Anomaly detection in hot forming processes using hybrid modeling

Hot forming is a widely used manufacturing process of crash-relevant structural components with complex geometries. In this paper a method is presented to detect anomalies during the hot forming process giving indications of possible quality defects. The method is based on a physical model of the thermal energy transfer between hot blank, pressing tool and cooling water. The model is built using lumped heat capacities and virtual heat resistances between the components. Within this model the temperature profiles of these components are simulated using input data from real process sensors. The physical model is then combined with a neural network to form a hybrid model. The neural network is trained using the input data and a parameter optimization algorithm and determines the currently optimal parameters of the physical model. During the production, a comparison of the simulated and the real temperature profile reveals anomalies in the hardening process. This way indications for potential quality defects are gained.