Scholarlink

<p>Weld quality inspection is critical for structural safety in automotive and aerospace manufacturing; however, conventional non-destructive testing (NDT) faces challenges from subjective interpretation and class imbalance. We present <i>WeldNet</i>, a multimodal deep learning framework that combines attention-based sensor fusion with Morphologically Guided Synthetic Augmentation (MGSA) for the quality assessment of resistance spot welds (RSW). Using a publicly available dataset of 495 RSW samples with surface images, infrared thermography, and process parameters, <i>WeldNet</i> achieved 95.14% accuracy (±0.85%) and 0.854 Macro-F1 (±0.012) through 5-fold cross-validation, significantly outperforming the Synthetic Minority Over-sampling Technique (SMOTE), focal loss, and ensemble baselines (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(p&lt;0.001\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>p</mi> <mo>&lt;</mo> <mn>0.001</mn> </mrow> </math></EquationSource> </InlineEquation>). Synthetic augmentation validated by certified NDT practitioners using Cohen’s Kappa statistic (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\kappa =0.82\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>κ</mi> <mo>=</mo> <mn>0.82</mn> </mrow> </math></EquationSource> </InlineEquation>) improved minority class detection by 12.4 percentage points, whereas multimodal fusion provided 27–58% gains over single-modality approaches. With a 9.2&#xa0;ms neural network inference time, <i>WeldNet</i> shows proof-of-concept feasibility for process-integrated inspection, although multi-site validation across diverse production environments remains essential before industrial deployment.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

WeldNet: a multimodal deep learning framework with validated synthetic augmentation for resistance spot weld quality assessment

  • 期刊论文

  • 发表日期：

    2026年4月13日

查看全文
  • Shapna Rani Edwin Raj,
  • Sowmyalakshmi Ravindran

摘要

Weld quality inspection is critical for structural safety in automotive and aerospace manufacturing; however, conventional non-destructive testing (NDT) faces challenges from subjective interpretation and class imbalance. We present WeldNet, a multimodal deep learning framework that combines attention-based sensor fusion with Morphologically Guided Synthetic Augmentation (MGSA) for the quality assessment of resistance spot welds (RSW). Using a publicly available dataset of 495 RSW samples with surface images, infrared thermography, and process parameters, WeldNet achieved 95.14% accuracy (±0.85%) and 0.854 Macro-F1 (±0.012) through 5-fold cross-validation, significantly outperforming the Synthetic Minority Over-sampling Technique (SMOTE), focal loss, and ensemble baselines ( \(p<0.001\) p < 0.001 ). Synthetic augmentation validated by certified NDT practitioners using Cohen’s Kappa statistic ( \(\kappa =0.82\) κ = 0.82 ) improved minority class detection by 12.4 percentage points, whereas multimodal fusion provided 27–58% gains over single-modality approaches. With a 9.2 ms neural network inference time, WeldNet shows proof-of-concept feasibility for process-integrated inspection, although multi-site validation across diverse production environments remains essential before industrial deployment.