Development of a Surface Defect Severeness Classification on Deep Drawing Simulation Results

This paper presents a data generation and neural network approach for video classification of cosmetic defects on simulated automobile outer surface parts. Cosmetic defects occur in the deep drawing process, where high forces are applied to a sheet metal. The deformation process at the deep drawing can be simulated and the results can be used for further investigation. This work presents a specific mesh discretization and numerical settings for a high part surface quality data generation within a rendered virtual environment. To comple-ment the introduced method, virtual environment settings are investigated. The generated input data is further used for training of a Convolutional Recurrent Neural Network (CRNN) to predict a severeness classification on part quality imperfections. This CRNN shows promising results with a prediction accuracy of 86% for a conservative classification. The term conservative classification means the detection of the correct severeness class or one severeness class higher.