摘要
文章针对工业场景中多参与方数据隐私保护与跨工况故障诊断模型泛化能力不足的问题,提出一种基于数据多样性增强的联邦故障诊断方法。该方法在各参与方本地训练阶段筛选高可信特征并上传至中央服务器,通过融合多源特征生成多样化虚拟特征后下发至各参与方用于扩展训练,从而在不共享原始数据的前提下增强模型的跨工况泛化能力。在滚动轴承数据集上的实验结果表明,Fed-DDE在多个诊断任务中平均准确率最高,部分任务超过90%,验证了其在保护数据隐私的同时提升故障诊断性能的有效性。
The paper addresses the issues of insufficient data privacy protection and limited generalization ability of cross-condition fault diagnosis models in industrial scenarios.It proposes a federated fault diagnosis method based on data diversity enhancement.During the local training phase,this method selects highly credible features from each participant and uploads them to the central server.By fusing multiple source features to generate diverse virtual features,these are then distributed to each participant for extended training.This approach enhances the model's generalization ability across different conditions without sharing original data.Experimental results on a rolling bearing dataset show that Fed-DDE achieves the highest average accuracy across multiple diagnostic tasks,with some tasks exceeding 90%.This verifies its effectiveness in enhancing fault diagnosis performance while protecting data privacy.
作者
蒋欣军
王宝华
吴国兴
殷戈
邓艾东
Jiang Xinjun;Wang Baohua;Wu Guoxing;Yin Ge;Deng Aidong(CHN ENERGY ChangZhou Second Electric Power CO.,Ltd,ChangZhou 213002 China;CHN ENERGY Science and Technology Research Institute Co.,Ltd,Nanjing 210046 China;National Engineering Research Center of Power Generation Control and Safety,Nanjing 211102 China)
出处
《信息化研究》
2025年第6期16-20,共5页
INFORMATIZATION RESEARCH
基金
国家能源集团项目GJNY-23-68:1000MW机组高效灵活低碳智能发电技术集成创新示范。
关键词
故障诊断
联邦学习
数据增强
未知工况
Fault diagnosis
Federated learning
Data enhancement
unknown condition
