Track Before Detect: A Novel Approach For Unsupervised Anomaly Detection In Time Series

Abstract

The need for robust unsupervised anomaly detection techniques in streaming data increases rapidly in today’s era of smart devices. Many existing anomaly detection methods have difficulties to detect anomalies in streaming data since most of them are designed to use all features of the data which are not applicable in a streaming context such as IoT. To address this problem, we present a novel unsupervised anomaly detection approach (Track Before Detect) for time series data. Track Before Detect (TBD) is capable of detecting a wide range of anomalies such as point anomalies and collective anomalies. In addition, it can differentiate between anomalous behavior and environmental changes in time series data in an unsupervised setting and without affecting the running system. Experiments based on real world data sets demonstrate that TBD succeeded in detecting anomalies in time series data and outperformed existing methods.