Research On Performance Optimization Of Massive MIMO Channel State Information Feedback System Based On Deep Learning

The advent of the fifth-generation(5G)communication system is driven by the increasing number of users and connected devices,as well as the growing demand for higher transmission rates brought on by the Internet of Things(Io T)and virtual reality(VR)applications.In order to meet the performance requirements of 5G,the massive multiple-input multiple-output(MIMO)technology can greatly boost the spectral efficiency,energy efficiency,system capacity and coverage by deploying large-scale antenna arrays at base stations(BS),utilizing spatial multiplexing and diversity.However,the above advantages are highly dependent on the accurate channel state information(CSI).The uplink CSI can be estimated by sending a pilot signal from the user equipment(UE)and receiving it at the BS.The challenge arises in the acquiring of the downlink CSI.Since the channel reciprocity is no longer available in frequency division duplex(FDD)systems,the UE needs to estimate the downlink CSI from the received pilot signal and feed it back to the BS via the feedback links.In massive MIMO systems,the feedback overhead consumes a large amount of bandwidth resources and will increases linearly with the number of antennas,presenting a major challenge for implementing massive MIMO in FDD systems.To achieve effective downlink CSI limited feedback in FDD massive MIMO systems,in this thesis,a new CSI feedback network based on deep learning method is proposed for CSI compression and reconstruction.The main contents are as follows:To begin with,this thesis analyzes the task of the CSI feedback network at UE,and adopts a local self-attention mechanism to make the neural network focus on the remote correlations in CSI during training and helps to extract valid information from the original CSI matrix.Simulation results show that the local self-attention mechanism can effectively improve the compression quality of CSI without significantly increasing the network complexity and burden on UE.Secondly,based on the CSI reconstruction process,this thesis analyzes the characteristics of the neural network deployed at BS,and proposes to globally deploy dense connections in the entire decoder’s feature extraction module.This enables information sharing from front to back in the CSI reconstruction network and more efficient feature fusion among the different parts.The simulation results show that the global dense connection significantly improves the quality of CSI reconstruction.In addition,this thesis introduces a vector quantization(VQ)module into the CSI feedback network to simulate the actual feedback process.In terms of the design of the codebook for VQ,the codebook is trained with the CSI feedback network as part of the parameters.The simulation results show that the deep learning-based VQ scheme effectively reduces quantization errors.Finally,the thesis designs a unified CSI feedback network,which can conduct multiple-rate CSI compression and reconstruction by deploying only one neural network.The trained network is capable of handling CSI feedback under different compression rates without changing any parameter.Simulation results show that the proposed network can effectively reduce the storage and training costs of multiple-rate compression CSI feedback networks.