Study Of Massive MIMO Precoding Algorithm Based On Deep Learning

Massive multiple-input multiple-output(MIMO)technology is one of the most critical technologies of the 5th generation mobile communication system(5G).By configuring a large number of antennas in the base station(BS)to provide services for multiple user equipments(UE)at the same time,the spatial freedom of the system is greatly improved.Because of its advantages in system throughput,spectrum utilization and energy efficiency,massive MIMO technology has become a research hotspot in the field of communication.However,with the expansion of antenna array scale,it brings computational complexity difficulties to the signal precoding of downlink BS terminal.In this paper,the signal precoding algorithm in the massive MIMO system is studied in order to maintain excellent precoding performance under the condition of low complexity.In recent years,more and more researchers have successfully applied deep learning frameworks to wireless communication systems,mainly focusing on the physical layer,such as resource allocation,signal detection,signal precoding,etc.Deep learning-based communication field research mainly has two ideas.One of the ideas is a data-driven way of training the communication system as a black box network,but it needs to pay the price of high training cost and high computing overhead.The other is a model-driven approach that makes full use of the knowledge of communication to construct network models for training.This thesis builds neural networks based on the deep learning model-driven approach to conduct in-depth research and analysis on the signal precoding of massive MIMO.The main contents are as follows:Firstly,the precoding algorithms of the massive MIMO system in quantization are studied.Focusing on the nonlinear precoding algorithm with one bit quantization accuracy,this paper deduces in detail how to solve the nearest vector problem by using alternating minimization under the framework of biconvex relaxation,and combines the biconvex relaxation algorithm with the deep neural network to establish the model-driven biconvex relaxation neural network precoding algorithm model.The algorithm expands the iterative formula of the original algorithm to form a deep multi-layer connected neural network,and optimizes the iterative parameters by training large numbers of batches of data,and finally achieves the purpose of improving the precoding performance.Secondly,the massive MIMO linear precoding algorithms are further studied.To solve the problem of high inverse complexity of high dimensional matrices,the Neumann series expansion algorithm and the Richardson iterative precoding algorithm are studied.The Richardson iterative algorithm based on regularied zero forcing precoding is combined with deep learning framework to construct a deep neural network by expanding the iterative formula and adding a trainable parameter.The complexity of the proposed algorithm is analyzed theoretically,and the performance simulation under different configurations is given,which proves that the proposed algorithm has low complexity and maintains high precoding performance in the massive MIMO system,and verifies the feasibility and effectiveness of the proposed scheme.