A Robust Bayesian Framework for Degradation State Identification in the Presence of Outliers

ABSTRACT Accurate degradation state estimation is critical for predictive maintenance, yet it is often compromised by measurement outliers and parameter uncertainty. Existing methods either assume Gaussian measurement errors, which are sensitive to outliers, or overlook parameter uncertainty, leading to overconfident predictions. To address these challenges, we propose a Bayesian online degradation state estimation framework that integrates robust error modeling with parameter uncertainty quantification. Specifically, we model measurement errors using a Student's‐ distribution to handle outliers and employ variational Bayes with Laplace and Gamma approximations to estimate posterior distributions of degradation states and parameters efficiently. This framework enables real‐time updates, ensuring adaptability to dynamic operating conditions. Furthermore, based on the estimated degradation states, we derive real‐time remaining useful life predictions and dynamic maintenance strategies under a cost function model. Numerical experiments and case studies demonstrate the framework's robustness, computational efficiency, and practical applicability.