Research On Hybrid Precoding Algorithm Based On Deep Learning In THz-MIMO System

Terahertz(THz)communication is considered as the key technology in the future6 G wireless system because of its huge bandwidth and ultra-high propagation rate,and it has attracted more attention in academic.In order to overcome the severe path loss fading and alleviate the high-power consumption,the THz massive multi-input multioutput(MIMO)system with hybrid precoding technology has been widely used in THz communication.However,due to the larger bandwidth and the larger number of antennas equipped at the base station(BS),the beam width is very narrow,the physical direction deviation between the physical direction and the target physical direction of the beam on different subcarriers will be significantly increased,and will be completely split into independent physical directions.This phenomenon,known as beam split,will lead to serious loss of array gain,which can not be ignored in THz massive MIMO systems.In addition,the solution of complex optimization problems with hybrid precoding has always been a key problem,which becomes more challenging in the case of imperfect channel state information(CSI).In view of the beam split,by introducing a delay network between the radio frequency(RF)chains and phase shifters(PSs),the frequency-independent beams generated by PSs can be transformed into frequency-dependent beams controlled by time delay and phase,thus compensating the array gain loss.However,the algorithm is based on the perfect CSI,it is very difficult to obtain perfect CSI in reality.Therefore,inspired by the deep learning method and the unsupervised deep learning method does not need a large amount of label data,we propose a hybrid precoding scheme based on unsupervised deep neural network(DNN)in the case of imperfect CSI,which trains the DNN model to learn the optimal hybrid precoding matrix so as to improve the performance of the system.Considering multi-user THz massive MIMO system,the algorithm based on fully connected neural network can not completely eliminate multi-user interference,which leads to the performance limitation of unsupervised learning method.In order to solve this problem,inspired by the attention mechanism,we propose a hybrid precoding scheme based on attention mechanism.The attention mechanism can extract the characteristics of multi-user interference,then the convolutional neural network(CNN)learns the optimal analog precoding matrix to maximize the system achievable rate.Through the simulation results,the proposed algorithm can achieve better achievable rate performance and the robustness has been verified.