电子学报 ›› 2022, Vol. 50 ›› Issue (5): 1211-1217.DOI: 10.12263/DZXB.20210723

• 学术论文 • 上一篇    下一篇

一种轻量化低复杂度的FDD大规模MIMO系统CSI反馈方法

廖勇, 李玉杰   

  1. 重庆大学微电子与通信工程学院,重庆 400044
  • 收稿日期:2021-06-07 修回日期:2021-12-27 出版日期:2022-05-25 发布日期:2022-06-18
  • 作者简介:廖 勇 男,1982年1月出生于四川自贡.现为重庆大学副研究员、博士生导师.主要研究方向为下一代无线通信、人工智能、区块链、量子计算及其在无线通信中的应用.E-mail: liaoy@cqu.edu.cn
    李玉杰 男,1996年7月出生于河南周口市.现为重庆大学微电子与通信工程学院研究生.主要研究方向为无线通信中的CSI反馈算法研究.E-mail: cleeshark@cqu.edu.cn
  • 基金资助:
    国家自然科学基金(61501066);重庆市自然科学基金项目(cstc2019jcyj-msxmX0017)

Lightweight and Low Complexity CSI Feedback Method for FDD Massive MIMO Systems

LIAO Yong, LI Yu-jie   

  1. School of Microelectronics and Communication Engineering,Chongqing University,Chongqing 400044,China
  • Received:2021-06-07 Revised:2021-12-27 Online:2022-05-25 Published:2022-06-18

摘要:

针对频分双工大规模多输入多输出(Multiple-Input Multiple-Output,MIMO)通信系统中信道状态信息(Channel State Information,CSI)反馈方法复杂度高、精度低和开销大的问题,本文提出了一种基于深度学习的低复杂度CSI反馈方法.该方法通过端到端的方式构建了一种从用户设备编码器到基站解码器相结合的网络结构.编解码器利用连续的平均池化层和上采样层完成特征图的降维和升维,同时引入深度可分离卷积神经网络减少网络参数量.在解码器部分,本文利用残差网络构建连续的拥有大卷积核的残差块逼近原始CSI矩阵.仿真结果表明,和已有的代表性方法相比,本文所提方法在归一化均方误差上有2dB~5dB的性能提升,在余弦相似度上也有2%~5%的提升,并且在时间复杂度和空间复杂度上均有更好的表现.

关键词: 频分双工, 大规模MIMO, 信道状态信息, 深度学习, 深度可分离卷积神经网络, 残差网络

Abstract:

Aiming at the problems of high complexity, low precision and high overhead of the channel state information(CSI) feedback method in frequency division duplexing massive multiple input multiple output(MIMO) communication system, this paper proposes a low complexity CSI feedback method based on deep learning. This method constructs a network structure from the encoder of user equipment to the decoder of base station in an end-to-end way. The codec uses a continuous average pooling layers and up sampling layers to complete the dimensionality reduction and dimensionality increase of the feature map, and introduces a depthwise separable convolutional neural network to reduce the amount of network parameters. In the decoder part, this paper uses the residual network to construct continuous residual blocks with large convolution kernel to approximate the original CSI matrix. Simulation results show that, compared with the existing representative methods, the method proposed in this paper has 2dB~5dB improvement in normalized mean square error, and 2%~5% improvement in cosine similarity, and it has better performance in time complexity and space complexity.

Key words: frequency division duplexing, massive MIMO(Multiple-Input Multiple-Output), channel state information, deep learning, depthwise sparable convolution neural network, residual network

中图分类号: