Spatio-Temporal Reinforcement Learning for Aero-Engine Remaining Useful Life Prediction

With advances in Industry 4.0, reliability and safety are emphasized in equipment health management. Remaining useful life (RUL) prediction is viewed as the key to health management, which allows proactive maintenance by analyzing historical data. However, existing approaches face significant challenges: model-driven approaches rely on accurate degradation modeling, while data-driven approaches ignore temporal correlation without interpretability. To address the issue, in this manuscript, we propose an expectation-informed spatio-temporal reinforcement learning (ESTRL) for aero-engine RUL prediction. Firstly, we design the ST network to fuse multi-sensor signals, which constructs degradation indicators balancing spatial dependence and temporal dynamics. Then, we introduce RL and design expected reward mechanisms to explicitly explain the correlations between short-term predictions and long-term degradation. The effectiveness of ESTRL is validated by a well-recognized dataset. Results indicate that ESTRL outperforms mainstream RUL prediction methods on both metrics with a promising application.