Multiscale Spatial-Temporal Bayesian Graph Conv-Transformer Based Distributed Fault Diagnosis for UAVs Swarm System

The fault diagnosis can improve the reliability and safety of unmanned aerial vehicles (UAVs) swarm systems. However, due to the fault propagation between UAVs, the fault features possess high nonlinearity and spatial-temporal coupled characteristics that are hard to be learned by data-driven fault diagnosis methods. Moreover, uncertainties induce untrustworthy diagnosis results. To solve the issues, a multiscale spatial-temporal Bayesian graph conv-transformer (MST-BGCT) is proposed for distributed fault diagnosis of UAVs swarm system. The MST-BGCT has three primary characteristics: 1) a spatial features extractor with graph attention network to both locally and globally mine spatial correlations among neighboring UAVs; 2) a convolutional Transformer-based temporal features extractor can further capture multiscale temporal-related fault features; and 3) these feature extractors are extended into Bayesian deep learning (BDL) framework to quantify uncertainty. The effectiveness and advantages of the proposed approach are illustrated by comparative experiments on a semi-physical simulation platform of fixed-wing UAVs swarm system under multiple situations of colored measurement noises.