Detection and Identification of Welding Defects Through Deep Learning Models

The use of welding components has grown significantly across a wide range of industries recently, making their quality and reliability an essential component of product design and manufacturing. Determining the flaws including gaps, misalignment, cracks, spatter, and other imperfections in weld components is a laborious task that calls for expert workers and time-consuming manual examinations. The proposed algorithm entails using a collection of pictures of weld components with and without flaws to train a deep learning architecture. The model is trained through the application of transfer learning methods, and is optimized to achieve higher accuracy and lower false positives. The proposed algorithm’s efficiency is evaluated with respect to its speed and accuracy. The objective is to evaluate the accuracy of a few of the deep learning models leading to combining manual tasks with artificial intelligence approaches. The inference is that You Look Only Once-version 5 (YOLOv5) architecture provides precision scores relating to dispersion, gaps, and misalignment of $\mathbf{2 0 \%}$, $\mathbf{2 5 \%}$, and $\mathbf{5 0 \%}$, with a mean average precision (mAP) of $50 \%-60 \%$ and YOLO $\mathbf{v} 8$ architecture provides precision scores of $39 \%$, $35 \%$ and $55 \%$ with a mAP of $55 \%-65 \%$ respectively. Considerable increase in accuracy together with time and resource savings is achieved by incorporating artificial intelligence for automatic detection and classification of welded components, which results in decrease of human intervention for quality control.