人-信息-物理协同下核电设备可演进式剩余寿命估计  

Evolvable Remaining Useful Life Estimation of Nuclear Power Equipment Under Human-cyber-physical Collaboration

作　　者：蒋翔宇 冯毅雄[1,2] 张志峰 宋秀菊[1] 洪兆溪 胡炳涛 谭建荣 JIANG Xiangyu;FENG Yixiong;ZHANG Zhifeng;SONG Xiuju;HONG Zhaoxi;HU Bingtao;TAN Jianrong(State Key Laboratory of Fluid Power and Mechatronic Systems,Zhejiang University,Hangzhou 310027;State Key Laboratory of Public Big Data,Guizhou University,Guiyang 550025;Ningbo Innovation Center,Zhejiang University,Ningbo 315100)

机构地区：[1]浙江大学流体动力基础件与机电系统全国重点实验室,杭州310027 [2]贵州大学省部共建公共大数据国家重点实验室,贵阳550025 [3]浙江大学宁波科创中心,宁波315100

出　　处：《机械工程学报》2025年第4期302-313,共12页Journal of Mechanical Engineering

基　　金：国家重点研发计划(2022YFB3402000);浙江省重点研发计划(2024C01207);国家自然科学基金(52105281,52205288)资助项目。

摘　　要：在工业5.0新时代,将设备物理实体与人类认知、信息技术融合,是推动故障预测与健康管理(Prognosticsandhealth management,PHM)智能增强的重要途径,剩余寿命估计作为PHM中的关键环节,为设备预测性维护提供时间裕量依据。受人的记忆和遗忘机制启发,提出一种人-信息-物理协同的剩余寿命估计可演进式模型框架,能够同时预测设备的连续状态和离散状态。模型采用基于实例的学习策略从时序数据中汲取并保留知识,无须事先假设固定的失效阈值和退化阶段,旨在随着样本的加入而逐渐调整趋于稳定,适合于设备小样本和先验知识不足的情况,或设备处于变化的运行环境中。模型底层采用核最小均方算法,使模型结构和参数可以在线更新;基于最近实例质心估计方法改进算法,将输入空间拓展为特征空间,随着输入数据的学习将空间自动划分为不同子区域,以实现在预测未来信号的同时获取健康状态信息,进而推导剩余寿命。为了使模型网络规模更加紧凑,引入在线矢量量化方法,通过消除模型冗余的基函数降低计算复杂度。将所提模型框架运用于某压水堆给水泵剩余寿命估计,验证了方法的有效性。In the new era of Industry 5.0,the integration of physical entities of equipment with human cognition and information technology is an important way to promote the intelligent of prognostics and health management(PHM).The estimation of residual life is a key link in PHM.It is the basis for the predictive maintenance of equipment.Inspired by the human memory and forgetting mechanism,an evolvable model framework of human-cyber-physical collaboration for remaining useful life estimation is proposed,which can simultaneously predict the continuous state and discrete state of the equipment.The model adopts instance-based learning to absorb and retain knowledge from time series data without assuming a fixed failure threshold and degradation stage in advance.The model aims to gradually adjust and become stable with incremental samples,which is suitable for the equipment with small-scale samples and insufficient prior knowledge,or the equipment in a changing operating environment.The kernel least mean square(KLMS)algorithm is used adapted as the underlying algorithm so that the model structure and parameters could be updated online.After KLMS expands the input space into a feature space,the nearest-instance centroid estimation(NICE)is used to improve it by automatically dividing the feature space into different sub-regions with the input data learning,to realize the prediction of future signals and obtain health status information at the same time.Then the remaining useful life is derived.In order to make the model network more compact,the online modified vector quantization(M-VQ)method is introduced to reduce the computational complexity by eliminating redundant basis functions of the model.The proposed model framework is applied to estimate the remaining life of the feed pump in a pressurized water reactor,and the effectiveness of the method is verified.

关 键 词：人-信息-物理协同 剩余寿命 同步预测 核最小方均 

分 类 号：TG156[金属学及工艺—热处理]