Near-surface defect detection in ultrasonic testing using domain-knowledge-informed self-supervised learning.

Area of Science:

• Materials Science
• Non-Destructive Testing
• Artificial Intelligence

Background:

• Artificial intelligence (AI) enhances ultrasonic testing (UT), but requires extensive labeled data, posing challenges.
• Conventional UT struggles with near-surface defect detection, focusing primarily on deeper flaws.
• Existing AI methods for UT face limitations due to data scarcity and near-surface defect challenges.

Purpose of the Study:

• Propose a novel, data-efficient AI method for near-surface defect detection in UT.
• Develop a self-supervised anomaly detection model incorporating domain knowledge for UT.
• Enable accurate defect detection and depth sizing without requiring labeled training data.

Main Methods:

• Generated synthetic faulty UT signals by fusing measured signals with back-wall reflections, respecting ultrasonic superposition principles.
• Developed a de-anomaly network to isolate subtle defect features within UT signals.
• Determined defect presence using the three-sigma rule on residual output and depth via time-of-flight calculation.

Main Results:

• The proposed method successfully detected near-surface defects in aluminum blocks under various conditions.
• Accurate defect depth determination was achieved using time-of-flight analysis on residual outputs.
• Qualitative and quantitative comparisons showed superior performance over existing methods for near-surface defect detection.

Conclusions:

• The novel self-supervised, domain-knowledge-integrated AI model effectively detects near-surface defects in UT without labeled data.
• The method demonstrates high accuracy in identifying defect presence and depth, addressing limitations of conventional UT.
• This approach offers a promising solution for enhanced non-destructive evaluation, particularly for challenging near-surface flaws.