Enhancing Weld Inspection Through Comparative Analysis of Traditional Algorithms and Deep Learning Approaches

Abstract

Automated inspection is vital in modern industrial manufacturing, optimizing production processes and ensuring product quality. Welding, a widely used joining technique, is susceptible to defects like porosity and cracks, compromising product reliability. Traditional nondestructive testing (NDT) methods suffer from inefficiency and limited accuracy. Many researchers have tried to apply deep learning for defect detection to address these limitations. This study compares traditional algorithms with deep learning methods, specifically evaluating the SwinUNet model for weld segmentation. The model achieves an impressive F1 score of 96.31, surpassing traditional algorithms. Feature analysis utilizing class activation maps confirms the model's robust recognition and generalization capabilities. Additionally, segmentation results for different welding defects were compared among various models, further substantiating the recognition capabilities of SwinUNet. The findings contribute to the automation of weld identification and segmentation, driving industrial production efficiency and enhancing defect detection.