The circulating water pump is a critical component of the circulating water system in the conventional island of nuclear power plants. As the key support of rotating parts in the pump unit, rolling bearings directly affect the operational reliability of the unit. Due to long-term continuous operation, load variation, and operating condition fluctuation, different bearings exhibit significant individual differences in health baseline, degradation amplitude, and feature evolution rate. Moreover, existing data-driven methods are mostly based on single-feature-scale modeling, making it difficult to effectively suppress the influence of individual differences on cross-bearing prediction. To address this issue, a dual-view GRU-GAT parallel network is proposed for rolling bearing remaining useful life (RUL) prediction. Specifically, a global normalization view that preserves the absolute feature level and a self-baseline standardization view that characterizes the change relative to the health baseline are constructed. GRU and a multi-scale temporal graph attention network are employed in parallel to extract the continuous temporal evolution of degradation features and the structural correlations among states at different time steps, respectively. The dual-branch features are then fused, and temporal attention pooling is applied to establish the mapping between degradation information and remaining useful life. Experiments are conducted on the public XJTU-SY bearing dataset and a self-built accelerated degradation dataset of cylindrical roller bearings, involving both cross-bearing tests under the same operating condition and cross-condition tests. The results show that the proposed method achieves lower MAE and RMSE than the compared methods (GCN, Transformer, and CNN-LSTM) on all four test bearings, verifying its effectiveness and providing a methodological reference for predictive maintenance of rolling bearings in nuclear power plant circulating water pumps.

Rolling bearing,Remaining useful life prediction,Graph attention network,Dual-view representation,Self-baseline standardization