Massive MIMO and Millimetre Wave Technologies: Design, Application and Integration with ML Techniques for 5G and Beyond Networks

Abstract

Massive Multiple-Input-Multiple-Output (mMIMO) and millimetre Wave (mmWave) are key enabling technologies to realise omnipresent, scalable and dynamic 5-th Generation (5G) and beyond networks that provide seamless wireless connectivity supporting enhanced data rates. Fundamental challenges for the full-scale deployments of these technologies are the possibility to integrate different services, e.g. wireless access and backhaul, and the support of applications like industrial robotics demanding a reliable wireless network service. Significant enablers for rapid adoption are the utilisation of self-Backhauling (s-BH) technologies to integrate access and backhaul services in the same spectrum resources and the adoption of Machine Learning (ML) techniques to add predictive features to the control operation of the network.

In this dissertation, we investigate a mMIMO-based s-BH network architectures to facilitate cost-effective network deployments of Small Cells (SCs), and we propose a ML-based beam recovery method to mitigate dynamic blockages and provide reliable wireless connections in mmWave indoor networks. Our work focuses on addressing the following research questions: (1) What is the potential of using mMIMO-based s-BH network architectures to provide backhaul links to SCs and favour Ultra Dense Network (UDN) deployments? (2) How to improve the mmWave network data rate stability by making the mmWave links more robust to blockage events? (3) How to design novel radio control methods by integrating mmWave systems with ML models to provide beam state predictions? We adopt several simulation and analytical tools, such as system-level simulations and stochastic geometry analysis, that evaluate large-scale networks performance. Moreover, we employ ML models such as Deep Neural Networks (DNN) to make predictions about the network environment.

First, we investigate through comprehensive Third Generation Partnership Project (3GPP)-based system-level simulations and analytical formulations the User Equipments (UEs) data rate performances of the mMIMO-based s-BH network architecture and compare these results to the mMIMO Direct Access (DA) network architecture, in which mMIMO-Base Stations (BSs) directly serve UEs in the absence of SCs. In this work, we propose a novel deployment of SCs, namely ad-hoc deployment, where SCs are positioned close to UEs to achieve Line-of-Sight (LoS) access links. Then, we evaluate the optimal partition between backhaul and access resources that maximises the end-to-end UE data rates. We show that ultra-dense SCs deployments supported by mMIMO s-BH provide rate improvements for cell-edge UEs that amount to 30% and a tenfold gain compared to mMIMO DA solutions with pilot reuse scheme 3 and reuse scheme 1, respectively. On the other hand, mMIMO s-BH underperforms mMIMO DA above the median value of the UE data rates when the effect of pilot contamination is less severe and DA links achieve a higher LoS probability.

Next, motivated by the need to improve the Sub-6 GHz links capacity further and to provide more stable wireless links, we moved our focus on the blockage problem in mmWave indoor networks, considering a DA network architecture. In this scenario, the mmWave links experience rapid and temporary fluctuations of the received signal power when they encounter blocking objects, such as humans and robots moving in the environment. We propose a novel method that performs the beam switching in advance based on the predictions provided by beam-specific DNN models and avoids the delay introduced by the method based on a detection threshold. We show that during the blocked intervals, the prediction-based method guarantees higher signal level stability and up to 82% data rate improvement to the detection-based method when blockers move at a speed of 2 m/s. In these conditions, the higher frequency of beam switching penalises the detection-based method's data rate while the prediction-based method achieves better results due to the ability of generalise well with different blocker speeds.

Finally, we show an example of a ML workflow for implementing the prediction-based method in the Open-Radio Access Network (O-RAN) architecture. This is because Software (SW)-based components and the use of a ML workflow can be the two main directions to facilitate tuning and optimisation of network parameters and the ML models training.