Spacecraft Fault Detection Identification & Causation Using Machine Learning

Machine learning-based Fault Detection, Identification and Recovery (FDIR) software will enable autonomous responses to faults onboard a spacecraft. By removing the human in the loop, anomalies can be addressed quickly, avoiding communication delays and long periods spent in “safe� mode during which the spacecraft cannot perform its mission. There are, however, several hurdles to machine learning-based FDIR, including the diversity of spacecraft and components that require algorithm tuning, the lack of labeled telemetry data from historical missions for training purposes, and the limited amount of telemetry data that includes faults (since anomalies are by definition rare). Verus Research proposes to supplement available satellite telemetry with simulated data to overcome these challenges. We will use unsupervised learning for anomaly detection on real data, and train supervised learning algorithms to classify faults using simulated data. Once deployed, the FDIR software will first detect anomalies and then identify their source. Unidentified anomalies will be studied by operators, labeled, and added to the training data to improve the algorithm, which is re-uploaded to the spacecraft. The entire SBIR effort will culminate in a software deliverable that can be tuned to specific spacecraft and is deployed onboard for use on orbit.