Spatially-informed deep learning for full-field ultrasonic nondestructive evaluation

This paper presents a deep-learning approach for estimating local material thickness from steady-state wavefield response measurements resulting from ultrasonic excitation of structural elements positioned at arbitrary angles from a scanning laser-Doppler vibrometer. Current processing techniques for such measurements can be unreliable because they do not account for warping and scaling of the response with respect to the scanned surface's position relative to the scanning vibrometer. To address this limitation, the proposed deep-learning model combines the full-field quadrature components of the measurement with an additional input for range data using an early fusion approach that provides the network with spatial context of the inspection environment. Additionally, the model employs deformable convolution layers that project locally onto the surface of the inspection specimen, improving the spatial invariance of the network. To train the model, a simulation pipeline was developed to generate steady-state ultrasonic response measurements and depth field measurements of aluminum alloy 6061 plates with random variations in thickness and orientation. The model was evaluated using both simulated and real-world measurements, which demonstrated that the combination of deformable convolution and early fusion outperformed other model configurations that either fully omitted or employed only one of these techniques. Furthermore, the model surpassed a non-machine learning baseline when evaluated on experimental measurements of two different plates across a range of scanning positions, orientations, and distances.