Enhanced Graph Deviation Networks for Anomaly Detection in Space Telemetry

In this paper we propose an enhanced Graph Deviation Network (GDN) architecture tailored for the analysis of time series telemetry data. While GDNs have worked well in industrial systems, spacecraft systems pose unique challenges that demand specialized adaptations. Our solution improves on the original GDN by using the sensors relationship representation, taking on a vector-based edge representations, and making use of several node embeddings in parallel to the original GDN layer to depict several sensor relationships. We demonstrate through the Mars Science Laboratory Anomaly Detection dataset (MSL) how these changes enables a better modeling of spacecraft-specific patterns and sensor relationships and with presenting an interpretable representation of sensors clustering. The results show improved anomaly detection performance while making it possible to interpret the learned sensor relationships.