Research On Channel State Information Feedback Method Of FDD Massive MIMO Systems Based On Deep Learning

The 5^thgeneration mobile communication system has three main application scenarios:enhanced mobile bandwidth（eMBB）,ultra-reliability and low latency communication（uRLLC）and massive machine type communication（mMTC）.These three application scenarios put forward new requirements for mobile communication technology.Massive multiple-input multiple-output（MIMO）communication system is a key technology to solve the above problems.Massive MIMO systems use massive antennas at the base station（BS）to provide services for users,which can improve link capacity and energy efficiency.However,the large number of antennas leads to the surge of channel state information（CSI）.In time division duplex mode,the uplink and downlink channels are reciprocal and the downlink CSI can be obtained through the reciprocity.In fact,there are problems such as calibration errors and hardware differences in the radio frequency link.Frequency division duplex（FDD）is the main mode of the current commercial cellular mobile system.Under the FDD mode,since the uplink and downlink channels do not have reciprocity,the feedback overhead is proportional to the number of antennas in the BS,the limited feedback scheme to reduce overhead becomes the focus of research.At present,the commonly used feedback scheme is vector quantization based on the codebook,but with the increase of the number of antennas,the complexity of the codebook design also increases greatly,and the amount of feedback also increases,which becomes a bottleneck in the application of massive MIMO systems.In recent years,with the rise of artificial intelligence technologies such as machine learning（ML）and deep learning（DL）,researchers have tried to integrate artificial intelligence technology with their industries and have achieved good results in some aspects.This paper introduces some applications of deep learning in current communication system and these applications provide reference for the CSI feedback scheme research in this thesis.This thesis integrates deep learning technology with CSI feedback of FDD large-scale MIMO system,and proposes some relevant CSI feedback schemes.The main contents are as follows.This thesis studies the CSI feedback algorithm of large-scale MIMO systems such as CsiNet.Based on the basic convolutional neural network,we propose a multi-scale and multi-channel convolutional neural network scheme named MRNet,in which,we also introduce the residual network,attention mechanism,dynamic learning rate,cavity convolution and other improved methods.Through the simulation of MRNet and the comparison with the current compressed sensing（CS）algorithm and CsiNet algorithm.The CSI reconstruction efficiency of the proposed MRNet algorithm is obviously higher than that of the comparison algorithms indifferent indoor and outdoor scenes with different compression rates,which proves the effectiveness of the MRNet algorithm in the reconstruction of CSI matrix performance.Furthermore,this thesis also proposes a spatiotemporal multi-scale three-dimensional convolution neural network named MS-3D Net in this scheme.Based on the time correlation of CSI,a set of CSI matrices with time step T are regarded as a video data.The traditional CNN-LSTM network mode is transformed into three-dimensional convolution for CSI compression and reconstruction.Improved methods such as convolution long short-term memory network（ConvLSTM）,attention mechanism and deep separable convolution are adopted in the proposed MS-3D Net scheme.In this thesis,MS-3D Net is simulated and compared with CsiNet,Rec CsiNet and other networks.The results show that MS-3D Net can significantly improve the reconstruction performance with a small amount of parameters.The simulation results demonstrated the feasibility and practicability of MS-3D Net network in CSI matrix compression and reconstruction.