Towards explainable anomaly detection for satellite telemetry

Abstract The increasing complexity of modern satellites and the growing amount of telemetry data available pose significant challenges for a safe and economic operation of satellites. To support the satellite engineers, traditional machine learning methods, including deep learning-based approaches, have shown promising results but lack intuitive explainability, hindering their adoption in operational settings. This paper presents a novel approach to anomaly detection and causal inference in satellite telemetry data, leveraging an ensemble of classical statistical models and deep learning architectures, combined with causal discovery techniques. We employ the Peter and Clark momentary conditional independence algorithm for identifying causal relationships with temporal dependencies and use its results as part of an in-depth root cause analysis enhancing anomaly detection. Our approach identifies potential anomalies and provides indications which satellite components cause the detected anomalies in order to facilitate interpretation by satellite operators. By integrating causal inference methods into anomaly detection pipelines, we aim to enhance explainability and facilitate decision-making in complex systems. This paper contributes to the growing body of work on anomaly detection and causal inference, highlighting the potential of combining machine learning and causation for improved operational performance.