Research On Spectrum Sensing Method Based On Residual Network

Spectrum sensing is the first step to solve the problem of low utilization of spectrum resources.The excellent performance of deep learning in classification problems brings new inspiration to the field of spectrum sensing.The existing spectrum sensing methods have the problems of insufficient utilization of signal characteristics,low sensing efficiency,sensitivity to noise uncertainty,and network degradation in deep networks.Aiming at the above problems,this paper studies the application of residual network algorithm in deep learning to complete the spectrum sensing process.Aiming at the problem that the received signal preprocessing method in the existing single-user spectrum sensing method is difficult to highlight the timefrequency characteristics of the spectrum signal,which is not conducive to feature extraction and the degradation of CNN network in the process of network deepening,the signal preprocessing and residual network ideas are applied to the single-user spectrum sensing model.A single-user spectrum sensing method is proposed.Firstly,the original received signal sample is preprocessed into a time-frequency matrix by short-time Fourier transform to highlight the time-frequency characteristics of the signal.Then,the preprocessed signal sample is used as the input of the designed improved residual network,and the shortcut connection is introduced in the design of the network model.At the same time,the dropout layer is added to the design of the improved residual module,and the over-fitting phenomenon is alleviated by discarding some hidden nodes in the network training.Through comprehensive experimental comparison,the proposed single-user spectrum sensing method has better performance than the traditional spectrum sensing method in terms of detection performance,robustness to noise power uncertainty,generalization ability under different modulation modes and unknown noise conditions.Under the condition of-19 d B SNR,the detection probability reaches 0.945 when the false alarm probability is 0.01.Due to the large number of secondary users in cooperative spectrum sensing,the existing cooperative spectrum sensing methods have the disadvantages of low detection efficiency and low detection probability under low signal-to-noise ratio.In this paper,a cooperative spectrum sensing method is proposed based on the characteristics of the received signal in cooperative spectrum sensing.By Cholesky decomposition of the covariance matrix of the original received signal,the upper triangular feature matrix is obtained,and the gray image is obtained by normalization.The extended convolution idea is used to design the extended residual network.By changing the expansion rate of the convolution kernel in the convolution process,the multi-scale information is obtained,and the information correlation between the secondary users in cooperative spectrum sensing is fully utilized.The experimental results show that compared with the existing cooperative spectrum sensing methods,the proposed method has more prominent performance in detection performance,model stability and other aspects.At the same time,it has higher detection probability in low SNR environment.When the SNR is-25 d B and the detection probability reaches 1,the false alarm probability is only 0.1.