Multistage health state estimation and uncertainty quantification for bearing remaining useful life via graph-guided representation learning

Bearing degradation follows stochastic and nonlinear trajectories, which makes the accurate prediction of remaining useful life (RUL) a critical challenge in industrial prognostics. Recent deep learning-based methods have achieved promising performance in bearing RUL prediction. However, they have limited ability to explicitly account for changes in degradation behavior during the prognostic process. Health-stage information has been employed to characterize degradation progression, but it is typically used only for stage estimation rather than guiding degradation representation learning or RUL prediction. To address this limitation, we propose the Stage-Aware Graph-guided Embedding-Uncertainty Quantification (SAGE-UQ) framework. The proposed framework consistently leverages health-stage information through degradation representation learning and uncertainty-aware RUL prediction. Specifically, graph-guided spectral representation learning and stage-aware temporal contrastive learning were employed to jointly estimate health-stage information. This estimated stage information is further exploited to learn degradation representations that capture health-stage transition characteristics. It is then leveraged to guide uncertainty-aware RUL predictions. The proposed framework explicitly incorporates health-stage transition ambiguity into predictive uncertainty. Experimental results demonstrate that SAGE-UQ consistently outperforms representative baselines and recent state-of-the-art methods. It achieves an average root-mean-square error (RMSE) reduction of approximately 17.4%. The lightweight architecture and low inference latency enhance its practicality for online prognostic applications.