Explaining Satellite Anomalies--Causal Inference for Space Operations

Abstract

The operation of satellites relies heavily on telemetry data, which has become increasingly complex due to the proliferation of parameters. Automatic anomaly detection and explanation are crucial for satellite operators to respond promptly to anomalies and ensure the reliability of their systems. This study aims to bring together classical machine learning methods and deep learning approaches in anomaly detection, with a focus on causal inference. For anomaly detection, we investigate the performance of the deep learning methods Graph-Augmented Normalising Flow (GANF) and Multi-Scale Temporal Variational Autoencoder (MST-VAE), as well as of the classical, density-based estimation Maximally Divergent Intervals (MDI) method. For causal inference, we apply two time series causal discovery algorithms, Peter and Clark Momentary Conditional Independence (PCMCI) and Joint Peter and Clark Momentary Conditional Independence (J-PCMCI), to identify causal relationships in the considered satellite telemetry data. Our methods are designed to provide explainable results and facilitate interpretation of the anomalies by satellite operators. We evaluate our approach using a use case study on satellite telemetry data collected during ground station contacts, incorporating telecommands given. This research contributes to the growing body of work on anomaly detection and causal inference in complex data sets, and advance our understanding of anomaly detection and causal inference.