Remaining Useful Life Prediction for Bearings Using Multiscale Temporal-Enhanced Deformable Attention With Soft Thresholding

The run-to-failure monitoring data of bearings can be utilized for modeling and analysis to predict their remaining useful life (RUL). Such data typically contain degradation features across multiple time scales and are often accompanied by noise interference. However, most existing deep learning (DL) methods struggle with effectively suppressing noise, dynamically capturing key degradation patterns, and integrating multiscale degradation features. To address these challenges, this study proposes a novel end-to-end RUL prediction method for bearings based on multiscale temporal-enhanced deformable attention with soft thresholding (MTEDAST). Specifically: 1) a soft-threshold denoising block based on an autoencoder (AE) is introduced, which dynamically adjusts the denoising threshold according to data characteristics, thereby effectively removing low-amplitude noise and enhancing key feature representations; 2) a temporal-enhanced deformable attention (TEDA) mechanism that incorporates temporal bias information into the attention mechanism is proposed, enabling it to dynamically adapt and capture critical temporal degradation patterns; and 3) a dynamic multiscale attention fusion (DMAF) block is designed, which employs element-wise multiplication to dynamically integrate multiscale features, enhancing informative scales while suppressing irrelevant ones. Experimental results on the PHM2012 and XJTU-SY datasets demonstrate that the proposed method outperforms other state-of-the-art approaches with an average improvement of 28.04% in the score metric and average reductions of 36.66% and 35.89% in mean absolute error (MAE) and root mean-square error (RMSE), respectively.