With the widespread use of UAVs in civil fields, their safety has gradually received more attention. Fault diagnosis and reconstruction technology are essential to ensure UAVs’ safe and stable operation. The UAV system has a complex structure, and the flight parameters collected are characterized by a wide variety and high dimensionality. Furthermore, there are multiple flight phases of the UAV performing the mission, which further adds to the complexity of the spatio-temporal related information of the flight parameters. Existing UAV fault diagnosis and reconstruction methods are more limited in handling spatio-temporal information and ignore the inherent dynamic characteristics of multi-stage flight data. This paper proposes a new multi-stage dynamic spatio-temporal information fault diagnosis and reconfiguration prediction network (MDSTDR). First, the graph is constructed using a dynamic adaptive adjacency matrix that considers the spatial correlation of complex multivariate features. Then, the spatio-temporal network is built by a graph convolutional network and a time-domain feature extraction module that considers long-term temporal features. The accuracy of fault diagnosis and predictive reconstruction is further improved by using a feedback training loop to focus model attention on samples with large reconstruction residuals. Finally, the performance of the proposed model is verified using actual flight data of the fixed-wing UAV. The experimental results indicate that MDSTDR can effectively diagnose and reconstruct multi-stage data faults, and its performance is better than the existing state-of-the-art (SOTA) method.
Fan et al. (Thu,) studied this question.