Analysis Of Primary User Signal Based On Deep Learning

In recent years,the rapid growth of wireless services has inspired the lack of wireless spectrum resources.To solve this problem,cognitive radio is proposed to achieve spectrum sharing by allowing secondary users reuse the licensed frequency bands without causing interference to primary user.The spectrum sharing can be performed more effectively with the knowledge of primary user signals.On the other hand,with the evolution of computer hardware and growth of big data,deep learning has been becoming a research hotspot currently.This thesis focuses on the application of deep learning in cognitive radio and uses deep neural networks to analyze the signals of primary user,which will benefit spectrum sharing and alleviate the lack of spectrum resources.The thesis firstly introduces the background,definition and key technologies of cognitive radio and describes deep learning and convolution neural networks in details at the same time.Secondly,the system model is given and the expression of received primary user signal is derived.Then,a series of data preprocessing methods suitable for deep neural networks are proposed.Moreover,the best method is selected for data representation of this thesis via experiments.Thirdly,the task of digital modulation recognition is discussed.Based on convolutional neural networks and preprocessed constellation data,a classifier for primary user modulation classification is trained.Compared with conventional cumulants and support vector machine based classification algorithms,this classifier avoids manual feature selection and has higher classification accuracy under low signal-to-noise conditions.Fourthly,this thesis studies the cognitive radio scenarios with inference and designs a deep learning based signal analysis algorithm for both primary user signal and inference.This algorithm jointly considers multiple signals,which can not only identify primary user signal but also analyze the modulation type of inference.Simulation results show that the designed algorithm has satisfactory performance in various scenarios.Finally,considering the task of signal-to-noise estimation for primary user signal,this thesis investigates a deep learning based estimation approach.This approach exploits the different patterns of constellation images at different signal-to-noise ratios to complete model training and result inferring.Compared with the existing M₂M₄ and SVR based methods,the approach of this thesis can reduce the normalized mean square error in signal-to-noise estimation.