Anomaly Detection in Spacecraft Telemetry Data using Graph Convolution Networks

Telemetry data anomaly detection is of great significance to guarantee the safe operation of spacecraft. However, the high dimensionality of telemetry variables and the strong correlation between variables pose a great challenge to multivariate anomaly detection. This paper proposes a Graph Convolution Network (GCN)-based anomaly detection method for telemetry data. In this method, GCN is proposed to extract correlation features between variables and learn an updateable correlation map. Then, Convolution Neural Network (CNN) extracts temporal information and combines correlation features for attention to obtain multivariate prediction results. In addition, a novel method of calculating single variable anomaly score and integrated anomaly score is proposed to locate anomalous variables. Finally, experiments on a real dataset are conducted, by which the proposed GCN-based approach is demonstrated to be effective and accurate in telemetry data anomaly detection. In addition, experiments also show our method preforms well in locating anomalous, providing interpretable for anomaly detection results.