setting the blast to increase and not knowing one of the fans were down or the feed of metal into the cupola could have different densities and could be difficult for the human loaders to approximate its weight Varying Parameters There was no research done on finding how the parameters changed non linearly In order to test for all possible cases a sine wave with an offset close to one was multiplied to the transfer functions except for the duplicate transfer function of the Smith predictor For each of the transfer functions the sine wave was at a different initial phase and all were at different frequencies With all at different frequencies the simulation could be run long enough so that all combinations of the parameters within a range could be studied 447 RESULTS The controller was tested under ideal conditions output noise input disturbance sensor fusion noise by varying the model parameters over a wide range and actual CMR time delay being different than that used in the Smith predictor The tests of the controller s robustness and dependability are described below with generated plots of the results 114 IDEAL CONDITIONS The model was tested without noise and with step inputs
65. every time it is needed In addition the number of sensors considered was coded as a dynamic variable whose value is read at the beginning of the execution cycle The code was also modified to handle a maximum of 10 sensors The fixed point code was tested successfully and produced results within the acceptable margin of error 35 2 41 Hardware implementation of the MSF code The reorganized and optimized sensor Multi sensor fusion was partitioned into modules and a data flow graphs of all these modules were created Currently we are working on measuring the execution time for each block and to translate these data flow graphs into VHDL code 2 12 Summary During the year 2000 the Hardware Team completed the FOGA implementation of the signal validation code Currently the group is working on the hardware implementation of the multi sensor fusion algorithm The group is also working on designing the communication protocols of the final system so that the FPGA system can be connected to the microprocessor board The group is confident that they will accomplish all the project goals in a timely fashion 36 Year 3 Accomplishments 3 Overview During the final year extended from 2001 through the first half of 2002 the 8 Hardware Team completed most of the project hardware and software as it was initially envisioned The multi sensor Self Validation SV algorithm was designed simulated and fully implemented in the FPGA h
66. from multiplier and continue processing of data Clock 77a Read the confidence of sensor x 1 from memory Done by the controller Clock 76 Start execution process for values in Set B registers Clock 7c Send sstd to 8 8 multiplier Clock 28a Read sensor data and std deviation from memory for sensor x 1 Clock 8b Load the output values corresponding to the set B into the output registers Clocks 5 through 8 are repeated for each of the remaining sensors 43 Span Vhd This module calculates the span of area in the sum trapezoid to determine the fused confidence The fused confidence is the area on either side of centroid in the sum trapezoid for a span of 3 times the minimum standard deviation TopMSF vhd This is the top evel module It generates all the needed control signals for other modules Trap2max vhd Prior to starting this module 2 clock cycles are spent for the selection of these points by the inp mux The operation of this module is as follow For all the forthcoming cycles the selection of points is performed in parallel with the execution of trap2max Clock 1 Set inputs to 8bit 8bit multiplier and perform the multiplication Clock 2a Reset all the registers by asserting Clr_hsum Clock 3 4 Set inputs to 1 6bit 8bit divider Perform division Clock 3 4b The height for the Oth sensor at trap index is stored in the accumulator register Step 3 4c Store the accumulated height in the h sum array at
67. in this report In this paper a measure of reliability of the sensor data called self confidence is obtained Self confidence is a measure of the agreement between the characteristics of current sensor data and historical 43 sensor data that are deemed valid This self confidence can be used for the detection isolation of faulty sensors An FL based system is developed based on same basic rules that characterize the sensor data namely 1 The data from the sensor should be within a valid range 2 The absolute value of the rate at which data varies should not be higher than a given threshold that is determined using historical data 3 The standard deviation of the sensor data within a certain window should be less than a given threshold and 4 The standard deviation of the sensor data should not be zero which would indicate a constant value This indicates that the sensor is not working properly These requirements are coded as rules in the fuzzy system The input variables used are the data rate of change in the data and the standard deviation of a certain window The membership functions for these input variables are defined by finding the variation and the trend in the historical data The membership functions are shown in Figure 2 4 The limits in the membership function namely MT1 MT2 etc are found from the processing of the historical data The deviation of the data from the curve is considered as the standard devia
68. index location specified by the index input Clock 5 8a Calculate the height for the 1st sensor at trap index Clock 5 8b Compare the current h sum with the previous h sum and store the max h_sum s index and value Repeat clocks 5 through 8 for all the remaining sensors Clock 9 12 Calculate the height for the 2nd sensor at trap index Repeat clocks 1 12 for all the forthcoming index points At Clock clkx 12 h_start h_end 4 perform the following 44 Step clkx 1 Store the index of the higher h_sum max_ind Clock clkx 2 Access the trap element with index max ind 1 and store it in trap_mx0 Clock clkx 4 Access the trap element with index max_ind 1 and store it in trap mxl Clock clkx 6 Access the trap element with index max ind Clock clkx 7a Calculate the mean of trap max_ind 1 and trap max ind and trap max ind 1 and trap max ind Clock clkx7b Select the mean of the element that is closer to trap max ind Number of clock cycles used in this module 12 h start h_end 11 The TopMSF project was created for the implementation of the MSF code with bus interface Busifc and memory interface Memifc The MSF code was then enhanced to run on the new Virtex FPGA board and again simulated successfully However it was not implemented and tested on the Virtex FPGA board with rest of the system due to insufficient time available at the end of the project period 3 5 MSF Block diagrams Block diag
69. level of the input signals The following paragraphs discuss changes and enhancements we made to the algorithm fixed point considerations changes made to membership functions and changes in the choice of the aggregation method The raw input signal values see Appendix are currently represented as floating point numbers in a data file later they will be input one at a time from the DAQ as fixed point numbers This representation is simplest for the Intelligent Algorithm Team in formulating algorithms and implementing them using Matlab as Matlab supports floating point arithmetic for maximum accuracy However as far as hardware implementation is concerned these numbers have to be converted into fixed point numbers as ultimately the output from the DAQ Board is in fixed point Using fixed point numbers also reduces the number of calculations to be carried out in the FPGAs thereby making it less complicated and less expensive Some of the membership functions provided by the Intelligent Algorithm Team exhibited very sharp rising and falling edges As small variations in floating point numbers cannot be represented adequately in fixed point such sharp transitions in inputs might produce large errors So the transitions were widened to represent changes in input more accurately using fixed point numbers 15 The fuzzy logic method adopted for aggregation was the Sugeno method This method was chosen over the Mamdani method as it is much l
70. level reflects how trustworthy the measurements are If the confidence is close to zero then the response should be slow If the confidence is close to one the response should be normal To test the integration of the sensor fusion in the controller design a relatively large disturbance is applied to the output with the cupola at steady state This disturbance represents a failure in the sensors rather than an actual disturbance Figure 4 12 and 119 Figure 4 13 plots of the inputs and outputs of the melt rate for the confidence levels of 0 9 and 0 5 respectively during the disturbance The confidence is taken to the fifth power and multiplied to the controller s rate of change for the blast rate The change in the melt rate due to the disturbance in the measurement at a 90 confidence is 0 038 tons hr and responds much like it would without sensor fusion The change in the melt rate for a 50 confidence was only 0 0038 The figures show that the reduced confidence slows the controller down so that the output changes are 10 of those at a confidence of 90 It does not stop the changes and if the sensors that cause the confidence to go down are not replaced or corrected the output could eventually reach an erroneous value With sensor fusion it is shown that the cupola s operators will have much more time to fix bad sensors before lowering the quality of the product The controller can be easily adjusted by changing the power that the
71. method can be found at http www cs rit edu atk Java sorting sorting html Reg_16 vhd This is a 16 bit register with enable and reset inputs Reg 8 vhd This is an 8 bit register with enable and reset inputs Sen_val vhd This module is used to validate the confidence of a sensor It operate as follows Clock 1a Read the confidence of sensor x from memory Done by the controller Clock 16 Load the conf into CONF A register Clock 22a Read sensor data and std deviation from memory for sensor x Clock 2b Load the sensor data and sensor std deviation into SDATA and SSTD A respectively Clock 23a Read the confidence of sensor x 1 from memory Done by the controller Clock 3b Load the conf into CONF Clock 3c Start execution process for values in Set A registers that includes CONF A SDATA A SSTD A registers 42 Clock 34 Send 58514 to the 8 8 multiplier Clock 4a Read sensor data and std deviation from memory for sensor x 1 Clock 4b Load the sensor data and sensor std deviation into SDATA B and SSTD B respectively Clock 4c Read the value on sq std bus from multiplier and continue processing of data Clock 5a Start execution process for values in Set B registers that include CONF B SDATA B SSTD B registers Clock 5b Send sstd to the 8bit 8bit multiplier Clock 5c Load the output values corresponding to the set A into the output registers Clock 6 Read the value on sq std bus
72. modification of SV code 25 2 3 Develop a Library of Basic Fixed Point Arithmetic Functions 25 2 4 Implementation of the SV Preprocessing Algorithm 26 2 5 Develop Architecture of SV Signal Processor Hardware 27 2 5 1 Select SV procedures for the hardware implementation 27 2 5 2 Separate constants from true 21 2 5 3 Simplify the fuzzy logic procedures 27 2 5 4 Create Block Diagrams is am eo EN race eee nete aen 28 2 5 5 Define Data Structure and Organization esses 28 2 5 6 Define finite state machine controllers 28 2 6 Develop Hardware Design of SV Signal processor 29 2 6 1 Code Hardware Blocks in VHDL imt er ette 29 2 6 2 Simulate Each Entity Code Separately eese 29 2 6 3 Design of the system interfaces eee eter eerte tom et ades 29 264 Desismof the PSM zo beste d ag mesa i 30 2 6 5 Add Blocks to Top Level VHDL Entity sees 30 2 6 6 Download Test and Debug Top Level SV Signal Processor 30 2 7 Develop Multi sensor SV Algorithm ceres 31 2 8 Develop Multi Sensor Fusion Algorithm 31 2 9 MSF C Code Optimization
73. no 11 pp 1543 November 1998 21 Specht D F Probabilistic Neural Networks Neural Networks November 1990 22 Ronald R Yager and Dimitar P Filev Essentials of Fuzzy Modeling and Control John Wiley amp Sons 1994 23 Jeff Frolik and Mohamed Abdelrahman Synreport of Quasi Redundant sensor Data A Probabilistic Approach n Proceedings of American Control Conference 2000 24 Hassan K Khalil Nonlinear Systems Second edition Prentice Hall Inc 1996 25 Mohamed Abdelrahman Kevin Moore Eric Larsen Denis Clark and Paul King Experimental Control of a Cupola Furnace n Proceedings of American Control Conference 1998 144 26 Pascal Gahinet Arkadi Nemiroviski Alan Laub and Mahmoud Chilali LMI Control toolbox 1 0 The Math Works Inc 27 Jeff Frolik C V Phanishankar and Steve Orth Fuzzy Rules for Automated Sensor Self Validation and Confidence Measure Proc of American Control Conference 2000 pp 2912 2916 28 Mohamed Abdelrahman Parameshwaran Kandasamy and Jeff Frolik Methodology for the Fusion of Redundant Sensors Proc of American Control Conference 2000 pp 2917 2922 29 Jeff Frolik and Mohamed Abdelrahman Synthesis of Quasi Redundant sensor Data Probabilistic Approach Proc Of American Control Conference 2000 pp 2922 2926 30 Vipin Vijayakumar Mohamed Abdelrahman Jeff Frolik 4 Convenient Methodology for the hardware implementation
74. of Quasi Redundant sensor Data Probabilistic Approach Proc Of American Control Conference 2000 pp 2922 2926 30 Vipin Vijayakumar Mohamed Abdelrahman Jeff Frolik Convenient Methodology for the hardware implementation of fusion of Quasi Redundant Sensors Proc 3274 South Eastern Symposium on System Theory Florida Mar 2000 pp 349 353 A APPEN DIOE S donors ean ae ED IR 54 Self Validation Example using 54 AV Matlab Pile 54 A 1 2 Matlab FIS File S V4 18 Men 57 A 1 3 Raw Data Input File 58 A 1 4 Matlab Input Output File sv4 in out txt eee 60 A 2 Self Validation Code Documents eere 70 A 2 1 Theoretical Timing Analysis Spreadsheet Sugeno 70 A 2 2 Preprocessed Output File 5 4 pp txt eese 72 A 2 3 Self confidence output file 5 4 22 74 Hardware Block Diagrams 77 A 4 FSM Controller ASM Charts 2 95 4 1 Self Validation Controller ASM 0 95 Timing pv pU 100 AU Self Validation 100 MSF Hardware Block Diagrams ceres 103 REFERENCE S EET 140 54 Appendices Self Validation Example using Matlab
75. of fusion of Quasi Redundant Sensors Proc 3274 South Eastern Symposium on System Theory Florida Mar 2000 pp 349 353 145
76. other MSF modules including the memory interface and the bus interface MSF top d vhd This module is the debugging version of the above MSF top level module Mux8B_2X1 vhd This module is an 8 bit 2 to 1 multiplexer Mux8B_7X1 vhd This module is an 8 bit 7 to 1 multiplexer Mux8B 10X1 vhd This module is 8 bit 10 to 1 multiplexer Pdiv16_b vhd This module is used to select the inputs of the 16bit 8 bit divider from two possible cases An 8 bit 2 to 1 multiplexer is used for this purpose The controller signals for the multiplexer are generated by the master controller Pdiv24 16 vhd This module has a 24 bit 16 bit divider This divider is used called by two modules the area module and the fus conf module 16 8 vhd This module has a 16 bit by 8 bit multiplier and a 2 by 1 multiplexer The MSF code requires 2 instantiation of a 16 bit by 8 bit multipliers The 2 by 1 multiplexer is used to select one set of these inputs for multiplication 41 Pmul8_8 vhd This module has 8 bit by 8 bit multiplier and a 4 by 1 multiplexer The MSF code requires 4 instantiation of an 8 bit by 8 bit multipliers To reduce the code size one multiplier is used to perform the 4 multiplication operations The multiplexer is used to select the inputs to the multiplier out the possible 4 sets of inputs Rearr_un_par vhd This module is used to sort all sensors data points in ascending order using the odd even sort method The C code of this
77. outValue FSNotO Ldconf Conf Reg CLK PBD 15 0 A ConfToPBD Ne 93 Eval5 P 1 FisEvaluate3 CLK MDin 15 0 E Sug Coeff Rego rule_out6 Z6 8 rule_out7 Z7 rule_out8 Z8 rule_out9 Z9 E ug Coeff A ee gt rule out Z0 rule out1 Z1 rule out3 Z3 rule out4 Z4 rule out5 75 rule out10 Z10 rule out11 211 Sug Coeff Reg2 E Sug Coeff d rule out2 Z22 94 Eval5 P 2 FisEval4 and FisEval5 Reg 12b 7 D totalw Divider 8 total w ZA 20b 12b 2 P 8b outvalue CLK firing Strength V 0 11 CIrTW total wf ZB SelRuleOut 3 P d rule 0 20 0 rule out1 Z1 4 rule out2 Z2 rule out3 Z3 Multiplier BX 8b I rule out4 Z4 16 Adder 20 Reg 20b 2727 Dtotalwf rule_out6 Z6 rule_out7 Z7 rule_out8 Z8 rule_out9 Z9 rule out10 Z10 rule out11 Z11 M CIrTWF rule out5 Z5 DO ON O Na FSM Controller ASM Charts 41 Self Validation Controller ASM Charts 95 ASM Chart 96 P 1 Cmd Handshake SV FPGA Controller Reset Idle 0001 PBRF PBR RFA 0 CR cIrMAC SetPA Read Cmd Reg Clear stuff
78. protocol has been appended with a time out facility to restart the system or continue from the point just before failure This task is 100 complete 2 2 4 Outline method for modification of SV code The outline for modifying the SV code to include the CPU end of the high level communication protocol has been completed The new routines that have to be now included in the C program are being written This task will be completed soon 2 3 Develop a Library of Basic Fixed Point Arithmetic Functions As previously stated the code uses the four basic arithmetic functions on fixed point data words of 8 16 24 and 32 bits To meet these computational requirements of the hardware implementation a VHDL library of the four basic operations was developed This library includes adders subtractors multipliers and dividers These operations were designed as functions so that they can be utilized through function calls Different designing techniques of these arithmetic units were utilized The following types were developed for adder subtractors 26 Ripple carry adders Carry Look ahead adders Look up table adders All of these designs were developed for 8 16 32 bit ranges Pipelined version of each of these types was also implemented Different designs of the dividers units were also developed These designs include Divider using Shift subtract algorithm Divider using single register algorithm These two designs were developed for 24 bi
79. rate 128 129 Chapter 5 5 Demonstration Plans 5 1 Introduction The demonstration plans aimed at illustrating the functionality of l PSC technology as it is applied to cupola iron melting furnaces The plans were carried out as proposed at a research facility operated by the US DOE in Albany Oregon ALRC ALRC operates an 18 research Cupola furnace equipped with state of the art instrumentation for measurement of various cupola parameters Moreover in order to carry out demo plans several new instrumentations such as a continuous immersion thermocouple and an ultrasonic radar were installed and tested on the furnace as promising technologies that could be recommended for use in the cupola foundries The parameters of importance to the current demonstration plans were 1 2 Iron Temperature Melt Rate Carbon content of molten iron Off gas temperature and composition Cupola back pressure Blast rate 130 7 Oxygen enrichment 8 Metal stream composition SCR 9 Coke to metal ratio CMR Instrumentation used for measurement of the above parameters included dip thermocouples continuous immersion thermocouple and optical Pyrometers for measurement of molten iron temperature ultrasonic radar for measurement of molten iron level electronic scale for measurement of molten iron weight thermal arrest equipment for quick measurement of carbon silicon and carbon equivalent of the molte
80. seconds for the and 30000 seconds the CMR The controller reacts to and corrects for these disturbances The outputs show a 20 error due to the time it takes to correct for the instantaneous changes in the inputs The plot shows a long history because the frequencies of the square waves were all different allowing worst possible cases to arise 117 dT ___ 1000 dMR _ _ 1000 dC __ 1000 eMR _ _ 1000 eC __ Error pe T Change in output 1000 dO2 100 dBlast _ _ dCMR deT 1000 deMR 1000 deC __ Rate of change in error CMR Change in input Figure 4 9 Step response under ideal conditions 1000 dMR 1000 dC 1000 eMR 1000 eC T ITOT _MR Change in outnut T ti ti 1000 dO2 100 dBlast dCMR deT 1000 deMR 1000 deC BR 9 C MR Rate of change x In error CMR i 118 Figure 4 10 Step response with noisy outputs dT ____ 1000 dMR 1000 dC 1000 dO2 100 dBlast dCMR BR MR CMR Figure 4 11 Step response with input disturbances generated with square waves 4 4 8 Integration of Sensor Fusion In Controller Design The sensor fusion technique provides the controller with a confidence level in the measurements of the output values As explained earlier the confidence
81. shown in Figure 3 24 This information on the expected temperature helps in deciding the correct value in the fusion process Using this estimated measure of the parameter along with the actual measure it can be observed that the effectiveness of the sensor fusion algorithm has improved This is illustrated in Figure 3 25 It can be seen that the fused value determined by the algorithm coincide to a good extent with the correct sensor for several instances of time However the algorithm still fails at multiple points Distibutions of Temperature at previous instant 0 08 Correct Sensor O O7 Erroneous Sensor 4 Fused Distribution O O6 2 O O5 4 4 H E p 680 720 TAO 760 780 Temperature 76 Figure 3 23 Distribution of Temperatures at the previous instant Final Distributions of Temperature 0 04 mig z Expected Distribution 0 035 Distribution of Sensor Reading Fused Distribution 0 03 0 025 0 02 0 015 0 01 0 005 1 660 680 700 720 740 760 780 Temperature Figure 3 24 Final Distribution of Temperature It was observed that the failure of the algorithm is accountable mainly to the effect the erroneous sensor has on the fused distribution Since the erroneous sensor has a very steady performance the fuzzy engine assigns it a very high confidence An improvement to the pr
82. shows the connections between the three modules The internal signals that connect the 3 modules namely 104 ifc 2 Port Mem Ifc and Signal Processor application are marked as 1 2 and 3 The signals that constitute each of these groups are listed here a BMxxx gt BMAI MAI BMModel BMR n I BMW n I BMCe n I BMDWel and MDinl b gt I PBD n I PBW n PBRF I c PMxxx gt PMA I PMDout I PMR n I PMW n I PMCe n I and PMDWe I 78 7 The following equations define the signals used in the block diagram a b RFen RF RFAdReg 2 RFtoBus RF RFtoProc PBRF d PBen PBRF PBA 2 PBW 8 The following equations define the signals used in the BusIfc block diagram a b BMR n BR n BMwrite Mem BMmode BMW n BW n BMwrite Mem BMmode BMDWe BMwrite Mem BMmode MDrd BR n BMwrite Mem BMmode BMce n Mem BMmode BDout BR BrdSel BDin BR n BrdSel 79 128k X 16 SRAM 17 N U8MA 7 0 U7MD 7 0 U8MD 7 0 y 15 8 U7OE n U8OE n U7WE_n U8WE_n U7CE_n 8CE_n in PC104 104 2 5 Buslfc Memlfc MA 1 0 BMxxx 1 a 7o BAD v 1 v 1 MD sy BDAT Rn IOW BW_n MWE n 2 BCS16 n 2 3 516 n MCE n Y v Signal Processor
83. that has been collected The data comes either from a text file as described in the previous section or it is being collected from a cupola in real time Once again be sure the correct modality is selected and declare the run name first To start interrogating the models press the Run Model Analysis button The VI keeps track of how many data sets it has processed and compares that number to the counter in the counter file If there are data sets that have not been processed the VI reads the next data set into the standard grammar and runs it through the models When a data set has been analyzed the analysis VI increments a counter If there is no additional data to analyze the VI waits and checks again a little later When there is no more data to collect either the cupola run is over or there is no more data from the text file press Return to Menu to end the analysis and close the VI 218 219 Online Analysis Computations vi at AFS model variable ANN Model variable Real Albany sensor Figure A 65 Figure 24 Analyze Collected Data Online Analysis Computations vi A 3 4 View Results The results of the analysis are viewed with this VI Declare the run name as before then press View Data The data is read in to the arrays and the variables and modalities are displayed in the ring boxes above the graph Any combination of inputs outputs parameters and modalities can be selected Up to three lines of dat
84. the AFS model and soon to the Neural Net model 226 227 Unfortunately most of the values in the sensor data file are expressed in British units the AFS model requires metric units The conversions are made by referring to the Default British Unit field in the standard grammar Sub VIs are called to make the conversion based on the name of the unit For example if the British unit is F for degrees Fahrenheit the function F vi is called This function converts from degrees Fahrenheit to Kelvin the default metric unit That is why the unit field must always be filled if it is empty lt blank gt vi is called and that file can not exist The conversion VIs are in fact quite simple to create so if a new variable with a new unit is added to the standard grammar sub VI by that name should be created to handle the conversion the unit conversion VIs are stored in DataAnalysis Online Analysis Unit Convert directory 15 5 Diagram Figure A 75 An example unit conversion diagram F vi 221 228 228 Section 2 Hardware Implementation 1 YEAR 1 ACCOMPLISHMENTS 5 1 1 COVER VIG est ED SD 5 1 2 Literature Search ee PN 8 1 3 Hardware Component Acquisitions e sseessooossooesooesoossssesssocesoossoossos 10 BOSE esee e e tee i ed RU tees 10 DYO b rn EEE uada i des 11 1353
85. to the three variables specified earlier namely iron temperature melt rate and Carbon content of the molten metal Controller Modality This modality uses information from the sensor fusion modality as well as planner modality to decide adjustments to the control parameters of the cupola These parameters included blast rate oxygen enrichment coke to metal ratio CMR and steel to cast ratio SCR 134 5 5 Results and Analysis of Demo Runs One of the runs was aimed at ensuring that the cupola instrumentation were working properly and the interface between the PSC and the data acquisition system as well as the integrated system at ALRC are working properly It was also used to test the effect of changing the CMR on the cupola operating conditions The subsequent runs aimed at demonstrating the ability of to control the carbon content of the molten iron by adjusting the composition of the iron stream the melt rate and temperature of the molten iron within an appropriate range The first of these runs illustrates the ability to change the carbon content of the molten iron from 2 8 Carbon to 3 while maintaining the metal temperature and the melt rate constant It also illustrates the ability of the 8 controller to reject disturbances in the form of unknown metal stream that is being introduced into the cupola Partial results of the first of these runs are shown in Figure 5 2 through Figure 5 7 135 3 20
86. value at oxygen addition value at 0 Add more input variables sensor Output Variables of Modality e at 0 rraw value at 0 Add more output variables Figure A 23 Change Variable List 186 187 E Change Variable_UpdateG vi Figure A 24 Change Variable In case the user wants to add more variables to either the input or the output of the modality then the user clicks the Add more input variables or Add more output variables button in Figure A 23 The dialog in Figure A 17 is shown again to add the input output variables When the user selects the Change Parameters option in Figure A 22 an interface as shown in Figure A 25 appears This interface allows the user to change the parameters associated with the Modality Group The contents in this dialog are same as the contents in Figure A 19 Path of VI that executes this group Input Nodes a Output Nodes a Figure A 25 Change Parameters of Modality Group 187 188 4 1 1 6 2 2 Sensor Parameters Setup Sensors monitoring the plant are subjected to self validation tests to ensure the reliability of the data being read by them These self validation tests require the creation of fuzzy FIS files Also other sensor parameters such as Standard Deviation etc need to be defined for each sensor The Sensor Parameters vi is programmed for assigning such sensor parameters The di
87. will be used when the confidence is low The controller is then 95 implemented by changing its parameters between those of fast controller and slow controller using the confidence as the weighting parameter between the two controllers A schematic diagram of the system with the controller designed is shown in Figure 4 2 The resulting expression for the controller parameter with the confidence as the weighing parameter is given below K 1 4 1 a f confidence 4 2 K State feedback matrix of dimension m x m is the number of output variables and n is the number of state variables Redundant WA Weighted Avg Pata K Controller Multiple Self Self Validation Defined Before Sensor Confid On each sensor 41 4 Fusion HEE eS Data 4 Redundant Estimate Confidence Sensors Controller Plant Output Figure 4 2 Schematic Diagram of the System with Sensor Fusion Integrated with the Controller 96 f Nonlinear piecewise continuous function operating on the minimum of the confidence on each state variable estimate from multiple sensor fusion State feedback matrix corresponding to fast controller m x n State feedback matrix corresponding to slow controller m x n The conditions that the nonlinear function f should satisfy to meet the controller requirements specified above are
88. 0 2400 00 Q b i0 Q PF HY YM TF OK 10 TFT DOK YO N DON TFT OO N FT OR Oo v YM WO ed y CN CN NOOO YO st cs Spout rv Datacast rv 2 rv Pyro rv Fused T v Figure 5 12 Changes in MR during Run 2 141 The second run aimed at showing the ability to quickly change the carbon content of the iron Two set points for the carbon were desired The first set point was 3 and the second set point was 2 8 It should be noted that the first set point took longer to achieve than the second set point as the cupola did not reach steady state when the controller was initially turned on The change in the Carbon content is shown in Figure 5 10 The change in set points was also accompanied by a requested change in the melt rate as shown in Figure 5 11 Figure 5 12 shows the change in the temperature of the molten iron during the run Figure 5 13 shows the change in the metal stream going into the cupola as suggested by 8 along with a disturbance in the form of pig iron stream replacing part of the cast iron The last run demonstrated the ability to drastically reduce the melt rate while maintaining the carbon content and the iron temperature within appropriate ranges This was achieved by a change in the CMR prior to the reduction in the blast rate The change in the CM
