基于卷积神经网络的无人作战飞机飞行轨迹实时预测

飞行轨迹预测是空战技术的一部分，预测方结合轨迹预测结果可以选择出更有预见性的机动。为快速、准确地获得无人作战飞机在未来时刻的位置，提出了基于卷积神经网络的飞行轨迹预测方法。原始动力学模型不能正确仿真滚转角有偏差的筋斗机动，采取限制角速度的方式对该模型进行改进；使用改进后的模型在不同条件下进行飞行仿真，得到大量轨迹样本；训练并测试具有不同层数和卷积核数的网络，从中找出预测误差最小的网络；对比卷积神经网络与长短时记忆网络、循环神经网络、全连接网络的运算速度和误差，结果表明：卷积神经网络预测方法在没有增加运算用时情况下，0.25 s后的平均预测误差在x轴方向约为4.2 m，y轴方向约为8.0 m，z轴方向约为19.5 m，且误差均小于其他3种方法。

CNN-based Real-time Prediction Method of Flight Trajectory of Unmanned Combat Aerial Vehicle

Trajectory prediction is part of air combat technology，and the predictors can select a more predictable maneuvering considerion of trajectory prediction results. A convolution neural network predicting method is proposed to obtain the position of unmanned combat aerial vehicle in a future time quickly and accurately. An improved model for limiting the angular velocity is presented since the original dynamic model can not correctly simulate the somersault maneuvering with roll angle deviation. The improved model is used for flight simulation under different conditions，and a large number of trajectory samples are obtained. The convolution neural networks with different layer number and convolution kernel number is trained and tested，and the network with the smallest prediction error is selected. Operational speed and error of the proposed method are compared with those of long short term memory neural network，recurrent neural network and fully connected neural network. The results show that the average prediction error of the proposed method is about 4.2 m on x axis，8.0 m on y axis and 19.5 m on z axis after 0.25 s without increasing operational time，and the errors are all smaller than those of the other methods.