【作者】 陈曦；

【导师】 严如强；

【作者基本信息】 东南大学 ， 仪器科学与技术， 2024， 博士

【摘要】 滚动轴承作为旋转机械的关键精密组件,其安全稳定运行对提升工业生产效率、降低维护成本、优化工作环境意义重大。因此,监测滚动轴承运行状况并预测其剩余使用寿命已成为设备运维领域的一项关键研究。在工业大数据与深度学习技术蓬勃发展的背景下,轴承剩余寿命预测研究迎来了前所未有的机遇与挑战。深度神经网络凭借强大的特征学习能力,可从海量数据中自动提取轴承性能退化特征,构建高精度预测模型,引领该领域技术革新。然而,随着数据价值不断凸显,工业数据隐私保护问题愈发关键。敏感数据的收集与处理,包括设备运行状态、潜在故障预警以及动态性能变化等,不仅涉及企业的核心信息,还关乎用户隐私安全,因此如何在保障数据隐私的同时推进滚动轴承剩余寿命预测方法的发展成为亟待解决的问题。联邦学习作为一种分布式深度学习新范式,能使不同数据持有者在不共享原始数据的前提下协同训练预测模型,为隐私保护下的滚动轴承剩余寿命预测开辟新路径。本文以非同工况运行下的滚动轴承为研究对象,紧密结合实际工业应用需求,深入探索并攻克在联邦学习框架下搭建剩余寿命预测模型所面临的异构信息难融合、孤岛资源难复用、客户端关系易忽视以及动态数据流建模遇瓶颈等一系列技术难题,为参与协同建模的非同工况轴承搭建个性化剩余寿命预测模型,并探索了针对新工况滚动轴承的高效建模策略,为遵从工业数据隐私保护的运维管理提供了技术支撑。本文的主要研究工作如下:(1)针对非同工况滚动轴承全寿命数据稀缺且退化模式多样情况下,直接运用联邦学习框架融合不同工况轴承监测信息、构建全局预测模型所面临的异构难题,分别从聚合过程及聚合后模型的深度优化两个方向出发,提出基于多客户端模型的软聚合方法与基于全局和本地模型的自适应聚合方法。前者通过模型插值的显式聚合技术,实现了各客户端模型间信息的有效融合与交流;后者则在模型元素层面进行精细化操作,促进了退化信息的深层次整合。两方法均可生成针对特定工况轴承的个性化剩余寿命预测模型。实验验证了所提两种个性化聚合方法在促进非同工况监测数据间有效信息流通、提升各客户端预测模型表现方面均具有积极作用。(2)针对现有客户端协同构建的模型缺乏可复用性的问题,提出了一种基于全局共享退化特征的滚动轴承剩余寿命预测方法。借助表示学习理念,将模型整体的训练过程拆分为全局共享退化特征提取模块与个性化特征提取模块的两步式训练过程。前端模块旨在捕获参与训练的所有客户端数据中的普遍性退化特征,后端模块则在全局共享特征的基础上,深度融合各客户端特有数据特征,以提取更具针对性的个性化退化表征。实验验证了该方法可高效整合不同工况的状态监测数据资源,有效捕获普适性的全局共享退化特征,并强化各客户端个性化模型的预测精度。此外,所提全局共享退化特征具有可扩展性。在面对新工况下滚动轴承的剩余寿命预测任务时,仅需以全局共享退化特征为基础,结合新工况的状态监测数据对本地模型进行部分更新,即可实现高效及精准建模。(3)针对非同工况下轴承联合建模时全局共享退化特征在聚合过程中出现的漂移问题,以及动态数据流场景下滚动轴承剩余寿命预测模型及其退化特征强化手段不成熟的现状,提出基于客户端相似度划分与持续学习的滚动轴承剩余寿命预测方法。通过计算各客户端初始模型特定层输出的余弦相似度,评估不同工况下轴承所在客户端的相关性,为每个客户端筛选出高度相似的伙伴客户端集合。进一步在这些群组内实施模型聚合,可降低非相关客户端对个性化建模的不利影响。联邦持续学习策略则能促使各参与客户端在不直接访问历史数据前提下,利用实时采集的监测数据流持续优化本地模型,并不断丰富共享退化特征。实验结果验证了所提方法显著促进了相似工况客户端间的知识交流,提高了既有客户端预测模型的构建效率和准确性。同时,强化后的共享退化特征在拓展客户端建模任务中能提供更可靠、高效的建模支持。(4)针对当前联邦学习框架在数据安全领域所面临的严峻挑战,尤其是通过权重传递可能导致的隐私泄露隐患,提出了一个完整的融合安全机制的滚动轴承剩余寿命预测方法框架。将Paillier同态加法加密算法作为可插拔组件融入所提出的上述建模方法中,在不损害现有方法所达成的预测精度的前提下,确保了各客户端本地模型更新的独立性以及中心服务器聚合结果的有效性,为构建强隐私保护下的非同工况滚动轴承剩余寿命预测模型提供了更为坚实和可信赖的技术支持。本文紧扣实际工业场景需求,在保护各客户端工业状态监测数据隐私的前提下,从既有客户端间的协同优化,到新增客户端的高效建模,再到动态数据流下模型与特征的强化,形成了一套符合数据使用合规性且充分考量工业应用特性的全面建模方案,从而实现非同工况下滚动轴承个性化剩余寿命预测。更多还原

