LSTM-based maneuver detection for resident space object catalog maintenance

Abstract Maintaining a catalog of Resident Space Objects involves a wide range of activities, exploiting measurement processing to estimate and update orbits, in an effort to achieve a comprehensive picture of the Near-Earth environment status in terms of human-made objects. The first step required to update a cataloged orbit is to correctly associate it with new observations. This task is particularly hampered by the growing presence of active satellites that are capable of maneuvering. Nonetheless, a large part of them perform routine maneuvers to preserve their orbit, defining very regular patterns. Given the significant quantity of past orbital and maneuvering data about tracked objects as the main by-product of catalog maintenance, the main focus of this work is to effectively exploit them by training Machine Learning models. The objective is to infer the probability of a target maneuver within a specific time horizon, using a Long Short-Term Memory Recurrent Neural Network, specialized in processing time sequences. Two distinct maneuver detection modules are developed and tested on real LEO object data to, respectively, extend a target’s control history and predict how likely a maneuver will happen in the immediate future.