DEEP LEARNING FOR WELD DEFECT DETECTION USING CNN OR VISION TRANSFORMERS FUSION DETECTION

Abstract

Automated weld defect detection is essential for maintaining structural safety standards in manufacturing sectors including automotive production pipeline construction and shipbuilding operations. Non-destructive testing (NDT) methods which exist today require specialized knowledge from human operators to perform their functions but this creates testing delays and produces unreliable results. This research paper introduces an innovative hybrid fusion framework which integrates Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) to achieve precise weld defect classification and localization. CNNs effectively identify local crack and porosity textures while ViTs use their self-attention system to understand both long-range connections and overall visual information about weld bead shapes and profile deviations. The proposed system uses a late-fusion method which assigns dynamic importance to different features while it processes radiographic data through dual CNN and ViT systems before combining their output. The fusion model achieved 97.8% accuracy on a proprietary weld radiography dataset containing 5200 images which showed four different defect types while standalone CNN and ViT models achieved 91.2% and 93.5% accuracy respectively. The ablation studies show that the hybrid method decreases false positive rates by 48% when compared to CNN-based systems and it enhances the identification of small non-porous defects. The results show that CNN-ViT fusion provides an advanced solution which works well for industrial weld inspections performed in real time across different environments.