Adaptive Federated Learning for Anomaly Detection in Satellite Telemetry

This paper presents a streamlined Federated Learning (FL) framework for anomaly detection in satellite telemetry, addressing limitations of centralized approaches for predictive maintenance in resource-constrained satellite networks. Evaluating FL models on the ESA-ADB dataset, the optimized LSTM with FedAvg + Fine-Tuning achieved an F0.5 score of 0.86, a precision of 0.93, and an AUC of 0.85, outperforming centralized models, which achieved a maximum F0.5 score of 0.63 and an AUC of 0.83. Additionally, FL significantly reduced communication costs, requiring only 1.8MB per round compared to the high overhead of centralized data transmission. Scalability analysis demonstrated stable performance up to 10 clients, with an F0.5 score of 0.87 and recall of 1.00. These findings validate FL as a practical, privacy-preserving, and scalable solution for onboard satellite anomaly detection.