基于物理信息神经网络的激光超声波场研究  

Research on laser ultrasonic wavefield based on physical-informed neural network

作　　者：颜鑫 应恺宁 戴鹭楠 谭钧夫 沈中华[2] 倪辰荫[1] YAN Xin;YING Kaining;DAI Lunan;TAN Junfu;SHEN Zhonghua;NI Chenyin(School of Electronic Engineering and Optoelectronic Technology,Nanjing University of Science&Technology,Nanjing 210094,China;College of science,Nanjing University of Science&Technology,Nanjing 210094,China;College of Computer Science and Engineering,Nanjing University of Science&Technology,Nanjing 210094,China)

机构地区：[1]南京理工大学电子工程与光电技术学院,南京210094 [2]南京理工大学理学院,南京210094 [3]南京理工大学计算机科学与工程学院,南京210094

出　　处：《激光技术》2024年第1期105-113,共9页Laser Technology

基　　金：江苏省研究生科研与实践创新计划资助项目(KYCX22_0420)。

摘　　要：为了解决传统神经网络技术过于依赖数据资源,同时也无法运用数据中暗含的物理先验知识等局限性问题,采用物理信息神经网络(PINN)、基于超声传播的波动方程,利用数值计算实验的数据训练出了正向激光超声单模态(表面波)波场的PINN模型;建立了反向求解激光超声单模态波场参数的PINN模型,并对激光超声波场进行了正向成像和反向参数推演。结果表明,当探测点不包含激发点时,正向PINN在数据量仅为10%的情况下可得到高精度的波场图像,相比于原波场下降了一个数量级;即使在包含激发点时,反向PINN利用25%的波场数据不仅可以重建波场,且不需要人为地分析就可以求解控制方程的参数,与原波场数据的参数误差均在5%以内;与传统神经网络相比,PINN通过加入符合激光超声特性的控制方程,降低了神经网络对于训练数据稀疏性的依赖;与传统的激光超声波场建模相比,PINN构建的物理模型更简单,可自动求得控制方程的参数,有着更好的鲁棒性。该研究可为波场重建和参数反演激光超声无损检测技术提供参考,在激光超声领域有着广泛的应用前景。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.

关 键 词：激光技术 激光超声 神经网络 偏微分方程 波场成像 

分 类 号：O426[理学—声学] O436[理学—物理] TP183[机械工程—光学工程]