Hybrid Framework for Automated Monitoring of Constellation-Scale Structures in LEO Mega-Constellations

To address the technical challenge of dynamically identifying shells and orbital planes in low Earth orbit (LEO) mega-constellations, this paper proposes a hybrid unsupervised-supervised framework and validates it on the Starlink constellation. The framework first employs Density-Based Spatial Clustering of Applications with Noise (DBSCAN) for robust shell partitioning and an adaptive right ascension of the ascending node (RAAN) phase-jump detection algorithm for initial orbital plane segmentation. To resolve complex scenarios where this unsupervised partitioning encounters ambiguity, a one-dimensional convolutional neural network (1D CNN) is introduced as a refinement module, automatically learning physical couplings among orbital elements for precise classification. Experiments demonstrate that this zero-prior, zero-parameter framework accurately reconstructs the orbital topology, with CNN refinement achieving a classification accuracy of 94.49% on densely spaced adjacent planes, outperforming multiple baseline models. The proposed approach establishes a scalable, automated paradigm for real-time monitoring and anomaly detection of LEO mega-constellation satellite networks, bridging orbital mechanics and non-terrestrial networks.