Efficient Precoding Based On Deep Leaning In Massive MIMO Systems

With the increase of various applications such as vehicle communication,highresolution augmented reality,3D video games and massive device connectivity,B5 G puts forward higher requirements for system capacity and transmission rate.In order to achieve this goal and requirement,millimeter wave,Massive MIMO and heterogeneous networks are widely studied and applied,and become the main deployment application scenarios of B5 G.The increasing demand of wireless communication system largely depends on the spectrum efficiency.Therefore,it is more urgent to study the frequency effect of B5 G new application scenarios than the traditional cellular network,and it has become one of the current research hotspots.At the same time,it is predicted that the carbon emissions of the information and communication technology industry will reach 14% of the global emissions in 2040,which is the main factor of global energy consumption.Therefore,B5 G system with low energy consumption design concept is imperative.Previous studies have shown that the combination of neural network model and precoding is an effective method to achieve high-frequency and high-energy precoding,which has become a research hotspot in recent years.Aiming at the problems of energy efficiency and real-time performance in multiuser scenarios of Massive MIMO heterogeneous networks,this paper proposes a fast precoding algorithm based on deep learning.Meanwhile,considering the influence of the number of antennas of different macro stations,cooperative small base stations,minimum user rate and number of users on the proposed algorithm,the system analysis and design are carried out.Theoretical analysis and simulation show that the proposed algorithm has higher energy efficiency and lower computation delay,and has better practicability.In millimeter wave communication system,in order to solve the problem that it is difficult to obtain training samples based on deep learning algorithm,this paper proposes an efficient hybrid precoding algorithm based on unsupervised learning.The algorithm redefines the loss function.The training samples only need the channel matrix,so the complex process of obtaining training samples is avoided.Simulation results show that the performance of the hybrid precoding algorithm proposed in this paper is equivalent to Mo-Alt Min algorithm under the full connection structure,and better than other comparison algorithm under the partial connection structure.In order to solve the problem of the performance of precoding algorithm based on convolutional neural network,a hybrid precoding algorithm based on graph neural network is proposed.By developing the characteristics of graph neural network,the defect of ignoring topology structure in extracting features of convolutional neural network is solved.Not only the neighborhood elements of channel matrix input are considered,but also the underlying topology formed by channel matrix is fully utilized.In this paper,the graph attention network is specially developed.The proposed algorithm can obtain accurate precoding vector without knowing all the information of graph structure.It overcomes the disadvantage of traditional graph neural network algorithm which needs to know all the information of graph structure.At the same time,the algorithm also introduces the multi head mechanism to improve the generalization ability and robustness of the network.Through simulation analysis,the algorithm proposed in this chapter can achieve good performance.