89. 00 fon Estimate 700 4 650 0 10 20 30 40 50 60 70 80 90 100 Figure 3 15 Estimated Value using PDF Considering Self Confidence 67 TC3 790 TCA 5 TC5 Estimate 780 770 760 750 52 54 56 58 60 62 64 Figure 3 16 Close of Figure 3 15 0 10 20 30 40 50 60 70 80 90 100 Figure 3 17 Confidence of the Estimate from PDF including the Self Confidence Multiple sensor fusion helps the feedback controller by giving a better estimate to the sensor s data but there might be conditions where even this estimate may be poor In other words multiple sensor fusion does not assure reliability at all conditions At these 68 conditions the feedback controller will fail degrading the performance of the system In the algorithm developed the reliability on the estimate is reflected by the confidence This measure of confidence can be used in a way to achieve a better performance of the feedback controller even when the estimate from the sensor fusion fails 34 A unified Framework for Multi Modal Sensor Fusion 3 41 Trend Fusion The independent sources of information for sensor fusion considered in this section include the real sensors themselves and or information regarding the trend of the measurand as provided by other sources such as models or virtual sensors Our goal here
90. 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 0 501960814 0 501960814 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 75 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 0 501960814 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 1 000000000 76 77 Hardware Block Diagrams The following notes help in understanding the block diagrams All names that start with a P are signals that connect with BUS Interface and Processor Interface respectively 4 block diagrams are used to describe the FPGA H W They are listed below a Top level Top vsd b Memory Interface Memifc vsd c Bus Interface Buslfc vsd d Register file RegFile vsd The Memory Interface depicts the 2 port Interface between RAM and Processor RAM and CPU The Bus Interface connects PC 104 Bus to Memory and SP and contains 8 x 16 Register File The Top Level
91. 1 table shows the RMS error between the original signal the corrupted and fused signal The details of the algorithm is given in 31 sampling difference between low and high sampling 1 32 0 20 40 60 80 100 120 140 Figure 3 35 Low sampling rate signal X n sensor failure 0 500 1000 1500 2000 2500 3000 3500 4000 4500 Figure 3 36 Corrupted high sampling rate signal X n 5 5 4 5 fused part fused part fused part 500 1000 1500 2000 2500 3000 3500 4000 4500 Figure 3 37 Fused signal Table 3 1 Mean square error of fusion methods Method MSE No Fusion corrupted signal 1 0856 Spline Interpolation 0 0273 Proposed Wavelet based Method 0 0063 91 92 Chapter 4 4 INTEGRATION OF MULTIPLE SENSOR FUSION IN CONTROLLER DESIGN The main focus of this project is to reduce the risk of a catastrophic response of a feedback control system when the feedback data from the sensors is not reliable while maintaining a reasonable performance of the control system An algorithm for multiple sensor fusion was presented in Chapter 3 Sensor fusion helps in improving the reliability of the measurement It does not however address the control problem when the data are known to be unreliable In certain conditions even multiple sensor fusion could produce an incorrect estimate So the problem still exists even if multiple sensor fusion is us
92. 1 1 3 0 33 1 6 1 98 C8 1 3 0 33 1 6 1 98 C9 1 3 0 33 1 55 18 15 P1 1 1 1 27 27 Grand Total 120 88 of times the function is called 7 FUNCTION fisArrayOperation BLOCK NAME PROBABILTY WEIGHT OF ITERATIONS Clock Cycles Weighted Clockcycles Col 1 2 3 52 1 1 1 3 3 1 1 1 3 43 129 n P1 1 1 1 41 Grand Total 132 of times the function is called 1 FUNCTION NAME fiscomputelnputMfValue BLOCK NAME PROBABILTY WEIGHT OF ITERATIONS Clock Cycles Weighted Clockcycles Col 1 2 3 114 1 1 1 1 1 3 1714 05 1714 05 114_115 1 1 3 2 2 7 231 88 1522 05 114_115_P5 1 1 1 1 169 88 Grand Total 1714 05 of times the function is called 1 FUNCTION NAME fisCompute TskRule Output BLOCK NAME PROBABILTY WEIGHT OF ITERATIONS Clock Cycles Weighted Clockcycles Col 1 2 3 116 1 1 1 146 146 116 117 1 1 4 36 144 116 117 58 4 1 1 34 Grand Total 146 of times the function is called FUNCTION fisComputeFiring Strength BLOCK NAME PROBABILTY 119 1 19 C18 1 2 119 C18 120 1 2 119 18 120 59 1 2 M9 18 120 19 1 2 1 3 M9 18 120 C20 1 2 1 3 M9 C18 120 C21 1 2 1 3 M9 C18 120 S10 1 2 119 C18 P6 1 2 119 C22 1 2 119 C22 120 1 2 119 22 120 511 1 2 M9 22 120 C23 1 24 3 M9 22 120 C24 11 2 1 3 M9 22 120 C25 1 24 3 M9 22 120 512 1 2 119 C22 P7 1 2 119 S13 1 121 1 of times the function is called FUNCTION fisEvaluate BLOCK NAME 515 P11 P12 s16 130 130_P13 C30 1 2 C30 132 C32 C32
93. 1 Parameshwaran Kandasamy Development of Sensor Fusion Algorithms for Redundant Sensors and Integration Controller Design Tennessee Technological university May 2000 Avinash Seegehalli Multi Dimensional Data Structure for Cupola Furnace Information Processing USU 2000 Jie Chen Detection and Extraction of Parallel Hardware During to VHDL Translation Tennessee Technological University May 2003 Sobha Sankaran Hardware Software Codesign Efficient Algorithms for Hardware Synthesis from C to VHDL Tennessee Technological University 2001 Srikala Vadlamani Comparison of Cordic Algorithms Implementation on FPGA Families Tennessee Technological University 2002 32 33 Chapter 2 2 MOTIVATION and OVERVIEW Feedback control systems have gained extreme importance in modern engineering world Feeding back the output has made it possible for systems to perform their assigned tasks with better reliability A number of control techniques have been developed to achieve the desired response from a feedback control system These techniques achieve accurate tracking of the system output along a specified reference value 1 There are also robust techniques that can achieve good performance even if the system is not modeled accurately 2 Robust feedback control system also reduces the sensitivity of the system with respect to the system parameter variation and external disturbances A schematic diagram of a general feedback co
94. 2 0 6993 0 6960 0 6885 0 6819 0 6780 0 6780 0 6808 0 6831 62 5 0350 0 6841 5 0940 0 6948 5 1540 0 6930 5 2170 0 6983 5 2740 0 6996 5 3360 0 7096 5 3940 0 7168 5 4560 0 7203 5 5140 0 7276 5 5760 0 7376 5 6330 0 7393 5 6960 0 7486 5 7530 0 7592 5 8160 0 7634 5 8730 0 7750 5 9360 0 7814 Pre processed data Filtered Temp Rate of Change Variance transpose med temp transpose rate of ch transpose vr ans 779 2346 0 0 779 2346 0 0 0322 779 2346 0 0 0341 779 2346 0 1 6676 779 1388 0 0015 12 3681 776 3767 0 0485 18 3832 771 7934 0 0716 23 5097 769 1202 0 0469 26 9804 765 8402 0 0556 21 4902 761 0989 0 0765 24 4485 759 6134 0 0240 35 1455 755 3357 0 0713 32 9003 748 8223 0 1163 33 1492 747 0019 0 0284 16 0921 746 9355 0 0011 10 7600 744 2286 0 0437 29 0435 740 3850 0 0641 43 1687 733 4850 0 1190 36 0950 731 3228 0 0349 28 1443 729 9124 0 0252 40 3018 724 1163 0 0935 51 5034 717 5365 712 7719 709 5170 706 4234 698 6740 692 4645 690 1521 684 3052 680 5815 679 3757 679 3757 679 3757 681 2703 681 9936 687 0741 693 6850 699 1591 702 9527 711 3555 717 8611 724 1016 730 0004 736 1996 742 0802 754 6400 761 7860 765 5183 769 7866 775 2564 776 2334 776 5442 778 4569 776 5442 777 0336 777 0336 774 5727 774 2689 773 5749 768 4979 767 3620 767 0927 764 5662 760 6639 756 6287 755 7172 754 2766 0 1134 0 0744 0 0571 0 0491 0 1360 0 0986 0 0392 0 1026
95. 3 9988 41 4333 42 0253 30 7168 33 6416 87 8179 78 7328 95 1722 96 1207 81 5600 104 9680 124 5435 107 5099 154 9003 138 6848 Output self confidence transpose conf ans 0 5000 0 5000 0 5000 0 5759 1 0000 65 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 0 9262 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 66 0 9454 0 7652 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 1 0000 67 68 1 0000 1 0000 1 0000 Rules showrule fis ans If Temp is high then self confidence is low 1 If Temp is low then self confidence is low 1 If Temp is ideal and Rate of Ch is Small and var is Normal then self confidencel is V high 1 If Temp is ideal and Rate of Ch is Very P then self confidencel 15 low 1 If Temp is ideal and Rate of Ch is Very N then self confidencel 15 low 1 If var is Constant then self confidencel 15
96. 322776 724 1163191 717 5365106 712 771911 709 5169645 706 4234273 698 6740478 692 4645316 690 1521471 684 3051548 1618 1675 1738 1795 1857 1914 1977 2035 2098 2154 2216 2279 2335 2398 2454 2516 2579 2636 2694 2757 2814 2877 2935 2998 3055 3118 3175 3238 3296 3359 3416 3479 3536 3594 3657 3716 3778 3836 3898 3955 4018 4075 4138 4195 4258 4315 680 5815166 679 3756931 675 8635272 678 1759117 681 270312 681 9936334 687 0741467 693 6850286 699 1591393 702 9526927 711 3554784 717 8610558 724 1016455 730 0004269 736 1995853 742 0802406 754 6399668 761 786003 765 5182727 769 7865598 775 2563547 778 4569227 776 5441776 776 2334429 780 2635 780 1469745 774 2689087 777 0335849 774 5727382 773 5749344 767 3619656 768 4978738 767 0926621 764 5662159 760 6639053 756 6286692 755 7171806 754 2765798 748 0687899 746 5332423 743 229096 740 0242123 730 3577719 730 4389082 722 5548769 720 7940466 59 4378 713 5910424 4435 710 0771502 4494 709 150988 4557 699 3032857 4615 695 9775606 4678 688 4638216 4735 681 8917815 4798 678 0239969 4856 678 0093234 4944 680 8007572 4975 683 0898366 5035 684 0997246 5094 694 8338841 5154 693 0462961 5217 698 3149765 5274 699 6347361 5336 709 5661641 5394 716 8192312 5456 720 3288077 5514 727 5723801 5576 737 6488177 5633 739 3267854 5696 748 6263861 5753 759 2077677 5816 763 4113184 5873 775 0069037 5936 781 3752399
97. 40 The original FPGA board APS X240 with the Xilinx XC4085 device about 80 000 gates was found to be too small to hold the rather large MSF processor and the 46 smaller SV processor So we purchased a replacement FPGA board APS V240 with a much larger device the Xilinx Virtex XCV800 with approximately 800 000 gates As a result some significant changes had to be made to our system design to accommodate the new device and its improved features The device download and configuration code on the CPU board was rewritten and tested with the new FPGA whose configuration details were different from the old FPGA Since the new FPGA had a significant amount of fast on chip memory we decided to eliminate the external memory chips which caused a significant redesign of the memory system and its interface with the SV and MSF processors Several other smaller changes were also implemented successfully on the new Virtex board 3 7 CPU Board We learned all the details necessary to effectively utilize the CPU board on this project As discussed earlier it interfaces to the Host PC via one serial line for all command status and raw data We also interfaced with a monitoring laptop PC using another serial line to allow us to effectively download and debug the CPU code written in C We revised the setup and communication portions of the SV and MSF code to run on the CPU board We implemented the high level communications protocol with the Ho
98. 51 B51 mfsRule52 B52 mfsRule60 B60 mfsRule61 B61 mfsRule62 B62 mfsRule70 B70 mfsRule71 B71 mfsRule72 B72 mfsRule80 B80 mfsRule81 B81 mfsRule82 B82 mfsRule90 B90 mfsRule91 B91 mfsRule92 B92 mfsRule100 B100 mfsRule101 B101 mfsRule102 B102 mfsRule110 B110 mfsRule111 B111 mfsRule112 B112 85 ComFire 00 mfsRule10 mfsRule20 mfsRule30 mfsRule40 mfsRule50 mfsRule60 mfsRule70 mfsRule80 mfsRule90 mfsRule100 mfsRule110 mfsRule01 mfsRule1 1 mfsRule21 mfsRule31 mfsRule41 mfsRule51 mfsRule61 mfsRule71 mfsRule81 mfsRule91 mfsRule101 mfsRule111 mfsRule02 mfsRule12 mfsRule22 mfsRule32 mfsRule42 mfsRule52 mfsRule62 mfsRule72 mfsRule82 mfsRule92 mfsRule102 mfsRule112 pfirin 86 th 2 mfsRuleO BO ON Na SelRule 3 0 mfsRule1 B1 co DY SelRule 3 0 mfsRule2 B2 o DO AO Na 87 ComFire Comparator fisComputeFiringStrength mfsRuleO BO 8 Comparator mfsRule1 B1 8 8 mfs1 lt 2 v 8 LdFS E Firing petrength Reg Firing Strength VO V11 Strength 8
99. 7 712 941 709 412 706 471 698 823 692 353 690 000 684 118 680 588 679 412 PROBABILTY 2 2 Preprocessed Output File sv4_pp txt 0 000 0 000 0 000 0 000 0 000 0 051 0 071 0 039 0 060 0 076 0 028 0 067 0 115 0 025 0 000 0 046 0 057 0 119 0 037 0 021 0 094 0 110 0 071 0 060 0 046 0 133 0 101 0 039 0 101 0 055 0 018 1 WEIGHT OF ITERATIONS 1 0 000 0 000 0 000 1 730 13 149 17 647 23 875 26 990 22 145 25 260 34 602 32 180 31 834 15 917 10 727 28 374 43 253 35 986 29 066 40 138 50 173 62 976 40 138 40 138 61 592 57 785 62 630 41 176 27 336 24 913 7 266 Clock Cycles 19 Grand Total Grand Total 19 19 Weighted Clockcycles Col 1 2 3 19 19 72 679 412 679 412 681 176 681 765 687 059 693 529 699 412 702 941 711 176 717 647 724 118 730 000 736 471 742 353 754 706 761 765 765 294 770 000 775 294 776 471 776 471 778 235 776 471 777 059 777 059 774 706 774 118 773 529 768 235 767 647 767 059 764 706 760 588 756 471 755 882 754 118 748 235 746 471 742 941 740 000 730 588 730 588 122 393 720 588 713 529 710 000 0 000 0 000 0 028 0 009 0 083 0 115 0 094 0 055 0 145 0 101 0 115 0 094 0 101 0 101 0 211 0 110 0 060 0 074 0 090 0 018 0 000 0 028 0 030 0 009 0 000 0 037 0 009 0 009 0 092 0 009 0 009 0 039 0 064 0 069 0 009 0 030 0 092 0 030 0 055
100. But the large errors are due to problems in the fuzzy logic membership function definitions and how they relate to the fixed point arithmetic which are currently being reconsidered We fully expect to fix the problems causing these large errors 1 5 2 Self Validation Preprocessing Code The algorithm devised by the Intelligent Algorithm team for processing the raw input values to determine the preprocessed input values was coded The raw input values represented using floating point numbers were converted into fixed point numbers The preprocessing was then carried out on these values to generate fixed point preprocessed input values 21 The code was successfully implemented using fixed point arithmetic using floating point scaling factor Scaling factors are constants determined from the range of the floating point numbers These factors are used to determine the fixed point equivalent for the floating point number A new method has been devised for implementing fixed point arithmetic using a fixed point scaling factor The implementation of the above method is now in progress The Matlab floating point output results see Appendix were compared with the ones generated by the fixed point code see Appendix The results were found to be compatible with lt 2 error which is within expected error bounds 1 6 Communication Software Development the CPU to Host Interface The CPU to host interface consists of a serial communication li
101. C gt oc o BW sx co teot o g co cf mm ere Ss 605 gt RN A A Figure 5 21 Detection of Bridging in the Cupola Changes Exit Temperature 148 Cupola_Press Figure 5 22 Detection of Bridging in the Cupola Changes in Cupola Pressure 149 5 2 Lr v ies 29 Figure 5 24 Cupola Always Provides Operational Challenges 150 151 Figure 5 26 Manual Sampling and Quick Analysis of Molten Iron Figure 5 27 Manual Measurement of Temperature of Molten Iron Figure 5 28 Optical Pyrometers for Continuous Measurements of Iron Temperature 152 Figure 5 29 Dip Thermocouple for Continuous Temperature Measurement Figure 5 30 Charging Deck of the Cupola at ALRC 153 Figure 5 31 Measurement of Melt rate Chemical Composition and Temperature Figure 5 32 Remote Monitoring and Control of the Cupola during Demo Runs 154 155 Chapter 6 6 1 Summary and Conclusions Section 1 of this report has reviewed major highlights of the project including the management activities development of algorithms for multiple sensor fusion and integration of sensing and control and demonstration runs on a cupola iron melting furnace in Albany research center The project involved in addition to the algorithms development the creation of a flexible software package based on obje
102. DOE s Albany Research Center ALRC Testing the developed for regulations of melt rate temperature and selected iron composition on the ALRC experimental cupola furnace 17 1 3 Deliverables The final goal of this project was the development of a system that is capable of controlling an industrial process effectively through the integration of information obtained through intelligent sensor fusion and intelligent control technologies The industry of interest in this project was the metal casting industry as represented by cupola iron melting furnaces However the developed technology is of generic type and hence applicable to several other industries The system architecture was divided into the following four major interacting components 1 object oriented generic architecture to integrate the developed software and hardware components 2 Generic algorithms for intelligent signal analysis and sensor model fusion 3 Development of supervisory structure for integration of intelligent sensor fusion data into the controller 4 Hardware implementation of intelligent signal analysis and fusion algorithms Table 1 1 lists the deliverables as they appeared in the proposal They are listed here for completeness As will be illustrated in the current report the objectives stated in the proposal have been achieved 18 Table 1 1 5 Project Tasks Task Description 1 Generic Structure for Integrating
103. Final Report Integrated Intelligent Industrial Process Sensing and Control Applied to and Demonstrated on Cupola Furnaces US Department of Energy Contract DE FC02 99CH10975 2 Tennessee Technological University Utah State University Idaho National Environmental Engineering Laboratory Albany Research Center Final Report Integrated Industrial Process Sensing and Control Applied to and Demonstrated on Cupola Furnaces US Department of Energy Contract DE FC02 99CH10975 Center for Manufacturing Research Tennessee Technological University Mohamed Abdelrahman Principal Investigator Roger Haggard and Wagdy Mahmoud Utah State University Kevin Moore INEEL Denis Clark and Eric Larsen Albany Research Center Paul King Section 1 Algorithms and Software Development Table of Contents ALGORITHMS AND SOFTWARE DEVELOPMENT 3 LIST OF FIGURES 2 9 T CHAPTER 1 25 setae 13 1 CHAPTER e 13 1 1 13 1 2 Objectives and Scope of the 15 1 3 pee 17 1 4 Project Organization Administration and Execution 18 1 4 1 Management Organization sse nennen 18 14 2
104. Implementation The Multi Sensor Fusion MSF processor was fully developed simulated and validated for the original FPGA board All the VHDL modules were developed These modules include Area_Top_R10 vhd This module calculates the area to the right and left of the centroid for every sensor up to 10 sensors Each sensor calculations requires 6 clock cycles as follows Clock 1 The points of the trapezoid for a sensor are selected Description of this operation is given in inp mux module Clock 22a The local controller compares the 4 points with the centroid and calculates the internal control signals sel sel Com Description of this operation is given in the control module Clock 2b the 8bit by 8bit multiplier is used 38 Clock 3 The 1698 multiplier multiplies the output of 8bit by 8bit multiplier with constant 102 Clocks 4 5 The 24bit 16bit divider divides the product from 16 8 multiplier with square of std of the sensor Clock 6 The quotient of the previous operation is accumulated by the accumulator Total number of cycles needed by this multiplier 6 no of sensors Busifc vhd This module contains the PC 104 bus interface for the FPGA signal process It defines the connections between the PC 104 bus memory interface register file and the SV MSF processor Centroid vhd This module is used to compute the centroid of all sensor trapezoids The module has an accumulator to add all the sensor data values T
105. It is composed of several modalities that handle the data as it is collected from the sensors fuse this data along with other data that are pertinent to the cupola operation such as data coming from other sensors virtual sensors models or expert systems MMSF The data is then fed to an intelligent controller which decides based on the required operational parameters which input variables to manipulate The required operational parameters are fed to the controller using the planner The planner can be used by the user to plan offline how the heat will be conducted However it can also be used online to make changes as appropriate to the heat plan 23 Control Room Figure 1 3 Overall System Vision for I3PSC applied to Cupola Furnaces 1 5 Evaluation based on Proposed objectives In fulfilling the proposed objectives the following has been achieved e Innovative sensor fusion algorithms based on a new concept has been developed implemented and tested These Algorithms allow for the fusion of quasi redundant sensors data and produces a best estimate and a parameter indicating the degree of confidence in the measurement The algorithms were presented in conference publication namely the American Control Conference ACC and 24 published in the prestigious journal of ZEEE Transactions on Instrumentation and Measurements The developed algorithms were improved to incorporate trend information as well as linguistic infor
106. R was accompanied by a change in the metal stream to compensate for the expected effect of the increase of CMR on the carbon content The main reduction in the melt rate was produced by a drastic cut in the blast rate and oxygen enrichment It should be noted that the forward change in the CMR reduce the required decrease in the blast rate and keeps the metal temperature within the 142 desired range after such drastic cut in the blast rate and oxygen enrichment These changes are illustrated in Figure 5 14 to Figure 5 20 Figure 5 14 shows the change in CMR in anticipation of the request for a change in the MR and the corresponding adjustments in the metal streams to compensate for that change Figure 5 16 and Figure 5 17 show the change in the blast rate and oxygen enrichment to achieve the desired MR QN TO DON s e xm e e o N TFT O O O x re 361 381 401 421 441 461 481 cast steel pig iron Figure 5 13 Metal Stream Changes control of C for Run 2 143 Figure 5 14 Forward Change in CMR to Achieve Large Change MR Run 3 Figure 5 15 Changes in Metal Stream to compensate for Change in CMR 144 Oxygen Enrichment A 0 Tr O DN
107. Redundant Sensors Proc of American Control Conference 2000 pp 2917 2922 29 Jeff Frolik and Mohamed Abdelrahman Synthesis of Quasi Redundant sensor Data Probabilistic Approach Proc Of American Control Conference 2000 pp 2922 2926 30 Vipin Vijayakumar Mohamed Abdelrahman Jeff Frolik 4 Convenient Methodology for the hardware implementation of fusion of Quasi Redundant 165 Sensors Proc 3274 South Eastern Symposium on System Theory Florida Mar 2000 pp 349 353 31 Mohamed Abdelrahman Min Luo and Jeff Frolik Wavelet Based Sensor Fusion for Data with Different Sampling Rates in Proceedings of American Control Conference Washington D C June 2001 32 Mohamed Abdelrahman et al Integrated Intelligent Industrial Process Sensing and Control Applied to and Demonstrated on Cupola Furnaces Progress Report Year 1 DOE contract DE FC02 99CH10975 March 2000 33 Mohamed Abdelrahman et al Integrated Intelligent Industrial Process Sensing and Control Applied to and Demonstrated on Cupola Furnaces Progress Report Year 2 DOE contract DE FC02 99CH10975 March 2001 34 Mohamed Abdelrahman et al Integrated Intelligent Industrial Process Sensing and Control Applied to and Demonstrated on Cupola Furnaces Progress Report Year 3 DOE contract DE FC02 99CH10975 March 2002 APPENDIX 8 CUPOLA INTERFACE APPLICATION USER MANUAL Version 2 1 January 4 2003 167 Table o
108. Sensor Fusion S3 3 1 1 Description of Parzen Estimator ee sett eda 53 3 1 2 Estimation of Measurand Value from PDF sss 54 S L3 Contidence in Estimate 5 occ aite nauta te caeci ee dier te a 57 32 Considering Self Confidence in Redundant Sensor Fusion 59 3 3 Application and 22 5545550 62 3 3 1 Results of the Sensor Fusion Methodology without Considering Self Confidence 62 3 3 2 Results of the Sensor Fusion Methodology Considering Self Confidence 65 3 4 unified Framework for Multi Modal Sensor Fusion 68 SAV Trend FUSION 68 3 4 2 Multiple Sensor PUSIOD sot render tt rat aee redet res 69 3415 Fusion based on Trend sz oec eed ood ed quen ct 72 3 4 4 Confidence based on agreement among the Sensors 77 3 4 5 Measure of Fused 80 SUMMA a mico ia 81 3 5 Fusion of Linguistic Sources 21 81 3 5 1 Linguistic Information on 81 3 5 2 Fusion of Linguistic Information on the Measurand 84 3 6 Wavelet Based Sensor Fusion for Data having Different Sampling Rates 88 2 01 ose
109. a can be placed on the graph If you want to select a different run for instance if you want to see how a previous run looked press the Reset Data button and declare a new run name and continue as before 219 220 View Data Structure vi File Edit Operate Project Windows Help gt Declare Number et Property Source Variable Run Name Graphs SAFS model Final Carbon Wen value Albany Carbon I Data 4 value 4 Real Albany 4 Offgas 02 Reset zin pata 18 0 16 0 14 0 Return to 120 S 10 0 8 0 6 0 4 0 20 0 0 20 D 1 1 1 1 1 1 1 0 50 100 150 200 250 300 350 400 450 500 550 Figure A 66 View Results View Data Structure vi A 4 Model Interfaces 4 1 AFS Model Interface The AFS Model is currently the most important model available much of the previous work was designed with the AFS model in mind although the interrogator should be easily applied to any model If used properly the AFS model interface requires no user intervention in order to run This is because of the potentially large number of model runs involved in a correlation analysis 220 221 AfsModel_ vi Figure A 67 AFS Model Interface Screen AfsModel vi When the model runs a DOS window opens and displays some error information about the numerical approximation The DOS window should be set to close upon execution this is done by op
110. a certain range The power control is a parameter that determines how fast the confidence effect over the controller rolls off as the confidence deviates from the high level 205 206 2717 43 271842 222113 266 25 gano Figure A 52 Change Controller Option Other Parameters The fourth tab Figure 3 10 displays a history of the charge that has been added to the furnace including the charge number time the coke to metal ratio steel to cast ratio as well as the set point for Carbon that was desired at the time the charge was added the predicted value of carbon when this charge reaches the melting zone and the actual value of Caron measure when this charge reaches the melting zone 206 207 Change the Control Option Here Figure A 53 Change Controller Options Charge Setup 207 208 2 TPSC Offline System User Manual Offline Analysis is concerned with producing Neural Network data sets doing correlation studies and interrogating models offline It consists of tools to setup and process the data sets and other tools to view the results in graphical format Offline GUI vi Offline analy Single Single Variable Correlation Multi Variable Correlation Nominal Multiple Correlation View Single Variable Graphs View Multi Variable Graphs View Nominal Multi Variable Graphs Figure A 54 Offline Analysis Menu Screen Offline A 2 1 Single Run
111. ace vi path Virtual Sensor interface vi path C I3psc Controller 2 C I3psc Plant 2 8 Modality Controller Sensors Modality Interfacel Albany Cupola Virtual Sensor _ Cocirolartoiast ana Selected Variables Melt Rate 2 Fusion INPUTS OUTPUTS melt rate Pyrometer Temperature Kalman PT 2nd Pyrometer Spout Temperature Melt Rate 2 metal temperature 3 Fusion interface vi path C I3psc Fusion _ Modality Fusion Interface Fusioninterface v Figure A 12 Modalities Variables and Interfaces Selection A 1 1 4 Select Variable Properties The user can select variable properties in the dialog shown in Figure 5 Highlight the properties in the parameter list to be added to the data and then click on Add to List button The Return to Menu button will close this dialog Parameter List raw value value Selected Parameters confidence es value standard deviation trend standard deviation trend trend confidence value confidence scaled value controller state Add to List eturn to Menu Figure A 13 Select Variable Properties 179 180 A 1 1 5 Save Setup Information After you select the modalities variables and interface for this system in Figure A 12 you should save the setup These setup information will be used in the following Modality Setup discussed here later Just
112. ality for monitoring trends of specific variables and alerting operator when certain patterns take place The software and data structure were designed to allow for easily incorporating other modalities and modifying the existing ones e The integrated system was successfully demonstrated on a research cupola facility in Albany Research Center Albany Oregon The demo involved successful interface of the developed system to the existing DAQ system monitoring and controlling the main parameters of the cupola furnace namely molten iron temperature melt rate and Carbon composition using manipulated variables namely oxygen enrichment blast rate steel cast ration and coke metal ratio The control system ability to achieve and maintain operational parameters as well as reject disturbances and minimize transition periods was illustrated A list of the papers supported by the project and published in refereed journals and conferences is presented in Appendix 1 A A list of academic theses supported by 26 the project is listed in Appendix 1 B Other information pertaining to the project achievements were presented in previous reports 32 34 1 6 Summary and Report Organization In summary the project has achieved all the proposed objectives starting from development of algorithms for sensor validation and fusion integration of sensing and control development of a package for integrating system components and a proof of concept of usi
