Research On Massive MIMO Hybrid Beamforming Based On Deep Learning

With the continuous development of communication technology,millimeter wave,massive multiple-input multiple-output(Massive MIMO)and beamforming have received extensive attention.The expansion of the frequency band to the millimeter wave band solves the problem of scarcity of spectrum resources,and the shorter wavelength of millimeter wave facilitates the integration of antennas,thus realizing massive MIMO technology.On the one hand,although the traditional full-digital beamforming has better performance,the cost is too high,while the analog beamforming has lower cost but poor performance,and the hybrid beamforming technology emerges as the times require.Hybrid beamforming achieves an effective compromise between system cost and performance,but the system performance is still insufficient compared to digital beamforming.On the other hand,affected by the hybrid architecture in actual communication,there is a rank-deficient problem in the channel,which will lead to further loss of system performance.Therefore,in this paper,under the millimeter wave massive MIMO system,a deep learning algorithm is used to solve the problem of insufficient hybrid beamforming performance and the rank-deficient of the practical communication channel.In order to further improve the performance of the hybrid beamforming algorithm,two Generative Adversarial Network(GAN)algorithms,which are digital beamforming algorithms based on conditional GAN and hybrid beamforming algorithms based on dual generators,are proposed to optimize the hybrid beamforming performance and maximize the system spectral efficiency based on multi-user massive MIMO hybrid architecture system.The digital beamforming algorithm based on conditional GAN takes the analog beamforming matrix as the input of the network to generate the required digital beamforming matrix.The hybrid beamforming algorithm based on Dual-Generator Generative Adversarial Network(DGGAN)takes the channel matrix as the input of the network,and the two generators generate the analog beamforming matrix and the digital beamforming matrix,respectively.The simulation results show that the spectral efficiency of the proposed two algorithms is better than other algorithms in the considered simulation conditions,and the performance is closer to the upper bound of full-digital beamforming.Especially the DGGAN method,its spectral efficiency is improved by 10.74%-20.59%compared with the traditional matrix-based algorithm,and compared with the existing deep learning network architecture,the performance is improved by 4.67%-8.31%.In order to combat the performance loss caused by the inherent channel rank-deficient of the practical hybrid architecture massive MIMO system,based on the multi-user massive MIMO hybrid architecture system under rank-deficient channels,two new GAN algorithms are proposed,which are a channel compensation algorithm based on GAN and a rank-deficient channel compensation and hybrid beamforming joint optimization scheme based on multi-generators,to realize the hybrid beamforming design under rank-deficient channels.Based on the rank-deficient channel compensation algorithm of GAN,the rank-deficient channel is used as the input of the network,and the generator compensates the missing information to obtain the compensated channel matrix.The joint optimization scheme of rank-deficient channel compensation and hybrid beamforming based on multi-generator GAN,the spectral efficiency is directly regarded as the optimization goal of the network,and the expected system performance is obtained.The simulation results show that the performance of the proposed two algorithms is better than other algorithms in the considered simulation conditions.The channel compensation algorithm improves the system performance by 38.81%-51.49%.The joint optimization algorithm improves the performance by 48.11%-77.79%compared with the traditional algorithms,and the performance is improved by 9.40%-52.44%compared with other deep learning algorithms,which reflects the robustness of the proposed algorithms.