Multivariate adaptive Brownian Motion-Particle Filter framework for remaining useful life prediction of nonlinear and state-noise coupled degradation process

Degradation modeling and remaining useful life (RUL) prediction are key techniques in prognostics and health management (PHM). However, existing frameworks driven by Brownian Motion struggle to effectively integrate nonlinear degradation features or accurately separate degradation state from the inherent model noise and observation noise, leading to prediction inaccuracies. To address these limitations, a multivariate adaptive Brownian Motion-Generic Particle Filter framework is proposed. The framework enhances prediction accuracy and reduces uncertainty by comprehensively coordinating the three stages of RUL prediction. In degradation modeling stage, multiple drift terms are introduced to accommodate complex nonlinear degradation patterns, while the first-order difference with nonlinear least square (NLS) provides accurate and robust initial parameter estimation. In parameter update stage, a state-noise coupled discrete state-space model is constructed, which considers the uncertainty of all hidden variables, enabling real-time adjustment of parameter distributions based on degradation data. Hidden variables are adaptively updated using two particle filter strategies tailored to different dimensions, mitigating particle degeneracy and improving the separation of true degradation state. In RUL solution stage, the RUL distribution is extended into a stochastic process to quantify prediction uncertainty. Simulation experiments demonstrate the framework’s ability to switch particle filter strategies based on dimensionality and illustrate its overall workflow. Validations on the mechanical torsion bars dataset and the MOSFET dataset demonstrate a 31.5 % improvement in average prediction accuracy over the baseline method.