Research On Intelligent Hybrid Beamforming Technology In Millimeter Wave Wireless Communication

In recent years,millimeter wave has attracted much attention because of its huge available bandwidth.Preliminary capacity estimates indicate that the large bandwidth of the millimeter wave will help meet the throughput requirements of future mobile networks.However,due to the high frequency and short wavelength of the millimeter wave signal,its diffraction ability is poor and the transmission loss is very serious.In order to deal with this problem,a large-scale antenna array beamforming technology can be used to concentrate the signal energy in the target direction to compensate for the path loss of millimeter wave communication.The traditional Sub-6G system mainly uses an all-digital beamforming architecture,in which each antenna unit is equipped with a dedicated radio frequency chain.However,in millimeter-wave massive MIMO systems,due to hardware cost and power consumption limitations,it is very difficult to implement such an architecture.Hybrid beamforming is a popular technique that can be used to partition beamforming into digital domain and analog domain.In the analog domain,multiple phase shifters are provided for each radio frequency chain to connect multiple antennas for analog beam forming? in the digital domain,digital beam forming is used to eliminate inter-stream interference.The traditional hybrid beamforming has problems such as high complexity,strong dependence on complete channel state information,and high beam training overhead.This article will combine deep learning and reinforcement learning to solve the above problems.The main tasks are as follows?1.Aiming at the high complexity of traditional hybrid beamforming algorithms,a hybrid beamforming algorithm based on convolutional neural networks is proposed.The problem of solving the analog beamforming matrix is constructed as a multi-label classification problem,and useful information is learned from the historical beamforming solution results,so that the algorithm can directly use the channel state information to obtain the corresponding optimal analog beamforming matrix.The simulation results show that the hybrid beamforming algorithm based on convolutional neural network can achieve performance similar to that of the all-digital beamforming algorithm.2.Aiming at the problem that most of the current hybrid beamforming algorithms have strong dependence on complete channel state information,a beamforming algorithm based on deep neural networks is proposed.The estimated channel information is used as the input of the network,and the loss function of the network is designed directly according to the real spectrum efficiency of the system,so that the network can learn how to perform beamforming under incomplete channel state information to maximize the system's spectrum efficiency.The simulation results show that,when there is no complete channel state information,the beamforming algorithm based on deep neural network has better robustness to the error of channel estimation than the traditional beamforming algorithm.3.Aiming at the problem of high overhead of traditional beam training methods,a beam training algorithm based on reinforcement learning DQN is proposed.The algorithm does not require prior information of the channel model,and does not need to collect training samples in advance.The beam training problem is established as a Markov decision process.Through interaction with the environment,the dynamic change information in the environment is learned in a trial and error manner,and the beam set used for beam training in each time slot can be adjusted adaptively.Thus,the beam training overhead is greatly reduced.The simulation results show that the beam training algorithm based on reinforcement learning DQN performs better than the traditional exhaustive search algorithm,considering the performance indicators of the beam alignment rate and the effective reachable rate of the system.