摘要
齿轮箱作为机械传动系统的关键组成部分,面对有限标记数据和数据隐私保护的需求,传统故障诊断方法存在数据泄露和模型精度不高等问题.因此,本研究提出了一种基于联邦学习的隐私保护框架,采用半监督原型网络与对比学习相结合的小样本齿轮箱故障诊断方法.首先,构建了DeceFL联邦学习框架,在每个客户端上使用有限数量的标记样本构建正负样本对.同时,采用对比学习的预训练方法,为自动编码器提供了初始化参数.随后,将自动编码器作为原型网络的特征映射函数,使用有限数量的标记样本计算类别原型.最后,通过原型细化方法对原型进行微调,减少异常数据的干扰,获得更加稳定和准确的原型.经过实际齿轮箱数据的验证,结果表明,本研究提出的隐私保护框架下的半监督小样本故障诊断方法能够仅使用极少数量的样本即实现更为出色的故障识别精度.这一研究为应对现实工业应用中的数据挑战提供了创新性的方法,推动齿轮箱故障诊断领域的进一步研究.
Gearboxes,as a key component of mechanical drive systems,encounter problems of data leakage and low model accuracy when employing traditional fault diagnosis methods in terms of limited labeled data and data privacy protection requirements.Therefore,this study proposes a privacy-preserving framework based on federated learning using a small-sample gearbox fault diagnosis method that combines semi-supervised prototype networks with comparative learning.Initially,the DeceFL federated learning framework is constructed to produce positive and negative sample pairs using a limited number of labeled samples for each client.Meanwhile,the pretraining method of contrast learning is used to provide initialization parameters for the autoencoder.Subsequently,the autoencoder is used as a feature mapping function for the prototype network to compute the category prototype using limited labeled samples.Finally,category prototype is fine-tuned using the prototype refinement method to reduce the interference of abnormal data and obtain a more stable and accurate prototype.After validation with real gearbox data,the results show that the proposed semi-supervised small-sample fault diagnosis method can achieve better fault recognition accuracy with very few samples.This research provides an innovative approach to address data challenges in real industrial applications and promotes further research in the gearbox fault diagnosis field.
作者
何心
段亚穷
王子栋
张永
HE Xin;DUAN YaQiong;WANG ZiDong;ZHANG Yong(School of Mechanical Science and Engineering,Huazhong University of Science and Technology,Wuhan 430074,China;School of Information Science and Engineering,Wuhan University of Science and Technology,Wuhan 430081,China;Department of Computer Science,Brunel University London,London UB83PH,United Kingdom;Yuanshi Intelligent Technology Co.,Ltd,Nantong 226000,China)
出处
《中国科学：技术科学》
EI
CSCD
北大核心
2024年第6期1105-1119,共15页
Scientia Sinica(Technologica)
基金
国家自然科学基金项目(批准号:62273264)
广西重点研发计划项目(编号:桂科AB22035023)资助。
关键词
齿轮箱
故障诊断
隐私保护
小样本
半监督学习
gearbox
troubleshooting
privacy protection
small sample
semi-supervised learning
