What question did this study set out to answer?

The aim is to improve defect identification in welding using a hybrid CNN-ViT framework for better accuracy and interpretability.

March 24, 2026Scientific Reports0 citationsOpen Access

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

KPKumar J. ParmarMarwadi University RJRahul JainPunjab Engineering College PAP. T. AnithaAcharya N. G. Ranga Agricultural University

Puntos clave

The aim is to improve defect identification in welding using a hybrid CNN-ViT framework for better accuracy and interpretability.
Developed a hybrid Convolutional Neural Network-Vision Transformer architecture.
Tested against a lightweight CNN baseline using the RIAWELC radiographic weld dataset.
Conducted misclassification-sensitive analysis utilizing confusion matrices and explainability methods like Grad-CAM.
Evaluated robustness with external GDXray dataset and fivefold cross-validation.
Hybrid CNN-ViT achieved 98.56% accuracy, outperforming the CNN baseline at 97.90%.
Misclassification rate improved by approximately 31%.
Global contextual modeling significantly enhanced defect discrimination.
Model showed strong performance on external data under different imaging conditions.

Resumen

Proper and valid identification of the defects in welding is also essential in the safety–critical manufacturing systems in order to maintain structural integrity. In general, even though deep learning-based inspection methods have demonstrated high classification accuracy, they frequently fail to distinguish similar-looking types of defects and misclassify them at high costs in practice. The paper presented an intelligent weld inspection framework, hybrid Convolutional Neural Network-Vision Transformer (CNN -ViT) architecture, aimed at improving reliability and interpretability, with potential relevance to Industry 5.0-oriented inspection systems. The suggested model was relatively tested against a tailored lightweight CNN baseline, provided that the same training conditions were implemented with the RIAWELC radiographic weld dataset. It is demonstrated that the hybrid CNN-ViT has the highest accuracy of 98.56%, compared to the CNN baseline (97.90%), at a relative difference of about 31% in the misclassification rate. Misclassification-sensitive analysis with confusion matrices aided by explainability methods with Grad-CAM and transformer self-attention maps proves that the global contextual modeling significantly enhances the discrimination of visually similar defects that include cracks, porosity, and absence of penetration. The trained model was also tested on external data (GDXray dataset) collected under different imaging conditions to test its robustness and it was found to perform well on generalization. Moreover, fivefold cross-validation was performed in order to evaluate how the model performs across different data partitions. These findings validate that the suggested hybrid attention-based framework, in addition to the accuracy improvement, leads to the decrease in the misclassification and the availability of the transparent decision explanations. In general, the suggested solution would be a technically robust and explainable framework that may support future Industry 5.0-aligned intelligent weld inspection systems.

Preguntar a la IA

Me gusta

Guardar

Ver artículo completo