ChanCrossFormer: A Ballistic Trajectory Prediction Model Integrating Channel Attention and Cross-Attention Mechanisms

Recent studies have shown that there is a certain degree of dynamic correlation among different categories of state information in ballistic trajectory data. Many transformer based models excessively fuses these different categories of state information, resulting in a decrease in the predictive performance of the model. While some new proposed channel independent strategies totally sever the correlation between these state information. To address this issue, this paper proposes a prediction model named ChanCrossFormer that integrates the channel attention and cross-attention mechanisms. The model adopts a dualstream architecture design: First, value embedding and position encoding are performed on different state information respectively, and the computational complexity is reduced through feature separation. Second, a channel attention module is introduced after the self-attention layer of the predictive variables to achieve dynamic weight allocation and selective enhancement of key feature channels. Finally, a deep interaction between different state information is established through the cross-attention mechanism to fully explore the potential association patterns among the features. Experimental results show that on the test set, the RMSE and MAE of this model are 20.5 % and 38.2 % lower than those of DLinear respectively. Ablation experiments further verify the effectiveness of the cross-attention and channel attention mechanisms, providing a new paradigm for ballistic trajectory prediction in complex scenarios.