113. ality groups vi X X x SensorParameters vi X X Planner setup vi x Monitor setup vi Charge setup vi X Controller setup vi X A 1 1 6 1 Declare Model Setup File Click Declare Model Setup File button on the Modality Specific Setup dialog Figure A 14 will open the dialog shown in Figure A 15 The VIs that perform the modality setup are defined here The setup VIs required by every modality are list in Table 1 Each VI s are selected by clicking the Look Up button The names of these VI s will be appeared in the pull down menu in the Modality Specific Setup dialog Figure A 14 Define Model Setup Files vi Modalities Monitor Real Alb PEE s 3 C MyD ocuments vipin CupolaProj LabVlEW_new Setup Modalities Modality Look Up A 5 C sMyDocuments svipinsCupolaProisLab VIEW new Monitoring ModalitysT rend Look Up Figure A 15 Declare Model Setup File 1 1 6 2 Run Setup VI Click Run Setup VI button on the Modality Specific Setup dialog Figure A 14 will run the modality setup VI selected in the list box 181 182 As shown in Table 1 Modality Groups VI and Sensor Parameters VI two main modality setup VIs that are required by most of the modalities Modality Groups vi is used to define groups in a modality and to delete or update the groups in the modality Sensor Parameters vi is used to set
114. alog for setting up sensor parameters is as shown in Figure A 26 Sensor Parameters vi Variables in Modality blast rate sensorl 221 sensor3 oxygen addition rate melt rate melt rate_SP Figure A 26 Sensor Parameters Setup The Variables in Modality list box shows the list of sensors in the modality To setup the parameters double clicking on any of the variables This opens a dialog as shown in Figure A 27 I Sensor Details vi Sensor Parameters Self Confidence Measure Standard Deviation Measure Save Exit Exit without saving Figure A 27 Sensor Parameters Interface Clicking on the Self Confidence Measure a dialog shown in Figure A 28 allows the user to set up self validation by creating FIS file or by assigning self 188 189 confidence If the user chooses to create FIS dialog as shown in Figure 29 appears The user can create the FIS file gt Confidence Menu vi Self Confidence gt New Create Fis File vi f us Figure A 29 Create Self validation Fuzzy FIS File Double clicking the Standard Deviation Measure in Figure A 27 a dialog as shown in Figure A 30 is opened to calculate the standard deviation values from history data or to assign the standard deviation value 189 190 gt Standard Deviation Menu vi Figure A 30 Standard Deviation Option Menu If the user selects to calculate th
115. an Sensor Data Std Dev Self Confidence Figure 3 9 Block Diagram of Multiple Sensor Fusion 62 block Diagram of the Multiple Sensor Fusion Algorithm Developed including the integration of self confidence is shown in Figure 3 9 3 3 Application and Testing In the following section the testing results are presented for the redundant sensor fusion methodologies presented in sections 3 1 and 3 2 The testing was performed using data from an experimental cupola iron melting furnace in Albany Oregon The system uses three temperature sensors that measure the temperature of the iron melt produced from the furnace These sensors were quasi redundant as explained in 23 The sensors data of two of the temperature sensors were translated using a linear regression relation to give an estimate of the third sensor The resulting data were then treated as if the sensors were redundant sensors 3 3 1 Results of the Sensor Fusion Methodology without Considering Self Confidence The results of testing the methodology of integrating redundant sensors presented in Section 3 1 are presented first Figure 3 10 shows the results of the test The data from one of the sensors TC5 were artificially perturbed by injecting sudden disturbance at t 10 minutes and high noise level into the sensor in the range t 40 to 70 minutes For comparison purposes an estimate of the measurand value using the average of the sensor data is presented in
116. ardware with significant speedup over a standard microprocessor implementation The Multi Sensor Fusion MSF algorithm was designed and simulated for the FPGA A new Virtex FPGA board was purchased and utilized in the system The CPU board was programmed and tested as the interface between the Host PC and the FPGA board New serial communication protocols and software were developed and tested An overall top level controlling application was created for the system user in Labview on the Host PC Details of these accomplishments are given in the following subsections 32 SV Implementation The multi sensor Self Validation processor was first simulated and validated using VHDL and Xilinx tools Next it was successfully implemented on the original FPGA board and then revised to work on the new Virtex FPGA board The complete system consisting of the Host PC running Labview code the CPU board running C code 37 and FPGA with the SV processor implemented with VHDL code was debugged and successfully tested 3 3 SV Speedup The final speedup calculated for the Self Validation processor on the FPGA was found to be significant Processing on the FPGA takes only 50 clock cycles versus 15782 clock cycles on a Pentium microprocessor giving a speedup of 315 Of course the total processing time also depends on the clock frequency which is higher on the Pentium but this is still an important accomplishment for the hardware team 34 MSF
117. are usually used to get a better estimate for the desired variable These techniques will be discussed in detail in 39 Chapter 2 schematic diagram of a closed loop system that utilizes the multiple sensor fusion is presented in Figure 2 2 Analytical Sensors Best Estimate Multiple Inferential Sensors Sensor Fusion 4 Physical Sensors qt gt Reference Error Controller Plant Inferential Input Detector Sensor Plant Analytical Model Sensor Figure 2 2 A Feedback Control System with Multiple Sensor Fusion The above techniques are used to reduce the sensitivity of the system performance with respect to sensors failures This is accomplished by not relying on a single sensor measurement For multiple sensor fusion or signal validation techniques to work satisfactorily certain conditions need to be satisfied These include for example the availability of redundant sensors an accurate plant model or known relations between variables Since most techniques still rely back on other sensors for the feedback value there will be situations where the feedback value is not reliable measure for the performance of the signal validation or multiple sensor fusion technique needs to be developed and utilized in the controller structure 36 2 2 Research Approach The research focus of this rep
118. ary features PC 104 8 bit and 16 bit bus compatible Pentium 166MHz processor 2 serial ports 1 EPP ECP parallel printer port 512 KB flash ROM 8 MB dynamic RAM 8 MB flash disk DOS compatible BIOS MS DOS 5 0 software Borland C 3 1 software and book 11 After receiving the board we studied its documentation and performed a basic operational test on the unit by connecting its monitor port to a PC and downloading some small programs for execution It passed all the tests and is working well at this time 1 3 2 DAQ Board We wanted a compact reliable 12 bit accuracy DAQ board for interfacing to the analog sensors and controllers of the cupola After searching through numerous sources we decided to purchase a data acquisition system the DaqBook 112 with optional DBK11A from IOtech Inc with these principle features Link to PC via standard or enhanced EPP parallel port 12 bit analog resolution 100 KHz sample rate 8 differential or 16 single ended analog inputs Expandable to 256 inputs 2 analog outputs Programmable gain of 1 2 4 or 8 per input channel 4 digital inputs and outputs Operate on 10 to 20 VDC power source AC adapter Packaged in suitable stand alone enclosure Screw terminal card with 40 terminal blocks for analog I O Drivers for Windows and DOS using C or C Driver for Labview 12 DaqView PostView data acquisition software We received this board but it has not yet been tested or in
119. asure of that sensor Such a distribution is constructed for each of the sensors and these distributions are added up and normalized The reading corresponding to the peak on the larger side of the centroid of the joint distribution is the fused measure The confidence in the fused value was considered as the area enclosed within three standard deviations on either side of the fused value 70 Multiple Sensor Fusion without Trend Information 800 r r Erroneous Sensor 7801 760 Fused 740 720 D3 700 Correct Sensor 680 Information from Trend Sensor 660 0 10 20 30 40 50 60 70 Instants of Time Figure 3 18 Multiple Sensor Fusion without trend information In the MSF algorithm the major emphasis was on the absolute measurements of the sensors and their self confidence There however are cases where trend information if available can provide useful information Consider for example a sequence of sensor measurements as shown in Figure 3 18 The fused value at points where all sensors meet would be of very high confidence and the confidence values at all other points would be low This is illustrated in Figure 3 19 Figure 3 20 shows the plots of Information regarding the trend In this figure we assume that we have an additional source of information on the trend but the algorithm does not use this information for the fusion proce
120. ata for one charge material at a time It is easiest to copy and paste the data from another application so that the spacing remains correct Be sure to double check that all the numbers are correct for the metal that is being declared When all the data is correct for the material press the green Save File button If there are more metals to declare fill in the fields as before otherwise press Done 223 224 Ex Create Metal Properties Figure A 70 Material Property File Creation Create Metal Properties vi 4 4 1 22 Material Selection Once all the metals are created the second option allows the user to select the metals that will be used At the upper left of the screen there is a field that says Search Pattern If you use a common parameter when declaring your metal names such as ALRC for the Albany Research Cupola you can use that key to filter out the material files that you won t be using Select the materials from the lists in the same order that they appear in the cin 264 file or else the AFS model won t execute properly 224 225 Material Selection vi Figure A 71 Material Selection Metal Selection vi A 4 1 2 3 Set Metal Mass The one parameter that can be varied from the original cin 264 file is the metal mass Use this option to change the mass of each material Figure A 72 Set Metal Mass Set Metal Mass vi A 4 2 AFS Preprocessor The AFS model has some interesti
121. ation and confidence measure in Proceedings of American Control Conference Chicago IL June 2000 Vipin Vijayakumar and Mohamed Abdelrahman A convenient methodology for the hardware implementation of fusion of quasi redundant sensors Proceedings of 32nd SSST Conference Tallahassee FL Mar 2000 pp 349 353 Param Kandasamy and Mohamed Abdelrahman A Methodology for Integrating Multiple Sensor Fusion in the Controller Design in Proceedings Of 32 SSST conference Tallahassee FL March 2000 pp 115 118 12 13 14 30 Mike Baswell and Mohamed Abdelrahman Intelligent Control of Cupola Furnaces in Proceedings of the 34th SSST conference Huntsville AL March 2002 pp 435 440 Wagdy Mahmoud Hardware Implementation of Automated Sensor Self validation System For Cupola Furnaces Proceedings of 31st conference on Computers and Industrial Engineering San Francisco CA Feb 2 4 2003 Mohamed Abdelrahman et al A Methodology For Multi Modal Sensor Fusion Incorporating Trend Information in Proceedings of 31st conference Computers and Industrial Engineering San Francisco CA Feb 2 4 2003 31 1 Theses supported by the project Min Luo Fusion of Multi resolution Sensors using Wavelet Transform Tennessee Technological university September 2001 Vipin Vijayakumar A Methodology for Multi Modal Sensor Fusion June 200
122. bound for the blast fraction is set at 270 which does not make sense because the blast fraction cannot surpass 100 Care should be made to insure appropriate data ranges 209 210 Correlation vi INPUTS coke ratio cupola diameter AES model variable ANN Model variable astr blast temperature Real Albany sensor total oxygen in blast ambient temperature relative humidit Figure A 56 Single Variable Correlation Correlation vi A 2 3 Multi Variable Correlation The Multi Variable Correlation application queries the selected numerical model for every possible combination within the input range As a result the number of model runs increases exponentially as variables are added and iterations per variable are increased The number of runs necessary n for a complete set is given by the number of variables selected v and the number of iterations per variable n i In figure 8 six variables are selected and set to run six times each resulting in 46 656 model runs On an average computer this would take about one month Keeping the total number of iterations below approximately 10 000 allows for completion in about a week 210 211 Multi Correlation vi Bis _ mam coke ratio cupola diameter diameter t bla Figure A 57 Multi Variable Correlation Multi Variable Correlation vi A 2 4 Nominal Mu
123. cation Clock 3 The 16bit 8bit multiplier multiplies the output of 8 8 multiplier with the constant 102 Clocks 4 5 The 24bit 16bit divider divides the product from 16 8 multiplier with square of std of the sensor Clock 6 Appropriate fus conx is selected based on the condition satisfied The resulted quotient is accumulated by the accumulator Total no of clock cycles needed for this module 6 no of sensors 1 The extra clock is used to divide the accumulated result with the number of sensors to calculate the fused confidence Inp mux vhd This module calculates the start base end base start top and end top points of the trapezoid of each sensor It also calculates the height of each of these trapezoids A 40 by 1 8 bit multiplexer is used to select the element of the trap 40 The module contains the multiplexers to select the various points of the trapezoid and the points of the array It takes 2 clock cycles to perform a single selection Memifc vhd The module contains memory access signals and bus address decoders This file was modified in order to use the Virtex BlockRAMs as the memory The MSF process can access 256 words starting at address 1000 10FF MSF Control vhd This module generates timing control signals needed for all datapath circuits It also generates control signals to interface with the system s memory and bus MSF top vhd This module is the top level module of the MSF code It calls all
124. ccurate measurement of the state variables that are used for controlling the performance of the plant It was soon realized that meeting these requirements is not possible in all situations and people tried to design controllers that are robust Many robust control design techniques were developed and these controllers were able to tolerate the variation in the model system parameters and state variable estimations But these controllers achieved them at the cost of performance 15 Adaptive controllers provide an answer to the problem Adaptive controllers are basically a controller whose control law adapts its own behavior as it learns about the process it is designed to control or as the process changes with its environment The field of adaptive control is a very wide and what is presented in this section is a brief introduction to adaptive control It is not intended to be a thorough literature review Adaptive control methods are classified into two broad categories 15 46 1 Indirect or explicit control basic requirement in this method is the availability of a design model but the parameters of the model are not known The plant parameters are estimated explicitly on line and the control parameters are then adjusted based on these estimations Indirect control methods utilize separate parameter identification and control schemes 2 Direct or implicit control This method does not assume the availability of the design model
125. center of each range This distribution is then combined with the distributions of the sensor data and normalized to get the fused Distribution from which the centroid and the confidences are calculated as discussed in earlier sections Figure 3 30 shows the additional linguistic trend information The effect of incorporating the linguistic trend information in the sensor fusion algorithm is illustrated in Figure 3 31 it can be observed that the performance of the fusion algorithm has improved when compared to Figure 3 29 84 Fused Trend with Linguistic Source 9277 Additional Source of Trend Trend of Sensor Data Fused Trend 0 25 10 20 30 40 50 60 70 Instants of Time Figure 3 30 Trends after considering Linguistic Source This increase in reliability of the Fused Trend further improves the calculation of the expected value of the parameter This causes the Fused value to be more reliable Sensor Readings and Fused Reading 800 780 760 740 720 45 WHE 700 680 L Sensor Value Correct Value Fused Value 10 20 40 50 60 Instants of Time 660 Figure 3 31 Sensor Fusion with Linguistic Trend Information 3 5 2 Fusion of Linguistic Information the Measurand Value 85 In this section the effectiveness of having expert system to enhance the fusion process by considering the measurand value Consider a case o
126. confidence is taken to 120 dT ___ 1000 dMR __ 1000 dC __ 1000 d02 __ 100 dBlast _ _ dCMR __ BR MR Figure 4 12 Response for melt rate confidence of 0 9 0 1 pulse for 600 seconds dT 1000 dMR _ 1000 dC 1000 dO2 ___ 100 dBlast dCMR BR MR Figure 4 13 Response for melt rate confidence of 0 5 and 0 1 pulse for 600 seconds 121 4 4 9 VARYING MODEL PARAMETERS A linear model has approximated the non linearity of the cupola furnace The experimental data shows that the model is good only for a narrow operating range This problem could be solved by designing many controllers and then use a look up table to choose the best controller for a certain operating point It would be better if one controller would work under all the normal operating ranges This is one reason for using fuzzy logic control that is it is robust In Figure 4 14 a sine wave disturbance with an offset of 1 125 and amplitude of 0 375 is multiplied to each of the nine transfer functions at different frequencies This varies the steady state transfer function response from 75 to 150 of the original value Since they are varied at different frequencies a worst case combination will align if the simulation runs long enough The non linearities in the actual cupola are not a problem for the fuzzy controller Figure 4 14 shows the results of a widely varying model Applying a sinusoidal varying gain to the nine individual t
127. ct oriented methodology that integrates the different components of the developed system The software package includes algorithms for offline analysis as well as online operation Details regarding the use of software package are provided in Appendix A The technical achievements of the project can be highlighted through the refereed journal and conference publications to interested professionals in the field of sensors and control as well as professionals within the metal casting industry Up to this point fourteen papers and seven theses that were supported by the project have been published The lists of papers and theses are provided in Appendices 1 A and 1 B Some of the technical details were not included in this report to protect the intellectual property of the participants 156 The developed system was tested in a series of demonstration runs These runs have demonstrated the ability of the system to be easily interfaced into an existing cupola foundry with its own data acquisition equipment sensors and networks be adapted to incorporate the available sensors and modalities fuse the available information sources and provide a best estimate as well as a confidence measure on the estimate monitor trends of individual variables as well as combination of variables and provide early warning on potential problems such as bridging that might be developing in the cupola integrate sensing and control algorithms
128. ction 28 With an industry estimate of 60 yield on castings this equates to the direct production 1 204 million tons of carbon generating 4 412 million tons of carbon dioxide per year This amounts to 1 2 of the total annual national production of green house gas 29 The cupola has maintained its competitiveness for several reasons Compared with competing technologies such as arc or induction melting the cupola uses the energy in coal more efficiently because it does not have to go through the intermediate step of producing electricity and the required coke making consumes little energy The combustion products in cupola melting are easily contained another advantage over arc melting The cupola is a relatively simple device that can be made in many sizes to suit the molten metal needs of foundries of various sizes While cupola melting is simple in principle burning coke with an air blast and melting metal the actual physical and chemical details of the process are quite complex and the phenomena occurring in the melt zone are difficult to measure directly 14 because of the aggressive chemical environment that exists inside the cupola Controlling these phenomena is desirable however for efficient energy use for producing iron of acceptable quality and for reducing the environmental impact of the melting process The inevitable random variations in charge composition blast effectiveness and even local meteorological conditions
129. d Existing Variable to a Modality Modify Existing Variable Plant cupola diameter Add New to Standard Grammar wal Sensor Sonos el ameter Delete a Variable from a Modality Monitor Blower Freq Delete a Variable from Standard Grammar Expert pressure drop normalized mass of air in blast amount of moisture in air normalized oxygen addition ir oxygen addition rate O2 Flow Rate blast temperature variable Selected modality Name blast fraction ccc c ns actual BR ratio Planner coke size Variable Description Local Variable Name total oxygen in blast ambient temperature weight of coke divided by the relative humidity weight of the metal in the charge SCR Metric Value Metric Unit nodes in mesh outer loop limit NUNG British value British Unit bottom to top sweeps top to bottom sweeps E insulated number of tuyeres cupola taper zalloysteel Figure 11 Modify Existing Standard Grammar There are three same tabs on these two dialogs On the Modality Options tab the user can create a new modality or delete an existing modality On the Variables Options tab the user can create a new variable or delete an existing variable in a modality On the Save Options the user can exit this function with or without saving the new or modified standard grammar A 1 1 3 Select Modalities amp Variable amp
130. dal membership functions for each input a large positive and negative a small positive and negative and an zero range Since there are two inputs error and change in error with five membership functions each there are twenty five rules relating the inputs and outputs of each fuzzy inference system used in this paper 4 41 Controller design In order to control the system properly some knowledge of the desired response for each output such as MR T and C and limitations on the ranges and possible rates of change of the available inputs such as O2 BR and CMR is required This is important since the controller might request changes in those input parameters that might not be achievable or might cause erroneous response of the cupola A normal settling time for a moderate change in melt rate was selected to be 5 minutes This means that small changes in MR could be achieved within 5 minutes Changes in molten metal temperature were also selected to be achieved in 5 minutes For changes in the C the long time delay associated with changes in the charge composition forces changes in the C to take a longer time period The time selected to achieve changes in C is 50 minutes 106 The process of designing the fuzzy controller is an iterative An initial guess was made to the membership function parameters the output values and the rules so that simulations could be performed Plots of the fuzzy inputs and system outputs were then
131. decreases In other words the region in which the true value could be with respect to the sensor reading becomes wider This could be reflected by scaling up the standard deviation of the Gaussian function used in building the PDF using the self confidences of the sensors Thus the PDF function becomes 59 2 3 4 where SC is the self confidence of the sensor and the rest of the parameters defined as in 3 1 Figure 3 7 and Figure 3 8 illustrate the effect of the change in self confidence over the shape of the Gaussian functions and hence the PDF It should be noted that this change is not used in the standard deviation used for finding the Confidence 0 4 T T T T 0 3 Self Confidence 1 4 E 0 2 4 09900 PDF ese 0 1 te Pd e Nm 4 e ee 9 e 0060046 6 8 10 12 14 16 18 20 22 Figure 3 7 Estimation of Measurand without Considering Self Confidence 61 0 4 0 3 Self Confidence 0 5 4 0 2 4 PDF 0 1 4 J Mais a ne 0 tl 4 6 8 10 12 14 16 18 20 22 Figure 3 8 Estimation of Measurand Considering Self Confidence Median Sensor Self enter GF Validation Std Dev R Estimate Self Confidence D Self Center GF Overall PDF n t Validation Std Dev Cumulative Confidence d a and m gt a Normalized n Medi
132. e Implementation 1 5 1 Self Validation Fuzzy Logic Code An introduction to the concept of fuzzy logic was presented earlier in this report The design and implementation of the fuzzy logic code for Self Validation began with 19 the literature search for fuzzy logic theory and implementations and with the study of the Matlab Fuzzy Logic Toolbox Matlab provides a very good general purpose C code implementation of fuzzy logic that closely resembles their Fuzzy Logic Toolbox Our initial version was derived from the Matlab C code corrected and simplified for our needs keeping only a subset of the membership implication aggregation and defuzzification functions Some of the original code that applied only to its use within Matlab was also eliminated greatly simplifying the code and increasing its speed and reliability Next the floating point code was replaced with fixed point code which is more suitable for hardware implementation in FPGAs The fixed point code uses only the four basic arithmetic functions on fixed point data words of 8 16 24 and 32 bits Both the Mamdani and Sugeno methods were implemented even though the project teams have since decided to use only the Sugeno method We have determined that the less computationally intensive Sugeno method is quite adequate for our application and thus we intend to only use this method in the future At this point the results from both the floating point and fixed point versions of
133. e Standard Deviation from historical data a dialog as shown in Figure A 31 appears for calculating the standard deviation E Standard Deviation Calculate vi Polynomial Order Sensor etse 5 Modality Ending Rec Valid Range T 1200 0 10000 1 1 1 1 1 1 1 1 1 1 1 1 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 6000 6476 Window Figure A 31 Standard Deviation Calculation 4 1 1 6 2 3 Other Specific Setup 190 191 There three more modality setup VIs namely trend monitor setup charge setup and planer setup which are used to setup the specific modalities 4 1 1 6 2 3 1 Trend Monitor Setup The Trend Monitor engine requires some parameters for its operation The Trend Monitor setup vi is programmed for assigning such parameters needed for calculating the trend The dialog of Trend Monitor Setup VI is as shown in Figure A 32 6 Trend Monitor Setup vi Monitor Modality Names Monitor 5 Real Albany sensor melt rate raw value Real Albany sensor metal temperature 1 raw value Assign Defaults Real Albany sensor sensorl raw value Real Albany sensor sensor2 raw value Real Albany sensor sensor3 raw value Plant Model metal temperature 1 rav value Plant Model melt rate raw value Plant Model Final Carbon raw value Set All As Defaults Figure A 32 Interface for setting up trend monitor parameters The interface shows the list of all grou
134. e the Control Option Here 203 A TPSC OFFLINE SYSTEM SER MANUAL e ee ees 208 4 2 1 Single Run 208 4 2 2 Single Variable Correlation 209 4 2 3 Multi Variable Correlation 210 4 2 4 Nominal Multiple Correlation 211 4 2 5 View Single Variable Graphs 212 4 2 6 View Multi Variable Graphs 213 2 7 View N M Correlation Graphs 214 ONLINE 615 216 4 3 1 Online Setup 216 A 3 2 Simulate Data Collection 217 4 3 3 Analyze Collected Data 218 4 3 4 View Results 219
135. ed This chapter starts with the study of a case where the multiple sensor fusion algorithm that was developed in Chapter 3 produces a bad estimate Then a methodology for the controller design to improve the system performance in these situations is developed The stability of the closed loop system with the developed controller is studied using Lyapunov stability theory in the final sections of the chapter A linear plant model is considered while studying the stability of the system with the controller 93 41 Motivation The multiple sensor fusion algorithm is developed based the fact that the correct estimate lies at the highest probability value as determined using the sensors data 0 4 0 3 Actual 0 2 Value 0 1 0 6 Figure 4 1 Wrong Estimate from the Multiple Sensor Fusion In the case shown in Figure 4 1 two sensors read the correct value near 10 while the other two read the wrong value near 14 The estimate from the sensor fusion can be seen to be closer to the wrong value just because the readings from the two sensors that have failed are closer to each other than those of the correct sensors Although the estimate from the sensor fusion algorithm was wrong it can be seen from the figure that the corresponding confidence is low This confidence can thus be used as an important parameter that can be integrated into the controller so as to improve
136. edu I3PSC 21 Industry ndustry Oversight End User ent Technical Development Figure 1 1 Graphical Representation of Project Organization Mohamed Abdelrahman Principal Investigator Jeff Frolic Mohamed Abdelrahman Faculty Investigator oi Faculty Investigator Graduate Research Assistant W Mahmoud Faculty Investigator Graduate Research Assistant Graduate Research Assistant Graduate suche Seeman s ae Research Dtm TTU Group Assistant TTU_Group1 TTU Group3 Figure 1 2 Detailed representation of Project Organization 22 1 4 4 Overall System Vision Figure 3 shows a schematic of the different components developed in this project and how they are tied together for a cupola furnace application The system is generally divided into online and offline components The offline analysis component is aimed at analyzing the data collected during a heat and plan for next heats This analysis is based mainly on cupola models The model currently in use is the model developed by the American Foundry Society AFS However the developed tool can be adapted to accept other models as they become available The online component is aimed at actual analysis and control of the cupola furnace during a heat
