A Stability-Driven Clustering Scheme for Uninterrupted Federated Learning in UAV Networks

Abstract

Federated Learning (FL) has been widely studied as a distributed learning paradigm for Unmanned Aerial Vehicle (UAV) networks because it can reduce communication overhead while preserving data privacy. However, high UAV mobility, heterogeneous residual energy, and dynamic communication link quality may cause the Cluster Head (CH) to exhaust its battery prematurely or be frequently replaced, thereby compromising the continuity of FL training rounds. To address these limitations, this paper proposes a stability-driven clustering scheme to support uninterrupted FL in UAV networks. The proposed scheme consists of two core components. The first is a CH selection algorithm that normalizes four metrics, namely residual energy, computational availability, connectivity, and mobility stability, and integrates them into a single stability score. The second is a history-window-based monitoring mechanism for CH and Cluster Member (CM) management. Furthermore, a pre-selected Vice-Head (VH), which serves as a backup CH, is combined with localized cluster split and merge operations. This design enables the proposed scheme to respond to dynamic events, such as CH failure and cluster-size variations, without interrupting FL aggregation. Comparative experiments were conducted on the FashionMNIST dataset under a non-independent and identically distributed (non-IID) partition setting. Compared with the bestperforming baseline in each metric, the proposed scheme reduced communication distance variance by approximately 41% and extended network lifetime by approximately 18%. In addition, the proposed scheme sustained training beyond the termination rounds of all baseline methods, yielding an improvement of approximately 6.7% in the FL accuracy. These results demonstrate that the proposed scheme effectively maintains cluster stability and FL training continuity.