基于迁移学习与加权多通道融合的齿轮箱故障诊断  

Gearbox fault diagnosis based on transfer learning and weighted multi-channel fusion

作　　者：侯召国 王华伟[1] 熊明兰 王峻洲 HOU Zhaoguo;WANG Huawei;XIONG Minglan;WANG Junzhou(School of Civil Aviation,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China)

机构地区：[1]南京航空航天大学民航学院,南京211106

出　　处：《振动与冲击》2023年第9期236-246,共11页Journal of Vibration and Shock

基　　金：国家自然科学基金和民航联合研究基金(U1833110)。

摘　　要：针对齿轮箱单一传感器故障识别精度波动大、数据利用率低、可靠性低及故障诊断模型在多工况下泛化能力不足等问题,提出了一种加权融合多通道数据与深度迁移模型的齿轮箱故障诊断方法。首先,为了充分挖掘齿轮箱多通道数据的信息,提出了基于信息熵加权的多通道融合方法,采用信息熵法计算各通道数据的融合权重,并对各通道的采样数据进行加权融合。其次,利用源域的融合数据对深度迁移模型进行预训练,将预训练得到的模型参数作为目标域模型的初始化参数,同时冻结目标域模型特征提取器的参数,并利用目标域的融合数据对目标域模型分类器的参数进行微调,实现深度迁移模型从源域到目标域的迁移以适应新的目标样本识别任务。最后,齿轮箱多工况迁移诊断试验结果表明,所提方法可有效用于齿轮箱的故障诊断,相比传统迁移学习方法平衡分布自适应算法(balanced distribution adaptation,BDA)、迁移成分分析(transfer component analysis,TCA)、联合分布自适应算法(joint distribution adaptation,JDA)、统计分布和几何空间联合调整算法(joint geometric and statistical alignment,JGSA)、测地线流式核算法(geodesic flow kernel,GFK)及深度迁移学习方法自适应批归一化(adaptive batch normalization,AdaBN)、多核最大均值差异(multi-kernel maximum mean discrepancy,MK-MMD)、深度卷积迁移学习网络(deep convolutional transfer learning network,DCTLN)这8种当前常用方法,具有更高的平均迁移诊断精度和变工况下良好的泛化性能。Here,aiming at problems of large fluctuation of fault recognition accuracy of single sensor of gearbox,low data utilization,low reliability and insufficient generalization ability of fault diagnosis model under multi-working condition,a gearbox fault diagnosis method based on weighted fusion of multi-channel data and deep transfer model was proposed.Firstly,in order to fully excavate information of multi-channel data of gearbox,a multi-channel fusion method based on information entropy weighting was proposed.The information entropy method was used to calculate fusion weights of various channels’data,and sampling data of various channels were weighted and fused.Secondly,fusion data of source domain were used to pre-train deep transfer model,the model’s parameters obtained with pre-training were taken as initialization parameters of target domain model,parameters of feature extractor of the target domain model were frozen,and fusion data of the target domain were used to fine-tune parameters of the target domain model’s classifier,realize transfer of the deep transfer model from source domain to target domain,and adapt to new target sample recognition task.Finally,gearbox multi-working condition transfer diagnosis test results showed that the proposed method can effectively be applied in gearbox fault diagnosis;compared with the traditional transfer learning methods balanced distribution adaptation(BDA),transfer component analysis(TCA),joint distribution adaptation(JDA),joint geometric and statistical alignment(JGSA)and geodesic flow kernel(GFK)and the deep transfer learning methods adaptive batch normalization(AdaBN),multi-kernel maximum mean discrepancy(MK-MMD)and deep convolutional transfer learning network(DCTLN)which are 8 currently commonly used methods,the proposed method has higher average transfer diagnosis accuracy and good generalization performance under variable working conditions.

关 键 词：故障诊断 齿轮箱 深度迁移模型 加权多通道融合 多工况 

分 类 号：TH165.3[机械工程—机械制造及自动化]