A Deep Learning Algorithm With High Accuracy And Low Complexity For Modulation Recognition

The rapidly developing wireless communication technologies require a variety of modulation methods for information transmission.However,complex and changeable wireless signal modulation methods have brought huge challenges to spectrum management,signal capture,communication countermeasures,etc.Therefore,it is necessary to design an efficient automatic modulation recognition algorithm,which is capable of accurately identifying the modulated wireless signals in a complex communication environment.An ideal modulation classifier should be both automatic and universal,while requiring minimal complexity for high accuracy recognition.The goal of this thesis is to design an automatically recognition algorithm,which has high accuracy and easy to deploy.Based on the deep learning neural networks,this thesis studies the recognition performance of a large-scale dataset consisting of eight digital modulated signals and two analog modulated signals.To pursue the high recognition accuracy using a modulation classifier,the effect of different representation forms of modulated signal on recognition accuracy is considered as well.Five conventional modulation recognition algorithms based on classical neural networks are thoroughly investigated firstly,and they are trained using two forms of signation representations.These classic neural networks are Convolutional Neural Networks,Recurrent Neural Networks,Densely Connected Networks,Long Short-Term Memory and Convolutional Long short-term memory fully connected Deep Neural Networks.From simulation experiments,we found that conventional neural network algorithms cannot accurately distinguish between lower-order and high-order QAM signals in the dataset.In order to further improve the recognition performance of the modulation classifier,this thesis proposes a novel neural network termed as Long short-term memory Convolutional fully connected Deep Neural Network,which gives full play to the advantages of Long Short-Term Memory,puts forward time domain features from modulated signals,and transmits them to the Convolutional Neural Network,and finally uses the full-connected network to map distributed features to the correspond sample space for recognition prediction.Numerous simulation results are provided to show that the network trained by polar form signals solves the problem of confusion between lower-order and high-order QAM signals,and the recognition accuracy in high signal-to-noise ratio can reach to 93.5%,which is better than state-of-the-art algorithms.As a hybrid model,the Long short-term memory Convolutional fully connected Deep Neural Network has high complexity,which means that the training of the model requires a lot of computation and storage requirements.To fix that,this thesis has made a lot of structural improvements and simplifications.The simulation results show that although the compact model has a slight loss of recognition performance,its trainable parameters are only 0.15% of the original model and the training cost is greatly reduced.This thesis studies a variety of ways to reduce the amount of model training data to further reduce training costs based on the compact model.According to the simulation results,both the resampling of modulated signal sample points by using uniform subsampling with rate of 1/2 and sampling training signal samples by using the Near Miss-3 under-sampling algorithm can achieve high recognition performance under high signal-to-noise ratio of the model,and achieve the goal of saving training cost.Finally,this thesis combines the uniform subsampling of 1/2 rate to resampling the modulated signal sample points and Near Miss-3 under-sampling algorithm to sample the training samples to train the compact model we designed.An idealized automatically classifier,which has 92.7% accuracy at high signal-to-noise ratio,is obtained using only 30% of the data of original training dataset.