| Abstract Scope |
Recent advancements in AI, particularly over the past three years, have sparked a resurgence of neural network research and development, especially through deep learning techniques involving recurrent and convolutional neural networks. These modern approaches offer a reduced reliance on data preprocessing and improved classification accuracy. However, their application to welding has so far been limited to narrowly defined and, oftentimes, unbalanced data sets. Despite growing enthusiasm for neural networks, significant barriers to widespread adoption remain—chief among them the need for thousands of training samples to achieve sufficient classification accuracy for a narrowly defined problem. Although commercialization appears likely in the near future—particularly in the automatic detection and classification of defects in weld radiography—this form of narrow AI will continue to hinder adoption in most commercial manufacturing environments. |