Bridging the Domain Gap - Visual Identification of Domain-Invariant Features in Time Series
Building a robust and reliable classifier to verify the functionality of electric motors is challenging. The classifier would need to be trained on huge amounts of real-world data from defective motors but this data does not exist. Instead, sufficiently enough data can be collected from the simulation of defective motors. However, using this simulated data to train the classifier and applying it to real-world measured data would not perform well. The gap between the simulated and the measured data negatively influences the classification. To overcome this gap, domain-invariant features can be used to train the classifier and improve its robustness and reliability by making it generalize better on the classification of the measured data. The main idea of this thesis was to build a visual-interactive approach to identify domaininvariant features. The measured and simulated data can be explored by domain experts, utilizing their knowledge. Although the main task of identifying the domain-invariant features was difficult to complete for domain experts, the implemented functionalities were considered to be essential with good usability. The visual-interactive tool enables the visual exploration of the data and a qualitative evaluation gives more insight into the thought process of the domain experts and their suggestions on refinements and extensions for the tool. The result of the evaluation is discussed and gives direction for further research and studies.
Darmstadt, TU, Master Thesis, 2022