Abstract
WiththegrowingdevelopmentofChinasmanufacturingindustry,theindustrial
fieldhasgraduallybecomeautomated,largeandsystematic.Intheprocessof
unmannedfactory,thecompositionofmachineryandequipmenthasbecomemoreand
morecomplicated,functionshavebecomemoreandmoreperfect,andthesafetyof
equipmenthasbeenreceivingincreasingattention.Faultdiagnosisisthemost
commonlyusedmethodforfeatureextractionfromvibrationsignalscollectedby
mechanicalequipment.However,todaysdevelopmentofmechanicalequipmenttends
tobehigh-precision,high-speed,andhigh-efficiency,accompaniedbycontinuous
developmentofdataacquisitionandstoragetechnologies.Thefaultsignalgradually
exhibitsthecharacteristicsofmechanicalbig-data.Traditionalfaultdiagnosis
methodsaredifficulttodealwithmassivefaultdata.Thedeeplearningalgorithmisa
branchofartificialintelligencebecauseofitsmulti-hiddenlayernetworkandadaptive
featureextractioncapability.Theabilitytominetheessentialcharacteristicsofthedata
anduseallthecharacteristicsoftheoriginalsignalwithoutdiscardingtheoriginaldata
informationaccuratelycharacterizesthecomplicatedmappingrelationshipbetweenthe
observeddataandthefaultcategorycomparedtothetraditionalmethod.Thispaper
dealswiththefailurebasedondeeplearning.Diagnostictechniquesarestudied.
Firstly,fromtheprincipleofDeepBeliefNets(DBN),theuseofstandard
handwrittendigitsetsforDBNrestrictionsThefeatureextractioncapabilitiesofthe
RestrictedBoltzmannMachine(RBM)partandthenetworkfine-tunedbyBPThe
classificationabilityisstudied,andtheinfluenceofthenumberofhiddenlayernodes