电子学报 ›› 2022, Vol. 50 ›› Issue (1): 79-88.DOI: 10.12263/DZXB.20201145

• 学术论文 • 上一篇    下一篇

基于深度多级残差网络的低信噪比下空频分组码识别方法

张聿远, 张立民, 闫文君   

  1. 海军航空大学信息融合研究所,山东 烟台 264001
  • 收稿日期:2020-10-18 修回日期:2021-01-07 出版日期:2022-01-25 发布日期:2022-01-25
  • 作者简介:闫文君 男,1986年出生,山东莱州人.现为海军航空大学副教授.主要研究方向为空时分组码识别.E-mail:wj_yan@foxmail.com
  • 基金资助:
    国家自然科学基金重大研究计划(91538201);泰山学者工程专项经费(Ts201511020)

A Space-Frequency Block Code Recognition Based on Deep Multilevel Residual Network with Low SNR

ZHANG Yu-yuan, ZHANG Li-min, YAN Wen-jun   

  1. Department of Information Fusion,Naval Astronautical University,Yantai,Shandong 264001,China
  • Received:2020-10-18 Revised:2021-01-07 Online:2022-01-25 Published:2022-01-25

摘要:

针对低信噪比下信号受噪声干扰强,空频分组码(Space-Frequency Block Code,SFBC)识别准确率低的问题,提出了一种基于时频分析与深度多级残差网络的SFBC自动识别方法.通过对互相关序列进行时频分析与降噪、非时钟同步拼接等预处理,以获取能够反映其本质特征的二维图像,适应不同接收端时延下的信号识别,构建带有多层跨越连接的深度多级残差网络以充分融合深浅层特征,实现SFBC识别.该方法不需要人为设定阈值和假设检验统计量,克服了传统算法人工提取特征存在的调参过程烦琐、专业经验要求高的缺陷,对低信噪比环境具有较强的适应性.在信噪比为-14dB时,该方法的识别准确率达到了95.8%.本文提出的特征转化和预处理方法,为基于特征提取的识别方法与深度学习技术相结合提供了新思路,其思想同样可应用于其他通信信号识别领域.

关键词: 空频分组码, 时频分析, 非时钟同步, 深度学习, 深度多级残差网络

Abstract:

Aiming at the problem of low recognition accuracy of space-frequency block code(SFBC) under low signal to noise ratio(SNR), an automatic recognition method of SFBC based on deep multilevel residual network(DMRN) is proposed. Through time frequency analysis of cross-correlation sequence, noise reduction and non-clock synchronization, the signal recognition can be adapted to different delay of the receiver and its essential characteristics can be reflected. DMRN with multi-layer spanning connections was constructed to fully integrate the features of deep and shallow layers to realize SFBC recognition. This method does not need to set thresholds and hypothesis testing statistics, and overcomes the defects of traditional algorithms in extracting features manually, for example, complex parameter adjustment process and high requirement of professional experience, and it has strong adaptability to low SNR environment. At -14dB, the recognition accuracy reaches 95.8%. The feature transformation and preprocessing methods proposed in this paper provide a new idea for the combination of feature extraction based recognition method and deep learning, which can also be applied to other fields of communication signal recognition.

Key words: space-frequency block code, time-frequency analysis, non-clock synchronization, deep learning, deep multilevel residual network

中图分类号: