Modélisation du comportement orbital nominal en utilisant l’apprentissage profond pour la détection de manoeuvres de satellites

Abstract

Accurate detection of satellite maneuvers, both in and out of the orbital plane, is an essential component of Space Situational Awareness (SSA). These maneuvers generally manifest themselves as anomalies in the time series of orbital parameters : semi-major axis and eccentricity for in-plane maneuvers, and inclination for out-of-plane maneuvers. This capability has become increasingly critical with the emergence of large satellite constellations, where accurate, realtime detection of maneuvers is essential to mitigate collision risks and prevent chain reaction scenarios. The diversity of propulsion systems used by satellites further complicates maneuver detection, requiring methods capable of adapting to a wide variety of thrust profiles and orbital behaviors. Although earlier work, such as that by Mukundan & Wang (2021), explored the use of orbital propagators combined with residual analysis to detect maneuvers, these methods often require satellite-specific ballistic coefficients and suffer from reduced accuracy due to unpredictable confounding variables such as solar activity. More recent studies have turned to the application of deep learning techniques to mitigate the impact of these perturbations, notably through physics-informed neural networks (PINNs) proposed by Varey et al. (2024), which integrate physical laws into the learning process. However, these approaches remain computationally resource-intensive and still have limited generalization capabilities. To address these challenges, we propose a new approach based on deep learning, consisting in training neural networks to predict the nominal, i.e. maneuver-free, behavior of space objects (Resident Space Objects, RSOs), from time series of debris Mean Orbital Elements (MOEs) extracted from the Special Perturbation (SP) database. The resulting predictions for satellites likely to maneuver can then be compared with actual measurements. Significant deviations, indicative of maneuvers, then appear as peaks in the residual time series, enabling unsupervised detection of anomalies characteristic of maneuvers, while maintaining a low false-positive rate. Two model variants are proposed for this application : MOE-Only models, which rely solely on past MOEs, and MOE+SW (Space Weather) models, which additionally incorporate exogenous space weather data, available for past and future time steps, to refine predictions. Experiments with several state-of-the-art time series prediction models demonstrate that the proposed methods enable both accurate and efficient orbit prediction, without the need for metadata (such as drag or solar radiation coefficients) nor intermediate integrations. These results highlight their strong potential for scalable maneuver detection applications with minimal prior information.

Date	15 Sept 2025
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Georges Ghazi (Supervisor) & Bruno Revelin (Co-supervisor)