Convolutional Neural Network for Quality Prediction in Resistance Spot Welding

This thesis aims to apply Convolutional Neural Networks (CNNs) to data generated by sensors to predict the quality of Resistance Spot Welding (RSW). RSW quality assessment was mainly based on destructive testing and manual inspections. These methods are expensive and not manageable for the real-time analysis required by an Industry 4.0 eco-system. Recent research has shifted toward machine learning and computational methods, which offer non-invasive and real-time evaluations, even though challenges persist in their application due to data limitations. In this study, we investigate the feasibility of using CNNs fed with the matrix generated by multiple sensors during the welding process with a complete black-box approach. The matrix size depends on the number of sensors and the sampling frequency. In RSW, the quality of a weld is crucially determined by its nugget diameter. The nugget directly influences the strength and stability of the weld, making it a critical predictor of structural integrity. The CNN model is developed in a real-case experiment according to the proposed idea and with the aim of predicting in real-time the weld nugget diameter and evaluating the performance metrics such as Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE). The analysis reveals that while CNNs hold potential for enhancing non-invasive quality assessment and for pure black-box application of data fusion and prediction, the limited size of the dataset in this study constrains the model’s ability to generalize and accurately predict weld quality. The results mainly show the importance of a robust dataset for training CNN models in manufacturing applications but at the same time, this research highlights the potential of CNNs to improve weld quality assessment based on more complex and larger databases.