State-Switching Stochastic Estimation Strategy for Multistage Degradation With Sensor Fusion

Predicting the remaining useful life (RUL) of industrial equipment is critical for reliability-centered maintenance, yet existing approaches often struggle to capture stochastic degradation behaviors and stage transitions that occur in real-world operations. Model-driven methods rely heavily on prior physical knowledge, while data-driven models lack interpretability and physical consistency. To bridge this gap, this paper proposes a physics-informed hybrid prognostic framework for multi-sensor systems undergoing multi-stage stochastic degradation. A Hidden Markov Model (HMM) is first employed to identify degradation stages, followed by a Wiener process that characterizes state-specific stochastic dynamics. A data fusion network extracts temporal features from multi-sensor data and estimates Wiener parameters through end-to-end optimization under a physics-constrained loss. This design enables simultaneous improvement of prediction accuracy and physical plausibility. Experiments on turbofan engine and nuclear pump datasets demonstrate that the proposed method achieves superior RUL prediction accuracy and robust uncertainty quantification compared with conventional hybrid and deep learning models.11.