Deep learning-based machine vision system for real-time edge fracture detection in the hole expansion test

초록

Edge fracture poses a critical challenge in sheet metal forming processes involving extruded holes, stretched flanges, or tight radius bends widely used in the automotive industry. Real-time detection of edge fractures in the hole expansion test, which is used to quantify the edge stretchability of sheet metal, has become both a technical challenge and an industrial necessity due to the diversity of materials. This study aims to develop an automated machine vision system for real-time detection of edge fractures and quantification of the hole expansion ratio during the hole expansion test. A deep learning-based vision pipeline is proposed, integrating dataset collection, dual-stage segmentation, and region-of-interest linearization to precisely capture the onset of cracks. This approach enables simultaneous recognizing edge fractures, quantifying the hole expansion ratio, and measuring crack gaps under varying surface reflectance and sheet thickness conditions. An ablation study evaluates the proposed approach against single-stage segmentation on the hole expansion test dataset prepared by hole punching method. Practical experimental validation confirms high segmentation accuracy and hole expansion ratio measurements with a maximum error of less than 2 % compared to manual evaluation. The findings offer a scalable and objective solution that overcomes the limitations of conventional monitoring methods and ensures consistent, real-time assessment across varying sheet materials.

키워드