Scholarlink

<p>Phased-array ultrasonic testing (PAUT) is widely used for weld inspection in aerospace and other safety-critical industries, but automated interpretation of A-scan signals is hindered by noise, weld geometry-induced echoes, and class imbalance among defect types. This study proposes a deep learning framework that classifies butt-weld A-scans into Normal, Linear defect, and Volumetric defect classes by combining a domain-knowledge-based two-stage preprocessing scheme with a convolutional neural network whose feature maps are conditioned on pre-calculated physical features through a feature-wise linear modulation (FiLM) layer. On 40 carbon-steel and stainless-steel specimens, tenfold cross-validation at the inspection-file level yields 87.66% mean accuracy and 81.02% macro F1, with preprocessing improving macro F1 by 10.98%p and the FiLM layer raising volumetric defect accuracy by 8.01%p. When applied as a binary defect detector by merging the two defect classes, the framework reaches 99.26% accuracy, a level well-suited to rapid screening of field inspection results.</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

Weld Defect Classification in Phased-Array Ultrasonic Testing Using a CNN with FiLM Layer

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月13日

查看全文
  • Jaeho Lee,
  • Chen Ciang Chia,
  • Jung-Ryul Lee

摘要

Phased-array ultrasonic testing (PAUT) is widely used for weld inspection in aerospace and other safety-critical industries, but automated interpretation of A-scan signals is hindered by noise, weld geometry-induced echoes, and class imbalance among defect types. This study proposes a deep learning framework that classifies butt-weld A-scans into Normal, Linear defect, and Volumetric defect classes by combining a domain-knowledge-based two-stage preprocessing scheme with a convolutional neural network whose feature maps are conditioned on pre-calculated physical features through a feature-wise linear modulation (FiLM) layer. On 40 carbon-steel and stainless-steel specimens, tenfold cross-validation at the inspection-file level yields 87.66% mean accuracy and 81.02% macro F1, with preprocessing improving macro F1 by 10.98%p and the FiLM layer raising volumetric defect accuracy by 8.01%p. When applied as a binary defect detector by merging the two defect classes, the framework reaches 99.26% accuracy, a level well-suited to rapid screening of field inspection results.