Scholarlink

<p>Laser cladding processes require reliable in-situ monitoring of melt-pool geometry to ensure process stability, deposition quality, and repeatability in industrial applications. This study investigates the use of deep learning–based object detection architectures, specifically YOLOv5 and SSD, for real-time estimation of weld-pool width and length using high dynamic range visible-spectrum imaging. Experiments were conducted using an IPG Ytterbium fiber laser under different laser cladding conditions, including variations in laser power, travel speed, and powder feed configurations. Image datasets acquired during deposition were used to train and evaluate the detection models under heterogeneous optical and operational conditions. Both architectures achieved robust spatial localization and reliable geometric estimation of the weld pool across multiple cladding scenarios. YOLOv5 achieved inference rates above 60 FPS while maintaining geometric accuracy comparable to SSD, demonstrating superior suitability for online real-time monitoring. The results show that object-detection approaches provide a computationally efficient alternative to segmentation-based methods for weld-pool monitoring, supporting future implementation in adaptive control and intelligent laser manufacturing systems.</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

Real-time weld pool monitoring in laser cladding using object detection models

  • 期刊论文

  • 发表日期：

    2026年5月27日

查看全文
  • Giamnfranko Segovia,
  • Viviana Meruane,
  • Gentry Wood,
  • Dmytro Havrylov,
  • Patricio F. Mendez

摘要

Laser cladding processes require reliable in-situ monitoring of melt-pool geometry to ensure process stability, deposition quality, and repeatability in industrial applications. This study investigates the use of deep learning–based object detection architectures, specifically YOLOv5 and SSD, for real-time estimation of weld-pool width and length using high dynamic range visible-spectrum imaging. Experiments were conducted using an IPG Ytterbium fiber laser under different laser cladding conditions, including variations in laser power, travel speed, and powder feed configurations. Image datasets acquired during deposition were used to train and evaluate the detection models under heterogeneous optical and operational conditions. Both architectures achieved robust spatial localization and reliable geometric estimation of the weld pool across multiple cladding scenarios. YOLOv5 achieved inference rates above 60 FPS while maintaining geometric accuracy comparable to SSD, demonstrating superior suitability for online real-time monitoring. The results show that object-detection approaches provide a computationally efficient alternative to segmentation-based methods for weld-pool monitoring, supporting future implementation in adaptive control and intelligent laser manufacturing systems.