基于注意力多尺度特征融合的频谱感知方法

A Spectrum Sensing Method Based on Attention Multi-scale Feature Fusion

随着无线通信和物联网(Internet of Things,IoT)设备的迅速增长,频谱资源短缺和电磁环境复杂性成为通信系统面临的挑战。频谱感知作为频谱管理的一项关键技术,使频谱资源短缺问题得到了缓解。卷积神经网络(Convolutional Neural Network,CNN)作为深度学习的代表,近年来在频谱感知任务中表现出色。为解决CNN实现频谱感知任务时感受野受限、多尺度信息融合和空间信息捕获等方面存在局限性的问题,提出了一种注意力多尺度特征融合CNN(Attention-Multi-Scale Feature Extraction-CNN,AMFE-CNN)模型,包含多尺度特征提取和注意力模块。多尺度特征提取利用膨胀卷积获取更大的时频感受野,注意力模块通过多重卷积和池化操作关注时频图的空间信息。实验结果表明,该模型在频谱感知任务中表现出色,提高了检测性能和泛化能力。

With the rapid growth of wireless communication and Internet of Things(IoT)devices,the scarcity of spectrum resources and the complexity of electromagnetic environments have emerged as significant challenges for communication systems.Spectrum sensing,as a key technology in the spectrum management,has alleviated the issue of spectrum resource scarcity.As a representative of deep learning,Convolutional Neural Network(CNN)have demonstrated remarkable performance in spectrum sensing tasks in recent years.To address limitations such as restricted receptive fields,multi-scale information fusion and spatial information capture in CNN-based spectrum sensing tasks,an Attention-Multi-Scale Feature Extraction-CNN(AMFE-CNN)model is proposed,which includes multi-scale feature extraction and attention modules.The multi-scale feature extraction utilizes dilated convolutions to capture a larger time-frequency receptive field,while the attention module focuses on spatial information in the time-frequency map through multiple convolution and pooling operations.Experimental results show that the proposed model performs excellently in spectrum sensing tasks,improving both detection performance and generalization capabilities.