A Study on Evaluation Methods for Manual Welding Qualification Test Specimens Using Deep Learning

This study explores the application of deep learning (DL) to evaluate welder qualification specimens. The aim is to establish an objective assessment method and reduce inspectors’ workload. The study employed a convolutional neural network (CNN) based on the ResNet architecture for image classification, focusing on medium-thickness bend test specimens from the Japan Welding Engineering Society (JWES) examination. Initial binary classification experiments distinguishing between ‘pass’ and ‘fail’ demonstrated that accuracy improved with training sample size, achieving over 90% on average. A confidence score enabled evaluation of classification reliability. Subsequently, multi-class classification was conducted to approximate testing standards. Due to data imbalance, the analysis was conducted on six representative defect categories instead of the original 13. The accuracy exceeded 70% with 40 images per class and further improved to 76.7% after additional training iterations. Furthermore, data augmentation led to an enhancement in performance, culminating in 97.1% accuracy when all available images were utilized. The proposed confidence score demonstrated efficacy in multi-class settings, enhancing decision-making. These findings demonstrate that CNN-based approaches can accurately and objectively evaluate welder qualification test specimens. The proposed methodology could improve inspection efficiency, reduce examiner workload, and contribute to the development of evaluation systems for welding certification.