A Hypersonic Target Trajectory Prediction Method Based on EGNN and Transformer

Abstract

To address the issues of long-term dependency and insufficient local feature extraction in traditional methods when processing hypersonic target trajectory data, this article proposes an innovative trajectory prediction method that integrates equivariant graph neural networks (EGNNs) and Transformer architecture. Specifically, by constructing dynamic graph structures to model the geometric motion characteristics of the target, EGNN uses an equivariant message-passing mechanism to extract spatial features with SE (3) covariance. Meanwhile, the Transformer, with its multihead attention mechanism and geometric correction attention module, explicitly captures the long-term spatiotemporal dependencies in the trajectory data. To further enhance the model’s performance, an improved whale optimization algorithm (IWOA) is proposed, which dynamically regulates the learning rate using Lyapunov stability theory and combines Hamiltonian dynamics to reconstruct the predation strategy, significantly improving global search ability and convergence efficiency. Additionally, the AdamW optimizer is used to independently handle the weight decay term, effectively suppressing overfitting. The experimental results show that the proposed method achieves a position prediction root-mean-square error (RMSE) of 532.1 m and a velocity prediction RMSE of 268.3 m/s on the Northwestern Polytechnical University (NPU) trajectory dataset, improving accuracy by 23.8% and 38.8%, respectively, compared to the next-best method. Moreover, the model’s parameter count (2.75 M) and computational cost (5.68 GFLOPs) are significantly lower than those of the comparison models. Ablation experiments verify the effectiveness of the EGNN equivariant feature, IWOA dynamic optimization mechanism, and AdamW regularization strategy, providing a solution that balances both accuracy and efficiency for hypersonic target trajectory prediction.