Ultrasonic characterization of small defects based on Res-ViT and unsupervised domain adaptation

Area of Science:

• Materials Science
• Non-Destructive Testing
• Artificial Intelligence

Background:

• Accurate characterization of small, inclined elliptical defects using ultrasound is challenging.
• Existing deep learning models like ResNet and ViT show promise but have limitations in generalizability.
• Domain adaptation techniques are crucial for improving model robustness in varying environments.

Purpose of the Study:

• To develop and evaluate a novel deep learning architecture for ultrasonic characterization of small, inclined elliptical defects.
• To enhance the generalizability and robustness of deep neural networks in noisy ultrasonic data.
• To reduce the uncertainty in characterizing defect parameters such as size, angle, and aspect ratio.

Main Methods:

• Performance evaluation of deep residual network (ResNet) and vision transformer (ViT).
• Proposal of a novel Res-ViT architecture fusing features from ResNet and ViT.
• Development of an unsupervised domain adaptation method using maximum mean discrepancy.
• Simulation studies at various noise levels to assess model robustness.
• Experimental validation on elliptical defects with large orientation angles.

Main Results:

• The proposed Res-ViT model demonstrated superior performance in characterizing defect parameters.
• Unsupervised domain adaptation significantly improved the model's generalizability in noisy environments.
• A 61% reduction in root-mean-square error (RMSE) for defect size was achieved compared to a benchmark method.
• The model effectively reduced characterization uncertainty for size, angle, and aspect ratio.

Conclusions:

• The Res-ViT architecture combined with unsupervised domain adaptation offers a robust solution for ultrasonic defect characterization.
• This approach significantly improves the accuracy and reliability of identifying small, inclined elliptical defects.
• The method shows strong potential for real-world non-destructive testing applications.