系统工程与电子技术 ›› 2021, Vol. 43 ›› Issue (11): 3086-3097.doi: 10.12305/j.issn.1001-506X.2021.11.07

• 电子技术 • 上一篇    下一篇

基于复数卷积-残差网络的雷达杂波幅度统计模型分类

张良, 杨威*, 李玮杰, 杨小琪, 刘永祥   

  1. 国防科学技术大学电子科学学院, 湖南 长沙 410073
  • 收稿日期:2021-05-24 出版日期:2021-11-01 发布日期:2021-11-12
  • 通讯作者: 杨威
  • 作者简介:张良(1988—), 男, 博士研究生, 主要研究方向为自然环境雷达回波分类|杨威(1985—), 男, 副教授, 博士, 主要研究方向智能感知与处理, 特别是雷达信号与信息处理|李玮杰(1996—), 男, 博士研究生, 主要研究方向为机器学习|杨小琪(1992—), 女, 博士研究生, 主要研究方向为雷达目标检测和阵列信号处理|刘永祥(1976—), 男, 教授, 博士, 主要研究方向为目标微动特性分析与识别
  • 基金资助:
    国家自然科学基金资助课题(61871384);国家自然科学基金资助课题(61921001)

Classification of radar clutter amplitude statistical model based on complex-valued convolutional-ResNet

Liang ZHANG, Wei YANG*, Weijie LI, Xiaoqi YANG, Yongxiang LIU   

  1. College of Electronic Science and Engineering, National University of Defense Technology, Changsha 410073, China
  • Received:2021-05-24 Online:2021-11-01 Published:2021-11-12
  • Contact: Wei YANG

摘要:

雷达杂波幅度统计模型分类是进行杂波背景下检测目标的重要步骤。雷达杂波原始数据通常是复数数据, 但现有杂波幅度统计模型分类研究都是在实数数据上完成的。复数数据同时包含幅度和相位信息, 更丰富的信息量有助于雷达杂波幅度统计模型分类。为此, 引入复数神经网络, 利用仿真杂波高分辨距离像(high resolution range profile, HRRP)复数数据, 对雷达杂波幅度统计模型分类问题进行研究, 完成了以下工作: 一是为构建复数最大池化层, 定义并改进了复数最大池化算法, 通过复数卷积神经网络(complex-valued convolutional neural networks, CV-CNN) 对杂波幅度统计模型的分类实验, 对比了两种复数最大池化算法和复数平均池化算法的分类效果, 实验结果表明复数最大池化算法的分类效果更好, 分类准确率为97.29%;二是为进一步提高分类准确率, 构建了复数卷积-残差网络(complex-valued convolution-ResNet, CV-CRN), 通过实验对比分析了CV-CRN的性能, 实验结果表明, CV-CRN的分类性能优于CV-CNN, 分类准确率达到98.84%, 并具有较好的鲁棒性。

关键词: 杂波分类, 高分辨距离像, 复数卷积-残差网络

Abstract:

The classification of clutter amplitude statistical model is an important step in detecting targets under clutter background. However, the raw clutter data from radar are usually complex numbers, and most of current research on classifying clutter amplitude statistical models focuses on real data. Complex data contain not only amplitude but phase information as well, and hence are beneficial to the classification problem we concern. In order to study the classification of radar clutter amplitude statistical model, we propose to use complex-valued neural network to process simulated complex clutter data of high resolution range profile (HRRP).The following work is completed: First, in order toconstructcomplex-valued pooling layer, the complex-valued maximum pooling algorithm is defined and improved based on python programming language, and the classification effect of two kinds of complex-valued maximum pooling algorithm and complex-valued average pooling algorithm is compared by using complex-valued convolutional neural networks (CV-CNN) to classify the statistical model of clutter amplitude. The experimental results show that the complex-valued maximum pooling algorithm is more effective and the classification accuracy reaches 97.29%. Second, in order to further improve the classification accuracy, we constructed the complex-valued convolution-ResNet (CV-CRN), then, compared and analyzed the performance of CV-CRN through the experiments of classification of clutter amplitude statistical model. Experimental results have revealed that, the classification performance of CV-CRN was better than that of CV-CNN, with classification accuracy rate being 98.84% and good robustness.

Key words: clutter classification, high resolution range profile (HRRP), complex-valued convolutional-resNet (CV-CRN)

中图分类号: