Anomaly Detection over Pseudo Period Data Streams Based on DTW Distance

Pseudo period data streams appear in a lot of applications, especially in monitoring domains. The anomalies detected over pseudo period data streams may possess significant domain knowledge which is worth to do further analysis. When Euclidean distance between time series changes greatly with the compared time series moving slightly along the time-axis, DTW (dynamic time warping) distance is suggested as a more robust distance than Euclidean distance. In this paper DTW distance is adopted as similarity measure of different wave sections in pseudo period data streams, and then the anomaly wave sections are defined, which have few historical similar counterparts based on that similarity measure. A nave algorithm is given to detect the anomaly wave sections by directly computing the DTW distance between the current wave section and all other wave sections in the historical dataset. However, the efficiency of the nave algorithm is very poor which limits its application. So a fast approximate algorithm based on the cluster index is proposed to speedup the nave method. Compared with the nave algorithm, this new method is much faster in speed and no big degrades in accuracy. Extensive experiments on the real dataset demonstrate the effectiveness of the proposed methods.