We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
基于物理信息神经网络的激光超声波场研究.
- Authors
颜 鑫; 应恺宁; 戴鹭楠; 谭钧夫; 沈中华; 倪辰荫
- Abstract
In recent years, non-destructive detection technology based on deep learning has developed rapidly, but traditional neural network technology relies too much on data resources and cannot use the physical prior knowledge implied in the data, which has many limitations. In order to solve this problem, physical-informed neural networks (PINN) were used in this paper. Based on the wave equation of ultrasonic propagation, the forward PINN model of laser ultrasonic single-mode (surface wave) wave field was trained by using the data of numerical calculation, and the inverse PINN model for solving laser ultrasonic single-mode wave field parameters was further established; therefore, the forward imaging and inverse parameter deduction of laser ultrasonic field were carried out. The results show that when detection points do not contain the excitation point, the forward PINN can obtain a high-precision wave field image when the data volume is only 10%, which is an order of magnitude lower than the original wave field; even when the excitation point is included, the reverse PINN can not only reconstruct the wave field by using 25% of the wave field data, but also solve the parameters of the control equation without artificial analysis, and the error of the parameters with the original wave field data is within 5%. Compared with the traditional laser ultrasonic field modeling, the physical model built by PINN is simpler, which can automatically obtain the parameters of the control equation and has better robustness. This research can provide a reference for wave field reconstruction and parameter inversion laser ultrasonic nondestructive testing technology, so PINN has broad application prospects in the field of laser ultrasonic.
- Publication
Laser Technology, 2024, Vol 48, Issue 1, p105
- ISSN
1001-3806
- Publication type
Article
- DOI
10.7510/jgjs.issn.1001-3806.2024.01.017