Identification of defects in a coating wedge based on ultrasonic non-destructive testing methods and convolutional neural networks

The paper deals with the identification of a crack-like defect in a coated wedge based on ultrasonic nondestructive testing. The authors propose an approach of defect identification followed by determination of its geometrical parameters. The approach is based on a shadowed ultrasonic nondestructive testing method combined with deep machine learning technologies. A wedge-shaped area is inspected for the presence of an internal defect. On one edge of the wedge there is a source of ultrasonic vibrations, on the opposite edge there is a receiver. Passing through the coating and body of the wedge, part of the signal is reflected from inhomogeneities and defects that may be present in it. The signal reaching the opposite edge of the wedge is read by the receiver. The received data is processed by a neural network model, which predicts the presence or absence of an internal defect and, if present, determines geometric parameters such as size and position. A finite element model of ultrasonic wave propagation inside the wedge is constructed. Special damping layers are used, due to which the influence of parasitic signal reflections and its further propagation into the wedge body is significantly reduced. Based on the built model, the shadow method of ultrasonic scanning is implemented. This method implies that on one side of the wedge are installed excitation devices, and on the opposite side - receiving devices. Several numerical experiments for various combinations of geometric parameters of the wedge and the defect have been performed using a distributed computing system. Based on the obtained data, a neural network model was built and trained, capable of identifying the defect and determining its characteristics. The input of the model is spectrograms of the readout signal, and the output is values characterizing the defect.