Research On Spectrum Prediction Algorithm In Cognitive Radio

With The rapid development of wireless communication field,spectrum is more and more crowded,and the spectrum resource becomes insufficient.Dr.Mitola puts forward to Cognitive radio which can perceive the current environment of communication.Cognitive radio can use decision tree algorithm to dynamically take advantege of idle spectrum according to environment,relevant parameters and frequency,so as to optimize the system performance and improve service rate.Spectrum prediction is a technology which can predict the future situation of PU(Primary User)occupied channel by mining and analyzing the historical data correlation of spectrum.Spectrum prediction can also reduce the collision probability of SU(Secondary User)and PU,energy and time loss,which can improve the performance of cognitive radio system as well,and finally realizes the full utilization of spectrum utilization.With the emergence of deep learning,neural networks can solve nonlinear problems without prior information of channel state.They are widely used in spectrum prediction.Firstly,the development status of spectrum prediction technology is introduced in detail.Then,this thesis analyses and contrasts the common spectrum prediction models,such as BP neural network,regression analysis and markov prediction algorithm,so as to emphasizes the importance of prediction for cognitive radio spectrum.This thesis elaborates the knowledge of queuing theory model and markov model,and simulates the user behaviors based on the two models by Matlab.The queuing theory model is used to model the cognitive radio channel,and the channel state data is generated as the input of the neural network.Firstly,LSTM(Long Short-Term Memory)neural network is used for data training by changing the channel environment and training parameters,the prediction performance of LSTM neural network is compared with traditional BP neural network.It is proved that LSTM algorithm can effectively solve the problems of fast convergence and gradient explosion in the training process of traditional neural network.Later,CNN(Convolutional Neural Networks)neural network is added for comparison with LSTM to analyze their advantages and disadvantages.Since the single-channel prediction is made before,the correlation among multiple channels of cognitive radio network can not be analyzed.In this thesis,the frequency spectrum data generated by M/M/N is used as simulation data to carry out multi-channels joint prediction on the basis of CNN,which analyses the influence of time domain correlation between multiple channels on the prediction error.It is proved that multi-channels joint prediction can effectively improve the prediction accuracy.