Matteo Rosa, Politecnico di Milano; Gaetano Calabrò, Politecnico di Milano; Pierluigi Di Lizia, Politecnico di Milano
https://doi.org/10.64861/QFPA9012
Keywords: Satellite maneuver detection; Space Domain Awareness (SDA); Space Situational Awareness (SSA); Transformer neural networks; Attention mechanisms; Two-Line Elements (TLE); Low Earth Orbit (LEO); Supervised learning; Orbital dynamics; Data-driven modeling.
Abstract:
The rapid growth of the Low Earth Orbit (LEO) environment has significantly increased the operational complexity of Space Situational Awareness (SSA) and Space Domain Awareness (SDA), particularly with respect to maneuver detection, catalog maintenance, and conjunction assessment. Within this context, reliable detection of satellite maneuvers using publicly available Two-Line Element (TLE) data remains a challenging problem. Accurate and timely identification of the maneuvers is essential to maintain consistent orbital predictions and mitigate collision risk. Recurrent neural networks, particularly Long Short-Term Memory (LSTM) architectures, have shown promising results in detecting maneuvers when trained and tested on the same satellite, but their performance often worsens when applied to previously unseen spacecraft data, limiting their operational scalability. This work focuses on two main research objectives. First, it aims to obtain reliable detection of maneuvers with reconstruction capabilities, going beyond the binary classification of events. Reconstruction means understanding the duration of the maneuver and the jump that the parameter makes due to the maneuver, since the label assigned to maneuvers is related to time. Second, it aims to generalize maneuver detection across multiple satellites, so that a model trained on a subset of spacecraft maintains its performance when applied to previously unseen satellites without the need for new training specific to each satellite. Addressing both objectives is essential for scalable SSA applications, particularly in the context of newly launched satellites or spacecraft that do not have historical maneuver data. The proposed approach is based on a supervised transformer architecture applied to time series derived from TLE data. The study initially focuses on LEO satellites and considers semi-major axis and inclination as input parameters. These parameters are chosen since they reflect both in-plane and out-of-plane orbital maneuvers. The neural network architecture consists of a transformer encoder incorporating multi-head self-attention followed by a decoder that produces time-dependent maneuver predictions. Unlike LSTM-based models, which process sequences sequentially and may struggle to preserve long-range contextual dependencies, the transformer directly models global temporal relationships through self-attention. This mechanism allows each time step within a window to be evaluated in relation to all others, facilitating improved modeling of extended maneuver signatures. A principal methodological contribution of this work is the introduction of a gradient maneuver mask. Traditional approaches typically rely on binary one-hot labeling, classifying time steps strictly as maneuver or non-maneuver. Such representation provides limited structural information considering maneuvers as isolated events. The proposed gradient mask assigns a high value to time steps within the maneuver interval and gradually decays to zero before and after the maneuver. This continuous representation embeds contextual and temporal structure into the training data, helping the model learn maneuver evolution rather than only event occurrence. Using this formulation, the learning objective is no longer simply to detect the maneuver event but aims to reconstruct the maneuver mask in terms of duration and jump of orbital parameters. The model is evaluated under two different scenarios. In the first scenario, both training and testing are performed on the same set of satellites, allowing the assessment of the performance of the model on spacecraft for which historical maneuver patterns have been observed during training. In the second scenario, the model is trained on a subset of satellites and subsequently tested on previously unseen spacecraft in order to evaluate its generalization capability across different satellite. For maneuver detection, a maneuver is considered correctly detected if the predicted maneuver interval overlaps any portion of the corresponding ground truth interval, without requiring an exact temporal match. Maneuver reconstruction is evaluated using the Intersection over Union (IoU) metric between the predicted and ground truth maneuver masks. A maneuver is considered successfully reconstructed if the IoU exceeds a threshold of 0.5. Reconstruction performance is then computed as the percentage of correctly reconstructed maneuvers with respect to the maneuvers that were correctly detected. Results demonstrate strong detection capability in both evaluation settings. The proposed model achieves comparable performance when evaluated on satellites included in the training set and on previously unseen spacecraft, demonstrating its ability to generalize across different satellite populations. In both evaluation scenarios, the model achieves maneuver detection F1-scores of approximately 0.95 without requiring retraining for individual satellites. These results indicate improved robustness and generalization relative to the state of the art. Regarding maneuver reconstruction, the proposed approach successfully reconstructs more than 90% of the correctly detected maneuvers, demonstrating its ability not only to identify maneuver occurrences but also to accurately recover their temporal extent. These results suggest that the proposed transformer-based architecture provides a scalable and operationally relevant solution for automated maneuver monitoring, directly supporting SSA and SDA applications involving newly launched satellites and expanding LEO populations.
Date of Conference: September 15-18, 2026
Track: AI for SDA Applications