Research On Cooperative Spectrum Sensing And Prediction Technology Based On Neural Network

With the increase of system capacity,spectrum resources are getting more and more scarce.The combination of Cognitive Radio(CR)and Non-orthogonal Multiple Access(NOMA)— CR-NOMA,can be used to allow more Primary Users(PU)to reuse the same resource block,and the opportunities for Secondary Users(SU)to dynamically access the spectrum are increased.However,the wireless environment in the CR-NOMA system is very complex,which brings great difficulties to the spectrum sensing and spectrum prediction of the SU.Traditional spectrum sensing algorithms and spectrum prediction algorithms are comprehensively analyzed in this thesis.Firstly,the spectrum sensing technology in CR system and CR-NOMA system is studied respectively.The principle of each algorithm is analyzed,and the advantages and disadvantages of each algorithm are pointed out.Secondly,various spectrum prediction algorithms based on neural network in CR system are studied.Their limitations are expounded in combination with the network structure of each algorithm.In view of the shortcomings of traditional algorithms,the spectrum sensing and spectrum prediction technologies in CR-NOMA system are involved in this thesis,which mainly include the following three aspects:(1)In view of the large number of computational parameters and low spectrum sensing accuracy of traditional spectrum sensing network,the advantages of the Shuffle Net and the Dense Net are combined in this thesis.A new cooperative spectrum sensing neural network— Shuffle Dense Net is designed.This network has few computational parameters,and its lightweight structure enables it to run directly on SU with low hardware conditions.In addition,the new network has a strong ability to transmit features in deep network,which can effectively improve the spectrum sensing accuracy in CR-NOMA system.(2)In order to deal with the complex wireless environment in the CR-NOMA system and the nonfixation of SU locations,a new distributed spectrum sensing scheme based on Shuffle Dense Net is proposed in this thesis.Firstly,the secondary user’s energy sensing results and their geographical location features are integrated in the cluster head.Then,the Shuffle Dense Net is used for spectrum sensing based on the features.The transmission delay between the base station and SU is avoided.And the spectrum access efficiency is improved in this distributed spectrum sensing model.The simulation results show that the proposed scheme can greatly reduce the number of the network computational parameters and improve the efficiency of spectrum sensing while ensuring the accuracy of spectrum sensing.The scheme can complete spectrum sensing task in scenarios where SU has mobility,and has wide applicability.(3)To alleviate the problems of low prediction dimension,narrow prediction frequency band and insufficient prediction accuracy of traditional algorithms,a spectrum prediction scheme based on Convolutional Gated Recurrent Unit(Conv GRU)Network is proposed in this thesis.The spectrum occupancy is integrated into a large-area spectrum occupancy image,and then the Conv GRU network is used to extract multi-dimensional features for multi-dimensional long-term spectrum prediction.In this scheme,a new decoder is designed,which can improve the network’s ability of extracting temporal features by integrating all steps’ output features and using them as the input of the prediction layer.The problem of insufficient long-term dependence of traditiolal spectrum prediction algorithms is solved.Simulation shows that this scheme improves the speed and the accuracy of prediction by combining the three-dimensional features of time,space and frequency for multi-slot long-term spectrum prediction.