Anomaly Detector

ML Anomaly Detector for Unknown Fault Modes

The ML Anomaly Detector for Unknown Fault Modes is a monitoring system implemented on radiation-hardened NOR flash and a dedicated machine learning accelerator. Drawing 2 W of power, it uses unsupervised autoencoders and isolation forests to identify anomalous system behaviors that do not match predefined fault signatures. Over extended lunar missions where uncharacterized fault modes inevitably emerge, the unit continuously learns new baseline behaviors across a 100-year operational lifetime. Operating with a 500,000-hour mean time between failures, it typically flags 50 anomalies per year for investigation, committing records to an anomaly archive while accepting authenticated updates.

ML anomaly detector identifying anomalous system behaviour not matching any known fault — flagging unknown fault modes for investigation.