A hierarchical neural network for multi-target cooperative intention recognition under single-satellite observation

To address the limitations of existing spacecraft intention recognition methods, which typically assume a fixed number of targets and restrict intention types to simple motion modes, this paper proposes a hierarchical recognition network for multi-spacecraft cooperative intentions under single-satellite observation. An intention framework comprising multiple mission types is first established, in which participating roles and the number of participants are explicitly defined, allowing the number of spacecraft involved in an intention to vary dynamically. Considering the practical constraints of observation target switching in single-satellite observation scenarios, a sparse measurement information dataset generation method is then developed. Based on this dataset, complex cooperative intentions are decomposed into orbital characteristics associated with individual spacecraft roles, and a Temporal Convolutional Network–Dynamic Gated Cross Network (TCN–DGCN) model is proposed. In the first layer, hierarchical TCN blocks are stacked to extract orbital features of multiple spacecraft from continuous observation data and to identify their corresponding role types. In the second layer, dynamic gating and feature crossover mechanisms are introduced to process the logits output from the first layer through parallel information flows, enabling the learning of role-combination features and the recognition of overall cooperative intention types. Simulation results demonstrate that the proposed model outperforms benchmark models, achieving up to an 8% improvement in recognition accuracy, while reducing the minimum test loss to approximately 50% of that of the benchmark methods. Further simulation analyses indicate that the proposed approach maintains superior performance in complex scenarios involving a large number of participating spacecraft.