A Cross‐Modal Gated Fusion Network Synergizing Physical Priors and Deep Learning for Robust Bearing Fault Diagnosis

Reliable fault diagnosis in rotating machinery is often hindered by poor cross‐individual generalization and the inherent black‐box nature of data‐driven models. To address these challenges, this paper introduces a hybrid diagnostic framework that synergizes explicit physical priors with deep spatiotemporal representations. The proposed architecture employs a dual‐stream strategy, integrating a residual multi‐scale CNN‐BiLSTM branch with an expert knowledge feature stream through a cross‐modal gated fusion mechanism. Unlike conventional concatenation, this gated module adaptively modulates the contribution of each feature stream, effectively resolving semantic inconsistencies between data‐driven and physics‐informed modalities. Comprehensive evaluations on a rigorous cross‐individual bearing fault benchmark demonstrate that the proposed framework consistently outperforms existing deep learning models. Further ablation studies and manifold visualizations confirm that the integration of physical priors enhances feature discriminability and model robustness. Moreover, an analysis of the gated coefficients reveals that the model dynamically adjusts its reliance on different feature streams in alignment with the bearing degradation process, thereby providing enhanced interpretability and clear evidence of its diagnostic superiority.