一种基于深度学习的FDD大规模MIMO系统CSI反馈方法

CSI Feedback Method Based on Deep Learning for FDD Massive MIMO Systems

针对频分双工(Frequency Division Duplexing,FDD)大规模多入多出(Multiple-Input Multiple-Output,MIMO)系统中现有信道状态信息(Channel State Information,CSI)反馈方法复杂度高、反馈精度低的问题,本文提出一种基于深度学习的CSI压缩反馈方法.该方法首先采用卷积神经网络(Convolutional Neural Network,CNN)提取信道特征矢量,然后利用最大池化(Maxpooling)网络压缩CSI,最后考虑到大规模MIMO信道存在空间相关性的特点,分别对单用户和多用户场景使用双向长短期记忆(Bidirectional Long Short-Term Memory,Bi-LSTM)网络和双向卷积长短期记忆(Bidirectional Convolutional Long Short-Term Memory,Bi-ConvLSTM)网络对CSI进行重构.本文利用大规模MIMO信道数据对所提的深度学习网络进行离线训练,该网络学习到的信道信息能充分表征信道的状态.仿真结果表明,与已有的典型CSI反馈方法相比,本文所提方法反馈精度更高,运行时间更短,系统性能提升明显.

Existing channel state information(CSI)feedback methods for frequency division duplexing(FDD)multiple-input multiple-output(MIMO)systems have high complexity and low feedback accuracy.In this paper,a deep learning-based CSI compression feedback method is proposed.The method first uses the convolutional neural network(CNN)to extract the channel feature vector,and then uses the maximum pooling(Maxpooling)network to compress the CSI.Finally,considering the spatial correlation of the massive MIMO channel,bidirectional long short-term memory(Bi-LSTM)network and bidirectional convolution long-term memory(Bi-ConvLSTM)network are used for single-user and multi-user scenarios respectively to recover the CSI.In this paper,the deep learning network is trained offline using massive MIMO channel data,the channel information learned by the network can fully characterize the states of the channel.The simulation results show that compared with the existing typical CSI feedback methods,the proposed method has higher feedback accuracy,shorter running time and better system performance.