摘要
自动调制识别是空间认知通信系统的关键技术,有助于实现自适应信号解调。深度神经网络虽然具有特征提取能力强的优势,但也存在参数众多、计算量大的问题,难以实现空间在轨应用。针对以上问题,提出了一种轻量化、高性能的卷积神经网络结构。网络先提取信号的同相正交相关特征,再提取时域特征,最后提取各通道特征均值进行分类。对11种调制方式分类的实验结果表明:当信噪比高于0 dB时,平均识别准确率能达到86.94%,较传统的高阶累积量的方法提高了31.54%;与目前高识别准确率的深度神经网络模型相比,仅使用不到10%的模型参数,在树莓派4B上计算速度平均提高了20倍。
Automatic modulation recognition is a key technology for spatial cognitive communication system,which helps to realize adaptive signal demodulation.Although the deep neural network has the advantage of strong feature extraction,it suffers from the problems of numerous parameters and large amount of calculation,and thus is difficult to be implemented in in-orbit applications.To mitigate these problems,we propose a lightweight,high-performance convolutional neural network structure.The network first extracts the in-phase and quadrature features of the signal,then the time domain features,and finally the mean value of each channel feature for classification.The experimental results of the classification of 11 modulation methods show that when the signal-to-noise ratio is higher than 0 dB,the average recognition accuracy can reach 86.94%,which is 31.54%higher than that of traditional cumulant methods.Compared with the current deep neural network model with high recognition accuracy,the network proposed uses only less than 10%of model parameters,and increases the calculation speed by an average of 20 times on Raspberry Pi 4 B.
作者
崔天舒
崔凯
黄永辉
赵文杰
安军社
CUI Tianshu;CUI Kai;HUANG Yonghui;ZHAO Wenjie;AN Junshe(Key Laboratory of Electronics and Information Technology for Space Systems,National Space Science Center,Chinese Academy of Sciences,Beijing 100190,China;School of Computer Science and Technology,University of Chinese Academy of Sciences,Beijing 100049,China)
出处
《北京航空航天大学学报》
EI
CAS
CSCD
北大核心
2022年第6期986-994,共9页
Journal of Beijing University of Aeronautics and Astronautics
基金
中国科学院复杂航天系统电子信息技术重点实验室自主部署基金(Y42613A32S)。
关键词
认知通信系统
调制方式识别
同相正交相关特征
卷积神经网络
深度学习
cognitive communication system
modulation recognition
in-phase and quadrature correlation feature
convolutional neural network
deep learning
