Accurate prediction of the remaining useful life (RUL) of rolling element bearings (REBs) in critical rotating equipment of nuclear power plants is of great significance for ensuring safe and stable unit operation and optimizing maintenance strategies. In the field of RUL prediction, various deep learning methods possess distinct technical advantages; however, traditional model fusion approaches often significantly increase network complexity and are prone to problems such as gradient vanishing during training. To address these challenges, this paper proposes an ensemble learning framework based on Dual-View Transformer and Multi-Scale Temporal Convolutional Network (TCN) for RUL prediction of REBs. First, health indicators of REBs are constructed via autoencoders, and the bearing degradation onset point is determined using the three-sigma (3σ) criterion. Second, two base models—Dual-View Transformer and Multi-Scale TCN—are built, where the former excels at capturing long-term temporal dependencies in data and the latter focuses on extracting local transient variation features; both output RUL prediction results independently. Finally, to overcome the insufficient fitting capability of existing meta-models for nonlinear data, a LinGate meta-model is designed to fuse the prediction results from the two base models, thereby achieving accurate RUL prediction for REBs. Comparative and ablation experiments conducted on the PHM2012 dataset validate the effectiveness and practical value of the proposed method. The method improves RUL prediction accuracy for REBs and provides support for health management and operation-maintenance decision-making of rotating machinery under complex working conditions such as nuclear power.

Rolling Element Bearing,Remaining useful life (RUL),Ensemble Learning,Temporal Convolutional Networks (TCN),Transformer