Cislunar Orbit Determination using CNN-based Constrained Admissible Region

This paper proposes a novel approach that integrates a convolutional neural network (CNN) with the constrained admissible region (CAR) method to improve orbit determination and tracklet association in cislunar space. Then the orbit family from which the observation value may originate is deduced by the presence of CNN-CAR, and the classification information is utilized as a prior to sequentially associate the track. The obtained CNN-CAR is integrated into the Initial Value Problem (IVP) optimization framework for associating tracklets. Simulation results demonstrate the effectiveness of the proposed method. The CNN-CAR-IVP algorithm exhibits exceptional performance in tracklet association, with both true positive (TP) and true negative (TN) rates exceeding 95%. Initial orbit determination achieves position accuracy around 10 kilometers. The findings have significant implications for space surveillance, collision risk assessment, and long-term catalog maintenance in the increasingly congested cislunar space environment.