Nowadays, sensors are cheap, easy to deploy and immediate to integrate into applications. Having a set of sensors emitting data streams, our concern in this PhD thesis is to build a sensor database for applications that require long-term storage.
Since huge amount of sensor data can be generated, selecting only relevant data to be saved for further usage, e.g. long-term query facilities, is still an issue. For instance, a query asking for the temperature of a given house per day along a year may have to consider every temperature value sent from different sensors in every room every minute. In such cases, approximating data in order to reduce the considered values (i.e. relevant values with respect to the application requirements) enhances such query efficiency.
Given sensor data streams, the key idea is to consider both spatio-temporal hierarchies and Spatio-Temporal Functional Dependencies as first class-citizens for designing sensor databases on top of any relational database management system.
We propose an axiomatisation of these dependencies and the associated attribute closure algorithm, leading to a new normalization algorithm.
Then, we propose an annotation of attributes with aggregate functions allowing to specify which summarized values are of interest in the data. We can automatically both produce a concrete SQL DB and configure the data loading process.
We have implemented a prototype to deal with both database design and data loading. This allowed us to conduct experiments with, synthetic and real-life, data streams from intelligent buildings.