Time Series Discords

Abstract

Time series discords, as introduced in by Keogh et al. [5] is described as the subsequence in the time series which is maximally different from the rest of the subsequences. Discovery of time series discords has been applied to several diverse domains including space shuttle telemetry, industry, and medicine [5] to detect anomalies in the data which can identify equipment failure, unusual patterns of activity and health problems. In this thesis we will examine the problem of finding time series discords, with detailed analysis of the problem and analysis of the effectiveness of prior work. Three different areas of discord discovery will be examined: Top Discord, Variable Length Discords, and Top-K Discords. In each of these areas, we strive to reduce the number or ease the selection of input parameters required by the end user. Emphasis is also placed on improved runtime and scalability of discord discovery methods.