融合深度学习的水电机组故障知识图谱构建与不确定推理诊断

This project will propose deep learning and knowledge graph integrated fault diagnosis theoretic hypothesis for hydropower generating unit (HGU), in which fault knowledge graph of HGU based on deep learning of isomeric data and multi-source knowledge will be built to fulfill the intelligent diagnosing of HGU faults. On aspect of knowledge learning of structured data, deep learning models for status monitoring data based on gradient optimization and error back-propagation will be built. In order to solve problems of diffusion of gradients, calculation complexity and generalization in model training, an unsupervised-supervised multi-phase training frame will be designed. As for knowledge learning of unstructured data, deep learning models based on conditional random fields method for text learning of testing repots will be proposed. For handling the problem of small sample learning for the models, the transfer learning method will be developed. With the multi-source fault knowledge learned from structured and unstructured data, the uniform distributed expression model for multi-source knowledge will be studied. And then, the multi-level rule enhanced fault knowledge graph based on prior knowledge learned from text and posterior knowledge learned from monitoring data will be built. Based on the knowledge graph, the naive Bayes reasoning network will be deduced to realize the uncertain reasoning of the knowledge graph for HGU faults. This project provides a novel style for fault diagnosis of HGU and the results will be valuable for popularization and application, because it will establish the novel intelligent fault diagnosis theory and methods with knowledge graph and promote the development of intelligence of hydropower generation.

本项目首次提出融合深度学习和知识图谱的水电机组智能故障诊断理论设想，建立以异构数据深度学习为基础、多源知识驱动的水电机组故障知识图谱，实现水电机组故障的智能诊断。在结构化数据知识学习方面，提出基于梯度优化和误差反馈的状态监测数据故障特征深度学习模型，设计无监督-有监督分阶段训练框架，解决模型的训练梯度弥散、计算复杂度和泛化问题；针对非结构化数据知识学习，建立基于随机条件场算法的水电机组文本故障知识深度学习模型，研究迁移学习理论，解决模型的小样本训练难题；以结构和非结构数据提取的故障知识为基础，提出多源知识的统一分布式表示范式，建立以文本知识先验、监测数据知识后验的多层次规则增强型多源故障知识图谱；在此基础上，推导朴素贝叶斯推理网络，实现故障知识图谱不确定推理诊断。本项目颠覆传统故障诊断模式，成果将创建以知识图谱为核心的新型智能诊断理论与方法体系，促进智慧水电发展，具有重大推广应用价值。

随着电力负荷的迅速增长、间歇性能源的大规模接入，水电站以其调峰填谷的独特运行特性，发挥着调节负荷、促进电力系统节能和维护电网安全稳定的功能，逐步成为我国电力系统有效的、不可或缺的调节工具。水电站在电力系统中担负着调峰填谷、调频调相、旋转备用等多重任务，开机、停机频繁，导致异常振动、结构疲劳、运行方式破坏等各种故障与事故出现机率增加。为此，本项目以水电机组故障推理诊断为主要研究目标，采用理论研究、实验研究与工程应用相结合的技术路线，利用多学科交叉融合的优势，从水电机组状态数据清洗与标记、水电机组状态数据故障特征提取、水电机组文本故障特征信息提取和多源知识图谱构建与诊断方法等四方面开展研究，构建了水电机组信息提取与多源知识图谱构建为核心的水电机组故障推理诊断方法与应用体系。具体表现在：（1）针对水电机组数据清洗难题，引入K均值算法，计算正常数据的特征值，实现异常数据点的检测与处理，针对水电机组状态数据标记难题，提出半监督动态“伪标签”学习模型；（2）提出用瞬时轴心轨迹来描述旋转机械轴系振动的瞬时状态，推求了瞬时轴心轨迹特征指标计算方法，提出了一种多元复信号变分模态分解算法，推导了多支承面轴心轨迹瞬时特征提取方法，进一步提出了转子振动三维瞬时轴心轨迹图，建立基于梯度优化和误差反馈算法的深度网络模型，提出了一种无监督-有监督两阶段训练方法解决梯度问题，实现水电机组故障特征的深度挖掘；（3）提出了一种基于注意力机制-双向门控循环单元和条件随机场的水电机组故障文本分类方法，实现了水电机组文本数据故障特征提取；（4）提出了水电机组故障知识图谱构建方案，提出了双层知识图谱更新方法，构建了双层知识图谱与半朴素贝叶斯结合的水电机组故障推理诊断模型。相关成果为水电机组的安全稳定运行提供了科学的理论指导，具有重要的工程应用价值。

内容获取失败，请点击重试