Adversarially-trained Graph Convolutional Recurrent Autoencoder for Spacecraft Anomaly Detection with Missing Data

Telemetry data anomaly detection has a pivotal role in spacecraft health management. It can provide great support for operation engineers and guarantee the safe operation of satellites. And the complex inter-coupling among sensors and the missing data condition are the two most significant features of satellite telemetry data. We need a powerful method that not only can efficiently and comprehensively capture the required information, but also needs to be able to adapt to missing data for anomaly detection. We propose an Adversarially-trained Graph Convolutional Recurrent Autoencoder (AGCR-AE), which can be well applied to spacecraft anomaly detection with missing data. A graph-based recurrent module is equipped to our AGCR-AE to capture both sensor correlations and time correlations. Meanwhile, an adversarially-trained autoencoder mechanism is used to reconstruct data, thus we can detect anomalies with missing data. Experiments reveal that AGCR-AE outperforms baselines and has outstanding performance.