h Naonal Conference on Elecrcal, Elecroncs and Compuer Engneerng (NCEECE The Analyss of he Thcknesspredcve Model Based on he SVM Xumng Zhao,a,Yan Wang,band Zhmn B,c School of Conrol Scence and Engneerng, Unversy of Jnan, Jnan Chna School of Conrol Scence and Engneerng, Unversy of Jnan, Jnan Chna School of Conrol Scence and Engneerng, Unversy of Jnan, Jnan Chna a emal: jd_zxm@6.com, bemal:76@qq.com, cemal:9696@qq.com Keywards: ho rollng; hckness predcon; SVM; AGC; Absrac. Ths paper presens a mehod whch mproves he hckness of he predcon model. A he same me, bulds up a SVM predcon model based on MATLAB smulaon expermen. By comparng he smulaon curves of he mechansm model and SVM regresson predcon, apparenly can be seen ha hcknessoupu precson effecvely mproved, whch s good o onlne denfcaon of he hckness conrol.. Inroducon Meallurgcal ndusry a home and abroad pay more aenon o he qualy of he plae srp, especally he dmensons, he hckness accuracy has become one of he mporan ndcaors. And precson problem of hckness conrol has resrced he produc qualy mprovemen. Through mprovng he predcve accuracy of srp hckness and achevng more precse conrol, he plae band seel producon play a crucal role. The hckness predced sysem s he mporan mehod of plae srp hckness accuracy based on mechansm model, s purpose s o realze he precse conrol onlne. However, envronmen s more complex due o he connuous srp producon, he mechansm model s dffcul o esablsh. The mechansm of he model we used s go afer gnorng and smplfyng some of he facors.in connuous producon, rollng echnology s more complex and nclude unknown facors of he uncerany, hese facors are dffcul o esablsh accurae mahemacal model. Therefore, he use of a large amoun of wealh n he process of acual producon daa modelng and conrol become nevable choce.. The SVM predcon model based on he hckness Vapnk pu forward a knd of novel machne learnng mehod of suppor vecor machne (SVM: he basc dea of SVM npu sample space s mapped o hghdmensonal feaure space by usng kernel funcons, o fnd an opmal classfcaon plane n hgh dmensonal space. Through SVM processng, o oban he nonlnear mappng relaonshp beween varable npu and varable oupu. I s ofen used when he suppor vecor machne (SVM algorhm n dealng wh small sample, nonlnear and funcon fng problems.in fac, he algorhm of suppor vecor machne (SVM s a convex quadrac opmzaon problem, he soluon s global opmal soluon. A presen, suppor vecor machne (SVM ge more successful applcaons n bonformacs face and gesure recognon, ex classfcaon and so on. Suppor vecor machne (SVM regresson algorhm of lnear and nonlnear. For lnear regresson problem, gven he ranng se {( x, y, ( x, y,, ( x, y } Among hem x R n y R Regresson lnear funcon can be used by f = ( x ω T x + b. I can be go he bes regresson funcon by solvng he mnmum value as formula(. 6. The auhors Publshed by Alans Press 9
ϕ( ωξξ,, = ω + C ξ ξ + ( = = where C s Penaly facor, ξ, ξ s Respecvely upper and lower bounds of slack varable. The consran as formula(. f ( x y x + ε y f ( x x + ε, =,, k ( xx, Lnear separable problem s also a convex quadrac opmzaon problem. To solve he problem, we nroduce he Lagrange funcon, as formula(. Lω, b, xx,, aa,, γλ, = where answerng o, ( ( ( ω C x + x a x + ε y + f x = = a x ε y f ( x xγ xλ = = + + + ( a a, γ, λ, =,, k. The Lagrange funcon L solve he mnmum value ω, b, ξ, ξ, as formula(. L L L L =, =, =, ω b ξ = ( ξ I can be descrbed by he formula above, as formula(. ( a a = = ω = ( a a x ( = C a γ = I can ge a convex quadrac opmzaon problem dual form by plugng he above formula n he Lagrange funcon, as formula(6. ω ( aa, = ( a a ( aj aj( x xj, j= (6 ( a a y ( a a + = j= As he consran condons n formula(7. = ( a a =, a ε, a C (7 Regresson funcon can be obaned by calculang he value of he formula and plugng n he formula (. n ( = ( ( (, f x = a a x x + b For nonlnear regresson problem, need o complee he followng wo aspecs: frs, wll realze he mappng from he orgnal sample space o a hghdmensonal space usng a nonlnear mappng funcon. Second, should analyze by usng he mehod of lnear regresson n he hgh dmenson space. So he key of deallng wh nonlnear regresson problem s o ask he mappng funcon, hus he problem become opmze he formula( of under he consran condons as formula(9. 96
