Home

Enhanced Control by Visualisation of Process Characteristics: Video

1. 100 200 300 400 500 200 300 500 700 Samples sampled every two seconds Samples sampled every two seconds a Using fixed background b Using customised background FIGURE 7 3 1 A comparison between valve position signal red flow signal green and extracted flow signal blue during Andreas test 7 3 ANDREAS TEST PROPERTIES 65 Another test was made to check if choking the flow could be noticed in the data extracted from the film sequence see Figure 7 3 1 for a simple comparison Both the fixed and customised backgrounds follow the valve position roughly and so does the flow signal We can see that the customised background fails slightly when the valve is totally closed On the other hand we can see that the customised background based algorithm behaves in the same way as the valve position signal during the third and fourth valve closings in contrast to the fixed background based algorithm which shows different behaviour 02F AN 0 25 f YA 63 sec 80 sec 0 3F YA 4 0 3 Ti E Correlation coefficient b 2 6 elation coefficient b o Z e wae aes x 04h 4 04b a 0 5 4 0 5 X 4 Z6 sec YA sec i i i A i i i i f i fi i 0 6 i i i i 500 400 300 200 100 0 100200300400 500 500 400 300 200 100 0 100 Time shift seconds Time shift seconds i 200 300 400 500 a With contr
2. 0 8 des 0 85 3 ser ot ost 4 oah j 0 4 A A E NIE Ny 3 0 24 il H 3 0 2 i H N A J 3 8 1 I I IN hh ith al F TER N SWAI Wi h n AR am Ain Ji M i W EL IJI YAWA YA ur i IV 4 H il ANA V i lf V of 02 4 0 2F H V W Vv y Vy E 2800 00 7500 300 T G w 200 300 200 500 ee 00 7300 200 Ta x w 200 300 700 500 a Control pressure first logging b Control pressure third logging i ja a 0 3 4 och 125 sec q 02 Mi N wie pls ost IMI I A TL A Ni L I y o2 V A A f Si US i 3 hath H VY Fy I LT i i y h INLINE VAA INDA nae Vwi AM ty vn M cl an IEIR z VM 3 EH by YAI VARA V gt i HIH zA 0 6 0 5 sec 4 08F U 2 st 800 00 300 200 a 00 0 100 200 300 200 500 ioo 00 00 00 1 00 v 1 00 200 300 700 500 Time shift seconds Time shift seconds c Control flow first logging d Control flow third logging 0 4 0 6 0 37 GO sec 7 Ps a 0 24 4 N Hf ell I A ie nett tefl i if NI UA i H i r YA i E ETa YI A OVE aa od g h i I II W E of y LAJA I peu ll I m i N OA Sa NV 1 WN A 8 Vv WI YA i Y aat Hg fed oo 8 ea i o3 F 7 V 0 4 V 0 6F IF 0 5 k 4 set 805 00 300 00 1 0 100 200 300 200 500 0800 00 00 00 1 00 1 00 300 300 700 500 ahi 5 sonds Timi sani
3. FIGURE 2 3 1 Coal powder injection control Flow Direction Coal Pipe 1 Concentration Velocity Mass Flow 1 Weight Corrected E9 Mass Flow 1 Mass Flow 2 Weight Multiplicator Mass Flow 3 Slag Vessel Weight Coal Vessel Weight FIGURE 2 3 2 A principle scheme for pulverised coal mass flow calculation feedback The control signal from each of the PID controllers is used to control a valve placed before the flowmeter on each pipe A control based on these terms can not be perfect having in mind the bad quality of the resulting measured calculated signals Depending on how the multiplicator changes the control signal will behave differently We will show later in Chapter 4 1 that some control signals have a strange behaviour 2 4 Video Surveillance The conditions surrounding the steel making procedure are rather rough The very high temperature of the flame and the high brightness from inside the blast furnace make life hard for those who want to control or study this process The existing cameras at Mefos can not handle the incoming high light intensity and they need to be protected from the heat A damping filter has to be employed to make 18 2 PROCESS DESCRIPTION the video picture viewable The filter itself is a dark green piece of thick glass that is placed a bit away from th
4. P i j p fe Hue a RGB model b HSI model FIGURE 5 1 1 RGB and HSI model representation 7 a Full color b Red compo c Green compo d Blue compo nent nent nent e Grey scale f Hue compo g Saturation h Intensity nent component component FIGURE 5 1 2 An image example with its decompositions in RGB HSI and grey scale matrices every matrix containing the values for one of the colours In our case every spot in a matrix holds a value representing the intensity of the particular colour For each spot in every of the three matrices we had eight bits available allowing 256 different values between 0 and 255 where 0 represents the lowest intensity and 255 the highest Using MATLAB the grabbed images could be transformed into 5 1 IMAGE DECOMPOSITION 37 this form Now further operations could be applied to them in order to retrieve vital information MATLAB allows also a transition from RGB to HSI and further manipulation of images As an example a nice picture taken outside Lule University of Technology representing the university s logotype engraved in a block of ice Figure 5 1 2 has been separated into its RGB components We can also see the same image in grey scale and its HSI components Dissection performed on this image will be used for a quick comparison to our grabbed images a Full color b Red compo c Green compo d Blue compo nent nent nent ct e Grey scale
5. To Linda and my parents Igor Abstract This master thesis based on work performed at Lule University of Technology in cooperation with Mefos is about measurement of pulverised coal flow injected into a blast furnace compensating for some of the usually used coke Coal is drawn from an injection vessel and transported under pressure with the help of nitrogen gas to a blast furnace It is blown through pipes to the tuyeres where it is injected into the iron making process Irregular coal supply to the furnace has bad influence on the quality of the produced iron so reliable control is needed In controlling the flow it is of great importance that the on line flow measurement is accurate Enhancing the existing measurement would be beneficial for the quality of the produced iron Therefore new means of blast furnace process surveillance and flow measurement using cam eras and image processing are studied The idea behind camera surveillance is also beneficial for estimation of other process parameters The main goal is obtaining relevant information from image data in order to estimate the pulverised coal flow Methods for achieving this are investigated and discussed A comparison to old measurement data is made Also validation of data retrieved with the help of image processing is mentioned It has been shown that video monitoring in conjunction with image processing is a feasible option when it comes to coal flow estimation The
6. f Hue compo g Saturation h Intensity nent component component FIGURE 5 1 3 An image taken with a green glass with its decom positions in RGB HSI and grey scale Now decomposing an arbitrary image from inside the blast furnace taken with a green glass in front of the camera gives images in Figure 5 1 3 For RGB the red and green part look fine whereas the blue one looks different due to the characteristics of the filter For HSI the hue and saturation buffers are very noisy and therefore hard to deal with The intensity component is quite alright though Hue and saturation can be compared to the ones in Figure 5 1 2 where no noise is present Finally the same was done to an arbitrary chosen image from the film sequence filmed with a transparent glass Figure 5 1 4 Blue component of the RGB looks better here than in the previous series It would therefore be advantageous to use transparent glass Evidently being able to use three channels instead of two would increase the redundancy when approximating the pulverised coal flow A restriction here are the cameras used The very bright light makes even the better cameras saturate That is clear in the saturation component in Figure 5 1 3 the lower area of the interior of the blast furnace and in Figure 5 1 4 it is visible on the edges around the peek hole Fortunately strictly measuring the flow should not be affected too much by cameras saturating for pixels surrounding the plume
7. 50 50 100 150 200 250 200 150 100 50 50 100 150 Time shift seconds e With flow signal First recording o Time shift seconds f With flow signal Third recording FIGURE 7 2 2 Correlation between the extracted flow signal and different measured signals 64 7 DATA EXTRACTION AND VALIDATION 0 3 T T 9 sec 0 25 4 0 2 2 0 15 4 Die A 0 05 4 Correlation coefficient 0 05 4 50 0 50 100 Time shift seconds 0 15 L L 200 150 100 150 FIGURE 7 2 3 Correlation between the extracted flow signal and the filtered flow signal from the third video recording now we were looking forward to see a higher correlation with the current data Here we have a problem since we know that the current measurement can not be trusted and the other measured data are related to the flow measurement A warning finger should be raised whenever there is slag injected together with coal because then the flowmeters really misbehave and do not serve their purpose any good Slag and coal powder can not be separated in our images with the current quality We have only seen slag injected from behind a green glass Adjusted levels a o ro 7 3 Andreas Test Properties gt IN F RT TIR ATE i d d W Lh Adjusted levels w
8. A tool to simplify the routine work in MATLAB had to be developed We found a tool called Danalyzer that was developed during another project 4 at Lulea University of Technology Danalyzer was in its early development stage version 0 1 and needed some work to make it suitable for our needs After further development we reached what we called version 0 2 A screenshot is in Figure 4 2 1 The Danalyzer essence is its portability and the ability to use for different purposes whenever there are signals involved that need to be analysed Danalyzer should combine ease of use with functionality and eliminate the tedious work to produce needed plots and perform data analysis Danalyzer today is capable of viewing and manipulating data in different ways the main features are Viewing of signals Browsing through the signals to view specific parts of them Several plots on one window Multiple windows with different plots simultaneously Analysis FFT PSD removal of trends Adding and subtracting signals Decimation of the signals Simplicity in creating data files Possibility of zooming and gridding Loading of signal sets Dnalyzer files version 0 1 URL http mir campus luth se washers work danalyzer CHAPTER 5 Image Processing In order to be able to evoke any useful information from the images we have sampled we need to know more about them A need for a close examination of the images general properties is as important as
9. Description First find the interesting area in the image then do some edge detection and morphological operations NSee also DYNCROP IMBACK function bg x y c r dynbg im th fim medfilt2 im Find where the interesting area is fim x y c r dyncrop fim if nargin fim edge fim 255 mean2 fim 4 7 255 else fim edge fim th end bwmorph bwmorph fim dilate 2 shrink 3 fim imconnect fim im bwfill fim holes A 11 bgmulti m Purpose Find the background in an image Synopsis y bgmulti im s n c layer accept YDescription Find the background in a series of images based on several consecutive images im is the image series s is the start point n is the number of images to detect their background and c is the wanted color layer layer is the number of 4 consecutive images to use and accept is the threshold for edge acceptance which is less than or equal to layer NSee also IMBACK function y bgmulti im s n c layer accept ok 0 for i O n 1 while ok lt layer e ok 1 edge im stok i c 255 mean2 im s ok i c 2 255 ok ok 1 end tmp sum e 3 gt accept 1 sti c bwmorph bwmorph tmp dilate 5 thin inf ok ok 1 for j 1 ok A 13 ALGOX M e j e jtl end end A 12 imarea m Purpose Calculate
10. Freguency c Pressure FIGURE 4 1 4 PSD plots for different measurements in pipe 1 second occasion Freguency is in Hz Nezt step in the study of the collected signals was to try different things like for example addition subtraction and other such operation on the different signals One of the more interesting things tried here was frequency analysis Frequency contents of our signals were analysed using Danalyzer just as all of the above analysis to make life simpler and save some time Studying signals collected on the 4th of June we found the same dominant frequency for pipe 2 and 3 in control and flow signals as shown in Figure 4 1 3 for pipe 2 This could point to poor control affecting the flow and making it fluctuate or the opposite if we assume that the controllers are the same in all the pipes but the flow measurement device behaves badly in different ways For our purposes though this was an excellent opportunity to check if the same frequency could be found in the processed images This matter will be investigated in Chapter 7 2 where all the data extraction is performed and thoroughly discussed We could not identify such control problems for pipe 1 see 32 4 DATA PROCESSING Figure 4 1 4 nor could they be found in the signals collected later which should not be misinterpreted as the control signal being unsurpassed
11. Johansson Andreas 1999 Model based leakage detection in a pressurized system Sweden Lulea University of Technology Licentiate thesis 1999 37 ISSN 1402 1757 7 Lorenz Juergen 1999 CVC Color Software tool for color recognition Stemmer Imaging GmbH URL http www cvc imaging com Cvc_ Tools Bildverarbeitung Color Farbmodelle_e htm 8 Low Adrian 1991 Introductory Computer vision and image processing England McGraw Hill Book Company UK Ltd ISBN 0 07 707403 3 9 Matrox Image Processing group 1998 Matrox Meteor Installation and hardware reference Canada Matrox Electronic systems Ltd Manual No 10529 MT 0110 10 McManus George J 1994 Replacing coke with pulverized coal Iron Age New Steel New York vol 10 issue 6 p 40 ISSN 1074 1690 11 O Hanlon J 1993 Injection of granular coal into the blast furnace Steel Times vol 221 issue 12 p 508 509 ISSN 0039 095X CODEN STLTA3 12 Ondrey Gerald Parkinson Gerald Moore Stephen 1995 Blast furnaces make way for new steel technology Chemical Engineering H W Wilson AST vol 102 p 37 ISSN 0009 2460 13 Oshnock T W 1995 Pulverized coal injection for blast furnace operation Iron amp Steel maker I amp SM vol 22 issue 4 p 49 50 ISSN 0275 8687 14 Piersol Bendat 1971 Random data Analysis and measurement procedures USA John Wiley amp Sons Inc ISBN 0 471 06470 X 15 Porat Boaz 1996 A course in digital si
12. considering computational efficiency To start with the mean values for the pixels representing the coal plume are found for both the horizontal and the vertical di rection The covariance matrix and the eigenvalues are calculated This is done in order to rotate the detected coal particle cloud with such an angle that it is standing upright i e the lower end is where the coal comes out of the tuyere see Figure 6 4 1 Now we can apply the actual algorithm itself The mean values for the pixels in the horizontal direction are calculated The volume of the left part and the right part are calculated separately and then added together to embody the total approximated volume This is done as an attempt to describe the body of revolution for the plume Calculating each half of the coal particle cloud is done in the following way First the number of pixels in each row on the left side i e left of the calculated middle line of the plume are stored in a vector Then each value in the vector is treated as a radius of the plume s left half at that particular row Using these radia we can calculate areas of as many half circles as there are values in the vector Summing all the areas we obtain a volume for the left half of the pulverised coal cloud Now the same procedure is applied to the right side of the plume In the end the two volumes are added to form the final result the plume s approximated volume Algorithm 14 illustrates the procedure 6 4
13. done donet1 if con x1 if v x1 con x1 1 else v x1 9 end end if con x2 if v x2 con x2 1 else v x2 9 end end sel find con end J line2pixel r e w c e w A 9 DYNCROP M ull sparse y x ones size x size i 1 size i 2 i A 7 imfilter m Purpose Filter out isolated pixels in an image Synopsis f imfilter i y Description The binary image p is converted into double if needed and the image will be scanned with a y by y square that filter out isolated spots according to y y is the sized of isolated spots NSee also FILTER2 BWMORPH IMREG function f imfilter i y f isa i uint8 i double i end m n size i for d 1 length y x y d f zeros m 2 x n 2 x f x 1 m x x 1 n x i c r v find f x0 ones 2xx 1 x0 2 2 x 2 2 x 0 for j 1 length c ci c j ri r j i x ci x ri x ri x sum sum f ci x ci x ri x ri x x0 0 f ci x cit x ri x rit x end i f x 1 m x x 1 n x end f i A 8 imback m Purpose Get an image background and make it black Synopsis bg imback i Description bg is a binary image where the black part represent the background and the white is off course the foreground The parameter t is the threshold used to detect the background edges NSee also BWMORPH IMCONNECT BWFILL EDGE function bg imback i t Try to Find the optimal threshold for each image ima
14. 1 INTRODUCTION obtained with a colour camera in order to use several independent estimation chan nels In this report we will focus on the first part i e coal flow but our results will hopefully be useful for the other targets too CHAPTER 2 Process Description We had an opportunity to work on LKAB s experimental blast furnace at Mefos 17 where we had a setup of cameras and a possibility to collect needed data In this chapter we will briefly describe the different parts of the plant and the coal injection part of the process FIGURE 2 0 1 A schematic overview of the plant at Mefos 2 1 The Blast Furnace The blast furnace at Mefos is marked with an A in Figure 2 0 1 A picture of the actual installation is in Figure 2 1 1 It has three tuyeres and a diameter at the tuyere level of 1 2m The working volume is 8 2 m3 The hot blast is produced in pebble heaters capable of supplying 1300 C of blast temperature The furnace is designed for operating with a top pressure of 1 5 bar It has a bell type charging system without movable armour The coal injection system see B in Figure 2 0 1 features individual control of coal flow for each tuyere Gas cleaning system part D of Figure 2 0 1 consists of dust catcher and electrostatic precipitator Material is transported to the top of the blast furnace through C as shown in Figure 2 0 1 A tapping machine with drill and mud gun is installed
