ATHMoS: Automated Telemetry Health Monitoring System at GSOC using Outlier Detection and Supervised Machine Learning

Abstract

Knowing which telemetry parameters are behaving accordingly and those which are behaving out of the ordinary is vital information for continued mission success. For a large amount of different parameters, it is not possible to monitor all of them manually. One of the simplest methods of monitoring the behavior of telemetry is the Out Of Limit (OOL) check, which monitors whether a value exceeds its upper or lower limit. A fundamental problem occurs when a telemetry parameter is showing signs of abnormal behavior; yet, the values are not extreme enough for the OOL-check to detect the problem. By the time the OOL threshold is reached, it could be too late for the operators to react. To solve this problem, the Automated Telemetry Health Monitoring System (ATHMoS) is in development at the German Space Operation Center (GSOC). At the heart of the framework is a novel algorithm for statistical outlier detection which makes use of the so-called Intrinsic Dimensionality (ID) of a data set. Using an ID measure as the core data mining technique allows us to not only run ATHMoS on a parameter by parameter basis, but also monitor and flag anomalies for multi-parameter interactions. By aggregating past telemetry data and employing these techniques, ATHMoS employs a supervised machine learning approach to construct three databases: Historic Nominal data, Recent Nominal data and past Anomaly data. Once new telemetry is received, the algorithm makes a distinction between nominal behaviour and new potentially dangerous behaviour; the latter of which is then flagged to mission engineers. ATHMoS continually learns to distinguish between new nominal behavior and true anomaly events throughout the mission lifetime. To this end, we present an overview of the algorithms ATHMoS uses as well an example where we successfully detected both previously unknown, and known anomalies for an ongoing mission at GSOC.