Residual life predictions for ball bearings based

MechanicalSystemsandSignalProcessingMechanicalSystemsandSignalProcessing21(2007)193–207Residuallifepredictionsforballbearingsbasedonself-organizingmapandbackpropagationneuralnetworkmethodsRunqingHuanga,
,LifengXia,XinglinLib,C.RichardLiuc,HaiQiud,JayLeedaDepartmentofIndustrialEngineering&Management,ShanghaiJiaoTongUniversity,No.1954,HuashanRd,Shanghai200030,PRChinabHangzhouBearingTest&ResearchCenter(HBRC),No.123,HuafengRd,Hangzhou310022,PRChinacSchoolofIndustrialEngineering,PurdueUniversity,315N.GrantStreet,WestLafayette,IN47907,USAdNSFI/UCRConIntelligentMaintenanceSystems(IMS),UniversityofCincinnati,OH45221,USAReceived23January2005;receivedinrevisedform12October2005;accepted21November2005Availableonline19January2006AbstractThispaperdealswithanewschemeforthepredictionofaballbearing’sremainingusefullifebasedonself-organizingmap(SOM)andbackpropagationneuralnetworkmethods.Oneofthekeycomponentsneededforeffectivebearinglifepredictionistheset-upofanappropriatedegradationindicatorfromabearing’sincipientdefectstagetoitsfinalfailure.Thisnewmethodisdifferentfromtheothersthathavebeenusedinthepast,inthatitusestheminimumquantisationerror(MQE)indicatorderivedfromSOM,whichistrainedbysixvibrationfeatures,includinganewdesigneddegradationindexforperformancedegradationassessment.Then,usingthisindicator,backpropagationneuralnetworksfocusingonthedegradationperiodscanbetrained.Thankstoweightapplicationtofailuretimes(WAFT)technology,ausefullifepredictionmodelforballbearingshasbeendevelopedsuccessfully.Finally,asetofacceleratedbearingrun-to-failureexperimentsiscarriedout,withtheexperimentalresultsshowingthatthenewproposedmethodsaregreatlysuperiortothose,basedonL10bearinglifeprediction,currentlybeingused.r2005ElsevierLtd.Allrightsreserved.Keywords:Residuallifeprediction;Self-organizingmap;Backpropagation;Neuralnetwork;Ballbearing;Prognostics1.IntroductionBearingsareamongthemostwidelyusedmachineelementsandarecriticaltoalmostallformsofrotarymachinery[1–4].Inordertopreventunexpectedbearingfailure,defectsthatcanoccurinbearingsshouldbedetectedasearlyaspossibletoavoidfatalbreakdownsofthemachinestowhichtheyaresocritical.Suchbreakdownscanleadtoacostlylapseinproduction,and/orevenhumancasualties.Anotheradditionalmeasurethatcanpreventbreakdownsisthecreationofareasonableandaccurateinspectionorreplacementschedule;ameasurethatwouldrelyontheeffectivepredictionofbearinglife.ARTICLEINPRESS:10.1016/j.ymssp.2005.11.008
Correspondingauthors.Tel.:[phone];fax:[phone].E-mailaddresses:[e-mail](R.Huang),[e-mail](L.Xi).Ofallbearingmonitoringtechniques,vibrationanalysisisthemostsuitableandeffective[5].Manypreviousstudieshavehelpedtodevelopwhathasnowbecomeawell-establishedtheoreticalfoundationforbearingdiagnostics.Thesestudieshavealsofurtheredthedevelopmentoftoolstocomprehensivelydescribebearingfailuremodes.Somehaveclassifiedbearingconditionsandfaultdiagnosisusingfuzzylogicconcepts[6],andneuralnetworksapproaches[7–9],whileothershavefocusedonwaveletanalysisapproachestodetectbearingfaults[2,3,10,11].Hybridapproaches,thatcombineneuralnetworksandfuzzylogic,arealsousedformonitoringanddiagnosticpurposes[3,12].Statisticalapproacheshavealsobeenusedtomodeltheeffectthatbearingdefectshaveontrendsofcertainstatisticalparameters[13,14],suchas,usingtherootmeansquare(RMS),CrestFactor,andKurtosis.Inrecentyears,inordertoreducecostsandshortenrepairtime,condition-basedpredictivemaintenancehasbecomeanefficientstrategyformodernindustries[1,2],whichnecessitatesadvancedtoolsinprognostics.Prognosticsistheuseofpredictivemaintenancepracticesandtoolstoanalysethetrendsofmachineperformanceagainstknownengineeringlimitsforthepurposeofdetecting,analysingandcorrectingproblemsbeforefailureoccurs.Moreadvancedprognosticsarefocusedonperformancedegradationmonitoringandassessment,sothatfailurescanbepredictedandprevented[15].Tofulfillthegoalofprognostics,threecrucialstepsareneeded.First,thedefectorabnormalityshouldbeabletobedetectedatanearlystage.Second,themachineorsystemperformanceshouldbeassessedrobustlyandtrackedcontinuously.Finally,theremainingusefullifeandpossiblefailuremodeofthemachineorsystemshouldbeeffectivelypredicted[1].Estimatingtheremainingusefullifeismostimportantinthesethreesteps,becausetheremainingusefullifedirectlyservesdecisionvariablesofprognostics.However,challengesineffectivelypredictingtheremainingusefullifeofballbearingsexist.Oneofthechallengesofbearinglifepredictionisfiguringouthowtoset-upanappropriatedegradationindicatorbasedonavibrationsignal.Previously,thetimefeatures,suchasRMS,Kurtosis,orCrestFactor,wereoftenchosen.Forexample,Ref.[16]usedtheRMSvalueandtheKurtosisFactorofthevibration,whileRef.[5]chosetheaverageoftheamplitudesofthedefectivefrequencyanditsfirstsixharmonics,overtime,asthedegradationindexofthrushballbearings.However,theseindexeseitherhavealowsensitivitytobearingincipientdefectordonotfit,verywell,thedeepgrooveballbearingunderhighlyacceleratedtest