基于支持向量机的故障智能诊断方法研究

华北电力大学（保定）博士学位论文基于支持向量机的故障智能诊断方法研究姓名：翟永杰申请学位级别：博士专业：热能工程指导教师：韩璞20040301i1——92FJ-SVMFJ-SVMFS-SVM3iiDDAGDAGSVMSVM4—SVRSVRSVRiiiAbstractStatisticalLearningTheoryisbasedonthetheoryfoundationandprovidesauniformframeworkforlearningsubjectoflimitedsamples.SupportVectorMachines(SVM)isamachine-learningalgorithmbasedonstatisticallearningtheory.Thisalgorithmaccomplishesthestructuralriskminimizationprinciple.ThefineperformanceofSupportVectorMachinestolimitedsamplesattractsattentionofinvestigatorsinfaultdiagnosisfield.Faultdiagnosisisalimitedsamplessubject.ThemostpredominanceofSVMisproperforlimitedsamplesdecision.Thenatureofthealgorithmisacquiringconnotativeclassinformationtogreatextentfromlimitedsamples.Fromthepointofgeneralization,SVMisthemorefavorableforthepracticalengineeringproblemasfaultdiagnosis.SVMhaseximiouspredominanceintheory,buttheresearchonapplicationiscomparativelydelayed.Basedonthis,theproblemofSVMintheapplicationoffaultdiagnosisontheaspectsofdatapretreatment,fuzzymethodwithuncertaininformation,therealizationofmulti-classSVManddiagnosisbasedonmodelwasdevelopedinthispaper.1Researchandapplicationonfeatureextractionforhigh-dimensiondataFirstly,therelationbetweenfeaturedimensionandclassifyresultwasanalyzedinthispart,whichshowsthattheincreaseoffeaturedimensionwillresultinworseclassifyresult.Thereasonforthismainlyliesinthedisturbanceoffalsefeature.Inthispaper,thelongitudinalcompressionthatappliedtocompressingfeaturedimensionswasmainlyresearched,andthismethodcaneliminatethefalseandretainthetrue,thatisfeatureextraction.Weintroducedthetypicalalgorithm--PrincipalComponentAnalysis(PCA)indetail.Asfarastheturbinefaultdiagnosisisconcerned,thedatasampleofsomeoriginalgroupnine-dimensionfrequencywasfeatureextracted,fromthetwo-dimensionprinciplecomponentmappingwecanconcludethatthefeaturedataafterextractedismoreseparable,whichsuggeststhatdatapreprocessinginpracticalfaultdiagnosisisfavorabletotherealizationofclassifyalgorithm,andprovideaneffectivelyfeasibledatapreprocessingmethodfortheturbinefaultdiagnosis.2Thefuzzyjudgmentsupportvectormachinealgorithmbasedonlossfunctionisproposed,andcomparedwithfuzzysamplesupportvectormachinealgorithm.Asfortheuncertaininformationinfaultdiagnosis,FJ-SVMbasedonlossfunctionwasproposedinthispaper,andfuzzymembershipgradesbasedonlosswasalsodefined.Furthermore,weeducedtheoptimalseparatinghyperplaneaftermodified,andjudgeivaccordingtothemaximummembershiprules.Fuzzysupportvectormachineconsideredthatdifferentwrongjudgmentresultinlossdifferenceinpractice.Atthesametime,ithaspreferablesensitivityfortheslightfaultofdeviceandforepartfaultdiagnosisinproductiveprocess.Practically,theselectionoffaultsampleisoftentypicalsamplewithobviousfeature.Therefore,thealgorithmthatusingtypicalsampletoformsupportvectormachineismorepractical.Justright,FJ-SVMisjustbasedonthisprinciple.Inthispaper,wealsointroducedFS-SVM.Fromthefuzzygradeswedirectlyeducedthisalgorithm.Weusetwokindsofalgorithmstosimulateexperiment,comparedtheresults,andanalyzedthedifferentemphasisintheapplicationofthesetwoalgorithms.3Discussingandanalyzingthecomposingofmulti-classsupportvectors,andresearchingthesetwotypicalalgorithms.Themethodofconversionfromtwo-classtomulti-classwasanalyzed.Combiningthedecisiontree,westudiedthetwokindsofmulti-classalgorithms:oneisDAGSVMbasedonDDAGstructure,andwediscussedtheselectionsofkernelfunctionparameters.Secondismulti-classalgorithmbasedonhierarchicalclusteringanddecisiontree.Weintroducedthetheoryandconcreterealization,andappliedthesetwomethodswiththelimitedsamples.4SensorfaultdiagnosisbasedonsupportvectorregressionisputforwardTheapplicationformofSVMinregressionsubjectisanalyzedandthesupportvectorregression(SVR)algorithmisleadout.IdentificationfordynamicsystemisdevelopedbasedonSVRandprovidedthemodelingfoundationforthefaultdiagnosisbasedonmodels.Inthispaper,forthesensorfaultdiagnosissystem,weproposedthesensorfaultdiagnosisbasedonthesupportvectorregression,anddesignedtheerrorgeneratorbasedonSVR.WiththeSVRsimulating,wegetfineresult.keywords:supportvectormachines,faultdiagnosis,featureextraction,fuzzytheory,decisiontree,systemidentification11.11967NASAMFPG[1]20021000MW70300MW[2][3]40—[4][1]2[5][6][7][8][9]ERM[10]1995VapnikTheNatureofStatisticalLearningTheory[11]StatisticalLearningTheorySLTSupportVectorMachinesSVMSLTSVM[12,13]SLTSVM[14-17][18][18,19]3[21]SVM1.21.2.12080[11]—Rosenblatt2060F.Rosenblatt[22]Rosenblatt4—2060—7019601986VCVC20801986BP[23]—2090,[24]Vapnik51.2.2[1]123456801989VenkatVenkataSubramanianKingChanKajiorHimmelblauHoskinshimmelblau19913TimoSorsa1993RBF(ART)Kohonen[2]1.31.3.1,7Vapnik20609090—VCVC1.3.219921995—BP8OptimalHyperplane1.3.391.3.4SVMSVMVapnik601971VapnikChervonenkisSVMVC1982VapnikSVM1992BoserGuyonandVapnik1993CortesVapnik1995VapnikSVM1997VapnikGokowichSmolaSVMSVMSVM[25]SVM[26]SVM[11][27]SVM[28]MITSVM[29][30][31][32][33][34][35][36]SVM[37][38][39][40][41][42][43]19992000[44][45]2000VapNikTheNatureofStatisticalLearningTheory[11]SVMSVM10PoyhonenS[46][47]JunFengGao[48]SVMSVM[49][50][51][52]SVMSVM[53][54]