Fatigue life prediction of composite materials using strain distribution images and a deep convolution neural network

Abstract

The damage process of composite materials, such as short fiber-reinforced plastics (SFRP), is complex. Therefore, it is necessary to accurately represent the damage process in fatigue life prediction. Herein, fatigue life prediction was conducted by combining the digital image correlation method, which is a non-destructive testing technique, with a convolutional neural network (CNN), using Xception as the network architecture. High prediction accuracy was obtained when training and testing were performed on the same SFRP specimens. In contrast, using different specimens for training and testing resulted in lower accuracy. This issue may be improved by increasing the number of specimens. The regions of interest in the model were visualized by Gradient-weighted Class Activation Mapping. Notably, the model indicated the breaking point as the region of interest from the early stages of the test. The breaking point was identified at an earlier stage by the CNN than by visual inspection, demonstrating the potential for a new method of damage observation.