期刊
INTERNATIONAL JOURNAL OF MECHANICAL SCIENCES
卷 230, 期 -, 页码 -出版社
PERGAMON-ELSEVIER SCIENCE LTD
DOI: 10.1016/j.ijmecsci.2022.107529
关键词
Structural health monitoring; Digital image correlation (DIC); Convolution neural networks (CNN); Semantic segmentation
资金
- National Natural Science Foundation of China [52075157]
- Natural Science Foundation of Hunan Province, China [2021JJ30122]
- Fundamental Research Funds for the Central Universities
A convolutional neural network based image semantic segmentation technique is proposed for pixel-level classification of DIC strain field images, achieving automated scrutiny and results free from human interference. By using a pre-trained CNN model as the backbone network through transfer learning algorithm, high accuracy in semantic segmentation can be obtained.
Vision-based damage detection in carbon fiber-reinforced plastic (CFRP) composites can be interfered by such factors as surface texture, stains and lighting. A digital image correlation (DIC) based surface strain monitoring technique, on the other hand, enables to track the change of strain distribution. It is promising to develop a new approach for online structural health monitoring (SHM), in which the DIC strain contours can be scrutinized automatically and the results are no longer substantially subjected to human interference. In this study, a convolutional neural network (CNN) based image semantic segmentation technique is proposed for pixel-level classification of DIC strain field images. A DeepLabv3+ encoder-decoder architecture combined with different feature extraction networks is investigated. The training dataset and validation of the model are obtained through finite element (FE) simulation. The images of quasi-static axial tensile strain field obtained from 2D-DIC are used to test the accuracy and efficiency of the trained CNN model. It is found that use of a pre-trained ResNet-50 CNN model as the backbone network of DeepLabv3+ architecture through a transfer learning algorithm can make the semantic segmentation results reach a mean intersection over union of 0.9236. The prediction accuracy of the semantic segmentation model trained from the FE data is comparable with that of the model trained from the experimental data, which demonstrates that the proposed machine learning approach for DIC measurement is cost-effective.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据