Transformer-based anomaly detection in P-LEO constellations: A dynamic graph approach.

Successful management of large space systems, such as the most recent Proliferated Low Earth Orbit (P-LEO) constellations, demands an increased level of autonomy and irregular behavior detection capability. Existing techniques for satellite network management rely on monitoring satellites individually, potentially failing to capture events shared across multiple units. In this work, we propose a pipeline for identifying anomalous connections in proliferated satellite networks. Our pipeline relies on representing a satellite constellation as a dynamic graph. Dynamic graphs can be conveniently exploited to represent P-LEO networks, as they allow to capture meaningful structural and temporal correlations. Such spatial and temporal information are first projected into an embedding space; next, this encoded representation is processed by a transformer-encoder network for the anomaly detection task. We conduct extensive analyses of the main problem parameters, including the temporal horizon, the structure of the graph, and the size of the constellation. To assess the general performance of the method, we perform a Monte Carlo analysis on 960 ground node pair connections, providing empirical evidence of the algorithm's generalization capability. The obtained results encourage the extension of the method to more complex, realistic modeling and kindred applications such as the on-edge detection problem. • Topology changes of P-LEO satellite networks are explored to detect anomalous links. • Spatial–temporal features of the network are leveraged to boost performance. • Evidence of algorithm generalization capability is collected via randomized simulations. [ABSTRACT FROM AUTHOR]