Vision-Assisted User Clustering for Robust mmWave-NOMA Systems

TL;DR

This paper proposes a vision-assisted user clustering approach for mmWave-NOMA systems that enhances robustness by reducing reliance on CSI, using camera feeds and deep learning to maintain effective clustering even with outdated or unavailable CSI.

Contribution

It introduces a novel vision-assisted clustering method that combines camera data and non-CSI feedback, improving robustness over traditional CSI-dependent approaches.

Findings

Vision-assisted clustering achieves comparable performance to CSI-based methods.

Clustering remains effective even with severely outdated or missing CSI.

Deep learning on camera images enables robust user grouping in mmWave-NOMA.

Abstract

When operated in the mmWave band, user channels get highly correlated which can be exploited in mmWave-NOMA systems to cluster a set of "correlated" users together. Identifying the set of users to cluster greatly affects the viability of NOMA systems. Typically, only channel state information (CSI) is used to make these clustering decisions. When any problem arises in accessing up-to-date and accurate CSI, user clustering will not properly function due to its hard-dependency on CSI, and obviously, this will negatively affect the robustness of the NOMA systems. To improve the robustness of the NOMA systems, we propose to utilize emerging trends such as location-aware and camera-equipped base stations (CBSs) which do not require any extra radio frequency resource consumption. Specifically, we explore three different dimensions of feedback that a CBS can benefit from to solve the user clustering problem, namely CSI-based feedback and non-CSI-based feedback, comprised of user equipment (UE) location and the CBS camera feed. We first investigate how the vision assistance of a CBS can be used in conjunction with other dimensions of feedback to make clustering decisions in various scenarios. Later, we provide a simple user case study to illustrate how to implement vision-assisted user clustering in mmWave-NOMA systems to improve robustness, in which a deep learning (DL) beam selection algorithm is trained on the images captured by the CBS to perform NOMA clustering. We demonstrate that user clustering without CSI can achieve comparable performance to accurate CSI-based solutions, and user clustering can continue to function without much performance loss even in the scenarios where CSI is severely outdated or not available at all.