基于ResNet的大型智能表面在毫米波系统中的应用

Application of ResNet-based Large Intelligent Surface in Millimeter Wave Systems

大型智能表面(Large Intelligent Surface,LIS)协助毫米波通信已经成为一种提高覆盖率和吞吐量的极具潜力的技术。为了设计LIS系统的反射波束,通常需要获取完美的级联信道状态信息。然而,由于LIS具有高维级联信道和大量无源反射元件,估计其级联信道状态信息一直是LIS的挑战之一。针对上述问题,提出一种基于残差神经网络(Residual Neural Network,ResNet)的LIS反射波束设计解决方案。该方案采用有源(连接LIS控制器的基带)和无源元件混合的LIS框架,只需要估计少量有源元件的信道状态信息便可以借助ResNet训练的网络模型预测最佳发射波束。与现有方法相比,所提出的ResNet网络可以减少训练开销,提升可实现速率,表现出更强的鲁棒性。

Large intelligent surface(LIS)assisting millimeter wave communication has been known as a promising technology to improve coverage and throughput.Designing the reflected beam of the LIS system relies on perfect cascade channel state information(CSI).However,it is very challenging to estimate cascaded CSI due to the fact that LIS has high-dimensional cascaded channels and massive passive reflection elements.To address the aforementioned problems,a residual neural network(ResNet)-based approach for designing reflection beam of LIS is proposed.The scheme adopts a hybrid LIS framework including active(baseband connected to LIS controller)and passive components,which requires to estimate the CSI of only a small number of active components.The best reflected beam can be predicted through trained ResNet model.Compared with the existing methods,the ResNet network can reduce training overhead,improve achievable rate as well as enhance robustness.