A wavelet-based system for event detection in online real-time sensor data

Abstract

Sensors are increasingly being used for continuous monitoring purposes, the process of which generates huge volumes of data that need to be mined for interesting events in real-time. The purpose of this research is to develop a method to identify these events, and to provide users with an architecture that will allow them to analyze events online and in real-time, to act upon them, and to archive them for future offline analysis. This thesis is divided into two major portions. The first discusses a general software architecture that performs the functions defined above. The architecture proposed assumes no prior knowledge of the data, and is capable of dealing with multi-source data feed from any type of sensor(s) on one end, and can handle multiple clients on the other. The second part of the thesis discusses a wavelet-based algorithm for detecting certain types of events in real-time in one-dimensional numeric time-series data. Wavelets were judged to be the most appropriate technique for analyzing random sensor signals for which no prior information is available. The wavelet-based method in addition allows users to delve into different levels of abstraction (based on varying time periods) while looking at the data, which cannot be done by any previous method for real-time event detection. This thesis also touches on the fundamental question of how one defines an event, which is more easily possible in a particular domain, for a specific purpose, but is much harder to do in a generic, domain-independent level.