Date of Award

2021

Document Type

Doctoral Thesis

Degree Name

Doctor of Philosophy

Department

Electrical & Electronic Engineering

First Advisor

Dr.-Ing. Bernd-Ludwig Wenning

Second Advisor

Prof. Dr.-Ing. Michael Kuhn

Third Advisor

Dr. Alan McGibney

Abstract

The rapid acceleration into the digital age brings a worldwide interconnectedness of an unprecedented level. This interconnectedness transforms the modern transportation network by linking its infrastructure components (e.g., signaling systems, traffic control systems, maintenance systems) together. Intelligent transportation systems (ITS) can utilize this newly emerged network infrastructure, improving traffic safety, traffic control, and reducing transportation-related environmental impacts, changing the way drivers, businesses and governments deal with road transport. It is expected that in the near future self-aware vehicles (e.g., trains, cars or trucks) will constitute the majority of the transportation infrastructure. Vehicles need to be able to reliably communicate critical sensor data to the back-end infrastructure, where systems on the cloud or the network's edge can utilize this data to deliver maximum safety, effective traffic management, or reliable asset monitoring. Moreover, the modern passenger expects the vehicle to offer uninterrupted entertainment services. Hence, a reliable uplink connectivity becomes necessity. For this reason, the vehicle of tomorrow needs not only be equipped with multiple wireless interfaces, but to have the means for reliable and efficient management of such interfaces. In this context, this thesis presents the adaptive similarity-based regressor (ASR) - an online machine learning method with similarity-based forgetting strategy. It is shown the approach can leverage commonly available parameters of commercial modems with application in ITS. The algorithm can adapt to the various properties of different mobile networks, while maintaining a fixed training set size. With as little as 24 training samples it outperforms state of the art traditional offline based machine learning (ML) algorithms by up to 2:4-times in terms of root-mean-squared error (RMSE), and up to 6:7-times in terms of mean-relative error (MRE). The ASR can be deployed across different mobile networks without the need of pre-training and is developed as a set of modular components and as such can be used as an enhancement for network or transport layer protocols. The modularity of the approach allows the ASR to enhance an intelligent interface selection policy called multi-armed bandit adaptive-similarity based regressor (MABASR) to select the best wireless link without performing active measurements. The policy enables an agent to learn the network properties from scratch and to react to unexpected network congestion ensuring end-to-end link reliability. By exploiting the linear correlation between uplink data rate and channel quality parameters the MABASR achieves up to 230-times lower performance deviation compared to other interface selection techniques. For the further improvement of the end-user experience the thesis proposes an early warning system for channel fading: The ASR is combined with a neural network (NN) for sequence prediction, which enables the forecasting of near future uplink data rates. To support future developments all datasets were published online and made available to the research community.

Creative Commons License

Creative Commons Attribution-Share Alike 4.0 International License
This work is licensed under a Creative Commons Attribution-Share Alike 4.0 International License.

Access Level

info:eu-repo/semantics/openAccess

Download

Share

COinS