Remaining useful life prediction under noisy conditions via adaptive degradation modeling incorporating memory effects

Accurate modeling of degradation processes is fundamental to reliable remaining useful life (RUL) prediction for mechanical systems. However, practical degradation processes are often affected by measurement uncertainty, non-Markovian degradation characteristics, and evolving degradation mechanisms, which may limit the prediction performance of existing methods. To address these challenges, this paper proposes an adaptive degradation modeling framework that jointly considers measurement uncertainty, memory effects, and online model adaptation. The proposed method first constructs an initial degradation model using the training units, with the model parameters estimated via maximum likelihood estimation (MLE). An improved cumulative sum (CUSUM) procedure is then employed to detect structural variation points by comparing the model predictions with online monitoring data. Once a structural variation is detected, the degradation model is adaptively updated using the newly available observations while retaining the stochastic degradation structure. An enhanced particle filtering (PF) algorithm is employed to recursively estimate the latent degradation state. For RUL prediction, the first hitting time (FHT) concept is adopted to derive an approximate probability density function (PDF) based on the adaptively updated degradation model. Experimental results obtained under both the internal data partitioning strategy and the official C-MAPSS training/testing protocol demonstrate that the proposed framework achieves improved predictive performance compared with the benchmark methods.