Detecting anomalies in data has become a frequent topic in both industry and academia. An anomaly may be defined informally as a "deviation from the rule", i.e., a data point which is substantially different from a collection of similar ones.
In the special case of industrial anomaly detection, this may be applied to images of industry products, signals of manufacturing robots, and anything else, where data is collected.
Kickstarted by the MVTec anomaly detection dataset (in short, MVTec AD; which may be found here), an entire community has formed around anomaly detection, where the Institute has contributed several state-of-the art research papers.
A variety of anomaly detection papers utilizing normalizing flows, a special deep learning technique, have been published. At the time of their release, they all represented a substantial advancement within their specific benchmarks.
Normalizing flows (NF) are able to accurately estimate the data density distribution, which enables the exact inference of a sample's likelihood. For this, they require special invertible neural network architectures.
Our NF-based models have been successfully used in both industry and academia to great success. This applies to both image data, as well as multivariate time series.
Other research looks towards capturing anomalies in more complex 3D objects. With modern techniques, such as 3D Gaussian Splatting, high-fidelty 3D models can be constructed using multi-view images of any object. Including this 3D information into current anomaly detection pipelines is also being explored by our researchers.
The source code is available for several of our methods, to stimulate research and reproducibility:
We have also released a dataset for multivariate time-series anomaly detection together with voraus robotik GmbH.
Have a look at the voraus-AD dataset.
