Progressive Cross-Model distillation with heterogeneous synergistic classifier for Few-Shot bearing fault diagnosis

In industrial scenarios, current signals are easier to collect than vibration signals, with more flexible sensor deployment. To address the challenges of source-free constraints, weak fault features, and limited fault samples in current signals for bearing fault diagnosis, and to achieve flexible fault diagnosis under complex working conditions, this paper proposes a progressive cross-model distillation method with heterogeneous synergistic classifier (PCMD-HSC) for few-shot bearing fault diagnosis. In this approach, a vibration-trained source model serves as a knowledge provider, guiding the target model to progressively align the response patterns of current features. Different from the knowledge transfer of conventional distillation methods, this approach transfers the inductive biases of the vibration modality to the target model in the form of response patterns, realizing cross-model response patterns learning and significantly enhancing the diagnostic performance of the target model on the current modality. The proposed heterogeneous synergistic classifier integrates the advantages of prior knowledge and metric learning, and performs synergistic decision-making via an inverse-uncertainty balancing fusion mechanism. It effectively addresses the problem of performance degradation of single classifier under few-shot conditions and improves the generalization ability and noise robustness of the model. Extensive experiments conducted on two bearing datasets covering multimodal data fully demonstrate that the proposed method achieves superior diagnostic accuracy and generalization in cross-condition scenarios.