Research On Spectrum Sensing Algorithm Based On Residual Neural Network

The traditional spectrum management is to allocate fixed spectrum resources to specific users while others can't use it,which is a waste of spectrum resources.The technology of cognitive radio is raised to improve the comprehensive utilization ratio of spectrum resources.It is an efficient technology for management and dynamic allocation of spectrum resource,and it is also a basic technology.The main purpose of this paper is to solve the problem of spectrum sensing based on the theory and algorithm in deep learning.On the one hand,this paper is constructed of the signal features by a simple stacked method for the spectrum sensing problem of the single user.According to the existing improved convolutional neural network methods,two network models,the residual neural network with increasing network depth and the inception network with increasing network width,are developed to train the spectrum sensing classifier.On the other hand,due to the cost of spectrum sensing for multiple secondary users is very expensive when collecting data information,a cooperative spectrum sensing with fewer secondary users is proposed,which uses data collected by a small number of secondary users to build signal features based on an idea of stack-based refactoring to get the cooperative spectrum sensing classifier by training the network.According to simulation experiment,whether it is in the spectrum sensing of single user or the cooperative spectrum sensing of a small number of secondary users,it is found that the detection performance of the two intelligent spectrum sensing is better than that of traditional spectrum sensing,and the performance of residual neural network algorithm is better.Because the model of residual neural network has a plenty of parameters,it is necessary to integrate residual neural network with Inception network and propose a fusion algorithm suitable for spectrum sensing to reduce the amount of parameters and the memory required in network training.The simulation experiment showed that the algorithm can reduce the amount of parameters while maintaining good performance.In the spectrum sensing of single user,the performance of the fusion algorithm is not as good as that of the residual neural network algorithm,but the detection probability of fusion algorithm can reach 89%when the signal-to-noise ratio(SNR)is-18dB.In the cooperative spectrum sensing of a small number of secondary users,the fusion algorithm has the similar performance as residual neural network algorithm,and the detection probability of which can reach 90.3%when the SNR is-16dB.