Ultrasonic Rough Crack Characterization Using Time-of-Flight Diffraction With Self-Attention Neural Network

Area of Science:

• Materials Science
• Nondestructive Evaluation (NDE)
• Artificial Intelligence in Engineering

Background:

• Time-of-flight diffraction (ToFD) is crucial for ultrasonic nondestructive evaluation (NDE), excelling at sizing smooth cracks but struggling with rough, irregular defects.
• The complex diffraction waves from naturally occurring rough defects challenge traditional ToFD analysis, impacting accurate characterization and sizing.
• Developing advanced methods is essential to overcome the limitations of current ToFD techniques for irregular defect geometries.

Purpose of the Study:

• To propose and validate a novel self-attention (SA) deep learning method for interpreting ToFD A-scan signals to accurately size rough defects.
• To enhance the model's performance on realistic defects by employing transfer learning (TL) with simulated and experimental data.
• To compare the deep learning approach's accuracy against conventional Hilbert peak-to-peak sizing methods.

Main Methods:

• A self-attention (SA) deep learning model was developed to interpret ToFD A-scan signals for rough defect characterization.
• High-fidelity finite-element (FE) simulations using Pogo software generated synthetic datasets for model training and testing.
• Transfer learning (TL) was applied to fine-tune the model, adapting it from Gaussian rough defects to realistic thermal fatigue defects.

Main Results:

• The SA deep learning method demonstrated significantly reduced uncertainty and error in characterizing rough defects.
• Experimental validation using 2-D rough crack samples from additive manufacturing confirmed the model's effectiveness.
• The deep learning approach outperformed the conventional Hilbert peak-to-peak sizing method in accuracy.

Conclusions:

• The proposed self-attention deep learning method offers a robust and accurate solution for sizing rough defects using ToFD.
• Transfer learning effectively improves the model's generalization capability for real-world defect characterization.
• This AI-driven approach represents a significant advancement in ultrasonic NDE for improved defect assessment.