Degradation-aware RUL prediction for rolling bearings via multiresolution variable-length patches and hierarchical KAN

Accurate remaining useful life (RUL) prediction of rolling bearings is important for predictive maintenance in smart manufacturing. However, bearing degradation signals collected in industrial environments are usually nonlinear, non-stationary, and locally fluctuating, which makes continuous full-life degradation trajectory modeling difficult. To address this issue, this paper proposes a degradation-aware multiresolution hierarchical learning framework for rolling bearing RUL prediction. First, root mean square (RMS) is selected as the health indicator according to the degradation trend comparison of multiple statistical features, and a two-stage degradation-aware RUL target is constructed to describe the transition from stable operation to accelerated degradation. Second, a multiresolution variable-length patch module is designed to adaptively adjust the temporal receptive field on RMS-based health indicator sequences, thereby representing both local vibration changes and long-term degradation trends. Third, a hierarchical temporal dependency management module based on Kolmogorov-Arnold Networks (KANs) is constructed to model intra-patch short-term fluctuations, inter-patch long-term degradation evolution, and cross-resolution interactions. Experiments on the XJTU-SY and IMS bearing datasets show that the proposed method achieves competitive and stable prediction performance compared with representative CNN-, RNN-, and Transformer-based baselines. Ablation studies, statistical significance analysis, and cross-bearing experiments further verify the effectiveness of the proposed modules.