Multiscale Transfer Learning Based Fault Diagnosis of Rolling Bearings

This paper aims to address two challenges for fault diagnosis of rolling bearings: full feature representation and feature alignment across working conditions. To tackle these issues, we propose a novel Multiscale Transfer Learning (MSTL) approach. The proposed model is capable of capturing, fusing, and aligning multiscale features across different working conditions. We introduce a novel multi-stream architecture to effectively process multiscale factors of raw signals, and the backend module incorporates a feature aligning and classification unit. Notably, our proposed model provides dynamic weights to optimize the extent of multiscale fusion, thus a task-related fusion mechanism can be adaptively achieved. To measure the difference between feature distributions under changing working conditions, we utilize the Wasserstein distance, which facilitates feature transferring during model learning. Our proposed MSTL method demonstrates superior performance compared to existing methods, resulting in improvements of 12.61% and 17.00% over conventional CNN methods on the CWRU and Paderborn datasets.