Channel Estimation With Deep Learning For Mm Wave Massive MIMO Systems

Massive multi-input multi-output(MIMO)millimeter-wave is one of the key research fields of future wireless mobile communications,which provides abundant spectrum resources and increases system capacity.There is no doubt that channel estimation plays an indispensable role in massive MIMO millimeter-wave systems.Therefore,considering the inadequacies of the existing channel estimation algorithms,the estimation algorithms of channel are studied based on deep learning for massive MIMO millimeter-wave systems in this thesis.1.In view of the performance defects of traditional compressive sensing algorithm its susceptibility to noise,inspired by the learned iterative shrinkage thresholding algorithm(LISTA),a learned iterative threshold shrinkage algorithm with memorys based on denoising convolutional neural network(DnCNN)is studied.Firstly,according to the angular domain sparsity of massive MIMO millimeter-wave channel,channel estimation is transformed into a sparse signal recovery problem.Furthermore,to avoid the threshold mapping error of LISTA in the signal recovery process,an iterative network is proposed based on a memory unit.Then,considering the noise,the DnCNN suitable for the system structure is integrated.Finally,the optimal parameters of the channel estimation network are obtained by model training.Simulation experiments show that the proposed channel estimation algorithm can improve performance compared with traditional sparse signal reconstruction algorithms in massive MIMO millimeter-wave systems.2.Considering the problem of joint estimation and matching of direction of arrival(DOA)and delay of traditional array signal processing algorithms,a channel estimation algorithm is studied based on multiple signal classification(MUSIC)and convolutional neural networks.The channel estimation algorithm connects multiple channel parameter estimation methods in series,and the channel matrix is calculated by estimating the channel parameters of each path.Specifically,in the estimation algorithm of delay,multiple parallel convolutional neural networks are used to establish the nonlinear relationship between the covariance matrix and the delay.Then,the classification and pairing of multi-path delays are completed after the DOA information estimated by the MUSIC algorithm is fused into the network.Finally,the channel gains are calculated by least squares,and all channel parameters are combined into a channel matrix.Simulation experiment results show that the proposed method can effectively solve the path matching problem of channel parameter estimation and has better performance.