WeldVGG: A VGG-Inspired Deep Learning Model for Weld Defect Classification from Radiographic Images with Visual...

Area of Science:

• Materials Science and Engineering
• Computer Science and Artificial Intelligence
• Non-Destructive Testing

Background:

• Manual weld inspection is subjective, inconsistent, and lacks scalability.
• Automated defect detection is crucial for efficient quality control in welded structures.
• Radiographic imagery is a key non-destructive testing method for weld integrity.

Purpose of the Study:

• To develop and evaluate WeldVGG, a deep learning model for automated weld defect classification using radiographic images.
• To enhance the interpretability of deep learning models in weld inspection.
• To benchmark WeldVGG against traditional machine learning and state-of-the-art deep learning models.

Main Methods:

• Training WeldVGG on the RIAWELC dataset, comprising X-ray weld images with four defect classes.
• Utilizing Grad-CAM++ for generating class-discriminative saliency maps to improve model interpretability.
• Performing stratified cross-validation and benchmarking against SVC, Random Forest, and MobileNetV3.
• Validating model generalization with tests on the GDXray dataset and using Wilcoxon signed-rank tests for statistical significance.

Main Results:

• WeldVGG achieved high classification accuracy and interpretability in weld defect detection.
• The model demonstrated strong performance on the RIAWELC dataset, outperforming traditional baselines.
• Grad-CAM++ provided visual validation for model predictions, enhancing trust and explainability.
• Positive generalization results were observed on the GDXray dataset.

Conclusions:

• WeldVGG presents a practical, scalable, and interpretable solution for automated weld defect classification.
• Deep learning models, like WeldVGG, can significantly improve the efficiency and reliability of weld inspection.
• The study highlights the importance of interpretable AI in critical industrial applications like quality control.