Damage localization method using ultrasonic lamb waves and Wav2Vec2.0 neural network

Abstract

In this paper, a Wav2Vec2.0 neural network based on an attention mechanism is proposed to locate defects in array ultrasonic testing signals. This method does not require knowledge of the a priori condition of the sample sound velocity or the feature extraction of ultrasonic scattering signals. First, an array piezoelectric ultrasonic testing system is used to detect a signal through hole defects at different positions in the plate structure. Then, three different neural networks—1D-CNN, Muti-Transformer, and Wav2Vec2.0—are used to locate the defects in the collected ultrasonic testing data. The performance of the network is verified with the data set collected through finite element simulation and the experimental system, and the identification accuracy and the calculation efficiency of different networks are compared and analyzed. To provide a solution for the poor balance of the experimental data set and the weak noise resistance of the simulation data set, a data set expansion method based on time domain transformation technology is proposed. The research results show that, the positioning accuracy of the Wav2Vec2.0 neural network proposed in this article is 98.46%, and the positioning accuracy is superior to Muti Transformer and ID-CNN.