15. 3 Approximated Shape Estimation Yet another algorithm for ap proximating the volume of the pulverised coal injected at any point in time and seen with the help of cameras is the algorithm we chose to call Approximated Shape Estimation Algorithm 15 6 4 ESTIMATION OF PLUME S VOLUME 57 a Before b After FIGURE 6 4 1 Plume image before and after rotation Algorithm 14 Rotalgo 15 16 17 18 19 20 21 22 23 rotalgo im n length im gt 0 740 y 0 c lt 0 vol 0 for i 1 to rows im for j 1 to columns im if im i j then y ey i TELE J Zaa i e end end x y m e gt cov mx m eigv calculate the eigenvalues of cov alpha 220 arctan f eigv 1 rim rotate im by alpha dias lt count the pixels of rim row wise for i 1 to length dias NED vol vol T 458 2 end end 58 6 ALGORITHMS Algorithm 15 Ellipalgo ellipalgo im 1 minaz find the minor axis length in im 2 majax find the major axis length in im 3 vol lt 3 T majaz minaz Approximated Shape Estimation is based on the observation of the plume s shape We could establish that the shape of the projection of the coal particle cloud was very close to elliptic Now assuming that our three dimensional object is symmetrical along its axes i e rugby ball look alike we can easily compute the volume of such a three dimensional body Finding both axes lengths for ou
16. Colours 0 00 0 10 Colours 0 10 0 20 Colours 0 20 0 30 Colours 0 30 0 40 Colours 0 40 0 50 Colours 0 50 0 60 Colours 0 20 0 30 Colours 0 30 0 40 Colours 0 40 0 50 Colours 0 50 0 60 Colours 0 60 0 70 Colours 0 70 0 80 Colours 0 80 0 90 Colours 0 90 1 00 Colours 0 60 0 70 Colours 0 70 0 80 Colours 0 80 0 90 Colours 0 90 1 00 e Blue component image without coal f Blue component image with coal FIGURE 5 2 5 RGB components based on images taken with a green glass Thresholded between 0 and 1 with steps of 0 1 44 5 IMAGE PROCESSING Full colour Quantized colours Colours 0 00 0 10 Colours 0 10 0 20 Full colour Quantized colours Colours 0 00 0 10 Colours 0 10 0 20 Colours 0 20 0 30 Colours 0 30 0 40 Colours 0 40 0 50 Colours 0 50 0 60 Colours 0 20 0 30 Colours 0 30 0 40 Colours 0 40 0 50 Colours 0 50 0 60 Colours 0 60 0 70 Colours 0 70 0 80 Colours 0 80 0 90 Colours 0 90 1 00 Colours 0 60 0 70 Colours 0 70 0 80 Colours 0 80 0 90 Colours 0 90 1 00 a Red component image without coal b Red component image with coal Full colour Quantized colours Colours 0 00 0 10 Colours 0 10 0 20 Full colour Quantized colours Colours 0 00 0 10 Colours 0 10 0 20 Colours 0 20 0 30 Colours 0 30 0 40 Colours 0 40 0 50 Colours 0 50 0 60 Colours 0 20 0 30 Colours 0 30 0 40 Colours 0 40 0 50 Colours 0 50 0 60 Colours 0 60 0 70 Colours 0 70 0 80 Colours 0 80 0 90 Colours 0 90 1 00 Colours 0 60 0 70 Colours 0 70 0 80 Colours 0 80 0 90 Colours 0 90 1 00 c Green component
17. DATA CHARACTERISTICS 63 0 5 0 5 0 4 13 sec 0 4 za ji 0 3 4 o3 a H 0 2 4 8 0 2 K f I a AA a as wih ll MJ otf J g ot p f reg Wai ill ju ily 5 V WW Wa I I IA JI li VA hy lilly va An NL ey Ji IN 7 i i y NN MM wt alee Wy IM We E JI Wa 4 zat J Lio aso o 50 0 50 100 150 200 250 ao Te 00 30 0 50 100 150 Time shift seconds Time shift seconds a With control signal First recording b With control signal Third recording 0 5 0 5 0 4 4 0 4F 11 sec LI 0 33 3 0 3F i i 0 2 q 02 A fork 8 kana Miji M SE dh E nt Ny a h 0 1 oH I i I nil sl slat i ki ul bt 4 FAN IN aul IN NaN TT lL a ly wad all ol Il afl i ji TI ith of inant JA Wid mi Mt ya HT ul Hi Hl r Pi W 01 F fi 4 if W j 030 00 0 o 50 700 750 200 250 Cio iso 00 E 0 30 100 150 Time shift seconds Time shift seconds c With pressure signal First recording d With pressure signal Third record ing 0 25 0 25 3 sec 0 2 i J 025 N 0 15 0 15 il ATY otf FH 5 4 oH ri yA f l 1 Wi I i z fr 0 05 N A he frn 4 005 f rd fos i li EV Ila WN fd i ih it I iji Ma AN AA I 3 PA or iJ RONNI M STEN AMT a T ey WW I 3 I we WEE WA We a WN 0 05 VING Vee ty g os J li I al i Ia cash _ 0 15 4 0 15 fd 0 2 4 0 24 H ae i i i bbe i i i Pee i i i YR i 200 150 100
18. METEORSFMT amp ifmt mem mmap caddr_t 0 off fb size PROT READ MAP PRIVATE dev off t 0 gettimeofday amp time amp zone ot double time tv sec time tv usec 1e6 pause fprintf stderr Press Enter to start grabbing getc stdin gettimeofday amp time amp zone for j 0 j lt noframes j 1 while double time tv sec time tv usec 1e6 lt ot gettimeofday amp time amp zone ioctl dev METEORCAPTUR amp cap fprintf stdout Image 03d sec d usec d n j time tv sec time tv_usec mmbuf unsigned char mem off frame_offset 0 85 86 B C CODE ptr buf unsigned char malloc size 3 for i 0 i lt size it 1 ptr 2 xmmbuf ptr 1 mmbuf ptr mmbuf ptr 3 mmbuf sprintf str pic 03d pnm j if fp fopen str w fprintf fp P c n Lulea University of Technology The Visualization Project n d d n255 n 26 COLS ROWS fwrite buf 3 sizeof char size fp fclose fp free buf ot double time tv sec time tv usec 1e6 pause munmap mem off fb_size close dev exit 0
19. N amp w f E uu tt pe mm a Extracted flow signal b Collected flow signal FIGURE 7 2 1 FFT of the extracted flow signal and the collected flow signal is obtained with the control signal and the worst one is found with the flow signal measurement The right side of the figure representing the third video signal shows exactly the same trends except for the absence of the Mefos carrier frequency which is not present in the collected signals on this occasion and also the time delays differ The delay was found to be very small a few seconds at the most Our flow estimation is delayed by approximately 11 13 seconds compared to the control signal The pressure on the other hand proceeds our extracted data by 9 11 seconds As a result the correlation peak differs by 2 seconds between he pressure and the control signal Comparing Figure 7 2 2 a nd b with Figure 4 1 5 c and d we notice that there is no clear negative correlation present in the first ones that is because our flow estimation has no direct feedback to the controller When it comes to the flow signals correlated with our estimated flow we see a clear difference in the flow measurement quality In the first recording we can register the flow 3 seconds before the flowmeter while in the third recording we can do the same 9 seconds before the flowmeter The negative correlation could be related to the different delays for the control flow and control est
20. The blast furnace is well equipped with sensors and measuring devices and an advanced system for process 13 14 2 PROCESS DESCRIPTION control Probes for taking material samples from the furnace during operation are being developed FIGURE 2 1 1 The blast furnace body LKAB use the furnace primarily for development of the next generation of blast furnace pellets The furnace performance shows that it is a good tool for other development projects An important area is recycling of waste oxides Injection of waste oxides is another research area as well as injection of slag formers The interesting parts of the plant are presented in Figure 2 1 2 Air Lock Vessel Blast Furnace Tuyeres Slag Vessel Coal Injection Vessel FIGURE 2 1 2 The most relevant parts for this project of the experimental blast furnace at Mefos 2 2 COAL INJECTION 15 FIGURE 2 2 1 Control screen for coal injection at Mefos 2 2 Coal Injection The coal injection arrangement Figure 2 2 1 consists of two coal vessels and three pipes each ending with a tuyere The three tuyeres are evenly spaced around the blast furnace as shown in Figure 2 2 2 Pulverized Coal Pipe 3 Video Camera Blast Furnace Web FIGURE 2 2 2 The three tuyeres are surrounding the blast furnace with the surveillance cameras supporting framework Looking again a
21. and hopefully will never be specially when she he deals with automatic control Free Software is widely available reliable supported and sufficient in most cases We used RTLinux as a platform for digitising the video tapes we recorded Using an ordinary S VHS video player a common PC old timer equipped with RTLinux a Matrox Meteor frame grabber card and a simple grabbing program see the source code in Appendix B 1 the work was accomplished The task here is to convert a full motion video PAL signal Phase Alteration Line running at 25 frames per second into single frames stored in a digital format The first step in the process of converting an analogue signal into a digital repre sentation is sampling This is accomplished by measuring the value of the analogue signal at regular intervals called samples These values are then encoded to provide a digital representation of the analogue signal The power spectral density PSD of the flow signals we have collected shows that in the worst case a sampling time of two seconds is sufficient Remember that we are not sampling to control we are just trying to recreate the flow signal in someway Two seconds might sound too fast but some of our flow signals have a significant frequency peak caused by bad control The same frequency peak has been found in the pressure and control sig nals See Figure 4 1 3 for a closer look at PSD signals related to pipe 2 taken form the second data collecting oc
22. cameras used for coal powder surveillance is that most cameras auto adjust to current conditions This means that a camera is constantly adjusting depending on the amount of coal present in the image due to noticeable changes in brightness This phenomenon will evidently harm a measurement of temperature which is highly dependent on colours being compared between two different frames It could also harm the coal flow estimation Having a fixed colour reference is valuable Even when no real changes occur in the colour of the flame an auto adjusting camera could perceive two different colours A remedy for this is if possible disallowing auto adjustment 8 2 SUGGESTIONS 71 Shutter time has to be constant and no changes made to the lens aperture while in operation Further if we are to do all computations needed in real time the most feasible solution is using digital cameras to avoid distortions in image quality by eliminating digitising and sampling effects It is of great help when comes to squashing the time needed to produce useful data from images 12 bit cameras which have appeared on the market could enhance the resolution and using several cameras for each tuyere would allow a more accurate measurement of the volume of the coal particle cloud More than one camera for each tuyere will complicate the problem therefore it is advisable to improve the algorithm developed for one camera unless a more exact coal estimation is needed Find
23. controllers are not working as they should Both the control signals 2 and 3 reach their lower bound frequently and are varying quite a lot which indicates bad control Well with a risk of being nasty we would say very bad control Examine Figure 4 1 1 for clarification Could this be a result of our non galvanically isolated data acquisition equipment A comparison between the control signals from the first and the second data collecting occasions which we carried out using the same cabling shows that the control signals do not saturate as often as the one in Figure 4 1 1 does This makes us conclude that we are not responsible for the bad control something else went haywire The controller for pipe 1 functions somewhat better at that given time Pressure as well as weight of the lock vessel remain constant for the whole time period For the injection vessel pressure we notice a sinusoidal looking curve with a diminishing amplitude whereas its weight is steadily falling because coal powder is persistently drawn from it and transferred to the blast furnace and no refillment of the coal vessel has taken place during logging time The constant negative slope of the signal suggests an even supply of coal to the blast furnace The second time we went to Mefos to collect signals we managed to get all thirteen of them i e the same twelve as before and also the pressure in injection pipe 1 Pressure signals for the three pipes look very much the same and t
24. ey yo MM MARI l i 0 5 Fi 4 er 4 7 sec a 0 6 i i i i i i i i i 500 400 300 200 100 0 100 200 300 400 500 500 400 300 200 100 0 100 200 300 400 500 Time shift seconds Time shift seconds a The valve after the pressure meter b The valve before the pressure meter FIGURE 7 3 3 Correlation between the extracted flow signal and the pressure signal during Andreas test the same before the pressuremeter will decrease it whereas the flow will always decrease In this case we need to separate the Andreas test into two cases which are outlined in Figure 7 3 3 The first shutting of the flow after the meter gave a strong negative correlation with 11 seconds delay before we discovered that some thing had happened in our images The second where shutting of the flow was before the meter led to a positive correlation after 37 seconds Seemingly the dif ference in the delay time is simply depending on the distance the coal has to travel before reaching the tuyere s mouth and get on the tape but we still think it is too long 0 6 T T T T T T T T 0 6 9 sec 2 sec oat J oak mR J f 0 2 Be i 4 0 2 j i 3 s r A 2 af 0 44 4 0 44 Z a 0 6 8 sec H i 4 o6 fi 49 sec 08 0 8 i fi f fi i i i i fi i f fi fi 200 150 100 50 0 50 100 150 200 250 300 200 100 0 100 200 300 Time shift seconds Time shift seconds Co
25. function j imconnect i When you are in trouble you need at least two objects in the image to connect 4 round length i 2 0 round size i 1 2 0 e v u imends i Remember that all the endpoints are not real endpoints r c find i dais the distance matrix between the endpoints for z 1 length e Osqrt r e r e z 72 c e c e z 72 end 77 78 A MATLAB CODE ul u count zeros 1 max u losers dfind d 0 mam d 1 m 1 length e on zeros 1 length e sel w ines length unique u done 0 while find con 0 mi setdiff m union sel losers bad 1 if isempty m1 break end while bad vi r1 c1 mip d m1 m1 xi m1 r1 x2 m1 c1 hif length find con 0 if length unique u ok 0 hhhhh bad 0 the next two rows is used if you want to connect every end in the last segment fucount ui x1 ucount ui x1 1 fucount ui x2 ucount ui x2 1 else ok u x1 u x2 end if ok m2 setdiff m1 x1 if isempty m2 break end vi r1 c1 mip d m2 m2 m3 setdiff m1 x2 v2 r2 c2 mip d m3 m3 if vi lt v2 mi m2 else mi m3 end else bad 0 end end w w x1 x2 ucount u1 x1 ucount ul x1 1 ucount ul x2 ucount ul x2 1 if ucount ul x1 losers losers find ui u1 x1 end if ucount ul x2 losers losers find ui u1 x2 end if lines done 2 up f ind u u x1 u up u x2 end
26. image without d Green component image with coal coal Full colour Quantized colours Colours 0 00 0 10 Colours 0 10 0 20 Full colour Quantized colours Colours 0 00 0 10 Colours 0 10 0 20 Colours 0 20 0 30 Colours 0 30 0 40 Colours 0 40 0 50 Colours 0 50 0 60 Colours 0 20 0 30 Colours 0 30 0 40 Colours 0 40 0 50 Colours 0 50 0 60 Colours 0 60 0 70 Colours 0 70 0 80 Colours 0 80 0 90 Colours 0 60 0 70 Colours 0 70 0 80 Colours 0 80 0 90 Colours 0 90 1 00 e Blue component image without coal f Blue component image with coal FIGURE 5 2 6 RGB components based on images taken with a transparent glass Thresholded between 0 and 1 with steps of 0 1 CHAPTER 6 Algorithms Here procedures for extracting information from video sequences will be dis cussed as well as a few useful algorithms designed to calculate changes in the pul verised coal flow Beginning with finding a static background used to mask out the coal plume we will move through an alternative method of finding a background for every image finding the coal plume and last we will introduce three algorithms for continuously approximating the volume of coal material injected i e a possible replacement measurement for the currently used flow measurement We will also try to point out possible drawbacks of the algorithms Starting with an image of coal injected into a blast furnace we need to end up with data that can be used in our steel making process We n
27. images include potential information for other purposes like determining the temperature of the flame and how well the coal is distributed inside the blast furnace This would solve some of the problems and eliminate obstacles caused by the nature of the steel making process Preface People have always asked us Why study automatic control The real question is Why not There are not many fields that affect our modern life as much as automatic control does of course mathematics and possibly physics are cornerstones in any nutritious study They are hard to compete with Applied science in all its forms is the way to go to enhance products and tools that are essential in today s society The achievements in the field of automatic control are surrounding us in our everyday lives no matter if we like it or not A lot of things out there are already done many more are waiting to be done There is also a lot of fine tuning to take care of which is sometimes even more challenging We wanted to be a part of this evolving development We want to thank Anders Grennberg without him we would not be closing loops these days This work is a part of a bigger project with involvement from the industrial and research world backed up by PROSA Centre for Process and System Automa tion Our master thesis was carried out at the Department of Computer Science and Electrical Engineering Control Engineering Group at Lule University of Tech nology in cooper
28. in June However in November a decision was made to try another slightly more advanced camera for filming the injection at tuyere 2 The camera used was also a Panasonic camera from the 400 series The difference was mainly the larger dynamic range and employment of a new Panasonic technique known as SuperDynamic Good reference manuals with the camera specifications could not be found The necessity of using a more advanced camera occurred to us when trying to record a film sequence without the dark green protecting glass Operators knew from experience that the existing cameras were not able to handle the intake of very bright light when the protecting glass was removed Removing the dimming filter transparent glass was mounted in its place The new camera used was performing better than the old one but 3 4 VIDEO DIGITISING 23 unfortunately it also reached the saturation point due to the extreme brightness of the light from inside the blast furnace Another problem discovered during the signal collecting session in November was the dynamic adjustment of the range This caused changes in the background of the picture depending on the amount of visible coal When there is a lot of coal the picture is percepted as darker by the camera soit adjusts to a darker picture when there is no coal the picture is brighter and again the camera adjusts accordingly As a result of these constant adjustments we get a background for the pulverised coal cloud th
