Automatic Defect Classification for Infrared Thermography in CFRP based on Deep Learning Dense Convolutional Neural Network

Abstract

Carbon fiber reinforced polymer (CFRP) is an important composite material widely used in aerospace and other industries. However, long-term service in harsh environments can lead to various defects such as debonding, delamination, water, cracks, etc. Therefore, it becomes imperative to conduct non-destructive testing (NDT) on CFRP to ensure its structural integrity and safety. Infrared thermography was employed for defect classification in CFRP laminate and CFRP honeycomb sandwich composites (HSC) by applied a convolutional neural networks (CNN). The proposed automatic defect classification method based on CNN is one of the goals of NDE 4.0 to apply advanced technologies (such as deep learning and AI) to improve NDT efficiency and accuracy. The infrared detection dataset consisted of five classes: debonding, water, delamination, crack, and health. To effectively expand the dataset, offline data augmentation technique were employed. A deep learning technique of Dense convolutional neural network (DCNN) were proposed to defect classification. AlexNet, VGG-16, ResNet-50 and DenseNet-121 based on transfer learning fine-tuning model was applied to classify debonding, water, delamination, crack and health. The classification results were analyzed by using a confusion matrix. The results shown that the accuracy of AlexNet, VGG-16, ResNet-50 and DenseNet-121 were 92.34%, 82.86%, 88.30%, 98.48%, respectively. DenseNet-121 demonstrates good performance in defect detection and recognition with an accuracy of 98.48%, and DenseNet-121 has high application potential in accurately classify and recognize defects in deep learning technique.