Missing signal reconstruction and aileron fault detection via generative adversarial learning

Aileron faults must be detected and identified quickly to improve the fixed-wing aircraft's safety and reliability. However, the low-frequency monitoring signals of the heterogeneous sensors are easily lost in transmission and sampling. Therefore, incomplete data will overlap fault features, leading to detection mistakes. This study proposes a generative adversarial learning framework to reconstruct missing signals based on the rest signals called Dual-hierarchical-variational Graph Autoencoder/ Generative Adversarial Network (DHVGAE/GAN). A pluggable reconstruction module is designed to enhance the generalization performance of the detection model. A dual spatiotemporal encoder (DSE) is constructed by a local graph attention network (LGAT) and Transformer-based feature extractors (TFEs) to extract common spatiotemporal features of missing and rest signals. A shared hierarchical variational autoencoder (HVAE) maps them to multiple conditionally independent latent distributions and generates reconstructed signals. The historical prior information of missing signals is transferred to the reconstruction module through adversarial learning. The detection model maintains high accuracy for data with missing signals relying on the recovered signals. The experimental results of the pre-launch and flight datasets verify the feasibility and effectiveness of the proposed method. © 2024 Elsevier B.V.