Context-aware dynamic designs for energy efficient mobile sensing -

Abstract

Technological advances in the past decade have driven a significant evolution of various technologies towards Internet of Things (IoT) in domains such as sensing, communications, and computing. In particular, today's smart mobile devices have become equipped with various specialized sensors and can be augmented with external wearable sensors to collect vital data. As a result, new context recognition applications have been developed to understand and analyze user's context such as activity, location or health conditions. Often times, multiple context applications are running simultaneously on smart devices placing strenuous demands on their battery-limited resources. As a result, to support the growing requirement and proliferation of such applications in IoT, there is a need to optimize the usage of the limited computing and sensing resources. To this end, this dissertation proposes a novel context-aware dynamic sensing framework that enhances the trade-off between energy consumption and accuracy in context detection and recognition. The goal is to have decisions customized for each context, and thus develop optimized context-aware designs for three aspects of sensor usage: how many samples to collect every time a sensor is triggered, when to schedule sensors' data collection, and which sensors to use. For the choice of sensors' samples, we propose two sampling mechanisms, where one is based on information theory, and the second is based on recent advances in deep learning. For sensor scheduling, we design an optimized real-time strategy based on the Viterbi algorithm with customized rewards to decide dynamically on when to trigger the sensors for data collection. Finally, for sensor selection, we develop a new mechanism to alleviate the energy limitations by considering synergy in sensor usage in multi-context setting. A new context ontology is built to support the framework selection strategy and the extraction of commonalities and differences between multiple context recognition models along with descriptive