英文翻译--神经网络PID在温度控制系统中的研究与仿真.doc
ExplorationAndSimulationofNeuralNetworkPIDInTemperatureControlSystemAbstract:ThispaperpresentsanewkindofintelligencePIDcontrolmethodonBPneuralnetworkandsomeofbasicconceptsaboutBPneuralnetwork.NeuralnetworkintelligencePIDcontrollerhasmanyadvancedpropertiescomparedwithtraditionalPIDcontroller.TheBPneuralnetworkPIDcontrolmethodisappliedtotemperaturecontrolsysteminindustryfield.Thesimulationresultsshowthatthecontrolmethodhashighcontrolaccuracy,strongadaptationandexcellentcontrolresults.Keywords:Neuralnetwork,PIDcontroller,Temperaturecontrolsystem1ForewordInindustrialprocesscontrol,PIDcontrolisabasiccontrolmethod,itsrobustness,simplestructure,easytoimplement,buttheconventionalPIDcontrolalsohasitsowndisadvantage,becausetheparametersofconventionalPIDcontrollerisbasedonbeingmathematicalmodelofcontrolledobjectidentified,whenthemathematicalmodeloftheobjectarechanging,non-lineartime,PIDparametersisnoteasyinaccordancewithitsactualsituationandmakeadjustments,theimpactofthequalitycontrolsothatthecontrolofthequalitycontrolsystemdecline.Especiallyinthepuretime-delaycharacteristicswiththeindustrialprocess,theconventionalPIDcontrolmoredifficulttomeettherequirementsofthecontrolaccuracy.Becauseofneuralnetworkswithself-organization,self-learning,adaptivecapacity,Inthispaper,basedonBPneuralnetworkPIDcontroller,sothatartificialneuralnetworkPIDcontrolwiththetraditionalcombinationofeachotherandjointlyimprovequalitycontrolandtothemethodinthetemperaturecontrolsystemusingthesimulationlanguageMatlabapplication.2BPneuralnetworkmodelandalgorithmconstitute2.1BPneuralnetworkmodelconstituteBPneuralnetworklearningprocessconstitutedmainlybytwostages:Thefirstphase(forwardpropagation),theinputsignalthroughtheinputlayer,hiddenlayerafterlayer-by-layertreatment,intheoutputlayeriscalculatedforeachneurontheactualoutputvalue.Thesecondstage(theprocessoferrorback-propagation),ifnotintheoutputlayerthedesiredoutputvalue,theactuallayer-by-layerrecursiveoutputanddesiredoutputofthemargin,andtherighttoadjustthebasisofthiserrorfactor.2.2TheneuralnetworkPIDcontrollerstructureandalgorithmInthetraditionalPIDcontrol,classicalincrementalPIDcontrolforms:u(k)=u(k-1)+pe(k)-e(k-1)+ie(k)+de(k)-2e(k-1)+e(k-2)Kp:proportionalcoefficienti=iop:Integralcoefficientodpd:DifferentialcoefficientSetupBPneuralnetworkPIDcontrollerstructure:r(k)e(k)u(k)y(k)+_y(k)Adaptiveinordertoachievedip,ofthepurpose,theoutputlayerforthethreeneurons,correspondingtodip,.Inputlayer,hiddenlayerneurons,thenumberofchargedobjectsinaccordancewiththecomplexityoffixed.Hiddenlayeractivationfunctionusedforthepositiveandnegativesymmetricalsigmoidfunction:xxxxeeeexxf)tanh()(Outputlayeractivationfunctionoftheuseofnon-negativesigmoidfunction:xxxeeexxg2)tanh(1)(Weassumethato31,o32,o33istheoutputofoutputlayer,whichcorrespondtop,i,d.Wetaketheperformanceindexfunctionasfollows:2)1()1(21kykrJWhentheactualoutputandthedeviationbetweenthedesiredoutput,thentheerrorback-propagation.Reversethespreadofthesubstanceisbyadjustingtheweightssothatthesmallestdeviation,itcanusethesteepestdescentmethod,errorfunctionbyanegativegradientdirectiontoalllevelsofneuronweightstoadjustoramend.Thenhave:NNPIDNNPlantNNArithmetic)1()3(kwli=-)()3()3(kwwJlili:Learningrate:MomentumofAvailablebythechainrule:)3(liwJ=)3()3()3()3()3()()()()()1()1(lillllwknetknetOOkukukykyJ=-e(k+1)3()3()3()3()3()()()()()1(lillllwknetknetOOkukukyOne:l=1,2,3SoBPneuralnetworkcanbetheoutputlayerweightsofthecalculationformula:)()()1()3()2()3()3(kwkOkwliilliOfwhich:)(*)()(*)()1(sgn()1()3(,)3()3(knetgkOkukukykelllBecauseofthePIDcontrol)()1(kukyalgorithminnormalcircumstancesareunknown,canbeusedtoreplacefunctionsymbols)()1(sgnkuky,andthroughadjustmentstocorrecterrors.Empathycanbehiddenlayerweightcoefficientcalculationformula:)()()2()1()2()2(kwkOwijjiijOfwhich:)()()3(31)3()2()2(kwknetflillii,Intheabovevarioustypes,theScorner(1),(2),(3)express,respectively,inputlayer,hiddenlayer,outputlayer,l:Thenumberofoutputlayerneuronsi:Thenumberofhiddenlayerneuronsj:Thenumberofinputlayerneurons)(1)(xgxgg2/)(12xffBasedontheabovecanbeBPneuralnetworkcontrolalgorithms:(1)determinetheneuralnetworkarchitecture,initializedweightsoneachfloor.Controlthevolumeofoutput,errorchecktheinitialvalue0.(2)ofthesamplingsystemhasbeen)(kr、)(ky.Calculatedbytheerror)()()(kykrke.AndundertheincrementalPIDalgorithmtotheerrorcomponentinputlayerasinput.(3)AccordingtoallfloorsoftheweightcoefficientsarecalculatedlayersBPneuralnetworkinputandoutput.Outputlayerweight,respectivelyKp、Ki、Kd.AccordingtoincrementalPIDcontrollerformulacanbeoutputu.(4)willserveuasthesupervisionofBPneuralnetworksignal,totheback-propagationalgorithmBP.Onlineaccordingtotheoutputlayer,hiddenlayerofthelearningalgorithmadjusttheweightsoneachfloor,sothattoachieveadaptiveadjustPIDcoefficients.(5)backto(2).3.InthetemperaturecontrolsystemsimulationexperimentIntheindustrialproductionprocess,controltheproductionprocessofallkinds,oftentothetemperatureoftheprocesssuchastimedelaycontroloftheprocess.Setthetemperaturecontrolwaschargedwiththeprocessoftransferfunctionis:)110)(140(3)(sssGSe60Thesimulationresultsasfollows:Figure(1)Figure(2)Figure(1)fortheconventionalPIDcontrol,Fig(2)FortheBPneuralnetworkPIDcontrol.FromthefigurewecanseethatconventionalPIDcontrolarisingfromovershootandtransitiontimethantheBPneuralnetworkPIDcontrolarisingfromovershootandthetransitiontimeismuchgreater,whichcanbeseenBPneuralnetworkPIDcontrolstrongself-adaptabilityandhighcontrolprecision.4.ConcludingremarksInthispaper,therootcontrolofBPneuralnetworkalgorithmtotimedelaythetemperaturecontrolsystemsimulation,experimentalresultsshowthattheBPneural