多传感器信息融合的轴承故障迁移诊断方法_包从望.pdfVIP

多传感器信息融合的轴承故障迁移诊断方法∗

摘要:在重型装备低速、重载、强噪声环境下,采用单一传感器难以全面获取轴承的故障诊断信息,导致故障识别率低、识别不稳定,

致使变工况下轴承故障迁移诊断失效。针对以上问题,提出了一种多传感器信息融合的轴承故障迁移诊断方法。首先,结合传感

器的通道数,构建了堆叠卷积神经网络(MCNNs)提取各个通道的故障特征;然后,在MCNNs中引入最小绝对收缩与选择算子

(Lasso),并通过网络反向传播完成了特征权值的更新,从而获得了多通道特征的融合;最后,利用源域数据对模型进行了训练,提

取了故障特征,并完成了特征融合,采用损失函数完成了模型参数的优化,将源域训练得到的模型结果作为目标域的初始模型,利

用目标域样本对初始模型的参数进行了微调,从而完成了模型迁移;并进行了信息融合效果、方法对比以及传感器信息采集属性的

性能实验。研究结果表明:传感器的安装位置对信息融合影响较大,MCNNs+Lasso方法具有较好的特征融合效果,平均迁移诊断

精度为99.03%,部分精度可达99.97%,在多个变工况的迁移任务中表现出较高迁移精度和良好的泛化性能。

关键词:滚动轴承;故障诊断;多传感器信息融合;堆叠卷积神经网络;最小绝对收缩与选择算子;迁移学习

Transferlearningmethodforshearerfaultdiagnosisundervariable

workingconditions

Abstract:Intheenvironmentoflowspeed,heavyloadandstrongnoiseofheavyequipment,itisdifficultforasinglesensortofullyobtain

bearingfaultdiagnosisinformation,resultinginlowfaultidentificationrateandunstableidentification,resultinginbearingfaultmigration

diagnosisfailureunderchangingworkingconditions.Aimingattheaboveproblems,abearingfaultmigrationdiagnosismethodbasedon

multi-sensorinformationfusionwasproposed.Firstly,multipleconvolutionneutralnetworks(MCNNs)wasconstructedcombinedwiththe

numberofsensorchannelstoextractthefaultfeaturesofeachchannel.Then,theleastabsoluteshrinkageandselectionoperator(Lasso)was

introducedintoMCNNs,andthefeatureweightwasupdatedthroughnetworkbackpropagation.Thusmulti-channelfeaturefusionwas

realized.Finally,thesourcedomaindatawasusedtotrainthemodelandextractfusionfeatures,andthemodelparameterswereoptimized

throughlossfunction.Themodelresultsobtainedfromthesourcedomaintrainingweretakenastheinitialmodelofthetargetdomain,and

theparametersoftheinitialmodelwerefine-tunedthroughthetargetdomainsamples,soastorealizemodelmigration.Finally,the

performanceexperimentsofinformationfusioneffect,method