A misclassification-aware explainable hybrid CNN-vision transformer framework for radiographic weld inspection

Area of Science:

• Materials Science and Engineering
• Artificial Intelligence
• Manufacturing Technology

Background:

• Accurate weld defect identification is crucial for structural integrity in safety-critical manufacturing.
• Existing deep learning methods struggle to differentiate visually similar weld defects, leading to costly misclassifications.
• There is a need for more reliable and interpretable weld inspection systems, particularly for Industry 5.0.

Purpose of the Study:

• To introduce a novel hybrid Convolutional Neural Network-Vision Transformer (CNN-ViT) architecture for intelligent weld inspection.
• To enhance the reliability and interpretability of automated weld defect detection systems.
• To improve the discrimination of visually similar defects and reduce misclassification rates.

Main Methods:

• A hybrid CNN-ViT architecture was developed and trained on the RIAWELC radiographic weld dataset.
• The proposed model was compared against a lightweight CNN baseline under identical training conditions.
• Explainability methods, including Grad-CAM and transformer self-attention maps, were used for misclassification analysis.

Main Results:

• The hybrid CNN-ViT achieved a superior accuracy of 98.56% compared to the CNN baseline (97.90%).
• A 31% relative reduction in the misclassification rate was observed with the hybrid model.
• Global contextual modeling in the CNN-ViT effectively distinguished visually similar defects like cracks and porosity.

Conclusions:

• The hybrid CNN-ViT framework offers improved accuracy and reduced misclassification for weld defect detection.
• The model provides transparent decision explanations, enhancing interpretability for industrial applications.
• This robust and explainable solution is well-suited for Industry 5.0-aligned intelligent weld inspection systems.