Ultrasonic Weld Quality Inspection Involving Strength Prediction and Defect Detection in Data-Constrained Training...

Area of Science:

• Manufacturing Engineering
• Materials Science
• Computer Vision

Background:

• Welding defects can lead to product failure, necessitating robust quality assessment methods.
• Current non-destructive testing (NDT) for weld defects often relies on expensive equipment like X-ray imaging.
• There is a need for cost-effective NDT solutions using readily available RGB cameras, despite limited public datasets.

Purpose of the Study:

• To develop a comprehensive weld quality assessment system using RGB images and deep learning.
• To enable accurate lap shear strength (LSS) prediction and visual weld defect detection from limited datasets.
• To create a cost-effective alternative to traditional NDT methods for weld inspection.

Main Methods:

• A convolutional autoencoder (CAE) was used to extract features from RGB images.
• A multimodal dataset was created by fusing image features with welding input parameter settings.
• Data augmentation techniques were employed to generate an ultrasonic weld defect dataset for training.
• Lap shear strength prediction and visual weld defect detection models were developed.

Main Results:

• Fusion of image features and input parameters reduced LSS prediction errors by 34% compared to using parameters alone.
• Visual weld defect detection achieved an accuracy of 74% on the limited dataset.
• The combined approach provides a complete weld quality inspection, enhancing reliability.

Conclusions:

• This research demonstrates the feasibility of complete weld quality assessment using RGB images and deep learning on extremely limited datasets.
• The developed multimodal approach significantly improves LSS prediction accuracy.
• The system offers a promising, cost-effective solution for non-destructive weld inspection, preventing premature joint failure.