Ensemble-based automatic generation of calibration models for complex sensor systems on example of non-destructive micro-magnetic materials characterization techniques

• Applicant: Professor Dr.-Ing. Andreas Kroll
• Subject Area: Automation, Mechatronics, Control Systems, Intelligent Technical Systems, Robotics
• Term: since 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 532921704

The surface condition of parts is effected by all current manufacturing processes. At the same time, the surface is the region of a component subjected to the highest stress. Therefore, the surface conditioning resulting from a manufacturing process is essential for the life expectancy and reliability of parts. Common materials characterization techniques for the near-surface condition require costly post-process laboratory measurements. In contrast, micro-magnetic materials characterization permits a non-destructive in-process measurement, e.g. of residual stress and hardness, and enables process-integrated automatic control of the part properties during manufacturing. The state of the art of calibrating micro-magnetic materials characterization systems are linear methods and an application-dependent, expert-based feature extraction and selection. However, for micro-magnetic materials characterization complex relationships between measured time-series (current/voltage) and target quantities (surface layer condition) have to be modelled. These dependencies change in general, in case the material or the target quantity is changed, such that each time a new costly calibration is required. Therefore, in this project algorithms for an automated feature extraction and selection as well as model generation are to be investigated, in order to obtain better calibration models with less efforts. For this purpose, a data-driven, ensemble-based technique is to be designed, which achieves high prediction accuracy by using nonlinear models. The necessity of time and cost intensive reference laboratory measurements will be reduced by combining data efficient model approaches and unsupervised and supervised learning methods. The ensemble approach permits achieving high prediction accuracy and eases the transfer to other applications. In addition, the prediction uncertainty will be quantified. The calibration technique will be designed such that other applications (manufacturing processes, materials) can easily be covered and it will be made publically available as a calibration toolbox.

• DFG Programme: Research Grants