10/2024 - Project explanation and research updates of the CausalAnomalies project
Anomaly detection is essential for several areas of application at the German Aerospace Center (Deutsches Zentrum für Luft- und Raumfahrt; DLR). In the event of deviations from expected patterns of a dynamic system, users can intervene to stop or adjust the procedures. When an anomaly is detected, it can indicate that the monitored system has entered a certain state, such as in the event of damage, which can be recognised and remedied through such intervention. However, understanding the causes of such anomalies is particularly important to model and predict the necessary procedures before the state of the system changes.
The CausalAnomalies project aims to develop new, causally explicable algorithms that detect anomalies. Finding abnormal states in time series data is a goal whenever sequential data is recorded and analysed. These data are often high-dimensional, error-prone, incomplete or not labelled. Furthermore, it is frequently unclear which abnormal behaviour within the data is relevant. Therefore, the first project goal is to explore different AI methods for anomaly detection. The next goal is to make these methods explicable in order to overcome the 'black box' nature of many AI methods. For this purpose, causal methods will be used and adapted. The developed methods will then be applied to various use cases at DLR and will be tested and optimised.
We, as the @DLR Institute of Data Science, are responsible for the development of the numerical methods for anomaly detection and special causal questions in addition to the project management. We work closely with the domain experts from the following five other DLR institutes:
The current status of our work at CausalAnomalies is:
- Development and implementation of a deep learning based anomaly detection algorithm for sequential data
- Development and implementation of conditional independence tests for mixed continuous and categorical variables for causal inference
- Completion of the use case-specific sprints, in which specific questions on anomaly detection and causal methods were developed. In addition, the numerical methods for anomaly detection and causal inference were tested in experiments to answer the respective questions.
Further information about the CausalAnomalies project can be found here.