Research On MIMO Systems Based On Distributed Learning

In the millimeter-wave massive multiple-input multiple-output(MIMO)system,the base station adjusts the amplitude and phase of the transmitted signal through beamforming technology in order to compensate for the path loss during signal propagation and enhance system capacity and spectrum utilization.However,traditional beamforming methods are not powerful enough to capture the growing complexity and diversity of modern wireless networks.Machine learning(ML)algorithms have unique advantages in extracting channel features and strong predictive performance for data with wireless channel corruption or defects.However,the conventional ML algorithms require a large number of datasets for model training,and the data collection and transmission bring huge transmission overhead,which is not conducive to protecting users’ data privacy and data security.To address these challenges,federated learning(FL)is introduced,where the BS trains models by collecting users’ local models instead of local datasets.FL avoids uploading large amounts of raw training data,which can greatly reduce transmission overhead and protect data privacy.The purpose of this thesis is to explore the application of FL in the field of channel estimation and beamforming.The main contents are as follows:A model-driven FL-based channel estimation algorithm is investigated in order to reduce the pilot overhead.The spatial channels are converted into beamspace channels by leveraging the angular domain sparsity of the millimeter-wave channels,and the channel estimation problem is described as a sparse channel recovery problem.To improve channel estimation performance,the FL-based learned approximate message passing(LAMP)scheme is proposed.The simulation results show that the proposed FLbased LAMP scheme outperforms the conventional orthogonal matching pursuit and approximate message passing channel estimation schemes,and the trained LAMP network has good channel estimation performance for channels of different number of paths.An FL-based hybrid beamforming method is proposed in order to reduce the complexity.To improve beamforming performance in imperfect channels,a convolutional neural network(CNN)is designed to predict the precoding vectors.Then,three FL algorithms are proposed for CNN model training considering the role of different model information(model gradients,parameters).In addition,the effect of transmission noise on model training is analyzed in this thesis.The simulation results show that the proposed three FL algorithms are capable of striking flexible tradeoffs among the achievable rate,the transmission overhead and the security performance metrics compared to traditional centralized machine learning algorithm.The joint FL-based channel estimation and beamforming are investigated in order to improve system performance.An FL-based channel estimation scheme is proposed,and a channel estimation neural network is designed to estimate the channel matrix from the received pilot signal.To maximize the beamforming gain,a hybrid beamforming neural network is designed to optimize the phase of the analog precoding matrix based on the estimated channel matrix.The simulation results demonstrate that the spectral efficiency of the system after joint channel estimation and beamforming outperforms the conventional perfect channel-based beamforming schemes.