Imbalanced Deep Transfer Network for Fault Diagnosis of High-Speed Train Traction Motor Bearings

Transfer learning-based fault diagnosis methods have been increasingly utilized for major equipment, including high-speed trains, turbine machines, and aircraft engines. However, most traditional transfer methods based on implicitly balanced data only consider feature shift. When applied to high-speed train traction motor bearing fault diagnosis, the cross-domain generalization ability of these transfer methods is weakened by label shift. Due to the complex operating conditions of high-speed trains, these transfer methods often fail under multiple operating conditions, resulting in reduced cross-domain diagnostic accuracy when faced with feature shift and label shift simultaneously. Therefore, we propose the imbalanced deep transfer network (IDTN) to tackle the aforementioned problem in cross-domain fault diagnosis of high-speed train traction motor bearings. Firstly, IDTN overcomes the influence of imbalanced distributions in source domain samples through deep imbalanced learning. Then, batch nuclear-norm maximization is introduced to enhance the prediction discriminability and diversity of the target domain samples. Finally, case studies of the high-speed train traction motor bearing fault dataset and the Case Western Reserve University bearing fault dataset are conducted. Experimental results prove the effectiveness and superiority of IDTN in the cross-domain fault diagnosis field with both feature shift and label shift.