Development of AI/ML Methods for Advanced Device Localization in Beyond 5G Systems

Abstract: This master thesis aims to investigate the positioning accuracy improvements of device localization by the implementation of AI/ML functionalities. In this project, we use the fifth-generation(5G) New Radio (NR) system and focus on indoor factories. This setup held under study is a relevant scenario for the Industrial Internet of Things within the industrial sector (IIoT), as the specific characteristics of this scenario create a disadvantageous environment for effective positioning. It is a typical deployment scenario standardized in Third Generation Partnership Project (3GPP) releases. The study is based on using a provided MATLAB simulator for generating the required information about the User Equipment (UE) locations and their channel measurements with the 5G base stations, the gNodeBs (gNBs). This simulator creates the channel model, environment geometry, and position reference signals aligned with the 3GPP. Pursuing the goal of mitigating the multipath propagation effects, different AI/ML methods have been developed in Python. Several AI/ML models have been explored investigating different inputs, such as the Channel Impulse Response or other significant channel features, as well as various model outputs, such as the direct UE position or intermediate angles or times of the radio signals. Consequently, this project evaluates the positioning performance of the assisted and direct AI/ML positioning versus the legacy methods in terms of accuracy and complexity while considering different deployment strategies. Different scenario configurations have been simulated regarding the generalization ability of the AI/ML methods evaluation. Furthermore, another objective has been studying the actual viability of AI/ML for 5G device positioning and in that case, which direction is more worthwhile for future investigation. Finally, based on the results of this simulation-based evaluation, it has been demonstrated that deploying AI/ML methods on the UE side is advantageous for improving the existing 5G location services in this particular scenario without requiring an excessive computational capacity. This thesis project has investigated several models of different natures and complexity, comparing their performance. Besides, the best generated AI/ML models show a general performance improvement versus the legacy methods from 80% of the distance error CDF. In the most severe Non-Line-of-Sight (NLOS) scenarios, the AI/ML methods have achieved an improvement of more than 10 meters for the 95% CDF compared with the legacy. To conclude, the AI/ML models achieve greater performance for most of the devices than legacy, also offering great results for heavily NLOS situations.