Rolling bearings are essential components of critical rotating equipment in nuclear power systems. Under harsh operating conditions and prolonged continuous operation, they are susceptible to progressive degradation, which can compromise the operational stability and safety of nuclear power plants. Consequently, remaining useful life (RUL) prediction for rolling bearings has become an important research topic in prognostics and health management (PHM) for nuclear power applications. To address the limited adaptability of conventional model fusion strategies to changes in degradation states and the complex nonlinear relationship between degradation features and RUL, this study proposes a hybrid prediction model integrating a temporal convolutional network (TCN) with Kolmogorov–Arnold networks (KANs). Specifically, the TCN extracts multiscale temporal information from degradation sequences within a given time window, while a front-end KAN learns a nonlinear state representation from the current degradation features. The gating module adaptively generates fusion weights based on the outputs of the two branches and uses these weights to combine the state and temporal representations following dimensional alignment. An independent downstream KAN then predicts RUL from the fused representation, enabling joint modeling of the current degradation state and historical temporal information. Comparative and ablation experiments were conducted on the IEEE PHM 2012 Challenge bearing dataset to evaluate the effectiveness of the proposed method. The experimental results indicate that the proposed model achieves good predictive accuracy and generalization performance.

Remaining useful life prediction,Rolling bearing,Temporal Convolutional Network,Kolmogorov-Arnold network