关键词: 无人机自组织网络, 韧性重构, 图注意力网络, 近端策略优化, 多智能体

Abstract: This paper investigates the resilient reconfiguration problem of flying ad hoc networks (FANETs) after topology damage. When external disturbances cause partial node failures, the network is required to reconstruct connectivity over the remaining active nodes under limited resources and restore its information-carrying capability as much as possible. In this paper, the network capacity under multi-hop transmission is adopted as the primary metric for network performance, which is defined as the sum of end-to-end effective capacities over all node pairs with bottleneck links and hop-count penalty taken into account. Based on this metric, the network topology, node geometric positions, and link communication resources are jointly optimized to improve the reconfigured network capacity while controlling reconstruction time and energy consumption. First, UAV states and resource budgets are treated as node features, while link conditions are treated as edge features and embedded into a graph attention network to characterize the network topology. Second, a hierarchical optimization framework is proposed, where the upper layer adopts a centralized proximal policy optimization method to generate an adjacency matrix for topology reconstruction, while the lower layer employs a multi-agent proximal policy optimization method to perform distributed displacement adjustment under a given topology. Finally, by maximizing the minimum link capacity as a link-level fairness surrogate objective, node transmit power and bandwidth are allocated to improve network capacity while controlling reconstruction time and cost. Simulation results demonstrate that, compared with other schemes, the proposed method can significantly restore the network capacity after topology damage while maintaining relatively low reconstruction time and energy consumption, thereby exhibiting favorable resilient reconfiguration performance.

Key words: flying ad hoc network, resilient reconfiguration, graph attention network, proximal policy optimization, multi-agent systems