A Physics‐Guided Shared‐Specific Stage‐Aware Prototype Transfer Network for Remaining Useful Life Prediction of Rolling Bearings

Accurate prediction of the remaining useful life (RUL) of rolling bearings under different operating conditions remains challenging, because variations in operating conditions distort the consistency of degradation representations across domains. Existing transfer learning methods often fail to explicitly separate transferable degradation semantics from residual information associated with specific operating conditions, while insufficiently modeling the dependence of bearing degradation on stage progression. This limitation makes it difficult to preserve the intrinsic temporal structure and may result in negative transfer. Moreover, most existing methods rely on post hoc interpretation strategies that cannot impose physical constraints during forward propagation, thereby limiting model interpretability. To address these issues, this study proposes a physics‐based shared stage‐aware prototype transfer network for cross‐condition bearing RUL prediction. The proposed framework uses separate shared and specific representation branches, applies stage‐aware prototype transfer in the shared space, and performs RUL regression using their gated fusion. By combining physics‐guided representation learning with stage‐aware semantic transfer, the proposed method enhances cross‐domain consistency while improving robustness and interpretability. Experiments conducted on the XJTU‐SY bearing dataset and an accelerated bearing life test platform demonstrate that the proposed method achieves superior prediction performance and satisfactory cross‐platform generalization capability. The results indicate that separating the transfer of shared degradation semantics from the final RUL regression provides an effective and interpretable approach for cross‐condition bearing RUL prediction.