TLE Prediction using Machine Learning for Satellite Maneuver Detection

The increasing number of spacecraft in orbit requires trajectory predictions that are timely, efficient, and highly accurate. Traditional methods, such as numerical propagators [1], are becoming less suitable due to their high computational demands and dependence on metadata, or parameters such as drag and radiation coefficients, which are difficult to predict. These coefficients, typically derived from recent trajectory data, require extensive computations and are subject to variability over time as they account effects of unmodeled phenomena. This variability compromises the reliability of predictions and poses challenges for predicting the trajectories of newly identified space objects, such as unknown satellites or debris resulting from collisions. Data-driven approaches, such as deep learning, provide a powerful alternative by capturing physical patterns without being disrupted by subtle local and global variations and without relying on explicit metadata. These methods enable efficient and robust predictions of spacecraft behavior, making them particularly well-suited for tasks such as orbit forecasting and maneuver detection. As the number of objects in orbit continues to increase, these capabilities become increasingly critical for ensuring effective and reliable space operations. This study applies a deep learning model to analyze Two-Line Elements (TLE) gathered from the SpaceTrack catalogue, with the goal of predicting the trajectories of the space objects. The focus is on Low Earth Orbit (LEO) satellites, as this region is not only the most congested but also the most susceptible to orbit variations, primarily due to residual atmospheric drag. Technical challenges arise from the diverse propulsion systems used by satellites in this region, including not exhaustively chemical and electric thrusters, but also solar wind or no propulsion at all. Additionally, TLE data is derived from various captors, and while data fusion and filtering techniques are applied, a certain level of noise remains inherent in the dataset. The trajectories are represented as a series of time-dependent points that describe the shape and position of an object’s orbit in the Earth-Centered Inertial (ECI) orbital frame, using the Keplerian orbital parameters. We use a deep learning model derived from the TSMixer architecture [2], which uses lightweight linear mixing operations across time and feature dimensions to efficiently leverage cross-variate information while reducing computational costs. This approach relies exclusively on time series of orbital parameters, removing the need for physical equations or metadata, such as drag or solar radiation coefficients. Furthermore, the model adapts dynamically by learning behaviours specific to distinct orbital regimes, enhancing its versatility and accuracy. Our results highlight the potential of data-driven techniques to outperform traditional methods in terms of computational efficiency and predictive accuracy. These approaches provide a scalable and robust solution for managing modern satellite operations and enabling Space Situational Awareness (SSA) at an industrial scale in the increasingly crowded and dynamic orbital environment. [1] Mukundan, A., & Wang, H.-C. (2021b). “Simplified approach to detect satellite maneuvers using TLE data and simplified perturbation model utilizing orbital element variation,” Applied Sciences, 11(21), 10181. https://doi.org/10.3390/app112110181 [2] Ekambaram, V., Jati, A., Nguyen, N., Sinthong, P., & Kalagnanam, J. (2023b). “TSMiXer: Lightweight MLP-Mixer model for multivariate Time series forecasting,” Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 459–469. https://doi.org/10.1145/3580305.3599533