Remaining useful life prediction of rolling bearings based on stage-weighted feature fusion and stage-wise domain alignment

Abstract

To address the domain distribution shift problem in remaining useful life (RUL) prediction of rolling bearings under cross-condition settings, this paper proposes a Stage-Weighted Domain Alignment Network. Most existing methods assume that deep features have stronger transferability, while paying limited attention to the different roles of stage feature representations in prediction and domain adaptation (DA). To this end, a multi-stage weighted network is developed to construct three-stage feature representations. The stage weights are generated based on the three-stage features of the source domain, enabling weighted selection among features from different stages. In the prediction branch, the stage weights are used to perform weighted fusion of multi-stage features for RUL regression. In the DA branch, multi-kernel maximum mean discrepancy alignment is performed separately on the features of each stage, and the same set of stage weights is used to adjust the alignment strength of different stage features. In this way, stage selection and stage-wise domain alignment are integrated into a unified framework. Experimental results show that the proposed model achieves an average RMSE of 0.1152, an average MAE of 0.0943, and an average R2 of 0.8435 across eight bidirectional cross-domain tasks. Compared with the single-stage and fixed-fusion variants, the proposed strategy achieves lower average errors and a higher average R2 under the eight evaluated cross-domain tasks.