To address the limited representation capability of single-modal signals and the lack of fault-mechanism constraints in conventional multimodal fusion, this paper proposes a physics-prior-enhanced dynamic vibration–current fusion method for motor fault diagnosis. Vibration and current signals are first transformed into time-frequency representations using the continuous wavelet transform (CWT), and different fusion levels are systematically evaluated to determine an effective fusion scheme. A dynamic feature fusion mechanism is then developed to adaptively generate spatial modality weights from joint vibration–current representations, enabling sample-dependent exploitation of complementary multimodal information. Furthermore, a fault-category-specific modality sensitivity prior is introduced as a physical soft constraint to guide the learned fusion weights toward fault-mechanism-consistent modality contributions, thereby coordinating data-driven feature learning with prior physical knowledge. Experimental results under 20 dB noise demonstrate robust fault classification performance. Incorporating the fault sensitivity prior further improves classification accuracy and F1-score while reducing the physical consistency loss, validating the effectiveness and interpretability of the proposed method.

motor fault diagnosis,fault sensitivity prior,vibration–current fusion,dynamic feature fusion,physics-prior constraint