Abstract
Laser welding of dissimilar materials, such as steel and copper, is highly susceptible to crack formation, which compromises joint integrity and service life. Traditional inspection techniques are slow, labor-intensive, and error-prone, underscoring the need for intelligent defect-prediction systems. This study utilizes a publicly available dataset of 360 weld cross-sections, generated using a definitive screening design (DSD) that encompasses six key process parameters: laser power, welding speed, angular orientation, focal position, gas flow rate, and sheet thickness. Multiple machine learning classifiers, including Decision Trees, Random Forests, Gradient Boosting, Support Vector Machines, and Neural Networks, were systematically evaluated using Orange data mining software with imbalance handling strategies. The novelty of this study lies in the application of the Orange data mining tool to address data imbalance in welding defect prediction and its optimization through a neural network framework, thereby enhancing both model reliability and predictive performance. Among them, a Multilayer Perceptron (MLP) neural network achieved the best performance, attaining 94.9% accuracy, 86.3% sensitivity, and 96.8% specificity, with an AUC of 0.961. The results establish neural networks as a robust and scalable tool for defect classification in steel-copper welding, offering a practical pathway for intelligent process monitoring and predictive quality assurance in Industry 4.0 manufacturing.