TF-MSWCT: a dual-channel fusion network for remaining useful life prediction of rolling bearings

Abstract As critical components in mechanical systems, rolling bearings require accurate Remaining useful life (RUL) prediction to ensure operational safety and optimize maintenance strategies. Given the limitations of existing approaches in multi-scale feature extraction, time–frequency information fusion, and computational efficiency, this paper proposed a novel time–frequency joint model, TF-MSWCT, which integrates multi-scale weighted convolution (MSWC) with a combined masked attention (CMA) mechanism. The model features parallel time- and frequency-domain branches, where the MSWC module adaptively extracts features across different scales. Moreover, the CMA module effectively captures long-term temporal dependencies. This architecture significantly enhances the representation of degradation features and improves prediction accuracy. Experimental results on two publicly available bearing datasets, FEMTO and XJTU-SY, demonstrate that TF-MSWCT consistently outperforms mainstream baselines in terms of mean absolute error (MAE) and root mean square error (RMSE), while maintaining superior computational efficiency and generalization capability. The proposed model offers a robust, accurate, and efficient solution for RUL predicting rolling bearings, with strong potential for practical industrial applications.