Ricarda-Samantha Götte successfully completed her doctorate on the topic of “Online estimation of model inaccuracies for automatic model adaptation while maintaining a physical-technical interpretability”. The first supervisor and DART follow-up group leader was Dr Julia Timmermann, the chair of the doctoral committee was Prof. Dr Ansgar Trächtler.
Abstract:
The aim of this thesis is the development of a model-based observer for input-affine, nonlinear systems that achieves a high estimation quality despite model inaccuracies. By additionally providing a parametric, physically interpretable representation of the model inaccuracies, an automated improvement of the model should be enabled. This thesis analyzes techniques of hybrid system identification such as physics-guided neural networks, as well as methods for compensating model inaccuracies within the observer design. Based on this analysis, a novel model-based observer is designed, which estimates states and model inaccuracies jointly and, in particular, obtains a parametric, physically interpretable representation of the inaccuracies. This consists of a linear combination of physically interpretable functions, whose associated parameters are modeled sparse and estimated in parallel to the system's states using an augmented state. The novelty of this thesis is thus the real-time capability to jointly estimate states and model inaccuracies in a physical-technical manner, on the basis of which an automated model adaption can be carried out. The application of the new methodology is particularly advantageous in the situation of occurring system changes since these can be estimated and identified at run time by the augmented observer.
