基于Swin-Transformer的多尺度多源域自适应轴承故障诊断  

Multi-scale and Multi-source Domain Adaptation Bearing Fault Diagnosis Based on Swin-Transformer

作　　者：周玉国 张志凯 张金超 于春风 周立俭 ZHOU Yuguo;ZHANG Zhikai;ZHANG Jinchao;YU Chunfeng;ZHOU Lijian(School of Information and Control Engineering,Qingdao University of Technology,Qingdao Shandong 266520,China;Department of Aeronautical Instrument and Electronic Control Engineering and Command,Naval Aviation University,Qingdao Shandong 266000,China)

机构地区：[1]青岛理工大学信息与控制工程学院,山东青岛266520 [2]海军航空大学航空仪电控制工程与指挥系,山东青岛266000

出　　处：《机床与液压》2025年第1期32-42,共11页Machine Tool & Hydraulics

基　　金：山东省自然科学基金项目(ZR2021QF113);青岛市科技局产业集群培育专项项目(23-1-2-qljh-6-gx)。

摘　　要：针对当前多源域自适应方法无法充分挖掘多源域中不同尺度故障信息的问题,提出一种基于Swin-Transformer(Swin-T)的多尺度多源域自适应轴承故障诊断方法。通过连续小波变换,获得振动信号在不同频带的特征。为更充分地利用多源域中不同尺度的故障信息,提出基于Swin-T的多尺度特征提取网络。为了减小各域之间的数据分布差异,构建基于最大均值差异的特征对齐网络,并根据不同尺度对分类的贡献赋予权值。此外,构建多尺度特征融合模块,对不同尺度的特征信息进行融合,得到故障特征集。最后,利用Softmax对特征集进行故障分类,并通过最小化多分类器预测差异损失得到最终分类结果。在凯斯西储大学和青岛理工大学轴承数据集上,该方法的故障分类准确度分别达到99.63%和99.40%。To solve the problem that the current multi-source domain adaptation(MSDA)method cannot be used to fully mine the fault information of different scales in the multi-source domain,a method based on Swin-Transformer(Swin-T)multi-scale multi-source domain adaptation bearing fault diagnosis method was proposed.The features of vibration signals in different frequency bands were obtained by continuous wavelet transform.In order to make full use of fault information with different scales in multi-source domains,a multi-scale feature extraction network based on Swin-T was proposed.In order to reduce the differences in the data distribution between different domains,a feature alignment network based on the maximum mean difference was designed,and the weights were assigned according to the contribution of the features with different scales for the classification.In addition,a multi-scale feature fusion module was constructed to fuse feature information at different scales to obtain a fault feature set.Finally,Softmax was used to classify faults and achieve the final consistency result by minimizing the multi-classifier predictions difference loss.The accuracy of fault classification on the Case Western Reserve University and Qingdao University of Technology bearing datasets reaches 99.63%and 99.40%,respectively.

关 键 词：轴承 故障诊断 多源域自适应 Swin-Transformer 多尺度特征提取 最大均值差异 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TH133.33[自动化与计算机技术—控制科学与工程]