Graph Transformer-based Deep Reinforcement Learning for Fault Diagnosis of UAV Swarm System Under Imbalanced Data

UAV Swarm Systems are in a healthy state most of the time and the occurrence time of severe faults is very short, leading to the imbalance of historical faulted flight data, which poses an intense challenge for data-driven fault diagnosis (FD). To overcome these problems, a graph Transformer-based deep reinforcement learning (GTDRL) is proposed for imbalanced FD of UAV Swarm Systems under imbalanced data. In the method, an FD agent is designed to perform the diagnostic action on one sample in each time step, and an imbalanced data environment is built to evaluate the diagnostic action and return an imbalanced reward to the agent. A new reward policy is designed to ease the imbalanced data problem by considering the interclass deviation level according to various sample sizes among different fault classes. In addition, a graph transformer is proposed as the deep Q network (DQN) to capture fault features of UAV swarm systems. Finally, experimental results validate the efficacy and superior diagnostic performance of the developed FD method based on a semi-physical simulation platform of a fixed-wing UAV swarm system.