Causal Explanation Enhanced Deep Reinforcement Learning for robust scheduling in degradation-prone flexible job shops

Within the flexible job shops characteristic of smart manufacturing and high-end equipment production, the robustness and long-term stability of scheduling systems have become core objectives alongside traditional production efficiency. However, in complex industrial scenarios, equipment failures are rarely purely exogenous stochastic events; rather, they are endogenous degradation processes driven by prolonged, high-load operation, manifesting as gradual performance deterioration and yield drops. Most existing Deep Reinforcement Learning (DRL) scheduling methods assume stochastic failures based on exogenous probability distributions, which biases agents toward greedy strategies to maximize short-term efficiency. This overdraws high-performance machines and weakens the conditions needed for preventive interventions. To address this issue and enhance decision credibility, this paper proposes a Causal Explanation Enhanced Deep Reinforcement Learning (CE-DRL) scheduling method. First, a scheduling environment encapsulating the physical degradation mechanisms is constructed to explicitly model the dynamic decay of operational health and the recovery process via opportunistic maintenance during idle periods. Second, to tackle the black-box nature of end-to-end DRL policies and provide logical justification for scheduling shifts, an embedded counterfactual necessity-attribution explainer is proposed to identify degradation-related machine features contributing to high-risk states. Finally, a causal consistency regularization mechanism aligns the policy focus with the counterfactual risk distribution, guiding the policy network to attend to machine health when degradation risk becomes critical. Extensive empirical evaluations demonstrate that, under load-induced degradation scenarios, CE-DRL encourages health-conscious load allocation and achieves competitive scheduling efficiency with a favorable efficiency-risk balance relative to the evaluated baselines.