Deep learning ultrasonic computed tomography for non-destructive testing of workpieces

Area of Science:

• Materials Science and Engineering
• Computer Science
• Signal Processing

Background:

• Traditional ultrasonic non-destructive testing (NDT) suffers from limitations in imaging speed and accuracy for industrial applications.
• Ultrasonic computed tomography (USCT), while promising for improved inversion accuracy, faces computational challenges hindering real-time industrial use.

Purpose of the Study:

• To develop an efficient and high-precision hardware and neural network system for industrial ultrasonic NDT.
• To enable rapid reconstruction of cross-sectional images from ultrasonic data acquisition to final output.

Main Methods:

• A hardware system for ultrasonic tomography was developed, controlling ultrasonic array elements for 360° scanning and signal acquisition.
• A novel neural network, HAU2Net, integrating attention mechanisms into U2Net architecture, was designed for image reconstruction.
• The network was trained with dynamically adjusted loss function constraints and underwent dimensionality reduction for efficient data processing.

Main Results:

• HAU2Net achieved rapid cross-sectional image reconstruction within milliseconds.
• Experimental results demonstrated superior performance of HAU2Net over existing methods (K-wave, SegNet, MgNO, U2Net) in PSNR/SSIM, model size, and prediction time.
• Robustness tests confirmed HAU2Net's effectiveness under noisy and real-world conditions.

Conclusions:

• The proposed HAU2Net system offers an efficient and high-precision solution for industrial ultrasonic NDT.
• This research presents significant theoretical and practical implications for advancing industrial inspection technologies.