Integrating frequency domain reconstruction with time-adaptive fusion network for rolling bearing remaining useful life prediction

Abstract Rolling bearing remaining useful life (RUL) prediction plays a critical role in equipment health management. However, noise in practical engineering applications can submerge early weak degradation features, making it difficult to capture early degradation patterns. This constrains the application of prediction methods across all degradation stages. To address these issues, this paper proposes a rolling bearing RUL prediction method based on a vibration signal reconstruction-enhanced method (VSRM) and a time-adaptive fusion network (TAFN). First, VSRM performs decoupled learning and non-linear reconstruction of the signal’s amplitude and phase spectra in the frequency domain. This effectively suppresses noise and significantly enhances the robustness and representational power of degradation features. Subsequently, TAFN introduces timestamp encoding as an explicit global temporal prior and designs a temporal-position-driven adaptive weighting mechanism. This achieves a dynamic fusion of global trends and local features, thereby resolving the lag problem that existing models face when capturing non-linear changes in degradation rates. In this work, the VSRM-TAFN framework is deeply integrated with several mainstream time-series prediction networks and comprehensively validated on two full-life-cycle rolling bearing datasets. The results demonstrate that the proposed VSRM-TAFN framework significantly improves the prediction accuracy of all mainstream time-series networks, achieving a minimum RMSE of 0.042 and a minimum MAE of 0.0327. This general architecture effectively overcomes the challenges posed by noise and non-linear degradation rate variations, providing an effective and universal solution for achieving high-robustness RUL prediction.