MachineryfaultdiagnosisbasedonfuzzymeasureandfuzzyintegraldatafusiontechniquesXiaofengLiuÃ,LinMa,JosephMathewCRCforIntegratedEngineeringAssetManagement,SchoolofEngineeringSystems,QueenslandUniversityofTechnology,GPOBox2434,BrisbaneQLD4001,Australiaarticleinfohistory:Received23July2007Receivedinrevisedform19July2008Accepted27July2008Availableonline3August2008Keywords:FuzzymeasuresFuzzyintegralsFuzzyc-meansDatafusionFaultdiagnosisabstractFuzzymeasureandfuzzyintegraltheoryareanoutgrowthofclassicalmeasuretheory.Fuzzymeasureandfuzzyintegraltheorytakeintoaccounttheimportanceofcriteriaandinteractionsamongthem,andhaveexcellentpotentialforapplicationssuchasclassification.Thispaperpresentsanoveldatafusionapproachformachineryfaultdiagnosisusingfuzzymeasuresandfuzzyintegrals.Theapproachconsistsofafeature-leveldatafusionmodelandadecision-leveldatafusionmodel.Thefuzzyc-meansanalysismethodwasemployedtoidentifytherelationsbetweenafeaturesetandafaultprototypetoestablishmappingsbetweenfeaturesandgivenfaults.Rollingelementbearingandelectricalmotorexperimentswereconductedtovalidatethemodels.Differentfeatureswereobtainedfromrecordedsignalsandthenfusedatbothfeatureanddecisionlevelsusingfuzzymeasureandfuzzyintegraldatafusiontechniquestoproducediagnosticresults.Theresultsshowedthattheproposedapproachperformsverywellforbearingandmotorfaultdiagnosis.2008ElsevierLtd.Allrightsreserved.1.IntroductionDuetotheincreasingcomplexityofmodernmachineryandtheconsequentgrowingdemandsforreliability,availability,safetyandcostefficiency,condition-basedmaintenance(CBMhasbecomeamainstreammaintenancestrategyinindustry.EffectiveCBMcanonlybeimplementedifaccuratemachinerydiagnosis/prognosisprogramsareinplace.Inturnthetaskoffaultdiagnosisandprognosisishighlydependentontheabilitytointerpretmulti-parameterdatabasedonadvancedsignalprocessingalgorithms.Variousmethodshavebeenappliedtomachineryconditionmonitoringandfaultdiagnosis,e.g.statisticalmethods,clusteringanalysis,neuralnetworks,model-basedmethods,geneticalgorithms,hybridsystems,etc.[1–6].Clusteringanalysisisaniterativepartitioningmethodwhichcanproduceanoptimaldiagnosticresult.Forclassificationpurposes,fuzzyclusteringanalysishasbeenshowntobemoreeffectivethantraditionalclusteringmethodsinhandlingfaultfeatureswhicharegenerallyimprecisewiththeboundariesamongdifferentfailuremodesusuallybeingambiguousintheirmappingspace.Fuzzymethodsareabletoclassifyfaultpatternsinanon-dichotomouswaywhichissimilartothewayhumanbeingsprocessvagueinformation[7].Thefuzzyclusteringanalysismethodregardsfaultcategoryasafuzzyset.Eachfaultofthatcategoryisassignedamembershipvaluerangingbetween0and1,inordertodescribetowhatdegreethefaultbelongstothatcluster.ContentslistsavailableatScienceDirectjournalhomepage:wwelsevier/locate/jnlabr/ymsspMechanicalSystemsandSignalProcessing0888-3270/$-seefrontmatter2008ElsevierLtd.Allrightsreserved.doi:10.1016/j.ymssp.2008.07.012Correspondingauthor.CurrentAddress:DownerEDIRail,OffCormorantRoad,KooragangIsland,NSW2304,Australia.Tel.:+61249200425;fax:+61249201401.E-mailaddresses:xf.liu@qut.edu.au,xiaofeng.liu@downeredirail.au(X.Liu.MechanicalSystemsandSignalProcessing23(2009690–700Fuzzyc-means(FCMisanoptimalfuzzyclusteringmethodwhichisusedtoproducec-partitions.FCMcomputestheclustercentresiterativelybyminimizingageneralizedlossfunction.FCMhasfoundapplicationsinselect...