A machine learning method for the orbit state classification of large LEO constellation satellites

Abstract

Recently, large LEO constellation satellites have received much attention from the aspect of Space Domain Awareness (SDA) due to their significant number and relatively complex orbit states. Low-thrust propulsion system technologies are widely used during the lifetime of such satellites, making it difficult for the work of TLE (Two Line Element) based maneuver detection. This study proposes a machine learning method for orbit state classification based on TLE data, considering the concept of the Auto-Encoder (AE) network and Decision Tree (DT). In this manuscript, the orbit states were classified into four categories: non-maneuver, controlled orbit raising, controlled orbit lowering, and controlled orbit maintenance. First, historical TLE data of multiple satellites of the STARLINK constellation system and uncontrolled objects in a similar orbital area was analyzed using an Auto Encoder network to identify nonlinear features. Second, the extracted feature was sent to the DT network for the training of the supervised classification model, which could classify the input into four categories of orbit states. Finally, extra targets that had not been considered during the modeling process were used to test the applicability of the proposed method on satellites of the same constellation system. The results suggest that the method we presented here was applicable for orbit state classification with a macro precision of 93.8% and a macro F1-score of 93.7%, according to our numerical experiments. Once set up, the proposed method presented a feasible orbit state classifier for satellites of the same constellation system, operating at a similar altitude.