Adaptive Wiener process modeling integrating physical causality and degradation states for high-reliability systems

Performance degradation of high-reliability equipment often involves challenges, particularly the coupling of multiple components and the concealment of early degradation data. These challenges make degradation modeling methods based on traditional data difficult to apply in engineering practice. To address these issues, this paper proposes a novel degradation modeling framework based on physics-informed neural networks. This comprises two modules: feature extraction of physics mechanisms and degradation feature extraction. The former employs a Bayesian causal attention mechanism to extract key physical causal features, capturing the coupled interaction effects among multiple components. The latter utilizes temporal convolutional networks to capture degradation state features, extracting subtle deterioration characteristics from time-series data. These are then fused into a joint representation and embedded in a nonlinear Wiener process to characterize complex degradation dynamics. For enhanced trainability and physical consistency, the Wiener equation is discretized using the Milstein method, and a physics residual regularization term is incorporated into the loss function. During online deployment, a two-stage update strategy is employed to dynamically refine model priors. Experimental validation based on an aerospace product dataset demonstrates that the proposed method achieves high prediction accuracy even under limited-sample conditions.