End-to-end prediction of weld penetration: A deep learning and transfer learning based method. (March 2021)

• Record Type: Journal Article
• Authors: Jiao, Wenhua; Wang, Qiyue; Cheng, Yongchao; Zhang, YuMing
• Is Part Of: Journal of manufacturing processes. Volume 63(2021)
• Page Start: 191
• Page End: 197
• Publication Date: 2021-03
• Subjects: End-to-end deep learning -- Convolutional neural network (CNN) -- Residual neural network (ResNet) -- Transfer learning -- Weld penetration; Production management -- Data processing -- Periodicals; Manufacturing processes -- Periodicals; Procestechnologie; Productietechniek; Production -- Gestion -- Informatique -- Périodiques; Fabrication -- Périodiques; Manufacturing processes; Production management -- Data processing; Periodicals; 670.5
• DOI: 10.1016/j.jmapro.2020.01.044
• Languages: English
• ISSNs: 1526-6125

Abstract

Abstract: Weld penetration identification is a long-standing and challenging problem due to the spatial limitation in sensing the back-side of weld joints in practical welding, and the key characteristic information of welding is difficult to be defined and extracted in the complex welding process. This paper proposes an end-to-end deep learning approach to predict the weld penetration status from top-side images during welding. In this method, a passive vision sensing system with two cameras is developed to monitor the top-side and back-bead information simultaneously. Then, the weld joints are classified as three classes i.e. under, desirable and excessive penetration depending on the back-bead width. Taking the top-side images as inputs and corresponding penetration status as labels, an end-to-end convolutional neural network (CNN) is designed and trained where the features are defined and extracted automatically. Testing experiments demonstrate 92.70 % as the classification accuracy. In order to increase the accuracy and training speed, a transfer learning approach based on residual neural network (ResNet) is developed. This ResNet-based model is pre-trained on ImageNet dataset to process a better feature extracting ability and its fully-connected layers are modified based on our own dataset. The experiments show that this transfer learning approach can decrease the training time with the prediction accuracy improving to 96.35 %.