Research On Joint Atenna Selection Hybrid Precoding Technology In MmWave Massive MIMO System

In the mmWave massive MIMO communication system,the use of Massive MIMO can make up for the loss in the millimeter wave transmission process.And the precoding technology can increase the system transmission rate by adjusting the transmitted signal phase and amplitude,thereby meeting the high transmission rate of 5G to provide strong technical support for 5G communications.However,the digital precoder uses a large number of RF links in Massive MIMO system resulting in immeasurable hardware complexity and energy consumption.It is useful to design a reasonable hybrid precoder which can reduce the use of RF links and simplify analog circuits.Antenna selection can further reduce the hardware complexity of the system due to there is a certain correlation between antennas.The antenna does not increase the spectrum efficiency much when they are used at the same time.Choosing a better antenna sub-array can reduce system complexity while losing less spectrum efficiency performance.Therefore,this thesis studies the hybrid precoding technology and antenna selection technology for mmWave massive MIMO system.The phase shifter in the fully connected hybrid precoding structure has unit modulus restrictions on the analog precoding matrix elements,which makes it difficult to solve the problem of hybrid precoding.Traditional hybrid precoding algorithms are usually difficult to achieve performance and complexity better at the same time.In response to this problem,this thesis uses a deep learning-based hybrid precoder design algorithm to construct a regression-based neural network structure,which realizes directly output the analog precoding/combination matrix when input channel matrix to the neural network.While the neural network obtains better performance,it also reduces the calculation time of the hybrid precoder design.In addition,the neural network is trained with unideal channel which add different noises to simulate the non-ideal estimated channel matrix,so that the neural network structure still has strong robustness in the case of non-ideal estimated channels.On the basis of hybrid precoding,antenna selection can further save cost and energy consumption.On the premise of a better performance,this thesis uses the deep learningbased antenna selection algorithm to transform the selection of antenna sub-arrays into neural network classification problems.The deep learning-based antenna selection algorithm can reduce the calculation time of antenna selection.What's more,it can directly output the antenna sub-array with better performance when inputting the channel.The deep learning-based antenna selection and hybrid precoding algorithm both obtain the result through the channel matrix.Therefore,we combine antenna selection with hybrid precoder design to build a neural network.Specifically,the antenna sub-arrays are input from the antenna selection neural network to the hybrid precoding neural network which can output he analog precoding/combination matrix.This thesis gives two antenna selection training data generation methods based on the number of antennas.Numerical experiments show that,the proposed antenna selection and hybrid precoding neural network structure provides a better antenna sub-array and hybrid precoder design in less time.