KAP-Former: a knowledge-aware physics-informed transformer for cross-domain few-shot bearing fault diagnosis

Abstract

Recent advances in deep metric learning and meta-learning have shown promise for few-shot bearing fault diagnosis, yet their cross-domain generalization remains limited under cross-sampling-rate conditions, cross-test-rig transfer, and strong background interference, mainly due to frequency-coordinate misalignment and attention drift. To address this issue, this paper proposes a knowledge-aware physics-informed Transformer, termed KAP-Former, for cross-domain few-shot bearing fault diagnosis. Specifically, a speed-adaptive physical embedding is developed to map frequency indices into the physical order domain according to the actual sampling rate and rotational speed, thereby improving feature consistency across operating conditions. An elastic physical attention gating module is further introduced to guide self-attention toward theoretically relevant fault harmonic regions via differentiable Gaussian masks, suppressing non-target components such as gear meshing interference and structural resonances. In addition, a physics prior-oriented constraint is designed to regularize the learned attention distribution toward the theoretical physical distribution, enhancing physical consistency and interpretability. Experiments on the CWRU, SEU, and HUST datasets demonstrate that KAP-Former consistently outperforms representative few-shot diagnosis methods in both same-domain and cross-domain tasks. In particular, it achieves F1-scores of 92.68% and 89.14% on the 1-shot CWRU →SEU and CWRU →HUST tasks, respectively.