Deep-Learning Approach to the Self-Piercing Riveting of Various Combinations of Steel and Aluminum Sheets

Abstract

Deep-learning architectures were employed to simulate the self-piercing riveting process of steel and aluminum sheets and predict the cross-sectional joint shape with a zero head height. Four steels (SPRC440, SPFC590DP, GI780DP, SGAFC980Y) and three aluminum alloys (Al5052, Al5754, Al5083) were considered as the materials for the top and bottom sheets, respectively. The key objective was to consider the material properties of these metal sheets (Young’s modulus, Poisson’s ratio, and ultimate tensile strength) in a deep-learning framework. Two deep-learning models were considered: In the first model, the properties of the top and bottom sheets were adopted as the scalar inputs, and in the second model, the three properties were graphically assigned to the three channels of the input image. Both the models generated a segmentation image of the cross-section. To assess the accuracy of the predictions, the generated images were compared with ground truth images, and three key geometrical factors (interlock, bottom thickness, and effective length) were measured. The first and second models achieved prediction accuracies of 91.95% and 92.22%, respectively.

Keywords

• Self-piercing riveting
• cross-sectional shape prediction
• deep learning
• segmentation map prediction
• material properties