ω ( aa, = ( a a ( aj aj k( x xj, j= ( a a y ( a a + = j= Where ω = ( a a ϕ( x ( ( (, ε (9, The regresson lnear funcon can be used by =. = f x = a a k x x + b Kernel funcon of suppor vecor machne use he gaussan kernel funcon and polynomal kernel funcon and sgmod kernel funcon, ec. Gaussan kernel funcon: K( xx, = exp x x σ Polynomal funcon: K( x, x (. d = xx + Sgmod kernel funcon: (, anh a (. K x x = xx + C K( xx, = + exp α ( xx. + C. The SVM predcon model based on malab smulaon expermen I should esablsh srp fnshng ex hckness predcon model by Lbsvm oolk. Algorhm process as shown n fgure. Daa preprocessng Cross valdaon o he opmum regresson Use he bes paragrameer ranng Fng predcon Fgure he SVM predcon model algorhm flow char The smulaon of MATLAB code: % % o fnd he bes parameer c and g of regresson problems: [besmse,besc,besg] = SVMcgForRegress(TS,TSX,,,,; [besmse,besc,besg] = SVMcgForRegress(TS,TSX,,,,,,.,.,.; % % shoukd use he bes parameers c and g for ranng of SVM: cmd = ['c ', numsr(besc, ' g ', numsr(besg, ' s p.']; model = svmran(ts,tsx,cmd; % % he SVM regresson predcon: [predc,mse] = svmpredc(ts,tsx,model; predc = mapmnmax('reverse',predc',tsps; The smulaon resuls and analyss: 97
SVRparameer selecon resul(dvew[grdsearchmehod] Bes c=9.96 g=. CVmse=.7.. MSE 6.. logg 6 logc 6 SVR parameer selecon resul(d[grdsearchmehod] Bes c= g=. CVmse=.6.. MSE 6.. logg Fgure Parameer selecon resul fgure of rough and fne hckness of seel The /...6... Real expor SVMforecas model predcon Fgure Thckness oupu curve of he predcon and he realy logc.6 Number of samples.. SVM predcon Mechansm Model error/mm.. Number of samples Fgure Thckness oupu error curve of he predcon and he realy 9
The relave error/%.... Number of samples SVMpredcon Mechansm model Fgure hckness oupu absolue error curve of he predcon and he realy By he smulaon resuls conras curve can be concluded ha he predcon algorhm based on suppor vecor machne (SVM can be a very good predcon, he expor of seel srp hckness and comparson wh acual measuremen resuls of error and absolue error generally smaller han mechansm model. Descrpon based on suppor vecor machne (SVM algorhm n opmal problems of srp seel hckness conrol has sgnfcan advanages. Concluson Ths paper esablshes he predcon model based on SVM and hckness. Bu for complex acual rollng process, 's jus aken he frs sep and sll have a lo of work o do.i can do layered wh a large number of hsorcal daa from he feld, and realze he offlne collecon sysem of layered modelng. Reference [] dongdong lu, Wang Yan. Based on RBF neural nework srp fnshng hckness predcon [J]. Journal of jnan unversy: naural scence edon, 6, ( :. [] yngyng lu, Wang Yan. The hckness of he ho rolled recognon and nellgen forecas [J]. Journal of jnan unversy: naural scence edon,, ( :. [] Ibssem Ladlan, Larb Houch, Lakhdar Djeml, e al. Modelng daly reference evaporanspraon (ET n he norh of Algera usng generalzed regresson neural neworks (GRNN and radal bass funcon neural neworks (RBFNN: a comparave sudy[j]. SprngerVerlag,,:67. [] Wllams James Mchael Zhang Deyong L Shuca, ec. The genec RBF neural nework based on rough se heory n he applcaon of rock burs predcon [J]. Rock and sol mechancs,, ( : 776. [] lou j w, facng souh, Roger wang. Based on he daa drven modelng and conrol of srp hckness [6] Mehod [J]. Journal of Bejng unversy of echnology,, ( : 99. [7] Bahua Zhang. Predcon of Fregh Ably n Counry Base on GRNN[J]. IFIP Inernaonal Federaon for Informaon Processng,,6:. 99