工作车间
计算机科学
强化学习
作业车间调度
调度(生产过程)
图形
流水车间调度
人工智能
工业工程
作业调度程序
数学优化
运筹学
分布式计算
机器学习
作者
Chenfei Liu,Yuan Tao,Wen Song,Zhuoyi Lin,Yaoxin Wu,Shujun Chen,Liang Gao,Haiyan Yin
标识
DOI:10.1016/j.aei.2026.105274
摘要
In human-centered flexible manufacturing systems under Industry 5.0, some human operations can be performed through cooperation among multiple workers. The selected worker number affects both processing time and total labor input, so scheduling involves three coupled decision dimensions: operation sequencing, resource assignment, and worker number selection. This study investigates a multi-objective flexible job shop scheduling problem with multi-worker cooperation, processing times that vary with the selected worker number, and transportation time. A mathematical model is formulated to minimize makespan and total labor input while describing the diminishing returns from multi-worker cooperation. To solve this problem, a preference-conditioned heterogeneous graph reinforcement learning method is proposed. The scheduling process is formulated as a Markov decision process, and a heterogeneous graph is used to encode operation precedence, resource states, station worker capacity, and transportation information. Based on this graph representation, a hierarchical policy network is designed with an operation-resource pair selection actor and a worker number selection actor to handle the additional worker number decision for human operations. Furthermore, the objective preference weight is embedded in both the state representation and the reward function, enabling a single model to generate multi-objective schedules under different objective preferences. Experiments on 130 instances constructed from a welding and assembly workshop show that the proposed method has advantages over ten classic and recent algorithms across multiple evaluation metrics. Preference interpolation and ablation analyses show that the preference input and hierarchical policy design help improve solution set quality. An industrial order case study further demonstrates that the method can generate executable schedules and has the potential to make fast decisions after disturbances.
科研通智能强力驱动
Strongly Powered by AbleSci AI