29. of the image and finds the interesting area in it before further processing Algorithm 9 Dynbg dynbg im th fim median filter im ifim find the interesting area in fim find th for ifim if th is not given e edge detect ifim according to th em enhance e using morphological operations cim lt connect the edges in em if needed bg fill cim s inside with 1 s TYSEN tee Coniko To summarise The best approach in order to find the background is stacking several consecutive images crop the resulting image edge detect it improve the edges by morphological operations connect any discontinuous edges and finally fill the inside with ones The result typically although exceptions exist looks like the one in Figure 6 1 5 which can be compared to Figure 6 1 1 FIGURE 6 1 5 A background using the suggested algorithm approach 6 2 Finding the Coal Plume Having nice background we can effortlessly find the coal plume What we basically have to do is just to subtract each image from its background and then 6 2 FINDING THE COAL PLUME 53 Original image Background image Original image Background image Masked image Plume image Masked image Plume image 3 a Based on an image taken with green b Based on an image taken with trans glass parent glass FIGURE 6 2 1 Plume eztraction threshold it In result we obtain the coal plume This operation is illustrated in Figure 6 2 1 There is some decision making prob
30. round ri round mc cpiz append c ci rpix append r b Algorithm 5 Imback mw NA imback im th eim edge detect im according to th mim clean up eim using morphological operations cim connect the objects mim bg fill cim s inside with 1 s Algorithm 6 Bgmulti oo 11 12 13 SONAR UV I H bgmulti ims layers ok 0 tmp 0 for i 1 to length ims layers 1 while ok lt layers tmp ims 1 end 1 end ok iJ tmp ok ok 1 end tmp mimp layers eim lt edge detect mtmp eims 1 end 1 end i lt clean up eim using morphological operations ok ok 1 tmp tmp ims 1 end 1 end i end 6 1 FINDING THE BACKGROUND 51 made about how many images to stack and when a pixel is counted as static or not Final polishing is done with some well chosen morphological operation sequence Algorithm 7 Bgmultiedge bgmultiedge ims layers accept 1 ok 0 2 for i 1 to length ims layers 1 3 while ok lt layers eims 1 end 1 end ok 1 edge detect ims 1 end 1 end ok i ok ok 1 end tmp lt sum eims layers pixel wise bg tmp gt accept 9 bgs 1 end 1 end i clean up bg using morphological operations 10 ok ok 1 11 eims shift left eims 12 end OO Oe OY In a well controlled plant there always exists a certain coal flow which makes this algorithm fail in finding the real background It simply assumes the areas close
31. saturation component and above 250 for the intensity component This is not true for the images taken with the transparent glass A clear impact of changing the glass can be seen which effects can not be addressed to the change of camera The intensity component graph is almost a copy of the RGB components graphs for the same images that is usually the case in a healthy image We can see where we can expect coal and where we do not A way to emphasise the difference between the coal and no coal in the images is to multiply each image histogram by the difference histogram we saw in the figures above before any further image processing in order to find the plume 5 2 2 Image Threshold To see if there is any definite threshold that can separate the background from the interesting elements the coal from gases and so forth we took the pictures we used above decomposed them into their RGB components quantised them linearly with ten steps and thresholded at those steps The steps used were between 0 and 1 with a step size of 0 1 Studying Figure 5 2 5 and 5 2 6 shows that threshold values can be examined in more detail Evidently 42 5 IMAGE PROCESSING Hue Component Hue Component 1000 2000 l 7 RN M UW PAPEETE ETEA nh 0 A Veen wA ok Yes yey portland See 1000 I 4 ly 2000 2000 3000 i i i i i 4000 i i i 0 50 1 150 250 300 0 50 100 150 250 300 Saturation Component Saturatio
32. suddenly our background would not fit the picture i e a new background mask would have to be found The same applies to zooming where the operators usually have the possibility to do Of course there is a simple cure to this problem fixing the cameras while the blast furnace is in use and banning the operators from doing any adjustment Alternatively one can have a procedure of taking fresh new coal free images every time the camera setup is changed A useful background image mask that can be used to remove all the uninteresting details is in Figure 6 1 1 6 1 2 Customised Background Approach The one background approach is sufficient in most the cases but it has several limitations If a worker happened to touch the camera housing causing some trembling then the extracted data se quence should be trashed until the camera settles An operator can control the camera position focus and zoom Any of these actions and any calculations based on one fixed background is faulty Most importantly we want a system that is as easy as possible and can automate most of the work with minimum interaction from the operators No one in the industry has time to grab a coal free image every time something changes in the camera setup or its surrounding environment A need of finding the background in every image is unavoidable Knowing that the colour of the coal is the same as or at least very close to the colour of the tuyere and the surrounding area makes fi
33. the images themselves As we mentioned before we used two different types of glass as filters during the video recording We also used two different types of cameras This makes the comparison we made here below not as bulletproof as we wanted it to be We will strive to explain most of the image processing carried out by inserting nice illustrative images and plots as a complement to the explanatory text In the very beginning it might be appropriate to explain two systems for representation of colour images RGB and HSI Later we will try to visualise the image content using different techniques 5 1 Image Decomposition Any description of the human visual system only serves to illustrate how far computer vision has to go before it approaches human ability In terms of image acquisition the eye is totally superior to any camera system yet developed People are continuously trying to improve the existing artificial vision systems One of the problems to be solved is how to represent colours Several systems have been invented for this purpose Here we will only discuss a couple of the most frequently used systems RGB and HSI 5 1 1 RGB and HSI spaces First the RGB system will be presented RGB stands for red green blue Using these three colours in different amounts almost any other colour can be produced On a screen mixing is done by having three adjacent dots one dot for each of the three colours If these dots are small enough and the ob
34. the other signals The strange behaviour in the beginning has been shown to be caused by tapping of iron during logging time as stated in Table 3 Although the time does not exactly match there is no indication that it could be caused by anything else The first tapping is visible in all the signals but the second one appears only in the control signal of pipe 2 It has nothing to do with the duration of the tapping because the effect on the signals is visible during the whole tapping time What happens in the blast furnace during the tapping is a very interesting topic to look into 4 1 ANALYSIS OF SIGNALS 29 Test signal and 31F1102 detrended with the sample means removed A dtrended plot for 31P1102 with the sample means removed T r r T r r r r T r r r A fi A he i anal 0 02 m E ay i Pig TE Poe PIC Sensor value Sensor value 25 3 i i i i i i i i i i i i i i i 3200 3400 3600 3800 4000 4200 4400 4600 4800 5000 5200 3600 3800 4000 4200 4400 4600 Sample Sample a Flow signal compared to valve posi b Pressure signal compared to valve tion position A dtrended plot for TEST005 with the sample means removed Sensor value o T E as J N NY V Mw i 24 al 3 fi i fi fi fi i i 3200 3400 3600 3800 4000 4200 4400 4600 4800 5000 5200 Sample c Control signal compared to va
35. the weighted area in an image Synopsis hh a imarea i Description A dark pixel have a higher weight than a light pixel Given an intensity image the maximum and minimum intensity values 0 are calculated and the pixels values are re mapped to reflect their weight The result is the sum of the weighted pixels YSee also 4 BWAREA function a imarea i In m size i r reshape i n m 1 x y v find r i1 0 if isempty v vmin min v vmax max v il double i double vmin 1 find il lt 0 il double vmax il i1 1 0 end sum sum i1 A 13 algox m Purpose Clean up in an BW image from uninteresting objects Synopsis res algox im th xc yc Description Identify the objects in the image and see which object is closest to the given point xc yc th is used to get rid of small object it is simply the number of white pixels in those objects YSee also BWMORPH unction res algox im th xc yc im bwlabel im a max max im r 1 if ma gt 1 for i 1 ma tmp im i s sum sum tmp if s lt th im tmp 0 else y x find tmp xt mean x yt mean y r i sqrt yt yc 72 xt xc 72 end end if length r gt 0 r find r 0 1000 v o min r res im 0 else res im end else res im end 81 82 A MATLAB CODE A 14 findflame m Purpose Fi
36. to possibly be related to the equipment used Although the data acquisition module was thought to be galvanically isolated Unfortunately it was found that was not the case Neglecting that fact collected data was analysed During both those logging sessions we used the same cabling with 250 Q resistances Wanting to verify what was concluded from the data collected in June another trip to Mefos was required This time not wanting to be be blamed for disturbing the blast furnace process once again Another data acquisition card had to be used The National Instruments signal collecting box turned out to have a card supporting collection of only eight signals at a time if they were to be isolated We also changed the cables used and switched to 50 9 resistances This was found to be a good solution after an examination of what went wrong with the blast furnace process during the June sessions at Mefos Of course a limitation like this meant arisk An omitted signal could later turn out to be of great importance We had to concentrate on two of the three tuyeres specifically pipe 2 and 3 The chosen signals have looking in Table 1 numbers 2 3 5 6 8 9 10 and 11 Motivation for this choice was that mass flow was wanted together with both the control signal and pressure signal Weight of the injection vessel as well as the pressure inside it were also taken into consideration on the expense of the control pressure and flow signals for tuyere 1 Co
37. to the tuyere s mouth to be part of the background when they do not change and the sharp edge between the tuyere and the coal is gone The result is a discon tinuous edge This effect is illustrated in Figure 6 1 4 The best thing we can do is connecting the discontinuous parts using the algorithm we discussed earlier but the tuyere shape is still not perfect FIGURE 6 1 4 Multiple layer edge detection The bigger an image is the more CPU and time is needed to process it Because the nature of our images a hot spot surrounded by a static dark part we need only to process the interesting part of them The area of interest is not always in the same location in a flexible system that is why we need to find the area in every single frame we want to process This is done with Algorithm 8 This algorithm basically scans the image for light pixels The rectangular area with a lot of light pixels inside is the most interesting part of the image Combining this with the previously mentioned background algorithms will make them faster 52 6 ALGORITHMS Algorithm 8 Dyncrop dyncrop im csum lt sum of the pixels in im column wise rsum lt sum of the pixels in im row wise cth calculate a column threshold rth calculate a row threshold cint csum gt cth rint rsum gt rth cim imfrint cint SSR ON E E Algorithm 9 is a modified version of Algorithm 5 It does pretty much the same thing beside it starts with median filtering
38. when the valve is closed which is an indication of a functioning controller The controller s feedback see Chapter 2 3 for further information is a reason for the bad response time unless Mefos engineers have designed it that way That proves that at least there is a reaction from the flowmeter when the coal flow to the furnace is drastically reduced What can be considered as strange is that looking at the pressure control and flow signals for pipe 3 there are no changes that could be directly connected to the changes in the flow for pipe 2 Since the coal leaves the injection vessel and is 4 1 ANALYSIS OF SIGNALS 31 then divided in three flows one could conclude that smaller flow in one of the pipes would mean more coal in the remaining pipes if the total flow was to be the same Here we can deduce that either the total flow decreased or that the flowmeters do not work very well The PSD for 31F101 decimated2 times The PSD for 31FV138 decimated times 40 30 30 206 20 a g 2 2 i 2 E ya JUN A i o 3 gt 107 Sa Ns 20 J magy AA V Us aa 30 MAMANA 4 304 4 ing J WA is i i i i ji i i i i 0 0 05 01 015 0 2 0 25 0 0 05 0 1 0 15 0 2 0 25 Freguency Freguency a Mass flow b Control The PSD for 31P1101 decimated2 times 30 T T 20 10 g ER mi r 10 F E G oy E 304 t Wy Wa al WAY aani acted Mad 50 fi hi i 0 0 05 01 0 15 0 2 0 25
39. 2 Edge detection containing some discontinuities Different threshold values can be used to find the edges in an image In our case the most prominent edges are those between the peek hole s edges with the tuyere and the inside of the furnace but also the coal particle cloud and the furnace interior If we could find a reasonable threshold that gives us the first mentioned edges then we are done This is the same problem as the one we mentioned in the simple algorithm we started with above Tests have shown that it is possible most of the time to choose a threshold that can be used to edge detect the image and keep parts of the most interesting edges The result of such edge detection looks like 48 6 ALGORITHMS the one in Figure 6 1 2 What we need to do in order to recreate our background is retie our disconnected edge parts Algorithm 2 Imends imends im 07 11 13 1 mask 17 00 19 23 29 31 01 02 03 2 order 04 00 05 06 07 08 eim 0 irows rows im icols lt columns im for r 1 to irows for c lt 1 to icols if im r c then tmp the nine pixels surrounding im r c isum lt sum mask tmp if isum Emask then dir lt get the order of isum in mask eim r c dir end 15 end 16 end 17 end 50 00 Sy Oh ety a BONO Connecting edge parts like these is an easy task for a human equipped with a pencil but it is harder for a computer to accomplish the same thing Algorithms exist to connect parts pix
40. 6 1 2 Customised Background Approach 6 2 Finding the Coal Plume 6 3 Estimation of Plume s Area 6 4 Estimation of Plume s Volume 6 4 1 Weighted Pixel Estimation 6 4 2 Rotated Plume Estimation 6 4 3 Approzimated Shape Estimation Chapter 7 Data Extraction and Validation 7 1 Relations Between Algorithms 11 13 13 15 16 16 16 17 19 19 20 22 23 27 27 33 34 35 35 35 35 38 38 41 45 45 45 46 52 55 55 55 56 56 59 59 10 CONTENTS 7 2 Extracted Data Characteristics 7 3 Andreas Test Properties Chapter 8 Conclusions and Suggestions 8 1 Conclusions 8 2 Suggestions Bibliography Appendiz A MATLAB Code A l mam m A 2 mip m A 3 imframe m A 4 imends m A 5 line2pixel m A 6 imconnect m A T imfilter m A 8 imback m A 9 dyncrop m A 10 dynbg m A 11 bgmulti m A 12 imarea m A 13 algox m A 14 findflame m A 15 imflame m A 16 algo2vol m A 17 evalvol m A 18 countpizels m A 19 findvolume2 m Appendix B C Code B 1 ssnap c 60 64 69 69 70 73 75 75 75 75 75 76 77 79 79 79 80 80 81 81 82 82 82 83 83 83 85 85 CHAPTER 1 Introduction Heavy industries are the backbone of our society any improvements in this area mean indirectly a better standard of living Steel and iron production is one of these industries Steel making has evolved dramatically since mankind learned how to produce it Yet there is still a lot to do because the p
41. For this a good resolution and adjustment of cameras is of great importance Image processing algorithms must be working in real time which could demand quite a lot of computational power especially if the number of tuyeres is large It would though certainly be an investment that pays off in the future 8 2 Suggestions Having discussed all encountered problems it would be nice to give some an swers These suggestions might be technically very hard to achieve but neverthe less complaining about things in the existing plant we owe some suggestions for improvements First of all the flow measurement should be improved using image processing Image processing could even be used alone in the future providing efficient and reliable image processing algorithms are fully developed A closer study of the behaviour of the pulverised coal in the pipes could also be enlightening It would for sure to some extent help understanding how the flow should be measured Other changes to be made in order to make the image processing and calculations involved as simple as possible are mostly concerning the cameras The green glass filter in front of each camera should be replaced by a transparent filter that lowers the intensity of light but does not distort it in any way Doing that we will gain a set of three channels with information which as mentioned would be good for redundancy reasons and at the same time do not expose cameras to excessive light Whe
