异构信号图融合驱动的水电机组劣化状态评估  

Heterogeneous Signal Graph Fusion-driven Hydropower Unit Degradation State Assessment

作　　者：张峰源 李玉鑫 费一涛 刘颉[1,2] 李浩亮 袁晓辉[1,2] 张勇传[1,2] ZHANG Feng-yuana;LI Yu-xin;FEI Yi-tao;LIU Jie;LI Hao-liang;YUAN Xiao-hui;ZHANG Yong-chuan(School of Civil and Hydraulic Engineering,Huazhong University of Science and Technology,Wuhan 430074,China;Hubei Key Laboratory of Digital Watershed Science and Technology,Huazhong University of Science and Technology,Wuhan 430074,China;Dongfang Electric Machinery Co.,Ltd.,Deyang 618000,China)

机构地区：[1]华中科技大学土木与水利工程学院,湖北武汉430074 [2]华中科技大学数字流域科学与技术湖北省重点实验室,湖北武汉430074 [3]东方电气集团东方电机有限公司,四川德阳618000

出　　处：《水电能源科学》2025年第4期190-194,共5页Water Resources and Power

基　　金：国家自然科学基金项目(52205104,U2340211);湖北省自然科学基金项目(2022CFB062);武汉市知识创新专项项目(基础研究项目);中国长江电力股份有限公司资助项目(Z242302026)。

摘　　要：传统基于单一种类监测信号分析的水电机组劣化状态评估存在机组状态表征不全面、评估时序敏感性差等不足,需深入研究异构监测信号特征挖掘与状态融合表征问题,提出一种异构信号图融合表征驱动的水电机组劣化状态评估方法。首先,嵌入工况信息-多源监测信号为节点特征,设计基于异构信号相似度阈值的边连接函数;然后,计算异构信号内部多重边连接关系,将同一时段内数据转换为差异化图空间结构;其次,融合图卷积网络和循环神经网络,构建兼顾时-空特征提取能力的机组健康基准模型,挖掘异构信号图中隐含的时空依赖关系以表征机组状态;进一步引入注意力机制融合图表征向量,输出预测理论健康信号值,并度量多维信号空间内预测值与实际值的距离,以评估机组综合劣化程度;最后,使用某水电机组实测数据验证所提方法能有效综合评估机组劣化。The degradation assessment of hydropower units considering only a single type of monitoring signal has some limitations,such as the partial representation of unit states,and poor sensitivity of time series assessment.Thus,it is necessary to study heterogeneous monitoring signals feature fusion-based unit state representation.In this paper,a heterogeneous signal graph fusion-driven hydropower unit degradation state assessment method is proposed.Firstly,both the working condition parameters and multi-source monitoring signals are embedded as node features,and the edge connection function based on the similarity threshold of heterogeneous signals is designed.Then,the multi-edge connections within heterogeneous signals are calculated,the same period data are converted into a variety of heterogeneous signal graphs with differentiated structures.Secondly,the graph convolutional network and recurrent neural network are sequentially combined to construct a unit health benchmark model(HBM)that takes into account spatial-temporal feature extraction ability.The latent dependence relationships in heterogeneous signal graphs are excavated by HBM,and output the graph representation vectors of unit state.Then,the attention mechanism is further introduced for graph vector fusion to realize the time sequence state representation.The distance between the predicted value and the actual value in the multidimensional signal space is measured to evaluate the comprehensive deterioration degree of the unit.Finally,the proposed method can effectively evaluate the deterioration of the unit comprehensively by using the measured data of a hydropower unit.

关 键 词：水电机组 劣化状态评估 多源异构信号 特征融合 图卷积网络 

分 类 号：TV734.21[水利工程—水利水电工程]