Scholarlink

<p>Laser powder bed fusion (LPBF) is a leading metal additive manufacturing (AM) technology widely used in aerospace, medical surveillance, and other critical fields that require high privacy standards. However, the complex forming process of LPBF often results in inevitable defects, limiting the stability and consistency of production quality. With advancements in artificial intelligence (AI), deep learning (DL) methods have significantly improved industrial monitoring, garnering attention for intelligent defect detection in LPBF. Nevertheless, applying DL-based defect detection across manufacturers with privacy requirements faces several challenges: (1) imbalance and heterogeneity of powder bed image defect data hinder the acquisition of high-quality datasets; (2) data sharing across different printers is impeded by privacy concerns, complicating cross-supplier collaboration; (3) solely using differential privacy (DP) methods struggles to balance model performance with privacy protection. To overcome these challenges, this study proposed a federated learning-based adaptive differential privacy (ADP-FL) framework that enables high-precision powder-spreading defect detection in LPBF monitoring while ensuring data privacy. Besides, we constructed a comprehensive dataset and validated robustness through transfer learning across different printers and materials. Furthermore, we simulated gradient inversion attacks to demonstrate that DP enhances data security. Extensive experimental results validate that our proposed approach significantly enhances the privacy-preserving capabilities of federated learning (FL) while maintaining high performance. Notably, ADP-FL effectively resolves the trade-off between data protection and model efficacy. For instance, ADP-FL improves the mean Intersection over Union (mIoU) by 2.7% compared to the standard DP approach.</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

A privacy-enhancing federated learning framework for cross-manufacturer LPBF powder bed defect identification

  • 期刊论文

  • 发表日期：

    2026年2月13日

查看全文
  • Yibiao Wang,
  • Jiafeng Tang,
  • Zhibin Zhao,
  • Chenxi Wang,
  • Xingwu Zhang,
  • Xuefeng Chen

摘要

Laser powder bed fusion (LPBF) is a leading metal additive manufacturing (AM) technology widely used in aerospace, medical surveillance, and other critical fields that require high privacy standards. However, the complex forming process of LPBF often results in inevitable defects, limiting the stability and consistency of production quality. With advancements in artificial intelligence (AI), deep learning (DL) methods have significantly improved industrial monitoring, garnering attention for intelligent defect detection in LPBF. Nevertheless, applying DL-based defect detection across manufacturers with privacy requirements faces several challenges: (1) imbalance and heterogeneity of powder bed image defect data hinder the acquisition of high-quality datasets; (2) data sharing across different printers is impeded by privacy concerns, complicating cross-supplier collaboration; (3) solely using differential privacy (DP) methods struggles to balance model performance with privacy protection. To overcome these challenges, this study proposed a federated learning-based adaptive differential privacy (ADP-FL) framework that enables high-precision powder-spreading defect detection in LPBF monitoring while ensuring data privacy. Besides, we constructed a comprehensive dataset and validated robustness through transfer learning across different printers and materials. Furthermore, we simulated gradient inversion attacks to demonstrate that DP enhances data security. Extensive experimental results validate that our proposed approach significantly enhances the privacy-preserving capabilities of federated learning (FL) while maintaining high performance. Notably, ADP-FL effectively resolves the trade-off between data protection and model efficacy. For instance, ADP-FL improves the mean Intersection over Union (mIoU) by 2.7% compared to the standard DP approach.