残差膨胀卷积结构下的多模态特征调制方式识别

Automatic Modulation Classification Based on Residual Dilated-convolution and Multi-feature

自动调制方式识别技术在通信领域有着不可或缺的作用,针对传统的卷积神经网络在信号分类问题中特征提取能力不足的问题,本文研究了一种利用多维度特征的端到端双流膨胀卷积神经网络来对调制信号进行分类的方法。该方法不仅利用原始采样信号,还利用输入信号的瞬时幅度和相位信息;原始IQ(In-phase and Quadrature,IQ)数据输入进神经网络后,网络首先通过内置的数据预处理模块对输入的IQ信号进行预处理,提取原始信号的幅度和相位信息,再将原始IQ信号和幅度相位两种特征信息分别通过两个并行的卷积神经网络结构分别进行特征提取;本文所设计的双流卷积神经网络模型中的膨胀残差网络分支利用卷积核的膨胀卷积特性,将膨胀卷积与残差网络结构相结合,在网络参数不变的情况下使得卷积核具有更大的感受野,同时也能够更好地结合上下文信息,另一个网络分支是将卷积神经网络与长短期记忆神经网络相串联,然后将两个并行卷积神经网络的输出特征向量进行矩阵相乘达到两种特征信息融合的目的。整个识别过程是基于端到端的,数据预处理模块内嵌到神经网络内部,由神经网络完成对数据的预处理,只需将原始的IQ数据直接送入神经网络即可;仿真实验结果显示相比较于单分支结构的卷积神经网络模型或者循环神经网络模型,本文所提出的基于残差膨胀卷积的双流网络结构在数据集RML2016.10a上识别准确率有了极大地提升,识别准确率最高能够达到85%,同时对于单分支结构无法识别的16QAM和64QAM两种信号,本文模型也具有一定的分类能力。

Automatic Modulation Classification(AMC)plays an important role in the field of communication.Aiming at the insufficient feature extraction ability of single traditional convolutional neural networks in signal classification problems,we exploit an end-to-end dilated convolutional neural network using multi-dimensional features to classify modulated signals.This method not only uses the original sampling signal,but also uses the instantaneous amplitude and phase information of the input signal;after the original IQ(In-phase and Quadrature)data is input into the neural network,the network first preprocesses the input IQ signal through the built-in data preprocessing module to extract the original signal Amplitude and phase information,and then the two kinds of feature information are respectively extracted through two parallel Convolutional Neural Network(CNN)structures.We exploit the dilated convolution to make the convolution network has a larger receptive field,and it also better combines contextual information.Meanwhile expansion convolution is combined with residual network structure.The output feature vectors are matrix-multiplied to achieve the purpose of fusion.The whole recognition process is based on end-to-end.The data preprocessing module is embedded in the neural network,so the neural network completes the data preprocessing.Thanks to the end-to-end structure,we only need to send the original IQ data directly to the neural network;The simulation experiment results show that compared with the convolutional neural network with a single structure,the recognition accuracy of the model proposed in this paper has been greatly improved on RML2016.10a dataset,and the highest recognition accuracy can reach 85%.At the same time,the model also has a certain ability to classify 16QAM and 64QAM signals that cannot be recognized by single branch structure.