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

Non-destructive testing (NDT) based on ultrasonics is widely used for internal defect detection. To enhance the efficiency of defect detection, Deep Learning (DL) has been incorporated into ultrasonic NDT workflows. However, conventional methods rely on computer-vision-based DL to identify defects from post-imaging data, and their performance is highly related to the quality of the image and ultrasonic data. Consequently, end-to-end strategies are more suitable, as they can reconstruct the velocity field directly from ultrasonic data and detect defects by the difference of velocities. End-to-end methods have demonstrated excellent performance in numerical tests. However, they lack direct and precise comparison between predictions and true models in field application. Thus, this study presented an end-to-end, imaging-free framework for inner void detection by a Fully Convolutional Network (FCN). This net was tested by numerical tests and model experiments. The methodology in this study includes FCN, SH-wave forward modeling, sparse traces frequency-wavenumber (F-K) filtering, and image processing. The sparse traces F-K filtering mitigates boundary reflections in model experiments, and image processing realizes the mapping from limited ultrasonic data to a high-resolution velocity model. The FCN was trained on 1,200 randomly generated void models, achieving a median Intersection over Union (IoU) of 0.70 (with a maximum of 0.92, a minimum of 0.11) in numerical tests. Furthermore, the model was validated using experimental data acquired via an array ultrasound device on a homogeneous glass block with preset voids. The FCN successfully predicted the internal velocity model directly from the experimental ultrasonic data, accurately identifying the spatial distribution of the voids. The model experiment results demonstrate a successful transition from synthetic data to physical measurement, highlighting the potential of imaging-free FCN for multi-scale void detection in engineering scenarios.