基于深度学习的多用户Massive MIMO预编码方法

Deep Learning-Based Precoding Design for Multi-user Massive MIMO Systems

在下行链路传输场景中,发射机处的功率分配和波束赋形设计至关重要。考虑一个多用户Massive MIMO系统中总功率约束下最大化加权和速率问题,经典的WMMSE算法可以获取次优解,但计算复杂度过高。为了降低计算复杂度,提出了一种基于深度学习的快速波束赋形设计方法,该方法可以离线训练深度神经网络,利用训练后的神经网络求解最优波束赋形解,只需要简单的线性和非线性操作即可完成。实验结果显示,该方法可以逼近WMMSE算法精度的90%以上,同时计算复杂度和时延也大大降低。

In a downlink transmission scenario,the design of power allocation and beam assignment at the transmitter is critical.This paper considers a problem of maximizing the weighting rate under the total power constraint in a multi-user massive MIMO system,where the classical weighted minimum mean square error (WMMSE) algorithm can obtain suboptimal solutions with high computational complexity.In order to reduce the computational complexity,this paper proposes a fast beamforming design method based on deep learning,which can train a deep neural network offline and use the trained neural network to solve the optimal beamforming with simple linear and nonlinear operations.Experimental results show that the method can approximate more than 90% accuracy of the WMMSE algorithm,while the computational complexity and delay are greatly reduced.