Logo

 
CSME 2024/10
Volume 45 No.5 : 453-460
 
Deep Learning-Based Welding Defect Identification Device: A Comparative Analysis and Superiority Assessment in Industrial Applications

Chia-Hung Laia
aDepartment of Intelligent Automation Engineering, National Chin-Yi University of Technology, Taiwan, 411030, ROC.


Abstract: Welding has broad applications. Nondestructive testing plays a crucial role in welding inspection. Every weld must be assessed for quality and the absence of defects, but such assessments must be done manually by an experienced operator. Thus, we formulated a deep-learning-based device that automatically conducts such assessments. We trained Welding Defects Net (WDNet) model with a small amount of images and adjusted the depth of convolutional layers and pooling layers during training. We also competitively evaluated several deep learning models for welding defect identification, the first in the literature to do so. In evaluation experiments, our device had accuracy rates as high as 97.8%, outperforming Visual Geometry Group 16 (VGG-16) and Residual Neural Network 50 (ResNet50) demonstrating promise for use in industrial settings.

Keywords:  CNN model, nondestructive testing, welding defects, image recognition.

*Corresponding author; e-mail: 
© 2024  CSME , ISSN 0257-9731 





TOP