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

Abstract

Traditional ultrasonic non-destructive testing (NDT) techniques face dual challenges in industrial applications: imaging accuracy and imaging speed. The application of ultrasonic computed tomography (USCT) in the medical field provides a new approach for industrial NDT. Although this technique improves inversion accuracy, the intensive computation required for inversion limits its potential for online applications in industrial testing. To address these issues, this paper proposes a hardware system for ultrasonic tomography and a neural network inversion method integrating attention mechanisms, enabling the complete process from ultrasonic data acquisition to rapid reconstruction of cross-sectional images. The testbed controls the ultrasonic array elements to emit and receive signals, performing 360° scanning of the object under test and obtaining a signal matrix. After dimensionality reduction, the data is input into HAU2Net, which provides a cross-sectional image within milliseconds. This network incorporates attention heads at different encoding and decoding layers of U2Net and is trained with dynamically adjusted loss function constraints. It effectively processes the tomographic ultrasonic data collected by the hardware platform, extracting features and reconstructing the images. Compared with methods such as K-wave, SegNet, MgNO, and U2Net, experimental results show that the proposed HAU2Net outperforms others in terms of PSNR/SSIM metrics, model size, and prediction time. Additionally, robustness tests under noise and real-world conditions further validate the superior performance of HAU2Net. This research provides an efficient and high-precision solution for industrial ultrasonic NDT, with significant theoretical and practical implications.

Keywords: Deep learning; Image reconstruction; Industrial dataset; Non-destructive testing; Ultrasound computed tomography (USCT).