Millimeter wave (mmwave) fixed wireless access is a key enabler of 5G and beyond small cell network deployment, exploiting the abundant mmwave spectrum to provide Gbps backhaul and access links. Large antenna arrays and extremely directional beamforming are necessary to combat the mmwave path loss. However, narrow beams increase sensitivity to physical perturbations caused by environmental factors. To address this issue, in this paper we propose a predictive transmit-receive beam alignment process. We construct an explicit mapping between transmit (or receive) beams and physical coordinates via a Gaussian process, which can incorporate environmental uncertainties. To make full use of underlying correlations between transmitter and receiver and accumulated experiences, we further construct a hierarchical Bayesian learning model and design an efficient beam predictive algorithm. To reduce dependency on physical position measurements, a reverse mapping that predicts physical coordinates from beam experiences is further constructed. The designed algorithms enjoy two folds of advantages. Firstly, thanks to Bayesian learning, a good performance can be achieved even for a small sample setting as low as 10 samples in our scenarios, which drastically reduces training time and is therefore very appealing for wireless communications. Secondly, in contrast to most existing algorithms that only predict one beam in each time-slot, the designed algorithms generate the most promising beam subset, which improves robustness to environment uncertainties. Simulation results demonstrate the effectiveness and superiority of the designed algorithms against the state of the art.

中文翻译：

---

毫米波固定无线接入链路中基于学习的预测性发射机-接收机波束对准


毫米波 (mmwave) 固定无线接入是 5G 及小型蜂窝网络部署之外的关键推动者，利用丰富的毫米波频谱提供 Gbps 回程和接入链路。为了应对毫米波路径损耗，需要大型天线阵列和极其定向的波束成形。然而，窄光束增加了对环境因素引起的物理扰动的敏感性。为了解决这个问题，在本文中，我们提出了一种预测性发射-接收波束对准过程。我们通过高斯过程构建发射（或接收）波束和物理坐标之间的显式映射，该映射可以包含环境不确定性。为了充分利用发射机和接收机之间的潜在相关性以及积累的经验，我们进一步构建了分层贝叶斯学习模型并设计了有效的波束预测算法。为了减少对物理位置测量的依赖，进一步构建了根据波束经验预测物理坐标的反向映射。设计的算法具有两个优点。首先，得益于贝叶斯学习，即使在我们的场景中低至 10 个样本的小样本设置也可以实现良好的性能，这大大减少了训练时间，因此对于无线通信非常有吸引力。其次，与大多数现有算法仅预测每个时隙中的一个波束相比，所设计的算法生成最有希望的波束子集，从而提高了对环境不确定性的鲁棒性。仿真结果证明了所设计算法相对于现有技术的有效性和优越性。