Case Study on Detecting Anomalies in CubeSat Telemetry Using Machine Learning Approach

Abstract

The increasing complexity of CubeSat missions and the volume of telemetry data they generate has heightened the need for advanced anomaly detection systems capable of identifying faults before they escalate into mission failures. Traditional threshold-based monitoring approaches fall short in capturing subtle, multivariate, and time-dependent anomalies inherent in satellite telemetry. This case study explores the application of machine learning (ML) techniques for anomaly detection in CubeSat telemetry, with a focus on evaluating recent research supported by publicly released benchmark datasets: OPSSAT‑AD and ESA‑ADB. These datasets provide real-world, labelled telemetry from operational ESA missions and have enabled systematic benchmarking of over 30 machine learning models. The study synthesises model performance across metrics such as F1-score and AUC, highlighting that temporal models like LSTMs and temporal convolutional networks (TCNs) consistently outperform classical methods in time-series tasks. Limitations in dataset continuity, anomaly sparsity, and generalisability are discussed, along with the need for explainable and resource-efficient models suitable for onboard deployment. The case concludes with a call for expanded benchmark datasets, real-time validation, and cross-disciplinary collaboration to ensure robust, interpretable, and mission-ready anomaly detection systems for future CubeSat applications.