【Abstract】 Rolling bearings are critical components in rotating machinery,and their safe and stable operation is essential for enhancing industrial production efficiency,reducing maintenance costs,and optimizing work environments.Therefore,monitoring the operating condition of rolling bearings and predicting their remaining useful life(RUL)has become key research in equipment maintenance.The rapid development of industrial big data and deep learning has brought opportunities and challenges to bearing RUL prediction.Deep neural networks,with their powerful feature learning capabilities,can automatically extract degradation characteristics from massive datasets to construct high-precision prediction models,driving technological innovation in this field.However,as the value of data becomes increasingly prominent,protecting industrial data privacy has become crucial.The collection and processing of sensitive data,including equipment operating conditions,potential fault warnings,and dynamic performance changes,involve core enterprise information and user privacy.Therefore,how to advance the development of rolling bearing RUL prediction methods while ensuring data privacy has become an urgent issue to address.Federated learning,a new distributed deep learning paradigm,enables different data holders to collaboratively train prediction models without sharing raw data,opening new avenues for privacy-preserving RUL prediction.This study focuses on rolling bearings operating under non-identical conditions,closely aligning with practical industrial needs.It explores and overcomes technical challenges in building RUL prediction models within the federated learning framework,including difficulties in integrating heterogeneous information,reusing isolated resources,accounting for client relationships,and modeling dynamic data streams.The research constructs personalized RUL prediction models for bearings under non-identical working conditions and explores efficient modeling strategies for new operating scenarios.The main works are as follows:(1)To address the challenge of data heterogeneity when using the federated learning framework to fuse monitoring information from bearings under different operating conditions,particularly in scenarios where full-life data is scarce and degradation patterns are diverse,this paper proposes two methods.The first is a personalized soft aggregation method,which uses explicit model interpolation techniques to facilitate effective information exchange between client models.The second is an adaptive local aggregation method that operates at the model element level,enabling deeper integration of degradation information.Both methods can generate personalized RUL prediction models tailored to specific operating conditions.Experimental results validate their efficacy in improving model performance and facilitating client information communication.(2)To address the lack of reusability of the models collaboratively constructed by clients,a RUL prediction method for rolling bearings is proposed based on global shared degradation representation(GSDR)in federated learning.Inspired by representation learning,the overall model training process is divided into a two-step procedure:a global shared degradation feature extraction module and a personalized feature extraction module.The front-end module aims to capture the universal degradation features from the data of all participating clients,while the back-end module integrates the specific data features of each client based on the GSDR to extract personalized degradation representations.Experiments validate that this method can effectively integrate the data resources of different operating conditions,capture the universally shared degradation features,and enhance the prediction accuracy of individual client models.Furthermore,the GSDR exhibits scalability.When facing the RUL prediction task of rolling bearings under new operating conditions,it is only necessary to update the local model based on the extracted GSDR with the state monitoring data of the new condition,to achieve efficient and accurate modeling.(3)To address the GSDR drift problem during the aggregation process,as well as the immaturity of model and degradation feature enhancement strategies in dynamic data stream scenarios,a method for RUL prediction based on client similarity and federated continual learning is proposed.By calculating the cosine similarity of the outputs in specific layers of the initial models at each client,the relevance of the bearings belonging to different clients under different operating conditions is evaluated,and a highly similar partner client set is selected for each client.Further model aggregation within these groups can reduce the adverse impact of irrelevant clients on personalized modeling.The federated continual learning strategy can enable each participating client to continuously optimize the local model and enrich the shared degradation features using the real-time monitoring data stream,without directly accessing historical data.Experimental results verify that the proposed method significantly promotes knowledge communication between clients under close operating conditions,and improves the construction efficiency and accuracy of the prediction model.Meanwhile,the enhanced GSDR can provide more reliable and efficient modeling support for expanding client modeling tasks.(4)To address the data security challenges in current federated learning methods,especially the risk of privacy leakage through weight transmission,a comprehensive RUL prediction framework for rolling bearings incorporating security mechanisms is proposed.The Paillier homomorphic additive encryption algorithm is integrated as a pluggable component into the aforementioned modeling methods,ensuring the independence of local model updates on each client and the validity of the aggregation results on the central server,without compromising predictive accuracy.This provides more solid and reliable technical support for constructing a RUL prediction model for rolling bearings under non-identical operating conditions.This research focuses on practical industrial needs,providing a comprehensive modeling solution compliant with data use regulations and fully considering industrial application characteristics,thereby achieving personalized RUL prediction for rolling bearings under non-identical operating conditions while safeguarding industrial monitoring data privacy.更多还原