42. IGITISING 25 FIGURE 3 4 2 Interlace phenomenon in a sampled frame from our video recording FIGURE 3 4 3 An image composed of two different frames couple of milliseconds the sampling time was 2 seconds so a strict time scheduling was not necessary and we wanted to make the program code for grabbing the frames easier Bear in mind the data acquisition programme used Lab View was running on Windows 95 sampling data every second We have every reason to assume that the sampling was not perfect on the millisecond level Because lack of a very good synchronisation signal as mentioned in Chapter 3 3 to start grabbing the images we had to rely on our extremely good reflex time 26 3 COLLECTING DATA and start grabbing when the synchronisation signal was visible on a monitor screen attached to our VCR Any mistakes here will result in a time delay when comparing the logged data with the one extracted from the sampled video recording which is hopefully much smaller than the overall time delay for the whole system A fully operational system based on real time image processing of the video signal should consider to deal with a short sampling time and enjoy using the real time features in RTLinux CHAPTER 4 Data Processing The collected data is of no use if it is not analysed with a critical eye Knowing the data characteristics and its limitations is of great help when it is later used in Chapter 7 for validating the extracted fl
43. ING A shame is that we never let the flow measurement settle or saturate before we changed the valve position The pressure signal in pipe 2 did also respond to our test strangling of the pipe reveals itself by an increment decrement in pressure As shown in Figure 4 1 2 b the pressure increases each time the valve after the flowmeter is closed and decreases each time the valve before the flowmeter is closed It seems that the pressure in the pipe can be kept constant when a valve is half closed The PSD for 31F 1102 decimated2 times The PSD for 31FV139 decimated2 times Power Spectrum Magnitude dB o Power Spectrum Magnitude dB iadi hs An WA n NA IN WA l Aand N A PA M re Ke Wwe es N S i V N ox ve 0 0 05 01 015 0 2 0 25 0 0 05 0 1 0 15 0 2 0 25 Freguency Freguency a Mass flow b Control The PSD for 31P1102 decimated2 times k Power Spectrum Magnitude dB i 3 ot a KAMWE KWAA i i 0 0 05 01 015 0 2 0 25 Frequency c Pressure FIGURE 4 1 3 PSD plots for different measurements in pipe 2 second occasion Frequency is in Hz good look at the control signal of pipe 2 presented in Figure 4 1 2 c ensures us about a reaction from the controller The response time is about 30 seconds Sometimes we were about to reopen the valve before the controller hits its upper limit The more important thing is that the control signal increases
44. Image Intensities 6000 4000 4000 2000 100 150 200 250 2000 2000 1000 1000 2000 1000 100 150 200 250 1500 1000 500 o 0 50 100 150 200 250 a Based on an image taken with a green b Based on an image taken with a green glass without coal glass with coal Coloured image Histogram of The Image Intensities Coloured image Histogram of The Image Intensities 6000 4000 2000 100 150 200 250 2000 1000 2000 1000 100 150 200 250 100 150 200 250 2000 1500 1000 500 1000 o o 0 50 100 150 200 250 0 50 100 150 200 250 c Based on an image taken with a trans d Based on an image taken with a parent glass without coal transparent glass with coal FIGURE 5 2 1 A comparison of image intensities in the RGB space enough the green and blue histograms for the coal free image do not reach above 220 those buffers are not saturated In fact the red component is not saturated either This is not the case in the image with coal except for the blue component Those shifts in the colour scale are quite interesting in temperature analysis and are not found in the images taken with green glass Notice also that in those images we do have something in the intermediate register between the brightest and the darkest hills This was not present in the pictures taken with the green glas
45. LULE 2000 076 TEKNISKA UNIVERSITET Enhanced Control by Visualisation of Process Characteristics Video Monitoring of Coal Powder Injection in a Blast Furnace Jihad Daoud Igor Nipl Civilingenj rsprogrammet Institutionen f r Systemteknik Avdelningen f r R eglerteknik 2000 076 ISSN 1402 1617 ISRN LTU EX 00 076 SE Enhanced Control by Visualisation of Process Characteristics Video Monitoring of Coal Powder Injection in a Blast Furnace m Jihad Daoud Igor Nipl 2000 02 29 Authors address LULE UNIVERSITY OF TECHNOLOGY DEPARTMENT OF COMPUTER SCIENCE AND ELECTRICAL ENGINEERING CONTROL ENGINEERING GROUP S 971 87 LULE SWEDEN jihdao 6 student luth se igonip 6 student luth se People come and go in our lives so does every coal particle in a blast furnace Some day you will become coal and if you are lucky you might be used in a blast furnace Nothing is static not us not you not any of the images we analysed Fighting against changes being static is like No opinion on that one A language can be used to control people a computer can be used to control much more It is only stupid people things that are easy to control but again everything is relative Do not ask us about the truth we are still searching When we find it you will know it or you are already dead If you are not dead you are too lazy or you know something we do not know To the poor people Jihad
46. Temperature measurements of the flame would be much more problematic Looking into the hue saturation and intensity buffers we can conclude that there is still noise present for hue and saturation although it looks sharper than previously The intensity component is not remarkably affected it look more alike the grey scale image than it did in the image taken with green glass 38 5 IMAGE PROCESSING ERILE a Full color b Red compo c Green compo d Blue compo nent nent nent e Grey scale f Hue compo g Saturation h Intensity nent component component FIGURE 5 1 4 An image taken with a transparent glass with its decompositions in RGB HSI and grey scale 5 2 Image Content An important issue is to find the difference between a coal free image and another one with coal A coal free image does not mean a gas free image nor does it mean an activity free image It is only an image that seems to have less coal than other images in general This differences might be viewable when looking at the intensity histograms of the picture Another issue is figuring out where to look for the plume the flame and the gases 5 2 1 Image Histograms n Figures 5 2 1 and 5 2 2 we see images with their corresponding histograms First we have RGB decomposition of the pictures and then HSI components The upper part of each figure is taken with a green glass and the lower part is taken with a transparent glass the left side repre
47. although not very favourable which will be discussed later The signals in the electrical cabinet were in the range of 4 20 mA a current circuit connection was employed for its robustness to signal noise not affecting the plant and its easiness Not surprisingly this project also demanded video recordings from the three cameras beside the previously mentioned signals These video signals could be found at another electrical cabinet at the plant Using ordinary video recorders video signals could be recorded for later usage and investigation 3 2 Measured Signals Knowing the situation presented in the previous section and the uncertainty of which signals to collect a decision was made to collect all of the 13 listed signals considered if possible This was done during the first data gathering on the 4th of June 1999 Although some problems arose in the blast furnace data was stored on a laptop and transferred to stationary computers standing in a small image processing lab at Lule University of Technology for later consideration A couple of days later sadly the pressure signal of injection pipe 1 was discovered to be faulty It was necessary to arrange another signal collecting session which was eventually done on the 10th of June 1999 This time all signals were fine except for one of the video signals Apparently we can not have them all Again there were problems with the furnace process during data gathering time which later turned out
48. anced it even more Next stop is on Figure 5 2 6 where we have images taken with a transparent glass The first thing to notice is that the plume is smaller Overexposure We believe that the plume is distorted Different thresholds do not give separate areas in general It is easier here to see the background There is almost no coal at 0 2 5 2 IMAGE CONTENT 43 Full colour Quantized colours Colours 0 00 0 10 Colours 0 10 0 20 Full colour Quantized colours Colours 0 00 0 10 Colours 0 10 0 20 Colours 0 20 0 30 Colours 0 30 0 40 Colours 0 40 0 50 Colours 0 50 0 60 Colours 0 20 0 30 Colours 0 30 0 40 Colours 0 40 0 50 Colours 0 50 0 60 Colours 0 60 0 70 Colours 0 70 0 80 Colours 0 80 0 90 Colours 0 90 1 00 Colours 0 60 0 70 Colours 0 70 0 80 Colours 0 80 0 90 Colours 0 90 1 00 a Red component image without coal b Red component image with coal Full colour Quantized colours Colours 0 00 0 10 Colours 0 10 0 20 Full colour Quantized colours Colours 0 00 0 10 Colours 0 10 0 20 Colours 0 20 0 30 Colours 0 30 0 40 Colours 0 40 0 50 Colours 0 50 0 60 Colours 0 20 0 30 Colours 0 30 0 40 Colours 0 40 0 50 Colours 0 50 0 60 Colours 0 60 0 70 Colours 0 70 0 80 Colours 0 80 0 90 Colours 0 90 1 00 Colours 0 60 0 70 Colours 0 70 0 80 Colours 0 80 0 90 Colours 0 90 1 00 c Green component image without d Green component image with coal coal Full colour Quantized colours Colours 0 00 0 10 Colours 0 10 0 20 Full colour Quantized colours
49. as test This test is also interesting in our case to see if we can detect the flow changes with image processing We had two valves which we could close to prevent the coal flow into the furnace The first valve was located directly after the injection vessel and before the flowmeter the other valve was placed after the flowmeter we refer to those as the valve before and the valve after relative to the flowmeter as illustrated in Figure 3 2 1 We opened and closed the valves repeatedly with different throughput rates The test duration was limited due to a desire for not affecting the plant or ending up with a plugged pipe We realised later that the time between the different actions was not far from the limit of being too short because the flow measurement is very strongly filtered see Chapter 4 1 for further insight Coal Injection Vessel Slag Vessel Controlled Valve Flow Meter a Valve After Valve Before FIGURE 3 2 1 The locations of valves along the injection pipe Table 4 shows the actions taken during the test the valve position we mention is actually the valve handle position and not the real valve opening rate position 22 3 COLLECTING DATA which is most likely non linear Notice that the valves were totally opened before and after the test Other signals as slag vessel weight and its pressure as well as weight correction multiplicator should have been collec
50. at is constantly changing Those changes are not great but can be irksome when for instance trying to retrieve information about the temperature of the flame Problems are in that case caused by the fact that a change in colour does not have to reflect a change in temperature which seems like a troublesome case to solve Maybe the easiest way to deal with this problem would be disabling the auto adjustment function in the cameras used Further another source of concern is the noise introduced by the poor quality video tapes and the video recorders themselves Watching a recorded sequence it is apparent for a human observer that unwanted noise is present Luckily the extent of this phenomenon is limited and it should not affect the outcome of later analysis too much All problems discussed are evidently of harm for the quality of images to be analysed Obviously a higher quality of images is better but it has to be pointed out that their quality is still well above what is needed for image processing to be performed in order to obtain vital information about the blast furnace process 3 4 Video Digitising The recorded video signals on the tapes needed to be converted into a usable format in order to process them in a computer Digitising the video signals was required It is always hard to choose a computer environment to work with in our case the choice was easy Microsoft s Windows family has never been a choice of a serious researcher engineer
51. ation with Mefos The work you hold in your hands is brought to you by two human beings but is a result of many more human participants People without whose knowledge and willingness to help you would not be able to read this report During the time we spent on this research we learned to know several people with different backgrounds from different companies gained more understanding of the complicated coal injection process in a blast furnace and improved our skills in image processing We had great help from our examiner Professor Alexander Medvedev and our supervisor Ph D Olov Marklund both at present working for the Department of Computer Science and Electrical Engineering at Lule University of Technology Thank you for offering us a part of your valuable time We would also like to thank Andreas Johansson Wolfgang Birk and other researchers at the Control Engineering Group Roland Lindfors at the AV centre Per M kikaltio and others at the Division of Industrial Electronics and Robotics Krister Engberg at the Division of Signal Processing for putting up with us The system administrators Mattias Pantzare and Jonas Stahre for their indispensable help All working at Lule University of Technology The Free Software Foundation offering the world the best they can achieve without them we would be dependent on commercial software except for MATLAB where we had no time to write the needed toolbox for Octave Not to forget the helpf
52. b 1 b 1 n a 1 a 1 b 1 b 1 amp ones 3 mn s sum sum d p y find 1 s if y if length wx lt mn wx mn 0 end wx mn wx mn 1 e e x v v y 1 dv lv yl u u mn end end A 5 line2pixel m Purpose Convert lines described by two end points into pixel s in an v image or elements in a matrix Synopsis A 6 IMCONNECT M r c line2pixel x y Description Found the pixels below a line and return thiere positions Don t worry about any warning like Warning Divide by zero it 4 has been taken care of NSee also CAPTURE ROIPOLY function r c line2pixel x y warning off r c i 2 2 length x j i 1 x1 x j x2 x i yl y j y2 y i y1 y2 x1 x2 b y2 m x2 for p 1 length b s x1 p t x2 p u y1 p v y2 p if u lt v ri uti v 1 elseif wv ri vti u 1 else ri end if s lt t ci sti t 1 elseif s gt t c1 t 1 s 1 else c1 end if isempty c1 ci s ones size r1 elseif isempty r1 ri u ones size c1 else c2 round ri b p m p r2 round m p c1 b p c1 c2 c1 ri r1 r2 end c c ci r r ri end A 6 imconnect m Purpose Connect discontinuous edges or pixels in a binary image Synopsis j imconnect i Description Find the discontinuous parts using imends and then connect the parts in a minimum way j is the new edge connected image See also IMENDS IMCHAIN
53. be connected to 4 The right answer is 3 and in this case 1 and 2 are the ghost endpoints They should either be removed from the endpoint list returned by Algorithm 2 or dealt with later in an algorithm that uses it Having the endpoints selected it is time to move on and connect them This is accomplished with Algorithm 3 based on the ideas presented earlier in this section 6 1 FINDING THE BACKGROUND 49 FIGURE 6 1 3 T connections and ghost endpoints Algorithm 3 Imconnect imconnect im oim find and identify the objects in im eim find the pixels representing the endpoints of the objects in oim dis calculate the distances between all the pixels in eim cim im while there are unconnected objects in oim pl p find the pixels with shortest distance in dis between non connected objects according to oim cim cim connect p1 and p2 update oim by making the objects that p and p2 belongs to one object 9 end SO eee iN ON Thus connecting the edges with a reasonable result is possible This algorithm starts with an image that has disconnected edge parts identifies them finds the endpoints and makes a list of the object they belong to calculates the distance for all possible connections between the endpoints watching out for ghost endpoints looks for the minimum distance makes a connection between the two parts and updates the endpoint list to regard the last connected parts as one part The algorithm does this unti
54. casion which can be compared to the ones related to pipe 1 also from the second occasion in Figure 4 1 4 As you can see the peaks in a b and c are below 0 2 Hz which means according to Nyquist s sampling 1 theorem that a sampling time of 2 5 seconds 553 should be sufficient Choosing 24 3 COLLECTING DATA 3 seconds sampling time is kind of adventurous because it will catch up with fre quencies up to 0 16 Hz which may result in some aliasing in our case We assumed that the characteristics in the collected signals would be reflected in the data to be extracted from the recorded video signal Actually the flow measurement device used a couple of seconds data filtering and the control system used at Mefos use a time base of 0 20 0 25 seconds Pole a 10 frames before capturing TEE b 10 captured fields of the frames above FIGURE 3 4 1 Interlace principle 3 Video is sampled and displayed such that only half the lines needed to create a picture are scanned at a particular instant in time A video frame in our case consists of two interlaced fields of 625 lines Interlace is the manner in which a video picture is composed scanning alternate lines to produce one field approximately every 1 50 of a second in PAL Two fields comprise one video frame As shown in Figure 3 4 1 if the upper sequence is captured by a conventional video camera the result will be the lower sequence which means that our frame c
55. ckground and Imweight for volume estimation at a threshold level of 0 45 We were lucky to have frequency distorted signals in our first data collecting session as mentioned in Chapter 4 1 This trait is helpful when looking at the signal dynamics A quick glance at the frequency content of the measured flow signal and the extracted one shows the same dominant frequency peaks as presented in Figure 7 2 1 The extracted signal is based on fewer samples which gives rougher impression compared to the collected signal but the frequency peak is there at 0 15 Hz By this we know that we are able to reconstruct the same frequency characteristic of the flow signal with our algorithms This was a first step in assuring that our measurement was valid At a later stage of the project we had to face the fact that the flow measurement quality was not satisfactory Lack of high correlation is likely due to bad existing measurement The measured signal is as explained corrected with a correction factor based on the weight signal This is done because the flow measurement is not good enough to be completely trusted In this case validation of the extracted data is not an easy task Searching for the truth we had to look at the other signals related to each tuyere Those are the pressure and the control signals As before both the first and the third video recordings have been inspected and the result is presented in Figure 7 1 1 The left side of the figure representi
56. correlation with the other algorithms demonstrating the non linear relation between them It is higher for small threshold values due to the fact that the plume is close to non existent and there is no bigger difference between area and volume estimations Volume estimation algorithms show high correlation with each other in general Comparing them using the different background algorithms till shows relatively strong correlation Figure 7 1 1 e reveals the previously mentioned fact that the blue component is not very nice The customised background tends to 59 60 7 DATA EXTRACTION AND VALIDATION get hard to detect resulting in vanishing plumes while the plumes using the other algorithm are not representative for the coal flow which results in a false strong correlation between them In this case we have to rely more on the red and green colour buffers Moving on to the right column of Figure 7 1 1 we notice that the third video recording is showing some different behaviour The correlation between area and volume estimation algorithms is weak as expected but uncovers the uneven quality between different threshold levels Looking at the correlation between the volume algorithms gives high overall correlation but again there are indications of the poor plume quality depending on thresholding level The lowest threshold levels give no plumes meanwhile high levels lean toward being fully white due to overexposure of the video The good plumes r
57. d be smaller than the image size in each direction end jezeros r c sti r s sti c s i sti r s st 1 c s A 4 imends m Purpose Find the end pixels in an binary images Synopsis e v u imends i Description 75 76 A MATLAB CODE An end in a binary image is a pixel connected with at most one other pixel to the neighbouring 8 pixels e is a vector of pixels order u is describing which line segment each pixel belong to and v is the direction in which way they are connected as it is described below zero means no connection as you may imagine we assume thin line segments 7123 2405 467 8 NSee also IMCHAIN IMFILTER function e v u imends i r c size i ii zeros r 2 c 2 i1 2 r 1 2 c 1 i r1 ci find il p 07 11 13 17 00 19 23 29 31 sort reshape p 9 1 This 1 below is if you don t want to treat single points as end points Comment it if you don t like it and don t forget to change v y into v v y 1 below 1 2 9 e v n i1 w 1 wx u for x 1 length r1 a ri x b c1 x d il a 1 at1 b 1 bt1 en n a 1 ati b 1 bt1 y unique en find en gt 1 f sum max y ones size y y up dowm for ij 1 length y down down find n y ij 1 end n down max y if length u for ij 1 length y up up find u y ij end u up max y end end mn mam en if mn w wtt mn W end na 1 a 1
58. e Component Blue Component 1000 T T 2000 T T T ili NI Pee cee eee 4 parn ae j SN eg A a pe Ne Wa a 4 2000F y jaa TE 500 i t i i 4000 0 50 100 150 200 250 300 0 50 100 150 200 250 300 a Difference of images taken with a b Difference of images taken with a green glass transparent glass FIGURE 5 2 3 Difference in the histograms in RGB space The x axes represent intensities and the y axes are the difference in pixels It might be hard to draw all the conclusions from the former figures Another way of seeing the same thing is subtracting histograms of the coal free images from those with coal The result is illustrated in Figure 5 2 3 and 5 2 4 The positive values represent intensities found in the images with coal and the negative values are those found in images without coal In the RGB space we see that the blue component of the images taken with a green glass is quite noisy Also the red one is a little bit so The three buffers in 5 2 3 a are different no specific characteristic on the other side 5 2 3 b are more or less three identical graphs Inspecting the HSI space the two histogram sets show different behaviour The noise in the hue and saturation components is clear Concentrating on the histograms of the images taken with the green glass we can see in an image with no coal values 50 65 that are missing in an image with coal for the hue component The same phenomenon is visible between 115 175 for the
59. e camera lens The cameras themselves are mounted about two meters from the outer wall of the blast furnace and are built in inside a protecting cover The light from inside the furnace is led to the cameras through peek holes in the wall near each tuyere via protecting pipes Furnace Pulverized Coal Pipe Wall Protecting Cover N Tuyere Filter N yA a ae Coal Plume Protecting Pipe Video Camera FIGURE 2 4 1 The video camera setup For a full comprehension of the whole video camera apparatus Figure 2 4 1 might be helpful for the devoured readers On its way to the lens light passes the previously mentioned glass filter The characteristics of this filter have not been examined in great detail but having a closer look at the three colour buffers shows for the human eye that the red and in particular the green light pass the filter almost unaffected while the blue light is filtered out to the extent that the blue buffer becomes nearly useless as discussed in Chapter 5 It is therefore desired to solve the filtering problem We had the possibility to use transparent glass instead of the green one which we believed would leave all the three colour buffers unspoilt By doing this we risked introducing overexposure to the video surveillance system Possible solutions will be discussed later in this report a With green glass b With transparent glass FIGURE 2 4 2 Sampl
60. e images taken with green glass as filter and with transparent glass An example of what is seen with the help of the cameras is in Figure 2 4 2 where the mouth of the tuyere is seen together with a dark elliptic shaped cloud the pulverised coal injected inside the furnace CHAPTER 3 Collecting Data Doing a project of this nature needs of course collaboration with the industry in order to improve things Simply there is a need of being at the field for investigating possible ways to get across suitable data to kick off the project People working at the plant know how to run it and how it behaves Collecting data does not only mean pure data measurements it also includes collecting the knowledge possessed by the plant workers Every detail is important every worker has something to tell that we probably need to know The knowledge we gained from people in the field is spread all over the report Below we are dealing mostly with measurement data collection 3 1 Available Signals and Equipment To start with we will shortly discuss which signals were available to us for fur ther study After an exhaustive investigation of the process and the parameters that were available and examining the signal collecting equipment we finally concluded what had to be done It was obvious right from the start that the coal flow signals were of importance for this project Also the nitrogen pressure in the pipes leading the pulverised coal to the blast f
61. eed to calculate coal flow flame temperature and coal distribution data While we are not handling the temperature and coal spreading parts in this report we will concentrate on the coal flow emphasising that the same ideas apply for all three parts Remember that we have colour images that means we have several information channels that can be combined after processing them separately The algorithms listed below handle a single information channel of the image if nothing else is mentioned explicitly All the discussed algorithms are implemented see Appendix and tested in MATLAB See Chapter 7 for further investigation of the implemented algorithms used with our data The steps from an image to a fully useful flow data can be outlined as follows Finding the image background Finding the coal plume Finding the plume s volume e Conversion between volume and flow measurement The first three points are dealt with next 6 1 Finding the Background A background is the dark area in the image representing the protecting pipe and the visible part of the tuyere Finding a suitable background is a first step in isolating the plume in every sampled image The need of a background will be explained in Chapter 6 2 The following discussion covers two different investigated approaches for finding the background mask a black and white image binary image We want to keep apart the use of two terms throughout this report plume and flame By p
62. els representing a regular shaped pattern What we have here is a circle with a part of a tuyere inside this is a tough one Closer examination has shown that we can assume that each two adjacent edge section s endpoints facing each other should be connected to create a continuous edge We can also look at the direction of each edge part s endpoints before we connect it to anther endpoint In this manner we try to trace through the edge parts in order to connect them all Finding those endpoints the edge part they belong to and the direction they are pointing in is done using Algorithm 2 This algorithm takes a binary image with disconnected edges and returns the possible endpoints and their direction encoded as a digit 0 8 where 0 is no connection and is imply an isolated pixel In this algorithm we look at the eight neighbouring pixels for each pixel before deciding anything A pixel is regarded as an endpoint if it is only connected to one of its neighbours We have to ensure ourselves that the edges are one pixel wide otherwise we need to look at more than the eight neighbouring pixels to determine whether a pixel is an endpoint or not A problem we still need to deal with are all the T connections as shown in Fig ure 6 1 3 those endpoints that are not to be connected are called ghost endpoints As you can see there is no doubt about how the green points should be connected 8 gt 7 and 6 5 It is not obvious knowing which point should
63. eport is a first reach for utilising computer vision systems in steel making industry by monitoring process parameters To round off this report we will briefly present some conclusions and then discuss a few suggestions for potential improvements and further work to be done 8 1 Conclusions Studying the whole process and developing algorithms based on image process ing for watching the process behaviour we came up with some conclusions They are all gathered here First we can say that the current flow measurement is not satisfactory or at least it could be improved using image processing This statement is based partly on what we have been told at Mefos but mostly on what we discovered comparing data extracted using image processing to the collected flow signals There should be a high correlation between those which could not be found We could only detect a lower degree of correlation Since this was quite a surprising result further tests were made In one of these the coal flow was choked and later the measured flow examined Our algorithms could detect changes in the coal flow We tried also to correlate the estimated volume to the control signal the correlation increased Another less sophisticated way of seeing that the present flow measurement is not very reliable was to simply find a film sequence where we saw that the flow was low in the beginning and high towards the end We also knew that the camera was not moved or zoomed dur
64. esidue somewhere in the middle of the threshold range where the small correlation dip could otherwise mislead any superficial examiner Our knowledge of the similarity between the colour buffers in this video signal is confirmed in 7 1 1 f in contrast to those in the previously examined video signal Here we have three useful colours that can be a part of flow estimation algorithms A good redundancy can be achieved 7 2 Extracted Data Characteristics The question now is if the algorithms applied to the video recordings give an accurate and representative measurement of the pulverised coal flow Close examination has been done on each of the resulting signals after running the different algorithms This has shown apart from what we have mentioned about the colour buffer quality and plume thresholding value in Chapter 7 1 that the fixed background gives a better result in general compared to the customised one This could be related to some bugs in our implementation of the Imconnect algorithm and our badly calculated estimated threshold values used in finding the background in conjunction with not using the best developed algorithm based on multiple images for finding the background The Imweight algorithm proved to be the most suitable among the tested ones although the other algorithms were not that bad and could be more useful with some modifications All of the following plots are based on the red buffer of the images using a fixed ba
65. ge type Nit is more accurate than any imthresh based on the mean value and or ithe standard deviation of the image This seems to be true at least for the stupid blue image we have e edge i t j imframe e 5 YA better imconnect could eliminate the next step The goal of this is to get a thin edge one pixel wide with no glitches bwmorph bwmorph j dilate 2 shrink 3 adjust this due to what trash you want to clean 3 is good for getting rid of single pixels yf imfilter b 3 c imconnect b g bwfill c holes A 9 dyncrop m Purpose Find the interesting area in an image and crop it Synopsis i x y r c dyncrop im YDescription Based on the standard deviation and the mean value of an image and where the light area are present the image 80 A MATLAB CODE is cropped and its crop parameters are returned i is the cropped image x and y are the start positions r and c are the dimensions of the cropped area NSee also IMCROP function i x y c r dyncrop im Find where the interesting area is xs sum im ys sum im m mean xs std xs 2 m mean ys std ys 2 3 xf find xs gt xm yf find ys gt ym x min xf x2 max xf y min yf y2 max yf c x2 x r y2 y imcrop im x y c r i A 10 dynbg m Purpose Find the image background Synopsis bg dynbg im th
66. gnal processing UK John Wiley amp Sons Inc ISBN 0 471 14961 6 16 Ramsey Technology inc 1996 Installation amp operation manual for Ramsey DMK 270 transmitter with Granucor REC 3956 REV B 10 96 PART NO 050622 17 Tottie Magnus 1999 LKAB s Experimental Blast Furnace at MEFOS Lulea URL http www lkab se english company mt pilot_mt_english html 1999 06 24 73 74 BIBLIOGRAP HY APPENDIX A MATLAB Code A l mam m Purpose Get the largest element independent of the matrix dimension Synopsis m mam x Description First establish the dimension of X and then get the largest element in it See also MAX MIN function m mam x length size x m x for i 1 d m max m end A 2 mip m Purpose NA Get the smallest element in a 2 D matrix and return its position Synopsis v r c mip x Description 4 First the smallest element the matrix and then calculate its position NSee also MAM MAX MIN function v r c mip x v a v c r a c min x min v A 3 imframe m Purpose Make the image boundary black Synopsis j imframe i s Description Ah If 2 s 1 is smaller than the number of columns and rows of the image then j is the same images as i with a black s pixels boundary YSee also YA function j imframe i s r c size i if r lt 2 s c lt 2 s error 2 s shoul
67. here is nothing distinctive about them except for the sudden variations seen throughout a 27 28 4 DATA PROCESSING A plain plot for 31FV138 A plain plot for 31FV139 3 2 T T 4 T T T 24 T T T im m iy MR il Il i vl W i i i i i i i i i i i i 0 500 1000 1500 2000 2500 3000 3500 o 500 1000 1500 2000 2500 3000 3500 Sample Sample a Pipe 1 b Pipe 2 A plain plot for 31FV140 i fi i fi i 1000 1500 2000 2500 3000 Sample c Pipe 3 FIGURE 4 1 1 Control signals first data logging occasion X axis is samples taken 1 per second and Y axis is sensor value in Volts section in the first part of the data This is not as obvious in the mass flow signals except for the flow signal for pipe 3 Actually it is hard to tell anything decisive studying the mass flow in pipes 2 and 3 Possibly the flow is slightly higher in the first part of the data The control signals are weird looking things which have very little in common Control signal for pipe 1 is following a constant line with quick high variations In control signal 2 we can discern that the signal drops after a while which resembles a little the behaviour of control signal 3 In the last mentioned signal again clear deviation from normal behaviour appears in the beginning Lastly looking at the injection vessel and air lock vessel signals it is clear that something happens exactly the same time as seen in
68. homogeneous media 16 Internally the device consists of a solids concentration sensor and a velocity sensor with two measurement points separated by a distance S based on the capacitive measuring principle The par ticle stream is measured in two points which the velocity transmitter correlates to find the closest similarity between them From this correlation function the transit time T from point one to point two can be determined In the solid concentra tion sensor the change in capacitance is proportional to the solids concentration this voltage signal is transformed into a frequency signal and is Pulse Frequency Modulated PFM The flow rate Q is given by Q C V Agensor where C is the concentration of the medium V its velocity and Asensor is the sensor cross section area and is calculated with Asensor Geneon where dsensor 18 the sensor diameter The concentration C Ae frem ferm K where K is an adaption factor to concentration sensor fpry is the measured frequency of PFM concentration signal fpru is the frequency of PFM signal at concentration zero and K is a calibration factor Finally the velocity is calculated with the well known formula V Not to forget that the both vessels are equipped with weight gauges and pressure meters Other important sensors are the pressure measurement devices in the coal injection pipes 2 3 2 Controllers Figure 2 3 1 is a schematic overview of the main con trollers Because the f
