Scholarlink

The growing demand for privacy-preserving, cyber-secure, and decentralized analytics in manufacturing and industrial environments has positioned Federated Learning (FL) as a powerful solution for collaborative model training without compromising data confidentiality and overcoming challenges of limited local data availability. This study investigates the FL approach for Time-Series Forecasting (TSF), comparing six deep learning architectures, including MLP, PatchTST, Informer, N-BEATS, DLinear, and LSTM models. The research implements a distributed learning framework, enabling collaborative model training while maintaining data privacy across different clients. Results demonstrate varying performance across different model architectures. PatchTST consistently achieved the best short-term performance, with a MSE as low as 0.00013 and MAPE of 3.11% on ETTm1 dataset. N-BEATS and DLinear also demonstrated strong and stable performance, especially for longer horizons. This work contributes to the field by comparing different deep learning architectures in a FL context for TSF and establishing practical guidelines and proof of the concept for implementing such systems in industrial settings. The findings have significant implications for applications in manufacturing, particularly in areas such as predictive maintenance, quality control, and process optimization.

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

Federated Learning for Smart Manufacturing: Evaluating Deep Learning Architectures for Time Series Forecasting in a Collaborative Framework

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Md Irfan Khan,
  • Mahady Rayhan,
  • Mojtaba A. Farahani,
  • Thorsten Wuest

摘要

The growing demand for privacy-preserving, cyber-secure, and decentralized analytics in manufacturing and industrial environments has positioned Federated Learning (FL) as a powerful solution for collaborative model training without compromising data confidentiality and overcoming challenges of limited local data availability. This study investigates the FL approach for Time-Series Forecasting (TSF), comparing six deep learning architectures, including MLP, PatchTST, Informer, N-BEATS, DLinear, and LSTM models. The research implements a distributed learning framework, enabling collaborative model training while maintaining data privacy across different clients. Results demonstrate varying performance across different model architectures. PatchTST consistently achieved the best short-term performance, with a MSE as low as 0.00013 and MAPE of 3.11% on ETTm1 dataset. N-BEATS and DLinear also demonstrated strong and stable performance, especially for longer horizons. This work contributes to the field by comparing different deep learning architectures in a FL context for TSF and establishing practical guidelines and proof of the concept for implementing such systems in industrial settings. The findings have significant implications for applications in manufacturing, particularly in areas such as predictive maintenance, quality control, and process optimization.