Contextualising sensors with linked data to improve relevancy, data quality and network adaptability
MetadataShow full item record
This item's downloads: 2036 (view details)
A sensor taken by itself, with no indication of what it is sensing, with what unit of measurement and where, would keep working without anybody being able to interpret its output. The sensor needs to be "contextualised". In this thesis we propose a sensor data modeling based on the Linked Data principles. We support its uptake by releasing LD4Sensors, a Web application that facilitates manual (GUI) and automated (REST API) sensor annotation, storage and retrieval using our data model. A side gain of this approach consists of automating the (currently manual) setup of several sensor settings, which will be presented in a demonstrative application and architecture. Finally, we demonstrate the advantages of using our data model to improve sensor relevancy and enrich Web data, within two different applications. First, we use our data model to improve the relevancy prediction of sensors during Daily Activity Logging tasks. Currently, Task Logging is performed by classifying the previously collected sensor readings in order to identify which task/s they were measuring. This approach has two downsides. First, the task identification happens after all the data has been recorded. Second, selecting which of the available sensors to query is difficult. Usually, sensors are selected according to their energy consumption or location. However, a more fine-grained selection would improve the system efficiency by reducing the amount of data to record an process while at the same time, saving energy. As the amount of Internet-connected objects increases and as we move towards ubiquitous computing web, the selection of on-demand information sources has become a significant requirement. In this thesis, I demonstrate that using our model based on Semantic Technologies and Distributional Semantic techniques we can identify which sensors to use during Task Logging while predicting the task being sensed in real-time. I compare my results with the other state-of-the-art Task Logging techniques showing the improvement. Second, we use our data model to enrich Web content in order to bridge the traditional Web and the Sensor Web. Traditional Web content is long-lived, as it most of the times lacks of real-time information. Sensors deployed pervasively throughout Smart Cities have the potential to be the source of such real-time information. However, querying all the sensor data sources is costly for they are distributed and live streaming high volumes of data. We realised the G-Sensing application that, as a Mozilla Firefox add-on, displays sensors data related to Google search results that represent real places. We demonstrate the feasibility and extensibility of our approach and the advantages it brings to the final user.