Anomalies can be the main carriers of significant and often critical information, and the identification of anomalies is a key task in many fields such as cyber security, intrusion detection, water quality monitoring, system health monitoring, environmental monitoring.
ACEMS researchers, led by Priyanga Dilini Talagala, have proposed a framework that provides early detection of anomalous series within a large collection of streaming time-series data. This includes a novel approach that adapts to non-stationarity in the time series. Using various synthetic and real world datasets from fiber optic cables, we demonstrate that our framework can work well in the presence of nonstationary environments and noisy time series from multi-modal typical classes.
The current framework is developed under the assumption that the measurements produced by sensors are one-dimensional. The rapid advances in hardware technology has made it possible for many sensors to capture multiple measurements simultaneously, leading ultimately to a collection of multidimensional multivariate streaming time-series data. An important open research problem is to extend our framework to handle such data.