Research On Downlink CSI Feedback Scheme In Massive MIMO Systems Based On Deep Learning

Massive Multiple-Input Multiple-Output(MIMO)technology has become one of the key technologies of 5G by its huge advantages on spectral efficiency and energy efficiency.Evolving from the MIMO technology,massive MIMO is characterized by a large-scale antenna array deployed at the base station.When the massive MIMO system works in frequency division duplex mode,since there is no reciprocity between uplink and downlink channels,the base station must obtain downlink channel state information(CSI)through feedback.And,the rise in the number of antennas will sharply increase the feedback overhead of the wireless communication system.Designing a CSI feedback scheme with both low feedback overhead and high reconstruction accuracy has become one of the key issues in the practical application of FDD mode massive MIMO.This thesis employs the deep learning(DL)technology widely used in the field of data science in recent years to study the compression and reconstruction of CSI.The following summarizes the specific research content of this article:Aiming at the problems of high CSI feedback overhead and high reconstruction delay caused by a large number of users in the multi-user massive MIMO system,a two-stage CSI feedback scheme based on the Adversarial AutoEncoder(AAE)is proposed.This scheme uses the method of adversarial training to explicitly incorporate user scheduling information into a part of the feedback codeword,so that the base station can perform user selection before reconstructing the complete CSI,thereby avoiding transmission and reconstruction of all users’ CSI,and significantly reducing the total feedback data and the number of CSI reconstructions.The simulation results show that the proposed scheme can maintain the CSI reconstruction accuracy with respect to the original CSI feedback network while transmitting user scheduling information,and the performance of gross throughput exceeds that of deep learning feedback scheme using maximal channel gain(MCG)user selection algorithms under the same feedback overhead.Aiming at the problem of multi-user beamforming based on limited CSI feedback,a multi-user CSI limited feedback scheme oriented to beamforming is proposed.In this scheme,the CSI reconstruction process is omitted,and the precoding matrix is directly generated using quantized codeword fed back by multiple users.In order to further compress the communication overhead caused by multiple feedbacks of CSI over a period of time,this paper proposes a time-varing channel feeback scheme based on Hidden Markov Model.In this scheme,a time feature fusion device is added on the base station to extract the historical feature of CSI and fuse it with the current short codeword to dynamically adjust the precoding matrix.The simulation results show that beamforming-oriented CSI limited feedback scheme proposed in this thesis is superior to the DL-based single-user CSI feedback scheme and the traditional codebook-based CSI feedback scheme in terms of throughput,time and space complexity.What’s more,the time-varing channel feedback scheme proposed in this thesis can effectively utilize the time correlation of CSI,and maintain high throughput performance with very lower feedback overhead.