Using Machine Learning to Automatically Detect Anomalous Spacecraft Behavior from Telemetry Data

This paper demonstrates three important results: (1) the ability to accurately classify anomalous spacecraft behavior using only traditional machine-learning algorithms, (2) the ability to successfully deploy a machine-learning model trained on one specific component of a satellite in the service of detecting anomalous behavior in all other instances of that particular component on the bus, and (3) the ability of a machine-learning engine to detect impending failure of a satellite component without having been trained on examples of such failures. Automated anomaly classification is shown using telemetry data from the Suomi National Polar-orbiting Partnership (NPP) weather satellite. Both supervised and semi-supervised approaches yield exceptionally high accuracy (i.e. approximately 99%) in detecting whether a given 120-minute window of telemetry data from an instrument's scan drive mechanism contains an anomalous event. The most successful supervised algorithm was found to be a random-forest classifier; in the case of the semi-supervised approach, a Gaussian elliptic envelope drawn in feature space performed best. An exhaustive feature-selection study revealed that the most successful quantities for performing classification could be derived from statistical moments of the 120-minute telemetry windows under consideration. Applying trained models to previously unseen components and anticipating component failures are both demonstrated using telemetry streams from the reaction wheels aboard the Kepler space telescope. Two of the four reaction wheels suffered failures during that satellite's mission. Since only two examples of failure were available, an unsupervised method (i.e. a Gaussian elliptic envelope with non-negligible contamination) was used to learn typical reaction-wheel behavior. The model was trained using only a small fraction of the data from a few months of telemetry from only one of the four reaction wheels. When applied to the telemetry streams from the other three wheels, this model - in conjunction with a time-based clustering algorithm - was able to successfully warn of elevated friction levels three to six months prior to failure.