Self-supervised contrast learning based UAV fault detection and interpretation with spatial-temporal information of multivariate flight data

Precise fault detection and interpretation can effectively enhance the safety of unmanned aerial vehicles (UAV) flight missions. However, sufficient fault data covering all fault modes is generally inaccessible, which poses a formidable challenge to the traditional supervised learning strategy. In this study, a novel UAV fault detection approach based on self-supervised contrast learning and spatial-temporal information of multivariate flight data is proposed. In the contrast learning task, a series of specific sample transformations are first designed, and the feature distribution of normal data can be modeled in self-supervised manner through comparing the similarity of different transformed samples. In above process, an auxiliary classification task that distinguishes different sample transformations is further introduced to facilitate the learning of critical features. In order to extract comprehensive spatial-temporal information from multivariate flight data, a multi-channel spatial-temporal encoder is designed in which two independent graph multi-head attention neural networks (GMAT) are implemented to mine the temporal features and multivariate spatial features, respectively. The extracted spatial-temporal features are then fused with the designed locally-enhanced token fusion module and the powerful multi-headed self-attention module. Finally, the occurrence of faults can be detected by comparing the reconstruction error with the fault threshold. With the box-plot analysis, the flight variables whose reconstruction errors are far from the overall distribution will be regarded as the possible fault sources to implement fault interpretation. Experimental results on the self-developed fixed-wing UAVs demonstrated the prominent performance of our method on fault detection and interpretation.