Abstract |
Over the last decades Wireless Sensor Networks (WSNs) have attracted great attention, as they constitute a key enabling technology for implementing sophisticated services in numerous application domains, including area and environmental sensing, health care monitoring, and industrial control systems. Despite their wide applicability, WSNs suffer from network and energy imperfections, which inevitably often lead to missing measurements. The resulting low volume of available data dramatically affects subsequent processing and learning tasks, such as detection of unusual events, clustering, and classification. In this thesis, we address the problem of missing WSN data by employing two non-conventional techniques, which are capable of recovering measurements
in a reliable fashion, namely: a)
Matrix Completion (MC), and b) Tensor Completion (TC). The key theoretical principle adopted is
that a complex signal can be recovered from a small number of random measurements, by
exploiting the underlying redundancies of the sensing data. However, this assumption is not
satisfied in real-life, and often, noisy datasets, which tend to be full rank. We tackle this limitation
by introducing the concept of appropriately forming the available data streams into low-rank 2D
and
3D structures, thereby enabling the utilization of MC and TC in the WSN domain.
To test the efficacy of our approach, we experiment on two prominent fields, namely WSN
-
based
Smart Water Management (SWM) and Human Activity Recognition (HAR). We synthesize their
respective processing and classification frameworks, which encapsulate our proposed modules
for data sampling, structuring, and recovery. These frameworks are evaluated against numerous
aspects, related to the quality of reconstruction on different
volumes of missing data, the accuracy
of subsequent analysis (e.g. classification), and the impact of sub-sampling on the network's
lifetime. Our analysis highlights the interaction of different recovery scenarios in terms of data
structuring and origin,
with several state-of-the-art classifiers. The results demonstrate that high
reconstruction accuracy can be achieved through the developed modules, even for the case of
extremely under-sampled, multi-modal streams of data, lacking up to 80% of their measurements.
|