Physics-informed acoustic–vibration fusion for all-ceramic bearing fault diagnosis under complex operating conditions

Complex noise and variable operating conditions substantially reduce the observability of fault-induced impact and modulation features in acoustic–vibration measurements of all-ceramic bearings and create modality-dependent differences in signal reliability. To address this problem, this study proposes PH-FusionNet, a physics-informed network for acoustic–vibration signal fusion and fault diagnosis. The proposed method follows a progressive framework consisting of local physics-prior enhancement, global representation stabilization, and reliability-aware fusion. First, a Laplace–Morlet physics-prior feature extraction layer is developed based on the impact–resonance mechanism to enhance fault-related features, including vibration impact responses and acoustic modulation wave packets. Subsequently, a long-range global competitive representation enhancement module is designed to aggregate long-range contextual information and recalibrate global responses, thereby strengthening fault-related representations while suppressing non-stationary noise interference. Finally, an impulse-guided gated fusion mechanism is introduced, in which vibration impact responses serve as structural references for selective acoustic-feature injection and vibration-conditioned fusion. Experimental results show that, across five representative noise conditions, PH-FusionNet achieved an average diagnostic accuracy of 86.84%, outperforming the second-best model by 3.48 percentage points, while attaining an average FNR of 13.16%. The model also retained comparatively robust diagnostic performance under asymmetric modality degradation and simulated speed-transition conditions, thereby demonstrating the effectiveness and robustness of the proposed method.