IBN-MixStyle Network With Dynamic Weighted Invariant Risk Minimization for Domain-Generalized Bearing Fault Diagnosis

In recent years, substantial progress has been made in cross-working condition fault diagnosis through transfer learning. However, a significant trade-off between diagnostic accuracy and generalization capability remains unresolved, especially in practical industrial scenarios where data from unseen domains is unavailable. To address this challenge, this paper proposes a novel domain generalization (DG) method that effectively integrates both discriminative and generalizable features for rolling bearing fault diagnosis under varying working conditions. Firstly, Fast Fourier Transform (FFT) is employed to extract phase-frequency representations, which are fused with time-domain signals to form a comprehensive input to the diagnostic network. Then a dual-branch architecture based on Instance-Batch Normalization (IBN) and MixStyle is constructed, the IBN captures discriminative features to enhance classification accuracy, and the MixStyle learns generalizable features to improve cross-domain adaptability. Furthermore, a Dynamic Weighting Invariant Risk Minimization (DWA-IRM) strategy is designed to assign higher importance to source domains with poor performance, thereby enforcing robust invariance constraints. The extensive experiments show that the proposed method achieves an average accuracy improvement of approximately 5.3% to 15.47% over state-of-the-art DG approaches, which demonstrate its effectiveness. This work offers a practical and scalable solution for intelligent fault diagnosis in highly variable industrial environments.