Learning hybrid model chains in milling processes

• Applicant: Professor Dr.-Ing. Christian Brecher
• Subject Area: Production Automation and Assembly Technology
• Term: since 2023
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 524834456

Due to a dynamic market and increasing individual customer needs, the digital shadow (DS) is becoming more and more important in regard to optimization and transparency of the current production. Previous work at the WZL on the development of this DS has shown that, with the help of a process-parallel material removal simulation based on machine data, the component quality can already be predicted during production. The time between production and the detection of tolerance violations can thus be considerably shortened. Furthermore, the DS enables a systematic analysis of the error causes with regard to individual influences, e.g. machine, tool and workpiece, as well as an increase in process transparency, which results in a high potential for a effective reduction of the occurring total error. The essential basis for this is a model chain that calculates displacements, forces and other intermediate variables based on machine-internal data. However, real industrial production processes are partly subject to stochastic process variables with highly complex interdependencies. To reduce the uncertainties of the DS, it is necessary to continuously adjust the model parameters along the model chain. Data-driven approaches such as machine learning (ML), in combination with knowledge-based approaches, offer potential for determining the residual uncertainties and continuously adjusting model parameters. The aim of this research project is to determine the residual uncertainties of the individual knowledge-based models with the help of a systematic data-driven evaluation along the root-cause chain in order to subsequently reduce them by continuously adjusting the corresponding parameters. The uncertainties of the submodels are identified and evaluated via a continuous quality comparison between the real and virtual workpiece by means of a feature-based observation and a partial factorial experiment plan. Machine learning algorithms are used here, which determine the residual uncertainties based on internal machine data and are fitted via quality data. The generated black-box models will be converted into grey-box models and explained to enable a better root cause analysis.

• DFG Programme: Research Grants