Effective detection of metal surface defects based on double-line laser ultrasonic with convolutional neural networks

Abstract

The time–frequency analysis by smooth Pseudo-Wigner-Ville distribution (SPWVD) is utilized for the double-line laser ultrasonic signal processing, and the effective detection of the metal surface defect is achieved. The double-line source laser is adopted for achieving more defects information. The simulation model by using finite element method is established in a steel plate with three typical metal surface defects (i.e. crack, air hole and surface scratch) in detail. Besides, in order to improve the time resolution and frequency resolution of the signal, the SPWVD method is mainly used. In addition, the deep learning defect classification model based on VGG convolutional neural network (CNN) is set up, also, the data enhancement method is adopted to extend training data and improve the defects detection properties. The results show that, for different types of metal surface defects with sub-millimeter size, the classification accuracy of crack, air holes and scratch surface are 94.6%, 94% and 94.6%, respectively. The SPWVD and CNN algorithm for processing the laser ultrasonic signal and defects classification supplies a useful way to get the defect information, which is helpful for the ultrasonic signal processing and material evaluation.