2023年第47卷第5期JournalofMechanicalTransmission基于深度迁移学习的齿轮故障诊断方法刘世豪1王细洋2龚廷恺2（1南昌航空大学飞行器工程学院，江西南昌330063）（2南昌航空大学通航学院，江西南昌330063）摘要针对齿轮故障样本欠缺问题，提出一种基于Hilbert-Huang谱和预训练VGG16模型的迁移学习故障诊断方法。对振动信号进行经验模态分解（EmpiricalModeDecomposition，EMD）得到本征模态函数（IntrinsicModeFunction，IMF），同时取相关系数最大的IMF做Hilbert变换，获取时频谱；利用预训练VGG16提取变负载下和各健康状态下齿轮的Hilbert-Huang谱图像特征；采用全局均值池化层取代VGG16模型部分全连接层，进行分类输出。实验结果表明，在少量的样本数据下，该方法的齿轮故障诊断准确率达到98.86%，优于TLCNN和TranVGG-19等迁移学习方法，证明了该方法在齿轮故障诊断中具有一定研究价值。关键词迁移学习VGG16模型Hilbert-Huang谱齿轮故障诊断全局均值池化GearFaultDiagnosisMethodBasedonDeepTransferLearningLiuShihao1WangXiyang2GongTingkai2(1SchoolofAircraftEngineering,NanchangHangkongUniversity,Nanchang330063,China)(2SchoolofNavigation,NanchangHangkongUniversity,Nanchang330063,China)AbstractAimingattheproblemofinsufficientgearfaultsamples,afaultdiagnosismethodoftransferlearningbasedonHilbert-Huangspectrumandpre-trainedVGG16modelisproposed.Firstly,theintrinsicmodefunction(IMF)isobtainedbyEmpiricalModeDecomposition(EMD)ofvibrationsignals,andthetimespectrumisobtainedbyHilberttransformofIMFwiththelargestcorrelationcoefficient.Secondly,pre-trainedVGG16isusedtoextractHilbert-Huangspectrumimagefeaturesofgearsundervaryingloadsandundervari⁃oushealthconditions.Finally,theglobalaveragepoolinglayerisusedtoreplacepartialfullconnectionlayerofVGG16modelforclassificationoutput.Experimentalresultsshowthatwithasmallamountofsampledata,theaccuracyofgearfaultdiagnosisreaches98.86%,whichisbetterthanthetransferlearningmethodssuchasTLCNNandTranVGG-19.Itisprovedthatthemethodpresentedinthispaperhassomeresearchvalueingearfaultdiagnosis.KeywordsTransferlearningVGG16Hilbert-HuangspectrumGearfaultdiagnosisGlobalaver⁃agepooling0引言在工业4.0背景下，工厂设备呈现复...
