基于UNet结构的大规模MIMO系统CSI反馈设计  

Design of UNet-based CSI Feedback in Massive MIMO System

作　　者：刘庆 李义 李璘 王有军 李康 熊林麟 王平 LIU Qing;LI Yi;LI Lin;WANG Youjun;LI Kang;XIONG Linlin;WANG Ping(Bijie Power Supply Bureau,Guizhou Power Grid Corporation,Bijie 551700,China;School of Communication and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)

机构地区：[1]贵州电网有限责任公司毕节供电局,贵州毕节551700 [2]重庆邮电大学通信与信息工程学院,重庆400065

出　　处：《电讯技术》2025年第3期371-377,共7页Telecommunication Engineering

基　　金：2022年贵州电网有限责任公司毕节供电局电力技术开发项目(0607002022030101SC00049)。

摘　　要：从用户端获取下行信道状态信息(Channel State Information,CSI)是频分双工(Frequency Division Duplex,FDD)模式下通信系统信息高效传输的关键,然而其反馈开销随着天线规模的增加而增大,给大规模多输入多输出(Multiple-Input Multiple-Output,MIMO)系统带来了重大挑战。针对此问题,提出了一种基于连续卷积和注意力机制的CSI反馈网络结构U-shaped Transformer Neural Network(UTNet)。首先,编码器和解码器分别采取编码与压缩同步、解码和重建同步的连续采样结构,实现特征提取和压缩。其次,在编码器的末端和解码器的开端分别插入Transformer模块,提取不同位置之间的关联信息。最后,通过调节CSI反馈网络参数实现对发送数据长度的控制,旨在实现CSI信号更加智能和高效的反馈。实验结果表明,在不同压缩率下UTNet的归一化均方误差(Normalized Mean Square Error,NMSE)低于-27.45 dB,相较于现有基于深度学习的方法,UTNet能在保持更高精度的同时反馈开销更小。The downlink channel state information(CSI)obtained from users is the key to efficient transmission of communication system information under frequency division duplex(FDD)mode.However,the feedback overhead increases with the antenna scale,which poses significant challenges to massive multiple-input multiple-output(MIMO)systems.To address this issue,a U-shaped Transformer neural network(UTNet)structure based on continuous convolution and attention mechanism is proposed for CSI feedback.First,the encoder and decoder adopt a continuous sampling structure of encoding and compression synchronization,decoding and reconstruction synchronization,to achieve feature extraction and compression.Second,Transformer modules are inserted at the end of the encoder and the beginning of the decoder to extract correlation information between different positions.Finally,the CSI feedback network parameters are adjusted to achieve control of the length of the transmitted data,aiming to achieve more intelligent and efficient feedback of CSI signals.The experimental results show that under different compression ratios,the normalized mean square error(NMSE)of UTNet is lower than-27.45 dB,indicating that UTNet can maintain higher accuracy while having a smaller feedback overhead compared with existing deep learning-based methods.

关 键 词：大规模MIMO CSI反馈 深度学习 Transformer模块 

分 类 号：TN929.5[电子电信—通信与信息系统]