基于神经常微分方程的机械故障诊断方法

打开文本图片集

中图分类号：TH133.33 文献标志码：A DOI：10.16385/j.cnki.issn.1004-4523.202307050

Abstract：Basedontheproblemsofpoorinterpretability，as wellasparameterincreaseand memoryconsumptioncaused byblind stackinglayers intraditionalfaultdiagnosis methodbasedondplearing，Neuralordinarydiferentialequation（NODE）isintro ducedintomechanicalfaultdiagnosis，theetworkstructureofNODEformachineryfaultdiagnosisisconstructed.Intheonstruct edstructure，thederivativesoftheparameterizedhiddenstatesoftheneuralnetworkareusedtoreplacethediscretesequencesof thespecifiedhiddenlayersByconstructinganonlinearrelationshipetwnfultdataandfulttypes，nordinarydiferentialequationsolvr（DEsolver）isusedtocompletetheclasificationofdiferentfaulttypes，andanendtoendfaultdiagnosismodelis formed.Theproposed methodisapliedtomechanicalfaultdiagnosis tobuildaspecificNODEnetwork model，andtheclassifica tiontaskofdiferentfaultcategoriesisaccomplishedtroughtheinputofultdata.Theconstructedmodelisappliedtothefaultdi agnosis ofspindlebearing intheaireraft engine，andcompared withthefaultdiagnosismethodbasedonresidual network model. Theexperimental resultsshow thattheconstructed modelandresidualnetwork model havesatisfactoryaccuracy.However，the constructed modelnotonlyreduces the memoryconsumption，but alsoreduces thenumberof model parameters byalmost five times.

Keywords：fault diagnosis；neural ordinary diferential equation；dynamics system；residual network

在大数据背景下，机械故障智能诊断的深入研究和应用迎来了新的机遇，特别是基于深度学习的机械故障诊断方法取得了很大的进展。（剩余10347字）