The-application-of-improved-BP-neural-network-in-t

TheapplicationofimprovedBPneuralnetworkintheenginefaultdiagnosisLUDi1,WANGJie11.CollegeofElectricalandElectronicEngineering,HarbinUniversityofScienceandTechnology,Harbin,Heilongjiang150080E-mail:[e-mail],[e-mail]:BPneuralnetworkisthecorepartofthefeedforwardnetwork,andembodiesthecoreandtheessenceofthepartsoftheartificialneuralnetwork.ThegoodnonlinearmappingabilityofBPneuralnetworkcanbeagoodapplicationinfaultdiagnosis.ButthetraditionalBPnetworkhasthetrendofforgettingoldsamplesduringthetrainingprocesswhenlearningnewsamples,andexiststhedefectoflowtrainingaccuracy.Aneuralnetworkalgorithmofincreasedstatefeedbackintheoutputlayerisdesignedinthispapertosolvetheproblemabove.TheimprovedBPalgorithmisusedinthefaultdiagnosisofautomotiveengine,theindexesoftheautomobileexhaustareusedastheinputsoftheneuralnetwork,theoutputscorrespondingtothedifferentmisfire.ThesimulationresultsshowtheproposedalgorithmcaneffectivelyimprovetheBPneuralnetworktrainingaccuracy,andmoreaccuratelytoachievemisfirediagnosis.KeyWords:improvedBPneuralnetwork,trainingaccuracy,misfirediagnosis11IntroductionTheconditionsofautomotiveengineoperatingareverycomplex,anditsfailurepresentscharacteristicsofmulti-site,multi-phenomenaandnon-linear,sothatthefaultdiagnosisbecomemoredifficult.Atpresentit’shardtoknowexactlywhatthefailuremechanismis,andalsocan'tdescribetheenginefaultsystemthroughtheenoughprecisemodel.Thereforerequiresustoconstructasystemthathasintelligence,canstudyfaultexamples,extractsfaultcharacteristicsfromalargenumberofsamples,andgaintheabilitytodetermineandpredict.Thedevelopmentofartificialneuralnetworktheoryprovidesawaytosolvethisproblem,Itsadaptivelearningabilitywillgreatlyimprovetherecognitionabilityandintelligentdegreeoffaultdiagnosissystem.Inrecentyears,manyexpertsandscholarsputforwardanumberofnewmethodsinneuralnetworkfaultdiagnosis,reference[1]presentedusingtheevidentialreasoningapproachexpertsystemtoapplytothefaultdiagnosisofengine,buttheaccuracyofthisfaultdiagnosismethoddependsontherichnessofexpertknowledgeintheknowledgebase.Reference[2]proposesasimulatedannealingalgorithmwithglobalsearchpropertytooptimizetheBPneuralnetworkforavoidingthelocalminimumandimprovingitsstability,butit’sapoorglobalsearchalgorithm,andsusceptibletotheeffectsofparameters.Inreference[3],antcolonysystemisanovelsimulatedevolutionaryalgorithm.Ithaspositivefeedback,distributedcomputation,anduseofaconstructivegreedyheroism.Thecombinationofantcolonyalgorithmwithneuralnetworkisadoptedinfaultdiagnosisofengineanditcanimprovetheefficiencyofoperation,however,thisalgorithmissuitableforthesearchpathproblem,thecalculationiscostly.Inreference[4],BayesianNetworkmodelforfuelsystemofcertainelectronicenginehasbeensetup,andthestrongestdependencyroutealgorithmhasbeenputforwardforThisisageneralprogramofHeilongjiangProvincialDepartmentofEducation,programnumber:11541062.backwardinference.However,thisapproachmustfirstdeterminetheaprioriprobabilityofthecauseofthefailureandtheconditionalprobabilitybetweenthereasonsandtheresultsofthemalfunction.Reference[5]presentedaheuristicalgorithm,itwasusedforattributereductionbasedontheimportanceofattributevaluestoreduceattribute,afaultdiagnosisapproachwasformedcombiningthefuzzyinformationsystemknowledgemethodwithBP-neuralnetworkoftheparticleswarmoptimization(PSO)algorithmtodiagnosethefaultoftheengine.However,duetothePSOalgorithmisakindofglobaloptimizationalgorithm,thewholeprocessoftrainingalgorithmrequiresalongtime.ThispaperpresentsanimprovedBPneuralnetworkalgorithm,itincreasesanewlayerofstatefeedbackfromtheoutputlayeronthebasisoftraditionalBPneuralnetworkmodel.Thisalgorithmcanovercometheshortcomingsoftraditionalnetworkforgettingoldsamplesduringthetrainingprocesswhenlearningnewsamples,andimprovetheaccuracyoftheneuralnetworktraining.2ImprovedBPNeuralNetwork2.1TraditionalBPneuralnetworkInneuralnetworksystem,themessageisexpressedbytheinterconnectedofmassneuronsandalltheconnectiveweights.Themethodforneuralnetworktoidentifyismainlythroughalargenumberofsamplesfortraining,withinternalnetworkadaptivealgorithmconstantlyadjustingtheweightstomeettherequirements.Stateidentifierisimpliedinthenetwork,specificallymanifestedininterconnectionformandweights.Intheuseofnetworkprocess,foraparticularinputmode,neuralnetworkproducestheoutputmodethroughtheforwardcalculation,andthentheparticularsolutioncanbeobtainedthroughthecomparisonandanalysisoftheoutputsignaltojudgethecategoryoftheinputsignal.BPnetworkisamultilayerfeedforwardneuralnetworkanditsalgorithminessenceischaracterizedbythequadraticsumofnetworkerrorasthetargetfunction,pressinggradientdescentmethodforsolvingthetargetfunctiontoreachitsminimumvalue.Atpresent,inthepracticalapplicationofartificialneural3URFHHGLQJVRIWKHVW&KLQHVH&RQWURO&RQIHUHQFH-XO\+HIHL&KLQD3352network,thevastmajorityofneuralnetworkmodelaretheBPnetworkoritsvariations.Itisalsothekerneltoforwardnetworkandembodiesthecoreoftheessenceofartificialneuralnetwork.2.2TheBPnetworkthatincreasedthestate