To meet the demands of on-orbit monitoring and space situational awareness, this
study proposes a Transformer-based relative motion pattern recognition method
based on the self-attention mechanism, addressing the issues of difficulty in
accurately characterizing the relative motion patterns of non-cooperative space
targets under short-arc observation conditions, strong dependence on prior
knowledge in traditional methods, and insufficient stability under noise
interference. Firstly, this study constructes a non-cooperative spacecraft
motion pattern action data set based on the Clohessy–Wiltshire (C-W) relative
motion equation. By designing typical initial conditions of relative motion,
trajectory samples including fixed-point following, hovering following,
oscillating following, coplanar circling, non-coplanar circling, approaching and
receding, etc. of multiple motion patterns are generated, and the relative
position and relative velocity sequences in the local orbital coordinate system
are used as model inputs. Subsequently, an motion pattern recognition network
composed of an input representation module, a Transformer encoding module, and a
classification output module is constructed. The trajectory sequence
representation is achieved through linear embedding and position encoding, and
the multi-head self-attention mechanism is used to capture the global
dependencies and key dynamic features between different time steps, thereby
completing the discrimination of the target motion pattern category. Finally,
model training and testing are carried out based on the constructed data set,
and experimental analysis is conducted from three aspects: training convergence,
classification performance in the absence of noise, and robustness in the
presence of noisy observations. The results show that the proposed method can
effectively learn the discriminative features in short arc segment relative
motion, and performs well in terms of accuracy, Macro-F1, and confusion matrix;
in the presence of Gaussian noise interference, it outperforms the Multi-Layer
Perceptron (MLP) baseline model while still maintaining clear category
boundaries and high recognition stability.