Research On Spectrum Sensing Technology Based On Deep Learning

Spectrum sensing is one of the important parts of cognitive radio(CR)to realize cognitive function.In order to improve spectrum utilization,deep learning spectrum sensing technology has become an important research direction in CR field.The convolutional neural network(CNN)spectrum sensing method,as one of the representative methods,can accomplish efficient and fast spectrum sensing tasks.However,the CNN spectrum sensing method is still faced with problems such as insufficient feature extraction ability,network structure to be optimized,and spectrum sensing accuracy to be improved.In view of the above problems,this dissertation makes an in-depth study on the single node and cooperative spectrum sensing method with the advantage of deep learning in image processing.The traditional single node spectrum sensing method based on CNN has the single-stream convolutional network structure and the shallow network structure which limits the ability of drawing the primary user(PU)feature.Aiming at these problems,the spectrum sensing method based on the residual cellular network(ResCelNet)is proposed in this dissertation,the dissertation proposes the ResCelNet.The two-stream convolution improves the feature extraction ability,the addition operation enhances the micro-feature information,and the residual learning is easy to train the deep spectrum sensing network.In this method,the received signals sampled by secondary users(SUs)are first integrated into a matrix and normalized into grayscale images,which are used as dataset of ResCelNet network.Then feature information of gray images are extracted from offline training data by two-stream convolution and residual learning.Finally,ResCelNet is used to identify and classify test data.To solve the problem that the feature extraction ability of traditional CNN cooperative spectrum sensing method is limited by shallow network structure and it is difficult to avoid gradient disappearance in deep network,the shortcut connections are added to traditional CNN spectrum sensing method to realize deeper network training,then a cooperative spectrum sensing method based on residual network(ResNet)is proposed in this dissertation.Firstly,the received signals sampled by multiple SUs are processed by covariance and normalized gray scale as the input of ResNet spectrum sensing network.Then,residual learning is used to train ResNet to extract the deep features of feature mapping graph.Finally,ResNet is used for spectrum sensing for test data.The numerical experiments suggest that the proposed ResCelNet method is superior to CNN and SVM when spectrum sensing is oriented toward single node.When the SNR is as low as-19 dB,the detection probability of the proposed method is 0.98 and the false alarm probability is 0.1,and it is suitable for different fading channels,different modulation modes and noise uncertainty.When spectrum sensing is oriented toward cooperative users,when the SNR is as low as-19 dB,the detection probability of ResNet method in this dissertation is up to 1.00 and the false alarm probability is as low as 0.01,and the hidden terminal problem is solved.