随着工业4.0时代的到来，工业网络面临着对实时性、可靠性、安全性和灵活性等方面更为严格的要求。然而，传统工业网络在动态配置、资源利用效率和管理复杂性方面存在明显不足。因此，软件定义网络（Software Defined Networking, SDN）应运而生，SDN通过将控制平面与数据平面解耦，显著提升了网络的灵活性、可扩展性以及可管理性。本文基于网络服务质量（Quality of Service, QoS）的三个核心指标，在SDN架构下对智能化路由策略进行研究，旨在优化QoS目标的同时提升整体网络性能。主要研究内容如下：

针对新兴网络业务快速发展带来的网络结构复杂化与服务质量要求多样化问题，提出了一种基于深度强化学习联合图神经网络的多智能体智能路由算法。采用深度Q网络构建策略模型，并引入消息传递神经网络对工业SDN网络的拓扑结构进行建模与特征提取，从而提升状态表示的准确性与模型的泛化能力。在此基础上，同时结合多智能体系统来处理有关QoS的多个优化目标，包括网络通信量、业务时延要求以及链路负载均衡。实验结果表明，所提算法相较于其他算法而言在平均网络通信量上提高了12.51%；在平均链路负载上提高了18.60%；在满足时延要求的平均业务流数量上提高了13.64%。同时，该算法能够高效地适应不同规模的网络环境，展现出较高的灵活性和兼容性。

由于业务时延要求、网络通信量、链路负载三个目标之间相互冲突，同时智能体间需要进行信息共享和协调，因此多智能体之间存在协作竞争关系。为提升多目标优化的适应性与整体性能，引入动态权重分配机制，可以根据网络实时状态动态调整各目标的权重，确保整体性能最优。实验结果表明，动态权重方法有效实现了多目标之间的整合，提升了网络的总体性能，保障了系统的稳定性和服务质量。

With the advent of Industry 4.0, industrial networks are facing increasingly stringent demands for real-time performance, reliability, security, and flexibility. However, traditional industrial networks suffer from limitations such as poor dynamic configurability, inefficient resource utilization, and high management complexity. To address these issues, Software-Defined Networking (SDN) has emerged as a promising solution by decoupling the control plane from the data plane, thereby significantly enhancing network flexibility, scalability, and manageability. This paper investigates intelligent routing strategies within the SDN framework, focusing on the optimization of three key Quality of Service (QoS) metrics to improve overall network performance. The main contributions are as follows:

To address the growing complexity of network structures and the increasing diversity of QoS requirements brought about by emerging network services, we propose a multi-agent intelligent routing algorithm that integrates deep reinforcement learning with graph neural networks. A Deep Q-Network is employed to construct the policy model, while a message-passing neural network is used to model the topology of industrial SDN networks and extract structural features, thereby improving the accuracy of state representations and enhancing model generalization. Furthermore, a multi-agent system is incorporated to jointly optimize multiple QoS objectives, including network traffic volume, service delay, and load balancing. Experimental results demonstrate that the proposed algorithm outperforms existing approaches by improving average network traffic volume by 12.51%, average link load by 18.60%, and the number of flows meeting delay requirements by 13.64%. In addition, the algorithm exhibits strong adaptability to networks of varying scales, reflecting high flexibility and compatibility.

Given the inherent conflicts among service delay, network traffic, and link load objectives, and the necessity for information sharing and coordination among agents, cooperative and competitive relationships naturally arise in the multi-agent system. To enhance adaptability in multi-objective optimization and improve overall network performance, a dynamic weight allocation mechanism is introduced. This mechanism adjusts the weights of each objective in real time based on the current network state, ensuring global performance optimization. Experimental results confirm that the proposed method effectively integrates multiple objectives, enhances overall network performance, and ensures system stability and service quality.