<p>Currently, there is a lack of integrated application solutions for the rapid identification, localization, and quantitative evaluation of internal defects in weld seams within industrial scenarios. Based on this, this study proposes a hybrid detection scheme that integrates You Only Look Once version 8 (YOLOv8) with the Cascade Feedforward Neural Network (CFNN), establishing an intelligent technical system that encompasses the entire workflow of defect quality inspection. In this scheme, YOLOv8 utilizes a lightweight feature pyramid network to enable accurate identification of defects; CFNN constructs a nonlinear mapping model between the pixel coordinates of defect features and their corresponding physical coordinates to achieve accurate localization; and a novel geometric integration approach is proposed to calculate the cross-sectional area of defect contours, thereby facilitating quantitative defect evaluation. The verification experimental results indicate that the defect recognition accuracy achieves 100%. The average errors for the lateral and vertical coordinates of the defects (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">${\text{W}}_{1}^{}$</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mtext>W</mtext> <mrow> <mn>1</mn> </mrow> <mrow> <mrow /> <mo>∗</mo> </mrow> </mmultiscripts> </math></EquationSource> </InlineEquation>,<InlineEquation ID="IEq2"> <EquationSource Format="TEX">${\text{W}}_{2}^{}$</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mtext>W</mtext> <mrow> <mn>2</mn> </mrow> <mrow> <mrow /> <mo>∗</mo> </mrow> </mmultiscripts> </math></EquationSource> </InlineEquation>) are 0.000332 mm (at the sixth decimal place) and 0.000576 mm (at the sixth decimal place), respectively, while the average error for the depth coordinate (<InlineEquation ID="IEq3"> <EquationSource Format="TEX">${\text{h}}^{}$</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow> <mtext>h</mtext> </mrow> <mrow /> <mrow> <mrow /> <mo>∗</mo> </mrow> </mmultiscripts> </math></EquationSource> </InlineEquation>) is 0.001577 mm. In the process of quantitative evaluation of defects, the average errors in the height (<InlineEquation ID="IEq4"> <EquationSource Format="TEX">${\text{H}}^{}$</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow> <mtext>H</mtext> </mrow> <mrow /> <mrow> <mrow /> <mo>∗</mo> </mrow> </mmultiscripts> </math></EquationSource> </InlineEquation>) and cross-sectional area (<InlineEquation ID="IEq5"> <EquationSource Format="TEX">${\text{S}}^{*}$</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow> <mtext>S</mtext> </mrow> <mrow /> <mrow> <mrow /> <mo>∗</mo> </mrow> </mmultiscripts> </math></EquationSource> </InlineEquation>) of weld defects are 0.00282 mm and 1.46 mm<sup>2</sup>, respectively. This study presents a comprehensive and integrated technical solution for the online detection of weld defects in industrial environments, which significantly advances the practical implementation of ultrasonic testing technology in actual industrial applications.</p>

Hybrid Deep Learning-Based Automatic Identification, Localization and Quantitative Evaluation of Internal Defects in Welds

期刊论文
发表日期：

2026年4月29日

查看全文

Qi Zheng,
Hao Wang,
Xiaohui Zhao,
Xiujun Wang,
Mengran Li,
Chao Chen

摘要

Currently, there is a lack of integrated application solutions for the rapid identification, localization, and quantitative evaluation of internal defects in weld seams within industrial scenarios. Based on this, this study proposes a hybrid detection scheme that integrates You Only Look Once version 8 (YOLOv8) with the Cascade Feedforward Neural Network (CFNN), establishing an intelligent technical system that encompasses the entire workflow of defect quality inspection. In this scheme, YOLOv8 utilizes a lightweight feature pyramid network to enable accurate identification of defects; CFNN constructs a nonlinear mapping model between the pixel coordinates of defect features and their corresponding physical coordinates to achieve accurate localization; and a novel geometric integration approach is proposed to calculate the cross-sectional area of defect contours, thereby facilitating quantitative defect evaluation. The verification experimental results indicate that the defect recognition accuracy achieves 100%. The average errors for the lateral and vertical coordinates of the defects ( ${\text{W}}_{1}^{*}$ $W_{1}^{*}$ , ${\text{W}}_{2}^{*}$ $W_{2}^{*}$ ) are 0.000332 mm (at the sixth decimal place) and 0.000576 mm (at the sixth decimal place), respectively, while the average error for the depth coordinate ( ${\text{h}}^{*}$ $h^{*}$ ) is 0.001577 mm. In the process of quantitative evaluation of defects, the average errors in the height ( ${\text{H}}^{*}$ $H^{*}$ ) and cross-sectional area ( ${\text{S}}^{*}$ $S^{*}$ ) of weld defects are 0.00282 mm and 1.46 mm², respectively. This study presents a comprehensive and integrated technical solution for the online detection of weld defects in industrial environments, which significantly advances the practical implementation of ultrasonic testing technology in actual industrial applications.