Scholarlink

<p>In manufacturing domains such as the aeronautical sector, where operational data can be sensitive, federated learning provides an architectural approach for collaborative model training across multiple production sites while keeping raw data local. This paper presents a comparative analysis of three widely used frameworks for Federated Learning: Flower, TensorFlow Federated (TFF) and PySyft. The evaluation focuses on deployment complexity and performance within a real-world industrial use case involving two large-scale machining centers dedicated to aeronautical component manufacturing, in particular to predict their power consumption. Key performance metrics, including loss, accuracy and F-score, alongside execution time and computational resource usage, are evaluated to capture the practical trade-offs of each framework under realistic deployment conditions. Results show that TFF achieves slightly higher predictive performance in terms of accuracy and F-score, while Flower and PySyft exhibit very similar model behavior. At the system level, Flower demonstrates lower training time in distributed settings compared to PySyft. In terms of usability, TFF simplifies single-machine deployment, while Flower better balances scalability in distributed settings. Based on these findings, the study provides practical guidance for selecting appropriate federated learning frameworks in industrial settings where data-locality requirements, heterogeneity, and infrastructure constraints are critical factors.</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

Comparison of federated learning frameworks for discrete event manufacturing systems

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月10日

查看全文
  • Endika Tapia,
  • Unai Lopez-Novoa,
  • Leonardo Sastoque-Pinilla,
  • Luis Norberto López-de-Lacalle

摘要

In manufacturing domains such as the aeronautical sector, where operational data can be sensitive, federated learning provides an architectural approach for collaborative model training across multiple production sites while keeping raw data local. This paper presents a comparative analysis of three widely used frameworks for Federated Learning: Flower, TensorFlow Federated (TFF) and PySyft. The evaluation focuses on deployment complexity and performance within a real-world industrial use case involving two large-scale machining centers dedicated to aeronautical component manufacturing, in particular to predict their power consumption. Key performance metrics, including loss, accuracy and F-score, alongside execution time and computational resource usage, are evaluated to capture the practical trade-offs of each framework under realistic deployment conditions. Results show that TFF achieves slightly higher predictive performance in terms of accuracy and F-score, while Flower and PySyft exhibit very similar model behavior. At the system level, Flower demonstrates lower training time in distributed settings compared to PySyft. In terms of usability, TFF simplifies single-machine deployment, while Flower better balances scalability in distributed settings. Based on these findings, the study provides practical guidance for selecting appropriate federated learning frameworks in industrial settings where data-locality requirements, heterogeneity, and infrastructure constraints are critical factors.