Energy Efficient Sensing for Unsupervised Event Detection in Real-Time

Abstract

General-purpose sensing offers a flexible usage and a wide range of Internet of Things (IoT) applications deployment. In order to achieve a general-purpose sensing system that is suitable for IoT applications, several design aspects such as performance, efficiency and usability, must be taken into consideration. The work of this thesis is focusing on implementing an energy efficient general-purpose sensing system that is based on unsupervised learning techniques for events labeling and classification. The system clusters raw data collected from a variety of events, like microwave, kettle and faucet running, etc., for classification. During the training phase, the system computes sensing polling periods, based on the rate of change in classes, that are then feed into a dynamic scheduler implemented on the sensor board in order to reduce energy consumption. The system is deployed in a one-bedroom apartment for raw data collection, and system evaluation. The results show that the mean accuracy of event classification is 83%, and sensor data polling is reduced in average by 95%, which translates to 90% energy saving, compared to the fixed polling period in the state-of-art approach.