Risk-sensitive joint inventory-maintenance strategy for bearing health management under prognostic uncertainty: an uncertainty-aware deep reinforcement learning approach

Traditional separation of inventory management and Prognostics and Health Management (PHM) often leads to resource misallocation. While Deep Reinforcement Learning (DRL) offers a promising solution for joint decision-making, standard agents typically treat Prognostics and Health Management predictions as deterministic ground truths. However, in real-world scenarios, remaining useful life (RUL) predictions inherently contain stochastic errors. Ignoring this uncertainty leads to risk-blind policies that fail to buffer against sudden failures when prediction confidence is low. To address this, this paper proposes an Uncertainty-Aware collaborative adaptive inventory strategy. First, we introduce a Bayesian uncertainty quantification mechanism using Monte Carlo Dropout to estimate not only the RUL value but also its prediction variance. Second, to overcome the agent's myopic behavior, a novel Asymmetric Cost-Aware Reward Shaping mechanism is designed. By strategically decoupling the training and evaluation reward functions-specifically by introducing safety stock penalties and attenuating holding costs during training-the agent is guided to establish robust inventory buffers against supply chain uncertainties. Simulation results demonstrate that the proposed Risk-Sensitive PPO strategy significantly outperforms deterministic baselines, reducing total costs by 40.3% under high-noise environments.