摘要
针对传统卷积神经网络(CNN)频谱感知方法没有充分利用特征图信息并且提取特征图的能力受限于浅层的网络结构等问题,通过在传统CNN频谱感知方法中添加密集连接,实现特征图信息重利用,同时在密集单元的两端加入捷径连接,实现更深层的网络训练,进而提出一种基于残差密集网络(ResDenNet)的频谱感知方法。该方法将频谱感知问题映射为图像二分类问题,首先对接收信号分割成矩阵并归一化灰度处理,得到的灰度图像作为网络的输入,然后通过密集学习和残差学习训练网络,最后将在线数据输入ResDenNet中,完成基于图像分类的频谱感知。数值实验表明,所提方法优于传统频谱感知方法,在信噪比低至-19 dB时,所提方法检测概率仍高达0.96,虚警概率低至0.1,同时具有更好的泛化能力。
Aiming at the problem that the traditional spectrum sensing method based on convolutional neural network(CNN) did not make full use of image feature and the ability of extracting the image feature was limited by the shallow network structure, a spectrum sensing method based on the residual dense network(ResDenNet) was proposed.By adding dense connections in the traditional neural network, the information reuse of the image feature was achieved.Meanwhile, shortcut connections were employed at both ends of the dense unit to implement deeper network training.The spectrum sensing problem was transformed into the image binary classification problem. Firstly, the received signals were integrated into a matrix, which was normalized and transformed by gray level. The obtained gray level images were used as the input of the network. Then, the network was trained through dense learning and residual learning. Finally, the online data was input into the ResDenNet and spectrum sensing was implemented based on image classification. The numerical experiments show that the proposed method is superior to the traditional ones in terms of performance. When the SNR is as low as-19 dB, the detection probability of the proposed method is still high up to 0.96 with a low false alarm probability of 0.1, while a better generalization ability is displayed.
作者
盖建新
薛宪峰
南瑞祥
吴静谊
GAI Jianxin;XUE Xianfeng;NAN Ruixiang;WU Jingyi(The Higher Educational Key Laboratory for Measuring&Control Technology and Instrumentations of Heilongjiang Province,Harbin University of Science and Technology,Harbin 150080,China)
出处
《通信学报》
EI
CSCD
北大核心
2021年第12期182-191,共10页
Journal on Communications
基金
国家自然科学基金资助项目(No.61501150)
黑龙江省自然科学基金资助项目(No.QC2014C074)
黑龙江省省属本科高校基本科研业务费基金资助项目(No.2018-KYYWF-1656)。