A Supervised Learning-Based Approach to Maneuver Detection Through TLE Data Mining

In the last decades, the Near-Earth environment has grown more and more crowded due to an exponential increase of launches and Space missions. As a direct consequence of that, fragmentations, impacts and debris have caused one of the most important issues of the commercial use of Space to rise: Space Traffic Management. In this framework, Space Surveillance and Tracking (SST) tries to address the problem by tracking and cataloguing Resident Space Objects through ground and space-based sensor networks. The collected data are processed and structured as catalogues in which new measurements are used to obtain orbit estimates. This kind of cataloguing activity implies the generation of a great deal of structured data. This setting is ideal to develop data-driven techniques and infer on the target and its pattern of life. The present work is a first step in this direction, specifically focusing on maneuver detection through Two-Line Element (TLE) time-series by means of a segmentation-aimed neural network. Starting with the extraction of meaningful orbital parameters from the TLE history of a target, the developed technique is capable of spotting maneuvering epochs across a given orbit history and even perform a first classification of general maneuver categories. This achievement is significant, since the available (and publicly accessible) data on maneuvering events are limited to few objects. Training with hybrid data and tests with real ones have been performed to asses accuracy and generalization capability. The results are promising, showing satisfactory performance on real data and proving how versatile this network architecture can be.