Online anomaly search in time series: significant online discords

The aim of this work is to obtain a useful anomaly definition for online analysis of time series. The idea is to develop an anomaly concept which is sustainable for long-lived and frequent streamings. As a solution, we provide an adaptation of the discord concept, which has been successfully used for anomaly detection on time series. An online approach implies the frequent processing of a data streaming for timely providing anomaly alerts. This requires a modification since discord search is not exactly decomposable in its original definition. With a statistical approach, allowing to rate the significance of the discords of each analysis, it has been possible to obtain a solution where the number of false positives is minimized. The new online anomalies are called significant online discords ( sod s). As a novel feature, sod search determines the quantity of anomalies in the time series under investigation. The search for sod s has been implemented and its properties validated with synthetic and real data. As a result, we found that sod s can be considered as a useful new tool for anomaly detection in fast streaming time series or Big Data contexts.