Parameter Estimation Of Multi-type Warheads Based On Convolutional Neural Networks

Space ballistic target recognition plays an important role in national defense systems.The middle segment is one of the most favorable stages in the whole flight of a ballistic target.With the development of radar technology and the advancement of science and technology,fretting is an inherent feature of ballistic target motion,which is more and more used in the identification of ballistic target.The narrowband radar has the advantages of long detection distance and sensitivity to the target micro-Doppler characteristics,so it has great recognition potential.Based on the above background,this paper aims to improve the geometric structure and the estimation accuracy of the moving parameters of the ballistic target by training the convolution neural network and improving the network structure and increasing the effective features.This paper mainly introduces the application of spatial ballistic target parameter estimation based on convolutional neural network through the following parts:The first part firstly models the ballistic target,deduces the theoretical formulas of the ballistic target spin,advance and chapter motion,and introduces the two most commonly used time-frequency analysis methods of Short-time Fourier Transform and Wigner-Ville.In the second part,based on the equivalent scattering center theory,the equivalent scattering center of flat cone,cylinder cone and skirt cone is analyzed.The change law of attitude angle during the flight of the ballistic target is deduced,and the Micro-Doppler theoretical curves of the three types of ballistic targets are obtained.Secondly,it introduces the theory of convolutional neural network,and gives the architecture of convolutional neural network in this paper.Finally,three kinds of ballistic target parameter estimation results are given: flat cone,cylinder cone and skirt cone.The third part studies how to further improve the estimation accuracy of three kinds of warhead parameters: flat cone,cylinder cone and skirt cone.the loss function of the convolutional neural network is improved and the weighted loss function is proposed.then considering the utilization of multi-polarization and multi-view,we propose weighted two-channel convolutional neural network and weighted double-layer convolutional neural network.Finally,the weighted double-layer two-channel convolution neural network is proposed.