Research On Modulation Recognition And Parameter Estimation Of Communication Signal Based On Deep Learning

Modulation identification and parameter estimation are key technologies in communication,which can identify the modulation information of the transmitted signal and provide the basis for subsequent demodulation operations.In cognitive radio systems,modulation identification technology can be used to avoid wasting spectrum as much as possible.In non-cooperative communication,accurate modulation recognition plays a key role in both modern warfare and social development.Deep learning has grown rapidly in recent years due to the explosive increase in computing speed of parallel processors such as Gpus.Compared with traditional methods,deep learning does not need to manually extract features,but set judgment rules to achieve better classification results.Therefore,the application of deep learning technology to modulation recognition and parameter estimation is conducive to improve the accuracy of recognition.The main content of this thesis is to study the application of neural network in modulation type identification,parameter estimation and multi-task learning based on deep learning technology.The main content is as follows:(1)Based on the layer structure and optimization algorithm of common neural networks,a convolutional neural network structure suitable for modulation recognition is designed,and the influence of different channel parameters on modulation recognition accuracy is discussed by simulating different channel environments.(2)The structure of modulation recognition network is optimized,and training and testing are carried out through residual connection,multi-scale convolution,introduction of timing branch,etc.,and the optimal network model is obtained by comparing the test results.(3)The modulation signal parameter estimation is realized.On the basis of the identification network,the two parameters of symbol rate and relative frequency offset are estimated accurately respectively by designing appropriate cost function and activation function.(4)Based on the correlation between modulation classification and parameter estimation task,the class output and parameter estimation of modulation signal are realized simultaneously by designing a multi-task learning network.