Research Of Automatic Modulation Classification Based On Convolutional Neural Networks

In the situation of an increasingly dense communication environment and endless modulation patterns of signals,the modulation classification becomes more and more difficult.It is very important for the application of radio communication to seek a new method of automatic modulation classification(AMC)with high accuracy and good timeliness.Convolutional neural networks have become a research focus in the field of deep learning due to its powerful ability in feature extraction and learning.Many researchers have tried to use convolutional neural networks in automatic modulation classification.However,these algorithms usually have simple network structures,weak feature extraction ability,and poor classification performance.Those networks have a large amount of parameters and long compute time.Therefore,on the basis of the existing research results,this thesis proposes two convolutional neural networks: convolutional neural networks based on raw in-phase and quadrature data(RID-CNN)and convolutional neural networks based on Time-Frequency characteristics(TFC-CNN).In this thesis,smaller convolution kernels are used in the both network models to improve the feature extraction efficiency and reduce the computational complexity.RID-CNN uses raw IQ data of the signal as model input,and removes the fully connected layer in the traditional method,which can greatly reduce the model parameters while ensuring the classification effect.TFC-CNN uses the signal time-frequency diagram as input.This thesis designs a new network structure called TFC structure,which improves the feature extraction ability and classification effect of TFC-CNN by changing the connection between different convolutional layers of the model.Focusing on the problem that single-path convolutional neural networks cannot fully extract data information and comprehensively learn the modulation characteristics of data,this thesis additionally proposes a dual-path convolutional neural networks(DP-CNN).DP-CNN combines the characteristics of RID-CNN and TFC-CNN,which can extract features of raw IQ data and time-frequency diagram,so as to obtain higher classification accuracy.At the same time,in the process of training all network models,regularization,batch normalization,and other optimization methods are applied to make the model training more stable and prevent the model from problems concerning optimization.The experimental results show that RID-CNN and TFC-CNN mentioned in this thesis have both superior classification performance and lesser compute time.DP-CNN proposed in this thesis not only has low computational complexity but also achieves high-precision automatic modulation classification.