Application of Convolutional Neural Networks to Control Quality of Resistance Spot Welding of Galvanized Steel Sheet

This study used convolutional neural networks (CNN) to manage the quality of resistance spot welding by categorizing photos of welds on galvanized steel sheets. The welding parameters included 19 cycles of welding time, 8.5 kA welding current, and 0.20 MPa electrode force. Endurance testing procedures were used to generate a dataset for the CNN model. Following that, weld surface photos were collected, nugget sizes were determined, shear strength was tested, the influence of zinc coating on the workpiece was investigated with a scanning electron microscope, and data was analyzed to classify the quality of the weld surface using K-fold cross-validation. The model was created with the pre-trained ResNet50 architecture and fine-tuning procedures. According to the research findings, the CNN model achieved the greatest accuracy of 93.93%, with precision, recall, and F1-Score values of 0.996, 0.998, and 0.997, respectively. The effect of the zinc coating was detected during the 270th welding cycle, revealing deformation of the electrode contact surface and melting of the zinc coating, which, when paired with the copper electrode, resulted in the creation of brass deposits on the electrode contact surface. This impact caused the nugget size to fall outside of permitted limits, reducing shear strength.