Adjacent Mean Difference (AMD) method for dynamic segmentation in time series anomaly detection

Abstract

Heuristic Ordered Time series Symbolic Aggregate approXimation (HOT SAX) is a well-known symbolic representation approach used to detect the abnormalities in the time series. Since HOT SAX allows dimensionality reduction and searches abnormalities with a heuristic algorithm, HOT SAX prevail in time series data analysis. However, because HOT SAX detects abnormalities through sliding a window of equal length, the search results of abnormalities would change when setting a different length of window and the optimal length is hard to define. Therefore, in this research, Adjacent Mean Difference (AMD) segmentation method was proposed to segment the data dynamically without setting any parameter. Essentially, AMD partitions data into multiple segments of different lengths based on the transitions between data points. After data segmentation, FastDTW was used to compare the distances between segments of different lengths. The experiments demonstrated that AMD is an easy and efficient method to segment data dynamically. And the comparison with HOT SAX shows that AMD can be used to detect abnormalities with better computational efficiency.