Time series representation for anomaly detection

Anomaly detection in time series has attracted a lot of attention in the last decade, and is still a hot topic in time series mining. However, time series are high dimensional and feature correlational, directly detecting anomaly patterns in its raw format is very expensive, in addition, different time series may have different lengths of anomaly patterns, and usually, the lengths of anomaly patterns is unknown. This paper presents a new conception key point and an algorithm of seeking key points, the algorithm uses key points to re-represent time series and still preserves its fundamental characteristics. Variable length method was used to segment re-represented time series into patterns and calculate anomaly scores of patterns. Anomaly patterns are identified by their anomaly scores automatically. The effectiveness of representational algorithm and anomaly detecting algorithm are demonstrated with both synthetic and standard datasets, and the experimental results confirm that our methods can identify anomaly patterns with different lengths and improve the speed of detecting algorithm greatly.