| The low latency and high bandwidth of mm Wave communication in 5G network make it possible to realize the high data rate and low latency requirements of Internet of Vehicles(IoV)communication.However,the large path loss and poor penetration of mm Wave make the application of mm Wave communication in IoV face many challenges;beamforming has become one of the key technologies for mm Wave communication to be applied to the IoV.Due to the advantages of strong environmental perception,deep learning applying to the IoV has become a research trend.In this article,we use deep learning to solve the mm Wave beamforming problem for IoV.The main work is as follows:First,this article uses ray tracing technology and simulation tools such as traffic simulators to model the IoV system.Where users can customize the IoV scenes and use the simulation platform to obtain simulation data such as corresponding received power by modifying parameters like vehicle speed,which provides a simulation data basis for the research of different IoV.Secondly,aiming at the low latency and environment perception requirements of beamforming in the suburban high-speed IoV scene,this article studies the beamforming problem based on codebook,and proposes a CBNet neural network to realize the design of the codebook,the proposed algorithm can impose constant modulus constraints and adaptively generate codebooks.The simulation results show that the proposed algorithm can realize environment perception.The performance of DFT codebook 64 beam can be achieved by using only 16 beams,which greatly reduces the training overhead.The proposed algorithm can reach 90%of the upper bound EGC of the system achievable rate,which improves the system achievable rate by 20%compared with the traditional DFT codebook.Finally,aiming at the complex channel of the urban IoV scene and the demand for high accuracy of beamforming.This article studies the adaptive beamforming,decouples the four-variable joint solution problem of the hybrid beamforming matrix into separate solutions of the precoder and combiner,and converts the design of the analog precoder and combiner into a regression problem.The proposed RFNet neural network generates the required matrix for beamforming.The simulation results show that the performance of the proposed algorithm reaches 98%of the best spectral efficiency,which improves the spectral efficiency of the system by 87%compared with the traditional OMP algorithm.In the future,we will focus on using machine learning to perform channel estimation on complex channels of the IoV faster and more accurately to meet the needs of time-varying channel information in the research of IoV. |
