Bayesian Graph Wavelet Convolution Transformer-Based Transfer Learning for Fault Diagnosis of UAV Swarm System Under Variable Operating Conditions

Unmanned aerial vehicle (UAV) swarm system flies under variable operating conditions. Nevertheless, due to high costs or limited operational time, it is impractical to collect large amounts of faulted flight data under new operating conditions, which limits the fault diagnosis (FD) performances of data-driven FD methods for UAV swarm system. The distributions of massive historical flight data from existing operating conditions are different from those of the flight data from new operating conditions, which cannot directly be used for FD under new operating conditions. To handle this issue, a spatial-temporal Bayesian graph wavelet convolution Transformer (ST-BGWCT) based multisource deep transfer network (MDTN) is proposed to use the existing fault data to improve the FD performances of UAV swarm system under new operating conditions with limited fault data. Specifically, the ST-BGWCT can adaptively extract spatial-temporal coupled fault features from flight data in different domains. Based on a convolution Transformer, a multisource cross convolution Transformer (MCCT) and a weight-sharing triple-branch cross-domain aligning convolution Transformer (TCACT) are developed to fuse multi-source domain features and align source-target domain features in a segment-bysegment way, respectively. Finally, the MDTN is extended into the Bayesian deep learning (BDL) framework to quantify uncertainty. Extensive experiments on a semi-physical simulation platform of a fixed-wing UAV swarm system illustrate that the proposed approach achieves good FD results and outperforms existing methods