Predictive Quality of Service for Enhanced Wireless Vehicular Applications

QC 20241014

Abstract

In recent years, the rapid advancement of emerging technologies has significantly fueled the expansion of the Internet of Things (IoT) and the increase of wireless connected devices in society. IoT devices with a safety-critical nature, such as remotely operated vehicles, demand a high quality of the communication service to function reliably. However, consistently achieving high quality of service (QoS) can be challenging for vehicles with high mobility due to changes in interference and the propagation environment, which may cause fluctuations in the statistical distributions that govern the channel and that model the wireless communication performance. This makes it very difficult to predict future communication conditions and performance, such as the wireless channel conditions and QoS metrics. Yet, by forecasting these factors, cellular networks can transition from a reactive approach and become more proactive. By anticipating performance degradations and allocating resources accordingly, safety-critical applications can operate without disruptions.

Unfortunately, the rapid fluctuations of the statistical parameters make it very difficult to predict them by model-based methods. Recently, research has highlighted the potential of machine learning (ML) to constitute models for predictive QoS (pQoS). ML algorithms can learn from vast amounts of data and identify complex patterns that may not be apparent from traditional methods. Its capability to adapt to new data by dynamical model updates makes ML particularly suitable for environments where the QoS is constantly changing. By leveraging ML for predictive purposes, network operators can ensure more efficient resource allocation and a robust network infrastructure. 

The first part of this thesis provides an essential overview of the dynamics in wireless communication systems, focusing on the wireless channel and QoS. The foundations of how ML learns from datasets along with an overview of popular deep neural networks (DNNs) are presented. We summarize the course of our research including a survey on wireless channel prediction, a novel pQoS model, and a QoS prediction framework along with a network digital twin (NDT). A summary of the principal contributions from our research concludes the overview of the thesis.

In the second part of this thesis, we report our major results. We introduce the innovative pQoS model, specifically for connected vehicles, which creates geographical segments, clusters the segments, optimizes the number of clusters, and trains a pQoS model for each cluster using federated learning (FL). We show how this predictive framework improves approaches commonly implemented in previous research, only considering one global predictive model. Moreover, an overview of wireless channel prediction is provided together with an extensive numerical evaluation of DNNs for the purpose of channel prediction, addressing the gap in previous research. Finally, a proof of concept of a real-time NDT based on experimental data is presented to predict the QoS in an enterprise process.

urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-354816