Data of Multi-agent Deep Reinforcement Learning Predictive Maintenance Strategy for Large-scale Production Line Clusters: Distributed Game Theory, Communication Limitations, and Elastic Collaboration

在超大规模智能工厂中，通信限制和任务耦合导致集中式维护失败，迫切需要一种基于多智能体深度强化学习的分布式协同预测性维护策略。文提出了一种基于分布式博弈推理和自适应协同机制的多智能体深度强化学习预测性维护策略：首先，建立基于设备拓扑和故障关联的异构网络环境;其次，利用图注意力网络（Graph Attention Network，GAT）高效表示通信约束下的局部状态;然后，本文在分布式纳什逼近博弈框架（DNAGF）中进行策略博弈和均衡更新。本文通过分层多智能体深度确定性策略梯度（H-MADDPG）进一步实现弹性协同训练。最终生成基于设备健康预测的分布式维护调度方案，实现自主、高效、稳健的预测性维护决策。实验表明，本文提出的策略实现了0.87的劳动力利用率和0.89的能源利用率，与传统方法相比有显著提升。在95%的高负载下，其响应时间保持较低，智能体协作的最高共识为92%。此时，策略波动仅为 0.05，证实了其高效协同和稳健性。In ultra-large-scale intelligent factories, communication constraints and task coupling lead to the failure of centralized maintenance, creating an urgent need for distributed collaborative predictive maintenance strategies based on multi-agent deep reinforcement learning. This paper proposes a multi-agent deep reinforcement learning-based predictive maintenance strategy integrating distributed game reasoning and adaptive collaborative mechanism: First, a heterogeneous network environment based on equipment topology and fault correlation is established; Second, the Graph Attention Network (GAT) is employed to efficiently represent local states under communication constraints; Then, strategy game and equilibrium update are conducted within the Distributed Nash Approximation Game Framework (DNAGF). This paper further realizes flexible collaborative training through the Hierarchical Multi-Agent Deep Deterministic Policy Gradient (H-MADDPG) algorithm. Finally, a distributed maintenance scheduling scheme based on equipment health prediction is generated, enabling autonomous, efficient and robust predictive maintenance decision-making. Experimental results show that the proposed strategy achieves a labor utilization rate of 0.87 and an energy utilization rate of 0.89, with significant improvements over traditional methods. Under 95% high load, its response time remains relatively low, and the highest consensus rate of agent collaboration reaches 92%. At this point, the policy fluctuation is only 0.05, which confirms its efficient collaboration and robustness.