【摘要】 针对频分双工(Frequency Division Duplex, FDD)模式下大规模多输入多输出(Multiple-Input Multiple-Output, MIMO)系统的最佳波束成形设计，存在信道状态信息(Channel State Information, CSI)反馈开销严重和单独考虑某一模块难以得到整体最优解的问题，提出了一种基于有限反馈的端到端神经网络架构。首先，在基站侧设计可优化导频，并通过稀疏信道发送给用户。然后，用户将接收到的导频通过深度神经网络量化为B比特的信息流，并反馈给基站。最后，基站利用反馈信息，通过图神经网络完成最佳波束成形矩阵的设计。该架构将导频设计、量化、反馈以及最佳波束成形等不同模块整体考虑、优化。实验结果表明，所提的端到端神经网络模型在导频长度和反馈开销有限的情况下，能实现的用户和速率已经达到了具有全CSI的传统波束成形方案的96%。更多还原

【Abstract】 For the optimal beamforming design of the massive multi-input multiple-output(MIMO) system in the frequency division duplex(FDD) mode, several challenges are encountered, including the heavy overhead of channel state information(CSI) feedback and the difficulty of attaining the overall optimal solution when considering individual modules in isolation.To solve above problems, an end-to-end neural network architecture is proposed.Firstly, the optimized pilots are designed at the base station side and sent to the user through a sparse channel.Then, the user quantizes the received pilots into B information bits through the deep neural network and provides feedback to the base station.Finally, the base station uses the feedback information to design the optimal beamforming matrix through the graph neural network.The architecture considers and optimizes different modules as a whole, including pilot design, quantization, feedback, and optimal beamforming.The experimental results show that the proposed end-to-end neural network model with short pilot sequences and limited feedback overhead can achieve 96% of the sum-rate of the traditional beamforming scheme with full CSI.更多还原