Dissertation DIS-2015-05

Bibliograph.
Daten
Philipp, Damian: Model-driven optimizations for public sensing systems.
Universität Stuttgart, Fakultät Informatik, Elektrotechnik und Informationstechnik, Dissertation (2015).
209 Seiten, englisch.
CR-Klassif.I.2.8 (Problem Solving, Control Methods, and Search)
C.2.4 (Distributed Systems)
KeywordsPublic Sensing; Multivariate Gaussian; Indoor Grammar; Energy Efficiency; Quality
Kurzfassung

The proliferation of modern smartphones such as the Apple iPhone or Google Android Phones has given rise to Public Sensing, a new paradigm for sensor data acquisition using spare resources of commodity smartphones. As smartphones are already deployed wherever there are people present, data collection is enabled at an urban scale. Having access to such a wealth of data facilitates the creation of applications depending on real-world information in a way that may have a lasting impact on our everyday life. However, creating large-scale Public Sensing systems is not without its challenges. On the data requesting side, an interface is required that allows to specify arbitrary sensing tasks independent of the mobility of participants and thus facilitates user acceptance of Public Sensing. On the data gathering side, as many people as possible must participate in the system and thus provide a sufficient amount of data. To this end, the system must conserve the resources shared by participants as much as possible, with the main concern being energy. Participants will withdraw from the system, when participating significantly impacts the battery life of their smartphone. We address the aforementioned issues in the context of two applications: Indoor map generation and large-scale environmental data acquisition. In the area of indoor map generation, we first address the problem of building an indoor map directly from odometry traces. In contrast to existing approaches, our focus is to extract the maximum amount of information from trace data without relying on additional features such as WiFi fingerprints. Furthermore, we present an approach to improve indoor maps derived from traces using a formal grammar encoding structural information about similar indoor areas. Using this grammar allows us to extend an incomplete trace-based map to a plausible layout for the entire floor while simultaneously improving the accuracy of floor plan objects observed by odometry traces. Our evaluations show that the accuracy of grammar-based maps in the worst-case is similar to the accuracy of trace-based maps in the best-case, thus proving the benefit of the grammar-based approach. To improve the energy efficiency of the mapping process, we furthermore present a generic quality model for trace-based indoor maps. This quality metric is used by a scheduling algorithm, instructing a participating device to disable its energyintensive sensors while it travels in an area that has been mapped with high quality already, enabling energy savings of up to 15%. In the area of large-scale environmental data acquisition, we first present the concept of virtual sensors as a mobility-independent abstraction layer. Applications configure virtual sensors to report a set of readings at a given sampling rate at a fixed position. The Public Sensing system then selects smartphones near the position of a virtual sensor to provide the actual data readings. Furthermore, we present several optimization approaches geared towards improving the energy efficiency of Public Sensing. In a local optimization, smartphones near each individual virtual sensor coordinate to determine which device should take a reading and thus avoid oversampling the virtual sensor. The local optimization can achieve a 99% increase in efficiency with the most efficient approaches and exhibits only about 10% decrease in result quality under worst conditions. Furthermore, we present a global optimization, where a data-driven model is used to identify the subset of most interesting virtual sensors. Data is obtained from this subset only, while readings for other virtual sensors are inferred from the model. To this end, we present a set of online learning and control algorithms that can create a model in just hours or even minutes and that continuously validate its accuracy. Evaluations show that the global optimization can save up to 80% of energy while providing inferred temperature readings matching an error-bound of 1°C up to 100% of the time.

Volltext und
andere Links
PDF (4137071 Bytes)
Abteilung(en)Universität Stuttgart, Institut für Parallele und Verteilte Systeme, Verteilte Systeme
BetreuerProf. Dr. rer. nat. Dr. h.c. Kurt Rothermel
Eingabedatum9. Dezember 2015
   Publ. Informatik