摘要：

由于有限的机载资源和续航能力, 无人机(UAV)在空中停留时间有限, 无法长时间连续执行计算密集型任务。为了满足军事行动、紧急救援等连续作业场景中UAV的不间断任务执行需求, 设计一种基于无线能量传输的多UAV边缘计算方法。采用一组具备无线能量传输和移动边缘计算能力的大型无人机作为空中边缘能量服务器(AEES), 为多个空中执勤UAV提供能量传输和边缘计算服务, 以提高UAV的任务执行效率。通过联合UAV的三维位置、电量和任务量信息, 建立多UAV能量与算力联合优化模型, 并采用多智能体深度Q网络(MADQN)算法实现AEES服务位置点和能量发射功率智能化决策, 以最大化固定服务时长内的系统吞吐量和能量传输效率, 同时最小化能耗。仿真结果表明, 所提出的MADQN方法有效地优化了AEES的服务位置和能耗, 能够高效地为UAV提供算力、能量等资源。与启发式学习算法和贪婪算法等其他基线方法相比, 明显提升了能量传输效益和系统吞吐量, 同时保证了能量传输、能耗和吞吐量等多个优化目标的平衡。

关键词: 多无人机, 动态资源分配, 深度强化学习, 无线功率传输, 移动边缘计算

Abstract:

Due to limited on-board resources and endurance, UAVs have limited residence time in the air and cannot perform computation-intensive tasks continuously. To satisfy the continuous task execution requirements of UAVs in continuous operation scenarios, a multi-UAV edge computing system based on wireless energy transfer is designed. This system uses a group of large UAVs as an Air-Edge Energy Server(AEES) to provide energy and computing resources, ensuring the efficient execution of the missions of the UAVs. A joint optimization model is established for multi-UAV energy and computing by combining the 3D position, power, and task volume information of the UAVs to maximize the system throughput and energy transfer efficiency, while minimizing the energy consumption of the AEES. A Multi-Agent Deep Q-Network(MADQN) algorithm is used to determine the service location point and energy emission power for the AEES to achieve the optimization goals. Simulation results show that the proposed MADQN algorithm effectively optimizes the service location and energy consumption of the AEES and that it can efficiently provide computing, energy, and other resources for the UAVs. Compared with heuristic learning algorithms and other baseline methods, such as the greedy algorithm, the proposed MADQN significantly improves the energy transfer efficiency and system throughput and successfully balances multiple optimization objectives, including energy transfer, energy consumption, and throughput.

Key words: multiple Unmanned Aerial Vehicle(UAV), dynamic resource allocation, Deep Reinforcement Learning(DRL), wireless power transfer, Mobile Edge Computing(MEC)