Scholarlink

<p>In numerous industrial applications, welded components are frequently subjected to loads with critical implications for their structural integrity and reliability, making weld defect detection and classification of paramount importance. In this paper, we propose a transfer learning for weld defect classification based on a new Xception Convolutional Neural Network (CNN) architecture. Our approach to address the challenge of unavailability of labelled welding data follows a two step procedure as we utilize pre-training ImageNet weights for features extraction and fine-tuning. We designed, enlarged and brought up a tailored development nitinol dataset containing mismanaged instances (crack; burn-through; undercut; defect-free) from Tungsten Inert Gas (TIG)-welded stainless-steel specimens. The fine-tuned Xception model yielded an overall classification accuracy of 94.3% with macro-averaged precision of 93.8%, recall of 93.1%, and F1-score reached at (model performance). These results show the effectiveness of the proposed approach over traditional machine learning and baseline deep learning methods, further supporting its potential for direct application in automated NDT within industrial quality inspection pipelines.</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 by transfer learning with xception neural network

  • 期刊论文

  • 发表日期：

    2026年7月22日

查看全文
  • M. Sathya,
  • Ramesh Vellaichamy,
  • T. Venish Kumar,
  • P. Vijayakumar,
  • S. Irudayaraj,
  • K. Ram Prasad,
  • B. Radha Krishnan

摘要

In numerous industrial applications, welded components are frequently subjected to loads with critical implications for their structural integrity and reliability, making weld defect detection and classification of paramount importance. In this paper, we propose a transfer learning for weld defect classification based on a new Xception Convolutional Neural Network (CNN) architecture. Our approach to address the challenge of unavailability of labelled welding data follows a two step procedure as we utilize pre-training ImageNet weights for features extraction and fine-tuning. We designed, enlarged and brought up a tailored development nitinol dataset containing mismanaged instances (crack; burn-through; undercut; defect-free) from Tungsten Inert Gas (TIG)-welded stainless-steel specimens. The fine-tuned Xception model yielded an overall classification accuracy of 94.3% with macro-averaged precision of 93.8%, recall of 93.1%, and F1-score reached at (model performance). These results show the effectiveness of the proposed approach over traditional machine learning and baseline deep learning methods, further supporting its potential for direct application in automated NDT within industrial quality inspection pipelines.