Title: TIG weld defect prediction from weld pool images using deep convolutional neural network and transfer learning

Authors: Rachna Verma; Arvind Kumar Verma

Addresses: Computer Section, Faculty of Science, JNV University, Jodhpur, Rajasthan, India ' Department of Production and Industrial Engineering, MBM University, Jodhpur, Rajasthan, India

Abstract: TIG welding is widely used in fabrication, but its quality depends on precise control of welding parameters. This study employs convolutional neural networks (CNNs) and transfer learning to predict welding defects from weld pool images. Six pre-trained CNNs (MobileNet, MobileNetV2, NasNetMobile, InceptionV3, ResNet50V2, and EfficientnetB0) are evaluated for their accuracy and real-time processing ability for a two-class problem (defective vs. non-defective welds) and a six-class problem (classifying good weld, burn through weld defect, contamination weld defect, lack of fusion weld defect, lack of shielding gas weld defect, and high travel speed weld defect). All the models except EfficientnetB0 achieved a very high accuracy. However, based on the inference time and memory size of the models, MobileNetV2 with 99.94% accuracy is recommended for developing an automated TIG welding systems, enabling real-time adjustment of parameters based on weld pool appearance, ensuring high-quality defect-free welds. [Submitted 29 January 2023; accepted 3 April 2024]

Keywords: TIG welding automation; weld defects; machine learning in TIG; pre-trained networks; CNN in TIG.

DOI: 10.1504/IJMR.2024.140279

International Journal of Manufacturing Research, 2024 Vol.19 No.2, pp.181 - 210

Received: 29 Jan 2023
Accepted: 03 Apr 2024
Published online: 01 Aug 2024 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article