用门控循环单元实时预测空战飞行轨迹  

Real-time prediction of air combat flight trajectory using GRU

作　　者：张宏鹏 黄长强[1] 轩永波 唐上钦 ZHANG Hongpeng;HUANG Changqiang;XUAN Yongbo;TANG Shangqin(Aeronautics Engineering College, Air Force Engineering University, Xi’an 710038, China;Air Force Research Institute, Beijing 100085, China)

机构地区：[1]空军工程大学航空工程学院,陕西西安710038 [2]空军研究院,北京100085

出　　处：《系统工程与电子技术》2020年第11期2546-2552,共7页Systems Engineering and Electronics

基　　金：国家自然科学基金(51579209)资助课题。

摘　　要：为了提高飞机飞行轨迹预测准确率、确保轨迹预测实时性,提出使用门控循环单元(gated recurrent unit,GRU)预测轨迹。对不同条件下的不同机动动作进行飞行仿真,得到大量轨迹样本。设计具有不同层数和神经元个数的网络,用得到的样本对其进行训练。选出在测试集上误差最小的网络结构。对比GRU网络、循环神经网络和反向传播网络的相对误差和预测用时。引入坐标变换矩阵,使轨迹预测不受航向和坐标系影响。对比3种方法在一段频繁变化的轨迹上的绝对误差。结果表明,所提方法的平均绝对误差在x轴上约为18 m,在y轴上约为11 m,在z轴上约为22 m,显著小于另外两种方法,且平均预测用时约为2.4 ms,满足实时性要求。To increase the accuracy of the aircraft flight trajectory prediction and ensure real-time perfor-mance of the trajectory prediction,a trajectory prediction method using gated recurrent unit(GRU)is proposed.Flight simulations are conducted under different conditions with different maneuvers and numerous trajectory samples are acquired.Neural networks with different number of layers and neurons are designed and trained by the acquired trajectory samples.The network structure with minimum error on the test set is selected.The re-lative error and prediction time cost of the GRU networks,recurrent neural networks and back propagation networks are compared.Transformation matrix of coordinate is introduced to make the trajectory prediction unaffected by course and coordinate system.The absolute error of the three methods in a frequently changing trajectory are compared.The results indicate that the average absolute error of the proposed method is 18 m in x-axis,11 m in y-axis and 22 m in z-axis approximately,and the error of the proposed method is significantly less than the other two methods and the average prediction time cost of it is about 2.4 ms,which meet the real-time requirements.

关 键 词：轨迹预测 门控循环单元 循环神经网络 空战仿真 坐标变换矩阵 

分 类 号：V247.5[航空宇航科学与技术—飞行器设计]