Advances in Neural Network assisted Tool Pressure Prediction

In car body tool engineering spotting patterns are used to validate the quality of the tool active surfaces. The objective is to display a homogeneous pressure distribution at defined drawing depths, as achieved by setting the parameters in the simulation accordingly. However, the qualitative evaluation of pressure distribution performed by human visual inspection of spotting patterns is not sufficient for quantitative analysis. It has been demonstrated that convolutional neural networks (CNNs) can predict pressure distributions from spotting patterns. This publication examines the impact of color quantity and contact pairing on the formation of spotting patterns. Likewise, the repeatability of spotting images is evaluated at the macro-, meso- and microscopic levels. A CNN based regression model estimates the absolute pressure distribution based on images of spotting patterns. The integral force, calculated from the estimated pressure distribution, is compared with the measured process force and used either for validation or as part of the CNN output post-processing. In this process, the CNN output is scaled to absolute values using the known total integral force, allowing the post-processing method to extrapolate the predicted pressure distribution effectively, even in untrained regions.