Ensemble-based automated calibration model creation for complex sensor systems using the example of non-destructive micromagnetic material characterization methods
Brief description
Brief description
The near-surface condition of a component is strongly influenced by all common manufacturing processes. At the same time, the surface is usually the most stressed area of a component. Therefore, surface conditioning by manufacturing processes is of crucial importance for the service life and reliability of components. Conventional material characterization methods for recording the near-surface condition require complex post-process laboratory measurements. In contrast, micromagnetic material characterization offers the possibility of non-destructive in-process measurement, e.g. of residual stress and hardness, and is therefore very well suited for use in a process-integrated control or regulation concept for component properties. Linear methods and the application-dependent, expert-based extraction and selection of characteristics form the state of the art for the calibration of micromagnetic measuring systems. In micromagnetic material characterization, complex relationships between the measured time series (current/voltage curves) and the target variables of the measurement (boundary zone properties) must be modelled. These correlations typically change when the material or the target variable is changed, so that a complex new calibration is necessary each time. This project therefore aims to algorithmically integrate automated feature extraction, selection and model creation in order to obtain better calibration models with little effort. For this purpose, a data-driven, ensemble-based method is to be designed in the project that achieves improved model quality through non-linear model approaches. The combination of data-efficient model approaches and supervised and unsupervised learning methods is intended to minimize the need for time-consuming and cost-intensive reference measurements for calibration. The ensemble approach should ensure a high prediction quality and transferability to other applications. In addition, the method is to be designed in such a way that it enables quantification of the prediction uncertainty. The method is designed in such a way that it can be used across applications (manufacturing processes, materials).
Publications
Felix Wittich, & Andreas Kroll (2025). Kalibriermodellerstellung und Merkmalsselektionfür die mikromagnetische Materialcharakterisierung mittels maschineller Lernverfahren. at – Automatisierungstechnik, 73(10), 778–790.
Felix Wittich, Farzad Rezazadeh, & Andreas Kroll (2025). Four Benchmark Datasets for Nonlinear Regression in Engineering Sciences. In Proceedings 35. Workshop Computational Intelligence: Berlin, 20.-21. November 2025.
Felix Wittich, & Andreas Kroll (2024). Datengetriebene Modellierung und Merkmalsselektion für die automatische Kalibrierung eines Sensorsystems für die mikromagnetische Materialcharakterisierung. In Proceedings 34. Workshop Computational Intelligence: Berlin, 21.-22. November 2024 (pp. 21–26).
Felix Wittich, Thomas Wegener, Alexander Liehr, Wolfgang Zinn, Artjom Bolender, Sebastian Degener, Thomas Niendorf, & Andreas Kroll (2025). Micromagnetic measurement of the surface layer state in hard turned 51CrV4 steel [Dataset]. Universität Kassel. doi.org/10.48662/DAKS-477
Project Researcher
Felix Wittich, MSc.
Period
as of 07/2024
Funding
German Research Foundation (DFG), project number 532921704