A cross-modal fusion network for equipment fault diagnosis exclusively adopting current signals

Abstract

Fault diagnosis based solely on current signals is attractive for practical deployment because of its low sensing cost, but its performance is easily degraded by weak fault representation and rapid speed variations. To address these issues, this study proposes a cross-modal feature collaborative matching distillation framework that transfers knowledge from an information-rich teacher model to a current-only student model. The proposed framework first employs a cross-modal feature alignment module to reduce the representation discrepancy between heterogeneous signals. An adversarial training strategy is then introduced to further constrain the distribution discrepancy between teacher and student features. In addition, hierarchical knowledge distillation transfers both intermediate feature knowledge and high-level decision knowledge, while a staged training strategy improves the stability of the distillation process. Experimental validation on an escalator bearing test bench and an electric heavy-duty truck drive-axle platform demonstrates that the proposed current-only model achieves diagnostic accuracies of 97.92% and 96.87%, respectively, while maintaining performance close to that of the information-rich teacher model. These results demonstrate the effectiveness of the proposed framework for low-cost fault diagnosis under variable-speed conditions.