Microcrack Defect Quantification Using a Focusing High-Order SH Guided Wave EMAT: The Physics-Informed Deep Neural Network GuwNet

It is challenging to apply deep learning in professional fields that lack big data support, especially in industrial structure health assessments using ultrasonic guided wave nondestructive testing (NDT) method. To solve this problem, one feasible solution is to introduce the concept of NDT physics into a deep neural network to compensate for the network's poor predictive abilities when trained on small datasets. Therefore, we propose a physics-informed deep neural network, named GuwNet, based on a unidirectional oblique-focusing (UOF) high-frequency, high-order shear horizontal guided wave electromagnetic acoustic transducer (EMAT) to quantify microcrack defects more accurately. First, the designed focused-transmission omnidirectional-reception UOF-EMAT can produce pure high-frequency, high-order guided waves. Through a circumferential arrangement of multiple receiving transducers, the maximum amount of information can be obtained regarding the reflected waves of the defect. This method solves the inherent problems of an EMAT (i.e., low energy conversion efficiency) and achieves effective detection of microcrack defects. Second, we study the quantification principle of microcrack defects suitable for UOF-EMAT, and propose a deep neural network using physical knowledge regarding this theory. We rationally design the network structure based on the quantitative principles and logic obtained from this article. In addition, feedback and feedforward loss functions suitable for evaluating different forms of variables are proposed to integrate the physical concepts of ultrasonic guided wave testing into the neural network training. Finally, we verify the performance of the proposed GuwNet based on the UOF-EMAT. Compared with traditional nonphysics-informed methods, the length, depth, and direction of the quantification errors are reduced to 0.127 mm, 0.279% d_t , and 1.843°, respectively, and the average quantification error is reduced by more than 80%.