Research On Modulation Recognition Method Based On Convolutional Neural Network

Modulation recognition is a key technology between signal detection and demodulation,and is the basis for further demodulation and decoding of received signals.It is widely applied in various military and civil scenarios.Traditional modulation recognition methods can be divided into two categories:likelihood-based and feature-based.The former cannot be applied to practical problems due to its dependence on prior information and high computational complexity,while the latter has well engineering application value but the recognition performance is too dependent on artificially constructed features.In recent years,modulation recognition algorithms based on deep learning have been extensively studied.Compared with traditional feature-based algorithms,deep learning recognition algorithms have the advantages of high classification accuracy and no need for artificial construction features.This thesis mainly focuses on modulation recognition method based on deep learning.The main work and innovations of this thesis are as follows:(1)The modulation recognition algorithm based on high-order cumulants and Fully Connected Neural Network(FCNN)and the modulation recognition algorithm based on Convolutional Neural Network(CNN)are studied.The simulation results show that compared with the feature extraction algorithm based on high-order cumulants,the deep learning algorithm has obvious advantages in recognition accuracy,but there are problems of high model complexity and relying on a large number of labeled samples for training.(2)To solve the problem of high model complexity in existing deep learning modulation recognition methods,a low complexity modulation recognition algorithm based on lightweight CNN is presented.In this thesis,a Depthwise Separable Convolution(DSC)residual structure is designed for feature extraction,which can prevent the vanishing gradient problem and reduce computational complexity.In addition,the global depthwise convolution algorithm is adopted for feature reconstruction,which further reduces the model complexity.Simulation results show that the proposed algorithm can significantly reduce model parameters and inference time compared with the recent works.(3)Aiming at the problem that supervised learning modulation recognition algorithm relies on a large number of labeled samples for training,a semi-supervised modulation recognition algorithm based on a small number of labeled samples and a large number of unlabeled samples is proposed.The algorithm is trained by a CNN that includes regularization methods.First,the error between the forward prediction of labeled sample and the label is calculated as supervised loss component,and secondly,the error between the twice forward predictions is calculated for each training samples as unsupervised loss.Finally,the weighted summation of supervised loss component and unsupervised loss component is adopted as the total loss function to drive network training.The simulation results show that the proposed algorithm can effectively utilize unlabeled samples to participate in network training to improve modulation recognition accuracy in the scenario where there are only a few labeled samples.