Research On Key Technologies Of Multi-User Massive MIMO Wireless Transmission For Intelligent Communications

Multi-Input Multi-Output(MIMO)technology,which can serve multiple users,is one of the key technologies in the fifth generation of mobile communications(5th Generation,5G),which enables the deployment of a large number of antennas at the Base Station(BS)to The system capacity,spectral efficiency,energy efficiency and communication reliability can be significantly improved by deploying a large number of antennas at the base station(BS)to serve multiple users with spatial resources at the same time and frequency resources.Although largescale MIMO has been widely used,there are still problems of high cost in practical deployment,including high channel estimation overhead and high computational complexity of multi-user scheduling.In recent years,artificial intelligence represented by Deep Learning(DL)has been developed rapidly and widely used in various communication scenarios,and the booming development of intelligent communication technology provides new ideas to solve these problems in large-scale MIMO.In this thesis,this thesis focus on the two key technologies of downlink channel reconstruction and multi-user scheduling in multi-antenna multi-user MIMO systems for intelligent communication.First,this thesis conducts an in-depth investigation on the existing downlink channel reconstruction and multi-user scheduling algorithms under large-scale MIMO systems,introduce the DL basics,and propose the research focus of this thesis.The existing traditional downlink channel reconstruction algorithms and multi-user scheduling algorithms are summarized,and the problems of high complexity and long running time of traditional downlink channel reconstruction algorithms and high computational complexity and slow speed of traditional multi-user scheduling algorithms are pointed out.Then this thesis introduces the basic knowledge of DL and investigate the latest DL-based downlink channel reconstruction and multi-user scheduling algorithms,which have initially shown their effectiveness and low cost,but there is still room for improvement.Second,to address the problems of high complexity of traditional high-precision downlink channel reconstruction algorithms and low reliability of existing DL-based downlink reconstruction algorithms,this thesis design a fast downlink channel reconstruction algorithm based on the target detection algorithm "You Only Look Once"(YOLO)in DL,YOLO-RecNet.The YOLO-RecNet network architecture is designed based on YOLOv3(the third generation of YOLO)by mapping the channel to a delay-angle domain image,locating the propagation path spots in the channel image,and converting them into frequency-independent parameters:angle and delay,effectively avoiding the complex estimation of frequency-independent parameters in traditional high-precision downlink reconstruction algorithms.Next,the downlink path complex gain is estimated at the user side and fed back to the base station to complete the downlink channel reconstruction.The simulation results show that the proposed YOLO-RecNet algorithm can complete the downlink channel reconstruction at the millisecond level with similar reconstruction accuracy and spectral efficiency as the high-precision Newtonian orthogonal matching tracking algorithm,which is effective and superior in real-time in the downlink channel reconstruction scheme of large-scale MIMO systems in frequency division duplex mode.Finally,the Attention-USNet algorithm,which introduces the attention mechanism,and the DeepSort-based time-varying channel user scheduling algorithm are designed to address the problems that traditional multi-user scheduling algorithms can hardly guarantee high performance and low complexity at the same time,and the performance of existing DL-based multiuser scheduling algorithms still has room for improvement.We describe the multi-user scheduling system model,demonstrate through analysis that the relationship between user channel vectors plays an important role in influencing the scheduling results,and illustrate the shortcomings of USNet designed based on fully connected neural networks in extracting such relational features.A fully connected neural network with a self-attentive layer is designed to improve the effectiveness of user relevance feature extraction,which can significantly reduce the number of network parameters and reduce the computation and running time while improving the system achievable rate performance.In addition,for the time-varying channel model,which requires higher real-time scheduling performance,we propose a motion prediction algorithm based on the target tracking algorithm DeepSort in DL.In addition,for the time-varying channel model with higher scheduling real-time requirements,this thesis propose a multi-user scheduling algorithm without estimating the channel at each moment based on the prediction function of the target tracking algorithm DeepSort in DL.The performance of the designed scheduling algorithm in terms of system achievable rate performance,number of network parameters,network computation,and running time is analyzed through simulation experiments under various scenarios.The simulation results demonstrate that the proposed Attention-USNet series algorithm can improve the reachable rate percentage of the scheduling system to more than 98%,and can significantly reduce the number of parameters and computation,which significantly reduces the time and computation cost of user scheduling.