Transfer-learning-aided defect prediction in simply shaped CFRP specimens based on stress distribution obtained from finite element analysis and infrared stress measurement
Yuta Kojima, Kenta Hirayama, Katsuhiro Endo, Yoshihisa Harada, Mayu Muramatsu
Abstract
In this paper, we propose a framework of nondestructive testing for predicting the 3D structure of internal defects in carbon-fiber-reinforced plastic (CFRP) from the distribution of the sum of principal stresses on surfaces (DSPSS) through transfer learning. DSPSS is obtained from both the finite element method and infrared stress measurement results. Infrared stress measurements are based on Kelvin’s theory to convert surface temperature changes to DSPSS changes. The machine learning model used in this framework is a 3D convolutional neural network (CNN). The transfer learning method employed in this framework is as follows. First, a CNN that predicts the 3D structure of defects is trained using the DSPSS dataset by the finite element method and the 3D structure of internal defects. DSPSS is used with noise that imitates the noise generated by experimental factors such as temperature fluctuations in infrared stress measurements and differences in physical properties between the polymer resin and the carbon fiber bundle of CFRP. Next, the CNN is trained using the DSPSS dataset obtained by infrared stress measurement and the 3D structure of defects. The accuracy of the trained CNN is evaluated using DSPSS infrared stress measurements. We discuss the factors that enable us to predict the 3D defect data from the two-dimensional DSPSS using a variational autoencoder. The proposed method makes it possible to estimate internal defect information.