A Novel Method for the Diagnosis of the Incipient

CircuitsSystSignalProcess(2010)29:577–600DOI10.1007/s00034-010-9160-1ANovelMethodfortheDiagnosisoftheIncipientFaultsinAnalogCircuitsBasedonLDAandHMMLijiaXu·JianguoHuang·HoujunWang·BingLongReceived:4April2008/Revised:28October2009/Publishedonline:10April2010©SpringerScience+BusinessMedia,LLC2010AbstractDiagnosisofincipientfaultsforelectronicsystems,especiallyforanalogcircuits,isveryimportant,yetverydifficult.Themethodsreportedintheliteratureareonlyeffectiveonhardfaults,i.e.,short-circuitoropen-circuitofthecomponents.Forasoftfault,thefaultcanonlybediagnosedundertheoccurrenceoflargevariationofcomponentparameters.Inthispaper,anovelmethodbasedonlineardiscrimi-nantanalysis(LDA)andhiddenMarkovmodel(HMM)isproposedforthediagnosisofincipientfaultsinanalogcircuits.Numericalsimulationsshowthattheproposedmethodcansignificantlyimprovetherecognitionperformance.First,toincludemorefaultinformation,threekindsoforiginalfeaturevectors,i.e.,voltage,autoregression-movingaverage(ARMA),andwavelet,areextractedfromtheanalogcircuits.Sub-sequently,LDAisusedtoreducethedimensionsoftheoriginalfeaturevectorsandremovetheirredundancy,andthus,theprocessedfeaturevectorsareobtained.TheLDAisfurtherusedtoprojectthreekindsoftheprocessedfeaturevectorstogether,toobtainthehybridfeaturevectors.Finally,thehybridfeaturevectorsareusedtoThisworkissupportedinpartbythedefensefoundationscientificresearchfundunderGrantA1420061264andnationalnaturalsciencefundunderGrant60673011.L.Xu()·J.Huang·H.Wang·B.LongSchoolofAutomationEngineeringofUniversityofElectronicScienceandTechnologyofChina,Chengdu610054,Chinae-mail:[e-mail]-mail:[e-mail]-mail:[e-mail]-mail:[e-mail]ingTechnologyInstituteofSichuanAgricultureUniversity,Ya’an625014,China578CircuitsSystSignalProcess(2010)29:577–600formtheobservationsequences,whicharesenttoHMMtoaccomplishthediagno-sisoftheincipientfaults.Theperformanceoftheproposedmethodistested,anditindicatesthatthemethodhasbetterrecognitioncapabilitythanthepopularlyusedbackpropagation(BP)network.KeywordsHMM·LDA·Featureextraction·Faultdiagnosis1IntroductionFaultdiagnosisrepresentsafundamentaltaskinthepreventivemaintenanceofanelectronicsystem.Accordingtothecollectedstatistics,mostpartofanelectronicsys-temisdigital(about80%),butabout80%ofthefaultsoccurintheanalogsegment[13].Whencomparedwiththedigitalcircuits,faultdiagnosisandtestingofana-logcircuitsareparticularlychallengingowingtothefollowingreasons[20]:(1)Theparametersofanalogcomponentsareusuallycontinuous,whichcouldchangefromzerotoinfinity.Itisimpossibletodefineanyunifiedfaultmodelforanalogcom-ponents.(2)Inpracticalanalogcircuittesting,theinformationusedinthediagnosisisnotsufficientbecauseoflackoftestnode.(3)Inaddition,toleranceeffectsoftheanalogcomponentsaredifficulttoeliminate.Thesefaultsareclassifiedas“softfaults”orparametricfaultsthatdonotchangethecircuit’stopology,andareduetothevariationsofthevalueofsomecircuitcomponents.Sincethe1970s,faultdiagnosisofanalogcircuitshasbecomeanactiveresearcharea,andsomeresearchmethodshavebeenreported[1,4,14,19],butmostofthereportedmethodsarebasedonbackpropagation(BP)network.Thedifferencemainlyliesintheirwaysofextractingthefeature.Forexample,inanearlierstudy[1],thesampledatafromtheoutputofananalogcircuitwerepreprocessedbywaveletde-composition,normalization,andprincipalcomponentanalysis(PCA)stepbysteptoreducethedimensionoftheinputfeatures.Inanotherstudy[19],thefaultfeatureswereextractedfromtheimpulseresponseoftheanalogcircuitbywaveletpacketdecomposition.Furthermore,thefaultfeaturesusedinanotherearlierstudy[14]wereobtainedfromtheautoregressivemovingaveragemodelofthetransferfunc-tion.Thesestudieshavedevelopedwell-establishedtheoreticalfoundationsandtoolstodetect“hardfaults”or“softfaults”withalargerangeofparametersvariation.However,moreandmoreattentionisbeingfocusedontheperformancedegradationmonitoringsothatfailurecanbepredictedandprevented,anditisveryimportanttoidentifythefaultatitsincipientstageandalerttheoperatorbeforeitdevelopsintoacatastrophicfailure.Thisisreferredasthecondition-basedmaintenance(CBM).Someresearches[3,8,17,18,21]haveexaminedtheCBM,includingfield-effecttransistors,powerconverters,printedcircuitboards,globalpositioningsystems,en-terpriseservers,etc.However,thediagnosisoftheincipientfaults,suchasshiftsinperformanceparameters,iscrucial,yetverydifficult.Byconsideringthatlineardiscriminantanalysis(LDA)isapopulartechniqueinextractingeffectivediscrimi-nativefeaturesandhiddenMarkovmodel(HMM)hasgoodrecognitioncapability,thisstudypresentsanovelmethodbasedonLDAandHMMtodiagnosetheincipi-entfaultsinanalogcircuits.Thismethodincludesthefollowingfoursteps:First,toCircuitsSystSignalProcess(2010)29:577–600579includemorefaultinformation,threekindsoforiginalfeaturevectors,i.e.,voltage,autoregression-movingaverage(ARMA),andwavelet,areextractedfromtheanalogcircuits.Second,theLDAisusedtoreducethedimensionsoftheoriginalfeaturevectorsandremovetheirredundancy,andthus,theprocessedfeaturevectorsareo