Streaming Analytics for Enhancing Anomaly Detection in Satellite State-of-Health Telemetry
Justin D. Li, Kevin Yarritu, Michael Bridges, Michael Shannon Eydenberg, Mayuri Shakamuri · 2024
Anomalous behavior can pose serious risks in the operation of complex, high-consequence systems. Detection is complicated and challenging, especially with high-dimensional data and varying data types, such as with satellite state-of-health (SOH) telemetry. Existing detection methods tend to perform poorly in such situations or suffer from satellite constraints on computational power and data availability. Current operational approaches rely on manual, retrospective analysis of system failures, which can result in lengthy response times, risking degraded capabilities and delays in satellite operations. We leverage streaming machine learning capabilities to develop a performant and scalable method for anomaly detection in multi-modal satellite SOH telemetry data. By building on a k-means clustering approach, our approach demonstrates anomaly detection capabilities that (1) update their joint, multivariate estimation of the current SOH variables in an online and continuous fashion, (2) operate in near real-time with reduced computational and data requirements, and (3) provide meaningful detections with an interpretable feature space mapping to relevant variables to support further operator diagnosis and response. Using real satellite SOH data as a case study, we demonstrate how our method can provide automated and adaptive diagnostic information, which can increase robustness in space operations through more rapid detection of satellite anomalies.