Sparse stage path graph Kolmogorov-Arnold Networks (KANs) for bearing remaining useful life prediction

Graph Neural Networks (GNNs) have been utilized to analyze monitoring data for predicting the remaining useful life (RUL) of bearings. Conventional methods typically aggregate degradation features through GNNs after converting the monitoring data into graphs. However, traditional GNNs face two primary challenges: (1) a single stage of graph data construction is insufficient to capture the complexity of bearing degradation, and (2) the effectiveness of these methods heavily depends on the selection of initial weights, making them susceptible to noise interference. To address these challenges, this paper proposes a novel sparse stage path graph Kolmogorov-Arnold network (SSPGKAN) model to enhance bearing RUL prediction. First, a siamese KAN model is developed to classify bearing degradation stages, and stage path graphs (SPGs) are constructed to aggregate the degradation features. Subsequently, the SSPGKAN model replaces traditional edges in the graph structure with learnable functions, overcoming the limitations of initial linear weight selection in GNNs and enabling dynamic transmission of graph data. Additionally, a sparse graph optimization method is used to reduce the impact of noise and accurately predict bearing RUL. Verification results from two test bench datasets demonstrate that the SSPGKAN model improves the prediction efficiency and accuracy of bearing RUL.