Exploration and Explanation of Anomalies in Multivariate Time Series

• Applicant: Professorin Dr. Heike Leitte
• Subject Area: Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing ; Technical Thermodynamics
• Term: since 2022
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 459419731

Anomaly detection (AD) in multivariate time series is a central task for data from chemical engineering (CPE). It supports the work of technical monitoring personnel in large-scale plants and paves the way for the digital transformation of the chemical industry. In this project we will develop methods that make the results of learning-based AD interpretable for humans. Understanding and interpretability form the basis for the integration of learning-based methods in safety-critical applications. We support these two points by developing exploration and explanation techniques that allow an understanding of the data, the anomalies detected in it and the corresponding detection algorithms used. The necessary research focuses on three core topics: 1. We explore analytical methods for monitoring anomalies in multivariate time series of CPE processes with dozen of noisy and heterogeneous sensor channels. 2. We integrate semantic explanation techniques for AD routines based on layer-wise relevance propagation (LRP). 3. We develop and integrate a knowledge database that stores the collected knowledge about the analysis process and makes it available via the monitoring software. The combination of these three components not only helps to identify and explain anomalies, but also provides support in eliminating them. In close cooperation with the A projects, we will expand the limits of anomaly detection in multivariate time series and work in cooperation with the B projection on the ML-based digital transformation of the chemical industry.

• DFG Programme: Research Units
• Subproject of: FOR 5359: KI-FOR: Deep Learning on Sparse Chemical Process Data
• Co-Investigator: Professor Dr. Klaus-Robert Müller