Author Login Editor-in-Chief Peer Review Editor Work Office Work

Computer Engineering ›› 2021, Vol. 47 ›› Issue (4): 108-114. doi: 10.19678/j.issn.1000-3428.0058472

• Mobile Internet and Communication Technology • Previous Articles     Next Articles

Small Sample Modulation Recognition Algorithm Based on Depth Cascade Siamese Network

FENG Lei, JIANG Lei, XU Hua, GOU Zezhong   

  1. Information and Navigation College, Air Force Engineering University, Xi'an 710077, China
  • Received:2020-05-29 Revised:2020-06-30 Published:2020-07-02

基于深度级联孪生网络的小样本调制识别算法

冯磊, 蒋磊, 许华, 苟泽中   

  1. 空军工程大学 信息与导航学院, 西安 710077
  • 作者简介:冯磊(1997-),男,硕士研究生,主研方向为通信信号处理、模式识别;蒋磊,副教授、博士;许华,教授、博士;苟泽中,硕士研究生。
  • 基金资助:
    国家自然科学基金(61601500)。

Abstract: The recognition accuracy of traditional modulation recognition algorithms based on deep learning is reduced in the case of small sample size.To solve the problem,this paper proposes a small sample modulation recognition algorithm for communication signal based on deep cascade Siamese network.Firstly,according to the spatial and temporal features of the sequence diagram of communication signals,the feature extraction module cascaded by Convolutional Neural Network(CNN) and Long Short Term Memory(LSTM) network is designed to map the original signal features to the feature space.At the same time,the extracted features are measured under the Siamese network architecture,and train the network with similarity constraints to avoid the problem of over fitting in the training.Finally,the modulation category of the to be tested samples is identified through the nearest neighbor classifier.Results of experiments on the public modulation dataset DeepSig show that compared with the traditional modulation recognition algorithms based on deep learning,the proposed algorithm can significantly reduce the number of samples required in training,and the recognition accuracy is higher under small sample conditions.

Key words: Convolutional Neural Network(CNN), Long Short Term Memory(LSTM) network, Siamese network, small sample, modulation recognition

摘要: 为解决传统基于深度学习的调制识别算法在小样本条件下识别准确率较低的问题,提出一种基于深度级联孪生网络的通信信号小样本调制识别算法。根据通信信号时序图的时空特性,设计由卷积神经网络和长短时记忆网络级联的特征提取模块将原始信号特征映射至特征空间,同时在孪生网络架构下对提取的特征进行距离度量并以相似性约束训练网络,避免特征提取模块在训练过程中出现过拟合现象,最终通过最近邻分类器识别待测样本的调制类别。在DeepSig公开调制数据集上的实验结果表明,与传统基于深度学习的调制识别算法相比,该算法能有效降低训练过程中所需的样本量,且在小样本条件下的识别准确率更高。

关键词: 卷积神经网络, 长短时记忆网络, 孪生网络, 小样本, 调制识别

CLC Number: