Classifying character focuses for fracture process zone description in mounted carbon or epoxy covers with a convolutional neural system
Autour(s)
- Bing Pan, Lixuan Zhang, Chang Li, Don Chen, Zheng Xiang, Lee Chen
Abstract
This paper presents a novel X-ray Computed Tomography (CT) image analysis method to characterize the Fracture Process Zone (FPZ) in scaled centre-notched quasi-isotropic carbon/epoxy laminates. A total of 61 CT images of a small specimen were used to fine-tune a pre-trained Convolutional Neural Network (CNN) (i.e., VGG16) to classify fibre orientations. The proposed CNN model achieves a 100% accuracy when tested on the CT images of the same scale as the training set. However, the accuracy drops to a maximum of 84% when tested on unlabelled images of the specimens having larger scales potentially due to their lower resolutions. Another code was developed to automatically measure the size of the FPZ based on the CNN identified 0◦ plies in the largest specimen which agrees well with the manual measurement (on average within 3.3%). The whole classification and measurement process can be automated without human intervention.