Online anomaly detection in high-dimensional data streams

In a wide range of applications, monitoring data streams for faults or changes in behavior (called anomalies) is of great importance.

A common scenario is that an object, say an industrial system, is measured regularly over time, for example by a large network of sensors, giving rise to a stream of high-dimensional data from which to infer the presence, timing, and location of anomalies. A complication is often that “normal behavior” depends on the context in which it is observed, which can for instance be the operational mode of the system, or external environmental conditions affecting it. The aim of this project is to explore and develop new machine learning methodology and theoretical foundations for reliable, online detection of contextual anomalies in streaming multivariate time-series data.