基于联邦深度网络的设备跨域协同运维方法
Method of cross-domain collaborative operation and maintenance for equipment based on federated deep network
李威 1李健俊 1单宇翔 1李俊杰 2张微3
作者信息
- 1. 浙江中烟工业有限责任公司 信息中心,浙江 杭州 311500
- 2. 浙江大学工程师学院,浙江 杭州 310004;杭州安恒信息技术股份有限公司,浙江 杭州 310051
- 3. 杭州安恒信息技术股份有限公司,浙江 杭州 310051
- 折叠
摘要
设备的运维方法对于提高制造系统的安全性和可靠性至关重要,尽管现有的基于深度学习的方法在复杂设备的状态监测模型构建中得到了广泛应用,但是这类方法依赖于海量有标签的高质量数据.在实际运维过程中,不同设备间的数据隔离与数据孤岛问题阻碍了智能运维模型的协同建立.因此,文中提出了一种基于联邦深度网络的设备跨域协同运维方法,通过基于联邦学习的分布式训练及集中式聚合方法,实现了在数据安全前提下不同设备的跨域协同运维;在本地模型训练过程中,结合自注意力(Self-Attention)机制和双向长短期记忆(BiLSTM),一种基于Self-Attention-BiL-STM的特征提取方法被设计用于潜在特征的提取,能够提取时间序列数据的双向时间特性和注意力表达能力.此外,通过考虑本地训练模型的差异性,设计了一种自适应聚合策略以改善全局模型聚合的权重.最后,以某工业设备为例,进行故障诊断试验,结果表明:所提方法能够有效协同各个设备的模型,实现高效准确的设备跨域协同运维.
Abstract
The method of equipment maintenance is crucial to improve security and reliability of manufacturing systems.Cur-rently,although the methods based on deep learning are widely applied to construct the state monitoring models of complex equip-ment,they rely on a large amount of labeled high-quality data.In the actual operation and maintenance processes,such problems as data isolation and data silo between different equipment hinder the collaborative establishment of intelligent maintenance mod-els.Therefore,in this article a method of cross-domain collaborative operation and maintenance for equipment is proposed based on the federated deep network.With the help of federated learning,the methods of distributed training and centralized aggregation are used to achieve cross-domain collaborative operation and maintenance of different equipment,while data security is ensured.In the process of training local models,in combination with Self-Attention and Bidirectional Long Short-Term Memory,a Self-At-tention-BiLSTM method of feature extraction is designed for latent features,which can extract the time series data's bidirectional temporal characteristics as well as such abilities as attention and expression.Furthermore,with the diversity of training local mod-els taken into consideration,a self-adaptive aggregation strategy is designed,so as to improve the weight of global model aggrega-tion.Finally,with an industrial device as the example,a series of experiment are conducted on fault diagnosis.The results show that this method can effectively collaborate the models of different equipment,thus ensuring efficient and accurate cross-domain collaborative operation and maintenance.
关键词
信息安全/联邦学习/故障诊断/自适应Key words
information security/federated learning/fault diagnosis/self-adaptation引用本文复制引用
出版年
2024
NETL
NSTL
万方数据