DEEP LEARNING BASED INTELLIGENT FAULT DIAGNOSIS OF ELECTRIC DRIVE SYSTEM

Abstract

Variablefrequencydrivefedinductionmotorsareutilizedindifferentcommercialandindustrialappli- cations suchasrenewableenergy,smartgridandelectricvehicleduetoitsreliableoperationandfast dynamic response.However,thecontinuousoperationofinverterfedinductionmotordrivesisadaunt- ing criticaltaskowingtotheincreasingloaddemandsassociatedwithhighharmonics,whichcouldlead to irrecoverablefailuresofthemachinewindings,andhighmaintenancecost.Furthermore,suchfailures may behazardoustohumanlifeinmilitaryoraerospaceapplications.Therefore,thecontinuousmoni- toring ofthepowerelectronicsdrivefedinductionmachinecomponentsorrunningatestonthemachine periodically issubstantialtominimizethetheeconomicallosses.Also,itiscrucialtoensurethesafety of thecontrolledsystemandtopreventthetheuncontrolledoutagesinthetechnologicalprocessesdueto machine failures.Inaddition,theconditionmonitoringschemereducesthemaintenancecosts,extends the machinelifeandincreasestheavailability,efficiencyandproductivity.Furthermore,theincipient faultdetection,whichbasicallyaimstodiagnosethefaultatanearlystage,preventingthemaintenanceto be scheduledformachinesthatmightnotordinarilybedueforservice,andmayalsopreventanextended period ofdowntime,causedbyextensivemotorfailures. Advancedautomationsystemsorprocessesinindustryrequireinverterfedinductionmotordrivesystem, which demandscautiouscoordinationofvariablefrequencyinverterwiththeinductionmotor.Inthe present-day industries,consistentfunctioningofnumerouselectricdrivesisessentialforoverallplant operation, andfailuresofsuchdrivesorportionofdrivesmustbeprotectedaheadoffaultoccurrence. The importanceofincipientfaultdetectionisthecost-effectivenessthatisaccomplishedbydiagnosing potential failuresatitsveryinception.Thefinanciallossesduetomachinefailures((i.e.statorwinding, bearing vibrationandovertemperature)arecostlywhichwillputlimitationsonelectricmachinedrives, when theyareconsideredincertainapplicationswheresafetyandreliabilityarecritical. Inter turnshortcircuitfault(ITSCF)relatedtostatorwindingfailuresisoneofthemaincontributor of theoverallelectricmachinefaults.Theresearchworkcarriedoutinthisthesisismainlyfocused on incipientinterturnshortcircuitfaultdetectionanddiagnosisinvariablefrequencyinductionmotor driveinpresenceoftheharmonicscausedbythehighswitchingfrequencyoftheinverter,aswellasthe ambiguous conditionsrelatedtotheloadvariations.Highelectricalandmagneticforceswithincreased environmental,thermalandmechanicalstressesleadtohigherprobabilityoftheinterturnshortcircuit fault.Thisfaultisoneofthemostsevereandfastevolvingfaults,whichmakesitimportanttodetect shortly afteritsoccurrence.Interturnshortcircuitfaultofinverter-driveninductionmotorrequiresfurther investigationinordertoovercomevariouschallengessuchastheriseinnoiseandthethermalstresses which leadtoinsulationdeteriorationofstatorwindings.Therefore,thegoalistoinvestigatetheITSCF of inverter-fedinductionmotor,whichhasnotyetbeenexploreddeeper. Numerous faultmonitoringtechniquesforinductionmotordrivescanbebroadlycategorizedasmodel based techniques,signalprocessingtechniques,anddatadriventechniques.Incaseofmodelbased techniques, accuratemodelsofthefaultymachineareessentiallyrequiredforachievingreliablefault diagnosis. Sometimes,itbecomescumbersometoobtainaccuratemodelsofthefaultymachines,toef- fect earlyandunambiguousdiagnosisofthefault.Thesignalanalysismethodswhichusethefrequency spectrum ofthelinecurrenttodetecttheinductionmachinefaults,arewidelyfoundintheliterature. Formoreaccuratefaultdetection,theadvancedsignalprocessingtechniquessuchaswaveletfamilyand synchrosqueezing transformareemployed.Intelligentfaultdiagnosis(IFD)usingdata-drivenapproaches havebeenutilizedforonlineconditionmonitoringandfaultdiagnosisininductionmachines.Neverthe- less, IFDtechniquesstillhavelimitationsinanalyzingbigdataduetotheirincapabilityofself-learning. That is,IFDapproachesdependonthehand-craftedfeatureextractionfromtherawdata,whichistime- intensiveandarduoustask.Thus,deeplearning(DL)baseddata-driventechniquesarefeasiblesolution to overcometheaforementionedproblemsofIFDmethods,andithasattractedreasonableattentioninthe area ofinductionmachinefaultdiagnosis.DLapproacheshaveautomaticfeaturelearningabilityfrom the measureddata,whichinturnshrunkthecomputationaltimerelatedtotheadditionalstepoffeature extraction.Also,DLtechniqueshaveself-learningcapabilitytoprocessmassivedataandalsohaveabil- ity toperformend-to-endstructurewhichcanlearnhigh-levelrepresentationfromrawdataandpredict the targetsautomatically. This researchworkisconcentratedmainlyontheinterturnshortcircuitfault(ITSCF)detectionand identification atanearlystageforinductionmotordrivenbyinverterusingdeeplearningapproaches.All the proposedmethodologieshavebeenextensivelyevaluatedinmultipleexperimentsfordifferentload variations,variousfrequenciesofthedriveanddifferentfaultseveritylevelsusingalaboratoryhardware prototype. Thus,ithasbeenrealisticallydemonstrated,themeritsofalltheproposedfaultdetectionand isolation schemes.