Robust Spatial-Temporal Autoencoder for Unsupervised Anomaly Detection of Unmanned Aerial Vehicle With Flight Data
Recently, the safety and reliability of unmanned aerial vehicles (UAVs) have gained increasing interest, and data-driven anomaly detection methods have been widely studied. However, it is still challenging to build an accurate and reliable detector due to the scarcity of labeled data and complex spatial-temporal characteristics of flight data with noises and disturbances. To this end, this article proposes an autoencoder-based unsupervised anomaly detection framework with multivariate flight data without labeled information. Specifically, we designed a new robust spatial-temporal autoencoder (RSTAE) model based on the temporal convolution network (TCN), the convolution neural network (CNN), and the attention mechanism to extract the complicated spatial-temporal correlations in multivariate flight data. Instead of using the traditional mean square error (mse) loss function, a modified loss function based on maximum correntropy criteria (MCC) is introduced to enhance the robustness of our RSTAE model during its training process. To further improve the anomaly detection performance, a dynamic threshold strategy is used. Experimental results on real flight data demonstrate the superior performance of the proposed method compared with several autoencoder-based methods in terms of six evaluation metrics.