Deep Learning Based Statistical Beamforming And User Scheduling For Massive MIMO Donwlink System

Massive multiple-input multiple-output(MIMO)has been recognized as one of the most potential technologies for future wireless communication systems.It can greatly improve the throughput of the system without additional bandwidth,thus has attracted significant attention in recent years.However,there are some inherent limitations in big data processing and ultra-high-speed communications for the existing technologies of massive MIMO systems,therefore the more efficient solutions are urgently needed.Recently,deep learning methods,as a rising star in the field of artificial intelligence(AI),have been regarded as an effective tool when dealing with increasingly complex physical layer of wireless communications.Therefore,this thesis explores the statistical precoding and user scheduling schemes of massive MIMO downlink transmission systems based on deep learning.Firstly,this thesis briefly summarizes the basic knowledge of deep learning.We introduce the forward propagation of deep neural network(DNN)as well as the activation process of neurons.After that,the basic modules of convolutional neural network and their corresponding functions are described.Furthermore,we introduce the loss function that guides the model to update its parameters.Moreover,a back propagation(BP)algorithm for model optimization is derived,and the mini-batch gradient descent(MBGD)method as well as the adaptive moment estimation(Adam)optimization algorithm is presented.Next,the deep learning based beamforming algorithms exploiting merely statistical channel state information(CSI)are investigated for single-cell full-dimension MIMO(FD-MIMO)downlink transmission systems.By maximizing the lower bound of average signal-to-leakage-and-noise ratio(SLNR),the optimal beamforming vector and its corresponding index of each user is derived over correlated Rician fading channels.To reduce the computation delays,the obtaining of the beamforming index is modeled as a classification problem in deep learning method,and the beamforming index is predicted according to the statistical CSI of each user.Then,a data-driven as well as two model-driven deep learning based beamforming neural networks is proposed.These network models are trained off-line to grasp the mapping of statistical CSI to beamforming index in supervised way.In the online prediction stage,the three proposed deep learning based algorithms can predict the optimal beamforming vector of the user through statistical CSI with almost no performance loss and relatively low latency.In addition,the beamforming algorithms based on model-driven deep neural network have better performance and require less computing time than the algorithm based on data-driven deep neural network.Finally,the user scheduling algorithm based on deep learning is studied for single-cell massive MIMO downlink transmission system over correlated Rician fading channels.First,based on the optimal beamforming vector obtained previously,we derive an approximate expression for the ergodic sum rate,and thus get a user scheduling algorithm which maximizes the ergodic sum rate approximation through exhaustive search.However,this exhaustive search algorithm is of extremely high computational complexity.Therefore,we try to solve the user scheduling problem using deep learning method.To solve the problem of output dimension explosion when treating the user scheduling problem as a multi classification problem,we model it as a multi-label classification problem.In the off-line training stage,the proposed scheduling network learns the user scheduling algorithm from the normalized signal and interference pattern of the users,the corresponding training samples are generated by the exhaustive search user scheduling algorithm that maximizes the ergodic sum rate approximation.In the online prediction stage,the probability of each user being scheduled is predicted by the user scheduling network according to the input normalized signal and interference pattern.Finally,the scheduled users are determined by the output probability vector.Simulation results show that the proposed user scheduling network can achieve comparable throughput to the exhaustive search algorithm with much lower computing delay.Besides,the trained user scheduling network only exploits statistical CSI,which greatly reduces the CSI feedback overhead.Moreover,it has good robustness to different channel environments and different numbers of transmit antenna.