- 无标题文档
查看论文信息

论文中文题名:

 基于残差神经网络和注意力机制的 OFDM 频谱感知    

姓名:

 孟琦峰    

学号:

 21207040026    

保密级别:

 公开    

论文语种:

 chi    

学科代码:

 081002    

学科名称:

 工学 - 信息与通信工程 - 信号与信息处理    

学生类型:

 硕士    

学位级别:

 工学硕士    

学位年度:

 2024    

培养单位:

 西安科技大学    

院系:

 通信与信息工程学院    

专业:

 信息与通信工程    

研究方向:

 无线通信技术    

第一导师姓名:

 王安义    

第一导师单位:

 西安科技大学    

论文提交日期:

 2024-06-12    

论文答辩日期:

 2024-05-31    

论文外文题名:

 OFDM spectrum sensing based on residual neural networks and attention mechanisms    

论文中文关键词:

 认知无线电 ; 频谱感知 ; 深度学习 ; 残差神经网络 ; 卷积注意力机制    

论文外文关键词:

 cognitive radio ; spectrum sensing ; deep learning ; residual neural networks ; convolutional block attention module    

论文中文摘要:

随着无线设备数量的迅速增加,频谱资源日益紧张。鉴于频谱的珍贵性,必须寻求有效利用的途径。认知无线电(CR)是缓解这一难题的解决方案之一,它能够识别和访问许可频谱中未被充分利用的频谱空洞。频谱感知作为CR实现其各项功能的基础,使得当主用户(PU)占用特定频段时,次用户(SU)能够避让此频段。本文的研究重点在于运用深度学习领域的图像处理理论与算法,解决单节点和协作频谱感知算法的问题。

(1)针对传统卷积神经网络(CNN)存在的问题,如梯度消失和特征提取不足,为提升其在频谱感知中的特征提取能力,采用残差神经网络和卷积注意力机制模块构建了基准模型,形成单节点频谱感知算法。在此基础上,通过将残差神经网络和卷积注意力机制结合起来,提出了一种名为ResNet-CBAM的频谱感知方法。该方法将频谱感知问题转化为图像的二分类问题。通过分析OFDM信号的循环自相关特征,并对其进行灰度处理,生成循环自相关灰度图像。再通过训练改进后的残差神经网络提取灰度图像的深层特征,并用测试集验证训练后的频谱感知模型。

(2)针对CNN模型对特征的提取存在限制,通过增加网络层数以提高特征提取能力可能导致在训练过程中发生梯度消失现象和单节点频谱感知存在阴影衰落和隐藏终端等问题。因此,提出了一种基于残差密集网络和卷积注意力机制(RDN-CBAM)的OFDM协作频谱感知方法。该方法通过将信号的时域信息堆叠构造成二维矩阵,对矩阵进行归一化,映射为灰度图像,并将灰度图像划分为测试集和训练集。利用训练集对神经网络进行训练,以提取灰度图像的深层特征。再将测试集输入训练好的神经网络频谱感知模型中,完成频谱感知,把问题转化为图像分类任务。

实验结果显示,针对单节点频谱感知,所提出的ResNet-CBAM频谱感知方法在性能上优于CNN和SVM方法。在信噪比降至-20dB时,该方法的检测概率达到87%,虚警概率仅为1%。面向协作用户的频谱感知场景中,在信噪比降至-20dB时,本文提出的RDN-CBAM谱感知方法展现出卓越的性能,检测概率高达90%,虚警概率仅为1%。

论文外文摘要:

With the rapid increase in the number of wireless devices, spectrum resources are becoming increasingly scarce. Given the preciousness of spectrum, effective utilization approaches must be sought. Cognitive Radio (CR) is one of the solutions to alleviate this problem, as it can identify and access spectrum white spaces that are not fully utilized within licensed spectrum. Spectrum sensing, as the basis for CR to achieve its various functions, allows secondary users (SU) to vacate specific frequency bands when primary users (PU) occupy them. The focus of this study is on applying theories and algorithms of image processing in the field of deep learning to address issues in single-node and collaborative spectrum sensing algorithms.

(1) In response to the issues present in traditional Convolutional Neural Networks (CNNs), such as gradient vanishing and inadequate feature extraction, a baseline model was constructed using Residual Neural Networks and Convolutional Block Attention Module (CBAM) to enhance its feature extraction capability in spectrum perception, forming a single-node spectrum perception algorithm. Building upon this foundation, a spectrum perception method named ResNet-CBAM was proposed by combining Residual Neural Networks with the Convolutional Block Attention Module. This method transforms the spectrum perception problem into a binary classification problem for images. By analyzing the cyclic autocorrelation features of Orthogonal Frequency Division Multiplexing (OFDM) signals and processing them into grayscale, cyclic autocorrelation grayscale images are generated. Subsequently, by training an improved Residual Neural Network to extract deep features from grayscale images and validating the trained spectrum perception model using a test set.

(2) Recognizing the limitations of CNN models in feature extraction and the potential issues arising from increasing network depth to enhance feature extraction, such as gradient vanishing during training and the existence of shadow fading and hidden terminals in

 

single-node spectrum perception, a novel OFDM cooperative spectrum perception method named Residual Dense Network with Convolutional Attention Mechanism (RDN-CBAM) is proposed. This method constructs a two-dimensional matrix by stacking the temporal information of signals, normalizes the matrix, maps it into a grayscale image, and divides the grayscale image into training and test sets. The neural network is then trained using the training set to extract deep features from the grayscale image. Finally, the test set is input into the trained neural network spectrum perception model to complete spectrum perception, thereby transforming the problem into an image classification task.

The experimental results demonstrate that, for single-node spectrum sensing, the proposed ResNet-CBAM spectrum sensing method outperforms CNN and SVM methods in terms of performance. When the signal-to-noise ratio drops to -20dB, this method achieves a detection probability of 87%, with a false alarm probability of only 1%. In the spectrum sensing scenario targeting collaborative users, when the signal-to-noise ratio decreases to -20dB, the RDN-CBAM spectrum sensing method proposed in this paper exhibits outstanding performance, with a detection probability reaching 90% and a false alarm probability of only 1%.

中图分类号:

 TM25    

开放日期:

 2024-06-12    

无标题文档

   建议浏览器: 谷歌 火狐 360请用极速模式,双核浏览器请用极速模式