1.1 What is a neural network? 1.2 Historical background 1.3 Why use neural networks? 1.4 Neural networks ersus conentional co!"uters - a co!"arison 2. Hu!an and #rti$icial Neurones - inestigating the si!ilarities 6-% 2.1 How the Hu!an &rain 'earns? 2.2 (ro! Hu!an Neurones to #rti$icial Neurones 3. )echni*ues and a""lication +-13 I!"le!entation o$ #NN on ,teel -late -rocess .ontrol 3.1 Network /esign 3.2 Network -ro"erty 3.3 ,i!ulation 3.4 ,i!ulation $or &last (urnace 3.0 ,i!ulation $or other de"art!ents in steel "late $low "rocess 4. (uture sco"e 13-14 0. .onclusion 14 6. 1e$erences. 10 #""endi2 16-1+ 1 Abstract: With the adent o$ reolutionary steel !aking "rocess3 de!and $or better *uality o$ steel "late has increased. 1eduction in re4ection o$ the $inal "roduct out"ut o$ steel "late "rocess has been an area o$ concern. Neural network is a "ower$ul tool in achieing this target. 5anu$acturing "rocess o$ steel "lates inole arious co!"le2ities .)here are a""ro2i!ately 06 di$$erent grades o$ steel !anu$actured. -roducing steel "lates with better *uality standards with higher acce"tance leel has been a challenge $or steel industries. Neural network with its i!!ense "otential is a $easible tool $or reducing the re4ections in the out"ut o$ steel "lates. #rti$icial neural network with hidden layers with the hel" o$ 5at 'ab so$tware is used to a""ro2i!ate the $unction de"endency o$ the e$$ectie "ara!eters o$ steel "lates 2 1. Introduction: 1.1 What is a Neural Network? #n #rti$icial Neural Network 7#NN8 is an in$or!ation "rocessing "aradig! that is ins"ired by the way biological nerous syste!s3 such as the brain3 "rocess in$or!ation. )he key ele!ent o$ this "aradig! is the noel structure o$ the in$or!ation "rocessing syste!. It is co!"osed o$ a large nu!ber o$ highly interconnected "rocessing ele!ents 7neurones8 working in unison to sole s"eci$ic "roble!s. #NNs3 like "eo"le3 learn by e2a!"le. #n #NN is con$igured $or a s"eci$ic a""lication3 such as "attern recognition or data classi$ication3 through a learning "rocess. 'earning in biological syste!s inoles ad4ust!ents to the syna"tic connections that e2ist between the neurones. )his is true o$ #NNs as well. 1.2 Historical backround Neural network si!ulations a""ear to be a recent deelo"!ent. Howeer3 this $ield was established be$ore the adent o$ co!"uters3 and has suried at least one !a4or setback and seeral eras. 5any i!"ortant adances hae been boosted by the use o$ ine2"ensie co!"uter e!ulations. (ollowing an initial "eriod o$ enthusias!3 the $ield suried a "eriod o$ $rustration and disre"ute. /uring this "eriod when $unding and "ro$essional su""ort was !ini!al3 i!"ortant adances were !ade by relatiely $ew researchers. )hese "ioneers were able to deelo" conincing technology which sur"assed the li!itations identi$ied by 5insky and -a"ert. 5insky and -a"ert3 "ublished a book 7in 19698 in which they su!!ed u" a general $eeling o$ $rustration 7against neural networks8 a!ong researchers3 and was thus acce"ted by !ost without $urther analysis. .urrently3 the neural network $ield en4oys a resurgence o$ interest and a corres"onding increase in $unding. )he $irst arti$icial neuron was "roduced in 1943 by the neuro"hysiologist Warren 5c.ulloch and the logician Walter -its. &ut the technology aailable at that ti!e did not allow the! to do too !uch. 3 1.! Wh" use neural networks? Neural