协方差矩阵分解的CNN协作频谱感知

Cooperative Spectrum Sensing with Covariance Matrix Decomposition and CNN

针对信号采样协方差矩阵主特征信息未被充分利用及检测门限不稳等问题,提出了协方差分解的卷积神经网络(Convolutional Neural Network,CNN)协作频谱感知方案。首先,对接收信号采样矩阵的协方差矩阵Cholesky分解后,进行统计量的计算,以充分提取两种信号的特征。其次,累计多个次用户所得统计量,组成统计量特征矩阵,以提高协作性及检测精度。最后,利用CNN对高维矩阵的特征提取性,经训练数据集训练得到CNN分类模型,并用测试数据集预测获得频谱结果。实验表明:相对支持向量机(SVM)、传统CNN等算法,所提算法训练精度高,延时小;当虚警概率为0.1,且信噪比为-15 dB时,所提方案检测概率分别比传统CNN和SVM类方法高约60%和69%。

According to the underutilization of main characteristic in signal sampling covariance matrix and an unstable detection threshold,a cooperative spectrum sensing method is proposed by covariance matrix decomposition of convolutional neural networks(CNN).Firstly,the covariance matrix of the received signal sampling matrix is Cholesky decomposed and the statistics are calculated to fully extract the characteristics of the two signals.Secondly,the statistics obtained by multiple secondary users are accumulated to form a statistical characteristic matrix for improving both collaboration and detection accuracy.Finally,by using the feature extraction ability of CNN to high-dimensional matrix,the CNN classification model is trained by the training dataset,and the spectrum results are predicted by the test dataset.Experimental results show that the proposed algorithm has higher training accuracy and shorter experimental time compared with SVM,traditional CNN and other algorithms.Given false alarm probability of 0.1,the detection probability of the proposed scheme outperforms the traditional CNN and SVM by 60%and 69%at SNR of-15 dB,respectively.