Rolling bearing fault diagnosis is crucial for ensuring industrial production safety. However, traditional methods exhibit limited generalization under varying loads, speed fluctuations, and multi-source domain distribution shifts. To address the challenges of source domain conflict, contribution imbalance, and global-local alignment in multi-source domain adaptive fault diagnosis, this paper proposes a wavelet-based subdomain-enhanced multi-source domain adaptive network. The network comprises three core modules: a shared feature extraction module, a private feature enhancement module, and an adaptive feature fusion module. The shared feature extraction module employs multi-scale wavelet convolution to capture multi-scale frequency characteristics, thereby enhancing feature representation capability. The private feature enhancement module integrates residual connections with a hybrid attention mechanism to emphasize critical fault-related information, while a domain discriminator extracts domain-invariant features. The adaptive feature fusion module achieves dynamic weight allocation and collaborative integration of multi-source domain knowledge. Experimental validations based on the bearing datasets from the University of Ottawa in Canada and Huazhong University of Science and Technology (HUST) demonstrate that the proposed method can achieve efficient inter-domain alignment, robust feature extraction, and accurate fault diagnosis. Additionally, it has fewer model parameters and outstanding advantages in training efficiency.
Duan et al. (2026) studied this question.