Fault diagnosis of actuators in UAV swarm based on personalized federated graph networks

An unmanned aerial vehicle (UAV) swarm consists of collaborative drones working collectively to achieve specific objectives. As a complex engineering system, it requires stable multi-drone coordination, reliability and robust information security to address failures and external disturbances. Actuator fault diagnosis is a critical component of UAV flight control systems ensuring reliable mission execution. For UAV flight control systems, actuator faults such as motor malfunctions and blade damage require precise diagnosis to maintain operational reliability. However, actuator fault diagnosis in UAV swarm remains challenging owing to the heterogeneous and privacy-sensitive nature of multi-source data as well as the scarcity of fault samples. Aiming to solve the aforementioned problems, a personalized federated learning (FL) framework based on graph attention neural networks (P-FedGAT) has been proposed. P-FedGAT effectively protects data privacy through the FL approach. Firstly, to overcome the challenge of limited samples, P-FedGAT utilizes GATs to extract multi-sensor features on each client. Additionally, a novel hypernetwork is employed to dynamically generate personalized adjacency matrices while FedProx regularization coefficients are utilized to balance client model personalization and global model generalization. Furthermore, a knowledge distillation strategy with a customized distillation loss function is introduced to fetch knowledge from the global model to local models effectively, thereby mitigating model heterogeneity. Comparative analysis on public UAVFD and RflyMAD datasets show that the proposed P-FedGAT outperforms mainstream methods such as FedAvg, FedProx, FedGCN, FedPer, and FedTP improving fault diagnosis accuracy by 1.12%–20.86% (p  <  0.05) and 2.26%–13.15% (p  <  0.05), with statistical significance verified via t-test. The proposed P-FedGAT demonstrates optimal performance across a range of client numbers (from 6–20) and under highly non-independent and identically distributed data distributions.