networks3 with their re!arkable ability to derie !eaning $ro! co!"licated or i!"recise data3 can be used to e2tract "atterns and detect trends that are too co!"le2 to be noticed by either hu!ans or other co!"uter techni*ues. # trained neural network can be thought o$ as an :e2"ert: in the category o$ in$or!ation it has been gien to analyse. )his e2"ert can then be used to "roide "ro4ections gien new situations o$ interest and answer :what i$: *uestions. ;ther adantages include< #da"tie learning< #n ability to learn how to do tasks based on the data gien $or training or initial e2"erience. ,el$-;rganisation< #n #NN can create its own organisation or re"resentation o$ the in$or!ation it receies during learning ti!e. 1eal )i!e ;"eration< #NN co!"utations !ay be carried out in "arallel3 and s"ecial hardware deices are being designed and !anu$actured which take adantage o$ this ca"ability. (ault )olerance ia 1edundant In$or!ation .oding< -artial destruction o$ a network leads to the corres"onding degradation o$ "er$or!ance. Howeer3 so!e network ca"abilities !ay be retained een with !a4or network da!age. 1.# Neural networks $ersus con$entional co%&uters Neural networks take a di$$erent a""roach to "roble! soling than that o$ conentional co!"uters. .onentional co!"uters use an algorith!ic a""roach i.e. the co!"uter $ollows a set o$ instructions in order to sole a "roble!. =nless the s"eci$ic ste"s that the co!"uter needs to $ollow are known the co!"uter cannot sole the "roble!. )hat restricts the "roble! soling ca"ability o$ conentional co!"uters to "roble!s that we already understand and know how to sole. &ut co!"uters would be so !uch !ore use$ul i$ they could do things that we don>t e2actly know how to do. 4 Neural networks "rocess in$or!ation in a si!ilar way the hu!an brain does. )he network is co!"osed o$ a large nu!ber o$ highly interconnected "rocessing ele!ents7neurones8 working in "arallel to sole a s"eci$ic "roble!. Neural networks learn by e2a!"le. )hey cannot be "rogra!!ed to "er$or! a s"eci$ic task. )he e2a!"les !ust be selected care$ully otherwise use$ul ti!e is wasted or een worse the network !ight be $unctioning incorrectly. )he disadantage is that because the network $inds out how to sole the "roble! by itsel$3 its o"eration can be un"redictable. ;n the other hand3 conentional co!"uters use a cognitie a""roach to "roble! soling? the way the "roble! is to soled !ust be known and stated in s!all una!biguous instructions. )hese instructions are then conerted to a high leel language "rogra! and then into !achine code that the co!"uter can understand. )hese !achines are totally "redictable? i$ anything goes wrong is due to a so$tware or hardware $ault. Neural networks and conentional algorith!ic co!"uters are not in co!"etition but co!"le!ent each other. )here are tasks are !ore suited to an algorith!ic a""roach like arith!etic o"erations and tasks that are !ore suited to neural networks. @en !ore3 a large nu!ber o$ tasks3 re*uire syste!s that use a co!bination o$ the two a""roaches 7nor!ally a conentional co!"uter is used to su"erise the neural network8 in order to "er$or! at !a2i!u! e$$iciency. 2. Hu%an and Arti'icial Neurones ( in$estiatin the si%ilarities 2.1 How the Hu%an )rain *earns? 