Research On Channel State Information Feedback Method Of Millimeter-wave Massive MIMO System

When the millimeter wave massive multiple-input multiple-output(Massive MIMO)system adopts the frequency division duplex mode,the base station needs to use the channel state information(CSI)fed back by the terminal for precoding to match the fading characteristics of the channel.However,the configuration of large-scale antennas brings about the problem of occupying uplink spectrum resources when high-dimensional CSI information is fed back.This thesis focuses on the CSI feedback technology of the millimeter wave Massive MIMO system,and the main research is as follows:1.Aiming at the feedback problem of CSI under different compression ratios,a network with residual channel characteristic attention(RCCA)as the core is proposed to solve the problem of high CSI feedback overhead.The network is composed of the channel characteristic attention mechanism structure that extracts the relationship between the two channel images formed by the real and imaginary parts,the convolutional neural network that extracts the amplitude and phase relationship between the adjacent elements of the channel,and the short-hop connection.In order to improve the accuracy of CSI recovery,multiple RCCAs are cascaded and connected through long hops to form a multi-layer residul(MLR)network.RCCA solves the problem that the existing deep learning method separates the real and imaginary parts when processing complex channel matrices(to form a two-channel image)and ignores the problem between the real and imaginary parts of the channel.It is mainly used to extract the correlation characteristic information of amplitude and phase between adjacent elements of millimeter waves Massive MIMO system beam domain channel.The simulation results under different datasets show that the proposed RCCA structure can effectively learn channel characteristics to reduce the amount of CSI feedback,and at the same time,it can accurately restore the signal under different compression ratios,and it has better generalization performance under different datasets.2.Aiming at the effect of noise on signal recovery performance under different signal-to-noise ratios,the data-driven and compressed sensing scheme are combined to propose a learning measurement matrix and channel element distribution characteristics(LMMCEDC)feedback methods to reduce the amount of CSI feedback.LMMCEDC is divided into two stages.In the first stage,the measurement matrix in the compressed sensing method is learned in a data-driven manner,and then the feedback CSI is channel compressed through the learned measurement matrix.The second stage combines the measurement matrix and beam channel element distribution characteristics to complete the recovery of CSI.This solution is based on the sparseness of the beam-domain channel of the millimeter wave Massive MIMO system,and converts CSI feedback and reconstruction into the problem of sparse signal compression and recovery by compressed sensing.The simulation results show that under different signal-to-noise ratios,the LMMCEDC feedback method proposed in this thesis can effectively reduce the system's feedback overhead and carry out accurate reconstruction while ensuring the feedback overhead.