Orbital motion intention recognition for space non-cooperative targets based on incomplete time series data

Abstract

This study establishes a method for recognizing the intentions of non-cooperative targets in orbital dynamics using incomplete time series data. By leveraging a relative orbital dynamics model, 38 distinct motion intentions are delineated, representing the largest variety identified in current research. A neural network-based intention recognition method is developed, with a comprehensive comparative analysis focusing on network architecture and types of time series data. The network architectures explored include long short-term memory networks, gated recurrent unit networks, and self-attention structures, while the time series data encompass position, angular, and distance measurements. Experimental findings reveal that the combination of gated recurrent unit networks with attention mechanisms achieves the highest performance in intention recognition, reaching a maximum accuracy of 95%. Both position and angular measurement time series demonstrate exceptional performance, with recognition accuracies exceeding 93%, utilizing the least amount of information in current studies. The proposed method maintains accuracies of 89% and 94% under position measurement noise and nonlinear orbital dynamics disturbances, respectively, compared to only 17% and 43% for methods based on orbital dynamics condition matching. Additionally, an analysis of intention observability is conducted to enhance model interpretability. This approach holds significant promise for applications such as spacecraft collision avoidance and non-cooperative space target capture.