Scholarlink

<p>This study proposes a multimodal CNN-Transformer model that fuses acoustic and visual information to perform ahead-of-time temporal prediction of weld penetration states in tungsten inert gas (TIG) butt-joint welding of flat plates. Instead of conventional frame-wise classification or two-stage video forecasting followed by regression, the task is formulated as time-indexed multiclass prediction at a specified lead time, while retaining state-transition samples to reflect continuous penetration evolution. Because molten pool images are often degraded by arc radiation, spatter, and thermal effects, synchronized acoustic signals are incorporated as complementary cues and transformed into acoustic spectrograms using the short-time Fourier transform. Spatial features are extracted from molten pool images and acoustic spectrograms by two CNN branches, fused through cross-attention and gated fusion, and then modeled temporally by a Transformer to predict the future penetration state. The model achieves accuracies of 93.85%, 96.23%, and 94.73% at lead times of 0.05&#xa0;s, 0.5&#xa0;s, and 1.0&#xa0;s, respectively, for three classes including under penetration, full penetration, and burn through. Under different process-parameter settings, accuracies of 89.64%, 93.38%, and 87.94% are maintained, indicating good generalization and robustness. The approach provides a practical basis for online monitoring and quality control in TIG welding.</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

Online welding quality monitoring via vision-acoustic fusion with future-state prediction

  • 期刊论文

  • 发表日期：

    2026年2月1日

查看全文
  • Yiyang Shua,
  • Peng Gao,
  • Jing Han,
  • Lianfa Bai,
  • Zhuang Zhao

摘要

This study proposes a multimodal CNN-Transformer model that fuses acoustic and visual information to perform ahead-of-time temporal prediction of weld penetration states in tungsten inert gas (TIG) butt-joint welding of flat plates. Instead of conventional frame-wise classification or two-stage video forecasting followed by regression, the task is formulated as time-indexed multiclass prediction at a specified lead time, while retaining state-transition samples to reflect continuous penetration evolution. Because molten pool images are often degraded by arc radiation, spatter, and thermal effects, synchronized acoustic signals are incorporated as complementary cues and transformed into acoustic spectrograms using the short-time Fourier transform. Spatial features are extracted from molten pool images and acoustic spectrograms by two CNN branches, fused through cross-attention and gated fusion, and then modeled temporally by a Transformer to predict the future penetration state. The model achieves accuracies of 93.85%, 96.23%, and 94.73% at lead times of 0.05 s, 0.5 s, and 1.0 s, respectively, for three classes including under penetration, full penetration, and burn through. Under different process-parameter settings, accuracies of 89.64%, 93.38%, and 87.94% are maintained, indicating good generalization and robustness. The approach provides a practical basis for online monitoring and quality control in TIG welding.