With the development of the internet of things (IoT), the concept of an edge network has been gradually expanding to other fields including internet of vehicles, mobile communication networks and smart grids. Because the resources of terminals are limited, the long-distance movements of users will increase the running costs of the services that are offloaded to edge servers, and even the services on terminals will stop running. Another problem is that resource shortages or hardware failures of these edge networks can affect the service migration policy. In this paper, a novel service migration method based on state adaptation and deep reinforcement learning is proposed to efficiently overcome network failures. Before migration, we define four edge network states to discuss the migration policy and adopt the two-dimensional movement around the edge servers to adapt to the applications scenarios of our work. Then, we use the satisfiability modulo theory (SMT) method to solve the candidate space of migration policies based on cost constraints, delay constraints and available resource capacity constraints to shorten the interruption time. Finally, the service migration problem can be transformed into the optimal destination server and low-cost migration path problem based on the Markov decision process by the deep Q-learning (DQN) algorithm. Moreover, we theoretically prove the rate of convergence in the learning rate function of our algorithm to improve the convergence rate. Our experimental results demonstrate that our proposed service migration mechanism can effectively shorten the delays from service interruptions, and better avoid the impact of edge network failure on the migration results and, thus, improve the users’ satisfaction.

随着物联网（IoT）的发展，边缘网络的概念已逐渐扩展到其他领域，包括车辆互联网，移动通信网络和智能电网。由于终端资源有限，用户的长途移动将增加卸载到边缘服务器的服务的运行成本，甚至终端上的服务也将停止运行。另一个问题是这些边缘网络的资源短缺或硬件故障会影响服务迁移策略。为了有效克服网络故障，提出了一种基于状态自适应和深度强化学习的服务迁移新方法。在迁移之前，我们定义了四个边缘网络状态来讨论迁移策略，并采用围绕边缘服务器的二维移动以适应我们工作的应用场景。然后，我们基于成本约束，时延约束和可用资源容量约束，使用可满足性模理论（SMT）方法求解迁移策略的候选空间，以缩短中断时间。最后，基于深度Q学习（DQN）算法，可以基于马尔可夫决策过程将服务迁移问题转化为最优目标服务器和低成本迁移路径问题。此外，我们从理论上证明了算法学习速率函数中的收敛速度，从而提高了收敛速度。