A Study on the Development of a Welding Quality Evaluation System Based on Data Augmentation Techniques

Abstract

This study reports on the development of an artificial intelligence-based vision inspection system designed for the real-time detection of humping defects in welding processes, where data collection is often limited. While higher welding speeds induce flow instability in the weld pool-leading to humping defects and compromising the mechanical reliability of joints-traditional non-destructive testing (NDT) methods, such as radiographic testing (RT), face limitations in real-time, in-line monitoring. To address this, this research introduces a subfolder-level batch random crop algorithm that extracts random positions while preserving the aspect ratio, thereby augmenting a small number of original images into a large-scale dataset. Transfer learning was then applied to the augmented data using ResNet-50, a pre-trained model optimized for extracting complex visual features, to maximize defect classification performance. When the developed system was evaluated in an actual welding environment, it accurately identified normal Bead on Plate (BOP) beads from real-time video with an accuracy of 97.2%. Furthermore, it successfully detected narrow-gap humping defects-characterized by complex shapes formed within confined spaces-with a high reliability of 90.5%, while suppressing the probability of misclassification into other classes to below 5%.