137. egrating the sensor fusion in the controller design was presented The stability of the designed controller was rigorously proven Traditional controllers however are more suitable for linear systems with well defined models Although linear models in a specified range of operation can describe cupola furnaces they are in general nonlinear systems with a lot of uncertainty in the inputs In this section an alternative to the previously designed traditional controller is given The alternative design is based on fuzzy control principles Using the pairing of cupola inputs to outputs as CMR C and BR MR three fuzzy controllers were designed The fuzzy controller is composed of a fuzzy inference system and an integrator The fuzzy inference system suggests changes to the inputs in order to achieve the desired changes in the outputs The integrator accumulates these changes and presents it to the cupola The inputs to each fuzzy inference system are 1 the error which represents the difference between the desired output and the current output and 2 the change in the error averaged over a period of time The outputs of the fuzzy inference system of each controller can take on five values large positive LgPos small positive SmPos zero small negative SmNeg or large negative LgNeg These values represent the required change in the 105 input to achieve the desired change in output Each fuzzy inference system contains five trapezoi
138. el files are data files These files are TestInput1 txt and TestOutput1 txt One set of Inputs and Expected Outputs TestInput2 txt and TestOutput2 txt Another set of Inputs and Expected Outputs TestInput3 txt and TestOutput3 txt Another set of Inputs and Expected Outputs The wk1 files can be used to test the m files as the m files read the wk1 files The Hardware Group studied analyzed and validated the MATLAB code for the purpose of understanding its functionality and the enable the development of the hardware implementation code The algorithm is simple and generic for n sensors It reflects the degree of agreement among sensors measure the reliability of the measured estimate and hence minimizes the effect of failed sensors A sensor s median reading that a deviate sharply from the centroid of the n membership functions of the sensors readings is assigned a lower confidence For sampling period the algorithm reads the median filtered temperature and the corresponding confidence of each of the n sensors from the outputs of the self validation algorithm Using the sensor reading the system uses also the historical variance measure of the sensors For each sensor the algorithm calculates the four parameters of a trapezoid membership function such that the area of 33 each of the trapezoid is the same unit trapezoids symmetrical around the sensor reading The heights of all trapezoids are added toge
139. elay plant offset Observing Figure 4 15 a 25 overshoot can be seen due to the 600 second difference between the cupola s and the Smith predictor s time delay Figure 4 16 shows the controller s reaction to a 600 second in the same difference which increases the settling time by 100 These results are good since the C remained within 25 of the required change in set point after a 2000 second settling time A 600 second difference in the time delays would indicate a serious problem and would probably not occur in the actual cupola Therefore this controller is acceptable for an average time delay given for the Smith predictor 124 dT 1000 dMR 1000 dO2 100 dBlast BR CMR MR Figure 4 16 Smith predictor with 600 second plant time delay offset If a function is available for calculating and updating the Smith predictor s time delay then this controller is conservative and should be adjusted to lower the settling time 4 4 11 COMBINING ALL NOISES AND DISTURBANCES Figure 4 17 is a plot of all the above noises and disturbances added in This represents the cupola in a worst case scenario where many disturbances are simultaneously taking place The Smith predictor s time delay is 200 seconds more than the cupola s and the sensor fusion confidence is 50 The figure shows the result of applying all the noises 125 disturbances and varying parameters to the system The controller work
140. ements for sequence of events during execution They are also help to determine how to group functions into hardware blocks and which sequential functions are reusable These block diagrams for the hardware implemented functions are shown in the Appendix 2 5 5 Define Data Structure and Organization The data structure of the FIS parameters for all sensors that are calculated during the set up phase and the intermediate results need to stored on the on the FPGA board s SRAM The data structure and their organization within the SRAM have been defined A register file was also designed to store the commands 2 5 6 Define finite state machine controllers The different states of a Finite State Machine FSM controller were defined The FSM is used to control all the timing details for the operation of the entire design The defining ASM chart is shown in the Appendix 29 2 6 Develop Hardware Design of SV Signal processor 2 6 1 Code Hardware Blocks VHDL Each block of code was coded in VHDL using the data flow graphs to help guide the sequencing of operations All arithmetic operation was done through function calls to the entities of the arithmetic functions library The coding was done in a way that allows pipelining and parallelism of operations 2 6 2 Simulate Each Entity Code Separately Each code was simulated separately to ensure its proper functionality Special debugging codes were inserted to allow monitoring of internal sig
141. en using the centroid would be affected by the faulty sensor Figure 3 4 shows that the peak value does not always correspond to the value on which most sensors agree at all cases 51 0 4 0 3 0 4 0 3 0 2 22 Figure 3 4 Comparison of Estimate with Peak Values 31 3 Confidence in Estimate The estimate value from the previous procedure takes into account the agreement between the sensor data However the estimated value does not explicitly reflect the 58 degree of agreement between the sensors and hence the confidence in the estimated value The agreement between the sensors is reflected in the width of the PDF function estimated according to the process previously described Thus the confidence is calculated using the area of the PDF that is enclosed within three standard deviations on each side of the estimated measurand value In the ideal case where all the sensors agree exactly this will be approximately equal to one As the agreement between the sensors decrease this area will decrease as well This is illustrated in Figure 3 5 and Figure 3 6 The confidence is related to the PDF function width according to the relation Estimate 3o Confidence 59 3 Estimate 30 where o is the maximum standard deviation of the parametric functions used in forming the PDF 0 4 T T T T m Integration Limits 0 3 0 2 0 1
142. en press the update button If the value of the controller output is to be changed the user enters the desired value at the appropriate control The check box charge is used to indicate that a charge is added to the furnace The charge number and time is displayed in the two indicators as marked on top of the indicator The two controls CMR and SCR are used to manually enter the actual coke to metal ratio and steel to cast ration that went in the latest charge added to the furnace The T MR Importance Scalar values determines is used by the controller to determine how to adjust the BR and Oxygen since the desired values for both parameters might not be achievable simultaneously A higher value would favor one parameter over the other The indicators at the bottom of the screen display important parameters regarding the charges The control Charges to Fill Cupola indicates the average number of charges that can fill the cupola The second Tab Figure A 51 displays a set of parameters used by the controller including the desired set points or temperature melt rate and the future SP values for the same parameters as well as Carbon This is important for parameters that require adjustment of charges to avoid the delay resulting from the melting time through the furnace The delay time is estimated approximately as the time it takes to melt the number of charges inside the furnace The set of controls marked with Nominal indicate the starting steady state
143. ening a DOS window selecting the properties button and checking the box marked close on exit This is to avoid having 400 DOS windows open at the end of a correlation A 4 1 1 AFS Setup There are a few parameters that must be defined at setup in order for the AFS model to work This is done using the Additional Model Specific Setup option on the Setup menu A 4 1 1 1 Define AFS File Paths 221 222 Define Afs File ae Figure A 68 Define AFS File Paths Define Afs File Paths vi The Fortran executable file needs to be fully declared at this point The fortran file is called from the DOS prompt therefore the file path needs to be compatible with DOS tules The path to the executable should not have any spaces in it DOS does not handle directory names with spaces in them The AFS model creates a large number of data files the second field defines where they will be written This may or may not be the same directory where the other data files were placed the Excel files and data structure files You may want to choose a different directory to keep the files organized There will be combinations of inputs that cause the numerical model to be non convergent The AFS Model interface waits for the numerical model to finish writing the output files before reading the values if the model hangs so will the interface To avoid this situation the interface times out after a set time limit That time limit
144. enry and D W Clarke The Self Validating sensor Rationale definitions and examples Control Eng Practice vol 1 no 4 pp 585 610 1993 12 T M Tsai and H P Chou Sensor fault detection with the single sensor parity relation Nuclear Science and Engineering vol 114 pp 141 1993 51 13 Mathieu Mercadal Sensor Failure detection using Generalized Parity relations for Flexible Structures Journal of Guidance Control and Dynamics vol 12 no 1 Feb 1989 14 Jeff Frolik C V PhaniShankar and Steve Orth Fuzzy Rules for Automated Sensor Self Validation and Confidence Measure n Proceedings of American Control Conference June 2000 15 Bernard Friedland Advanced Control System Design Prentice Hall Inc New Jersey 1996 16 and B Wittenmark Adaptive Control Addison Wesley Publishing Co Reading MA 1989 17 Liu Hsu Aldayr D de Araujo Ramon R Costa Analysis and design of I O based variable structure adaptive control input output variable structure model reference adaptive control systems JEEE Transactions on Automatic Control vol 39 no 1 pp 4 Jan 1994 18 E Burdet A Codourey Evaluation of parametric and nonparametric nonlinear adaptive controllers Nonlinear controllers Robotica vol 16 no 1 1998 19 Judith Hocherman Frommer Sanjeev R Kulkarni Peter J Ramadge Controller switching based on output prediction errors JEEE Transactions on Automatic C
145. eral graduate students MS benefited from this project through the financial support for their work on the project Sobha Sankaran was a key contributor to the detailed design and implementation of the Self Validation software and hardware Srikala Vadlamani was a key designer of the Multi Sensor Fusion software and hardware under the supervision of Dr Mahmoud Jie Ellen Chen was responsible for implementing the control and communication software on the Host PC to interface with the CPU board for both SV and MSF Each of these students contributed significantly to the development of the PSC hardware system 49 3 10 2 Future Recommendations We recommend that the following tasks be undertaken to fully utilize the results of the current project Complete implementation and testing of the MSF processor on the FPGA board controlled by the CPU board Then combine both processors SV and MSF on the same FPGA running simultaneously Complete the hardware software interface between the CPU Board and the Data Acquisition System DAQ to allow direct autonomous acquisition and processing of the Cupola data without burdening the Host PC Test the completely integrated system with the Cupola and live data 50 REFERENCES 1 Nagrath LJ and Gopal M Control Systems Engineering Second Edition New Age International P Ltd Publishers 1995 2 Maciejowski J M Multivariable Feedback Design Addison Wesley Publishers Ltd 1990
146. ess complicated and easier to compute Specifically the Sugeno constant method was adopted for processing over the linear method as it requires fewer parameters to be specified in the FIS file and it is also much easier to compute with no loss in accuracy 1 4 2 Self Validation Preprocessing The fuzzy logic code was written to accept preprocessed inputs consisting of median filtered temperature variance of temperature and rate of change of temperature However the raw data that is available to us consists of values of time and temperature measured from the sensors Initially the Intelligent Algorithms Team did the preprocessing in an Excel spreadsheet It was then converted to Matlab code see Appendix for an example Matlab M file which we had to convert to C code The paragraphs below discuss our changes and enhancements to the preprocessing algorithm fixed point considerations and changes made in the algorithm to simplify it For reasons previously specified the code was changed to use only fixed point numbers The input signal values which are floating point numbers at this time are converted into fixed point numbers before the preprocessing is carried out The preprocessed values derived from the raw input data time and temperature are the median filtered temperature rate of change of temperature and the variance of the temperature These values are determined using the following formulae 16 Median filtered tem
147. evious algorithm would consider the degree of agreement among the trend sensors This agreement would be used to modify the self confidence of each sensor This is presented in the following section 77 Multiple Sensor Fusion considering Trend 800 r r r r 780 760 740 720 D3 WAED 700 680 L Correct Sensor Erroneous Sensor Fused Temperature 0 10 20 30 40 50 60 70 Instants of Time 660 Figure 3 25 Multiple Sensor Fusion Considering Trend 3 4 4 Confidence based on agreement among the Sensors A calculation is proposed wherein the self confidence measure of each sensor is modified based on the degree of agreement among the trends of the sensors Using the Parzen estimator the fused distribution is estimated and the area enclosed by each of the sensor within three standard deviation from the fused value is calculated The self confidence measure of each sensor is modified using the formula confidence max min SC ae 0 5 max Area TotalArea Where SC Self Confidence of the Sensor Area area enclosed by i sensor in Fused Distribution 78 Total Area area enclosed in three standard deviation around the fused value An arbitrary minimum value of 0 5 was assigned to the sensor confidence In developing the normalized distribution of the fused data the contribution of the erroneous sensor is reduced when compared to the contribution made by othe
148. f Content sa tune ikii 166 AA IPSC ONDINE SYSTEM USER MANUAL ce 169 4 1 1 Setup Application 174 A 1 1 1 Define Standard Grammar 175 A 1 1 2 Create Modality Standard Grammar 177 A 1 1 3 Select Modalities amp Variable amp Interface 178 A 1 1 4 Select Variable Properties 179 A 1 1 5 Save Setup Information 180 A 1 1 6 Modality Setup 180 A 1 1 6 1 Declare Model Setup 181 A 1 1 6 2 Run Setup 181 A 1 1 7 Done Setting 198 4 1 2 F PSC 199 A 1 2 1 Cupola Operation Monitor 200 A 1 2 2 Chang
149. f sensor failure as shown in Figure 3 32 In this case the sensor performs satisfactorily till a particular instant and from then on it has a shift in its readings But the sensor still continues to have the similar trend as indicated by the other linguistic sources for trend As a result of this the algorithm tends to start following erroneous reading provided by the sensor In this section a similar methodology as that of the linguistic source on trend is considered for the parameter value as well The operator provides the algorithm with a linguistic value which as before has a predefined range The ranges for the measurand values could for example be Very low 667 5 692 5 Low 680 715 Normal 692 5 737 5 High 715 760 Very high 750 5 775 These ranges are obtained from the historical data considering the behavior of the parameter Figure 3 33 shows that the performance of algorithm after considering linguistic information on the parameter also 43 WED 700 680 800 760 740 720 86 Failure of Sensor Sensor Reading Fused Reading Actual Measure 20 30 40 50 60 Instants of Time Figure 3 32 Another Case of Sensor Failure Linguistic Information on Parameter Value Sensor Reading Fused Reading mee Actual Measure 4 10 20 30 40 50 7 Instants of Time Figure 3 33 Multi M
150. f sensors using data from the sensor that is being validated This is discussed in the next section 42 2 3 3 Self Validation The technique of validating a sensor using the historical data from that sensor alone is called self validation These self validating sensors are called intelligent sensors and many researches are taking place to create intelligent measurements Yang and Clarke in their section 10 have defined the self validating sensors their rationale and how they can evolve into intelligent measurements Initial research in this area started by considering the invalid data of sensors as noise and hence using filtering techniques for the self validation Kalman filter was found to give good results for self validation A detailed description of self validating sensors was given in Henry and Clarke 11 The research by Tsai and Chou 12 uses the correlation of system dynamics with multistep readings of a sensor Using historical data of the sensors for self validation was used by Mercadal 13 This reference paper uses the historical data to create an analytical model for the sensor depending upon its previous values The actual sensor data are then validated by comparing it with the value predicted from the model The paper 14 develops a fuzzy based self validating algorithm based on the validated historical data of the sensor The algorithm developed in this paper is described in detail as it is implemented and used as a part
151. f the data structure of the system The structure of a standard grammar file is shown in Figure A 8 The following rules must be followed in order to run the system properly 1 The first row bears a descriptive name of the column or of the modality The first six columns are fixed They contain different information Starting from column seven the modalities appear In the example shown in Figure A 8 the modalities are Planner Controller Plant Virtual Sensor Fusion Monitor and Expert The order of the modalities in the standard grammar presents the execution order of the modalities in the system So the modalities should be in a specified order Up to eight modalities can be included in the system 2 The second row has INPUT which is a flag value to determine the start of the input variable in each modality At the end of the list of input variables there is a blank row and a row with OUTPUT which is a flag to determine the beginning of the output variable in each modality 3 The name in the first column is a globe name representing a variable however each modality can have a local variable name that corresponds to the variable in the first column of the standard grammar This local variable name is used in the current implementation to interface the system to the DAQ To define an existing standard grammar file double click on Define Standard Grammar in Setup Menu Figure A 6 The dialog in Figure A 7 will be opened
152. ght line blocks 18 A conditional block includes conditional branches and multiple execution paths Based upon the results of the decisions made in these conditions program control flows through different paths of execution As the average execution time is required a certain weight was associated with each of these blocks The weight associated depends on the probability that a particular path of execution is followed Generally we assumed all branches had equal probability and equal weight A procedure block either a subroutine or a function is called by other statements elsewhere in the code The number of clock cycles taken for a call and return from the procedure block was included in the calculations An assembly language file of the SV C code was generated This was the equivalent assembly language code for every line in the C code The number of clock cycles required by the instructions was then found from the instruction set description in the Pentium Microprocessor data book An MS Excel file was created to document the results see Appendix It shows the number of clocks consumed for the execution of every block defined in the code The total processing time for the code is the sum of the individual block clock cycle counts It was found that each input to the Mamdani style fuzzy logic code required 55 000 clock cycles and each input to the Sugeno style fuzzy logic code required 15 000 clock cycles 1 5 Self Validation Softwar
153. gorithm The remaining interfaces CPU to FPGA and CPU to DAQ will also be developed and tested during this year The effort in 2001 contract year 3 will be centered on refining the signal processing system and demonstrating its operation within a working cupola environment As discussed earlier there have been some failures encountered and numerous bugs have been fixed Most of the development has taken longer than expected due to unexpected problems and complications that have arisen But we are confident in our ability to overcome the problems and feel that we are definitely on the road to successful completion of this project Year 2 Accomplishments 2 Overview During the year 2000 the IPSC Hardware Team completed the hardware design and testing of the self validation algorithm and began working on both the hardware implementation of the multi sensor fusion algorithm and the communication protocols of 24 the ultimate form of the hardware system the next few sections we will present the details of all tasks performed by the hardware group in the second year 2000 2 2 Communication Protocols As stated earlier the final project will have three different communication protocols Currently the group is working on the user PC and the FPGA board communication protocol The work done can be summarized as follows 2 2 1 Develop low level communication protocol A low level communication protocol that controls the commu
154. hat satisfies the Lyapunov criteria for both and is sufficient to prove the stability of the closed loop system where K is given by 4 1 The above proof for stability is based on the assumptions that the state of the system is known exactly This is not always true in the closed loop system Multiple sensor fusion gives a good estimate for the measurand However there exists an uncertainty in the state of the system as illustrated in Figure 4 3 The stability conditions should incorporate this uncertainty in the state of the system Theorem 4 1 is extended to include the uncertainty in the states Theorem 4 2 Let us consider a linear time varying system given by the system equation X t Ax t Bu t 4 12 u t K t x t g x where g x is the uncertainty in the output If the system satisfies 1 The conditions of Theorem 4 1 2 The Lyapunov function V x Px satisfies kx s vq 100 0 4 0 3 0 2 Assume to be True 0 1 Figure 4 3 Uncertainty in the Estimate from Multiple Sensor Fusion Stag amp gt gt 0 4 13 VI20 VY x eD W x is a positive definite function and jis a positive constant 3 The uncertainty is piecewise continous and locally Lipschitz in x and satisfies the following conditions v 4 14 where the bound on 6p and 0 is given by the expressions a 2 PBK
155. he MR is at steady state and an increase in the MR is 107 requested error in the MR is positive and change in the error is zero then according to rule 23 in Appendix 4 A the blast rate should be increased at its maximum rate Another example is if the error in MR is zero but the output is heading toward an overshoot because the rate of change in the error is a large positive value then the change in the BR should either be a small negative according to rule 15 in Appendix 4 A One of the main objectives of this research was to obtain a controller that could be used for any cupola not just the Albany Research Center s experimental cupola in Albany Oregon Therefore equations were developed to reconfigure the fuzzy membership function and output parameters These use the steady state parameters and the limitations on the system inputs to achieve the reconfiguration of the parameters jo jo aridi mrsma mrlg mrsmb mrss mrss mrsmb mrlg Figure 4 5 Membership functions of the error in melt rate eMRate 108 dmrsma pus dmrlg dmrsmb Ta dmrsmb dmrlg Figure 4 6 Membership functions of the change in error for the melt rate deMRate 109 4 4 2 Smith Predictor Due to the nature of the cupola in loading the fuel at the top and burning at the bottom a long time delay exists between changing the CMR and that change affecting the melti
156. he centroid is calculates as the result of dividing the resultant sum by the no of sensors The divider is implemented in the pdiv16 8 module The centroid is executed in parallel with val sen module the sensor data read in the val sen module are added by the accumulator in this module Conf control vhd This module is used to generate control signals for the self confidence process Control vhd This module is a local controller of the datapath of the area module It performs the six comparisons needed to determine the location of the centroid with respect to the trapezoid points of all sensors Div 8 vhd This module is an 8bit 8 bit divider Div 16 8 vhd This module is a 16bit 8 bit divider 39 Div_16_16 vhd This module is 16bit 16 bit divider Div24_16UNS vhd The module performs unsigned division of 24 bit 16 bit A quotient of 8 bits is generated fus_con_top vhd This module is used to calculate the fused confidence which is the area within a span of 3 min STD on either side of the fused value The operation of this module requires 6 clock cycles as follows Clock 1 The 4 points of the trapezoid for a sensor are selected Clock 2a The local controller compares the 4 points with the limits of evaluation start conf and end conf Local controller calculates the internal control signals selfuscon selmul The conf_control module describes this operation Clock 2b The 8bit 8bit multiplier performs multipli
157. ic FL is also used for signal validation using redundant sensors The advantage of using FL is that the strict boundary posed by the numerical sensor data can be replaced with linguistic terms 6 Analytical redundancy is used in situations where physical redundancy is not possible Analytical redundancy is created using a model for the process Neural networks have been used to create the analytical redundancy using historical data of the process 8 Combinations of fuzzy and neural systems have been used to create analytical redundancy for a specific sensor Another type of redundancy is created using inferential sensors The redundancy is obtained from using sensors that measure other variables and the relationship between the variables and the variable of intent Genetic algorithm is used to find empirically the variables best suited for use in the inferential redundancy while neural based fuzzy system is trained to estimate the monitored sensor signal 9 These analytical and inferential redundancy are then treated as physical redundancy and used in validating physical sensor data There are many difficulties in creating physical or analytical redundancy for sensor signals like increased cost complexity in hardware implementation for the sensors and uncertainty in modeling the plant Moreover the reliability of the sensors that are used for the redundant measurement cannot be assumed Hence few researchers started to work on the validation o
158. ine System Top Level Menu Three options are listed in this menu namely Set up application Run and Quit If you need to run for a new application you should double click on Set up application The dialog as shown in Figure A 6 is popped up The procedure of setting up an application will be introduced in section 2 If the user wants to run an existing application double click on Run The dialog Figure A 44 is popped up The running of the system will be introduced in section 3 Double clicking on Quit will exit the system The complete flow chart of using IPSC is given in Figure A 3 Figure 4 shows the procedure of setting up a new application and Figure A 5 shows the procedure for setting up a modality The modalities that represent different system functions such as data acquisition data fusions controller etc are the components to build the system As shown in Figure A 2 each modality contains a set of variables and each variable has an associated set of properties that get calculated by the application This data structure is a 169 170 3 D parallepoid with modalities variable and properties representing of the axis The time represents the 4 axis 0 1 Properties Properties Properties Variables Variables d
