西安电子科技大学学报 ›› 2023, Vol. 50 ›› Issue (3): 40-49.doi: 10.19665/j.issn1001-2400.2023.03.004

• 面向IT3.0的感通算融合6G关键技术专题 • 上一篇    下一篇

CNN-Transformer轻量级智能调制识别算法

杨静雅1,2,3(),齐彦丽1,2,3(),周一青1,2,3(),赵登攀4(),王尚权1,2,3(),石晶林1,2,3()   

  1. 1.中国科学院计算技术研究所 处理器芯片全国重点实验室,北京 100190
    2.中国科学院大学 计算机科学与技术学院,北京 100049
    3.中国科学院计算技术研究所 移动计算与新型终端北京市重点实验室,北京 100190
    4.中国人民解放军77646部队,西藏自治区 日喀则 858600
  • 收稿日期:2022-12-16 出版日期:2023-06-20 发布日期:2023-10-13
  • 通讯作者: 周一青
  • 作者简介:杨静雅(1995—),女,中国科学院计算技术研究所博士研究生,E-mail:yangjingya19b@ict.ac.cn;|齐彦丽(1991—),女,助理研究员,E-mail:qiyanli@ict.ac.cn;|赵登攀(1982—),男,中国人民解放军77646部队,E-mail:375914615@qq.com;|王尚权(2000—),男,中国科学院计算技术研究所硕士研究生,E-mail:13207502540@163.com;|石晶林(1972—),男,研究员,E-mail:sjl@ict.ac.cn
  • 基金资助:
    国家重点研发计划(2020YFB1807800)

Algorithm for recognition of lightweight intelligent modulation based on the CNN-transformer networks

YANG Jingya1,2,3(),QI Yanli1,2,3(),ZHOU Yiqing1,2,3(),ZHAO Dengpan4(),WANG Shangquan1,2,3(),SHI Jinglin1,2,3()   

  1. 1. State Key Lab of Processors,Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100190,China
    2. School of Computer Science and Technology,University of Chinese Academy of Sciences,Beijing 100049,China
    3. Beijing Key Laboratory of Mobile Computing and Pervasive Device,Institute of Computing Technology, Chinese Academy of Sciences,Beijing 100190,China
    4. Unit 77646 of PLA,Shigatse 858600,China
  • Received:2022-12-16 Online:2023-06-20 Published:2023-10-13
  • Contact: Yiqing ZHOU

摘要:

现有基于深度学习的调制识别方法,当存在噪声及不确定信道干扰影响时识别率较低,同时由于模型参数量大,难以直接应用到移动终端。针对该问题,提出一种基于卷积神经网络(CNN)和Transformer的轻量级智能调制识别方法。该方法首先利用卷积神经网络对信号进行局部信息特征提取,然后利用卷积神经网络通道注意力和Transformer时域注意力模块分别从信号的通道和时域两个维度关注最有利于识别的特征,降低信道或噪声等的影响,以提升识别率。所提方法可以适应多种信号表征,如原始IQ信号、幅度相位信号及变换域特征。仿真表明,在RadioML2016.10b数据集上,相较现有基于卷积神经网络的方法,所提方法的平均识别率提升了约8%~12%,相比基于残差神经网络和长短时记忆网络的方法,参数量降低了约90%~92%,计算量降低了约83%~93%。实验结果验证了所提方法增加模型分类精度的同时,有效地降低了模型的参数量和计算量。

关键词: 调制识别, 通道注意力, 时域注意力, 轻量级网络

Abstract:

Existing modulation recognition methods based on deep learning have the problems of low recognition accuracy under the influence of noise and uncertain channel interference,and are difficult to apply to mobile terminals due to a large number of parameters.This paper proposes a lightweight modulation recognition method based on the Convolutional Neural Network (CNN) and Transformer to solve the above problems.In order to improve the accuracy,the CNN is first used to extract the local features of the signal.Then,the CNN-based channel attention and Transformer-based temporal attention modules are used to focus on the features that are most conducive to recognition from the two dimensions of the signal channel and time domain,respectively,while reducing the impact of the channel,noise,etc.The proposed method can be applied to a variety of signal representations,such as raw IQ signals,amplitude-phase signals,and transform domain features.Simulation shows that on the RadioML2016.10b dataset,compared with the existing convolutional network methods,the average recognition accuracy of the proposed method is increased by 8%~12%.Compared with the methods based on the residual neural network and long-term memory network,the number of parameters is reduced by 90%~92%,and the amount of calculation is reduced by about 83%~93%.Experimental results show that the proposed method can improve the accuracy of model classification while effectively reducing the number of parameters and the amount of calculation.

Key words: modulation recognition, channel attention, temporal attention, lightweight network

中图分类号: 

  • TN911.7