5uch is still unknown about how the brain trains itsel$ to "rocess in$or!ation3 so theories abound. In the hu!an brain3 a ty"ical neuron collects signals $ro! others through a host o$ $ine structures called dendrites. )he neuron sends out s"ikes o$ electrical actiity through a long3 thin stand known as an a2on3 which s"lits into thousands o$ branches. #t the end o$ each branch3 a structure called a syna"se conerts the actiity $ro! the a2on into electrical e$$ects that inhibit or e2cite actiity $ro! the a2on into electrical e$$ects that inhibit or e2cite actiity in the connected neurones. When a neuron receies e2citatory in"ut that is su$$iciently large co!"ared with its inhibitory in"ut3 it sends a s"ike o$ electrical actiity down its a2on. 'earning occurs by changing the e$$ectieness o$ the syna"ses so that the in$luence o$ one neuron on another changes. 5
.o!"onents o$ a neuron
)he syna"se
2.2 +ro% Hu%an Neurones to Arti'icial Neurones We conduct these neural networks by $irst trying to deduce the essential $eatures o$ neurones and their interconnections. We then ty"ically "rogra! a co!"uter to si!ulate these $eatures. Howeer because our knowledge o$ neurones is inco!"lete and our co!"uting "ower is li!ited3 our !odels are necessarily gross idealisations o$ real networks o$ neurones. . )he neuron !odel !. Techni,ues and a&&lication I!"le!entation o$ #NN on ,teel -late -rocess .ontrol )he "roble! can be $or!ulated in si!ilar a""roach as in an ;"erations 1esearch !odel. )here are certain controllable "ara!eters which are key decisions to be taken $or the o"ti!u! "rocess out"ut . )hese controllable "ara!eters are actually altered in the "rocess o$ 6 Network 1 &last (urnace In"ut 7=ncontrollable8 "ara!eters A 423 ;ut"ut 7.ontrollable8 "ara!eters A 3 !anu$acturing di$$erent grades o$ steel "late !aking or to get di$$erent s"eci$ications. )here are certain constrains in the "rocess $low which needs to be considered while taking the key decisions and there are arious resource sti"ulations in these constraints which are the uncontrollable "ara!eters. )he -late !anu$acturing "rocess is considered $ro! the &last (urnace to -late 5ill and se"arate networks are trained $or di$$erent "rocess de"art!ents as shown.
-rocess (low .hart !.1 Network -esin #ll the networks are trained and si!ulated in the Neural Network )oolbo2 o$ 5atlab so$tware. @ach network has one in"ut layer3 one hidden layer 3 one out"ut layer and the layer details are as $ollows. 18 )he nu!ber o$ neurons in the in"ut layer is e*ual to the nu!ber o$ training in"ut and these are the uncontrollable "ara!eters $or the de"art!ent in the "rocess $low. 28 )he nu!ber o$ neurons in the out"ut layer is e*ual to the nu!ber o$ target "ara!eters and these are the controllable "ara!eters on which the decision is to be taken. 8 38 )he nu!ber o$ neurons in the hidden layer has been ke"t e*ual to the nu!ber neurons in the in"ut layer. Network )y"e B (eed $orward &ack -ro"agation )rans$er (unction B 'ayer 1 B tansig3 'ayer 2 B "urelin !.2 Network .ro&erties 18 &last (urnace B Network No 1 Neural Network ob4ect< #rchitecture< Nu! In"uts< 1 nu! 'ayers< 2 bias .onnect< C1? 1D in"ut .onnect< C1? 6D layer .onnect< C6 6? 1 6D out"ut .onnect< C6 1D target .onnect< C6 1D nu! ;ut"uts< 17read- only8 nu! )argets< 17read-only8 nu! In"ut /elays< 6 7read-only8 nu! 'ayer /elays< 6 7read-only8 ,ub ob4ect structures< in"uts< E121 cellF o$ in"uts layers< E221 cellF o$ layers out"uts< E122 cellF containing 1 out"ut targets< E122 cellF containing 1 target biases< E221 cellF containing 2 biases in"ut Weights< E221 cellF containing 1 in"ut weight layer Weights< E222 cellF containing 1 layer weight $unctions< ada"t (cn< >trains> init (cn< >initlay> "er$or! (cn< >!se> train (cn< >trainl!> -ara!eters< #da"t -ara!< ."asses init -ara!< 7none8 "er$or! -ara!< 7none8 train -ara!< .e"ochs3 .goal3 .!a2G$ail3 .!e!Greduc3 .!inGgrad3 .!u3 .!uGdec3 .!uGinc3 .!uG!a23 .show3 .ti!e Weight and bias alues< IW< E221 cellF containing 1 in"ut weight !atri2 9 'W< E222 cellF containing 1 layer weight !atri2 b< E221 cellF containing 2 bias ectors ;ther< =ser data< 7user stu$$8 !.! Network Trainin )raining in"uts 7controllable "ara!eters8 are a""lied to the in"ut layer o$ the network3 and desired out"uts are co!"ared at the out"ut layer. )he di$$erence between the out"ut o$ the $inal layer and the desired out"ut is back-"ro"agated to the "reious layer7s83 usually !odi$ied by the deriatie o$ the trans$er $unction3 and the connection weights are nor!ally ad4usted using the /elta 1ule. )his "rocess "roceeds $or the "reious layer7s8 until the in"ut layer is reached. )raining -ara!eters< )raining (unction B trainl!3 @"och 7set alue8 A1663 Hoal 7set alue8 A63 -er$or!ance (unction - 5,@ )raining $or each network was done with batches o$ 0 sets o$ data and with 1666 set o$ data within the s"eci$ied range. )he training is done in such a way that the new set o$ "rocess "redictions can be a""ro"riate done during the si!ulation and the details o$ releant "ara!eters $or training can be better understood by going through the subse*uent section on si!ulation. !.# /i%ulation ,i!ulation hel"s us to $ind out the a""ro"riate "ara!eters o$ a new steel "late with di$$erent s"eci$ication. @ach de"art!ent in the $lowchart is taken one at a ti!e. ,i!ulation is done with 10 sets new data $or which re*uired ariable "ara!eter in "roduct s"eci$ication is already deter!ined as "er custo!er re*uire!ent and an attribute o$ acce"tance o$ 166I is taken. )he identi$ied uncontrollable "ara!eters are listed in )able-I. )he a""ro"riate resources to be utiliJed with their corres"onding re$erence nu!bers are also !entioned !.0 /i%ulation 'or )last +urnace 10 &last (urnace being a de"art!ent catering to arious other "roduct re*uire!ents other than steel "lates3 the custo!er s"eci$ication is not included. )he arious target "ara!eters re*uired $or Hot !etal3 &last (urnace slag3 &last (urnace gas and the "ara!eters relating to the $ed !aterials are taken as uncontrollable 71e$ )able No.BI8. .are needs to be taken $or a""ro"riateness o$ "osted data while si!ulating the de"art!ent under consideration. )he "ara!eters like the Hot 5etal s"eci$ication $or &last (urnace 7"reious de"art!ent83 considered as targeted alue3 can now be taken in actual obtained alue $or the subse*uent "rocesses3 i$ the test results are known by that ti!e. ,i!ulation $or other de"art!ents in steel "late $low "rocess (or the &last (urnace3 .ontinuous cast and -late 5ill de"art!ents the arious releant resource sti"ulations are considered along with the $ollowing 18 )he arious ele!ents in "ercentage3 which constitute the che!ical co!"osition3 are the ariables and the e2"ected alues in the "rocessed "roduct need to be !entioned in the uncontrollable "ara!eter. 28 ,o!e o$ the !echanical "ro"erties like =),3 K, L @longation are o$ ariable ty"e while the I!"act or the bend is attribute where the acce"tance is taken as "ercentage. 38 (or internal de$ects3 the non destructie testing by =ltrasonic the acce"tance is taken as "ercentage. )he network out"ut and targeted alue is "lotted on gra"h $or select "ara!eters are dis"layed in (ig. #. +uture sco&e: )here are large sets o$ "rocess data in thousands aailable in the conentional "lants like the steel "lants and these hae the "otential to train the e$$icient network. In steel !anu$acturing the data within the "roduct is in ery s!all ranges and batch training o$ 0 sets at a ti!e is $ound ery e$$ectie. Within a "roduct $a!ily 7e.g. steel "late83 the !ore the data aailability $or each "roduct3 the network beco!es !ore e$$icient. &atch training o$ 0 sets each is a""licable in this case also. 11 ,till the "rocess $low can be !ade !ore detailed by taking all the o"erations which are releant to the "rocess. 7e.g. -ara!eters $or argon rinsing is not taken in the aboe case study o$ steel "late.8 #.1+ile 1anae%ent /"ste% # $ile !anage!ent syste! based on the steel "late !aking "rocess releance can be designed $or each grade o$ steel "late with se"arate sets o$ three $iles o$ trained networks and one trained network $or blast $urnace. (or si!ulation based on the re$erence o$ grade and "rocess de"art!ent3 a""ro"riate network in the releant $older should get selected. )he in"ut alues can be $ed through $ront end inter$ace so$tware and this gets "rocessed at the back end so$tware. )he network out"ut can be obtained back in the $ront end again. ;nce the so$tware is in its i!"le!entation "hase3 eery network out"ut as si!ulation results is "ro"osed to be "re$erably etted by the "lant e2"erts till a consistent con$idence leel is attained. )he actual obtained alues and !aintained alues in "ro"er $or!at o$ controllable and uncontrollable "ara!eter should go as training to the network as eternal "rocess. 0. Conclusion: # "ersistent a""roach using #NN is a $easible tool to reduce the nu!ber o$ re4ections $ro! a !anu$acturing "rocess. I$ a "rocess is ca"able o$ "roducing a "roduct with *uality con$or!ance o$ at least 2 sig!a leels3 then it i!"lies it has "otential to achiee si2 sig!a leels within a reasonable ti!e "eriod. )here is "ossibility o$ reducing the de$ects3 which got into "roduct because o$ one or !ore "ara!eters. )he real challenge is to list all the releant "ara!eters and analyJe which "ara!eters are in the hands o$ the "roduction "ersonnel and which "ara!eters can not be controlled. /uring training the hu!an interention $or altering 12 (ile Network .onerter (ile Network .ontinuous .ast (ile Network -late 5ill .onerter (ile Network &last (urnace 1. (older $or Hrade I, 2662 (ile Network .onerter (ile Network .ontinuous .ast (ile Network -late 5ill .onerter 2. (older #,)5 - # - 365 (ile Network .onerter (ile Network .ontinuous .ast (ile Network -late 5ill .onerter 06. (older ,#I' HI))@N 696#1 In&ut and Out&ut Inter'ace 'or /i%ulation +i 1.! +I*E 1ANA2E1ENT /3/TE1 +O4 NE54A* NETWO46 +O4 /TEE* .