基于神经网络的降落伞落点散布快速生成方法  

A Neural Network Based Method to Rapidly Generate the Dispersion of Parachute′s Landing Positions

作　　者：郭鲲 花飞 曹润清 GUO Kun;HUA Fei;CAO Run-qing(Xi′an Aeronautics Computing Technique Research Institute,AVIC,Xi′an 710000,China)

机构地区：[1]航空工业西安航空计算技术研究所,陕西西安710000

出　　处：《航空计算技术》2023年第1期58-61,76,共5页Aeronautical Computing Technique

基　　金：航空科学基金项目资助(2018ZC31003)。

摘　　要：为了解决使用蒙特卡洛法计算降落伞落点散布慢的问题,提出了组合使用降落伞运动模型和神经网络产生落点散布,并利用硬件加速神经网络推理过程的方法。降落伞运动模型用于产生散布中心,神经网络用于产生相对于散布中心的落点。验证了两种不同结构的神经网络:多层感知器和转置卷积神经网络。发现降落伞运动模型结合多层感知器计算得到的降落伞落点散布和使用蒙特卡洛法计算得到的落点散布最为接近。同时,降落伞运动模型和多层感知器或者转置卷积神经网络组合使用都可以达到实时计算降落伞落点散布的效果。因此,提出的方法可以较好地应用于高时效性场景下的降落伞落点散布计算。It is conventional to use Monte Carlo method to calculate the dispersion of parachute′s landing positions.However,this method takes too much time.To tackle this problem,a neural network based method is proposed,which combines a parachute dynamic model with a neural network.In addition,the hardware acceleration technique is applied to speed up the inference of the neural network.The parachute dynamic model is used to calculate the center of the dispersion and the neural network is used to generate parachute′s landing positions surrounding the center.Two types of neural networks,Multi-layer perceptron(MLP) and transposed convolution neural network(TCNN),are trained and tested for the proposed method in this paper.Experiment results show that landing positions generated by MLP are more similar to those generated by the Monte Carlo method.In addition,the proposed method,no matter which type of the neural network is used,can generate landing positions in real-time.Therefore,the method proposed in this paper can be applied to generate the dispersion of parachute′s landing positions when real-time is critical.

关 键 词：降落伞落点散布 多层感知器 转置卷积神经网络 圆概率误差 

分 类 号：V244.216[航空宇航科学与技术—飞行器设计]