69. i bg NOOR ON eH We have been looking at two algorithms in order to obtain a good result Algorithm 10 is used when we do not know where the background mask fits in the image First we threshold the given image find where the background best fits and then we remove the unwanted objects by multiplying the thresholded image with the background and voila the plume is there The second algorithm Algorithm 11 is for the case when we know where to place the background on the image Also here we start with thresholding the image use morphological operations to get rid of unwanted objects and finish with morphological filtering Algorithm 12 in case the earlier morphological operations were not enough This filtering can also be executed as the last stage of Algorithm 10 Algorithm 11 Implume implume im th bg 7 y 1 nim lt bg im gt th 2 mim clean up nim using morphological operations 3 fim identify the flame in mim knowing its originate from 2 y Algorithm 12 Plumeident plumeident im th x y 1 lim identify the objects in im 2 for i 1 to the number of objects in lim 3 isum sum lim i 4 if isum gt th 5 xc yc find the centre of limfi 6 distfi y ye a xc 7 else 8 distfi 0 9 end 10 end 11 mindist min dist 40 12 fim lim mindist The need of morphological filtering arose when we started to process the video signal from our third signal collecting sess
70. imated flow the difference between them is approximately the maximum value in the first minus the maximum value in the second which is about 13 seconds Applying the filter identified in Chapter 4 1 1 to the flow signal from the third video recording and correlating it with the extracted flow measurement resulted in the plot in Figure 7 2 3 he positive correlation noticed in Figure 7 2 2 f has increased and moved from 9 to 9 seconds because of the filter influence Also we can clearly see a suppression in the negative correlation we had at 13 seconds This makes the correlation plot look much more similar to those in Figure7 2 2 b nd Figure 7 2 2 d The good job done by the filter should not be overestimated remembering that filtering followed by inverse filtering can distort the signal This could be noticed because the signal fluctuates with higher frequency as Figure 4 1 6 shows Data was filtered using different low pass filters to get rid of the high frequency fluctuations consisting of pure noise This was mainly visible in the algorithms using a customised background approach Now the correlation not very surprisingly increased slightly to reach above 0 5 Unfortunately for a closer validation we would need to know the exact amount of coal injected at any time However for Correlation coefficient Correlation coefficient Correlation coefficient 7 2 EXTRACTED
71. in total From there four algorithms Area Imweight Rotalgo and Ellipalgo were used for data extraction Of course we applied all these on each colour buffer Ending up with 2 2 11 4 3 528 different runs Calculations performed on pictures as is often the case in image processing were computationally demanding and time consuming Today s computers not the one we used offer the needed speed This in combination with an optimised code implementation and suitable choice of programming language compared to the MATLAB code we used would boost the data extraction speed in order to allow real time execution Extracted data had to be analysed in a cautious way so that nothing was omitted Having estimated the coal flow using images it was close at hand to search for a correlation between our estimation of the flow and different measured signals at the plant 7 1 Relations Between Algorithms Having our four algorithms we want to build an opinion about how the algo rithms are related to each other Figure 7 1 1 shows some three dimensional plots describing the main correlation features according to colour buffers and threshold levels The colours are running from 1 to 3 and represent red green and blue respectively Thresholds are ranging between 1 and 11 from the smallest to the largest As you can see there are big differences between the two video signals The left column representing the first video capturing The Area algorithm shows small
72. ing a functional model for the coal cloud behaviour would contribute to enhancing the coal flow measurement as well A good model for the plume flame would be of great help Having a good volume estimation is essential A closer investigation of what exactly happens when slag is injected besides pure coal powder is advised Slag injection makes the whole injection process even more complicated The case when slag is injected with coal should be studied and compared to the pure coal injection 72 8 CONCLUSIONS AND SUGGESTIONS Bibliography 1 Agarwal Jay Brown Francis Chin David 1996 The future supply of coke New Steel H W Wilson AST vol 12 p 88 ISSN 1074 1690 2 Birk Wolfgang 1997 Pressure and flow control of a pulverized coal injection vessel Sweden Lule University of Technology Master thesis 1997 045 38 41 ISSN 1402 1617 3 Creek Patricia Moccia Don 1996 Digital media programming guide Silicon Graphics Inc Document Number 007 1799 060 URL http autarch loni ucla edu ebt bin nph dweb dynaweb SGI_Developer DMediaDev_PG Generic _ BookText View 3 cs fullhtml pt 4 4 Daoud Jihad Durand Cyril Mock Nico Nipl Igor Sayyahfar Zohreh 1999 Opti mazation of washing control at AssiDoman Kraftliner in Pite Sweden Lule University of Technology Automatic control project 5 Jain Anil K 1989 Fundamentals of digital image processing USA Pentice Hall Inc ISBN 0 13 332578 4 6
73. ing filming time Now the present flow measurement did not change at all whereas data recovered from images indicated the observed rise in coal particle flow Unless coal flow suddenly changed path right after injection to the blast furnace the current coal flow measurement is inadequate The missing link here is still a conversion between calculated volume and useful flow measurement Another conclusion was that using the green glass as a protecting filter for the cameras skews the colour buffers specially the blue one Since the cameras are colour cameras we lose valuable information in one of the channels The red and green buffers look decent despite the just mentioned filter But of course all information is valuable so being able to use the blue buffer as well as the two others would be advisable Still investigating the cameras we can state that if image processing is to be applied not making it too complicated a static image is needed Algorithms have been developed to avoid this problem and those are comparable with the assumption of static images An additional obstacle is that the plume is filmed from an angle This makes it slightly harder to estimate the volume of it Strict camera positions have to be used in order to be able to rotate the plume and calculate its exact volume 69 70 8 CONCLUSIONS AND SUGGESTIONS Looking into the future not only do we want to estimate the coal flow but also study other process parameters
74. ion where the images tend to be overex posed and thresholding gave sometimes several unwanted objects beside the plume itself You can see the difference for yourself in Figure 6 2 3 The idea behind the algorithm is first identifying the different objects in the image followed by removing the smallest of them according to a given value Then by knowing the tuyere s position in the image the distance between that point and 6 4 ESTIMATION OF PLUME S VOLUME 55 a Before using the algorithm b After using the algorithm FIGURE 6 2 3 Morphological filtering using Algorithm 12 centre of every object left can be calculated to finally choose the object with the shortest distance 6 3 Estimation of Plume s Area Having isolated the plume from the rest of the image we could easily calculate its size in pixels The number of white pixels in the picture is determined and the resulting value describes the size of the coal particle cloud and its density We have to be aware of that flow is given by the volume of the particle cloud As an a priori study we can use the area of the plume to represent pulverised coal flow 6 4 Estimation of Plume s Volume The area of a two dimensional plume in the film sequence is easily calculated however the three dimensional pulverised coal body is what we are looking for Estimating the volume of the coal cloud is not an easy task First of all we see the cloud from an angle Secondly we do not know
75. l we have only one object in our image with no disconnections An auxiliary algorithm Algorithm 4 is used to draw a line between two endpoints in an image given their coordinates Back to our initial subject finding the background Algorithm 5 lists the needed steps Edge detecting morphological filtering connecting any disconnected edges and finally filling the resulting object with ones to get the black and white background In a short run the background image is static the happenings are restricted to the inner parts of the furnace If we stack several succeeding images the static part of the images will be dominating The background can now be regarded as the most significant pixels those that are repeatedly found in most of the images within the stacked bunch It is now easy to find the edges surrounding the interesting area Algorithm 6 shows how this is done Similar approach is made in Algorithm 7 apart from the fact that we stack the edges found in the images instead of the images themselves Decisions have to be 50 6 ALGORITHMS Algorithm 4 line2pixels NNN NPR eB eee ee ee NE OOANDURWNH OO OO AOS CoG PRL COSI wa line2pixels x1 y1 x2 y2 me fet be y2 m x2 rel c if yl lt y2 r yi y2 elseif y2 gt y1 r y2 yl end if cl lt 22 c tl 2 elseif 72 gt z1 c z2 x1 end if c nil c 1 length r lt 21 elseif r nil r 1 length c y1 else ci 4
76. lem here What should we regard as the coal plume We know for sure that there is no sharp edge between coal and the gases inside the furnace Also it is not obvious what threshold value should be chosen Depending on the threshold value chosen the coal cloud has different sizes and sometimes even slightly different shapes although it is always close to being elliptic Choosing the correct threshold involves distinguishing between coal and gases in the image which is not completely unequivocal Figure 6 2 2 shows the same plume extracted with different thresholds Notice well plumes with high thresholds are not of any use in the video series taken with transparent glass but are quite nice in the green glass case a eee am ewe eames fare ees we New data threshold 0 25 New data threshold 0 30 New data threshold 0 35 di a New data threshold 0 40 New data threshold 0 45 New data threshold 0 50 an New data threshold 0 55 New data threshold 0 60 New data threshold 0 65 a Based on an image taken with a green b Based on an image with a transparent glass glass FIGURE 6 2 2 Size of the plume depending on threshold value 54 6 ALGORITHMS Algorithm 10 Findplume findplume im bg bwth th bwim lt im gt bwth csum lt sum of the pixels in bwim column wise rsum lt sum of the pixels in bwim row wise cint csum gt th rint lt rsum gt th roi bwimrint cint fim 1 ro
77. low signal following the control signal after a delay of about 25 seconds Notice also that the negative correlation in the third signal logging is much stronger than the one in the first one implying a stronger feedback coupling The pressure flow correlation plots are very much alike the control flow correlation plots with almost the same delays translated by about 2 seconds which confirms what we saw in the control pressure and control flow plots Having just these signals to study it is hard to draw further conclusions They were mainly collected in order to at a later stage be compared to data extracted from recorded film sequences Thus just as means to verify our sequitures from image processing of the video Flow signal in pipe 2 before and after filtering Flow signal in pipe 2 before and after filtering compared to valve position 9 2 E 5 wa Measurement Volt N i i i i i i i i i i i i i i i i i i i 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 5500 3400 3600 3800 4000 4200 4400 4600 4800 Samples 1 second Samples 1 second a Filtered and unfiltered signal b Filtered unfiltered and valve position signals FIGURE 4 1 6 Comparison between the filtered flow signal green unfiltered flow signal blue and the valve position signal red 4 1 1 Filter Identification We saw earlier that the flow measurement did not respond in the way we expected it
78. lowmeter is not reliable the on line flow measurement is multiplied by a correction factor calculated according to the injection vessel weight loss deviation during a certain period of time See Figure 2 3 2 for details Do ing so the idea of on line measurement is lost in a short term perspective while it is relatively accurate considering a longer period of time The corrected flow measurement itself is the output of a PI controller using 1 as its setup value and feedback with the flowmeters to coal vessel weight ratio Before dividing the flow measurement by the vessel weight both signals are windowed with a window length of 10 minutes before performing the division This is done because the scale used in the vessel has a limited sensitivity and a big error margin if compared to the flow measurement during a short time The coal injection is controlled with three PID controllers one for each tuyere An operator is supposed to feed the system with the desired amount of coal flow needed in the furnace the amount is divided by three and the result serves as a setup value for each of the controllers The output from the PI controller is used as 2 4 VIDEO SURVEILLANCE 17 Operator DIV Coal Injection Flow Meter Vessel Valve ui PID nf Tuyere 9 Slag Vessel 96 PID 9 6 PID
79. ls belonging to the thresholded plume the same importance is clearly unfair Common sense suggests weighting pixels so that darker pixels are weighted heavier Although this seems to be a reasonable solution there are no definite rules for this kind of weighting In our case it came to be more or less all about trial and error for how the weighting should be arranged More knowledge about the spreading of coal injected into the blast furnace and metallurgical competence would probably help when deciding the weighting Algorithm 13 Imweight imweight im imax max im imin min im 4 0 nim imaz im imin warea sum nim lt imaz pe SO Net Giving pixels weights according to their shade can be seen as trying to establish a measure of the plume s volume A darker pixel means more depth a lighter less depth in the inwards direction of the image It has to be made clear that there is nothing indicating a linear weighting of pixels A logarithmic or inverse logarithmic weighting might be a better approach One could also consider a look up table solution which would be suitable for storing and recovering weights in a future implementation of a fully developed algorithm Algorithm 13 is based on a linear weighting 6 4 2 Rotated Plume Estimation Rotating the two dimensional projec tion of the plume we can approximate the coal particle cloud volume by estimating the plume s body of revolution The algorithm here might seem a little awkward
80. lume we mean the coal particle cloud being injected through a tuyere meanwhile flame refers to the surrounding burning area which might include gases 6 1 1 One Background Approach The first background approach is to search for an image with a small amount of coal powder visible Reasonably one would argue that the best way of finding the plume would be to find the right threshold values first finding the background and then extracting the plume The shortcoming of this method is obvious in our case In our images the plume the visible part of the tuyere and everything around the opening in the blast furnace 45 46 6 ALGORITHMS wall is very dark Under those conditions there is no simple way of distinguishing the coal plume from other parts The same problem is encountered in both RGB buffers as well as HSI buffers FIGURE 6 1 1 A typical fixed background image Establishing that thresholding is not sufficient means of recovering the back ground from our images we have to think of slightly more advanced solutions One possibility is to look for an image containing very small quantities of coal and then thresholding to obtain a nice background without any trash As long as we can find an image not containing too much coal this works The drawbacks are that it might be hard to find an appropriate image and that the cameras are not fixed which means that if someone would accidently bump into one of them the whole image would move and
81. lve po sition FIGURE 4 1 2 Measured signals in pipe 2 blue compared to valve position red during Andreas test X axis is samples taken 1 per second and Y axis is sensor value in Volts During the November visit to Mefos only eight signals were collected due to the previously mentioned hardware limitations One of these signals turned out to be of no use namely the injection vessel weight signal Pressure in the vessel seems to be relatively constant Now looking at the other signals we have to keep in mind that we choked the pulverised coal flow for injection pipe 2 in Andreas test as we mentioned in Chapter 3 2 Strangling the flow could only be done for short periods of time not to block pipe 2 permanently needing a serious intervention in the system Not very surprisingly inspecting the flow signal 2 we can unambiguously see distinct changes in the flow as Figure 4 1 2 a clearly shows A very interesting thing is the delay in the flow measurement signal it seems to be caused by some filter this made us start scratching our heads Later this discovery will be treated in Chapter 4 1 1 Another interesting observation is that the flow is not really affected until a valve is totally opened or totally closed and we have an overshoot when a valve is totally opened A possible reason is that the valves have a non linear characteristics and that as shown below the flow can be kept as long as the pressure is constant 30 4 DATA PROCESS
