基于Bahdanau注意力机制的大口径火炮双向GRU轨迹预测  

Bidirectional GRU trajectory prediction for large-caliber artillery based on bahdanau attention mechanism

作　　者：彭晨洋 陈龙淼[1] 张鸣洋 PENG Chenyang;CHEN Longmiao;ZHANG Mingyang(School of Mechanical Engineering,Nanjing University of Science and Technology,Nanjing 210094,China)

机构地区：[1]南京理工大学机械工程学院,南京210094

出　　处：《兵器装备工程学报》2024年第7期56-64,共9页Journal of Ordnance Equipment Engineering

基　　金：国家自然科学基金项目(U2141246)。

摘　　要：针对大口径火炮弹丸轨迹预测问题,提出一种基于Bahdanau注意力机制与双向门控循环单元网络的序列到序列轨迹预测模型(S2S ATT-BiGRU)。通过不同条件下六自由度弹丸运动模型仿真,得到大量弹丸轨迹数据样本,并利用滑动窗口法和差分法构造仿真数据集;将所提轨迹预测模型与GRU模型在仿真数据集下进行实验。研究结果表明,S2S ATT-BiGRU模型在输入时间为0.5 s预测未来时间为5 s的情况下,其MSE为13.89 m^(2),MAE为1.84 m, MAPE为0.29%,其预测精度远高于GRU模型。在其余情况下S2S ATT-BiGRU模型的预测结果均优于GRU模型。这表明S2S ATT-BiGRU模型具有更强的存储输入轨迹序列信息和自适应关注重要输入轨迹信息的能力,为弹丸轨迹预测研究提供有利参考。In response to the trajectory prediction problem of large-caliber artillery projectiles,a sequence-to-sequence trajectory prediction model(S2S ATT-BiGRU)based on the Bahdanau attention mechanism and bidirectional gated recurrent unit network is proposed.A large number of projectile trajectory data samples are obtained through the simulation of the 6-degree-of-freedom projectile motion model under different conditions,and the simulated dataset is constructed using the sliding window method and difference method.The proposed trajectory prediction model is compared with GRU model through simulation experiments on the simulated dataset.The research results show that the S2S ATT-BiGRU model,when predicting the future time of 5 s based on an input time of 0.5 s,has an MSE of 13.89 m^(2),an MAE of 1.84 m,and an MAPE of 0.29%,which is significantly more accurate than GRU model.In most cases,the S2S ATT-BiGRU model’s predictions are superior to GRU model.These findings suggest that the S2S ATT-BiGRU model has a stronger ability to store input trajectory sequence information and adaptively focus on important input trajectory information,providing a favorable reference for projectile trajectory prediction research.

关 键 词：轨迹预测 弹道模型 S2S ATT-BiGRU模型 深度学习 

分 类 号：TJ012.3[兵器科学与技术—兵器发射理论与技术]