| In this paper,combined with deep learning methods,the channel estimation,beam training and beam tracking for millimeter wave(mm Wave)massive multiple-input multiple-output systems are studied.The detailed work is concluded as follows.Considering the low rank problem of the traditional least squared channel estimation method in the wideband mm Wave scenario,a time-domain channel estimation method exploiting block sparsity is proposed,which utilizes the block sparsity to estimates the time-domain channel by iteratively searching for the index of the best nonzero block achieving the largest projection of the current residual.Compared with the simulation performance of the existing method,this method can achieve significant performance improvement.In this paper,considering the high overhead in beam training for mm Wave massive MIMO systems,the nonlinear relationship between beam codewords caused by channel power leakage is utilized,and the beam training problem is modeled as a multi-classification problem.Two schemes of beam training based on deep neural network(DNN)are proposed.In the first scheme,several beam pairs selected from the codebook are used for channel measurement,and an offline trained DNN is used to directly predict the optimal beam pair.In the second scheme,beam training is divided into two stages: initial test and additional test.In the initial testing phase,the trained DNN is used to obtain a probability vector.In the additional test phase,a small number of beam pairs are selected to perform additional beam training according to the probability vector obtained from the initial test,and then jointly determine the final beam pair based on the results of the two stages.The simulation results show that the schemes proposed in this paper can significantly reduce the overhead of beam training with a small loss of achievable rate when using narrow beams.This paper also studies the beam tracking problem in the mm Wave massive MIMO time-varying channel model.The beam tracking algorithm based on Extended Kalman Filter requires a known angle time-varying model,so there are certain limitations in practical applications.To address the above problem,a beam tracking algorithm based on Q-learning is proposed.In this algorithm,the beam tracking is divided into two stages.In the first stage,a set of predesigned codewords is used as candidate beams.The beam switching direction is selected based on the Q-learning and the beam tracking is performed by switching the candidate beams.In the second stage,in order to improve the accuracy of beam tracking,additional measurements are performed around the beam codeword obtained in the first stage,and the beam codeword with the highest received signal strength is selected as the result of beam tracking.The simulation results show that the proposed beam tracking algorithm based on Q-learning can accurately perform the beam tracking with low complexity under different angular varying rates. |
