Anomaly and change point detection for time series with concept drift

Abstract Anomaly detection is one of the most important research contents in time series data analysis, which is widely used in many fields. In real world, the environment is usually dynamically changing, and the distribution of data changes over time, namely concept drift. The accuracy of static anomaly detection methods is bound to be reduced by concept drift. In addition, there is a sudden concept drift, which is manifested as a abrupt variation in a data point that changes the statistical properties of data. Such a point is called a change point, and it has very similar behavior to an anomaly. However, the existing methods cannot distinguish between anomaly and change point, so the existence of change point will affect the result of anomaly detection. In this paper, we propose an unsupervised method to simultaneously detect anomaly and change point for time series with concept drift. The method is based on the fluctuation features of data and converts the original data into the rate of change of data. It not only solves the concept drift, but also effectively detects and distinguishes anomalies and change points. Experiments on both public and synthetic datasets show that compared with the state-of-the-art anomaly detection methods, our method is superior to most of the existing works and significantly superior to existing methods for change point detection. It fully demonstrates the superiority of our method in detecting anomalies and change points simultaneously.