摘要
在认知无线电(Cognitive Radio, CR)中,频谱感知(Spectrum Sensing, SS)是支持动态频谱分配、提高频谱利用率的关键技术。传统的SS方法容易受到噪声不确定的影响,导致在低信噪比(Signal to Noise Ratio, SNR)环境下检测准确率较低且计算参数量较大。针对这些问题,提出了基于信号处理(Signal Processing, SP)特征的鲁棒多尺度神经网络(Robust-Multiscale Neural Network, R-MsNN)的SS方法,结合多尺度神经网络和门控循环单元(Gated Recurrent Unit, GRU)的优点,有效地解决了SS中面临的挑战。多尺度神经网络从底层到高层逐渐提取抽象的特征,以增强信号识别的鲁棒性。GRU有选择性地保留和遗忘过去时间信息,从而更好地捕获长期依赖性和时间关系。为了验证R-MsNN的泛化能力,实验将从广义高斯分布(Generalized Gaussian Distribution, GGD)生成的噪声样本作为噪声模型1,以及从未占用的调频广播信道中收集到的实验数据作为噪声模型2的环境下,分别与不同的深度神经网络(Deep Neural Network, DNN)架构进行了SS性能比较。实验结果表明,采用组合SP特征训练的R-MsNN的平均检测概率与最优模型相比,在4种不同参数下的噪声模型1中分别提高了1.74%、2.55%、2.08%、1.59%,在噪声模型2中提高了1.72%。此外,与GRU相比,R-MsNN的参数量减少了一半。由此说明,采用组合SP特征训练的R-MsNN在多种复杂噪声环境下均具有很强的鲁棒性,且能够满足SS任务中高检测概率和低参数量的双重需求。
features from the bottom layer to the top layer to enhance the robustness of signal recognition.GRU selectively retains and forgets past temporal information,thereby better capturing long-term dependencies and temporal relationships.To validate the generalization capability of R-MsNN,the SS performance of R-MsNN is compared with that of different Deep Neural Network(DNN)architectures in two different noise models:one generated from a Generalized Gaussian Distribution(GGD)as noise model 1,and experimental data collected from unoccupied frequency modulation broadcasting channels as noise model 2.The experimental results demonstrate that,compared to the optimal model,R-MsNN trained with a combination of SP features achieved an average detection probability increase of 1.74%,2.55%,2.08%,and 1.59%for four different noise models 1,and a 1.72%increase for noise model 2.Additionally,compared to GRU,R-MsNN has half the number of parameters.This indicates that R-MsNN trained with a combination of SP features,exhibits robustness in a variety of complex noise environments and can meet the dual requirements of high detection probability and low parameter number in SS tasks.
作者
孟水仙
闫森
王树彬
MENG Shuixian;YAN Sen;WANG Shubin(Inner Mongolia Autonomous Region Radio Monitoring Station,Hohhot 010011,China;School of Electronic Information Engineering,Inner Mongolia University,Hohhot 010021,China)
出处
《无线电工程》
2024年第8期1854-1861,共8页
Radio Engineering
基金
国家自然科学基金(62361048)。
