基于深度残差网络的自动调制识别方法研究

自动调制识别是电磁环境特性分析的关键问题,而传统方法多基于人工设计特征进行识别,数据特征表示和判别分析能力有限。为此提出一种新颖的深度神经网络特征表示方法进行调制识别任务。首先,利用递归神经网络结构对电磁信号序列进行表示,建立了基于多层双向GRU网络结构的识别方法。其次,从一维空间卷积表示序列的角度思考,建立了基于深度残差卷积网络的调制识别方法。最后,针对加性高斯白噪声信道的调制方式仿真数据集,将提出的方法与典型神经网络模型如多层感知器、卷积神经网络进行了对比实验。实验结果表明,提出的方法在自动调制识别方面具备更强的特征表示能力和竞争力,有利于推动深度学习在自动调制识别领域的应用。

Automatic Modulation Recognition with Deep Residual Network

Automatic modulation recognition(AMR)is a key issue in the analysis of electromagnetic environment characteristics.However,existed methods use hand-crafted features for AMR,which is limited in feature representation capacity.In this paper,based on the shortcomings of the traditional automatic modulation recognition methods,two novel data-driven feature representation methods based on deep neural network are proposed.Firstly,the recurrent neural network was considered to represent the signal sequence,and a multi-layer bidirectional Gated Recurrent Unit network structure was proposed.Secondly,from the view of one-dimensional spatial convolution can representation local sequence,an AMR method based on deep residual convolution network was proposed.Finally,the proposed methods were compared with typical neural network models such as multilayer perceptron and convolutional neural network in the simulation dataset with additive Gaussian white noises.The experimental results show the proposed methods exhibit a discriminate representation and compare performance.Our research is a step towards practical deep learning application in AMR.