A Novel Method for Time Series Anomaly Detection based on Segmentation and Clustering

There have been several algorithms for anomaly detection in time series data. However, most of them suffer from high computational cost and hence can not suit real world applications well. In this paper, we propose a novel method for time series anomaly detection. In this method, first, subsequence candidates are extracted from time series using a segmentation method. These candidates are then transformed into other subsequences with the same length and input for an incremental clustering algorithm. Finally, we identify anomalous patterns by using an anomaly score. The experimental results show that our approach is much more efficient than the HOT SAX algorithm while the anomalous patterns discovered by the proposed method match those by the Brute-Force one.