系统工程与电子技术 ›› 2022, Vol. 44 ›› Issue (6): 2019-2026.doi: 10.12305/j.issn.1001-506X.2022.06.30

• 通信与网络 • 上一篇    下一篇

基于残差生成对抗网络的调制识别算法

秦博伟*, 蒋磊, 许华, 齐子森   

  1. 空军工程大学信息与导航学院, 陕西 西安 710077
  • 收稿日期:2021-05-11 出版日期:2022-05-30 发布日期:2022-05-30
  • 通讯作者: 秦博伟
  • 作者简介:秦博伟 (1998—), 男, 硕士研究生, 主要研究方向为通信信号处理、机器学习、模式识别|蒋磊 (1974—), 男, 副教授, 博士, 主要研究方向为通信信号处理、电子对抗、模式识别|许华 (1976—), 男, 教授, 博士, 主要研究方向为通信信号处理、盲信号处理、通信对抗|齐子森 (1982—), 男, 副教授, 博士, 主要研究方向为跳频信号处理、阵列天线

Modulation recognition algorithm based on residual generation adversarial network

Bowei QIN*, Lei JIANG, Hua XU, Zisen QI   

  1. Information and Navigation College, Air Force Engineering University, Xi'an 710077, China
  • Received:2021-05-11 Online:2022-05-30 Published:2022-05-30
  • Contact: Bowei QIN

摘要:

针对小样本条件下通信信号识别准确率不高、网络训练困难的问题, 本文提出一种基于残差生成对抗网络的调制识别算法。首先, 设计一种以Leakyrelu作为隐藏层激活函数的新残差单元, 使得网络对输入为负值的数据也可以进行梯度计算; 然后, 将新残差单元组成的残差网络和卷积神经网络作为本文算法的基本网络结构, 使用卷积步幅为1的非对称小卷积核, 更好地提取信号的边缘特征信息; 最后, 用Dropout代替池化操作, 并选择Adam梯度优化算法以交替迭代方式完成网络训练。仿真实验结果表明, 小样本条件下, 残差生成对抗网络算法复杂度明显降低, 信噪比(signal to noise ratio, SNR)在0 dB以上时, 对10种调制信号的识别准确率可以达到91%, 验证了所提方法的有效性。

关键词: 残差网络, 生成式对抗网络, 调制识别, 小样本

Abstract:

Aiming at the problem of low accuracy of communication signal recognition and difficult training of network in few-shot learning, a modulation recognition algorithm based on the residual generative adversarial network is proposed. Firstly, this algorithm designs a new residual block unit using Leakyrelu as the activation function of the hidden layer, which can also calculate the gradient for the negative input data. Then, the residual network composed of the new residual units and the convolutional neural network are used as the basic network structure of this proposed algorithm. Moreover, we apply the asymmetric small convolutional kernel with a convolutional step of 1 to better extract the edge feature information of the signal. Finally, Dropout is adopted to replace the pooling operation, and Adam gradient optimization algorithm is selected to complete the network training in an alternate iteration mode. Experimental results show that under the condition of a few-shot learning, the recognition accuracy of the residual generation adversarial network algorithm for 10 modulated signals can reach 91% when the signal to noise ratio (SNR) is higher than 0 dB, which verifies the effectiveness of the proposed method.

Key words: residual network (Resnet), generative adversarial network, modulation recognition, few-shot

中图分类号: