Research Of Deep Learning-Based Antenna Selection And Hybrid Beamforming In Millimeter-Wave Massive MIMO Systems

Millimeter wave massive MIMO is one of the most promising research directions to support future mobile communications.However,the cost and complexity of the hardware increase significantly due to the multiple expensive RF links equipped in massive MIMO systems.Antenna selection techniques can simplify the hardware structure by optimization strategies to select some antennas of high performance for communication while ensuring system performance;meanwhile,hybrid beamforming techniques use low-dimensional digital beamforming and high-dimensional analog beamforming to greatly reduce the number of RF links required,which in combination with antenna selection techniques can further reduce system complexity and improve signal transmission quality.However,the current traditional model-oriented antenna selection and hybrid beamforming algorithms have the disadvantage of high computational complexity.For this reason,a deep learning-based joint optimization scheme for antenna selection and hybrid beamforming is proposed in this paper,and the specific research content of this paper is as follows:(1)An antenna selection algorithm based on a two-dimensional convolutional neural network is proposed.A convolutional neural network is constructed and the network is trained using different implementations of the channel matrix.After the training is completed,input the channel matrix and the network can output the subarray that maximizes the spectral efficiency.The proposed deep learning-based antenna selection algorithm is shown to be close to the optimal antenna selection algorithm using exhaustive search in terms of performance,but with lower computational time complexity.(2)A hybrid beamforming algorithm based on a two-dimensional convolutional neural network is proposed.After the optimal subarray selection for the antenna is determined,the corresponding partial channel matrix is fed into the second convolutional neural network constructed.The convolutional neural network then predicts the best RF beamformer and calculates the corresponding baseband beamformer.This neural network reduces the estimation of the channel matrix compared to conventional solutions,while its achievable system spectral efficiency performance is close to that achievable by the SVD fully digital optimal unconstrained beamforming algorithm.