Forming quality prediction of aluminum alloy self-piercing riveted joints based on deep learning

To address complex simulation modeling, high experimental costs, and low efficiency, a data-driven prediction method for the forming quality of self-piercing riveted (SPR) joints was proposed. Firstly, image segmentation technology was employed to preprocess the cross-sectional images of SPR joints obtained from experiments, generating categorized cross-sectional images for model training. These images meticulously labelled different parts of the rivet, upper plate, lower plate, and background, providing the model with abundant visual information. A dataset based on experimental images was established to build a deep learning model with a conditional generative adversarial network architecture based on convolutional neural networks. The thicknesses of the upper and lower plates and the length of the rivet were integrated into a vector containing three scalar values as the input for this network model. During the model training phase, 15 different combinations of riveting process parameters were designed, and the model was trained and predicted in two stages. Through extensive training with experimental data and parameter adjustments, the learning effect of the model was continuously optimized. In the prediction phase, the model's output was compared and analyzed with the experimentally obtained cross-sectional images of joints, with a focus on two key geometric parameters: rivet spread and residual bottom thickness. The results indicate that the trained deep learning model can accurately predict the cross-sectional shape of joints for various combinations of base plate thickness and rivet length, with average prediction accuracies of 92.03% and 92.48% for rivet spread and residual thickness, respectively. The developed model for predicting the forming quality of SPR joints demonstrates high accuracy, thereby enhancing the efficiency of the joining process design.