中国机械工程 ›› 2026, Vol. 37 ›› Issue (10): 2511-2519.DOI: 10.3969/j.issn.1004-132X.2026.10.014

• 系统能量管理与回收利用 • 上一篇    下一篇

基于双注意力网络深度强化学习的柔性作业车间节能调度方法

曾丹1,2, 杨武1,2, 杜彦斌1,2(), 吕岩1,2, 李聪波3   

  1. 1.重庆工商大学智能装备绿色设计与制造重庆市重点实验室, 重庆, 400067
    2.重庆工商大学机械工程学院, 重庆, 400067
    3.重庆大学机械与运载工程学院, 重庆, 400044
  • 收稿日期:2025-11-18 出版日期:2026-10-25 发布日期:2026-11-10
  • 通讯作者: 杜彦斌
  • 作者简介:曾丹,女,1993年生,博士、讲师。研究方向为生命周期评价、绿色智能制造
    杜彦斌*(通信作者),男,1982年生,教授、博士。研究方向为绿色制造与再制造。E-mail: duzi2009@163.com。
  • 基金资助:
    重庆市自然科学基金(CSTB2022NSCQ-MSX1337);重庆工商大学高层次引进人才科研启动费(2256014)

Energy-efficient Flexible Job Shop Scheduling Method via Dual Attention Network Based Deep Reinforcement Learning

ZENG Dan1,2, YANG Wu1,2, DU Yanbin1,2(), LYU Yan1,2, LI Congbo3   

  1. 1.Chongqing Key Laboratory of Green Design and Manufacturing of Intelligent Equipment,Chongqing Technology and Business University,Chongqing,400067
    2.College of Mechanical Engineering,Chongqing Technology and Business University,Chongqing,400067
    3.College of Mechanical and Vehicle Engineering,Chongqing University,Chongqing,400044
  • Received:2025-11-18 Online:2026-10-25 Published:2026-11-10
  • Contact: DU Yanbin

摘要:

针对柔性作业车间节能调度研究中工序-机器复杂关系建模不足,算法效率与解质量难以兼顾,冗余信息干扰决策等问题,以最小化最大完工时间与机器总能耗为优化目标,提出一种基于双注意力网络深度强化学习的调度优化方法。首先,构建双目标调度问题数学模型并转化为马尔可夫决策过程;设计包含工序消息Agent与机器消息Agent的双注意力网络,用于分别建模工序优先级约束和机器竞争关系与加工资格,精准捕捉工序-机器复杂关联关系并滤除冗余调度信息。其次,在深度强化学习框架下将工序选择与机器分配整合为联合决策问题,通过双目标协同优化机制生成工序-机器对的优先级概率分布,平衡运行效率与解质量。最后,在合成数据集与Mk1~Mk7基准数据集上开展对比实验,结果表明所提方法综合性能优势显著,可有效平衡解质量与运行效率,兼具实用性与优化性能。

关键词: 柔性作业车间节能调度问题, 深度强化学习, 双注意力网络, 多目标优化

Abstract:

To address the issues in existing research on the EEFJSP, including inadequate modeling of complex operation-machine relationships, the persistent trade-off between algorithm efficiency and solution quality, and decision-making interference from redundant information, this paper proposed a scheduling optimization method based on deep reinforcement learning (DRL) with a dual attention network (DAN), targeting the minimization of makespan and total machine energy consumption. First, the bi-objective scheduling problem was mathematically formulated and further transformed into a Markov decision process, upon which a DAN consisting of an operation message agent and a machine message attention agent was developed to separately characterize operation precedence constraints, machine competition relationships and processing eligibility, enabling accurate capture of intricate operation-machine correlations while eliminating redundant scheduling information. Subsequently, within a DRL framework, operation selection and machine assignment were integrated into a joint decision-making problem, where a bi-objective collaborative optimization mechanism generated the priority probability distribution of operation-machine pairs to achieve a well-balanced trade-off between operation efficiency and solution quality. Finally, comparative experiments on synthetic datasets and the Mk1-Mk7 public benchmark datasets verified that the proposed method achieved prominent comprehensive performance, effectively reconciling solution quality with operation efficiency and delivering both strong practicability and superior optimization capability.

Key words: energy-efficient flexible job shop scheduling problem(EEFJSP), deep reinforcement learning, dual attention network, multi-objective optimization

中图分类号: