Dynamic Production Scheduling Method for Pyrotechnic Product Lines Based on FS Simulation Data and Deep Reinforcement Learning

To address the difficulty of balancing schedule feasibility, dynamic disturbance response, and execution stability for pyrotechnic product lines under coupled hard constraints from equipment, labor, logistics, and safety regulations, a dynamic scheduling method based on FS simulation data and deep reinforcement learning is proposed. An FS discrete-event simulation is used to construct the training and validation environment, and repeatable dynamic samples are generated through controlled disturbance injection. Operations, machines, labor posts, and safety spaces are represented as a heterogeneous scheduling graph; state features are extracted by a graph neural network; the action space of the PPO policy is constrained by action masking and feasibility verification; and event-triggered rolling rescheduling is employed for fast local updates. Simulation validation indicates that executable schedules can be generated without violating stringent safety constraints. Compared with single dispatching rules, metaheuristic methods, and a basic learning-based baseline, the main advantage lies in the integrated trade-off among plan stability, urgent-order response, and online rescheduling quality after dynamic disturbances. The results demonstrate that the proposed method provides feasible intelligent decision support for strongly constrained dynamic scheduling of pyrotechnic product lines.