82. n Component 1000 T T T T 2000 T 1000F sofy 1 L N Dy IA PANIE ER AS Wi Paji yu Ra 10004 W Vyts 500 i i i za i i i 50 100 150 200 250 300 0 50 100 150 200 250 300 Intensity Component Intensity Component 5000 y 5000 T T ea aa oe eo ao V V 10000 i i i i i sa i i i i i 0 50 100 150 200 250 300 0 50 100 150 200 250 300 a Difference of images taken with a b Difference of images taken with a green glass transparent glass FIGURE 5 2 4 Difference in the histograms in HSI space The x axes represent intensities and the y axes are the difference in pixels the plume s size and to some extent also the shape changes are dependent on the chosen threshold value This is seen when looking at what parts of the image belong to which shade for the three colour buffers Figure 5 2 5 is based on images taken with a green glass In general the images show that the coal including the background is found between 0 1 and 0 5 The background should be separated from the coal somewhere in between 0 1 and 0 2 and the gases lie between 0 5 and 0 9 The foreground is seen at the lower part of the images The blue components are kind of special the last two thresholds are reserved for the foreground That could be something related to the temperature of the foreground The plume has a different shape here and is getting bigger There are clearly elements that have a breakthrough in the blue component Maybe the green glass has enh
83. n it comes to video recording synchronisation a good synchronisation sig nal should be created if delays have to be determined with great accuracy This should trig the cameras VCRs and the data acquisition box An external trig signal is preferred Using the same trig signal to initiate grabbing is a good idea It is a good idea to change Andreas test from closing valves to changing the setup value of the coal powder flow This is easier to change and more sensitive to small changes Changing the setup value in a manner that guarantees different transitions between low high flows will improve the test keeping in mind that the changes have to be held for periods longer than the system delays As we discussed before a static picture makes analysis less complicated There fore cameras should be zoomed on the interesting area of the picture and then mounted in some way so this picture remains static during the whole time we are interested in surveilling coal injection If the cameras were fixed as well as other movable parts like the tuyeres also the task of compensating for the angle from which the plume is seen would be a one time calculation When the angles of the cameras can be changed they can be moved backward and forward and the tuyeres are not fixed then we have a situation where the angle from which the plume is seen is constantly changing which of course does not make life any easier A somewhat different problem to be solved concerning
84. nd flame in a picture when the background is available but the interesting region in the image is unknown Synopsis Y4 flame findflame im mask thi th2 Description Given an image a background mask smaller than the image and threshold values it is possible to a flame in the picture thi is 4 the value used to threshold for the flame and th2 is used to position the mask on the image See also 4 IMFLAME function out findflame im mask thi th2 I1 Ji size mask im double im gt th1 i find sum im gt th2 j find sum im gt th2 iround max vi min vi 2 jround max vj min vj 2 ri im i 11 2 i 11 2 1 j J1 2 j J1 2 1 out 1 roi mask A 15 imflame m Purpose Find a certain intensity level representing the flame Synopsis 1 imflame I th BG x y Description The background BG image is subtracted from the image I after thresholding at th where th is 0 0 1 0 The resulting image is modified and cleaned with some morphological operation x and y represent the end of the tuyere where the coal is spread out See also EDGE IM2BW BWMORPH FINDFLAME function 1 imflame I th BG x y l double BG double im2bw I th bwmorph bwmorph 1 gt 0 erode 2 dilate 2 1 algox 1 200 x y A 16 algo2vol m Purpose Calculate the approximate volume of a 2 D coal particle cloud Synopsis v algo2vol plym Desc
85. nd values are given calculation is performed 4 for the whole length of the horizontal axis NSee also EVALVOL FINDVOLUME2 function vector countpixels plym b e r c size plym if nargin b 1 e c end if r r 1 end if e vector 0 else for l 1 r vector 1 length find plym 1 b e 1 end end A 19 findvolume2 m Purpose Calculate the approximate volume of a 2 D coal particle cloud using the rotation algorithm Synopsis v findvolume2 plym Description Find the volume of a particle cloud using countpixels and evalvol To start 4 with the mean values for the on pixels are found for both horizontal and vertical direction The covariance matrix is calculated and then the eigenvalues for it This is done in order to rotate the plume with such an angle that it is standing upright i e the bottom end is where it comes out of the tuyere Now the algorithm can be applied on the plume The mean for the on pixels is calculated for the horizontal direction When this is done the volume volume of the left part and the right part are calculated separately using 4 the functions countpixels and then evalvol NSee also COUNTPIXELS EVALVOL COV EIG function volume findvolume2 bwplymim ycoord xcoord find bwplymim 83 84 A MATLAB CODE if length ycoord gt 0 v ycoord xcoord m mean v c cov ycoord xcoord v d eig c alpha 180 atan v 1 2 v 1 1
86. nding the background a hard task Several algorithms have been looked at in order to evaluate which one gives most satisfactory performance The easiest thing to do is just thresholding the images and doing some morpho logical operations to clean up unwanted trash and fix the shape of the background 6 1 FINDING THE BACKGROUND 47 Algorithm 1 Imfilter imfilter im y pim lt zero pad im with 2 y pixels around im mask lt yxy zero matrix surrounded by ones irows rows im icols columns im for r y 1 to trows y for c y 1 to icols y if pim r c then tmp pimfr y rt y c y cty mask if sum tmp 0 then pimfr y rt y c y cty 0 end 12 end 13 end 14 end SO CONT OY Cee Ne HO Typical cleaning up operations are erosion dilation shrinking and filing The problem with this simple algorithm outline is that we believe from our exami nation of the images in Chapter 5 that there is no such a threshold that could guarantee us a good background image Compromises have to be made and more advanced algorithms have to be implemented in order to eliminate the bad effects of the compromises Filtering can be used to remove any unwanted small objects that are not possible to eliminate with the previously mentioned operations Algorithm 1 does such kind of filtering it looks for islands isolated pixels and removes them according to the filter parameter which decides the size of the unwanted objects FIGURE 6 1
87. ng the first video signal is as usual characterised by its frequency content Best correlation Correlation coefficient Correlation coefficient Correlation coefficient e ooo 8 Se 8 7 2 EXTRACTED Colour 1 2 Threshold level a Rotalgo background vs First recording 8 8 Colour 2 Threshold level c Rotalgo vs Imweight both with fixed background First recording Colour Threshold level e Ellipalgo with customized back ground vs Area with fixed background First recording DATA CHARACTERISTICS Area both with fixed 61 Correlation coefficient Threshold level b Rotalgo vs Area both with fixed background Third recording 0 98 0 96 fi ce p bd R Correlation coefficient Colour 1 Threshold level d Rotalgo vs Imweight both with fixed background Third recording o o e R2 BR o a Lot ots o E EEO ibe 2 Correlation coefficient e o pL Threshold level f Ellipalgo with customized background vs Area with fixed background Third recording FIGURE 7 1 1 Correlation between different algorithms 62 7 DATA EXTRACTION AND VALIDATION INN N JI all baal 0 05 0 0 05 01 0 15 0 2 0 25 0 25 02 0 Frequency Hz Frequency Hz Amplitud ra a
88. ntrol pressure and flow signals in pipes 2 and 3 were thought to be sufficient regarding the small number of signals that could be gathered using the available equipment The main reason for this logging session was to record a new video signal of the injection without using a green glass filter instead we changed the glass in front of the camera in pipe 2 to a transparent one The glass 3 2 MEASURED SIGNALS 21 in front of the other camera pipe 3 remained unchanged Bad luck and possible wiring problems resulted in a missed signal this time it was the coal injection vessel weight A summary list can be found in Table 2 Other activities at the plant linked to our work can be found in Table 3 Collected Signals Tuyeres Recorded 509 06 01 7 51 05 18 47 23 999 06 10 12 45 46 14 49 08 999 11 18 13 24 34 15 04 24 2 3 5 6 8 9 10 TABLE 2 Data collection occasions A REN 18 47 18 51 20 5 8 04 13 19 14 30 14 36 z 3 No E Boo 105145 TABLE 3 Activities and setup values during the eine ime During the third logging session we had an opportunity to disturb the coal flow in one of the pipes We chose pipe 2 This test series originated from a need of data for an extra validation of achieved results in a different project ran by Ph D student Andreas Johansson related to his research in clogging detection in a pressurised system 6 at Lule University of Technology We refer to this test series as Andre
89. ol signal b With pressure signal 0 5 T ia Ve 0 3 A oa 5 E 0 1 3 OF 5 of 041 o2t ed VJ bal Jas sec 0 4 i i i i i i i 500 400 300 200 200 300 400 500 100 100 Time shift seconds c With flow signal FIGURE 7 3 2 Correlation between the extracted flow signal and different measured signals during Andreas test We found it suitable to highlight the correlation between the collected signals and the extracted one during the Andreas test as we did before for the other video sequences Because the valve closing takes place on different valves that are placed differently relative to the measurement device we will have different plots compared with those we got earlier As Figure 7 3 2 illustrates the correlation with the control signal is strongly negative at 6 seconds that is a delay of 6 seconds in the control signal after an action of closing or opening has been taken Looking further into the properties of the pressure signal no particular corre lation can be found This does not surprise however since we know that choking the coal flow after the pressuremeter will result in increased pressure while doing 66 7 DATA EXTRACTION AND VALIDATION 0 3 A ae J A A A 0 2 FA i on ee i eee FA OFA pi f EJ Correlation coefficient b b R 2 YI i i i Correlation coefficient S gt R r y JU
90. onsists of two different fields captured at different times Thisis a drawback in digitising analogue video signals that use interlace We had to separate each captured frame into two fields but this did not affect our results A reason for that is We sampled at a very low rate compared to the field rate and it is not very likely that much of the image in our case has changed in 1 50 seconds Figure 3 4 2 shows how this effect is present in our case Another important factor when it comes to digitising is the frame grabber equipment The used frame grabber card a PCI Matrox Meteor should be con sidered as appropriate in our case The card could make 24 bpp at 768 576 pixels resolution This gave us 256 256 256 RGB colours At the time this project started the device driver available for the grabber card for Linux was not of a very high quality We were forced due to some possible hardware limitations to increase the delay time for the card in order to make the desired resolution men tioned above This could sometimes lead to unwanted effects which resulted in grabbing an image combined of two frames see Figure 3 4 3 We had to accept this bug for the time being specially because it does not appear so often and could not have a major influence on the total result We did not use the real time functionality in RTLinux for several reasons The program we wrote for grabbing the video frames gave a good synchronisation a 3 4 VIDEO D
91. ow measurement Decisions based on the analysed data can then be taken with good accuracy To make this analysis fast and practical a small user interface was written in MATLAB It was given the name Danalyzer from the two words data and analyse and will be described later in Chapter 4 2 4 1 Analysis of Signals Performing the analysis of collected signals it is close at hand to start by looking at the plain signals comparing their shapes and trends Trying to find possible sim ilarities which could reveal interconnections and dependencies between the signals is basically the first step To start with the pressure signals were viewed First signal collecting occasion gave us only the last twelve signals according to their numbers in Table 1 As mentioned before the pressure in injection pipe 1 was not stored due to hardware problems The other two pressure signals looked quite the same the pressure in injection pipe 3 being slightly higher than the pressure in injection pipe 2 We do not think that should be the case The level not considering the fast fluctuations was quite constant Then the mass flow for the three pipes was examined It seemed to be almost the same in all pipes although of course small differences could be noticed Moving over to the control signals it could be noticed that the control signal for pipe 1 was behaving in a much better way than control signals for pipes 2 and 3 Looking at those signals it is clear that the
92. pi rotated imrotate bwplymim alpha crop gt 0 i j find rotated im mean i jm mean j rotax round jm if rotax gt 0 left countpixels rotated 1 rotax right countpixels rotated rotax 1 size rotated 2 else left 0 right 0 end 1l evalvol left r evalvol right volume ltr else volume 0 end APPENDIX B C Code B 1 ssnap c include lt fcntl h gt include lt stdlib h gt include lt unistd h gt include lt stdio h gt include lt sys mman h gt include lt sys ioctl h gt include lt sys time h gt include ioctl meteor h define ROWS 576 define COLS 768 define DEV dev mmetfgrab0 define SRC METEOR_INPUT_DEVO define OFMT METEOR_GEO_RGB24 define IFMT METEOR_FMT_PAL define CAP METEOR_CAP_SINGLE define FRAMES 1 int main int argc char argv FILE fp stdout int noframes 1 float pause 5 int src SRC ifmt IFMT cap CAP int size COLS ROWS struct meteor_frame_offset off struct meteor_geomet geo char mem buf str ii int dev i j unsigned char ptr mmbuf struct timeval time struct timezone zone double ot if arge gt 0 noframes atoi t argv if arge gt 0 pause atof argv dev open DEV O_RDONLY geo rows ROWS geo columns COLS geo frames FRAMES geo oformat OFMT ioctl dev METEORSETGEO amp geo ioctl dev METEORGFROFF off ioctl dev METEORSINPUT amp src ioctl dev
93. r projection computation of the volume is reduced to a straightforward calculation volume 3 m a b where a is the length of minor axis and b the length of major azis CHAPTER 7 Data Extraction and Validation Image processing had to be used for recovering data from the recorded image sequences To start with recorded video signals were digitised according to Chapter 3 4 Performing tests on data from all three data collecting occasions and after some initial stages have shown similar results on video signals from different tuyeres We could restrict the final tests to one tuyere from the first and last video recordings Using the procedures and algorithms developed in Chapter 6 we could extract data from the images We tried tuyere number 2 from the first and third video recordings representing two independent data sets because they had different characteristics Film sequences were sampled with 2 second intervals and were 600 images long i e the studied sampled sequences were 20 minutes long except for the sequence of Andreas test which consisted of 650 images corresponding to about 22 minutes The earlier explained background extracting methods fixed and customised were used to find the backgrounds in order to isolate the coal cloud in each picture The coal cloud was thresholded at eleven different levels from 0 2 up to 0 7 with step size of 0 05 regarding 0 as the darkest level and 1 as the brightest with 256 grey scale levels
94. reflects the amount of pulverised coal present We see a clear colour shift when changing the glass The histograms of the hue and saturation are not fully occupied a possible opportunity for histogram stretching appears We can see which values are the most freguent ones when coal appears in an image but the problem is that those values are not only reserved for the coal Other elements in the image share the same space Looking at the height and the width is the first approach Also we could think of comparing the low intensity part of the histogram to the high intensity part This could be done by forming a ratio between those after dividing the histogram in half A high ratio value would indicate coal whereas a low one would mean that a minor amount of coal is present It has to be pointed out that again we do not 5 2 IMAGE CONTENT 41 know which part is coal and which are just gases formed inside the blast furnace since both seem to be present in the same place in the histograms Red Component Red Component 2000 4000 WA 2000 if on 4 if E E 0 anan Aa MSM ra er ae AIN a i J 0 TR ee 2000 y 1000 i i i008 0 50 100 150 200 250 300 0 50 100 150 200 250 300 Green Component Green Component 5000 T T 4000 T T 2000 Se SSE a N k JI Oo At ii TTT rere rr 5000 i 2000 V 210000 i i i ania i i i 0 50 100 150 200 250 300 0 50 100 150 200 250 300 Blu
95. ription Assuming that the projection of the plume is always oval we can use a simple algorithm for calculation of the volume All we need is the length and the width of the plume a is the width of the plume b is the length of the plume NSee also IMFEATURE function volume algo2vol plym x y find plym if x gt 0 master imfeature plym MinorAxisLength a master MinorAxisLength master imfeature plym MajorAxisLength b master MajorAxisLength area pi a b volume 4 areax a 3 else volume 0 end A 19 FINDVOLUME2 M A 17 evalvol m Purpose Evaluation of volume given a vector of radia Synopsis 4 v evalvol vector Description 4 Find the volume of half circles given their radia in a vector Then add all these values together to get an approximate volume for one half of the body of revolution NSee also COUNTPIXELS FINDVOLUME2 function vol evalvol vector vol 0 for n 1 length vector vol vol pi vector n 2 2 end A 18 countpixels m Purpose Count the on pixels in every row of an image and store the result as numbers in a vector Synopsis vector countpixels plym begincolumn endcolumn Description Given a binary image of a plume the beginning and the end columns I for which we want the calculation to be performed a vector of number of f on pixels in each row between these two columns is calculated If no start or e
