Complex-product manufacturing workshops are characterized by diverse process routes, heterogeneous machine capabilities, tight coupling between processing and assembly, and stringent parent–child kitting constraints. These characteristics make fixed-priority rules insufficient for coordinating job release, machine competition, and assembly waiting. To address scheduling in a production process that connects front-end processing with back-end fixed-position assembly, this study proposes an attention-based proximal policy optimization method. First, the manufacturing process is formulated as a staged model comprising a front-end hybrid flow-shop processing stage and a back-end fixed-position assembly stage. The model captures operation precedence, machine heterogeneity, stage transitions, and kitting constraints. Next, a reinforcement-learning scheduling framework is established by defining the state space, action space, and dynamic action mask, thereby incorporating operation sequences, machine eligibility, resource occupancy, and assembly release constraints into sequential decision making. Furthermore, an A-PPO policy that combines local operation attention with machine-competition attention is designed to select feasible job–equipment-unit matching actions. An industrial engineering case shows that A-PPO achieves a makespan of 3823.7, outperforming six fixed-priority rules (Rule 2–Rule 7) and a genetic algorithm (GA) baseline. GA achieves a makespan of 3875.3, whereas the best rule-based methods achieve 3986.4. Compared with GA, A-PPO reduces the makespan by 1.33%; compared with Rule 6 and Rule 7, it reduces the makespan by 4.08%; and compared with the average of the six rules, it reduces the makespan by 18.95%. The results demonstrate that the proposed method shortens order completion cycles and supports intelligent scheduling in complex-product manufacturing workshops. The proposed method is currently applicable to phase-oriented complex-product manufacturing workshops with structured process routes, equipment capabilities, and parent–child assembly constraints; its transferability across layouts and enterprises requires further validation.
Wang et al. (Thu,) studied this question.