159. ineering vol 114 pp 141 1993 142 13 Mathieu Mercadal Sensor Failure detection using Generalized Parity relations for Flexible Structures Journal of Guidance Control and Dynamics vol 12 no 1 Feb 1989 14 Jeff Frolik C V PhaniShankar and Steve Orth Fuzzy Rules for Automated Sensor Self Validation and Confidence Measure Jn Proceedings of American Control Conference June 2000 15 Bernard Friedland Advanced Control System Design Prentice Hall Inc New Jersey 1996 16 KJ Astrom and B Wittenmark Adaptive Control Addison Wesley Publishing Co Reading MA 1989 17 Liu Hsu Aldayr D de Araujo Ramon R Costa Analysis and design of I O based variable structure adaptive control input output variable structure model reference adaptive control systems JEEE Transactions on Automatic Control vol 39 no 1 pp 4 Jan 1994 18 E Burdet A Codourey Evaluation of parametric and nonparametric nonlinear adaptive controllers Nonlinear controllers Robotica vol 16 no 1 1998 143 19 Judith Hocherman Frommer Sanjeev R Kulkarni Peter J Ramadge Controller switching based on output prediction errors IEEE Transactions on Automatic Control vol 43 no 5 pp 596 May 1998 20 Michel Barbeau Froduald Kabanza Richard St Denis A method for the synreport of controllers to handle safety liveness and real time constraints IEEE Transactions on Automatic Control vol 43
160. into Input and Output Variables 183 184 Once Done a dialog to name the created group appears This dialog also allows the parameters of the group to be added and the path of the engine VI to be defined This dialog is as shown in Figure A 19 Change Association vi Group Name MT mean Path of VI that executes this group a C I13psc Fusion Modality Fusion MultipleSensorFusion vi Variables associated with this group Input Nodes a Kalman PT 4 Virtual Sensor 2nd Pyrometer io gooo E m Output Nodes 5 Pyrometer Temperature Figure A 19 Add Group Parameters In this dialog the user can name a group define the path of the engine VI and give the variables associated with that group The detail explanations are as follows 1 While defining the group name for a fusion modality group a suffix is attached to indicate what kind of Standard Deviation STD value will be used estimating the fusion value The format is GroupName max max means using the maxim STD value min means minimum STD value and mean means the average STD value The default mode is mean Each modality can be executed on different computer The computer is identified by IP address In the Path amp IP of VI that executes this group control you can enter the IP address of the computer and the path of that engine VI on that computer
161. is integrated to reflect the total change in the system input The value of the integrator is bounded to prevent integral windup Each integrator output then passes its value to three transfer functions obtained from the transfer function matrix Each system output receives three values one from each integrator after passing through a transfer function These values are added to give the total change in the output Error signals which represent deviation of the actual signal from the desired output are one of the inputs that get fed to the controllers The change in the error signal is averaged over a period of time and is sent to the controller as a second input Figure 4 8 illustrates the layout 112 To other From Operating outputs other point inputs Transfer function 1 Transfer function 2 Transfer function 3 Desired change Figure 4 8 Simulation layout 4 4 6 NOISE DISTURBANCES AND VARYING PARAMETERS The actual system is non linear and has been approximated by a linear system with constant parameters in the transfer matrix Sensors that measure the system outputs are subject to noise and the system inputs may not perform exactly as the controller demands The simulation must take these factors into consideration Noise Gaussian noise was added to the system outputs before sampling for feedback This represents the fact that the sensors do not measure the outputs perfectly For such a harsh
162. is set here and is dependent on the machine processor speed A little experimentation should be done in order to determine the best setting for each individual computer For a 300 MHz PC 150 200 seconds seemed necessary for an 800 MHz PC 100 seconds is sufficient A 4 1 1 2 Charge Selection The AFS model can accept up to 10 charge materials There are numerous conditions that these materials must meet in order for the model to run correctly The following 222 223 procedure is the best that I come up with for allowing the user to vary the charge material The first step is to use the original model interface the AFS interface to define the charges to be used See the documentation with that program for information on how to do this Run the AFS model using the new charge makeup then print the cin 264 file that is created The file that our interface creates must exactly match the other 4 1 2 Metal Selection Option Menu Metal Data GUL vi File Edit Operate Project Windows Help gt Material Selection Options Define New Metal Select Metals Set Metal Mass Set Current Configuration as Default Return to Main Menu Figure A 69 Material Selection Options Metal Data GUI vi 4 4 1 2 1 Create Material Property Files Select the first option to open a screen that creates a new metal This screen has fields for all the variable names that the AFS model needs Use the cin 264 file to fill in the d
163. is the development of a methodology wherein information regarding trend as is fused with the absolute measurements from the sensors Motivation for developing this trend fusion algorithm stems from the fact that the estimates of virtual sensors and models developed for the cupola furnace were found to provide more accurate information on the measurand trend rather than on its value This section is arranged as follows a brief description of the previous work on multiple sensor fusion is presented This is followed by description of the algorithm of fusion based on trend Simulations that illustrate the algorithm and its effectiveness are presented throughout the paper 69 3 4 3 Multiple Sensor Fusion The process of multiple sensor fusion MSF based on Parzen estimator presented in the previous sections 28 provides an algorithm for fusing data from multiple sensors In this methodology no emphasis is given on the trend of the system Based on the measure provided by the sensor the reading is fed to a fuzzy engine 27 The fuzzy engine looks at the median value the rate of change and the variance of the parameter and assigns it a confidence measure The fuzzy engine assigns each sensor self confidence value based on the agreement between current and historical behavior 27 In the MSF algorithm a trapezoidal distribution is constructed around each sensor measurement 30 The spread of the distribution depends on the slef confidence me
164. l be less affected by a sensor failure Some of the algorithms that were studied in Chapter 2 produces an estimate that represents the value that most sensors agree But these techniques do not specify the degree of agreement on the estimate by the sensors Hence a multiple sensor fusion algorithm that reflects the degree of agreement among sensors will be more appropriate 53 In this chapter multiple sensor fusion algorithm that produces measure of the confidence in the estimated value of the measurand is developed The confidence measure reflects the degree of agreement among the sensors First the discussion on redundant sensor fusion algorithm and how a measure of confidence in the estimate is produced are presented Next the integration of the self confidence into the multiple sensor fusion algorithms to mitigate the effect of sensor failure are presented The multiple sensor fusion algorithm is tested using data from an experimental run of a cupola furnace A comparison of the results with that of the averaging method is also discussed A unified framework for multi modal sensor fusion is also presented in this chapter of the report 3 1 Parzen like Methodology for Redundant Sensor Fusion 3 1 1 Description of Parzen Estimator The Parzen estimator is a nonparametric method for estimating probability density functions without making any assumptions about the nature of the distribution 21 Given a set of sensor data the Pa
165. l6d8 2 ee 2 0 8 141648 m k fus conf 16 Ae 16 16 8 divl6 _8out noofsensors nye ACD puc ivl6 8ou 258 Footy 1 D lk azan 24 16 D div24 160out UU squ std b mull6 8 is NOTE Sel6d 8 Inputs from 0 centroid 15 Trap Height 2 Fused Conf 116 st conf 117 Fused Conf P1 ce ae conf lt st_b amp amp en_conf gt en_b 0 Comp ey b 8 fus conf 0 258 454 gt st conf st conf st b 1 st b Comp m 8 A fus _conf 1 255 gt 2 st conf UN Comp Pr ue amp st conf en b 2 stt E E 8 8 st conf st b re gt 0 8 255 fus _ conf 2 divin 24 _ 5 8 8 st conf 3 st conf st t SP or 7 8 153 Jus _ conf 3 8 jg st st t ete 4 t 8 8 st conf cl E 8 102 fus _ conf 4 mulin8 8 gt st st conf t en t Ea Comp AT 8 102 AS _ conf 5 118 Fused Conf P2 st conf en t 8 Cop st t amp amp st conf en b 6 en b 8 en b 20 st conf i div fus 6 st SO k st conf b en PUN Comp E ES _ conf 0 7 A 8 en b EN Comp Ae i 8 aa E Conf en t EM mom 8 en b
166. ller variable then click on OK The parameters setup dialog will popped up as shown in Figure A 43 196 Controller Setup vi Controller Names total o vaen in blas blast rate coke ratio SCR Figure A 42 Controller setup Control Parameters vi Control Parameter total oxygen in blast Control Parameter FIS Path i 24 Plant K 0 00 Plant Tau Seconds Delay Minutes 0 Schimdt Predictor Delay between E and dE minutes jo Sample Time minutes Initial Output of Controller 9 3 0 00 Control Parameter Nominal value 9 0 0000 Control Parameter Lower Limit 3 0 0000 Control Parameter Upper Limit 3 0 00000 Control Parameter Limits Figure A 43 Setup the cotrollel parameters 197 197 198 A 1 1 7 Done Setting Up After setting up all the modalities for the new application double click Done Setting Up will save all the modality setup and close the application setup window Figure 6 The system returns back to the top level menu Figure 1 198 A 1 2 TPSC Running After setting up the new application double click Run on the panel shown in Figure A 1 The dialog of IPSC Running Figure A 44 is popped up I3PSC Runing Modality File Names 3 0 Desired Sample Intervall 2 12 0000 7 Actual Sample Interval 0 0000 Write Data to file GUI Machine Names Figure A 44 PPSC Running
167. low 1 If Temp is low and Rate of Ch is Very P then self confidencel is V low 1 If Temp is low and Rate of Ch is Very N then self confidencel is V low 1 If Temp is high and Rate of Ch is Very then self confidencel is V low 1 If Temp is high and Rate of Ch is Very P then self confidencel is V low 1 If var is High Noise then self confidencel 15 low 1 If var is Constant then self confidencel 15 low 1 Make plots of fis file figure plotmf fis input 1 title Temperature input figure plotmf fis input 2 title Rate of Change input figure plotmf fis input 3 title Variance input figure plotfis fis diary off 69 70 2 Self Validation Code Documents A 2 1 Theoretical Timing Analysis Spreadsheet Sugeno TIMING CALCULATIONS FOR EXEC C Sugeno COMMENT Naming Convention The first alphabet in the block names denote the block type C denotes Conditional block 5 denotes Straight line block denotes Iteration block The functions are listed in the order in which they appear in exec c NOTE fisPrintVariables is not included here as it dosen t include any computations FUNCTION NAME fisTrapezoidMf BLOCK NAME PROBABILTY WEIGHT OF ITERATIONS Clock cycles Weighted Clockcycles Col1 Col2 Col3 Col 1 2 3 51 1 1 1 10 10 C1 1 1 1 13 13 C2 1 1 1 13 13 C3 1 1 1 13 13 1 3 0 33 1 6 1 98 C5 1 3 0 33 1 6 1 98 C6 1 3 0 33 1 57 18 8
168. ltiple Correlation A nominal valued multiple variable correlation refers to the variation of one parameter while all other parameters are held at their default value A correlation in this manner allows the user to see the results of a change in one variable as all the others are held constant There is also a cost advantage to this manner of correlation in that more variables can be selected and the number of iterations per variable can be increased without the exponential relationship number of runs is simply the product of the number of variables and the number of iterations 211 212 Nominal Multi Conelationvi a gt Figure A 58 Nominal Multiple Correlation Nominal Multi Correlation vi A 2 5 View Single Variable Graphs of the correlation applications create tab delimited spreadsheet files They are assigned the file extension xls for easy importation to Microsoft Excel A list of these files in the output directory path declared at setup 1s displayed in the list box To view the contents of a file select the file and press Read New File To view a single variable correlation you must use the View Single Graphs option otherwise the data is not in the correct format The file will still be read but it won t make any sense For the single variable correlation the input variable is displayed and the available outputs are shown in the list box When a variable i
169. mation This allows for the fusion of information from sources other than physical sensors such as virtual sensors models and expert system information Generic algorithms for the integration of sensing and control based on the previously developed algorithms for sensor fusion were developed and implemented The developed generic algorithms for sensor fusion and integration of sensing and control represent advances in basic science The researchers have also presented application specifically for cupola furnaces These results were presented at professional conferences with audience interested in the advancement of melting methods Algorithms for the implementation of the sensor validation and multiple sensor fusion algorithms on hardware were developed simulated and tested A generic data structure and an object oriented based software package were developed for the incorporation of the different algorithms The current package incorporates the following modules o A Data Acquisition modality for interfacing with existing data acquisition system in a cupola or other industrial plant 25 Aplanner modality where a plan for the cupola heat can be developed A sensor fusion modality o A virtual sensors modality for predicting values of some important parameters based on other system measurements controller modality for producing suggested values for the manipulated variables based on the system requirements o A monitoring mod
170. meters and are asymptotically stable with the same Lyapunov function in 4 5 Hence the Lyapunov equations for these systems will be satisfied Equations 4 6 and 4 7 gives the respective Lyapunov equation A BK P P A BK O AA 6 98 A BK 0 4 7 where Q and Q are positive definite matrices For the controller K given by Equation 4 1 the derivative of the Lyapunov function is V 2 x A BK P P A AA 8 Substituting 4 1 in 4 8 gives V x A B aK P P A B aK 1 Writing A aA 1 a A and separating the and K terms gives V x o A BK P P A BK 1 a A BK P P A BK x 4 9 which by 4 6 and 4 7 gives aQi 1 a Q2 4 10 The derivative of the Lyapunov function is hence ox Qx 12 ax Qux 4 11 In expression 4 11 both the terms in RHS are negative since 0 lt lt 1 Hence V is always negative Q E D The closed loop system with the controller designed in section 4 2 can be proved asymptotically stable by the direct application of the above theorem As the confidence in the estimate changes the controller K changes with time But the controller parameters 99 are bounded by K for high speed and for low speed and the intermediate value varies between these bounds Finding a single positive definite matrix P for the Lyapunov function in 4 5 t
171. mpMethod min AggMethod max DefuzzMethod wtaver Input Name Temp Range 0 1500 NumMFs 3 MF 1 low trapmf 0 0 672 95 677 MF2 ideal trapmf 672 96 677 781 9 785 94 MF3 high trapmf 781 9 785 94 1500 1500 Input2 Name Rate of Ch Range 3 5 3 5 NumMFs 3 0 26 0 35 3 5 3 5 MF2 Very 3 5 3 5 0 35 0 26 MF3 Small trapmf 0 35 0 26 0 26 0 35 Input3 Name var Range 0 2025 NumMFs 3 MF 1 Constant trapmf 0 0 0 4 0804 MF2 Normal trapmf 1 0201 4 0804 184 4164 251 2225 MF3 High Noise trapmf 184 4164 251 2225 2652 25 2756 25 Output Name self confidencel Range 0 1 NumMFs 4 MF1 V_low linear 0 0 0 0 1 MF2 low linear 0 0 0 0 5 57 MF3 high linear 0 0 0 0 75 MF4 V_high linear 0 0 0 1 Rules 300 2 D 100 2 1 232 4 1 210 2 1 220 2 1 001 2 1 110 1 1 120 1 1 320 1 1 310 1 1 003 2 1 001 2 1 p Lm v A 1 3 Raw Data Input File RawIn txt 56 118 176 238 295 359 416 475 537 599 659 715 779 837 899 959 1017 1079 1135 1197 1255 1319 1376 1439 1496 1559 779 2346227 779 6359885 779 1388128 776 3767261 771 7933885 769 1202064 765 8402284 761 098934 759 6134492 755 3356674 748 8223216 746 9354711 747 0019339 744 2286262 740 3850099 733 484972 729 9123854 731
172. mputeInputMF Value FisTrapezoidMF FisComputeFiringStrength FisArrayOperation FisMin and FisComputeTSKRuleOutput These procedures compute the membership functions of the input parameters i e convert crisp inputs to fuzzy inputs apply the inputs to the fuzzy rule set aggregate the outputs of each rule and compute the outputs The output of each preprocessed set represents the computed confidence in the corresponding sensor measurement 2 5 2 Separate constants from true Variables An analysis study of all the code s variables showed that some of the variables are calculated or defined in the setup part and their values do not change during the execution part Hence they will be implemented once in the setup not during the execution in hardware 2 5 3 Simplify the fuzzy logic procedures To simplify the fuzzy logic code for the hardware implementation variables that do not change during the execution were converted into constants In addition all error 28 checking Mamdani code has been removed from the fuzzy logic code The error checking code was not needed if the Fuzzy Inference Structure FIS file which is the input file to the setup part is correct The Mamdani code was not needed since we choose to implement Sugeno type code After the verification of the code revision it was tested successfully 2 5 4 Create Block Diagrams A block diagram for each function was created They are used to determine the requir
173. n iron Moreover other parameters such as CMR and SCR were manually monitored and calculated The demonstration plans aimed at addressing the following questions 1 Can PPSC system be successfully interfaced and integrated into an existing cupola with its own instrumentation and data acquisition system with minimal effort 2 Can the IPSC system provide reliable information regarding the cupola parameters and state of operation 3 IPSC system be used to integrate sensing and control algorithms in order to provide an automatic control system that can successfully operate cupola furnaces in order to avoid some of the problems that currently occur in cupola foundries 131 The last question is the most important as its success automatically indicates positive answer to the first two questions The PPSC system was interfaced through an Ethernet network connection to the existing Data Acquisition Computer at Albany Research Center ALRCDAQ A special software module was written to specifically exchange the important cupola parameters between the IPSC computer and the ALRCDAQ This software module is what needs to be customized if the PSC system is to be used at a different facility This arrangement ensured that no changes to the ALRCDAQ were required and that any changes to the number of parameters monitored can be done quickly This arrangement is illustrated in Figure 5 1 132 Cupola Cupola Computer Master
174. n BW_n 15 0 RO 2 0 RFAdReg 1 0 RFtoBus RFtoProc aN 16 CmdReg LD CLK LdCmd SV FPGA Controller 84 Sensor 8 8 CMD IncMAC 2 Controller 3 0 FSM CLK 4 Inc CLR CNT CNT Counter CLK Q CLK Delay Reg Q 4 SelRuleOut 3 0 16 SelRule 3 0 Busy PBD 15 0 StatToPBD CcC 7 0 6bit Memory Access Counter CLK gt PBRF PBW n 2 0 ConffoPBD PB RF Ifc m PMR n PMW PMCe n lfc LdIMFPOa LdIMFP2a LdiMFPor 934 a gt LdIMFP2b LdSugo LdSug2 em gt LdiV Load from RF I LdimfVal00 01 02 10 11 12 20 21 22 LdFS LdFSn LdTw LdTwf LdConf ClrTw ClrTwF Load Internals SetPB 5 CLK Q gt SellmfpB R IMFval00 A00 IMFvalOO A01 IMFval00 A02 IMFval00 A10 8 IMFval00 A11 g IMFval00 12 IMFval00 20 8 IMFvalOO A21 8 IMFvalOO A22 8 255 CompFire HardWired Signals fisComputeFiringStrength mfsRule00 B00 mfsRule01 B01 mfsRule02 B02 mfsRule10 B10 mfsRule11 B11 mfsRule12 B12 mfsRule20 B20 mfsRule21 B21 mfsRule22 B22 mfsRule30 B30 mfsRule31 B31 mfsRule32 B32 mfsRule40 B40 mfsRule41 B41 mfsRule42 B42 mfsRule50 B50 mfsRule
175. nal Processing System Hardware its interfaces Signal Processing System Hardware Sensor Data Commands amp Controls Cupola Sensors amp Controllers Over the next two years the Hardware Team will complete the project s hardware implementation The Self Validation algorithm will be implemented and tested in programmable logic The Sensor Fusion algorithm will be analyzed re written and implemented in programmable logic The communication interfaces to the FPGA board DAQ board will be designed and tested thus completing the entire hardware software system 12 Literature Search The Hardware Team conducted a lengthy literature search in different software and hardware areas The project requires knowledge in many software and hardware areas some of which are outside the Hardware Team areas of expertise In order to prepare us for the hardware implementation of the software algorithms the hardware group members needed to educate themselves about fuzzy logic neural networks and data acquisition This educational process included studying and understanding the basic definitions terminology and some of the theories and algorithms in the fuzzy logic and neural networks areas The features of the two basic types of fuzzy logic Mamdani and Sugeno were studied The advantages and disadvantages of various neural networks implementation techniques their learning processes and their operation were also
176. nal advisory board for the project under the direction of the AFS This board had representatives from foundries and industrial control companies and will serve to assess the progress of the project and the achievement of its goals Collaborators arranged several meetings with the advisory board over the period of the project These meeting were coordinated with meetings of the AFS cupola steering committee The purposes of these meetings were to review the status of the completion of the project exchange ideas among collaborators and external advisory board and to inform the industry about the benefits of the technology and its potential advantages 20 1 4 3 Coordination of Teams Efforts Coordinating the efforts among the teams working on the project was the responsibility of the principal investigator This coordination was achieved through continuous communication through Use of Email and Phone as needed to address individual teams concerns problems or achievements Emails could be addressed to a specific team leader or to the PI Conference calls were scheduled as needed among TTU investigators and investigators from USU and INEEL to discuss the progress and coordinate the efforts The meetings with the advisory boards were used to have technical meeting among the technical developers at TTU USU INEEL and ALRC A web site and ftp sites were developed where technical materials were exchanged among the collaborators www ece tntech
177. nals during simulation After fixing all the code problems the blocks were synthesized to check the amount of logic blocks needed for their implementations The number of clock cycles needed for each block was recorded The arithmetic functions were modified to the exact number of bits needed in the implementation 2 6 3 Design of the system interfaces The system interfaces include a bus interface a memory interface and a register file The FPGA bus interface represents the connection between the PC and the FPGA boards connected to the PC 104 bus Different signals are used to select one of the FPGA boards and to communicate read from or write to with either the processes on the FPGA boards or with the on board memory The memory interface connects the on board memory with the FPGA processes or the host machine 30 2 6 4 Design of the FSM The design of an efficient FSM requires intimate knowledge of the sequence of operations and the number of clock cycles needed for each operation It coordinates the use of both the address and data buses and memories by the FPGA processes and the host machine and controls all the system s activities The controller also issues various load store and Read Write enable signals 2 6 5 Add All Blocks to Top Level VHDL Entity After successfully testing all the signal validation system components they were integrated with the top level VHDL entity The top level entity acts as the main function in a C c
178. ndow you select the diagnostic input nodes using ADD REMOVE can delete the parameter you ve selected In the Name box you can name your diagnostic This name will be display on the upper right corner of the Cupola Operation Monitor with the diagnostic alarm led 202 203 gt Set 4parameters monitor vi Alarm For Bridging Virtual Sensor Kalman PT rav Plant Final Carbon raw value app 1 REMOVE ok Figure A 49 Setup Diagnostic Parameters A 1 2 2 Change the Control Option Here If a controller is set in the system the dialog of Change the Control Operation Here 15 popped up during the running and it will stay on the desktop So you can change the control operation at any time There are four tabs on the Change the Control Option Here dialog gt Change the Control Option Here li ey gt lt File Edit Operate Tools Browse Window Help Other Charges ro Fill Cupola Charges to Average t 15 gi 5 00 Charging Simulation Figure A 50 Change Controller Option Outputs of Controller 203 204 The first tab is as shown in Figure A 50 It displays information regarding the controller output It also allows the user to override the controller suggestions by choosing to operate in the manual mode To change the setting from Manual to Automatic or the reverse The user presses the Change button then selects the desired setting th
179. ne connecting the CPU board to the host PC and the associated communication software Using standard PC type RS232 serial ports it is intended to run at the maximum standard speed of 115200 bits per second The host PC is the user s interface to this entire signal processing system The purpose of the CPU to host interface is to transfer control commands from the host PC to the CPU board for execution and to transfer status and computed results back from the CPU board to the host PC To make this interface robust a special error detecting correcting communication protocol was designed based on a combination of pre defined message blocks character counts checksums handshaking and retransmission requests 22 A preliminary version of this communication software has been written code was written initially to implement the complete communication protocol But then changes were made in the code to enable the receiver to wait for the sender or vice versa for an infinite length of time without timing out The purpose of including this change was to temporarily avoid error reports due to certain error conditions The testing was carried out first between 2 PCs and then between a PC and the CPU board Only limited success could be achieved regarding the implementation of the protocol It was observed that the code properly supported serial communication only up to a bit rate of 9600 bits per second The expected bit rate of 115200 could n
180. ng FPGA to implement sensor fusion algorithms project has supported the development of basic science in the form of publications in professional refereed journals and conferences as well as practical and applied science with reference to cupola furnace as evidenced by demonstration using cupola furnace data and models and actual demo plans on a research cupola This report is divided into two sections Section 1 describes a subset of the developed algorithms and results of demonstration runs Chapter 1 summarizes the project organization objectives and results of the project Chapters 3 and 4 provide a description of the basic algorithms that were developed for sensor fusion and control More information can be found in the published work listed in Appendices 1 A and 1 B A brief description of the developed software package is provided in the form of a user manual in Appendix A of Section One Section 2 describes the hardware implementation Chapters 1 3 describe work accomplished during project years Chapter 4 gives a summary of the 27 work accomplished and future recommendations The section ends with Appendices that describes more details of the algorithms hardware implementation 28 1 List of Publications Supported by Project Jeff Frolik Mohamed Abdelrahman and Param Kandasamy A confidence Based Approach to the self validation fusion and reconstruction of quasi redundant sensors IEEE Tra
181. ng input parameters that do not correspond directly to commonly used industry terms A preprocessor was written that converts the common industry terms a list is included below to the eccentric terms required by the AFS 225 226 model If both variables are chosen for inclusion in the database in a correlation the user 1s asked to choose which variable to use as the input value If this proves to be a hindrance perhaps a more elegant solution can be reached The figure shows the option screen pressing the button by the desired variable selects that variable and the corresponding value is passed along as an input Fix Variable Conflicts vi Figure A 73 Conflicting Variable Resolution Fix Variable Conflicts vi The list of conflicting variables is AFS Model Variable Common Industry Variable normalized mass of air in blast blast rate normalized oxygen addition in blast oxygen addition rate coke in charge coke ratio amount of moisture in air relative humidity The AFS Model creates a large list of output files most of them are not relevant to the model interrogator so the majority are deleted at the end of the run A 5 Real Sensors Interface The Real Sensors modality does not have a set interface It is intended to be used during online analysis to map data in the sensor text file produced at the Albany cupola to the correct location within the standard grammar The standard grammar is then sent to both
182. ng zone In order to accommodate that time delay a predictive model based strategy is utilized This was suggested first by Smith Smith 1957 and thus the method is referred to as a Smith predictor The Smith predictor is utilized with the C controller The schematic diagram of the system with the Smith predictor is shown in Figure 4 7 As shown in this figure an estimate of the input to the cupola is fed to a duplicate transfer function relating the CMR to the C The signal from the duplicate transfer function has a time delay applied to it and then is subtracted from the C output signal The resulting signal is then added to the undelayed signal from the duplicate transfer function The final result is the feedback signal for the system The time delay in the Smith predictor is representative of the best known or average delay measured anil Transfer ontro function 1 Integrato Signal 8 varies To other Hom outputs other inputs Transfer function 2 varies Transfer function Transfer function 3 varies Feedback Signal Figure 4 7 Implementing a Smith predictor 110 4 4 3 Integration of Sensor Fusion in Controller Design The harsh environment at the output of the cupola results in sensor drift and failure By using several sensors and the sensor fusion technique developed earlier a confidence between zero and one can be determined that reflects the accuracy of the output value It is desirable to use the
