Machine Learning-based Stability Assessment and Change Detection for Geosynchronous Satellites

Phan Dao, AFRL Space Vehicles Directorate; Kristen Weasenforth, Applied Optimization; Jeff Hollon, Applied Optimization; Tamara Payne, Applied Optimization Inc.; Kimberly Kinateder, Wright State University; Adam Kruchten, University of Pittsburgh

Keywords: Machine learning, Bayesian changepoint, stabilized GEO, automation, photometry

Abstract:

Analysts have been able to inspect the light curve of a geosynchronous satellite to assess its stability if it’s three-axis-stabilized or periodicity if it’s tumbling. However, with the large volume of data collected persistently with wide field of view sensors, manual inspection by humans may not be sustainable. It’s desirable to automate the stability assessment to (a) classify the satellite as stable or unstable and when possible (b) pinpoint the moment it transitions from being stable to being unstable. In this paper, we will show how such as an automated system is developed. We found the Random Forest (RF) of Decision Trees classifier to be sufficiently robust and accurate as a solution for (a). For each light curve, multiple features are used by RF. A feature is usually a single value calculated from the light curve. A discussion of the optimal features to use with RF will be provided. Stability assessment accuracy better than 97% has been attained in our test. Once the RF algorithm has detected the first night the satellite becomes unstable, we then try to find the precise time of the change. If the transition occurs during observation, it’s likely that the signature shows a change. A Bayesian change point detection algorithm is used to pinpoint the change. It’s not a coincidence that the entire process of classifying light curves and finding change points is based on two Machine Learning algorithms.  In observation-based science, as shown in this example, the objectives are often to associate measurements with an unknown state or to infer a change in the hidden parameters of a model. Those objectives happen to be among the most common problems pursued by the Machine Learning researchers.

Date of Conference: September 11-14, 2018

Track: Poster

View Paper