*ATE/ (ile Network .onerter (ile Network .ontinuous .ast (ile Network -late 5ill .onerter 1aster Network /et All 2rades the weights and biases reduces i$ all the releant uncontrolled "ara!eters at the in"ut and re*uisite nu!ber controllable "ara!eters are taken into consideration (or any 5anu$acturing -rocess the aailable in$or!ation can be conerted into nu!erical data and organiJed data handling with a "ro"er training !ethodology will eole an a""ro"riate correlation between the arious controllable and uncontrollable "ara!eters aailable as historical data. )his correlation will be a boon $or the old and conentional "lants i$ they are willing to record all the releant "ara!eters een $ro! now onwards. 4e'erences: 1. &hat3 N. M.3 -. #. 5inder!an3 ). N. 5c#oy3 and N. Wang3 1996. O5odeling .he!ical -rocess ,yste!s ia Neural .o!"utation3P I@@@ .ontrol ,yste!s 5agaJine3 "". 24-36. 2. @. 1.3 and .. #. Massiliadis3 1990. O#rti$icial Intelligance< ,tarting to 1ealiJe Its -ractical -ro!ise3P Chemical Engineering Progress3 "". 22-36. 3. (an3 H. ).3 and Wu3 ,. 5.3 19903 .ase studies on !odelling !anu$acturing "rocesses using arti$icial neural networks. Nournal o$ @ngineering $or Industry3 11%3 412-41%. 4. Hra!ckow3 5. Nansen3 ). -o""e3 .. ,chQ$$ner and H. ,Rrgel 7199%8. Neural networks $or "rocess control in steel !anu$acturing. In< -roc. I@@@ I.#,,- 9%3 100-10+. 0. Na!es #. (ree!an L /aid 5.ska"ura3 Neural network algorith!s3 and "rogra!!ing 7 -#< -earson education #sia3 266283 ". 1%31+392396 6. N.&. ;lieira et al.3 :=se o$ Neural Networks in the -late -roduction ,cheduling at =,I5IN#,3: ,e!inSrio de 'a!inaTUo 7-orto #legre< #ssociaTUo &rasileira de 5etais3 199283 "". 319-339. %. 5aha4an3 5.3 O ,tatistical Vuality .ontrol P 3 /han"at 1ai L .o. 7-8 'td. 2660.3 "". 064 069. +. 5ukesh3 /.3 199%. OHate ,tatistics? )ry Neural Networks3P .he!ical @ngineering -rogress3 "". 96-164. 9. N. -ort!ann et al.3 :#""lication o$ Neural Networks in 1olling 5ill #uto!ation3: Iron and ,teel @ngineer 728 7199083 "". 33-36. 16. -ort!ann3 N.(.3 /. 'indho$$3 H. ,Rrgel3 and ;. Hra!ckow 719908. #""lication o$ neural networks in rolling !ill auto!ation. Iron and steel engineer3 %27283 33-36. 13 11. 1Rscheisen3 5.3 1. Ho$!ann and M. )res" 719928. Neural control $or rolling !ills< incor"orating do!ain theories to oerco!e data de$iciency. In< -roc. NI-, 43 609-666. 12. ,essel!ann3 1.3 (. Wahlers3 H. WRrcher3 and ). -o""e 719908. ;"ti!iJation o$ the electrode control syste! with neural networks. In< -roc. 0th @uro". ,teel .ongress3 266-2%1 13. ,chlang3 5.3 ). -o""e3 and ;. Hra!ckow 719968. Neural networks $or steel !anu$acturing. I@@@ @2"ert Intelligent ,yste!s3 117483 +-16. A&&endi7: Table I BLAST FURNACE Controllable (Target) Uncontrollable ( Input ) Operation 1 Top Preure at! "#$ to 1#% Iron ore 1 Fe & '(#)!in Para!eter ) Blat ti!e *r +) ) SiO) & )#%,-."#% / Te!perature C 1%"".1'"" / P & "#1!a0 ( Al)O/-SiO) & "#$$!a0 % Si1e !! 1" to (" Li!etone ' CaO & /2#%,-.1 3 SiO) & '#%,-."#)% $ 4gO & $#%,-."#% 2 Si1e !! ' to %" 4anganee Ore 1" 4n & /"!in 11 SiO) & /"!in 1) Al)O/ & %!a0 1/ P & "#/!a0 1( Si1e !! )% to $" Sinter 1% Fe & %"4in 1' FeO & 1" 13 SiO) & '#"" 1$ Al)O/ & /#"" 12 CaO & 1(.1% )" 4gO & ( to % )1 Cao-SiO) Baicit5 )#" to )#1 Co6e Anal5i )) Blat No# 1 to 3 Furnace Re7 )/ A* & 1% to 1' )( 84 & "#/ to "#( )% 4 & %,-."#% )' S & "#% to "#' )3 C & 3% to $" Blat Furnace 9a ( Targete: ) )$ CO & )% TO )' )2 CO) & 1% TO 1' /" N) & %% TO %3 /1 ;) & ) TO / 14 Blat Furnace Slag ( Targete: ) /) SiO) & /) to // // Al)O) & 12 to )" /( CaO & /1 to // /% 4gO & 2 to 1" /' 4nO & < 1 /3 CaO-SiO) Baicit5 "#2$ ;ot 4etal /$ Si & "#% to "#$ /2 4n & "#$ to 1 (" S & "#"( !a0 (1 P & "#)" (Targete:) () C & (#"" Co%&arison ra&h: 15