Site-Specific Beam Alignment without Explicit Channel Knowledge via Deep Learning

Quickly and accurately steering the highly directional beams of a millimeter-wave (mmWave) base station (BS) toward the intended user(s) is a challenging task known as the beam alignment problem. Prior work has identified machine learning (ML) as a promising tool for developing fast, effective beam alignment solutions through site-specific optimization. However, these approaches typically rely on training datasets containing the propagation channel information between the BS and virtually every possible user within the site. Acquiring such datasets is impractical through channel estimation and time-consuming when using ray-tracing tools. In this paper, we propose a novel ML-driven framework that relies solely on measurements of received signal power, collected by users during standard 5G beam sweeping procedures, for training. By leveraging these beam sweeping measurements, our model implicitly learns features of the surrounding environment, enabling the construction of a sparse probing beam codebook and a corresponding beam selector for rapid beam alignment. Remarkably, our approach matches the speed and effectiveness of state-of-the-art methods trained on perfect explicit channel data, even when the beam sweeping measurements used for training are noisy.