Multiple Source Domain Transfer Fault Diagnosis Method in Rolling Bearing Under Variable Working Conditions

Through domain adaptation fault diagnosis methods, it is possible to effectively address the issue of insufficient utilization of multiple similar datasets caused by significant variations in the vibration data distribution of bearings under different working conditions in actual industrial production. Typically, labeled fault samples cover various working conditions (i.e., multiple source domains), but the diagnostic knowledge provided by each individual source domain is limited. Therefore, a multi-source domain transfer fault diagnosis method based on adversarial transfer (MSDTFDM) is proposed. Initially, shared multi-wavelet convolutional feature extractors are employed to extract fault knowledge from multiple source domains. Subsequently, Maximum Mean Discrepancy is utilized to achieve global alignment between samples, and a multi-source correction prototype category alignment method is introduced. Through adversarial transfer learning, comprehensive transfer of fault knowledge from multiple source domains to the target domain is achieved, enhancing the model fault diagnosis performance under variable working conditions. The effectiveness of the proposed method is validated through six sets of cross-condition fault diagnosis experiments on cylindrical roller bearings operating at constant and variable speeds.