| 张洁,李继庚.基于XGBoost与联邦迁移学习融合的造纸水分异常预测研究[J].中国造纸,2026,45(3):180-189 |
二维码（扫一下试试看！） |
| 基于XGBoost与联邦迁移学习融合的造纸水分异常预测研究 |
| Study on Paper Moisture Anomaly Prediction Based on XGBoost and Federated Transfer Learning Integration |
| 收稿日期：2025-09-06 修订日期：2025-10-09 |
| DOI：10.11980/j.issn.0254-508X.2026.03.021 |
| 关键词: 水分异常预测 联邦学习 迁移学习 XGBoost 领域对抗神经网络 |
| Key Words:water anomaly prediction federated learning transfer learning XGBoost domain adversarial neural network |
| 基金项目: |
|
| 摘要点击次数: 218 |
| 全文下载次数: 141 |
| 摘要:在造纸工业智能化进程中，纸张水分含量的精准预测对质量控制和能效优化至关重要。传统方法因单个企业样本稀缺且模式单一，模型泛化能力有限，同时难以满足数据隐私保护需求。为此，本研究提出了一种基于XGBoost的联邦学习框架，通过分布式架构保障数据隐私，并引入领域对抗神经网络进行跨域特征对齐，以提升泛化能力。经2家造纸厂数据验证，该全局模型准确率达91.74%，各项指标均显著优于本地模型及其他对比算法，与集中式XGBoost性能差距不足1%。特征对齐后构建的2种特征组合虽性能略有降低（≤1.5%），但验证了跨域稳定性和协同增益。 |
| Abstract:In the intelligentization of the paper industry， accurate prediction of paper moisture content is crucial for quality control and energy efficiency optimization. Traditional methods suffer from limited model generalizability due to scarce and homogeneous single-enterprise samples， while also failing to meet data privacy protection requirements. To address these issues， this paper proposed a federated learning framework based on XGBoost， which ensured data privacy through a distributed architecture and incorporated a Domain-Adversarial Neural Network for cross-domain feature alignment to enhance generalization capability. Validated with data from two paper mills， the global model achieved an accuracy of 91.74%， with all performance metrics significantly outperforming local models and other comparative algorithms. The performance gap with centralized XGBoost was less than 1%. Although the two feature combinations constructed after feature alignment showed a slight performance decrease (≤1.5%)， they demonstrated cross-domain stability and collaborative benefits. |
| 查看全文
HTML
查看/发表评论 下载PDF阅读器 |