183. nication between the User PC and the CPU Board has been designed According to this communication protocol all information passes between the User PC and the signal processor in the form of messages For generality and versatility a single message consists of 0 or more bytes of information data preceded and followed by certain message control bytes Byte 1 usually represents the function code which defines the message class and meaning of data Byte 2 and 3 specify the length of the data block which could be 1 255 or 1 65535 The information following these bytes in finally followed by the error detection byte which is the checksum of all other bytes 2 2 2 Write and test the low level communication code for initialization transmit and receive The developed low level communication code was written in C and was tested between two CPUs The system was initialized to operate at a particular baud rate and then it was used to transmit and receive a string of data The system was tested successfully for various baud rates including the operation baud rate of 115200bps 25 2 2 3 Define specifications for high level communication protocol details for MS SV program for User Interface Following the low level communication protocol a high level communication protocol was designed This protocol specifies the format of the messages that are exchanged between the user PC and the CPU board The sequence of the messages has also been decided upon The
184. nsaction on Instrumentation and Measurement Vol 50 No 6 December 2001 Mohamed Abdelrahman and Param Kandasamy Integration of Multiple Sensor Fusion In Controller Design Accepted for Publication in the Transactions of Instrumentation Society of America 2002 Mike Baswell and Mohamed Abdelrahman Fuzzy Control Of A Cupola Iron Melting Furnace To Appear in Transactions of American Foundry Society 2003 Mohamed Abdelrahman and Param Kandasamy Integration Of Intelligent Industrial Process Sensing and Control for Cupola Iron Melting Furnace in proceedings of the 7 Mechatronic Forum International Conference Atlanta 6 8 September Atlanta GA 2000 10 11 29 Mike Baswell and Mohamed Abdelrahman Fuzzy Control Of Cupola Iron Melting Furnace AFS Congress Kansas City MO May 2002 Min Luo and Mohamed Abdelrahman Wavelet Based Sensor Fusion for Data with Different Sampling Rates in Proceedings of American Control Conference Washington D C June 2001 Mohamed Abdelrahman et al Methodology For Fusion Of Redundant Sensors in Proceedings of American Control Conference Chicago IL June 2000 Jeff Frolik and Mohamed Abdelrahman Synthesis of quasi redundant sensor data a probabilistic approach in Proceedings of American Control Conference Chicago IL June 2000 Steve Orth Jeff Frolik and Mohamed Abdelrahman Fuzzy rules for automated sensor self valid
185. nstants k are Amin P k2 Amax P This satisfies the condition in Theorem 5 1 of 24 lt min T fen Thus all the conditions for the Theorem 5 1 in Error Reference source not found are satisfied Applying the theorem and corollary 5 3 in Error Reference source not found completes the proof Q E D This theorem can be applied to the closed loop system developed As mentioned before the system satisfies all the conditions of the Theorem 4 1 Thus all the conditions for the Theorem 4 2 are satisfied The closed loop with the controller designed as specified in Equation 4 1 is stable within the Ball B The asymptotic stability of the closed loop system with respect to the origin cannot be specified Thus the confidence parameter from the multiple sensor fusion algorithm presented in Chapter 3 is integrated into the controller design and the closed loop system with such a controller is proved to be asymptotically stable though not to the origin but to a ball of radius r The confidence from the multiple sensor fusion is integrated into the controller to prevent the degradation of the system s performance when the multiple 104 sensor fusion fails The next chapter deals with the implementation of this theory in a multi variable feedback control system and simulates the performance of the system 4 4 Fuzzy Controller In the previous sections a traditional controller was designed and a procedure for int
186. ntrol system is shown in Figure 2 1 Reference Input Error Output Controller Plant Detector Feedback Element Figure 2 1 Schematic Diagram of a Feedback Control System 34 2 1 Motivation Sensors are used to measure and feedback output data in feedback control systems The feedback data are used to decide the necessary control action The performance of a feedback control system depends heavily on the reliability of the sensors readings There are different reasons why the sensor data may not be reliable These reasons include 1 Sensors may be prone to high levels of noise and disturbances during measurement and transmission of the data 2 Sensors characteristics may vary with changes in environmental parameters such as the temperature humidity or due to aging 3 Accurate measurement of some variables may not be possible due to the physical nature of the process and 4 Failure of electronic circuitry of the sensor There are several methods available to increase the reliability of process measurements using redundant sensors The redundancy may be achieved through physical sensors analytical sensors or inferential sensors Analytical sensors depend on a model of the physical process to estimate the value of the intended system parameter Inferential sensors utilize other output variables to infer estimates for different variables Techniques such as signal validation and multiple sensor fusion
187. o two sections for our convenience specifically referred to as the fuzzy logic and preprocessing portions Work is proceeding on evaluating the execution timing of the Self Validation algorithms to assist us in choosing the optimal functions for hardware implementation versus software implementation The Self Validation algorithms were supplied to the Hardware Team in the form of high level Matlab and Excel code We decided to first implement them in faster lower level C language code mostly using fixed point arithmetic and then convert portions to even faster fixed point hardware implementations in programmable logic on the FPGA boards After some optimizations and debugging effort the C coded versions of the fuzzy logic and preprocessing portions were successfully verified against the original Matlab and Excel results The CPU board will eventually have to communicate in three ways 1 To the host computer for the user s interface to the system 2 To the DAQ board for data acquisition from the sensors and 3 To the FPGA boards for signal processing computations The CPU to Host interface consists of an error detecting correcting serial communication protocol and its implementation as C code executing on the CPU board and on the host computer The protocol has been specified and a limited version of the C code has been written and verified Work on the other two communication interfaces has not yet begun The following figure shows the Sig
188. odal Sensor Fusion with Linguistic Sources 87 It can be observed that the algorithm with this linguistic source of information on the parameter value provides a very reliable fused value 88 3 6 Wavelet Based Sensor Fusion for Data having Different Sampling Rates 3 6 Introduction Data obtained from numerous sensors can be used to provide more reliable evaluation of physical data than a single sensor In many industrial settings several sources of data regarding a certain parameter may be collected However information from these sources may not always be available at the same points in time due to physical limitations In a cupola furnace for example the temperature of the molten iron is measured both using a thermocouple and a pyrometer The thermocouple TC measurements are made on a physical sample extracted from the furnace output and thus performed relatively infrequently as compared to the near continuous collection of pyrometer data However the pyrometer data is considered to be a less reliable measure and is susceptible to gross corruption Figure 1 shows a sample of cupola temperature data obtained using a pyrometer Pyro Temp and a thermocouple Bath Temp For this data the sampling ratio between pyrometer and TC data is approximately 16 1 Our objective in this fusion algorithm is to provide a generic methodology to fuse two data sources with different time resolution as motivated by this previous example In this wo
189. ode The entire hardware code was simulated and synthesized successfully A bit map file representing the hardware design was then generated 2 6 6 Download Test and Debug Top Level SV Signal Processor The C code was modified clean up the preprocessing steps and to include functions that download the hardware design onto the FPGA to load the system parameters onto memories on the FPGA boards to read from the on board memories to send data to the FPGA processes and read their results Some of the signal validation code was replaced with FPGA communication tasks These tasks include send data send command and get result Self Confidence The modifications were done in many stages In the early stages A C code was developed to allow testing the on board memories by writing data onto them and then reading their contents Currently the system uses 64 31 bytes of on board memory sensor and requires about 50 cycles to process each preprocessed set of inputs 2 7 Develop Multi sensor SV Algorithm The Self validation system was originally tested for data acquired from a single sensor However in the Cupola system more than a single sensor s data would be actually fed into the self validation system So the single sensor self validation system was expanded to be a multi sensor self validation system For the testing purpose the multi sensor self validation system accepts the input values of several sensors as a single file where
190. ode was enhanced through the following modifications 34 calculation of the height sensors trapezoids has been optimized This version assumed that the number of fused sensors data is three constant All the static values like number of sensors number of samples std_deviation were made as dynamic values that are read from an input file The entire code has been optimized by removing redundant calculations All needed value were calculated once at the beginning of the code execution and used as many times as needed The third version was enhancements included replacing all data structures to arrays for faster computations The code for three sensors was also modified to limit the number of points considered to 13 4 n 1 instead of the original 86 points This has resulted in a significant saving of the computational requirements of the code Each of these versions was coded and tested The results obtained form the last version have a maximum error of 2 which is acceptable for our application 2 10 Develop the MSF fixed point code As explained before the floating point code is not suitable for the hardware implementation In the final MSF version the code was implemented as a fixed point code of 8 16 24 and 32 bit resolutions The code was also optimized such that the calculation of any variable is done only once However to reduce the accumulation of round off errors some intermediate variables were calculates
191. ofs for stability Their tolerance to large parameter variations has made them more suitable for many industrial applications 2 5 Conclusions Most of the literatures in multiple sensor fusion exist for detection purposes and are developed for target or enemy detection in military based research Few literatures are available on fusing redundant sensors for non military applications These are commonly based on averaging the redundant data Kalman and Bayesian methods are based on probability density function PDF Kalman filtering technique however needs a good model of the system which is not always available The Bayesian method considers two data points at a time for a confidence measure and also involves lot of matrix manipulations The approximate agreement approach explains the advantages of finding agreement between the sensors One other factor that is required is the degree of agreement between the sensors on the estimate value There was no literature discussing an algorithm to find such a measure Adaptive control methods are available to improve performance The controller parameters are adapted based on the system parameter variation environment changes and even with performance However not much of research exists in the area of adapting the controllers based upon the sensor reliability 50 This chapter discussed some of the self validation techniques multiple sensor fusion algorithms and adaptive control approaches The
192. ogy especially under current economic conditions Certain issues need to be considered towards achieving acceptance of the system in cupola foundries Although the IPSC system was designed to be generic certain modalities such as virtual sensors and the automatic controller need to be setup to address the specific needs of a foundry and thus would require the investment of time and resources Sensors for monitoring of key parameters in a foundry such as temperature and chemical composition have to be installed and operated if not already available A training period for personnel in the foundry would be necessary The investigators have used and continue to use professional meeting and personal contacts to increase awareness of the cupola foundries to the benefits of the TPSC and the possible economic and environmental impact of its utilization Avenues for support of the first industrial implementation of a cupola foundry using private as well as government funds are currently being explored From a different perspective 8 was intended to be generic and applicable to other applications that require the integration of sensing and control Thus another 159 avenue to pursue is to seek funding for the adaptation of the developed system in other applications within the scope of the industries of the future 160 161 REFERENCES 1 Nagrath LJ and Gopal Control Systems Engineering Second Edition New Age Internati
193. ology for Sensor Fusion using 74 Figure 3 22 Distributions of Trends and Fused Trend sess 75 Figure 3 23 Distribution of Temperatures at the previous 1 76 Figure 3 24 Final Distribution of Temperature 76 Figure 3 25 Multiple Sensor Fusion Considering Trend ses 77 Figure 3 26 Distributions of Trend after accounting for agreement between sensor trends PUN 78 Figure 3 27 Multiple Sensor Fusion after 79 Figure 3 27 Fused Contidence PIobs eet eed iem der 80 27 Fail re of d DeBsOE iios secu onerata ii s cedula Tei 82 Figure 3 27 Trends after considering Linguistic Source 2 2 84 Figure 3 27 Sensor Fusion with Linguistic Trend 84 Figure 5 27 Another Case of Sensor FallUtG ee essa cit eee Is pei 86 Figure 3 27 Multi Modal Sensor Fusion with Linguistic Sources esses 86 Figure 3 27 Cupola temperature datas oae eds 89 Figure 3 28 Low sampling rate signal X n aii esee oae eo 90 Figure 4 1 Wrong Estimate from the Multiple Sensor Fusion 93 Figure 4 2 Schematic Diagram of the System with Sensor Fusion Integrated with
194. on during Run 883 146 Figure 5 21 Detection of Bridging in Cupola Changes in Exit 147 Figure 5 22 Detection of Bridging in the Cupola Changes in Cupola 148 Figure 5 23 Opening the Tap hole at ALRC Cupola 149 Figure 5 24 Cupola Always Provides Operational 149 Figure 5 25 An Overview of ALRC Research 150 Figure 5 26 Manual Sampling and Quick Analysis of Molten 151 Figure 5 27 Manual Measurement of Temperature of Molten Iron 151 Figure 5 28 Optical Pyrometers for Continuous Measurements of Iron Temperature 151 Figure 5 29 A Dip Thermocouple for Continuous Temperature Measurement 152 Figure 5 30 Charging Deck of the Cupola at ALRC see 152 Figure 5 31 Measurement of Melt rate Chemical Composition and Temperature 153 Figure 5 32 Remote Monitoring and Control of the Cupola during Demo Runs 153 1 13 Chapter 1 11 Introduction The cupola furnace is used by the iron foundry industry to melt scrap steel cast iron and alloying materials into a consistent grade of iron for casting purposes There are approximately 400 cupolas in the United States which accounts for 70 of cast iron produ
195. onal P Ltd Publishers 1995 2 Maciejowski J M Multivariable Feedback Design Addison Wesley Publishers Ltd 1990 3 Richard R Brooks and S S Iyengar Multi Sensor Fusion Fundamentals and Applications with Software Prentice Hall Inc New Jersey 1998 4 Ren C Luo and Michael G Kay Multiple Integration and Fusion in Intelligent Systems JEEE Transactions on Systems Man and Cybernetics vol 19 no 5 September 1989 5 R C Luo M Lin and R S Scherp Dynamic multi sensor data fusion system for intelligent robots JEEE Journal Robotics and Automation vol RA 4 no 4 pp 385 396 1988 6 Keith E Holbert A Sharif Heger and Nahrul K Alang Rashid Redundant Sensor Validation by Using Fuzzy Logic Nuclear Science and Engineering vol 118 pp 54 64 1994 7 Asok Ray and Rogelio Luck An Introduction to sensor Signal Validation in Redundant Measurement Systems JEEE Control Systems Magazine vol 11 no 2 pp 43 Feb 01 1991 162 8 Marcello R Napolitano Charles Neppach Van Casdorph Steve Naylor Mario Innocenti and Giovanni Silvestri Neural Network Based Scheme for Sensor Failure Detection Identification and Accomodation Journal of Guidance Control and Dynamics vol 18 no 6 Dec 1995 9 Mohamed Abdelrahman and Senthil Subramaniam An Intelligent Signal Validation System for Cupola Furnace Part 1 and Part 2 American Control Conference San Diego 1999 10 Janice C Yang and David
196. ontrol vol 43 no 5 pp 596 May 1998 20 Michel Barbeau Froduald Kabanza Richard St Denis A method for the synreport of controllers to handle safety liveness and real time constraints IEEE Transactions on Automatic Control vol 43 no 11 pp 1543 November 1998 21 Specht D F Probabilistic Neural Networks Neural Networks November 1990 22 Ronald R Yager and Dimitar P Filev Essentials of Fuzzy Modeling and Control John Wiley amp Sons 1994 23 Jeff Frolik and Mohamed Abdelrahman Synreport of Quasi Redundant sensor Data A Probabilistic Approach n Proceedings of American Control Conference 2000 24 Hassan K Khalil Nonlinear Systems Second edition Prentice Hall Inc 1996 25 Mohamed Abdelrahman Kevin Moore Eric Larsen Denis Clark and Paul King Experimental Control of a Cupola Furnace n Proceedings of American Control Conference 1998 52 26 Pascal Gahinet Arkadi Nemiroviski Alan Laub and Mahmoud Chilali LMI Control toolbox 1 0 The Math Works Inc 27 Jeff Frolik C V Phanishankar and Steve Orth Fuzzy Rules for Automated Sensor Self Validation and Confidence Measure Proc of American Control Conference 2000 pp 2912 2916 28 Mohamed Abdelrahman Parameshwaran Kandasamy and Jeff Frolik A Methodology for the Fusion of Redundant Sensors Proc of American Control Conference 2000 pp 2917 2922 29 Jeff Frolik and Mohamed Abdelrahman Synthesis
197. operation values The controls marked with scalar are adjustable parameters that can be changed if the response of the controller is not satisfactory The user is given the option of overriding the confidence values calculated by the senor fusion and supplying a constant value The is done using the switch marker Assign Conf Calc Conf Finally the user can override reading coming from the fusion modality and pass to the controller another set of readings by using the Manual Readings selector 204 205 2700 00 gt 2700 00 2000 00 2700 00 2000 00 275 00 025 200 2000 2600 00 2200 00 6 Figure A 51 Change Control Option Set Points The third Tab Figure 3 9 is used by the user to set minimum and maximum values for different parameters in the system The array marked with K is used to supply the expected steady state gain matrix of the cupola This is calculated experimentally or using the AFS model The vector array marked Tau supplies the time constant that can be used with the matrix K to form a dynamic linear model for predicting cupola response The controls marked with update are used to indicated how often the controller updates the corresponding parameter The number supplied is given in terms of the number of samples The Confidence Effect Control set of parameters are used to increase sensitivity to the confidence values within
198. ort is to develop a methodology to prevent the performance degradation of an automatic control system due to unreliable sensor data The suggested solution to the problem is twofold 1 The development of a multiple sensor fusion algorithm that can produce a best estimate and reliability measure for the estimate of the sensor data 2 The development of a controller structure which utilizes the estimate and the reliability measure to change its performance so as to prevent costly mistakes The methodology developed should reduce the sensitivity of the system to the sensor data when the reliability of the sensor data is found to be low This is achieved by changing the controller s dependability on the sensor signal according to the reliability measure from the multiple sensor fusion A block diagram of the feedback control system to be developed is shown in Figure 2 3 It resembles Figure 2 2 but for the additional flow of information the confidence from the multiple sensor fusion block to the controller 37 Analytical Sensors Best Estimate Multiple Inferential Sensors Sensor Redundant Fusion lt Physical Confidence Sensors c Reference Error Controller Plant Inferential Input Detector Sensor Plant Analytical Model Sensor Figure 2 3 Block Diagram of Proposed System The problem considered in
199. oses a methodology that acquires linguistic information from an expert system and converts it to numerical form that can be fused along with the other numerical information sources on trend Consider a parameter being monitored by a single sensor that gives information about its value Considering that the sensor fails at some particular instant as shown in Figure 3 29 We assume an additional source of trend information The result of the sensor fusion presented earlier is shown in the same figure It partially corrects the faulty sensor readings but the performance is still not satisfactory Failure of a Sensor r 700 Acutal Value Sensor Reading Fused Value 10 20 30 40 50 60 70 Instants of Time 680 0 Figure 3 29 Failure of a Sensor 83 Considering that we linguistic source of information on trend Each linguistic variable has a pre defined range of measurand trend values These ranges are defined based on the behavior of measurand An example of a set of ranges defined for the trend of a measurand could be Sharply decreasing 0 1752 0 0584 Decreasing 0 1168 0 Steady 0 0584 0 0584 Increasing 0 0 1168 Sharply increasing 0 0584 0 1752 So the linguistic information provided by the operator or expert system is converted to give the operating range of operation of the trend A Triangular distribution is constructed around this range with the peak at the
200. ot be achieved More work will be done to improve this code to reach full speed using the complete protocol 17 Summary The Hardware Team has successfully begun development of the Signal Processing System We completed the literature search developed the system architecture and purchased a CPU board DAQ board and multiple FPGA boards to build the system We analyzed and refined the first of the system algorithms created by the Intelligent Algorithms Team to enhance its suitability for hardware implementation This Self Validation algorithm consisting of two major functions preprocessing and fuzzy logic was first implemented entirely in C code and then tested successfully for compatibility with the original Matlab version The effort to measure the code s execution time is almost complete at this time This is needed to determine which functions are most computationally intensive and thus need to be executed on the FPGA 23 hardware Finally preliminary version of the CPU to Host interface has been designed and tested During the first few months of 2000 contract year 2 the Hardware Team plans to complete the remaining details of the software implementation and timing analysis of the Self Validation algorithm and begin its implementation on the FPGA boards By the end of 2000 we plan to perform similar algorithm analysis software implementation timing analysis and FPGA hardware implementation of the Sensor Fusion al
201. p if ch in time 0 rate of ch 0 else rate of ch j ch in temp ch in time end finding the Variance of the five value with respect to the median the third input to Fuzzy system 0 for 1 1 5 var var temp l curr temp 2 end vr j var 5 End Preprocessing Execute Fuzzy Logic next Getting the Self confidence from the fuzzy system conf j evalfis med_temp j rate of ch j vr j fis end figure plot a 2 title Raw Temperature figure plot med temp title Median Filtered Temperature figure plot rate of ch title Rate of Change figure plot vr 56 title Variance figure plot conf title Confidence echo on diary in out txt RLH 11 18 99 Execute sv4 m using sv4 fis TpDtl wk1 raw input data Number of data points sz 1 Pre processed data Filtered Temp Rate of Change Variance transpose med temp transpose rate of ch transpose vr Output self confidence transpose conf 96 Rules showrule fis Make plots of fis file figure plotmf fis input 1 title Temperature input figure plotmf fis input 2 title Rate of Change input figure plotmf fis input 3 title Variance input figure plotfis fis Ee Ae diary off echo off 2 Matlab FIS File SV4 fis System Name SV4 Type sugeno Version 2 0 NumlInputs 3 NumOutputs 1 NumRules 12 AndMethod min OrMethod max I
202. perature u T T where is the measured temperature AT 4 Temperature Rate of change r where t is the measured time At t t i i l Variance o ly r i s 5 j 0 Originally one of the preprocessed results was standard deviation instead of variance However it was changed to variance since it is relatively complicated and expensive to evaluate the square root of a number needed for standard deviation in a hardware implementation 1 4 3 Self Validation Execution Timing The processing time of the Self Validation SV code can be reduced if a part of the code is implemented on FPGAs Parts of the code that have longer processing times are being identified The following two methods are being used to determine the processing time for the code namely hardware based timing measurement and theoretical timing analysis 1 4 31 Determination of Timing with Hardware Timer First we tried to calculate the processing time using a standard hardware timer chip called a Programmable Interval Timer PIT C code was written to configure the operating modes of the PIT We achieved partial success in this method However it had several problems First the processing time measurements need a resolution of a few microseconds The PIT could not measure with such fine resolution Second the code 17 was running in the Microsoft WINDOWS environment Background processes
203. plant with unknown parameters were made to track the reference signal through switched nonlinear feedback control strategy Many controllers were designed and the controllers are selected online through a performance evaluation procedure that uses the output prediction error The paper also discusses sufficient conditions under which the closed loop control system is exponentially stable This approach achieved asymptotically stable control and the results of this approach were illustrated with three examples Automatic synthesizing of controllers other than gain scheduling was used in 20 The paper describes a method that automatically derives controllers The controllers were derived for timed discrete event systems with non terminating behavior modeled by timed transition graphs The specifications of control requirements were expressed by metric temporal logic MTL formulas The syntheses of the controllers were performed by using a forward chaining search and a control directed backtracking The synreport process does not require explicit storage of an entire transition structure This feature of automatic synthesizing of the controllers for the above procedure of switching controllers may compliment each other for obtaining superior performance from an adaptive controller 49 Adaptive controllers give better performance even when the system parameters or the environment changes Adaptive controllers have gained importance with rigorous pro
204. problem of improving the performance of the system even under the failure of sensors is solved using adaptive control approach The self validation technique and multiple sensor fusion algorithm is used to decide upon the adaptation of the controller The self validation technique developed in Year 1 of the project was reviewed in section 2 4 In Chapter 3 the developed methodology for redundant as well as multi modal sensor fusion is presented 51 52 Chapter 3 3 MULTIPLE SENSOR FUSION In Chapter 2 several multiple sensor fusion algorithms were discussed Sensor fusion is used to reduce the effect of a sensor failure over the operation of the system considered The signals from sensors are fused to get a better estimate of the measurand value Thus sensor fusion helps in improving the reliability of the measurements that primarily affects the performance of a system This is especially true in the case of feedback control systems Among the multiple sensor fusion algorithms discussed in Chapter 2 many techniques build on averaging the redundant sensors readings However averaging the sensors data would still mean that a failed sensor would affect the estimate value So the factor that should be considered in the sensor fusion is the confidence in the data obtained from each sensor This is the concept that was introduced in section 2 2 as the self confidence A multiple sensor fusion algorithm incorporating this self confidence wil
