Robust Millimeter Beamforming via Self-Supervised Hybrid Deep Learning
Fenghao Zhu, Bohao Wang, Zhaohui Yang, Chongwen Huang, Zhaoyang Zhang, George C. Alexandropoulos, Chau Yuen, Mérouane Debbah
Abstract
Beamforming with large-scale antenna arrays has been widely considered in wireless communications research in recent years, playing a significant role in fifth generation (5G) networks, as also expected to happen in their upcoming sixth generation (6G). To improve its performance, various techniques have been leveraged, e.g., optimization schemes and deep learning. Although the late deployment of deep learning approaches has been proven quite attractive in certain scenarios, it has been showcased that when the environment or the dataset changes, the performance of supervised learning gets severely degraded. Therefore, the design of effective neural networks for beamforming, exhibiting strong robustness, is an open research area for intelligent wireless communication systems. In this paper, we propose a robust self-supervised deep neural network for beamforming, which is tested with two different datasets emulating various wireless deployment scenarios. Our simulation results demonstrate that the proposed self-supervised network with hybrid learning performs sufficiently well in both the DeepMIMO and the new WAIR-D datasets, exhibiting strong robustness under various environments.