REPOSITORY > RESULTS

Abstract

Due to the nature of the wireless transmission medium, wireless communications are characterised
by notably larger losses of data packets than wired communications. The quality
of wireless links is highly dependent on channel variations, interference and even transceiver
imperfections. Such link uncertainty instigated the development of numerous techniques
that can withstand uncertain conditions by adjusting the parameters of the wireless link
to achieve higher reliability or selecting a more reliable alternative wireless link for data
transmission. These techniques rely on effective wireless link quality estimation.
Analytical approach, initially used for link quality estimation, was soon complemented
and superseded by statistical and more recently machine learning approaches based on
empirical data traces, resulting in data-driven models. Statistical approach fits the model
to the underlying distribution of the specific property or behaviour under investigation,
whereas machine learning models transfer link quality estimation into classification problem,
potentially taking into account multiple phenomena of the link, thus being better
suited for the real-world wireless links that exhibit dynamic behaviour and are often subject
to various transient phenomena. Observation of wireless link quality is important also
from the perspective of early anomaly detection, especially in large scale industrial or commercial
deployments. Automatic detection of malfunctions, caused by software, hardware,
or external factors in dynamic operating environment, can be an important asset to reduce
unexpected maintenance downtime, and consequently financial losses.
In this dissertation we are concerned with classification of wireless links using machine
learning techniques to support link quality estimation and anomaly detection. Our main
attention is given to the challenges of designing wireless link classifiers based on machine
learning techniques. In the first part, focused on link quality estimation, we perform
in-depth quantitative research on how each step of feature engineering, data engineering
and algorithm tuning influences the estimation performance. We pay special attention
to improving the detection of minority classes, i.e. less frequent data samples in the
wireless link dataset. We propose a new supervised classifier for link quality estimation
that improves detection of minority class by over 40% through feature selection, and by
over 20% through data re-sampling strategies, without any significant impact on detecting
majority classes.
The second part of the dissertation is focused on wireless link anomaly detection, where
we define four basic types of anomalies that occur on wireless links and describe their symptoms
and probable causes. With a systematic quantitative approach, we investigate the
performance of two threshold-based approaches, three supervised and three unsupervised
reference machine learning algorithms. We show that the performance of supervised approaches
may be dominant, however, certain unsupervised approaches combined with deep
learning autoencoders for input features come close to the performance of supervised approaches
while not requiring annotated data, which may prove as a significant advantage.