低信噪比下基于分裂EfficientNet网络的雷达信号调制方式识别  

Radar Signal Modulation Recognition Based on Split EfficientNet Under Low Signal-to-Noise Ratio

作　　者：李钦 刘伟 牛朝阳 宝音图 惠周勃 LI Qin;LIU Wei;NIU Chao-yang;BAO Yin-tu;HUI Zhou-bo(PLA Strategic Support Force Information Engineering University,Zhengzhou,Henan 450001,China)

机构地区：[1]中国人民解放军战略支援部队信息工程大学,河南郑州450001

出　　处：《电子学报》2023年第3期675-686,共12页Acta Electronica Sinica

摘　　要：针对低信噪比条件下复杂多类雷达信号调制方式识别率低的问题,本文提出了一种基于时频分析和深度学习的雷达信号调制方式识别方法.利用CTFD(Cohen class Time-Frequency Distribution)时频分析将信号时域波形变换为二维时频图像,更清晰地表征信号特征;采用灰度化和双三次插值运算等方法对时频图预处理,实现图像通道数和尺寸的减少,以降低深度学习模型数据输入量;进一步调整输入输出通道数构建小型EfficientNet网络,再由多个小型网络并行处理构建分裂网络EfficientNet-B0-Split3,将时频图像输入网络实现雷达信号调制方式识别.实验结果表明,在信噪比为-8 dB时,新方法对17类不同调制方式的雷达信号整体识别率可达97.1%,相对于扩张残差网络提高约2.4个百分点;在信噪比为-10 dB时,识别率可达92.1%,相对于EfficientNet提高约0.7个百分点,提升了低信噪比条件下复杂多类雷达信号调制方式识别率.Aiming at the low recognition rate of complex and multi-class radar signal modulation under the condition of low signal-to-noise ratio(SNR),this paper proposes a radar signal modulation recognition based on time-frequency analy⁃sis and deep learning.Using Cohen class time-frequency distribution(CTFD),transform the signal time-domain waveform into a two-dimensional time-frequency image to characterize the signal characteristics more clearly.Using methods such as grayscale and bicubic interpolation,preprocess the time-frequency image to reduce the number and size of image channels and the amount of data input for deep learning models.Further,adjusting the number of input and output channels to build a small EfficientNet.Then multiple small networks are processed in parallel to construct a split network EfficientNet-B0-Split3.The network gets the inputted time-frequency images and realizes the radar signal modulation recognition.The ex⁃perimental results show that when the SNR is−8 dB,the overall recognition rate of the new method for radar signals with 17 different modulation methods can reach 97.1%,which is about 2.4 percentage points higher than dilated residual network(DRN);when the SNR is−10 dB,the recognition rate can reach 92.1%,which is about 0.7 percentage points higher than Ef⁃ficientNet,improves the recognition rate of complex and multi-class radar signal modulation methods under the condition of low SNR.

关 键 词：雷达信号 调制识别 时频分析 深度学习 分裂EfficientNet 

分 类 号：TN971[电子电信—信号与信息处理]