About the project
The central goal of online anomaly detection is to flag anomalous segments in a multivariate time series where new observations arrive continuously. This stands in contrast to offline anomaly detection where the entire time series is available beforehand for analysis. This setting is less challenging and much more well-studied than the online setting.
Online anomaly detection puts strict demands on the computational complexity and adaptability of the algorithms. Furthermore, classic parametric algorithms rely on specifying the type of anomaly that is observed (e.g. change in mean and variance). If this is misspecified, it can lead to a significant loss in detection power.
The goal of this project will therefore be to develop new deep learning based algorithms that learn to detect deviations from the baseline time series. A central part will be to investigate how to train these models in a truly online fashion to adapt to an evolving baseline (possibly integrating domain knowledge). If possible, rigorous theoretical guarantees will be established.
Project team
- PhD Fellow Sebastian Grau Nielsen, University of Oslo
- Professor Ingrid K. Glad, University of Oslo
- Professor Michael Kampffmeyer, UiT The Arctic University of Norway
- Senior Research Scientist Martin Tveten, Norwegian Computing Centre
- Professor Idris Eckley, Lancaster University
- Postdoctoral Fellow Per August Jarval Moen, University of Oslo