基于矩阵乘积态的机械故障诊断方法研究黄文静,李志农*(无损检测技术教育部重点实验室(南昌航空大学),南昌330063)[摘要]在机械故障诊断中,针对传统神经网络处理高阶数据难度大、网络参数多、耗费大量计算资源的不足,提出了一种基于矩阵乘积态的张量网络故障诊断方法。通过输入高阶张量故障数据到矩阵乘积态故障诊断模型中,将高阶张量表示为多个低阶张量,从而简化数据结构和参数量。为了验证该方法的有效性,将其应用在齿轮的故障诊断中,并与传统的卷积神经网络故障诊断模型进行对比。同时,验证了键维度对模型准确率的影响。结果表明:所提模型的键维度会影响模型准确率,键维度为16的模型准确率高于键维度为8的模型准确率;该模型在减小数据复杂度的同时,还可以识别不同故障类型,准确率达到90%,比传统的卷积神经网络故障诊断模型性能更好。[关键词]高阶张量;张量网络;矩阵乘积态;故障诊断[中图分类号]TP181;TH17[文献标志码]Adoi:10.3969/j.issn.1673-6214.2023.03.002[文章编号]1673-6214(2023)03-0149-06ResearchonMechanicalFaultDiagnosisMethodBasedonMatrixProductStateHUANGWen-jing,LIZhi-nong*(KeyLaboratoryofNondestructiveTesting(MinistryofEducation),NanchangHangkongUniversity,Nanchang330063,China)Abstract:Inmechanicalfaultdiagnosis,itisdifficultfortraditionalneuralnetworkstoprocesshigh-leveldata,andmanynetworkparametersconsumealotofcomputingresources.Therefore,thispaperproposesatensornetworkfaultdiagnosismethodbasedonmatrixproductstate.Byinputtinghigh-ordertensorfaultdataintothematrixproductstatefaultdiagnosismodel,thehigh-ordertensorisrepresentedasmultiplelow-ordertensors,thussimplifyingthedatastructureandreducingtheparameternumber.Inordertoverifytheeffectivenessofthemethod,itisappliedtothefaultdiagnosisofgearsandcomparedwiththetraditionalconvolutionalneuralnetworkfaultdiagnosismodel.Moreover,theeffectofbonddimensionontheaccuracyofthemodelwasassessed.Theexperimentalresultsshowthatthebonddimensionoftheproposedmodelaffectsthemodelaccuracy,demonstratingahigheraccuracywhenthebonddimensionis16incomparisontothatofthemodelwithabonddimensionof8.Whilereducingthedatacomplexity,themodelcanalsoidentifydifferentfaulttypeswithanac...