BsNet: A Deep Learning-Based Beam Selection Method for mmWave Communications

Abstract

Millimeter wave (mmWave) techniques have attracted much attention in recent years owing to features such as substantial bandwidth for communication, and it has applications in radar systems and location applications. To compensate for the severe path loss in mmWave bands, beamforming techniques with a massive antenna array are usually employed to provide high directivity. However, the resulting high-gain and narrow pencil beam make the beam alignment costlier and much more difficult. Hence, conducting beam alignment with a low overhead becomes critical. Herein, we propose a promising solution that does not require channel knowledge and treats the beam selection as an image reconstruction problem; thus, deep neural networks can be employed to operate the beam domain image reconstruction. This approach can be divided into two stages: off-line training and on- line prediction. The overhead of the on-line beam selection can be significantly reduced via off-line Eigen-beam extraction without degrading the beamforming performance. Simulations are conducted to confirm the performance of the proposed framework in scalability and robustness.