小样本学习驱动的无线频谱状态感知  

Wireless Spectrum Status Sensing Driven by Few-Shot Learning

作　　者：申滨[1,2] 李月 王欣 王紫昕 SHEN Bin;LI Yue;WANG Xin;WANG Zixin(School of Communications and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China;Chongqing Key Laboratory of Mobile Communications Technology,Chongqing 400065,China)

机构地区：[1]重庆邮电大学通信与信息工程学院,重庆400065 [2]移动通信技术重庆市重点实验室,重庆400065

出　　处：《电子与信息学报》2024年第4期1231-1239,共9页Journal of Electronics & Information Technology

基　　金：国家自然科学基金(62371082)。

摘　　要：无线频谱状态感知是实现无线频谱资源高效利用及各种用频系统和谐共存的先决条件之一。针对复杂无线传播环境下获取的频谱观测往往存在数据稀疏性、数据类别分布不稳定、标记数据严重不足的情况,该文提出基于插值和小样本学习(FSL)分类的无线频谱状态感知方法。首先,对捕获的稀疏频谱观测数据插值,构建频谱状态地图,作为频谱状态分类器的输入数据。其次,针对频谱数据类别分布不稳定、数据量严重不足的问题,基于小样本学习方法,利用嵌入模块和度量模块协同工作,以实现快速精确的频谱状态分类。具体地,利用嵌入模块将频谱数据映射到嵌入空间,提取频谱数据中的隐含特征;在度量模块的设计中,分别提出基于原型和基于样例的两种类别表示方式,通过计算待分类样本与类别之间的相似度判断待分类样本类别。最后,为了确保分类模型克服测试样本数量少导致过拟合问题,设置A-way B-shot任务训练模型。仿真结果表明,与传统机器学习方法相比,本文模型可以在低信噪比条件下进行精准分类;同时,在测试集样本数很少的情况下,或者在测试集中出现在训练集从未见到的新类时,所训练的模型也可以精准快速判别无线频谱的场景类别。Wireless spectrum status sensing is one of the prerequisites for achieving efficient utilization of spectrum resources and harmonious coexistence among systems.A spectrum sensing scheme based on interpolation and Few-Shot Learning(FSL)classification is proposed to address the sparsity of spectrum data,unstable distribution of data categories,and severe shortage of labeled data in complex wireless propagation environments.Firstly,the sparsely distributed observation data is interpolated and a spectral status map is constructed as the input data to the spectral status classifier.Then,for the cases where the distributions of data categories are unstable and the amount of data is severely insufficient,a few-shot learning-based classification algorithm is proposed,incorporating the embedding modules and measurement modules to realize fast and accurate spectrum status classification.Specifically,the embedding module is used to map spectral data to the embedding space and extract hidden image features from the spectral data.In the measurement module,two category representation methods,prototype-based and sample-based,are proposed to determine the category of the samples by calculating the similarity between the samples and the categories.Finally,an A-way B-shot task training model is set to ensure that the classification model will not cause overfitting problems due to the small number of test samples.Simulation results show that compared with traditional machine learning methods,the proposed model can achieve accurate classification under low signal-to-noise ratio conditions.In addition,it can quickly distinguish the categories of radiation source activity scenarios even when the number of samples in the test set is small or when new classes that have never been seen in the training set appear in the test set.

关 键 词：频谱状态感知 频谱状态地图 插值 小样本学习 

分 类 号：TN911.73[电子电信—通信与信息系统]