Research On Hybrid Precoding Technology Based On Machine Learning In Millimeter Wave Massive MIMO Systems

The fifth generation(5G)system,as a new generation of mobile system to provide communication needs facing the society in 2020,further enriches the application scenarios.Among them,millimeter wave(mm Wave)and massive multiple input and multiple output(m MIMO)technologies play an important role in 5G.The abundant spectrum resources of the mm Wave frequency band and the m MIMO system that can provide ultra-high array gain complement each other.Therefore,the combination of mm Wave and m MIMO technology have become the most critical part in 5G.In order to improve the reception performance,the precoding technology is adopted at the base station.The traditional fully digital precoding,where a dedicated RF chain is configured for each antenna,leads to unbearable hardware costs and power consumption.Thereby the hybrid precoding architecture have received widespread attention.As a science of artificial intelligence,machine learning has a significant advantage in solving nonlinear problems and reducing time complexity.To this end,this thesis focuses on the research of hybrid precoding design in mm Wave m MIMO systems based on machine learning.The main contents are as follows:1.For the single user mm Wave m MIMO systems,a hybrid precoding design scheme based on convolutional neural networks(CNN)is presented.In this scheme,on the basis of decoupled receiver transmitter hybrid precoding and hybrid combiner design,the analog precoding and digital precoding are designed separately to maximize the spectrum efficiency of the system.At the same time,the problem of selecting the optimal analog precoding from the codebook is transformed into a classification problem,and then CNN network is constructed and trained to predict the optimal RF precoding and combination matrix.On the basis of analog precoding,the digital precoding matrix is designed.The simulation results show that compared with the traditional mixed precoding scheme based on OMP,the CNN-based hybrid precoding scheme has lower time complexity while ensuring the improvement of spectrum efficiency and the reduction of bit error rate.2.For the multiuser mmwave m MIMO systems,a hybrid precoding scheme based on histogram of oriented gridients and support vector machine(HOG-SVM)is proposed.In this scheme,the RF precoding matrix and RF combination matrix are designed by exhaustively searching the codebook to maximize the user's expected received signal power.Then,the channel sample eigenvector is extracted by HOG,and the index values of the selected RF precoding matrix and RF combination matrix are used as labels to train SVM classifier,and then the optimal RF precoding matrix and RF combination matrix are predicted.On the basis of this,the digital precoding matrix is designed by ZF criterion with the equivalent channel matrix.The simulation results show that the scheme can achieve better spectral efficiency and lower time complexity than the traditional optimization based and exhaustive search scheme.