Ultrasonic tomography imaging enhancement approach based on deep convolutional neural networks

Area of Science:

• Materials Science
• Non-destructive Testing
• Nuclear Engineering

Background:

• Structural health monitoring of containment liner plates (CLP) is crucial for nuclear power plant safety.
• Ultrasonic tomographic imaging, using methods like RAPID, can detect hidden defects.
• Lamb wave testing faces challenges due to multimodal dispersion, complicating mode selection.

Purpose of the Study:

• To enhance ultrasonic tomographic images of CLPs by addressing blurring issues.
• To improve the precision and clarity of defect detection in CLPs.
• To develop a cost-effective and efficient imaging technique for CLP structural integrity.

Main Methods:

• Sensitivity analysis was used to select the appropriate Lamb wave mode (S0).
• Deep learning architecture, specifically U-Net, was employed for image segmentation.
• Transfer learning was utilized due to the limited availability of training data for the U-Net model.

Main Results:

• The S0 Lamb wave mode was identified as the most sensitive for CLP inspection.
• U-Net architecture successfully segmented experimental ultrasonic tomographic images.
• Deep learning approaches effectively eliminated blurred zones in tomographic images, revealing clear defect edges.

Conclusions:

• Deep learning, particularly U-Net with transfer learning, significantly enhances ultrasonic tomographic imaging of CLPs.
• The developed method improves defect visualization, aiding in more accurate structural health assessments.
• This approach offers a promising solution for reliable and efficient non-destructive evaluation in nuclear facilities.