基于深度学习的智能表面毫米波MIMO信道估计
Deep Learning-based Intelligent Surface Millimeter-wave MIMO Channel Estimation
张思伟 1袁德成 2王国刚1
作者信息
- 1. 辽宁省化工控制技术重点实验室,辽宁沈阳 110142;沈阳化工大学 信息工程学院,辽宁 沈阳 110142
- 2. 辽宁省化工控制技术重点实验室,辽宁沈阳 110142
- 折叠
摘要
研究了一个基于深度学习的大型智能表面(Large Intelligent Surface,LIS)毫米波多输入多输出(Multiple-Input Multiple-Output,MIMO)系统.为了克服波长和阵列间距相差较大的信号传输问题,传统的均匀线性阵列(Uniform Linear Array,ULA)被替代为均匀平面阵列(Uniform Planar Array,UPA).提出了一种基于改进的双卷积神经网络——ChannelNet的信道估计方法.采用最小二乘(Least Squares,LS)算法获取初始化的信道信息,使用ChannelNet获得更高精度的信道信息,并重点探究了在UPA结构下的表现.通过与LS算法和多层感知器算法进行比较.结果表明,该算法在信道估计精度和效率方面均优于以上2种算法,且使用UPA结构的ChannelNet算法相对于使用ULA结构的表现更好.该方法在毫米波MIMO信道估计方面具有更好的性能.
Abstract
A deep learning-based Large Intelligent Surface(LIS)millimeter-wave Multiple-Input Multiple-Output(MIMO)system is investigated.In order to overcome the challenge of signal transmission with significant differences in wavelength and array spacing,the conventional Uniform Linear Array(ULA)has been replaced by a Uniform Planar Array(UPA).An improved dual-convolutional neural network-ChannelNet algorithm based method for channel estimation is proposed.Initially,the least squares algorithm is employed to obtain the initial channel information.Subsequently,ChannelNet is utilized to achieve higher precision in channel estimation,with a particular focus on its performance within the context of the UPA structure.By comparing ChannelNet algorithm with the least squares algorithm and the multilayer perceptron algorithm,the results indicate that ChannelNet algorithm outperforms both the least squares algorithm and the multilayer perceptron algorithm in terms of channel estimation accuracy and efficiency.Furthermore,the performance of the ChannelNet algorithm utilizing the UPA structure is superior to that of using the ULA structure.This indicates that the proposed method exhibits superior performance in millimeter-wave MIMO channel estimation.
关键词
大型智能表面/信道估计/ChannelNet/均匀线性阵列/均匀平面阵列Key words
LIS/channel estimation/ChannelNet/ULA/UPA引用本文复制引用
基金项目
国家重点研发计划(2018YFB1700200)
出版年
2024
国家科技期刊平台
NSTL
万方数据