Application of deep learning in nondestructive evaluation of metal microstructural grain size

In order to solve the problem of low detection accuracy caused by the lack of artificial prior knowledge, the complexity of grain scattering mechanisms and the limitations of ultrasonic evaluation methods, this paper proposes an intelligent method for the grain size evaluation of metals based on the Convolutional Neural Network (CNN) by laser-ultrasonics. The pulse echo signal transmitted by the bottom surface of the titanium alloy sample was obtained by the laser ultrasonic detection system. The wavelet time-frequency graphs of laser ultrasonic waveforms were obtained based on complex Morlet wavelet basis functions with different center frequencies and bandwidths. They were chosen as the input of the CNN model constructed by the Visual Geometry Group Network (VGGNet) with a large number of small convolutional kernels and pooling layers. The VGGNet model was trained, optimized and verified by the wavelet time-frequency graph sets. The results have shown that the model can accurately predict the grain size range of metals.