Transferable Deep Learning for Multivariate Spacecraft Telemetry Anomaly Detection

Abstract

As spacecraft complexity and satellite deployment rates increase, automated anomaly detection is critical to ensuring mission safety. A majority of traditional model designs and evaluations are tested on datasets that are augmented or not representative of operational realities in spaceflight. This paper addresses this through its utilization of the European Space Agency Anomaly Detection Benchmark (ESA-ADB), a dataset characterized by sub-2% anomaly density, irregular sampling, and intentional telecommand-driven state changes. Time-series telemetry data is spatially encoded using 2D Gramian Angular Fields (GAFs) and CNN architectures are evaluated through a comparison of a shallow model versus deep residual networks (ResNet), assessed for the ability to classify anomalous and nominal telemetry segments. Stacked multichannel GAFs outperform single-channel approaches by capturing subsystem co-activation patterns, doubling the F0.5 score on both datasets they are tested on. While spatial encoding models achieve near-perfect event-wise recall, precision remains fundamentally bounded by the misclassification of telecommand driven rare nominal events.