摘要
简述了利用深层卷积神经网络进行自动调制识别(Automatic Modulation Recognition,AMR)的进展,并结合其模型在基准数据集上的实验表明,大多数不依赖于先验知识的特征提取模型容易忽略模型参数量大、计算复杂度高的问题,因此将工作重点集中在保持高精确度的同时轻量化模型。利用多信道深度学习模型,从时间和空间的角度有效提取特征,搭建以卷积神经网络(Convolution Neural Networks,CNN)和门控循环单元(Gating Recurrent Unit,GRU)为特征提取层的深层学习框架,可以在现有高识别度模型的识别效果上有略微提升,具有高效的收敛速度,且减少了40%以上的参数体积,在训练时间和测试时间上更有优势。该方法在RadioML2016.10a数据集0 dB以上信噪比条件下的识别精度保持在90%以上。
This paper briefly describes the advances in AMR(Automatic Modulation Recognition)techniques using deep convolutional neural networks.Experiments of the model on benchmark datasets indicate that most feature extraction models that do not rely on prior knowledge tend to ignore the problems of large number of model parameters and high computational complexity.Therefore,this paper focuses the work on lightweighting the model while maintaining high accuracy.It utilizes a multi-channel deep learning model to effectively extract features from the perspective of time and space,and builds a deep learning framework with a CNN(Convolutional Neural Networks)and a GRU(Gated Recurrent Unit)as the feature extraction layer,so that it can slightly improve the recognition effect of the existing high-recognition model with efficient convergence speed and reduce the parameter volume by more than 40%,which has more advantages in training time and test time.The recognition accuracy of this method in RadioML2016.10a is maintained at more than 90%at a signal-to-noise ratio of more than 0 dB.
作者
彭钰琳
文红
侯文静
石伟宏
严地宝
PENG Yulin;WEN Hong;HOU Wenjing;SHI Weihong;YAN Dibao(University of Electronic Science and Technology of China,Chengdu Sichuan 611731,China)
出处
《通信技术》
2023年第6期714-718,共5页
Communications Technology
基金
四川省自然科学基金(2022NSFSC0946)。
关键词
自动调制识别
深度学习
卷积神经网络
轻量级
automatic modulation recognition
deep learning
convolutional neural network
lightweight
