Deep-Learning-Based Predictive Architectures for Self-Piercing Riveting Process

Abstract

Deep-learning architectures were developed for the self-piercing riveting (SPR) process to predict the cross-sectional shape from the scalar input of the punch force. Traditionally, the SPR process is studied using a physic-based approach, including finite element modeling, but in this study, a data-driven approach consisting of two supervised deep-learning models was proposed. The first model was used for data transformation from an optical microscopic image to a material segmentation map, which characterizes the shape and location of the two sheets and the rivet by applying a convolutional neural network (CNN)based deep-learning structure. To validate the developed models, two types of sheet combinations were tested, namely, carbon-fiber-reinforced plastic (CFRP) and galvanized dual-phase steel (GA590DP) sheets, and steel alloy (SPFC590DP) and aluminum alloy (Al5052-H32) sheets. The transformation was performed with a mean intersection-over-union of 98.50% and a mean pixel accuracy of 99.78%. The next model, which was a novel generative model based on a CNN and conditional generative adversarial network with residual blocks, was then trained to predict the cross-sectional shape from the input punch force. The predicted cross-sectional shapes were compared with the experimental results of SPR. The overall accuracy was 94.20% for CFRP-GA590DP and 96.31% for SPFC590DP-Al5052, with respect to three key geometrical indexes, namely, rivet head height, interlock length, and bottom thickness.

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

“…Figure 4 shows the structure of Model 1 to predict the crosssectional shape of the specimen after the SPR process. A previously proposed generative model (scalar-to-seg generator) [13] was adopted, which was based on the CNN and conditional GAN (cGAN) [18] architecture with residual blocks. As shown in Fig.…”

“…4, to enable effective learning with deeper layers. More details regarding the adopted model can be found in [13].…”

“…Notably, unlike the previous model [13], a condition vector composed of the material properties of the top and bottom sheet materials, was used as the input of the generator instead of the scalar punch forces. Because the punch force was not an input parameter, it was set such that the head height was zero.…”

“…Recently, Oh et al [13] proposed a deep-learning model to predict the cross-sectional shape from the input punch force of the SPR process, and the three key geometric factors in the cross-section were measured to assess the prediction quality. Deep-learning models offer several advantages over other simulation methods such as the FEM; specifically, deeplearning models do not require the implementation of any assumptions, and because these approaches are data-driven, all complex problems can be solved with sufficient data.…”

“…Considering these aspects, in this study, based on the authors' previous work [13], two deep-learning models were developed, which could consider the material properties of the top and bottom sheet materials (steel and aluminum alloy, respectively) in a deep-learning framework and predict the cross-sectional joint shape for various sheet combinations. As the first model, the previously developed generative model (known as the scalar-to-seg model [13]), which was based on a convolution neural network (CNN) [14] and generative adversarial network (GAN) [15] with residual blocks [16] was adopted. However, in contrast to the previous model, which adopted the punch forces as the inputs, the material properties of the top and bottom sheets were considered as the inputs.…”