Predictive Analysis from numerical and experimental data in press hardening

Abstract

Machine learning, big data and deep learning are today’s catchphrases for how to improve reliability and productivity of your manufacturing equipment. Production companies implement a large number of sensors to record every activity within their production lines and learn as much as possible about their running processes in order to predict shifting product properties and to prevent stoppage due to failure. The successful application of machine learning algorithms to predict machine and process behavior depends on a reliable and balanced database. Since the foremost goal of every manufacturing business is to make sound parts and to avoid defects, there is a large amount of data available for smoothly running processes but only very little for failure production. One approach to solve this imbalance would be to link the production line data with simulation data. Simulation models allow for computing failure parts with no additional costs and therefore enable the exploration of the entire parameter space. We conducted press-hardening experiments with a variation of process parameters for a structural car body part on the press hardening line at Fraunhofer IWU. As an evaluation criterion, we measured the hardness of the final part at critical spots. In order to expand the experimental data, we applied FE simulations to the entire press hardening process chain. The paper explains limitations of the model and elaborates on its parameterization. As a final task, we applied a basic machine-learning algorithm to both experimental and numerical data as well as to their combination in order to evaluate the data space expansion through simulations. The results obtained through machine learning indicate significant differences in the prediction of part quality for solely experimental data and its combination with simulation data. This is especially true for press hardening because of non-linear system behavior and a large amount of uncertain and hard-to-identify parameters. We found that the most challenging parts of uniting measured and simulated data is not only to create simulations with appropriate accuracy, which allow for a meaningful extrapolation of the parameter space, but also to compare simulation and production data based on the same criterion and to have stable simulation models for the entire parameter range.