Machine learning-assisted mmWave beam management

Millimeter wave (mmWave) devices need to leverage highly directional beamforming (BF) to overcome the higher isotropic path loss. On the other hand, such narrow beams are sensitive to the propagation conditions including blockage and reflections. As a result, beam management – finding and maintaining good analog BF directions – is critical to enabling communication at the mmWave spectrum. This dissertation will focus on designing beam management solutions for mmWave systems that can find near-optimal beams with low overhead and latency.

In the first part of this dissertation, a machine learning (ML)-aided beam alignment method is proposed where ML models are trained to predict candidate beams and serving base stations (BSs) using only the location information of user equipments (UEs) as context information. At the cost of only a small overhead in uplink feedback of a UE’s coordinates through lower-frequency links, the proposed method can reduce the search space by approximately 4× for the optimal BS and over 10× for the optimal beam, even in a dynamic environment with imperfect UE coordinates. A dataset modeling a realistic, generalizable environment is created using a state-of-the-art commercial ray-tracing software and published to train and validate the ML models.

To further enhance the ease of adoption without modifications to the existing cellular network standards, a 5G-compatible beam alignment method that uses a site-specific probing codebook to predict candidate beams is proposed in the second part of this dissertation. The probing codebook and the beam predictor are jointly trained with a novel neural network (NN) architecture. By sweeping a small learned codebook that is adapted to the propagation environment, the proposed NN beam predictor can accurately select the optimal narrow beam while reducing the beam sweeping overhead by as much as 14× in challenging non-line-of-sight scenarios.

The third part of this dissertation further explores the idea of site-specific probing, and proposes a grid-free beam alignment approach that uses the measurements of a few probing beams to directly compute arbitrary BF weights for each UE from the continuous search space. The probing beams and the beam synthesizer functions are jointly trained in a novel deep learning pipeline so that UEs can both be discovered with high probability and achieve high BF gain. The proposed method is better than the exhaustive search by orders of magnitude in terms of the trade-off between signal-to-noise ratio (SNR) and beam alignment speed. It also improves upon the approach proposed in the second part by eliminating the per-UE search and achieving higher SNR than standard codebooks of narrow beams.