Related Experiment Videos
TIG welding defect detection using ResNet and Random Forest
Vu Quang Huy1, Vu Minh Thuan2, Hoang Van Huong2
1Faculty of Advanced Education, Ho Chi Minh City University of Technology and Engineering, HCM City, Vietnam.
Plos One
|July 30, 2026
Summary
This study introduces an automated method using ResNet50 and Random Forest for identifying orbital TIG welding defects. The approach achieves 98% accuracy, significantly improving defect classification in precision manufacturing.
Area of Science:
- Materials Science
- Mechanical Engineering
- Computer Science
Background:
- Orbital TIG welding is crucial for structural integrity in precision manufacturing.
- Manual inspection of weld defects is challenging due to variations in defect characteristics, leading to time-consuming evaluations.
- Automated defect detection is needed to enhance efficiency and reliability.
Purpose of the Study:
- To develop and evaluate a novel method for accurate classification of orbital TIG welding defects.
- To improve upon existing methods for weld defect detection using deep learning and machine learning techniques.
Main Methods:
- A hybrid approach combining ResNet50 for deep feature extraction and Random Forest for classification was proposed.
- A dataset of 3,219 orbital TIG weld images was utilized, categorized into normal, lack of fusion, overheating, and uneven weld classes.
- Five-fold cross-validation was employed to assess the model's performance.
Main Results:
- The proposed ResNet50-Random Forest method achieved an overall accuracy of 98% in classifying weld defects.
- This hybrid model outperformed various Convolutional Neural Network (CNN) architectures (VGG16, VGG19, ResNet18, ResNet50) and traditional machine learning models (SVM, Random Forest).
- ResNet50 effectively learned hierarchical features, while Random Forest provided robust classification.
Conclusions:
- The developed method offers a highly accurate and efficient solution for automated orbital TIG welding defect detection.
- This approach has the potential to significantly enhance quality control in precision manufacturing by reducing manual inspection burdens.
- The combination of deep feature extraction and robust classification provides a powerful tool for analyzing complex weld images.