Graph Neural Network Assisted Spectrum Resource Optimisation for UAV Swarm

Unmanned aerial vehicles (UAVs) serving as aerial base stations have attracted enormous attention in dense cellular network, disaster relief, sixth generation mobile networks, etc. However, the efficiency is obstructed by scarce spectrum resources, especially in massive UAV swarms. This paper investigates a graph neural network-based spectrum resource optimisation algorithm to formulate the channel access and transmit power of UAVs with the consideration of both spectrum efficiency (SE) and energy efficiency (EE). We first construct a domain knowledge graph of UAV swarm (KG-UAVs) to manage the multi-source heterogeneous information and transform the multi-objective optimisation problem into a knowledge graph completion problem. Then a novel attribute fusion graph attention transformer network (AFGATrN) is proposed to complete the missing part in KG-UAVS, which consists of an attribute aware relational graph attention network encoder and a transformer based channel and power prediction decoder. Extensive simulation on both public and domain datasets demonstrates that, the proposed AFGATrN with a rapid convergence speed not only attains more practical spectrum resource allocation scheme with partial channel distribution information (CDI), but also significantly outperforms the other five existing algorithms in terms of the computation time and the trade-off between the SE and EE performance of the UAVs.