About this Abstract |
Meeting |
2021 AWS Professional Program
|
Symposium
|
2021 AWS Professional Program
|
Presentation Title |
Automatic Welding Defect Classification Based on CNN of Different Structure |
Author(s) |
Xiaopeng Wang, Xu Wang, Zidong Lin, Xinghua Yu |
On-Site Speaker (Planned) |
Xiaopeng Wang |
Abstract Scope |
Tube and pipe welds in power generation industries requires intensive NDE, especially radiography. Current radiographic inspection primarily relies on human knowledge. It is costly and time-consuming. Recent advancement of deep learning technology could be used for automatic radiographic inspection. Welding defect inspection was explored by using the classification algorithm based on convolutional neural network(CNN). Thousands of digital radiographic images of tube weld acquired by double wall double images method were collected and manually classified, including the images with or without defect. In this work, the effect of CNN architecture of different structure, the ratio between the defect images and good images, and the sample size on the classification accuracy were investigated. It was suggested that the deeper depth of network and the residual block in the CNN architecture increase prediction accuracy. Whats more, compared with the sample ratio, the sample data size has a greater impact on the prediciton accuracy.
Key words: Welding defect classification, Deep learning, Dataset |
Proceedings Inclusion? |
Definite: Other |