Beam Training in mmWave Vehicular Systems: Machine Learning for Decoupling Beam Selection

TL;DR

This work tackles the overhead of codebook-based mmWave beam training in dynamic vehicular networks by introducing location-based ML methods to decouple beam selection at the BS and UE. It analyzes three scenarios—two decoupled (with and without UE location) and one coupled—using throughput-ratio and ATR-based ML models trained on ray-traced urban-channel data at 28 GHz. The results show that decoupled beam selection with UE location nearly matches joint BS-UE beam selection, while decoupled without location can be compensated by clustering-based strategies to regain performance, especially as more beam pairs are swept. The findings suggest practical, scalable beam-training strategies for heterogeneous vehicular devices that leverage readily available localization and robust statistical techniques to cope with dynamic urban channels.

Abstract

Codebook-based beam selection is one approach for configuring millimeter wave communication links. The overhead required to reconfigure the transmit and receive beam pair, though, increases in highly dynamic vehicular communication systems. Location information coupled with machine learning (ML) beam recommendation is one way to reduce the overhead of beam pair selection. In this paper, we develop ML-based location-aided approaches to decouple the beam selection between the user equipment (UE) and the base station (BS). We quantify the performance gaps due to decoupling beam selection and also disaggregating the UE's location information from the BS. Our simulation results show that decoupling beam selection with available location information at the BS performs comparable to joint beam pair selection at the BS. Moreover, decoupled beam selection without location closely approaches the performance of beam pair selection at the BS when sufficient beam pairs are swept.

Figures (5)

• Figure 1: Multi-output throughput ratio regression models are shown. Model (A) represents the regression model for scenario 1. Model (B) represents the regression model for the BS or a UE in the scenario 2 and a UE in the scenario 3, where M denotes the beam codebooks, $\mathcal{F}$ or $\mathcal{W}$.
• Figure 2: One snapshot of the urban street environment created in Blender and simulated in Sionna. It represents the region of interest in the urban street. The buses are blockers and support NLOS paths from the BS and UEs.
• Figure 3: Throughput ratio comparison for the three scenarios. The number of UE beams, $|\mathcal{S}_\text{w}|$ is set to 5 for the decoupled scenarios. The increasing number of beam pairs is due to the increasing number of selected BS beams. Decoupling beam selection with location has a minor throughput ratio decrease of less than $5\%$. Disaggregating location information has a higher decrease, whereas the proposed solutions achieve a comparable throughput ratio in sweeping at least $100$ beam pairs that are $\approx 10\%$ of total beam pairs.
• Figure 4: Throughput ratios for scenario 2 and scenario 3. The horizontal and vertical axes represent the number of the selected UE and BS beams. Heatmaps are useful to identify the required number of BS and UE beams to achieve a certain performance. For example, $4$ BS beams and $3$ UE beams are sufficient to achieve $90\%$ throughput ratio in DBSwL whereas DBSwoL requires $19$ BS beams and $3$ UE beams, which is equivalent to say $45$ more beam pairs are required compared to DBSwL.
• Figure 5: Misalignment probability comparison for the three scenarios. The number of UE beams, $|\mathcal{S}_\text{w}|$ is set to 5 for the decoupled scenarios. There is a large gap between the coupled scenario 1 and decoupled scenario 2 and 3. The decoupled scenarios might not necessarily yield the best beam pair since they are designed to achieve a high throughput ratio, not the best beam pair.