Innovative ML-based Methods for Automated On-board Spacecraft Anomaly Detection

Spacecraft health monitoring is an important task to assure the mission operational life. For this purpose, a variety of telemetry data are analyzed to detect anomalies that can lead to failures and cause irreversible damage to on-board devices. In this paper we propose and analyze different ML-based methods that contribute to the generation of an intelligent anomaly detector capable of identifying anomalies in spacecraft telemetries, with a particular attention to the memory footprint of each method. Finally, we investigated how to model the abnormal behaviors during the validation/test phase, exploiting different families of possible configurations for anomaly injection. The achieved results, after several tuning setups, suggest that all of the adopted methods are suitable for implementation when observing relatively short intervals of time. Instead, when longer time windows are used, the score-based detectors outperform the proximity/density-based ones, despite being more resource hungry.