205. ps and the variables in a group of the trend monitor modality All groups of this modality are listed in the menu ring Monitor Modality Names When a particular group is selected all variables in that group get listed in the list box below the menu ring Double clicking on any one of the variables the dialog in Figure A 33 is shown The user can setup parameters of the trend monitor engine The window length indicates the number of sample points to be used to calculate the trend Slow changing variables can be assigned a longer window length The Trend Monitor also provides with an option wherein it throws an alert when certain trends are encountered The Available Cases list box lists all the possible trend cases and allows the user to add cases to be watched out by using the Add and Remove button The added cases appear in the Cases to be watched out for The Preview shows the trend of the variable for the case selected 191 192 gt Setup ariables vi x Plant Pyrometer Available Cases Cases to Watch out for Length i 0 5 E i ling Time Threshold MT 50 20 MR 0 1 20 FC 0 1 20 1jo o0100 Figure A 33 Assign Parameters for trend monitor The Assign Defaults button in the Figure A 32 dialog allows the user to setup the default parameters The interface is exactly the same as shown in Figure A 33 4 1 1 6 2 3 2 Charge setup gt Charge Setup _vi x Average time E
206. r sensors This can be observed from Figure 3 26 It can be observed that the distribution of erroneous sensor spreads further and its contribution to the final fused distribution is reduced Distributions on Individual Trends and Fused Trend Correct Sensor Additional Information on Trend 100 Erroneous Sensor Fused Distribution 0 1 0 05 o 0 05 0 1 Rate of Change Figure 3 26 Distributions of Trend after accounting for agreement between sensor trends Using the new confidence measure the fusion distribution is recomputed Since the self confidence of the each sensor is dependent on the degree of agreement between all the sensors the effect of the erroneous sensor is largely eliminated and there is a great improvement in the performance of the sensor fusion algorithm From the combined 79 distribution the fused value is the argument of the peak of the distribution the larger side of the centroid of the distribution The equations are as shown below xPDF x dx Centroid PDF X Measurand Estimate arg Peak PDF x Where x is the parameter whose fused value is to be estimated PDF x is the estimated density function of the parameter The developed fusion algorithm incorporating trend was tested for various sets of data and the results were in agreement to those expected It was also observed that the performance of the sensor fusion algorithm co
207. radar radar radar reda Cupola Ext_ metaltemper 201990716 sursze o Cupola Press pressure dro 0 1339 m S satoC So0 scimatGoF Press Cupola Pres Manual Melt of rol 19541 tonner 214951 tonih Manual Melt Manual Manual Melt Manual Melt Manual Manual Figure A 8 Example of Standard Grammar A 1 1 2 Create Modality Standard Grammar The Standard Grammar can be created or modified using this function Double click on Create Modality Standard Grammar in Setup Menu Figure A 6 will open the dialog shown in Figure 9 clicking on button Create New will open the dialog Figure A 10 to create a new standard grammar Clicking on button Modify Existing will open the dialog in Figure A 11 to modify the existing standard grammar Create Standard Grammar Modality Options Variable Options Save Options Save Save Return to File 5 Menu Save Standard Grammar Save Standard Grammar with or Return to Mei with Original File Name with a New File Name without Saving Figure A 9 Create Modify Standard Grammar Create Standard Grammar Figure A 10 Create New Standard Grammar 178 Create Standard Modality Options Variable Options Save Options Standard Variable Names standard grammar INPUTS Ad
208. rams for all the VHDL modules were created These diagrams were used to optimize the reuse of primitive components such as the multipliers and the dividers They were also used to show the sequence of operation for each module These block diagrams include Area Comp This diagram represents the comparison between the left and right of the sensors centroid 45 1 AreaP4 These four diagrams illustrate the computation of trapezoid area Centroid This diagram illustrates the calculation of the centroid Dividers This diagram shows the components needed for the division operation and the sequence of their use 1 5 These five diagrams show the sequence of calculation needed to compute the fused value Inpmux This diagram illustrates the use of multiplexers to choose between different input sets to the dividers or the multipliers Multipliers This diagram shows the design of multipliers with different bit width Rearrange This diagram pictorially shows the sort operation of the N 4 1 points in an ascending order It calls the sort module to perform the sort operation Sort This diagram shows the details of the sort operation Traphtp1 Traphtp5 These five diagrams show the calculation of the height of the trapezoids Ue This module shows the design of a comparator unit Validatepl Validatep4 these four diagrams are used to show the sequence of Validating sensors data 3 6 Virtex FPGA Board APS V2
209. ransfer functions varies the model The frequencies are all different in order to study the worst case scenario The controller performs excellent in this test The one case at 12000 seconds indicates the melt rate cannot be controlled At this point the input plot shows the blast rate is at its maximum level therefore the desired operating point cannot be reached 122 dT 1000 dMR _ 1000 dC 1000 dO2 100 dBlast _ _ dCMR BR HF f Figure 4 14 The results of varying the model parameters 4 4 10 Varying Pure Time Delay of the CMR The Smith predictor depends on previous knowledge of the time it takes for the charge to burn down to the output level It can be an average for the range of the cupola s operation or it could be a function of the inputs and outputs Even if it were given by a function the penetrability of the charge by the blast cannot be absolutely known clumping of charge materials in the cupola and coke consistency can all lead to inaccuracy in the calculation Therefore results with the cupola given a 2400 second time delay with the Smith predictor given an 1800 second time delay is given in Figure 4 15 123 Figure 4 16 is the results of a pure time delay of 1200 seconds in the cupola and 1800 seconds in the Smith predictor dT 1000 dMR 1000 dO2 100 dBlast _ _ Bog BR MR CMR Figure 4 15 Smith predictor with a 600 second time d
210. ratures about serial and parallel communication protocols The study helped the group understanding the features requirements capabilities and limitations of various communication protocols Based on this study a serial communication protocol was designed to control the traffic between two computers We researched the commercially available microprocessor boards These boards were evaluated based on the type and speed of the microprocessor the size of RAM available on board the type and size of their external buses the software used in downloading programs onto the board and their cost We also researched commercially available DAQ cards These cards were evaluated based on their sampling rate number of input channels number of output channels programmability and cost The studies about FPGA boards microprocessor boards and DAQ boards helped the Hardware Team in preparing the specification list for purchasing these devices 10 1 3 Hardware Component Acquisitions We selected purchased and tested an appropriate CPU board DAQ board and four FPGA boards for this project Following are their descriptions 1 31 CPU Board We wanted a compact inexpensive and fully PC compatible which eases software development CPU board based on the common PC 104 bus After considerable research we decided to purchase a microprocessor board the SBC2586 166 with options 2586 30 8 2586 25 BO BC3 1 from Micro Sys with these prim
211. rch an algorithm was developed for estimating the measurand value The algorithm is an integration of the peak and centroid methods 1 2 Find the range X which contains 95 of the PDF energy Find the centroid of the PDF using x PDFdx Centroid X PDFdx X Find the area on each side of the centroid The estimate 15 found as the value of measurand that corresponds to the supremum of the PDF on that side of the centroid that has the higher area thus Measurand Estimate arg Sup PDF 3 2 The function arg corresponds to finding the x co ordinate at which the maximum value occurs in the PDF This procedure is illustrated in Error Reference source not found 56 0 2 Centroi 0 15 01 Estimate 0 05 0 0 5 10 15 20 25 Figure 3 2 Estimation of the Measurand Value This particular method of finding the estimate is found to be more advantageous than other methods as explained in this section The estimate of the multiple sensor fusion algorithm should be the value on which most of sensors agree and at the same time the estimate should not be adversely affected by invalid sensors Other methods for estimating the measured value from the PDF such as centroid and peak allow faulty sensors to have an effect on the estimate or may not give the most probable value on which the sensors agree Figure 3 3 shows how the estimate if chos
212. re 5 19 Changes in Iron Temperature deg F during Run 3 146 Carbon 11 21 31 41 51 61 71 81 91 101 111 121 131 141 151 161 171 181 191 201 211 221 231 241 251 261 271 281 291 301 311 321 Figure 5 20 Changes in Carbon during Run 3 Figure 5 21 and Figure 5 22 illustrate a different capability of the PPSC As we mentioned earlier one of the modalities of IPSC is a monitoring modality that can be directed to monitor the trends of specific variables This modality can also be directed to monitor for a set conditions on multiple variables including specified trends and absolute values In the case illustrated here the monitoring modality detects the occurrence of a bridging condition in the cupola through the monitoring of two parameters namely the cupola exit temperature and the cupola back pressure These variables are easily measured and continuously monitored The two variables as shown in Figure 5 21 and Figure 5 22 show a simultaneous increase during the marked window The simultaneous increase of both variables is a good indicator of 147 the occurrence of bridging the cupola The operator could thus be alerted for the bridging and an action to alleviate the problem Exit Temperature 1800 1600 1400 1200 1000 800 600 400 200 ry gt Lo 5 D ce
213. rget for downloading and also to specify which test to perform The program was used successfully to test the FPGA system 1 44 Analysis and Validation of Algorithms The algorithms for self validation of sensor data were obtained from the Intelligent Algorithm Group in the form of Matlab and Excel files A sample set of Matlab files are shown in the Appendices The algorithms were then analyzed and validated for the purpose of successful implementation in hardware The following sections discuss more in detail about the self validation fuzzy logic and preprocessing algorithms applied to the sensor data 1 41 Self Validation Fuzzy Logic Fuzzy logic is a convenient way of mapping an input space to an output space Fuzzy inference is the process of formulating the mapping from a given input to an output using fuzzy logic The Matlab Fuzzy Logic Tool Box supports five parts of the fuzzy inference process that includes Fuzzification of the input variable e Application of the fuzzy operator e Performing the implication operation e Aggregation of outputs e Defuzzification 14 Fuzzy logic is used for self validation of the sensor data FIS file text file specifies the inputs and outputs of the fuzzy logic system the type of fuzzy logic the range and shape of membership functions and every other detail about the fuzzy logic system see Appendix Self validation using fuzzy logic on sensor data is used to determine the confidence
214. riables and an execution engine The information inside the system is organized as a data structure Node is the basic component in the data structure Modality variable and property are three elements of a node All the nodes with the time form the four dimension data structure Therefore to build a system for a new application the user needs setup the modalities the variables in the modality and the property of the variables The user can setup these in the dialog shown in Figure A 6 gt Setup Application Setup Menu Define Standard Grammar Create Modify Standard Grammar Set Output File Location Select Modalities amp Variables amp Interface Select Variable Properties Save Setup Information Set Default Values View Error Messages Done Setting Up PS Double on any of the options to make zeledion Figure A 6 Application Setup Menu The main procedure of setting up a new application is followed Define Standard Grammar refer to section A 1 1 1 Create Modify Standard Grammar refer to section A 1 1 2 Select Modalities amp Variable amp Interface refer to section A 1 1 3 Select Variable Properties refer to section A 1 1 4 Save Setup Information refer to section A 1 1 5 Modality Setup refer to section A 1 1 6 aD dA des po c Done Setting Up refer to section A 1 1 7 175 A 1 1 1 Define Standard Grammar Standard Grammar is the basic reference to describe the structure o
215. rk the pyrometer data is said to have high time resolution and the TC is said to have as low time resolution sensor The assumption made is that the reliability of low rate signal is higher than high rate signal 89 Under this scenario the fused data will be both higher in accuracy and have higher time resolution than can be gleaned from either source on its own 1560 1520 Pyro_Temp Bath Temp 1480 A E 1440 1400 i f nr MI 1360 1320 1280 1240 1200 1 22 43 64 85 106 127 148 169 190 211 232 253 274 295 316 337 358 Figure 3 34 Cupola temperature data Wavelet transforms can be used to project the data features into different levels of time resolution The fusion process is thus performed at the appropriate time base of resolution common to data from sensors having different sampling rates By fusing the data features of different levels the sampling rate difference between two data sources can be compensated Figure 3 35 Figure 3 35 show an example of such situation in which a multi rate fusion algorithm would be needed The high sampling rate signal is corrupted while the low sampling rate signal is not The result of a wavelet based fusion algorithm is shown in Figure 3 36 The effectiveness of the developed algorithm is 90 further discussed in Table 3
216. rman Frommer Sanjeev Kulkarni Peter J Ramadge Controller switching based on output prediction errors IEEE Transactions on Automatic Control vol 43 no 5 pp 596 May 1998 20 Michel Barbeau Froduald Kabanza Richard St Denis A method for the synreport of controllers to handle safety liveness and real time constraints JEEE Transactions on Automatic Control vol 43 no 11 pp 1543 November 1998 21 Specht D F Probabilistic Neural Networks Neural Networks November 1990 22 Ronald R Yager and Dimitar P Filev Essentials of Fuzzy Modeling and Control John Wiley amp Sons 1994 164 23 Jeff Frolik and Mohamed Abdelrahman Synreport of Quasi Redundant sensor Data A Probabilistic Approach In Proceedings of American Control Conference 2000 24 Hassan K Khalil Nonlinear Systems Second edition Prentice Hall Inc 1996 25 Mohamed Abdelrahman Kevin Moore Eric Larsen Denis Clark and Paul King Experimental Control of a Cupola Furnace In Proceedings of American Control Conference 1998 26 Pascal Gahinet Arkadi Nemiroviski Alan Laub and Mahmoud Chilali LMI Control toolbox 1 0 The Math Works Inc 27 Jeff Frolik C V Phanishankar and Steve Orth Fuzzy Rules for Automated Sensor Self Validation and Confidence Measure Proc of American Control Conference 2000 pp 2912 2916 28 Mohamed Abdelrahman Parameshwaran Kandasamy and Jeff Frolik A Methodology for the Fusion of
217. rning the software tools we needed and the signal processing methods and hardware implementation techniques available We also researched the possible system organizations communication requirements and commercial boards available for the embedded microcomputer CPU data acquisition interface and programmable logic FPGA Field Programmable Gate Array needed for computationally intensive tasks After making decisions about our functional and cost requirements we then selected and purchased the appropriate commercial boards one CPU one DAQ and four FPGA boards Basic testing and familiarization work was done on the CPU and FPGA boards while the DAQ board has not yet been tested At this time the overall system consists of two algorithms Self Validation and Sensor Fusion The Self Validation algorithm whose hardware implementation is now in progress inputs the raw time temperature measurements from sensors derives some characteristic quantities and filtered outputs and then applies fuzzy logic to determine a self confidence value for each sensor The Sensor Fusion algorithm whose hardware implementation work has not yet begun combines the filtered inputs and self confidence values from several sensors into one robust value As the first step toward hardware implementation we analyzed validated and refined the Self Validation algorithms that were supplied by the Intelligent Algorithms Team The algorithms were divided int
218. rr2_7 gt UE E 8 8 8 nae E 8 gt f gt tar 7 7 tarr2 9 8 U E 8 aro _ tarr2 10 gt fp tar m 11 A U E 3 10 8 8 8 tars _ gt og tar3 11 EP d tarr2 12 A 127 128 Trap ht p1 COMP t t b t gt b trap lt 51 b or trap en b COMP trap 25st b and trap st t COMP trap 2st t and trap lt en t trap 2st trap lt 1 trap en t and trap amp en 8b 8b 16b hl Ae d AoH ml LAS 8 255 8 Jae 8b 85 9 16b 16 2 8 g 255 7g 28b 0 1 2 3 129 Trap_ht p2 130 Ld hsumarr 16 16 D fot h sum 0 clear clr hsum 4 hsum hl 8 1 ht_ par 16b 16b ht par sum 16 16 height 8 gt p D sum l 2 D 2 8 16b 16 16 PARA h2 clk gp clear clr clr hsum 16 16 gt fo h_sum 40 clk clear clr_hsum 131 h ind z 132 Trap_ht p3 int_ind from master controller h sum int ind max COMP 4 16 Ld 16 l D 16 max sig 16 clear clr _hsum max ind sig max ind 32 clk clear clr
219. rt of it is show in the next page y Done Setting Up Figure A 4 Procedure of Setting Up Application 172 172 Return to Menu Declare Model Setup Files _ Need declare new setup file diu Run Setup VI Y Y Y Y ensor i Parameters Modality Groups Du Charge Setup Planner setup Setup y Y Y Y Self Standard Save and Exit without Delete Define confidence Deviation Exit saving Quit Update Modality Measure Measure Modality Parameters v Add Variables in a group Add Group v v y i Modify Delete Quit Modality modality cna Y ange Change Done Matlab le Parameters List v Figure A 5 Procedure of Modality Setup Setup Application The main purpose of setting up an application is to setup the modalities to build a TPSC system PPSC contains a generic interface in order to include different modalities required by the application Planner Controller Plant Virtual Sensor Fusion Monitor and Expert are eight modalities that have already been given A system can be built with these modalities Every modality has several groups Each group has input variables output va
220. rzen estimator utilizes parametric functions such as Gaussian functions that are centered at each of the sensors readings The functions are then added up and normalized The resulting Probability Density Function PDF reflects the distribution of the sensors data The PDF energy is more concentrated where more data points exist This is illustrated in Figure 3 1 54 0 2 0 15 AN X Sensor Value 01r r p Cumulative PDF 2 Individual Gaussian Fn 0 05 7 0 HH 5 10 15 20 25 Figure 3 1 Individual Gaussian Functions and the Cumulative PDF For this research a Gaussian function GF is selected as the parametric function The mean value of the GF is equal to the sensor reading and the standard deviation is CT estimated from the noise level in sensor 14 The PDF is given by 2 20 PDF x where is the number of sensors is the k sensor data and o is the standard deviation The parameter o is estimated based on the standard deviation of the noise associated with each of the sensors considered 3 1 3 Estimation of Measurand Value from PDF The estimate value of the measurand is calculated from the PDF obtained as There are several ways to get an estimate of the explained in previous section 55 measurand value similar to defuzzification methods such as average centroid maximum and sum of the maximum 22 In this resea
221. s This is accomplished by using a probability density function PDF around each sensor s data This PDF around each sensor s data is used to find the distance from other sensors data This distance measure is stored as a matrix for each sensor which are combined later to find a combined matrix from which the optimal fusion estimate is found This method of having an individual matrix and forming a combined large matrix that is reduced to get optimal value is called the Bayesian approach 5 Many others also approach the multiple sensor fusion problem by finding the best combination of sensors that are to be fused The search is based on the distance between the sensors each sensor s failure rate and its previous data Algorithms like neural based search and genetic algorithms were used Chapter 10 3 The combination of selected sensors is then usually averaged to find the estimate The performance of these approaches depends on the search algorithm These are best suited for decision making sensor fusion problems As an extension to the above search first and then fuse multiple sensor fusion is implemented using approximate agreement approach in Chapter 11 3 The approach first establishes an agreement set on each sensor data This is done by each sensor 40 broadcasting its value to other sensors Each sensor then forms the agreement set based on the distance from other sensor data This agreement set helps in eliminating invalid sen
222. s SmPosMR and deMRate is LgNegdMR then ChangeBlast is SmPosBR 17 If eMRate is SmPosMR and deMRate is SmNegdMR then ChangeBlast is SmPosBR 18 If eMRate is SmPosMR and deMRate is ZerodMR then ChangeBlast is SmPosBR 19 If eMRate is SmPosMR and deMRate is SmPosdMR then ChangeBlast is SmPosBR 20 If eMRate is SmPosMR and deMRate is LgPosdMR then ChangeBlast is SmPosBR 21 If eMRate is LgPosMR and deMRate is LgNegdMR then ChangeBlast is SmPosBR 22 If eMRate is LgPosMR and deMRate is SmNegdMR then ChangeBlast is SmPosBR 23 If eMRate is LgPosMR and deMRate is ZerodMR then ChangeBlast is LgPosBR 24 If eMRate is LgPosMR and deMRate is SmPosdMR then ChangeBlast is LgPosBR 25 If eMRate is LgPosMR and deMRate is LgPosdMR then ChangeBlast is LgPosBR Soo SU Bur or The fuzzy inputs are eMRate for the error in the melt rate and deMRate for the rate of change in the error The membership function names are LgNegMR and SmNegMR for large and small negative melt rates LgPosMR and SmPosMR for large and small positive melt rates LgNegdMR and SmNegdMR for large and small negative rate of change in the melt rates LgPosdMR and SmPosdMR for large and small positive rate of change in the melt rates dealMR is about a zero error ZerodMR is about a zero rate of 127 change in the error of the melt rate The outputs are constants for large or small negative or positive and zero changes in the blast
223. s satisfactorily under these extreme conditions dT ____ 1000 dMR _ 1000 dC MR 1000 dO2 100 dBlast _ _ dCMR BR Figure 4 17 System Performance Under Effect of All Disturbances 126 4 If eMRate is LgNegMR and deMRate 15 LgNegdMR then ChangeBlast is LgNegBR If eMRate is LgNegMR and deMRate is SmNegdMR then ChangeBlast is LgNegBR If eMRate is LgNegMR and deMRate is ZerodMR then ChangeBlast is LgNegBR If eMRate is LgNegMR and deMRate is SmPosdMR then ChangeBlast 15 SmNegBR If eMRate is LgNegMR and deMRate is LgPosdMR then ChangeBlast is SmNegBR If eMRate is SmNegMR and deMRate 15 LgNegdMR then ChangeBlast is SmNegBR If eMRate is SmNegMR and deMRate is SmNegdMR then ChangeBlast is SmNegBR If eMRate 15 SmNegMR and deMRate is ZerodMR then ChangeBlast is SmNegBR If eMRate 15 SmNegMR and deMRate is SmPosdMR then ChangeBlast is SmNegBR 10 If eMRate is SmNegMR and deMRate is LgPosdMR then ChangeBlast is SmNegBR 11 If eMRate is IdealMR and deMRate is LgNegdMR then ChangeBlast is SmPosBR 12 If eMRate is IdealMR and deMRate is SmNegdMR then ChangeBlast is ZeroBR 13 If eMRate is IdealMR and deMRate is ZerodMR then ChangeBlast is ZeroBR 14 If eMRate is IdealMR and deMRate is SmPosdMR then ChangeBlast is ZeroBR 15 If eMRate is IdealMR and deMRate is LgPosdMR then ChangeBlast is SnNegBR 16 If eMRate i
224. s selected the graph is displayed When done press Close Window and the application will close 212 213 View Output Graphs vi File Edi Operate Project Windows Help gt Excel Files Output File Directory Experiment xls 4IC Datadnalysis D ata NNData xls nominal sls Real fs xls Read New Fie RealSensor xls SimRef xls Close Window Ie Independent Parameter Variable Correlation Correlation 20 Testnmy xls blast temperature 0 9 Output Parameters melt rate metal temperature 1 metal temperature 2 Final Carbon Final Silicon Final Manganese 08 200 4000 00 1000 ENR gt 7 Figure 59 View Single Variable Correlation Graph View Output Graphs vi A 2 6 View Multi Variable Graphs A multi variable correlation has a vast amount of information in its database This graph viewer is designed to cut small slices out of the data and display it on the graph There are two graphs The input variable is displayed above the graph and the output variable is displayed to the left Any combination of inputs and outputs can be selected from the ring boxes of the input variables are shown to the right of the graphs along with the values for each variable that were used in the correlation This allows the user to tinker with the various inputs and view what happens to the variable being graphed 213 214 gt View Multi Graph File Edit Operate Projec
225. signal processing methods software tools and hardware devices available We selected and purchased appropriate hardware CPU Board FPGA boards DAQ board and found effective ways to utilize them An overall system architecture was created We developed software 48 implementations of the two major algorithms Self Validation and Multi Sensor Fusion to verify their proper operation We replaced the original floating point arithmetic with fixed point versions for efficient hardware implementation A library of fixed point arithmetic routines was created in VHDL We designed and implemented a sophisticated communication structure to tie together all the system elements Many enhancements were made to the original algorithms to enhance their performance in hardware We developed detailed architectures for their implementation into hardware We partitioned the algorithms into two sections one section contained computationally intensive portions destined for hardware implementation and the other section continued to be implemented in software running on a standard microprocessor The hardware portion of the SV and MSF processors was implemented in VHDL and simulated Two new subsystems were designed specifically for the FPGA including a PC 104 bus interface and a memory block Finally the SV processor was fully implemented and tested on the FPGA utilizing all of the system components In addition to the technical contributions of this project sev
226. sing and control system or PSC for short for the foundry cupola the primary industrial process used for producing cast iron However the 17 5 is generic enabling cross cutting technology that can be broadly applied to advanced process sensing and control problems in the ferrous metal casting industries as well as in other industrial environments The project addressed two main objectives A Development of a generic architecture for the integrated intelligent industrial process sensing and control system The proposed 17 PSC architecture is characterized by e Intelligent signal processing capabilities and sensor fusion methodologies e Intelligent algorithms for hybrid model fusion e Methodologies for integrating intelligent signal analysis and sensor and model fusion algorithms with intelligent model based control methodologies An object oriented generic architecture for integrating all system components e Implementation of the intelligent signal processing and sensor fusion algorithms through hardware realization using reconfigurable logic B Demonstration of the application of 17 to the specific industrial setting of cupola iron melting furnaces The demonstration will include 16 Testing of the developed algorithms using experimental data and static and dynamic models available from a production cupola and the ALRC research cupola Implementation of the developed algorithms on the 18 inch research cupola at
227. sor data and find the estimate on which most sensors agree This approach requires 3t 1 sensors with 1 giving accurate reading where t is the number of faulty sensors The mean of the agreement set after removing t lower and t higher data gives the estimate This algorithm is again best suited for binary decision making Target or no Target Multiple sensor fusion techniques use the redundant data and come up with one value Each sensor data have an effect on the final estimate A failed sensor will have adverse effect on the estimate if not removed So it is necessary to validate the sensor data before fusing the redundant data and remove the sensor The next section discusses some of the signal validation techniques that achieve this 2 3 2 Signal Validation Signal validation is a technique by which the sensor s signal is validated for its accuracy Signal validation may involve all or one of the following detection isolation and characterization of faulty sensors 6 Most of the initial researches depended on finding an additional measure for the sensor either by having redundant sensors or by producing an analytical redundancy to the sensor data by using a model for the process A detailed survey of the signal validation using redundant sensors based on statistical methods is described in Ray and Luck 7 The statistical approach is based on the difference between the current sensor data and other redundant sensor data Fuzzy 41 log
