Detecting Variable Length Anomaly Patterns in Time Series Data

Abstract

The anomaly detection algorithm, developed by Leng et al. (2008), can detect anomaly patterns of variable lengths in time series. This method consists of two stages: the first is segmenting time series; the next is calculating anomaly factor of each pattern and then judging whether a pattern is anomaly or not based on its anomaly factor. Since the lengths of patterns can be different from each other, this algorithm uses Dynamic Time Warping (DTW) as distance measure between the patterns. Due to DTW, the algorithm leads to high computational complexity. In this paper, to improve the above mentioned algorithm, we apply homothetic transformation to convert every pair of patterns of different lengths into the same length so that we can easily calculate Euclidean distance between them. This modification accelerates the anomaly detection algorithm remarkably and makes it workable on large time series.