@article{Yin2026, 
author = {Zhuoqing Yin and Yue Wang and Weiwei Cai and Rui Zhong},
title = {Relative motion recognition of space non-cooperative objects via angle-only measurements and deep neural networks},
year = {2026},
journal = {Astrodynamics},
keywords = {space non-cooperative object, relative motion recognition, trajectory geometric, deep neural network},
url = {https://www.sciopen.com/article/10.1007/s42064-026-0334-4},
doi = {10.1007/s42064-026-0334-4},
abstract = {This study presents a method for recognizing relative orbital motion of space non-cooperative objects using angle-only sequence measurements. A six-class taxonomy of relative motion is established for short-term observation and early warning decision-making scenarios based on orbital dynamical characteristics. A hierarchical classification strategy is first developed based on geometric properties of trajectories propagated under a two-body orbital dynamics model, enabling automatic generation of physically consistent motion labels. The influence of observation window length, geometric thresholds, and two-body versus Clohessy–Wiltshire (CW) modeling differences is systematically analyzed. A BiGRU-attention network is constructed, with hyperparameters optimized through comparative training experiments. Comparative experiments with recurrent neural network (RNN), long short-term memory network (LSTM), and bidirectional gated recurrent unit with self attention (BiGRU-SA) models are also conducted to validate the effectiveness of the proposed architecture. Simulation results show that the proposed model achieves the best performance under a 1-h observation window, reaching a recognition accuracy of 90.52%, demonstrating strong capability for short-term angle-based motion discrimination. The proposed method maintains robust performance under measurement noise and systematic sensor bias, achieving accuracies of 86.2%, and under J2 orbital perturbations, achieving an accuracy of 89.41%. It also achieves reliable generalization under ±200 km orbital radius variations, with accuracy remaining above 76%. This method provides a robust and intelligent solution for monitoring and threat assessment of space non-cooperative objects. Furthermore, the framework enables maneuver detection and intention update within approximately one observation window (about 1 h), supporting near-real-time space situational awareness.}
}