Fault Detection and Identification
Concerning “Fault Detection and Identification”, we are conducting research together with our scientific partners on the following topics, among others:
Fast (Online) Algorithms that Learn Partial System Models
Algorithms from robust statistics show good results for estimating parameters in the presence of outliers (faults), in particular, algorithms with a high breakdown point. Furthermore, algorithms with a high breakdown point for machine learning problems are designed, e.g. learning robust decision trees, neural networks, and support vector machines.
Fault Detection and Prediction
Robust and fast algorithms (see above) are the basis for outlier detection and fault detection by residual analysis. Residual analysis can be seen as an analysis of confidence regions. Other methods to estimate and analyze confidence regions are density level estimation and statistical depth estimations and implicitly outlier detection. The purpose of confidence regions is to distinguish between normal and abnormal system states. This requires stochastic methods since system observations are random. To predict the time of fault occurrences, we employ event time analysis.
Our focus is on applying methods like robust stochastic optimization and automated planning and to look for alternatives that work best for our use cases. The quality of predictions will be most critical for achieving good optimization results in practice. Robust optimization algorithms are especially designed to work with uncertain predictions.
Selected publications are summarized in the strategic project dasRES.