Research On Several Technologies In Massive MIMO Systems Based On Deep Learning

Millimeter wave(mm Wave)and Massive Multiple-Input Multiple-Output(MIMO)technologies are currently hot topics in the field of communication.mm Wave has poor penetration ability and faster attenuation,while Massive MIMO technology improves system reliability and transmission rate by equipping base stations with a large number of antenna arrays to combat signal fading caused by mm Wave.Low-frequency communication systems generally use a full digital precoding,while for mm Wave Massive MIMO systems,adopting a full digital precoding approach would result in a significant hardware overhead.Hybrid precoding design can reduce the number of Radio Frequency(RF)links and ensure the performance.Obtaining accurate Channel State Information(CSI)becomes a critical issue for assessing the performance of Massive MIMO systems.Traditional channel estimation algorithms based on pilots have been the focus of attention,but due to the large number of antennas in Massive MIMO systems,the computational complexity also increases,and estimating the downlink channel becomes challenging in Frequency Division Duplexing(FDD)systems due to the lack of reciprocity.In this paper,a deep learning based channel estimation neural network is proposed,which utilizes the spatial partial reciprocity of FDD systems and variational Bayesian autoencoders to extract the hidden variables of uplink and downlink channels and perform cross alignment and distribution alignment to achieve downlink channel estimation in FDD systems.Simulation results show that under non-ideal spatial reciprocity conditions,the proposed network has better estimation performance compared to traditional algorithms..Traditional hybrid beamforming designs typically use optimization methods to find the hybrid beamforming matrix that satisfies the optimal objective function.In this paper,a deep learning network for hybrid precoding is proposed,which takes pilot signals as inputs and outputs the phase of the analog precoding matrix.A loss function designed based on spectral efficiency is used for optimization through backpropagation,thus achieving the design of the analog precoding matrix.On top of the analog precoding matrix,the optimal solution of the digital precoding matrix is computed using the zero-forcing algorithm.Furthermore,an deep neural network is proposed,which embeds the loss function designed based on spectral efficiency into the traditional beamforming design.And the iterative algorithm is divided into two steps,in the first step,assuming that the merged matrix is known,the suboptimal solution of the hybrid precoding matrix is obtained through the deep neural network;in the second step,based on the suboptimal solution of the precoding matrix,the suboptimal solution of the merged combining matrix is obtained through the second deep neural network.This process is repeated from the first step to the second step,and simulation results show that the obtained hybrid beamforming matrix becomes more excellent in performance as the number of iterations increases,and compared with other algorithms,it can achieve better Bit Error Rate(BER)performance.The deep neural network proposed in this paper demonstrates relatively superior performance in the corresponding aspect,and holds certain research value.