A Robust Double‐Stage Feature Selection and TCN‐iT‐KAN Hybrid Framework for Remaining Useful Life Prediction of Rolling Bearings
Abstract
The accurate prediction of the remaining useful life (RUL) of rolling bearings remains a challenging problem because degradation representations are often influenced by noise, feature redundancy, and the limited run‐to‐failure data. To overcome these challenges, this study develops a leakage‐aware prognostic framework that integrates a robust double‐stage feature selection strategy with a hybrid parallel predictor. The prediction model combines a parallel temporal convolution network, an inverse‐style Transformer, and a Kolmogorov–Arnold network head (TCN‐iT‐KAN). After identifying the first prediction time, a multi‐domain candidate feature bank is constructed from the time, envelope, frequency, wavelet packet, and discrete wavelet transform domains. Then, a double‐stage feature selection procedure is performed on the training set only to preserve informative features and avoid data leakage between data splits. The selected feature sequences are then used as input to a parallel prediction model, where the TCN captures local and multi‐scale temporal patterns, the inverse Transformer branch models global dependencies between features, and the KAN head performs nonlinear regression. Experimental results on the XJTU‐SY rolling bearing dataset, using a predefined data split, show that the proposed framework generally outperforms the benchmark models. The findings from ablation studies, branch sensitivity analysis, and post‐hoc feature analysis further show that these improvements are due to the combined effect of train‐only feature screening, parallel modeling of temporal feature interactions, and the KAN regression head.