基于SimAM注意力机制的轴承故障迁移诊断模型.pdfVIP

基于SimAM注意力机制的轴承故障迁移诊断模型∗

摘要:针对轴承故障在跨工况迁移诊断时,其域不变特征难以提取,易出现模型过拟合这一问题,提出了一种基于无参数注意力模

块(SimAM)的轴承故障迁移诊断方法。首先,以一维卷积神经网络作为基本框架,利用自适应批量归一化(AdaBN)对各输出层进

行了归一化处理,经两层卷积层和两层池化层后,对输出特征进行了随机节点失活操作;然后,利用改进后的参数化修正线性单元

(PReLU)激活函数自适应提取负值输入权值系数,分别以交叉熵损失函数监督训练有标签的源域数据,以均方对数误差(MSLE)作

为损失函数训练无标签的目标数据;最后,利用自制实验台数据和凯斯西储轴承公开数据对模型进行了验证,分别以不同的单一工

况作为源域,其余工况作为目标域进行了迁移诊断任务研究。研究结果表明:基于SimAM的轴承故障迁移诊断方具有较好的域不

变特征提取的性能,且所提特征具有较好的聚类效果;自制实验台中的平均迁移精度在89.1%以上,最高均值可达97.85%,CWRU

数据集中的平均迁移精度达98.68%。该成果可为后续轴承故障由实验向工业现场的迁移诊断奠定基础。

关键词:轴承故障诊断;迁移学习;无参数注意力机制;自适应批量归一化;参数化修正线性单元;均方对数误差;卷积神经网络

RearingfaulttransferdiagnosismodelbasedonSimAMattentionmechanism

Abstract:Aimingattheproblemthatdomaininvariantfeaturesaredifficulttoextractandmodeloverfittingiseasytooccurinbearingfault

migrationdiagnosisduringcross-workingconditionmigrationdiagnosis,abearingfaultmigrationdiagnosismethodbasedonasimple

parameter-freeattentionmodule(SimAM)wasproposed.Firstly,theone-dimensionalconvolutionalneuralnetworkwasusedasthebasic

framework,andadaptivebatchnormalization(AdaBN)wasusedtonormalizeeachoutputlayer.Aftertwoconvolutionallayersandtwo

poolinglayers,theoutputfeaturesweredeactivatedbyrandomnodes.Then,theimprovedparametricrectifiedlinearunit(PReLU)

activationfunctionwasusedtoadaptivelyextractthenegativeinputweightcoefficient,andthecross-entropylossfunctionwasusedtomonitor

thetrainedlabeledsourcedomaindataandthemeansquaredlogarithmicerror(MSLE)wasusedasthelossfunctiontotraintheunlabeled

targetdata.Finally,themodelwasverifiedbytheself-madeexperimentalbenchdataandtheopendataofCaseWesternReservebearing.

Differentsingleworkingconditionsweretakenasthesourcedomain,andtheotherworkingconditionsweretakenasthetargetdomaintocarry

outthemigrationdiagnosistask.Theexperimentalresultsshowthattheproposedmethodhasgooddomaininvariantfeatureextraction

performa