Millimeter Wave Beam Selection in Time-varying Channels with Orientation Changes and Lateral Mobility

Abstract

Beamforming enables millimeter-wave (mmWave) communications to achieve high data rates in 5G and beyond systems. This requires the use of many narrow directional beams at both the transmitter and receiver to overcome the adverse propagation conditions in mmWave channels. However, accurate beam alignment incurs significant training overhead. Changes in user device orientation and mobility can quickly result in beam misalignment, reducing the data rate. They also make the beam gains a non-stationary random processes. We first present a novel modified bivariate Nakagami-m (MBN) model to tractably and accurately characterize the joint, non-stationary statistics of the channel gains seen at the times of measurement and data transmission. This model accurately captures the widely used mmWave spatial channel model (SCM), which is realistic but uses a geometry-based method to construct the channel. We use the MBN model to propose a near-optimal, practically amenable bound based selection (PABS) rule. Our approach captures several pertinent aspects about the spatial channel model and 5G, such as transmission of periodic bursts of reference signals, feedback from the user to enable the base station to select its transmit beam, and the faster pace of updating the data rate compared to the transmit-receive beam pair. The PABS rule markedly outperforms the widely used conventional power-based selection rule and is less sensitive to user orientation changes. However, this method needs angle of arrival (AoA) at the user equipment (UE). We then propose a comprehensive and novel approach called latent Thompson sampling-based beam selection (LTBS), which combines latent Thompson sampling to track the AoA as a latent state, receive beam subset selection based on the sampled AoA in a manner compliant with the 5G new radio standard, rate adaptation, and data beam selection based on predicted throughput. We propose two variants of LTBS that trade-off between complexity and accuracy in modeling beam pair gains. The prior update and channel gain prediction in one of the variants are based on SCM. We propose variations that employ windowing to also tackle lateral user mobility, which alters the AoA and the channel statistics. Our numerical results show that the proposed methods track the AoA in a manner robust to user orientation changes and provide higher average data rates compared to conventional and state-of-the-art learning-based beam selection methods.