基于深度学习的空空导弹多类攻击区实时解算

现代空战日趋复杂，传统攻击区只能提供导弹发射的限度，无法满足现代空战决策的需求。基于目标机水平逃逸角度的最大攻击区、50°攻击区、90°攻击区、水平不可逃逸攻击区、最小攻击区等5类攻击区，针对现有攻击区解算方法无法同时解算多种攻击区的问题，提出多函数深度拟合网络(MFDFN)模型，以实现多种攻击区的同时解算。设计了改进的进退法解算流程，通过弹道仿真获取攻击区数据样本库。根据多函数拟合网络的特点，设计了“整体预训练+局部微调”训练策略，并对网络进行有监督训练。仿真结果表明：采用“整体预训练+局部微调”训练策略的MFDFN比传统网络不仅提高了计算实时性，而且很大程度上提高了计算准确性，其平均相对误差低至0.27%，平均绝对误差低至58.81 m；MFDFN模型具有较强的可靠性和实用性。

Real-time Computing of Air-to-air Missile Multiple Capture Zones Based on Deep Learning

Modern air combat is becoming more and more complex，and the traditional attack zones can only provide the limits of missile launch，which cannot meet the needs of modern air combat decision-making. For this reason，five types of attack zones, maximun attack zone, 50° attack zone, 90° attack zone, horizontal unescapable attack zone, and minimun attack zone, based on escape angles of enemy aircraft are proposed. The existing solving method of attack zone cannot be used to simultaneously solve the problem of multiple attack zones.A multi-function deep fitting network (MFDFN) is proposed to realize the simultaneous solution of multiple attack zones. Firstly，an improved advance-retreat method is designed，and the sample library of capture zone is obtained through ballistic simulation. According to the characteristics of multi-function fitting network，a training strategy，called “overall pre-training and local fine-tuning”，is presented，by which network is supervised trained.The simulated results show that the MFDFN using the “overall pre-training and local fine-tuning” training strategy reduces the computing time while greatly improving the computing accuracy. The average relative error is about 0.27%，and the average absolute error is about 58.81 meters，which proves that the model is reliable and practical.