Remaining useful life prediction of rotating machinery based on transformer-KAN

Abstract Remaining useful life (RUL) prediction is critical for ensuring the safe and efficient operation of rotating machinery, which plays a vital role in industrial production. However, the traditional transformer model, relying on a simple linear output in its final layer, is limited in capturing nonlinear relationships. The nonlinear Wiener process (WP) model, despite its strengths in characterizing stochastic degradation, fails to fully capture the intricate time-series features in equipment operation data. To address these challenges, this paper introduces an improved method for RUL prediction which combines transformer and Kolmogorov–Arnold network (KAN) and utilizes WP for uncertainty quantification. The transformer can efficiently capture the long-term dependencies in equipment degradation data through its powerful sequential modeling capability and global attention mechanism, and its combination with the KAN further enhances the flexibility and accuracy of feature extraction. In this paper, the transformer-KAN model is proposed as a drift function for WP. Meanwhile, this paper utilizes the first hit time to derive an approximate expression for the RUL probability density function and estimate the drift and diffusion coefficients of the transformer-KAN-WP model. The method is validated on the bearing dataset and the tool holder power head degradation data and compared and analyzed with other commonly used methods, which proves the effectiveness of the method in improving the prediction accuracy.