Adaptive Resource Scheduling Framework Based on Dynamic Priority Assignment for Advanced Manufacturing Systems
Abstract
Advanced manufacturing systems require scheduling methods that respond to changing task arrivals, deadline pressures, queue congestion, and resource availability. This study proposes an Adaptive Dynamic Priority Assignment (Adaptive-DPA) framework in which a proximal policy optimization (PPO) policy adjusts four scheduling weights while task selection remains governed by an explicit Priority Index (PI). The framework was evaluated in sequential, parallel, flexible, and cell-based production environments under workload targets of 80, 120, and 160 completed operations using static dispatching rules and a direct-dispatch PPO scheduler as comparators. The results show environment-dependent behavior: fixed-weight PI remains competitive in sequential and parallel production, whereas Adaptive-DPA provides clearer descriptive benefits in flexible and cell-based production. Paired analyses across the five evaluation seeds did not show statistically significant differences after Holm correction; therefore, the performance results were interpreted descriptively, rather than as statistical evidence of universal superiority. A controlled reward-ablation experiment further shows that the effects of local counterfactual and pressure-alignment contributions depend on the production environment and workload and that the complete reward is not uniformly preferable. The computational-overhead analysis shows that Adaptive-DPA mean dispatch-decision latency ranges from 1.844 to 3.873 ms, which is higher than that of the static rules and Direct-PPO but remains within the millisecond range in the evaluated simulation setting. These findings characterize the performance, interpretability, and computational trade-offs of learning-based weight adaptation, while preserving an explicit and traceable priority calculation.