系统工程与电子技术 ›› 2026, Vol. 48 ›› Issue (10): 3363-3371.doi: 10.12305/j.issn.1001-506X.2026.10.09

• 传感器与信号处理 • 上一篇    

基于改进DenseNet的LPI雷达特征提取算法

杜赟1(), 刘松涛2(), 彭锐晖1(), 白雪涛1(), 郭宇1()   

  1. 1. 哈尔滨工程大学信息与通信工程学院,黑龙江 哈尔滨 150001
    2. 海军大连舰艇学院信息系统系,辽宁 大连 116018
  • 收稿日期:2025-05-29 接受日期:2026-01-16 出版日期:2026-10-25 发布日期:2026-09-30
  • 通讯作者: 刘松涛 E-mail:dy3318883623@163.com;navylst@163.com;pengruihui@hrbeu.edu.cn;seadesert@heru.edu.cn;2647086116@qq.com
  • 作者简介:杜 赟(2000—),男,硕士研究生,主要研究方向为电子对抗技术及应用
    彭锐晖(1979—),男,教授,博士,主要研究方向为多源融合探测、雷达抗干扰与目标特性技术
    白雪涛(2000—),男,硕士研究生,主要研究方向为脑电信号处理、分类与识别技术
    郭 宇(2001—),男,硕士研究生,主要研究方向为电子对抗技术及应用

LPI radar feature extraction algorithm based on improved DenseNet

Yun Du1(), Songtao Liu2(), Ruihui Peng1(), Xuetao Bai1(), Yu Guo1()   

  1. 1. School of Information and Communication Engineering,Harbin Engineering University,Harbin 150001,China
    2. Department of Information System,Dalian Naval Academy,Dalian 116018,China
  • Received:2025-05-29 Accepted:2026-01-16 Online:2026-10-25 Published:2026-09-30
  • Contact: Songtao Liu E-mail:dy3318883623@163.com;navylst@163.com;pengruihui@hrbeu.edu.cn;seadesert@heru.edu.cn;2647086116@qq.com

摘要:

为提升低截获概率(low probability of intercept,LPI)雷达多分量信号背景下的特征提取能力,提出一种基于改进稠密神经网络(densely connected convolution networks,DenseNet)的特征提取算法。所提算法以DenseNet169为主干网络,结合优化的卷积块注意力模块,并将传统空间注意力机制和通道注意力机制的串行结构改进为并行结构。同时,对信号二维时频图像进行降噪滤波处理,以增强模型对关键信息的表征能力。实验结果表明,对比7种传统网络结构,改进DenseNet算法在模拟真实信道环境的背景下可以实现更好的识别效果,在SNR=−8 dB时对常见的13类LPI雷达信号识别的精确率、召回率以及F1-Score均超过了97%,在−8 dB≤SNR≤10 dB的背景下平均准确率达到了99.22%。所提算法在LPI雷达信号深度时频特征提取与识别方面具有有效性。

关键词: 低截获概率雷达, 特征提取, 稠密神经网络, 注意力机制

Abstract:

To enhance feature extraction capabilities in multi-component signal backgrounds of low probability of intercept (LPI) radars, a feature extraction algorithm based on improved densely connected convolutional network (DenseNet) is proposed. The proposed algorithm utilizes DenseNet169 as the backbone network, incorporating an optimized convolutional block attention module (CBAM). The serial structure of traditional spatial attention mechanism (SAM) and channel attention mechanism (CAM) are transformed into a parallel architecture. Additionally, denoising filtering is applied to the two-dimensional time-frequency images of the signals to enhance the model’s representation capability of critical information. Experimental results demonstrate that the improved DenseNet algorithm achieves superior recognition performance compared to seven traditional network architectures under simulated real-channel conditions. At SNR=−8 dB, it achieves accuracy, recall, and F1-Score exceeding 97% for 13 common LPI radar signal classes. In −8 dB≤SNR≤10 dB range, the average accuracy reaches 99.22%. The proposed algorithm demonstrates effectiveness in the extraction and recognition of deep time-frequency for LPI radar signals.

Key words: low probability of intercept (LPI) radar, feature extraction, densely connected convolution network, attention mechanism

中图分类号: