A Collaborative Scheduling Approach for Sheet Metal Workshops in Printing Equipment Ovens Based on Graph Attention Reinforcement Learning

To address the collaborative optimization problem caused by the strong coupling between 2D sheet metal nesting and flexible shop floor scheduling in the sheet metal manufacturing process for color-separation ovens of satellite-type flexographic printing presses, this paper proposes a collaborative optimization method for nesting and scheduling based on a dual-flow graph attention network and proximal policy optimization. This problem involves inherent structural and resource symmetry in manufacturing operations and is further complicated by process precedence constraints, multi-resource competition, human–machine collaboration, assembly dependencies, and the dynamic coupling between nesting decisions and downstream production takt. Based on the composition of oven components and actual production methods, a collaborative optimization model integrating nesting and scheduling was constructed; a discrete-event simulation-driven joint scheduling environment was established to uniformly model the nesting, cutting, flexible machining, and assembly processes. On this basis, the collaborative nesting–scheduling process was formalized as a Markov decision process, clearly defining the state space, action space, reward function, and state transition mechanism. To enhance the state representation capabilities of the reinforcement learning agent in a high-dimensional discrete action space and under strongly constrained dynamic scheduling scenarios, this paper embeds a dual-stream graph attention network into the PPO framework to develop the DS-GAT-PPO algorithm. This algorithm simultaneously captures spatial nesting relationships among parts as well as temporal dynamic features such as equipment load, worker fatigue, and production takt time, thereby generating feasible and efficient joint scheduling plans. Experimental results show that the proposed method can reduce completion time by approximately 10.9–33.8% while maintaining a high sheet utilization rate; in a comparison of reinforcement learning algorithms, the proposed method achieved better completion times in most test cases and reduced the number of convergence steps by approximately 6.67–66.67%, validating its effectiveness in improving solution quality, convergence efficiency, and scheduling stability. These findings provide a foundation for further research and practical applications of collaborative nesting–scheduling optimization in dynamic manufacturing environments.