Scholarlink

<p>Powder bed uniformity is a critical determinant of process stability and part quality in laser powder bed fusion (LPBF) additive manufacturing. Irregularities in the powder layer can accumulate during the layer-by-layer melting process, leading to defect formation that compromises structural integrity and may ultimately cause process failure. Thus, real-time detection of both powder bed non-uniformity and resulting workpiece defects is essential. However, conventional imaging systems often fall short in providing high-resolution, real-time monitoring over large and complex build areas due to limited spatial resolution and insufficient sensitivity to subtle features. To overcome these limitations, we introduce a high-resolution online monitoring system that integrates a contact image sensor (CIS) into the recoater assembly, coupled with a YOLOv9-based defect detection model augmented by a wavelet feature fusion (WFF) module. The CIS performs line-scan imaging synchronized with the recoater motion, delivering uniform, distortion-free images with exceptional spatial consistency. This setup enables the capture of fine-scale defects that were previously undetectable—especially contamination-induced anomalies such as mixed powders, oxidized agglomerates, and foreign inclusions—which are typically characterized by faint boundaries, low textural contrast, and complex morphologies. The proposed YOLOv9-WFF model employs multiscale frequency-domain decomposition to enrich low-frequency feature representation, allowing efficient and precise classification and localization of nine distinct types of powder bed defects. Experimental results show that the model achieves superior detection performance in terms of accuracy, recall, and mAP, while maintaining a lightweight architecture with reduced computational overhead. Furthermore, the CIS-based system exhibits strong robustness against ambient lighting variations and supports straightforward integration and scalability across various machine platforms. Its compact and modular design makes it highly suitable for wide field-of-view, high-definition industrial powder bed monitoring, providing a transferable and extensible solution for next-generation quality assurance in LPBF.</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

High-precision online monitoring of powder bed defects of additive manufacturing using deep learning with wavelet feature fusion

  • 期刊论文

  • 发表日期：

    2026年3月31日

查看全文
  • Wenhao Tian,
  • Pinghua Yang,
  • Guangping Guo,
  • Linda Ke,
  • Quan Wei,
  • Fang Dong,
  • Sheng Liu,
  • Chen Zhang

摘要

Powder bed uniformity is a critical determinant of process stability and part quality in laser powder bed fusion (LPBF) additive manufacturing. Irregularities in the powder layer can accumulate during the layer-by-layer melting process, leading to defect formation that compromises structural integrity and may ultimately cause process failure. Thus, real-time detection of both powder bed non-uniformity and resulting workpiece defects is essential. However, conventional imaging systems often fall short in providing high-resolution, real-time monitoring over large and complex build areas due to limited spatial resolution and insufficient sensitivity to subtle features. To overcome these limitations, we introduce a high-resolution online monitoring system that integrates a contact image sensor (CIS) into the recoater assembly, coupled with a YOLOv9-based defect detection model augmented by a wavelet feature fusion (WFF) module. The CIS performs line-scan imaging synchronized with the recoater motion, delivering uniform, distortion-free images with exceptional spatial consistency. This setup enables the capture of fine-scale defects that were previously undetectable—especially contamination-induced anomalies such as mixed powders, oxidized agglomerates, and foreign inclusions—which are typically characterized by faint boundaries, low textural contrast, and complex morphologies. The proposed YOLOv9-WFF model employs multiscale frequency-domain decomposition to enrich low-frequency feature representation, allowing efficient and precise classification and localization of nine distinct types of powder bed defects. Experimental results show that the model achieves superior detection performance in terms of accuracy, recall, and mAP, while maintaining a lightweight architecture with reduced computational overhead. Furthermore, the CIS-based system exhibits strong robustness against ambient lighting variations and supports straightforward integration and scalability across various machine platforms. Its compact and modular design makes it highly suitable for wide field-of-view, high-definition industrial powder bed monitoring, providing a transferable and extensible solution for next-generation quality assurance in LPBF.