Resource-constrained continual adaptation for online aero-engine remaining useful life prediction under non-stationary operating conditions

Abstract

Online remaining useful life prediction for aero-engines is challenged by evolving operating conditions and degradation patterns, which may cause static offline models to lose reliability after deployment. Full retraining and unrestricted historical replay are also impractical for resource-constrained prognostic systems. This study proposes a continual-adaptation framework based on a function-aware heterogeneous constraint strategy. Representation-related parameters are protected through orthogonal gradient descent, whereas dynamics-related parameters are regularized using dynamically updated Fisher information. Residual-driven replay is further introduced to prioritize difficult samples under limited buffer capacity. The framework is evaluated on the four C-MAPSS subsets using an engine-level streaming protocol and five independent random seeds. After five adaptation stages, the proposed method reduces root mean square error (RMSE) by 9.3%, 5.0%, 6.5%, and 2.7% on FD001–FD004, respectively, relative to the pretrained model. It also achieves the lowest final RMSE among representative continual-learning baselines and reduces the mean RMSE-based forgetting from 0.13 cycles for the strongest external baseline to 0.07 cycles. The average update time is 11.86 s per incremental batch, with a peak GPU memory usage of approximately 300.5 MB. These results demonstrate an effective balance among predictive accuracy, knowledge retention, and bounded resource overhead under non-stationary degradation conditions.