Frank M. Knowledge-Driven Harmonization of Sensor Observations 2020 torrent

Textbook in PDF format

The rise of the Internet of Things (IoT) leads to an unprecedented number of continuous sensor observations that are available as IoT data streams. It seems to be obvious to employ this new source of data for better founded decision support in various domains. However, harmonization of such observations is a labor-intensive task due to heterogeneity in format, syntax, and semantics. We therefore aim to reduce the effort for such harmonization tasks by employing a knowledge-driven approach. In order to avoid having to build up a new knowledge base for each harmonization task, we pursue the idea of exploiting the large body of formalized public knowledge represented as statements in Linked Open Data (LOD) for this purpose.
This approach reveals three challenges: i) we have to establish trust for at least a subset of LOD in order to ensure that statements employed for the harmonization process are consistent and trustworthy, ii) we have to handle sensor observations contained in IoT data streams with respect to the dimensions of volume, veracity,
velocity, and variety and iii) we have to address varying data requirements thatare given for varying use cases and target decision support systems (DSSs). We address these challenges by i) enabling knowledge workers to collaboratively curate and annotate knowledge and leverage it using common knowledge pub­lished as LOD, ii) mapping key-value tuples of observations contained in IoT data streams to meaningful and validated triples on-the-fly, and iii) providing dynamic harmonization workflows that automatically adapt to the requirements of different data consumers based on the context knowledge of an observation.
Our approach is evaluated within the domain of geographical information systems (GISs). The results show that i) the informative value of knowledge bases can be leveraged by LOD if knowledge about schema and provenance is evaluated precisely, ii) mapping, transformation, and validation of continuos environmental observations can be efficiently provided using current stream processing tech­ nologies and iii) machine learning algorithms are suited to dynamically compose efficient preprocessing workflows that meet varying requirements of various data consumers