Dual-stream learning with inverted transformer and BiGRU for rolling bearing remaining useful life prediction

Abstract

Abstract Predicting the remaining useful life (RUL) of rolling bearings is essential for ensuring the reliability and safety of rotating machinery. However, accurate RUL prediction remains challenging due to the non-stationary degradation behavior of bearings operating under complex and time-varying conditions. Existing deep learning approaches often suffer from two key limitations: conventional Transformer architectures treat time steps as tokens, potentially mixing heterogeneous physical meanings across variables, while static feature fusion strategies lack adaptability to different degradation stages. To overcome these challenges, this study proposes a dual-stream framework that decouples spatial correlations from temporal degradation trends. Specifically, an Inverted Transformer is employed to model cross-variable relationships among multiple sensor signals, enabling effective extraction of spatial feature dependencies. Meanwhile, a Bidirectional Gated Recurrent Unit (BiGRU) is introduced to capture long-term temporal evolution and cumulative damage progression. Furthermore, a dynamic gating mechanism is developed to adaptively fuse spatial and temporal representations according to the current degradation stage, thereby enhancing prediction robustness under varying operating conditions. Experimental results on bearing degradation datasets demonstrate that the proposed method achieves competitive performance, reaching an average coefficient of determination ( $$R^2$$ ) of 0.8706. In addition, SHAP-based interpretability analysis indicates that the model focuses on high-frequency degradation indicators closely associated with bearing fatigue failure, providing physically consistent insights into the degradation process. The proposed framework offers a reliable and interpretable solution for predictive maintenance and intelligent condition monitoring of industrial rotating machinery.