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

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]