Learning Long-Range Relationships for Temporal Aircraft Anomaly Detection

Time-series classification for anomaly detection in calibrating aircraft sensors is crucial to ensuring aviation security. Nevertheless, the lengthy temporal span of sensor data causes difficulties in extracting global information dependence and the limited number of samples can easily cause model overfitting. To tackle these problems, we propose SMDA-Net, the stratified multiscale representation learning network with automatic data augmentation for long-range time-series modeling. Specifically, we design a stratified structure to extract multiscale characteristics of time-series, wherein we develop an encoder with an efficient self-attention block for ultra-long sequences. Meanwhile, we present a scheme via learning to weight the contribution of the augmented samples to the loss for automatic data augmentation to improve the generalization ability of our model. Extensive experiments indicate that our model exhibits high performance on Flights dataset and exceeds state-of-the-art methods on 18 long-range time-series datasets. Moreover, we verify the effectiveness of our method through ablation study and visualization analysis.