96. rocess itself is quite complicated and not fully understood New technology has contributed in many ways to improve the steel making procedure where involvement of people with different backgrounds and academic knowledge is essential In existing blast furnaces there are many problems which remain unsolved despite many years of thorough research New improvements and breakthroughs are made every day but there will always be more work to be done due to the sophisticated process nature Efficiency quality environmental issues and cost reduction requirements are the main objectives The fuel used in the furnace is one of the targets Changing the kind of fuel used has shown very good results Traditionally coke is used Many other alternative fuels 10 have been tested such as pulverised coal natural gas oil but also waste materials The future supply of coke 1 is another problem that might lead to steadily increasing prices Pulverised coal has become a good alternative It is 30 40 cheaper and more environmentally friendly 12 than coke Using pulverised coal resulted in a 40 saving in coke requirements at British Steel Scunthorpe works 11 13 In addition pulverised coal has a quicker impact on the reaction in the active zone of the furnace Beside choosing an alternative fuel steelmakers need a better overview of the process Controlling the product quality relies on identifying the process param eters from a metallurgical point of
97. rrelation coefficient o fi Correlation coefficient o T ari a The valve after the flow meter b The valve before the flow meter FIGURE 7 3 4 Correlation between the extracted flow signal and the flow during Andreas test Lastly we will without being inconvenient torture the flow measurement and see fit will confess Figure 7 3 2 d shows a positive correlation at 9 seconds which conflicts with our previous result of 9 seconds Again here we had a reason to separate the extracted signal in two parts This has increased the correlation in general Plot a in Figure 7 3 4 gave us a delay of 2 seconds implying that we are 7 3 ANDREAS TEST PROPERTIES 67 able to detect the changes in the coal flow before the meter does The meter seems to continue measuring the flow for a while after the actual flow cut off On the contrary plot b in the same figure tell us that the correlation maximum occurs at 29 seconds Now the flowmeter is faster than our calculated flow We guess that the flowmeter is very dependent on the carrier medium 68 7 DATA EXTRACTION AND VALIDATION CHAPTER 8 Conclusions and Suggestions Image processing has hopefully proven to be an asset amongst other tools that can be used for enhancing the monitoring of the blast furnace process This is quite a new research area which most probably will eventually lead to an improved surveillance and control of the process What is described in this r
98. s It might be an effect of better camera dynamics The only difference here is that green glass images contain more noise especially for the blue component Moving over to the HSI domain it is easy to see that the hue and saturation parts are noisy possibly due to overexposure Use of different types of glass does not affect the level of noise considerably The intensity part is however not disturbed 40 5 IMAGE PROCESSING Coloured image Histogram of The Image Intensities Coloured image Histogram of The Image Intensities 4000 4000 2000 2000 1000 1000 100 150 200 250 100 150 200 250 2000 1000 o 100 150 200 250 a Based on an image taken with a green b Based on an image taken with a green glass without coal glass with coal Coloured image Histogram of The Image Intensities Coloured image Histogram of The Image Intensities 4000 2000 2000 2000 4000 1000 2000 1000 200 250 100 150 200 250 2000 2000 1000 1000 o o o 0 50 100 150 200 250 0 50 100 150 200 250 c Based on an image taken with a trans d Based on an image taken with a parent glass without coal transparent glass with coal FIGURE 5 2 2 A comparison of image intensities in the HSI space It is easily seen that the first hill in the histogram for the intensities
99. sents coal free images and the right side represents images with coal In the RGB domain Figure 5 2 1 searching the histograms of the images we can make some observations Starting with the green glass images and their RGB decomposition the green component is saturated in both images From the red components we can conclude that gases have an intensity value just below 50 in those images The blue component of the coal free image has other characteristics than the other components but the red and green components are slightly similar If we assume that the green glass is not harmful to the image contents then we have an unanswered question What do we see there We will try to answer this question later in Chapter 5 2 2 We notice a slight shift from the high intensities to the low ones when we move from a coal free image to one with coal This is because obviously the first hill represents mostly the coal in the picture The conclusion is that we can by comparing histograms determine if there is coal present or not This phenomenon is even seen in the blue component which is the most noisy one Comparing the images taken with the transparent glass and the green glass we notice that the width of the first hill in the histograms changes The first hill in the histogram located at around value 50 is much wider when coal is present Interesting 5 2 IMAGE CONTENT 39 Coloured image Histogram of The Image Intensities Coloured image Histogram of The
100. server is far enough away this gives the appearance of one pixel colour rather than three adjacent ones These three colours are primary colours mixing any combination of two of them does not make the third In fact any three colours could be used providing they are independent i e primary RGB is the most widely spread colour system HSI stands for hue saturation intensity HSI may be regarded as the same space as RGB but represented in a different coordinate system Hue is effectively a measure of the wavelength of the main colour i e there are different numbers representing different colours E g lets say 0 represents red then move all the way through different colours up to 256 in case we have 256 colours which again represents red Imagine a coloured disc that warps around here 256 is mapped to 0 Saturation is a measure of hue in each spot If saturation is 0 then the final colour is without hue i e it consists of white light only Intensity is simply a measure of brightness of each pixel Consult Figure 5 1 1 for a visualisation of the colour spaces 5 1 2 Image Quality Colour image capture involves the capture of three images simultaneously With RGB an early industry standard intensity of each of red green and blue has to be measured for each spot These are then stored in three 35 36 5 IMAGE PROCESSING i SEE uu are Sn H i Pai j r ae Hy L Au p WI gt Saturation oo i G S p i p
101. sonds e Pressure flow first logging f Pressure flow third logging FIGURE 4 1 5 Correlation between pressure flow and control sig nals for the first and third collected data 4 1 ANALYSIS OF SIGNALS 33 By looking at the correlation between different signals we can see how hard they are coupled and see if there exist any delays Figure 4 1 5 shows clear correlation between the control and pressure signals The control signal does not follow the flow signal as well as the pressure does This indicates that either the control is bad or there is something else beside the flow signal that affects the control signal which we will see is true here or both Almost the same correlation is found between the pressure and the flow signals The same figure shows delays between the different signals Notice that the the correlation plots are different if you compare those from the first nd the last occasion there is a clear frequency present in the Figure 4 1 5 a c and e It is most likely the frequency we have pointed out earlier that we call Mefos carrier frequency Beside the difference in the frequency the delays seem to be pretty much the same in both the first and third data collection The pressure is very dependent on the control signal with a delay of 3 seconds In the control flow correlation we see the clear feedback effect represented as a negative correlation with a very short delay and the positive correlation with a f
102. t Figure 2 1 2 The upper vessel the one that is filled with coal when needed works as an airlock vessel used to pressurise the coal vessel below From the lower vessel called the injection vessel coal is divided and distributed under pressure through pipes to the tuyeres in the blast furnace with the help of nitrogen which serves as a carrier medium The left vessel is used to inject slag 16 2 PROCESS DESCRIPTION when desired A big problem is determining the behaviour of the coal particles travelling through the pipes That is because of the various size of the particles turbulence effects in the pipes and pipes characteristics Coal particles can clog in a pipe which can disturb the process before being discovered Another problem is leakage of the carrier gas Solution for the latter is proposed in 6 2 3 Current Control The existing control of the pulverised coal flow to the blast furnace is based on a continuous on line measurement of the coal flow itself Although this is true it is not the whole truth It has been shown that the flow measurement device is not very accurate that is why the current control is dependent on a weight measurement of the injection vessel 2 3 1 Sensors Mainly we can talk about three flow measurement devices every one of them connected to pipes transporting pulverised coal to the tuyeres The devices used are Ramsey DMK 270 industrial mass flow rate and velocity mea suring instruments for non
103. ted but this was not discovered before it was too late after 1 2 closed after 1 1 opened after 1 1 closed 6 1499 00 after 1 2 dosed e 1433 00 after 1 1 opened o aoo before 1 2 closed TABLE 4 Actions and time schedule during Andreas test 3 3 Video Recording At the same time as data was collected using the data collecting module a video recording had to be made This was clearly a sensitive issue to deal with The time delay between the start of collecting other data and recording the video signals from the cameras had to be determined in some way because those signals were to be compared during later analysis The only viable alternative was to start gathering data at a certain point in time then start the video recording and marking a fixed point in time on the film sequence by manually dimming the light to the cameras for a couple of seconds That way the time delay was restored later when analysing data Using a more sophisticated synchronisation method would be a better option if it was not such a lengthy procedure remembering that the electrical cabinets and the cameras are apart from each other making a complex wiring scheme doomed to fail in such environment and also having in mind the quite small benefits employing an approach different from the simple solution used Video cameras presently installed at Mefos are of the type Panasonic WV CL 410 These cameras were used at both occasions
104. to do when applying Andreas test Because we had no access to documentation regarding the flowmeter and the control system in general at the time we ran into this we started to investigate the case A filter was a good way to start in Connecting a filter to a measurement device 34 4 DATA PROCESSING 1 z 1 e 22 3679 is quite common A possible filter could be H z obtained by z e 22 3679 identification Inverting the filter and filtering the flow measurement with it should hopefully recreate the supposed original signal Figure 4 1 6 a hows how the filtered signal is related to the original signal and Figure 4 1 6 b shows how this fits with the valve position after zooming on the interesting parts A further knowledge of the control system has shown that as mentioned in Chapter 2 3 the feedback signal sent to the controller is not the pure flow measurement we have examined here Having a look in the user manual for the flowmeter the filters are integrated in the sub measurement used in the device in order to calculate the flow For our knowledge the measurement device can not be separated into a flow measurement and a filter 4 2 Danalyzer File Analysis Measurement 1 Measurement 2 Pressure injection pipe 1 Pressure injection pipe I The Visualization Project Zoom Clear Stat UA 04 Jun 1999 17 54 08 Grid Hold Stop ff 04 Jun 1999 18 47 23 FIGURE 4 2 1 A screenshot of Danalyzer
105. ul people at Mefos LKAB Securitas and Bj rn Olsson at SSAB Re Tek members for keeping up the spirit of being atrocious and taking the computers we borrowed before we finished the project El Tek members 1PROSA s home page URL http www sm luth se csee prosa html Department of Computer Science and Electrical Engineering s home page URL http www sm luth se 3Mefos home page URL http www mefos se Free Software Foundation s home page URL http www fsf org 5Octave s home page URL http www che wisc edu octave octave html 7 8 PREFACE for making it possible to buy provisions during the time we spent writing this report Last but not least we want to thank our friends for their psychological support Thank you all you made us do it Contents Chapter 1 Introduction Chapter 2 Process Description 2 1 The Blast Furnace 2 2 Coal Injection 2 3 Current Control 2 3 1 Sensors 2 3 2 Controllers 2 4 Video Surveillance Chapter 3 Collecting Data 3 1 Available Signals and Equipment 3 2 Measured Signals 3 3 Video Recording 3 4 Video Digitising Chapter 4 Data Processing 4 1 Analysis of Signals 4 1 1 Filter Identification 4 2 Danalyzer Chapter 5 Image Processing 5 1 Image Decomposition 5 1 1 RGB and HSI spaces 5 1 2 Image Quality 5 2 Image Content 5 2 1 Image Histograms 5 2 2 Image Threshold Chapter 6 Algorithms 6 1 Finding the Background 6 1 1 One Background Approach
106. urnace was thought to be of interest Further some other signals were considered in case they would later show to be significant for the study of the coal powder flow For convenience Table 1 with available signals is attached Of course there were many other signals available but we did not think that they were of interest 209 a e 302 31FVi40 Control sig pipe 3 0100 020 M 031 zr Hao springs Pressure Took Yess 0 10 bar TABLE 1 The considered signals and then de oron o ome 021007 e s Unfortunately the collected signals could not be taken straight from Mefos computer system The reason for this was simply the deficient capacity of the presently used system In this case it was necessary to wire a signal collecting 19 20 3 COLLECTING DATA equipment into an electrical cabinet consisting of a National Instrument data ac quisition box SCXI 2000 equipped with SCXI 1120 an 8 channel isolation amplifier and SCXI 1100 a 32 channels programmable amplifier with gains Those modules contained an IO board and an analogue to digital converter with amplifiers In to tal we had two different opportunities to log signals we did that on three different dates The 8 channel card was used in the third logging session while the other one was what we had in the first and second session A portable computer was used to store data with the help of the data collecting programme LabView This was the best option available
107. view and how well the process is controlled Controlling the process beside unidentified process parameters runs into problems related to flow measurement temperature measurement and fuel distribution in the blast furnace which are hard to deal with using old fashion techniques because of the high process complexity and the very demanding environment Quick development in computer hardware has opened new perspectives and possibilities At present it is an easy task to process a large amount of data to extract useful information in order to control and supervise the steel production in a blast furnace One way to do this is the use of cameras pointed to the pulverised coal outlet from the tuyeres into the furnace Camera images can be analysed in order to determine different important process parameters The goal is to calculate high quality parameters that reflect what happens in the furnace This will make life easier for metallurgical and automatic control people A prestudy 2 has shown that there is a significant relation between the pul verised coal mass flow estimation and the size of the coal plume in the analysed video recorded series of images The recording was done with a black and white camera Deeper investigation is required to verify the results and find algorithms for calculation of the coal plume volume and coal flow estimation but also tem perature and coal powder distribution It is also interesting to study the result 11 12
108. what form this cloud possesses in three dimensions having only access to its two dimensional projection We have to make an assumption about its shape Three different approaches to estimate the volume of the pulverised coal cloud will therefore be discussed here None of these gives us the actual flow Using these algorithms we will obtain numbers representing the flow which can be compared to numbers in the same series i e we can conclude if the flow is increasing decreasing or staying at the same level However small future adjustments should allow using these values for real measurement where the flow is given in grammes per second What still has to be done is deciding what value corresponds to what flow which should be a part of a further study Being able to translate evaluated values to grammes per second we will have a good measurement of the flow just using image processing 6 4 1 Weighted Pixel Estimation The first more advanced attempt to make an approximation of the volume of the pulverised coal cloud seen in the grabbed images discussed in this report will be the Weighted Pixel Estimation Weighted Pixel Estimation is based on the fact that the cloud is not uniformly dark This implies that in the middle of the cloud there is more coal than on the edges where it is not at all as dark We can establish that the size of the cloud 56 6 ALGORITHMS is dependent on the chosen value for thresholding of the image Giving all the pixe

Enhanced Control by Visualisation of Process Characteristics: Video

Contents

Download Pdf Manuals

Related Search

Related Contents