Scholarlink

<p>Arc-sound monitoring in wire-arc additive manufacturing (WAAM) provides millisecond-scale sensitivity to process instability, but its interpretation is easily confounded by geometry-dependent propagation, structural resonance, and environmental noise. To address this process- and geometry-coupled monitoring challenge, this study proposes a knowledge-intensive acoustic-electric operator-learning framework for unsupervised lack-of-fusion (LOF) detection. Synchronized voltage-current (VI) windows, local trajectory descriptors, and arc-sound signals are aligned and encoded as normalized log-Mel spectrogram tiles. A geometry-aware DeepONet is trained using only healthy data to learn the conditional mapping from electrical excitation and local geometry to the expected acoustic response. LOF is then detected from condition-aware residuals between measured and predicted spectra, which are further exploited for anomaly scoring and CNN-based discrimination. Across three WAAM process windows, DeepONet achieves the best reconstruction performance, with an MAE of 0.101, MSE of 0.066, SSIM of 0.550, R<sup>2</sup> of 0.802, and PSNR of 11.91&#xa0;dB. The DeepONet + CNN detector further attains an accuracy of 0.962, F1 of 0.883, AUC of 0.992, and AP of 0.962. These results indicate that the proposed framework provides an interpretable route for LOF monitoring and spatially resolved defect localization under the tested geometry-varying WAAM conditions.</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

Knowledge-intensive detection framework for unsupervised lack of fusion monitoring in wire-arc additive manufacturing via an acoustic-electric-coupled operator learning

  • 期刊论文

  • 发表日期：

    2026年7月27日

查看全文
  • Fengshuo Liu,
  • Haochen Mu,
  • Donghong Ding,
  • Jianping Zhao,
  • Xiaowei Wang

摘要

Arc-sound monitoring in wire-arc additive manufacturing (WAAM) provides millisecond-scale sensitivity to process instability, but its interpretation is easily confounded by geometry-dependent propagation, structural resonance, and environmental noise. To address this process- and geometry-coupled monitoring challenge, this study proposes a knowledge-intensive acoustic-electric operator-learning framework for unsupervised lack-of-fusion (LOF) detection. Synchronized voltage-current (VI) windows, local trajectory descriptors, and arc-sound signals are aligned and encoded as normalized log-Mel spectrogram tiles. A geometry-aware DeepONet is trained using only healthy data to learn the conditional mapping from electrical excitation and local geometry to the expected acoustic response. LOF is then detected from condition-aware residuals between measured and predicted spectra, which are further exploited for anomaly scoring and CNN-based discrimination. Across three WAAM process windows, DeepONet achieves the best reconstruction performance, with an MAE of 0.101, MSE of 0.066, SSIM of 0.550, R2 of 0.802, and PSNR of 11.91 dB. The DeepONet + CNN detector further attains an accuracy of 0.962, F1 of 0.883, AUC of 0.992, and AP of 0.962. These results indicate that the proposed framework provides an interpretable route for LOF monitoring and spatially resolved defect localization under the tested geometry-varying WAAM conditions.