关键词: 故障诊断, 滚动轴承, 迁移学习, 对比对齐, 子域自适应

Abstract: The fault data of bearings exhibited significant distribution discrepancies under varying operating conditions， relatively low diagnostic accuracy was resulted in practical fault detection models. Additionally， most existing research on cross-domain bearing fault diagnosis primarily emphasized inter-domain alignment and intra-class comparison， while neglecting the influences of interactions between subdomains. Therefore， a cross-domain fault diagnosis method of bearings was proposed based on joint subdomain contrast alignment. In order to highlight the fault features， the bearing vibration signals were transformed into time-frequency graph by short-time Fourier transform， and the fault features were obtained by inputting them into the feature extraction module. Domain adaptation methods achieved cross-domain recognition by transferring knowledge learned from the source domain to the target domain. During the domain adaptation processes， a joint subdomain contrast alignment strategy was used to bring samples from the same subdomain closer together while separating samples from different subdomains， which aligned the subdomain distributions of the same class samples among the source and target domains， thereby enhancing the models generalization ability in the target domain. Resnet34 was used as the feature extraction network on the model architecture， and the maximum mean difference was used at the output of the network to align the global distribution of the source domain and the target domain. Compared with the classical domain adaptation methods， the experimental results on the bearing fault data set of Case Western Reserve University shows that the cross-domain fault diagnosis method of bearings based on joint subdomain contrast alignment has better feature transfer ability. 

Key words:  , fault diagnosis, rolling bearing, transfer learning, contrast alignment, subdomain adaptation