Risk-Aware Predictive Maintenance for Aero-Engines via CVaR-Guided Reinforcement Learning and LLM-Based Risk Abstraction

Abstract

A fundamental limitation of current predictive maintenance is the inability to transform prognostic uncertainty into actionable decisions, which often leads to either wasteful early interventions or catastrophic failures in aero-engines. To address this, we propose the Risk-Aware Predictive Maintenance (RAPM) framework consisting of four integrated components: a probabilistic stacked sparse autoencoder that extracts degradation features while providing distributional remaining useful life (RUL) predictions; a Monte Carlo Dropout module with Conditional Value at Risk (CVaR) that summarizes lower-tail predictive risk and produces a conservative RUL estimate for decision making; a locally deployed large language model (Qwen2.5-3B) that abstracts numerical uncertainty into discrete semantic risk levels, enabling interpretable reward shaping; and a Proximal Policy Optimization agent trained with curriculum learning to handle long-horizon sparse-reward maintenance optimization. On the NASA C-MAPSS FD001 benchmark, RAPM achieves a 96.0% optimal stopping rate with no late interventions or failure events compared with 90.2% for the strongest independent maintenance baseline. Under controlled substitution with LSTM-MCD, Transformer-MCD, CNN-MCD, and Deep Ensemble prognostic backbones, the same CVaR-LLM-PPO decision architecture maintains optimal stopping rates of 92.4–95.2% with no late interventions or failure events. Within the matched Rule-PPO ablation, CVaR increases the mean optimal stopping rate from 92.0% to 95.0% under unperturbed conditions, which is a gain of 3.0 percentage points. Under Gaussian sensor noise (σ=0.05), the gain reaches 9.0 percentage points, from 73.0% to 82.0%. These results demonstrate that the decision benefit of RAPM extends across independently trained prognostic models.