LoHi-WELD: A Novel Industrial Dataset for Weld Defect Detection and Classification, a Deep Learning Study, and Future Perspectives

Abstract

The automated inspection of weld beads is of great importance for many industrial processes. Failures may cause a loss of mechanical resistance of the weld bead and compromise the manufactured part. Several methods have been proposed in the literature to address this problem, and recently, methods based on deep learning have gained prominence in terms of performance and applicability. However, such methods require vast and reliable datasets for different real defects, which have yet to be available in recent literature. Hence, this paper presents LoHi-WELD, an original and public database to address the problem of weld defect detection and classification of four common types of defects — pores, deposits, discontinuities, and stains — with 3,022 real weld bead images manually annotated for visual inspection, composed by low and high-resolution images, acquired from a Metal Active Gas robotic welding industrial process. We also explore variations of a baseline deep architecture for the proposed dataset based on a YOLOv7 network and discuss several case analyses. We show that a lightweight architecture, ideal for industrial edge devices, can achieve up to 0.69 of mean average precision (mAP) considering a fine-grained defect classification and 0.77 mAP for a coarse classification. Open challenges are also presented, promoting future research and enabling robust solutions for industrial scenarios. The proposed dataset, architecture, and trained models are publicly available on https://github.com/SylvioBlock/LoHi-Weld.

Keywords

• Weld defect detection and classification
• deep learning
• gas metal arc welding (GMAW)
• metal active gas welding (MAG)
• weld bead industrial
• public dataset