Research On Channel State Information Feedback Technology Based On Deep Learning In Massive MIMO Systems

The multi-input multi-output(MIMO)technology is one of the key technologies of the 5th Generation(5G)mobile communication system,and it is expected to be applied and further enhanced in the next Generation mobile communication system.The increase in the number of antennas brings more available space resources to the system,thus bringing greater system capacity and higher spectral efficiency,but at the same time,the difficulty of acquiring Channel State Information(CSI)also increases significantly.Especially in the Frequency Division Duplex(FDD)system with no reciprocity between upstream and downstream,CSI must be acquired through feedback.Therefore,it is an urgent problem for massive MIMO systems to develop an efficient and accurate channel feedback scheme.In recent years,artificial intelligence technology led by Deep Learning(DL)has been developing rapidly and deeply integrated with various vertical industries,including communication industry,and has become one of the key technologies of the next generation communication system.As a result,intelligent communication has become one of the mainstream directions of future communication system development,and it also provides a new solution to the channel feedback problem of massive MIMO system.In this thesis,the optimization of channel feedback technology based on DL in massive MIMO system is studied.Firstly,this thesis explores the existing DL-based CSI feedback scheme in massive MIMO systems.This thesis describes the basic concept of DL technology,introduces three kinds of classical and two kinds of commonly used neural network structures and their application scenarios,and also introduces three traditional CSI feedback methods and their principles,advantages and disadvantages and application scenarios.Finally,the research on DL-based CSI feedback scheme at home and abroad is discussed in detail.The network structure,training method and performance evaluation of the first DL based CSI feedback network in FDD massive MIMO systems-CsiNet are mainly introduced.Then the subsequent improvement schemes based on CsiNet are reviewed from the introduction of time correlation,the introduction of upstream and downstream correlation,network structure optimization and other aspects,respectively.The existing problems and future development directions are also pointed out.Then,this thesis studies the DL-based channel feedback and quantization technology in FDD massive MIMO systems.This thesis analyzes the sparse characteristics of massive MIMO channel in the angle domain and time delay domain,establishes a massive MIMO system model in FDD mode,describes the process of channel feedback after adding quantization in the system model,and introduces the commonly used quantization methods and application scenarios.A DL network QCsiNet+,which improves the network structure according to the theory of convolutional neural network and adds quantization module and inverse quantization module,is designed.The method of data set generation,training strategy and parameter setting under this architecture are described.The simulation compares both the performance of the quantization module under different quantization methods,and the reconstruction performance of the whole network with the most representative CSI feedback network CsiNet under different scenarios,different quantization methods,different compression rates and different quantization bits.The simulation results show that the proposed technology can generate feedback bitstream for transmitting and storage,thus increasing the feasibility in the actual system.It eliminates the effects of the quantization error,therefore it has stronger robustness to quantization error.It also reduces the overhead of feedback bits and ensures the reconstruction accuracy,so it has good research significance.Finally,this thesis studies the DL-based joint channel estimation and feedback technology in FDD massive MIMO systems.This thesis establishes a massive MIMO system model in FDD mode,describes the process of channel estimation and feedback in this system model,and introduces the commonly-used traditional channel estimation algorithm and the DL-based super resolution network.By analyzing the relationship between channel estimation and feedback and super resolution,two joint channel estimation and feedback architectures,channel joint estimation and feedback network CEFnet and pilot direct feedback network PFnet,are designed,and the data set generation method,network parameter setting,training method and training parameter setting under the two architectures are described.The simulation first analyzes the independent performance of each subnet in detail,and then compares the reconstruction performance and network complexity of the two architectures under different scenarios,different compression rates and different quantization bits.The simulation results show that the two architectures have good reconstruction performance under the premise of complete implementation of CSI feedback,and have strong robustness to different environments and quantization errors,which shows the feasibility of the two networks in the actual communication system.