摘要
工业物联网助推机械故障诊断步入大数据时代,然而因各节点需要共享本地的私有数据而造成隐私泄露是当前工业物联网亟需解决的问题。联邦学习有望应用于工业物联网以实现在私有数据不离开本地存储的前提下,协同各节点训练诊断模型。然而,联邦学习面临着以下诸多挑战。首先,联邦学习的中心化架构极易引发单点故障。其次,工业物联网中各节点的故障数据通常是非独立同分布的,以致联邦学习难以收敛。再次,联邦学习缺乏防御手段来阻止恶意节点的攻击。最后,联邦学习需要激励机制来鼓励节点分享资源。针对这些挑战,提出了一种区块链和边缘计算赋能的联邦学习故障诊断框架,采用去中心化的模式保障工业物联网中机械设备故障数据的隐私和安全。在此框架中,构造了一种特征对比损失函数来解决非独立同分布问题,设计了一种拜占庭容错的评分机制来抵抗投毒攻击,并开发了一种基于信誉的激励算法来评估应给予节点的奖励。所提方法被应用于工业物联网中风力发电机的行星齿轮箱故障诊断模拟场景,在私有本地数据不泄露的前提下,展现出最优的综合性能。
The industrial Internet of Things(IIoT)promotes mechanical fault diagnosis into the era of big data,however,privacy leakage caused by the need to share local private data among IIoT nodes is an urgent problem to be solved.Federated learning(FL)is expected to be applied to IIoT,which enables nodes to collaboratively train diagnostic models without making private data leave local storage.However,there are many challenges faced by the FL.Firstly,the centralized architecture of FL is highly susceptible to single point of failure.Moreover,the fault data of nodes in the IIoT are usually not independent and identically distributed(non-IID),which makes it difficult for the FL to converge.In addition,the FL lacks defense measures to prevent attacks conducted by malicious nodes.Finally,the FL needs incentive mechanisms to encourage nodes to share resources.Aiming at the challenges introduced above,a blockchain and edge computing enabled FL fault diagnosis framework is proposed,which adopts a decentralized mode to ensure the privacy and security of mechanical equipment fault data in the IIoT.In the proposed framework,a feature-contrastive loss function is constructed to address the non-IID problem.A Byzantine-tolerance scoring mechanism is designed to resist poisonous attacks.A reputation-based incentive algorithm is developed to evaluate the rewards owed to nodes.The proposed method is applied to a simulation scenario of planetary gearbox fault diagnosis for wind turbines in the IIoT,demonstrating its optimal overall performance without the disclosure of local private data.
作者
邵海东
肖一鸣
闵志闪
韩淞宇
张海舟
SHAO Haidong;XIAO Yiming;MIN Zhishan;HAN Songyu;ZHANG Haizhou(College of Mechanical and Vehicle Engineering,Hunan University,Changsha 410082;Nanjing Research Institute of Electronics Technology,Nanjing 210039)
出处
《机械工程学报》
EI
CSCD
北大核心
2023年第21期283-292,共10页
Journal of Mechanical Engineering
基金
国家自然科学基金(52275104)
湖南省自然科学基金优秀青年科学基金(2021JJ20017)资助项目。
关键词
区块链
边缘计算
故障诊断
联邦学习
工业物联网
blockchain
edge computing
fault diagnosis
federated learning
industrial Internet of Things