228. ss Examining this figure it is clear that the erroneous sensor is effecting the calculation of the fused value The methodology 71 introduced in this section aims at handling such situations by using the available trend information Confidence plot of the Fused Reading 1 T T 0 9 4 0 8 4 M 07 4 1 0 6 d 0 5 4 0 4 0 10 20 30 40 50 60 70 Instants of Time Figure 3 19 Confidence Plot Trend Informations 0 15 Trend of Correct Sensor Trend from an Another Source 0 1 Trend of Erroneous Sensor 0 05 5 0 0 05 0 1 0 10 20 30 40 50 60 70 Instants of time Figure 3 20 Sources of Trend Information 72 3 4 3 Fusion based Trend As explained earlier the MSF algorithm presented in 27 considered only the agreement between the sensor values rather than their trend By examining Figure 3 18 it is obvious that one of the used sensors is trending differently Obviously this sensor is erroneous and its value should not be allowed to affect the fused value The trend fusion algorithm introduced thereafter minimizes the influence of that erroneous sensor on the fusion process by including the trend of the sensor as a source of information for the fusion process The algorithm proposed in this section looks at the trend of the parameter along with the measured value Based on the trend information provided by each of the real sensors a fused
229. st PC as presented elsewhere Of course the CPU board directly controls the Virtex FPGA board passing low level commands and raw data to it and receiving back the processed SV data at each sample time It essentially calls on the FPGA as a high speed hardware based subroutine implementing the signal processing algorithms 47 3 8 Communication A sophisticated error detecting and correcting high level communication protocol was defined and implemented over the serial communication line between the Host PC and the CPU board running at the full 115 200 bps speed The line passes all commands status input data and output data between the Host PC and the CPU board This protocol allows the system to automatically recover from transient communication errors and continue normal operation 3 9 Host PC Application Control and communication with the CPU board and FPGA were implemented as a Labview application on the Host PC giving a simple and effective interface to the system operator This application fully supports all aspects of both the SV and MSF downloading and processing tasks Each step in the process may be executed independently simplifying debugging and reducing the total execution time 3 10 Summary 3 10 1 Work Completed During this I3PSC project the Hardware Team successfully completed most of its planned work but had insufficient time to finish a few portions The team began with a literature search to study the possible
230. studied The group members used all available resources in obtaining information These resources included published papers books and the World Wide Web The group members studied and acquired experience in using some of the software commonly used in the fuzzy logic and the neural networks areas such MATLAB the fuzzy logic toolbox using MATLAB and the neural network toolbox using MATLAB The group also studied the use of LabView software which is often used with Data Acquisition Cards DAQs On the hardware side the group searched the published literature about hardware implementations of fuzzy logic and neural networks especially those implementations using reconfigurable logic The methods and techniques used in some of these implementations were studied and summarized for future use The group also studied the data books of programmable logic devices with concentration on the logic families produced by Xilinx and Altera The study educated the group especially the graduate students about the state of the art families of field programmable logic arrays FPGAs in terms of their logic capacities features structure interconnection and speed The programmable logic device study was complemented by another search of commercially available FPGA boards These boards were evaluated based on their logic capacity speed external RAM availability the width of their interface busses and cost The Hardware Team also searched the published lite
231. t 16 bit 16 bit 8 bit In addition the following techniques of multipliers designs were implemented Array multiplier Pipelined array multiplier Dadda Wallace multiplier Pipelined Dadda Wallace multiplier The multipliers were developed for 8 bit by 8 bit inputs All the designs in the VHDL library were coded simulated and synthesized These designed were optimized for the hardware implementation of the SV algorithm on the XC4085XLA FPGA chip A bit maps for the best designs were then generated downloaded onto the FPGA chip and successfully tested 2 4 Implementation of the SV Preprocessing Algorithm The modified preprocessing algorithm was written with fixed point code The code was successfully tested and evaluated with data from the Intelligent Algorithms group The code was then reorganized and split between the setup and execution file 27 codes A new input files were created for the signal Validation code After fixing numerous implementation problems the code was successfully tested and its results were validated Finally the process was completed after inserting appropriate documentation to the code 2 5 Develop Architecture of SV Signal Processor Hardware 2 5 1 Select SV procedures for the hardware implementation We decided to implement in the FPGA hardware all the procedure in the execution code that performs computations on the outputs from the preprocessing stage These procedures include FisEvaluate FisCo
232. t Windows Help gt Input Variables 1 coke ratio Output File Directory SIC Datadnalysis Data Excel Files NND ata xls Read New File nominal xls xls csse coke ralio 8 500000 RealSensor xl blast rate 2250000 Test sls Output Variables fant Testmy xls Offgas CO vi a aile blast temperature 300000000 3000 10000 11000 12000 13000 14000 Input Variables 2 coke ratio coke ratio 000000 Output Variables 2 blast rate 30050000 oxygen addition zo 000000 blast temperature 300 000000 3000 10000 11000 12000 13000 1400 Figure 60 View Multi Variable Corr Graphs View Multi Graph vi 2 7 View N M Correlation Graphs The nominal value multi variable correlation creates a spreadsheet file that shows the relationships between varying a large number of inputs and outputs The graph application is set up like a matrix The columns of graphs are all associated with the input parameter shown at the top of the column The rows are associated with the outputs shown on the left side of the row This configuration creates a four by four matrix of graphs showing the user the trends of many variables at once File selection is the same as before 214 w Nominal Graphs vi 215 File Edit Operate Project Windows
233. t up time the steary burn period the transition and the shutdowm by highlight the menu and clicking Run 193 194 165 Add Behaviors vi 3 Update melt rate gt Poo Emm Add Start amp Stop Time x Add Start Up Add Steady Burn 2 Add Transition Add Shutdown 0 40 Save Heat Plan Done i 9 00 nma D D 1 i 01 00 AM 01 00 AM 01 00 AM 01 00 AM 01 01 AM 01 01 01 01 01 01 01 01 01 01 Time Pyrometer Temperature Poto E 1750 00 1500 00 1250 00 8 1000 00 B 750 00 500 00 250 00 0 00 00 T T 1 1 01 00 01 00 01 00 01 00 01 00 01 01 1 01 01 01 01 01 01 01 01 01 01 Time 101 2nd Pyrometer Plot T T 1 01 00 AM 01 00 AM 01 01 Al 01 01 01 01 01 01 Figure A 36 Add new plan In the dialog shown in Figure A 37 user setups the start and stop data and time The startup points can be setup in the dialog in Figure A 38 Set Run Time Range vi File Edit Operate Tools Browse Window Help set run gt 13 Application Font 65 1 times a Start Date and Time January zoo Start Time AN Alo Afo d I End Time End Date and Time Run Lenath s January X 2 A200 0 4 1 Figure A 37 Setup the start and stop time of a plan 194 195 File Edit Operate Tools Browse Window Help
234. terfaced to the rest of the system 1 3 3 FPGA Boards We purchased four X240 FPGA boards from Associated Professional Systems APS Each of these boards is equipped with one Xilinx XLA4085 FPGA chip that has a logic capacity of 180K gates On each of these boards there are two 128Kx8 SRAM Static Random Access Memory chips The boards are PC 104 16 bit compatible The FPGA on each board can be configured from a PC 104 bus an EPROM or a parallel port using a Xilinx Xchecker cable Each board also has a socket for standard clock oscillator The documentation that came with these boards included an application example program in C and example FPGA configuration data The example code allows the user to download the configuration data onto the FPGA chip mounted on the board and to test the on board SRAM Using modified versions of the example code each of the four boards was tested An ISA Industry Standard Architecture carrier board was used to connect the X240 board to the ISA bus of a PC Three FPGA boards passed the initial test The failed FPGA board was sent back to the manufacturer where it was repaired Upon its return the board was tested successfully The four FPGA boards were then mounted on top of each other using the PC 104 bus to create the FPGA system The provided example code was modified to allow the user to communicate with all the FPGA boards The new program allows the user to 13 specify any of these boards as the ta
235. that if the output noise caused the change in error signal to enter the large change in error range in the fuzzy system then the inputs may change significantly to correct for it It then leads to an oscillating input and output As can be seen in Figure 4 10 the controller rejects output noise efficiently The inputs move seldom under the noise conditions introduced increasing the life of the actuators INPUT DISTURBANCE Input disturbances can be common in the cupola environment The scrap iron is typically lifted with a human operated front end loader Since the scrap iron will have varying densities there will be a disturbance at the input The blast rate can be affected by its ability to penetrate the charge in the cupola The configuration of the charge is constantly changing in the cupola causing a disturbance Mechanical problems in producing the blast can cause disturbances The oxygen enrichment will be disturbed if the blast is disturbed because it is a percentage increase of oxygen in the blast air Mechanical problems in the oxygen delivery can also cause disturbances Figure 4 11 shows plots of the inputs and outputs with the inputs disturbed The disturbances are in the form of a square wave This gives periodic negative and positive disturbances to the inputs Figure 4 11 reflects the ability of the controller to adapt to disturbances The frequencies of the square waves were taken to be 6000 seconds for the 116 9000
236. that would reflect a change in the operation of the cupola The temperature was requested to undergo a change of 50 C at 3000 seconds the melt rate to change by 0 1 tons hr at the same time A change of 0 1 in C was requested at 100 seconds The reason the changes in temperature and melt rate were requested at a much later time was that the CMR has a much longer settling time due to the long time delay of the charge With these change times the three outputs changed and settled at close to the same time Figure 4 9 shows the plots of the outputs inputs error and the change in error Most of the fuzzy parameters were fine tuned using these plots Figure 4 9 shows that the settling time is very close to 600 seconds for the melt rate and temperature The carbon has a settling time of 2000 seconds after a pure time delay of 1800 seconds This long settling time is necessary for the Smith predictor to work properly OUTPUT NOISE Gaussian noise was added to the output signal This represents the fact that the sensors are subject to extreme noise because of the nature of the cupola Even when averaging several sensors measuring the same output there is noise Figure 4 10 is the plots generated with noise Notice the error plot is basically the noise after the outputs reach steady state The change in error reflects a noisy output which had to be taken into 115 consideration when choosing the membership function parameters An example is
237. the Contolera Cp TM 95 Figure 4 3 Uncertainty in the Estimate from Multiple Sensor Fusion 100 Figure 4 4 Region Stability sacs qoa deii uod e RT eR Eae iA 101 Figure 4 5 Membership functions of the error in melt rate eMRate 108 Figure 4 6 Membership functions of the change in error for the melt rate deMRate 108 Figure 4 7 Implementing Smith predictors Oda 109 Figure 4 8 Simulation lay Outs at Oe e rade cea eta 112 11 Figure 4 9 Step response under ideal conditions 117 Figure 4 10 Step response with noisy outputs 118 Figure 4 11 Step response with input disturbances generated with square waves 118 Figure 4 12 Response for melt rate confidence of 0 9 and 0 1 pulse for 600 seconds 120 Figure 4 13 Response for melt rate confidence of 0 5 and 0 1 pulse for 600 seconds 120 Figure 4 14 The results of varying the model parameters sss 122 Figure 4 15 Smith predictor with a 600 second time delay plant offset 123 Figure 4 16 Smith predictor with 600 second plant time delay offset 124 Figure 4 17 System Performance Under Effect of All Disturbances 125 Figure 5 1 Configuration for Interfacing IBPSC with ALRC DAQ for Demo Runs 132 Figure 5 2 Control of Carbon Content Run Z2
238. the various sensors data points are presented as separate columns of temperature and time values The self validation system operates on one set of data for one sensor and determines the self confidence value for that sensor It then reads the next set of data for the next sensor and determines the self confidence for that sensor The output of this system is a set of self confidence values for the different input data of the various sensors The system was tested successfully and it can now handle up to 50 sensors The code will be modified to accept data directly from the sensors instead of a data file 2 8 Develop Multi Sensor Fusion Algorithm The Hardware Group Team received the Multi sensor fusion MSF algorithm from the Intelligent Algorithms group in the form of three of MATLAB files and 4 Excel files The MATLAB files represent MSF code These files are feeddata m Serves as the front end file that reads the data from a txt file and feeds it point by point to the sensor fusion function 32 ValueConf m This is the function that does the multiple sensor fusion It takes data at every instant from feeddata m file and returns the Fused Value and its confidence at that instant checkindex m this function is called by ValueConf m and it checks whether the maximum value returned by the max function is indeed at the center the trapezoid Max function in Matlab returns the maximum value and the index of its first occurrence The Exc
239. ther and the position of the maximum height is computed The value of the normalized area within 3 standard deviations is considered as the overall confidence On the mean time the algorithm group converted the MATLAB files into a floating point C code All sections of code that apply only to the MATLAB implementation including redundant error checking were removed from the floating point C code The C code was tested successfully and produces comparable results to that of the MATLAB code 2 9 MSF C Code Optimization Careful examination of the MSF code revealed that the many of the computations performed by the MSF algorithm were not ultimately required For example The code calculated the areas before and after the maximum height of the overall graph resulted from adding the trapezoids of all sensors by dividing the total range of the trapezoids into a fixed number of steps and calculating the areas between each step A large saving in the computational requirements could be achieved by computing the areas as the sum of triangles and rectangles The areas of trapezoids triangles and rectangles were calculated using basic algebra laws In addition to be faster and less demanding in terms of computational resources the new method was more accurate also it avoided many of the round off errors Three versions of The MSF C code were developed to optimize and adapt the floating point code for the hardware implementation In the first version the c
240. this report is that the performance of the feedback controller degrades when the feedback signal from the sensor data is unreliable The problem of increasing the reliability of the feedback signal was tackled in many ways The most common method used to increase the reliability of the feedback signal is multiple sensor fusion One other approach is to check the reliability of each sensor by using self validation In this chapter a quick review of some of these multiple sensor fusion and self validation techniques is presented A basic overview of adaptive controllers and some adaptive methods are also discussed 2 3 Multiple Sensor Fusion and Signal Validation 2 3 1 Multiple Sensor Fusion Sensor fusion is defined as the method to fuse or manipulate information from different sensors and come up with one value of interest These sensors may measure the 38 desired measurand may measure different values which should be combined to get the required information If the different sensors are measuring the same quantity then these sensors are called redundant sensors In this report multiple sensor fusion is constrained to mean only the fusion of redundant sensors There are many reasons why multiple sensor fusion is used Combining several sensors data will give more accurate information of a measurand improving the reliability of measurement data The measurement data become less sensitive to noise and disturbances that might not affect
241. tion of the sensor The block diagram of the self validation algorithm from the paper is shown in Figure 2 5 MemberShip Function1 Temperature High Low Ideal MT1 MT2 MT3 MT4 Membership Function2 Rate of Change in Temperature Very_Negative Small Very Positive 10 MR8 MR8 MR7 MR7 MR8 10 MR8 Membership Function 3 standard Deviation Constant Normal High Noise MS1 2MS1 MS2 MS3 10 MS3 Figure 2 4 Membership Functions Create Create Fuzzy Acquire Data Pre process Parameters from Membership Set Data Verified Data Functions from Set Parameters Acquire Real Pre Processing to Obtain Data m Median Filter gt calculate input M Fuzzy System gt Confidence Runtime parameters Figure 2 5 Block Diagram of the Self Validation Technique 44 45 The rules of the fuzzy system remains the same for all sensors while the membership function varies from one sensor to another sensor depending upon the sensor s historical data The fuzzy output gives the self confidence of the sensor The median output of the data is the sensor data output of the self validation 2 4 Adaptive controllers Controllers are designed based on the model of the plant Earlier control designers assumed exact knowledge of the plant and that the plant is modeled accurately These controllers demanded a
242. tions The 4 D array is stored in a data file that any of the applications data collection sensor fusion controller interrogator etc on the cupola network will be able to access A setup file is also created that defines what information is stored and its location in the data structure A third file is maintained that keeps count of how many data points have been collected and how many have been processed This is done so that each separate application can access the information it needs regardless of where the application is running A 3 1 Online Setup The online setup menu requires the user to input the current run name Setup then creates the three files discussed in the previous section with the run name as the file 216 217 name extensions are dsc for the data structure dsv for the data structure variable list and dsi for the data structure counter file The data lt run name gt dsv file is initialized with the setup data that was chosen during setup this file should not be modified by any other applications The lt run name gt dsc file is initialized with zeroes in the proper dimension sizes according to how many modalities variables and variable properties were selected The time dimension is initialized to five and expanded dynamically as the data points are collected The counter file is initialized to zero I Online Setup vi File Edit Operate Project Windows Help File Location
243. to provide a closed loop automatic control system that can aid in maintaining the important operational cupola parameters such as carbon content melt rate and iron temperature with specified boundaries at various conditions of operation requirements Specific examples that were illustrated included the ability of the system to change the carbon content quickly during a run while maintaining the temperature and varying the melt rate Another example showed the ability of the system to plan a large reduction in the melt rate while maintaining the carbon content and the temperature within acceptable ranges and 6 157 reduce the transition period to steady state operation by changing the initial charge setup in the cupola The demonstration runs the publications and the developed software package illustrate that the project have achieved the proposed objectives Full utilization of the developed algorithms software and hardware within the scope of the industries of the future depends on other factors that are technical and economical The next section discusses these issues in more details 158 6 2 Future Recommendations As we have mentioned earlier the PPSC system has achieved the technical objectives set at the start of the project The system was tested using a state of the art research cupola furnace It has not yet been adopted and tested by a commercial facility Cupola foundries are in general conservative in adopting new technol
244. uld be improved further by incorporating the fused value at previous instants This would be akin to low pass filtering of the data coming out of the sensor fusion module Figure 3 27 shows the performance of the final algorithm It is clear that the algorithm has picked the correct sensor for all instances of time Multiple Sensor Fusion 43 299 0 Correct Sensor Erroneous Sensor Fused Plot 10 20 30 40 50 60 Instants of Time Figure 3 27 Multiple Sensor Fusion after filtering 80 3 4 5 Measure Fused Confidence In this algorithm it can be observed that we have different types of information sources namely trend sensors and value Thus in evaluating the measure of the fused confidence it is necessary to weigh the confidence obtained from trend fusion and that obtained from value fusion The formula used to evaluate the overall confidence of the measurement is given by N TOC N FusedConfidence Where Number of Trend Sources Confidence of Fused Trend Ny Number of Sources of Value Cy Confidence of Fused Value Figure 3 28 illustrate the calculated overall confidence for the previous example Confidence Plot of the Fused Reading 0 9 0 7 0 6 Qo5 20500 0 5 I 1 0 3 0 10 20 30 40 50 60 70 Instants of time Figure 3
245. up the parameters of the sensors These two setup procedures will be introduced in the following section A 1 1 6 2 1 and section A 1 1 6 2 2 The rest setup VIs will also be introduced in section A 1 1 6 2 3 4 1 1 6 2 1 Modality Groups The dialog of Modality Groups vi is shown in Figure 16 IE Modality Groups vi Define Modality Parameters Delete Update Modalities Quit Figure A 16 Modality Groups Main Menu Two functions are offered in this dialog namely Define Modality Parameters and Delete Update Modalities Define Modality Parameters can define a new group in the modality and Delete Update Modalities can modify the existing group in a modality 4 1 1 6 2 1 1 Define Modality Parameters Double clicking the Define Modality Parameters option opens an dialog shown in Figure A 17 This dialog allows the user to add variables that form the inputs and outputs of the modality and add them as a group and also to create multiple such groups The user has to select the variable its associated modality and property and then click the Add Variable button to add that node to the modality Once all the input and output variables are added this set is classified as a Modality Group 182 183 Add_Variable_Group vi Available List Properties Modalities value trend coke ratio blast rate Real Albany sensor E Monitor Plant Model value value standard deviation trend standard deviation
246. used to tune the controller parameters For example an initial definition of a small change in the error or a large change in error is readjusted after looking at the response of the system during a simulation A narrow value for the ideal range would cause the system to be very sensitive to noise while a wide range for the ideal membership function would allow the system to deviate considerably from the desired output Using the simulations the fuzzy output parameters were chosen such that the settling times were close to 600 seconds the system inputs would not change too quickly and overshoots were minimized Figure 4 5 and Figure 4 6 are examples of the membership functions of the two inputs for the melt rate fuzzy inference system Figure 4 5 represents the error in the MR while Figure 4 6 represents the change in error in the MR An iterative method for changing the rules from the initial guess is similarly followed These rules were updated based on examining plots generated through simulations The list of rules for the melt rate controller is shown in Appendix 4 A The rules for the other two outputs of interest in this paper namely T and C are very similar Examining a subset of these rules illustrates the main idea behind the fuzzy controller The error in an output is positive when the set point is higher than the actual value and the rate of change in the error is negative if the error is decreasing and vice versa Consider a case when t
247. ution obtained from actual sensor measurements the sensor fusion algorithm is performed to obtain the fused value and a measure of confidence using the algorithms described in details in 27 and summarized in previous section 74 Measurand at current instant Real Sensors Fused Value Multiple Sensor Fusion Measurand at previous instant Confidence Measurand not relaible Expected Measure System Model rend of Measurand reliable Trend of Real Sensors Sensor Information on Trend Trend Measurand Fusion Estimator Fused Trend Figure 3 21 General Methodology for Sensor Fusion using Trend Figure 3 21 summarizes the sequence of steps proposed in the incorporation of trend in sensor fusion Figure 3 22 shows the distribution of the trends of three sources of information at one instant namely 60 The distribution of the fused trend is shown and this distribution is convoluted with the fused distribution of the temperature at the previous instant which is shown in Figure 3 23 75 Distributions on Individual Trends and Fused Trend 120 Correct Sensor Additional Information on Trend 100 Erroneous Sensor 4 Fused Distribution 80 40 J 20 4 0 0 06 0 04 0 02 0 0 02 0 04 0 06 0 08 Rate Change Figure 3 22 Distributions of Trends and Fused Trend The final temperature distribution is
248. utput Regs Read RF2 Input 1 Read RF3 Input 2 ASM Chart SV FPGA Controller P 4 Evaluate Conf Eval0 0080 LdFS IncCnt CirFSn CirTW CirTWf gt Eval 1 11 0100 LdFS LdFSn IncCnt LdTW LdTWf 0 Count 11 Eval 12 1 0200 LdFSn LdTW LdTWf DivOut 0400 IncCnt 0 Count 4 1 C LdConf OutConf 0800 PBRF PBW RFA 5 OutConf 99 Load Firing Strength 0 Load Firing Strength 1 11 Load Total 0 10 Load Total 11 Delay for Output Divider Store Conf RF5 100 A 5 Timing Diagrams 51 Self Validation Timing 101 SV FPGA Timing P 1 8MHz 1 5 10 15 20 ope s 0 in1 in2 RF R1 R2 R3 IMF Params Mem 0 1 2 3 M4 5 M6 M7 8 M10 M11 M12 M13 M14 M15 M16 00 A01 A02 A11 A12 A20 3 1 1 2 1 2 3 1 2 3 1 1 2 3 1 2 3 1 2 Eval Eval3 Count SelRule 0 1 2 4 5 6 7 SelRuleOut Notes 1 Command handshake at start end NOT shown 2 Register File RF and Memory Mem limited to 1 Rd or Wr per clock 3 Values available in regs at end of clock shown with label 4 8MHz 125ns clock 5 Allow 3 clocks 375 ns for divider to produce Axx 6 Allow 5 clocks 625 ns for divider to produce out 102 SV FPGA Timing P 2 25 30 35 40 45 8MHz
249. value for the trend is calculated using the Parzen estimator algorithm presented in 27 Using this trend and the previous distribution of the measurand at a previous instant an estimate of the current value of the measurand is estimated This estimated distribution is further used for the estimation of the final fused value of the measurand The sequence of steps used in this algorithm is as follows e Determine the measure of self confidence from the fuzzy engine 26 e Estimate the fused trend from the individual sensors using the Parzen Estimator Algorithm 27 73 Based on the parameter value at the previous instant and measure of trend an estimate of the parameter at this instant is calculated The distribution for the expected temperature is obtained from the distribution of the fused temperature at the previous instant and the distribution of the fused trend at the current instance The distribution of the expected value is obtained from the formula Pi pity At dt Where P Fuzzy distribution of the expected measurand value the i instant _ Fuzzy distribution of measurand at 1 1 instant dP dt rate of change trend At change in time The above equation can be considered as a fuzzy arithmetic operation with the result being the distribution for the expected temperature The distribution obtained for the expected measurand value is normalized Using this measurand estimate and the distrib
250. xpected of Charges between charges SCR SP 1500 211 00 50 00 3 00 Figure A 34 Setup Charges 192 193 This is designed to setup the number of charges that were the furnace dialog is as shown in Figure A 34 The the user can define the number of charges average time between charges CMR SCR and Expected C SP in this dialog 4 1 1 6 2 3 3 Planner setup gt Planner Load Existing Plan Create Emergency Behavoir Done Setting Up melt rate 3 1 00 En 01 00AM 01 30AM O200AM 02 30AM 03 00AM 0330AM 04 00AM 04 30AM 04 59AM 01 01 01201 01201 01201 01201 01201 01201 01201 01 01 Time Pyrometer Temperature _Plot 0 2000 00 1500 00 1000 00 500 00 0 00 1 1 1 1 1 1 1 1 01 00 01 30AM 02 00AM 02 30 03 00AM 03 30AM 04 00AM 04 30AM 05 00 AM 01701 01 01 01201 01201 01201 01201 01201 01201 01201 ZndPyometer Pott ERE 0 00 01 00AM 0 30AM 0200AM 02 30AM 01 01 01 01 01 01 01 01 04 30AM D5DODA 01 01 01 01 04 00 01 01 03 30 AM 01 01 03 00 01 01 Time Figure A 35 Planner Setup The planner setup VI produces the dialog in Figure A 35 to setup the planner Highlight on Create New Heat Plan then click on Run to create a new plan The dialog in Figure A 36 is used to add a new plan User can add the start and stop time the star

Applied to and Demonstrated on Cupola

Contents

Download Pdf Manuals

Related Search

Related Contents