Maneuver detection and duration estimation of non-cooperative spacecraft impulsive maneuvers based on hybrid learning

Abstract

Traditional satellite impulsive maneuver detection methods suffer from low accuracy, unreliable time estimation, dependence on manual parameters, and poor noise robustness. This paper proposes a hybrid neural network model to detect impulsive maneuvers and pinpoint the start and end of the maneuver window. First, the model takes ground-based radar observation time-series data and their first-order differences as input; residual analysis and first-order differencing preprocessing are applied to mitigate gross errors and random noise. Then, a CNN-LSTM-SA hybrid neural network is constructed: one-dimensional convolution extracts local abrupt features from orbital sequences, a long short-term memory (LSTM) network captures temporal evolution patterns, and self-attention mechanisms enhance feature weights during critical maneuver periods. Finally, a dual-output head architecture enables simultaneous determination of maneuver events and effective localization of time windows. Simulation experiments verify that the proposed method achieves an accuracy of 95.22%, an F1 score of 0.94, and a time window intersection over union (IoU) of 0.6673 in impulsive maneuver detection. It achieves zero false positives while maintaining a high recall rate, outperforming standalone CNN and LSTM models as well as existing dual-structure models overall.