A dual-module approach with sliding variable-length window for remaining useful life prediction in bearing

Deep Learning (DL) has significantly advanced Prognostics and Health Management (PHM) by enabling automatic feature extraction for Remaining Useful Life (RUL) prediction. However, bearing run-to-failure degradation typically evolves through multiple stages—Health Stage (HS), Slow Degradation Stage (SDS), and Accelerated Degradation Stage (ADS)—characterized by different degradation rates. It poses a challenge to traditional DL-based RUL prediction models to effectively capture the temporal features under multiple degradation rates. To address this issue, this paper proposes a Dual-Module with Sliding Variable-length Window (DM-SVW) approach. Specifically, within the proposed Dual-Module (DM) approach, the health module extracts informative temporal features from HS data and transfers the learned prior knowledge to the degradation module. Guided by this prior information, the degradation module models the overall degradation trajectory using DS data to predict the RUL. In addition, to further overcome the limitations of the traditional Sliding Window (SW) strategy, which requires explicit stage division within the DS, a Sliding Variable-length Window (SVW) strategy is employed to jointly model the SDS and ADS, thereby capturing global dependencies across both stages. To implement the proposed DM-SVW approach, a semi-supervised DL model, namely Time-Transformer and Denoise Auto-Encoder (TT-DAE), is developed. The model integrates a self-supervised DAE for robust spatial feature extraction and a supervised TT for RUL estimation. The effectiveness of the DM-SVW approach is validated on the XJTU-SY bearing dataset, a laboratory bearing dataset, and the N-CMAPSS turbofan engine dataset. Experimental results demonstrate that the DM-SVW approach achieves the lowest RMSE values (11.83 and 0.158 on XJTU-SY, 0.044 on the laboratory dataset). It outperforms DL models employing the traditional SW and SVW strategies, as well as two-stage DL models, achieving a 13.4% reduction in RMSE compared with the two-stage approach. Additional comparisons with L S T M - D A E SVW , B i L S T M - D A E SVW , G R U - D A E SVW , and Time-Transformer Encoder and DAE (TTE-DAE) further validate its superior capability to capture global dependencies across the SDS and ADS.