Abstract: In customized production, differences in process routes, machine procedure switching, and coupled resource allocation lead to dynamic operation-machine matching and make multi-objective scheduling difficult. To address this problem, this paper establishes a multi-objective scheduling model for customized production oriented to process matching and procedure switching, and proposes a direct scheduling decision-making method based on proximal policy optimization (PPO). The proposed method transforms the scheduling process into a Markov decision process. A multi-matrix state representation is constructed using job processing information and machine operation information, operation–machine combinations are used as action outputs, and an action mask mechanism is introduced to eliminate infeasible actions. An objective-oriented reward mechanism is further designed to guide policy learning. Experimental results show that the proposed method achieves better comprehensive scheduling performance than heuristic rules and metaheuristic algorithms on instances of different scales. It can effectively shorten the makespan, improve average machine utilization, and reduce the bottleneck machine load while maintaining high solution efficiency.