Cluster Head Detection for Hierarchical UAV Swarm With Graph Self-supervised Learning

In this paper, we study the cluster head detection problem of a hierarchical unmanned aerial vehicle (UAV) swarm network (USNET), consisting of multiple double-level clusters, from the perspective of the observers who are assumed not to know any internal information on the USNET. Specifically, we propose a multi-cluster graph attention self-supervised learning (MC-GASSL) framework for the observers. The MC-GASSL comprises two main algorithms, including the sequential metric learning (SML) algorithm for detecting the clusters in the USNET, and the graph attention self-supervised learning (GASSL) algorithm for detecting the cluster head in each cluster. The basic idea of the SML is to learn an embedding space where the UAVs can be easily clustered based on Euclidean distances, and the GASSL works by calculating the attention weights between UAVs and finding the attention center of the cluster. Extensive numerical results validate the effectiveness of the MC-GASSL. Moreover, the ablation study shows that the GASSL can detect the cluster head in single clusters with over 98% detection rate on average. The ablation study also shows that the clustering purity of the USNET with the SML exceeds that with traditional clustering algorithms by at least 10% on average.