A Machine Learning Approach to Modeling Satellite Behavior

To date, there are roughly 1,957 active satellites and 17,494 inactive and debris objects that are currently tracked orbiting the globe. Each satellite's orbit varies by shape, size, and orientation depending on the satellites distinct function. Due to their diverse orbits and quantity, it is not feasible for analysts to simultaneously monitor all satellites; however, machine learning approaches allow for the automation of satellite characterization for space situational awareness (SSA). This project presents preliminary evidence of how machine learning applications can be used to model satellite behavior. Satellite data were modeled using the machine learning approaches of supervised classification, unsupervised clustering, and supervised neural networks. Supervised models were used to determine satellite stability as a critical feature for satellite characterization. Similarly, unsupervised techniques provided insight into satellite characterization by geospatial location and detection of anomalous behavior and maneuvers. Deep learning, using a recurrent neural network (RNN), predicted sequential satellite maneuvers over time. Together, these satellite modeling results have implications for providing surveillance and security in space.