Using deep learning to identify the depth of metal surface defects with narrowband SAW signals

Abstract

An effective nondestructive testing technique that enables the detection and quantification of surface and subsurface defects is highly demanded for assuring the safety and reliability of materials. In this paper, we classified and regressed defects with different depths. A transient thermal grating (TTG) method induced by a pulsed laser is used to characterize the properties of surface defects. Based on thermo-elasticity theory, a theoretical model with different types of surface defects is established with the finite element method (FEM). Surface acoustic waves (SAWs) excited by the TTG method propagate at the surface of a sample with defects at different penetration depths and are measured by a vibrometer.By combining the simulation and experimental results, the time–frequency scalograms of 8200 images were obtained by wavelet transform analysis. 70 % of the images are randomly selected to train the residual network (Resnet) model, and the remaining data (2460 images), as well as 10 scalograms in an independent experiment, are used as previously unknown experimental signals to test the robustness of the Resnet model. The results indicate that the Resnet with scalograms via wavelet transform achieves a 96.21 % classification accuracy on the validation set, with a prediction accuracy of 90.00 % for the 10 scalograms. Moreover, compared with three widely used machine learning algorithms, VGG16, back propagation (BP), and support vector machine (SVM), this method achieves the highest classification accuracy. Finally, 500 data selected randomly from the previously trained experimental database and 10 data in an independent experiment are used for regression prediction. The result shows that the data are well located along or close to a straight line with a slope of 1. The statistical results suggest that 80.6 % (66.0 %) of the residuals of the experimental data fell within the size range ±10 μm (±5 μm). This work proves the feasibility and reliability of the combination of the Resnet model with wavelet transform and numerical simulation to identify the depth of surface defects using laser ultrasonics.