251国家重点研发计划(2020YFB1314000)资助收稿日期:2022–03–13;修回日期:2022–04–06北京大学学报(自然科学版)第59卷第2期2023年3月ActaScientiarumNaturaliumUniversitatisPekinensis,Vol.59,No.2(Mar.2023)doi:10.13209/j.0479-8023.2022.098基于频域降采样和CNN的轴承故障诊断方法周翔宇毛善君†李梅北京大学遥感与地理信息系统研究所,北京100871;†通信作者,E-mail:sjmao@pku.edu.cn摘要在工业领域,设备运行过程中采集的原始故障信号具有强噪声以及多工况的特点,现有的基于数据的轴承故障诊断模型的抗噪能力与泛化能力相对较弱。针对以上问题,提出一种基于频域降采样(down-sampling)和卷积神经网络(CNN)的轴承故障诊断方法Ds-CNN。频域降采样包含最大偏移降采样和噪声横截断两个部分,可以实现样本增强,降低样本在频域的差异性,同时减弱噪声对频域信号的影响。基于频域信号建立的CNN模型能够自动提取降采样后频域信号的故障特征,并完成对轴承故障的识别分类。实验结果表明,在强噪声环境和多工况条件下,与目前常用模型相比,Ds-CNN具有更高的识别准确率。关键词轴承故障诊断;深度学习;卷积神经网络(CNN);强噪声;多工况BearingFaultDiagnosisMethodBasedonDown-SamplinginFrequencyDomainandCNNZHOUXiangyu,MAOShanjun†,LIMeiInstituteofRemoteSensingandGeographicalInformationSystem,PekingUniversity,Beijing100871;†Correspondingauthor,E-mail:sjmao@pku.edu.cnAbstractIntheindustrialfield,theoriginalfaultsignalscollectedduringtheoperationoftheequipmenthavethecharacteristicsofstrongnoiseandmultipleworkingconditions.Mostofpreviousdata-drivenfaultdiagnosismethodsforbearingshaverelativelyweakanti-noiseabilityandgeneralizationability.Tosolvetheseproblems,anovelbearingfaultdiagnosismethodbasedondown-samplinginfrequencydomainandconvolutionalneuralnetwork(CNN),calledDs-CNN,isproposed.Down-samplinginfrequencydomainconsistsofmaximumdown-samplingwithbiasandnoisetransversetruncation,whichcanrealizedataaugmentation,reducethedifferencebetweensamplesinfrequencydomain,andreducetheinfluenceofnoiseonsignalsinfrequencydomain.TheCNNmodelbasedonfrequencydomainsignalscanautomaticallyextractfaultfeaturesfromsignalsafterdown-samplingandcompletetheidentificationandclassificationofbea...
