系统工程与电子技术 ›› 2020, Vol. 42 ›› Issue (8): 1841-1849.doi: 10.3969/j.issn.1001-506X.2020.08.26

• 通信与网络 • 上一篇    下一篇

基于TCNN-BiLSTM网络的调制识别算法

刘凯(), 张斌(), 黄青华()   

  1. 上海大学通信与信息工程学院, 上海 200444
  • 收稿日期:2019-12-17 出版日期:2020-07-25 发布日期:2020-07-27
  • 作者简介:刘凯 (1981-),男,副教授,硕士研究生导师,博士,主要研究方向为雷达信号处理、通信信号处理、室内定位技术和深度学习。E-mail:liukai@shu.edu.cn|张斌(1994-),男,硕士研究生,主要研究方向为通信信号处理、深度学习。E-mail:alexanderbin@shu.edu.cn|黄青华 (1978-),女,副研究员,硕士研究生导师,博士,主要研究方向为阵列信号处理、盲信号处理和3D音频。E-mail:qinghua@shu.edu.cn
  • 基金资助:
    国家自然科学基金(61571279)

Modulation recognition algorithm based on TCNN-BiLSTM

Kai LIU(), Bin ZHANG(), Qinghua HUANG()   

  1. School of Communication and Information Engineering, Shanghai University, Shanghai 200444, China
  • Received:2019-12-17 Online:2020-07-25 Published:2020-07-27
  • Supported by:
    国家自然科学基金(61571279)

摘要:

针对传统调制识别算法在低信噪比下识别率不高的情况,提出双路卷积神经网络级联双向长短时记忆(two-way convolutional neural network cascaded bidirectional long short-term memory, TCNN-BiLSTM)网络的调制识别算法。首先,该算法并联不同尺度卷积核的卷积层,提取调制信号不同维度的特征。然后,级联BiLSTM层,对多维特征构建LSTM时间模型。最后,使用softmax分类器完成识别。仿真实验表明,所提算法结构在加性高斯白噪声和特定信道参数的瑞利衰落信道下,性能要优于基于传统特征和其他网络结构的识别算法。在特定信道参数的瑞利衰落信道下信噪比低至6 dB时,该算法对6种数字调制信号的识别率仍可达到92%以上。

关键词: 调制识别, 并联网络, 卷积神经网络, 双向长短时记忆网络

Abstract:

For the traditional modulation recognition algorithm, the recognition rate is not high at low signal to noise ratio (SNR). The paper proposes a two-way convolutional neural network casaded bidirectional long short-term memory (TCNN-BiLSTM) network modulation recognition algorithm. Firstly, the algorithm parallelizes the convolutional layers with convolution kernels of different scales to extract features of different dimensions of the modulation signal. Then it cascades the BiLSTM layers to build LSTM time model for multi-dimensional features. Finally, a softmax classifier is used to complete the recognition. Simulation experiments show that the performance of the algorithm structure under additive Gaussian white noise and Rayleigh fading channels with specific channel parameters is better than the recognition algorithms based on traditional features and other network structures. When the SNR in the Rayleigh fading channel with specific channel parameters is as low as 6 dB, the recognition rate of the six digital modulation signals can still reach above 92%.

Key words: modulation recognition, parallel network, convolutional neural network, bidirectional long short-term memory network

中图分类号: