Weld Pool Analysis from Images Using Deep Learning Methods

Abstract

This paper presents a method for weld quality assessment based on static images of weld beads produced during cladding process. The dataset consisted of 400 images from real industrial cladding operations, equally divided into acceptable and unacceptable samples. Three convolutional neural network models were evaluated: a custom CNN, ResNet50, and MobileNetV2. MobileNetV2 achieved the best overall performance, with accuracy = 0.7625, precision = 0.8000, recall = 0.7000, F1-score = 0.7467, and AUC = 0.8375, while ResNet50 achieved the highest AUC = 0.8719. Grad-CAM analysis showed that the models focused on relevant weld regions, including fusion lines and structural irregularities. In addition, manually isolated weld pool regions were analyzed using six geometric features: area, ellipticity, symmetry_x, symmetry_y, offset_x, and offset_y. Statistical analysis showed that symmetry_x and offset_x were the most significant features related to weld quality. Among classical machine learning models trained on these features, Random Forest achieved the best results with accuracy = 0.603, F1-score = 0.582, AUC = 0.634. The results confirm that transfer learning is effective under low-data conditions, while weld pool geometry provides useful supplementary information but is not sufficient as a standalone input for reliable weld quality assessment.