Deep learning for imaging-free voids detection by ultrasonic data: bridging numerical data and model experiments

Area of Science:

• Materials Science
• Engineering
• Computer Science

Background:

• Non-destructive testing (NDT) using ultrasonics is crucial for internal defect detection.
• Deep Learning (DL) enhances ultrasonic NDT, but conventional methods depend heavily on image quality.
• End-to-end strategies offer improved defect detection by reconstructing velocity fields directly from ultrasonic data.

Purpose of the Study:

• To present an end-to-end, imaging-free framework for inner void detection using a Fully Convolutional Network (FCN).
• To validate the FCN's performance in both numerical simulations and physical model experiments.
• To demonstrate the potential of imaging-free DL for multi-scale void detection in engineering.

Main Methods:

• Utilized a Fully Convolutional Network (FCN) for direct velocity field reconstruction and defect detection.
• Employed SH-wave forward modeling and sparse traces frequency-wavenumber (F-K) filtering to process ultrasonic data.
• Integrated image processing techniques to map limited ultrasonic data to high-resolution velocity models.

Main Results:

• The FCN achieved a median Intersection over Union (IoU) of 0.70 in numerical tests with 1,200 void models.
• Experimental validation on a glass block demonstrated successful prediction of internal velocity models and accurate void localization.
• The framework showed a successful transition from synthetic data to physical measurements.

Conclusions:

• The imaging-free FCN framework provides a robust method for inner void detection directly from ultrasonic data.
• This approach overcomes limitations of image-dependent DL methods in ultrasonic NDT.
• The study highlights the potential of FCN for efficient and accurate multi-scale void detection in real-world engineering scenarios.