Scholarlink

<p>In this study, 132 AISI 316&#xa0;L foil samples were welded using different process parameter settings and an adjustable ring-mode laser beam source. A high-speed camera was employed for the in-process monitoring to observe the melt pool. Post-process images of the weld seam surfaces were recorded using a digital microscope and a profilometer to gather 2D and 3D information. Based on the measurements, this study proposed a two-stage decision-level sensor data fusion by combining in- and post-process images. The images were analyzed using Convolutional Neural Networks (CNNs), including the VGG19. These networks were either adapted to the application via a transfer-learning approach or retrained. The Gradient-weighted Class Activation Mapping method was used to visualize the classification decisions made by the CNNs and support the model selection. The VGG19 as the best model achieved a classification accuracy of 99.95% for the in-process image dataset and up to 97.22% on the digital microscope datasets. A two-stage sensor data fusion approach was finally implemented, including a segment-wise fusion of the model predictions for individual sub-images to obtain a single sample prediction in the first stage, followed by a Bayesian-based fusion in the second stage using the output sample probabilities. Using a limited test dataset of 26 samples, the multi-sensor-based fusion approach achieved correct classification of all the test samples (corresponding to a 95% confidence interval for the classification accuracy ranging from approximately 87% to 100%), thereby demonstrating the feasibility under controlled experimental 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

Multimodal sensor data fusion evaluating in- and post-process images for the quality control of laser-welded stainless-steel foils

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月9日

查看全文
  • Tony Weiss,
  • Paul Hoffmann,
  • Pawel Garkusha,
  • Fabian Vieltorf,
  • Michael F. Zaeh

摘要

In this study, 132 AISI 316 L foil samples were welded using different process parameter settings and an adjustable ring-mode laser beam source. A high-speed camera was employed for the in-process monitoring to observe the melt pool. Post-process images of the weld seam surfaces were recorded using a digital microscope and a profilometer to gather 2D and 3D information. Based on the measurements, this study proposed a two-stage decision-level sensor data fusion by combining in- and post-process images. The images were analyzed using Convolutional Neural Networks (CNNs), including the VGG19. These networks were either adapted to the application via a transfer-learning approach or retrained. The Gradient-weighted Class Activation Mapping method was used to visualize the classification decisions made by the CNNs and support the model selection. The VGG19 as the best model achieved a classification accuracy of 99.95% for the in-process image dataset and up to 97.22% on the digital microscope datasets. A two-stage sensor data fusion approach was finally implemented, including a segment-wise fusion of the model predictions for individual sub-images to obtain a single sample prediction in the first stage, followed by a Bayesian-based fusion in the second stage using the output sample probabilities. Using a limited test dataset of 26 samples, the multi-sensor-based fusion approach achieved correct classification of all the test samples (corresponding to a 95% confidence interval for the classification accuracy ranging from approximately 87% to 100%), thereby demonstrating the feasibility under controlled experimental conditions.