吴赛 王智慧 邵炜平 林春生 郑伟军 杨德龙



摘要 频谱分析的关键在于准确识别信号的调制方式,而常用的自动调制识别方法在低信噪比下的识别率低,并且能够识别的信号调制方式种类数少。基于此种情况,提出了一种基于循环谱和改进的深度神经网络的频谱分析方法。该方法使用卷积神经网络、长短时记忆和深度神经网络相结合的神经网络(CLDNN)并将循环谱特征作为该网络的原始输入特征。仿真结果显示所提出的方法在信噪比为-2 dB时能够达到90%的识别准确率,极大的提高了低信噪比情况下的信号识别性能。
关 键 词 自动调制分类;循环谱;神经网络;频谱感知
中图分类号 TN911.7 文献标志码 A
Cyclic spectrum and improved deep-neural-network based spectrum analysis method
WU Sai1, WANG Zhihui1, SHAO Weiping2, LIN Chunsheng3,
ZHENG Weijun2, YANG Denglong1
(1. China Electric Power Research Institute, Beijing 100192, China; 2. State Grid Zhejiang Electric Power Co., LTD, Hangzhou, Zhejiang 310007, China; 3. School of Information and Communication Engineering, Beijing University of Posts and Telecommunications, Beijing 100876, China )
Abstract The essential of the accurate spectrum sensing is the automatic modulation classification. The commonly used automatic modulation classification (AMC) schemes has inferior classification performance at low signal-to-noise ratio (SNR) scenario and only few modulation formats can be identified. On this situation, a novel cyclic spectrum and improved deep-neural-network based modulation classification scheme is proposed, where convolutional neural network and a long short-term memory assisted deep neural network (CLDNN) structure is utilized and the cyclic spectrum features are the inputs of the network. The simulation results verify that the proposed scheme achieves 90% recognition accuracy at -2 dB SNR, which has greatly improved the recognition accuracy at low SNR and outperforms other recent methods.
Key words automatic modulation classification; cyclic spectrum; neural network; spectrum sensing
無线通信以即时实现固定与移动、移动与移动的无障碍、无缝隙信息覆盖为目标,造成了无线通信业务的指数级增长。然而,承载业务的无线电频谱已经基本分配殆尽且十分拥挤,很难找到剩余的频段来满足日益增长的新业务需求。认知无线电技术通过频谱感知分析电磁环境中授权频段的空闲时段,将空闲时段动态的分配给次要用户。对频谱进行动态分配,极大的提高了频谱的利用效率。而信号调制方式的正确识别是频谱分析的前提,只有在正确的识别信号的调制方式之后,才能进一步进行频谱监测、管理、分配等任务。
现有的关于自动调制分类识别主要采用两类方法[1]:一类是基于最大似然决策理论的方法;另一类是基于特征提取分析的统计模式识别方法。……