Deep Learning-based Integrated Channel Estimation And Precoding In MmWave Massive MIMO

Millimeter wave has extremely rich spectrum resources.With the depletion of microwave frequency resources and the performance of current communication system approaching the theoretical value,it is a trend and necessity to study millimeter wave and higher frequency signal transmission.Large scale MIMO is the extension of MIMO technology in the fourth generation wireless communication technology.It not only improves the system capacity in space,but also improves the high loss of millimeter wave.However,the high-dimensional antenna array makes the conventional channel estimation and fully digital precoding schemes difficult to be afforded.In this context,hybrid precoding is proposed to solve the problem of high consumption of pure digital precoding.This thesis studies how to decompose the channel matrix with lower complexity under the background of massive MIMO,in which it is difficult to obtain channel state information and with high computational complexity.It shows that the method of randomized singular value decomposition can effectively approach the singular value decomposition.At the same time,two kinds of codebooks commonly used in hybrid precoding,beam steering codebook and general quantized codebook,are carefully studied,and their advantages and disadvantages are pointed out Then,the time-varying millimeter wave large-scale MIMO system is studied,and the difference between analog precoding and digital precoding is revealed.Most hybrid precoding methods assume that the system has perfect channel state information or even more additional information,which may be difficult to satisfy in practice,and the cost of obtaining perfect channel state information is too high in massive MIMO.In this thesis,the conventional channel estimation and hybrid precoding processes are comprehensively considered.By combining the randomized singular value decomposition with the communication process,the natural mapping between the calculation steps and the communication mechanism is found,and a semi-blind channel estimation and hybrid precoding integrated mechanism is proposed.Then,a deep learning network called autoencoder is used to extract channel statistical features,which speeds up the convergence of RSVD and makes the cost lower.This thesis further studies the hybrid precoding in time-varying millimeter wave massive MIMO.According to the time-varying difference between analog precoding and digital precoding,a double pilot hybrid precoding time-varying tracking method is proposed,and the two-dimensional convolutional long-term and short-term memory network is used to predict the analog precoding,so that the double pilot mechanism only needs to estimate the low dimensional equivalent channel in most cases.