电力中文文本数据挖掘技术及其在可靠性中的应用研究

发布时间：2018-09-04 08:09

【摘要】：随着电网智能化与信息化的建设,电网企业积累了大量的数据,逐渐构成了学术界和工业界共同关注的电力大数据。本文将电力系统全生产过程、电力设备全寿命过程中产生的数据,统称为全寿命数据,其类型涵盖了结构化和非结构化数据。目前在电力领域中,主要侧重于结构化数据挖掘的研究,也有针对图像识别的研究,但是,电力文本挖掘研究才刚刚起步。本文以获取可靠性的相关指标(统计指标、健康指数、广义故障率和可靠性)为目的,以电力设备资产管理为着眼点,重点对电力文本挖掘及多源数据的融合技术进行了深入研究,取得了如下成果。1.针对可靠性和资产管理的特点,给出了电网自然语言处理(Natural language processing for power grids,NLP4PG)的基础定义。编撰了专业词典、变压器本体词典和领域语料库,并将部分内容开源共享。融合了本体论与电力领域知识,建立了适用于NLP4PG的总框架。指出了NLP4PG的工作原理和潜在应用,并分析了若干语言特性和语言模型,填补了电力中文文本挖掘研究的空白。2.基于语义框架思想,将离线统计学习与在线语义规则的方法相结合,提出了适用于故障与缺陷分类统计的文本挖掘模型。解决了缺陷文本句子成分难以划分、数字量无法精确提取等问题,并实现了本体字典的半自动化构建技术。3.针对设备健康指数(Health Index)多源、多类型的特点,改进了kNN算法,提出了具备自主区间寻优能力的电力缺陷短文本分类模型。该模型实现了从故障缺陷文本到HI的自学习映射,能够在最大程度上挖掘同类设备的相似信息、融合同台设备的历史信息,使得HI更加合理、精准。在一定程度上,改变了运维人员主观评价故障/缺陷等级的方式,并符合电网企业精细化管理的发展要求。4.为了能够充分利用从全寿命数据中挖掘到的亚健康状态信息,同时处理多类型复发事件,提出了基于鞅过程和非参数理论的设备故障率预测模型。该理念完全区别于以往单一变量的故障率模型(基于时间或状态量的定参数分布函数),通过算例结果、鞅残差检验、灵敏度分析表明,该模型能够定量地、精确地描绘短期故障率,为剩余寿命预测、状态检修优化提供基础可靠性指标。5.针对智能变电站自动化系统(Smart Substation Automation System,SSAS)这类新型的、缺乏运维数据的场景,提出了基于网络层次分析法的广义可靠性评估模型。参考IEC 61850标准、设计文档和相关文献,构建了SSAS本体平台,包括物理本体、逻辑本体和信息流本体。对应地,将系统可靠性作为总指标,涵盖了物理可靠性、逻辑可靠性、信息流可靠性和系统经济性。四大指标下面包含了若干定量与定性的指标。通过算例结果与灵敏度分析表明,在缺乏运维数据的情况下,本模型能够用于新型系统的可靠性评估,为设计方案的遴选提供一种新思路。
[Abstract]:With the construction of intelligent and information power grid, power grid enterprises have accumulated a large amount of data, which has gradually become the common concern of the academic and industrial power big data. In this paper, the data generated in the whole production process of power system and in the whole life process of power equipment are referred to as the whole life data. The types of data include structured and unstructured data. At present, the research on structured data mining and image recognition is mainly focused on in the field of electric power. However, the research of power text mining is only in its infancy. The purpose of this paper is to obtain reliability related indicators (statistical index, health index, generalized failure rate and reliability), and to focus on the management of power equipment assets. The technology of power text mining and multi-source data fusion is studied in detail, and the following results are obtained. 1. According to the characteristics of reliability and asset management, the basic definition of natural language processing (Natural language processing for power grids,NLP4PG) of power grid is given. Professional dictionaries, transformer ontology dictionaries and domain corpus are compiled, and some contents are shared open source. The ontology and power domain knowledge are combined, and the general framework for NLP4PG is established. The working principle and potential application of NLP4PG are pointed out, and some language characteristics and language models are analyzed. Based on the idea of semantic framework, a text mining model for fault and defect classification statistics is proposed by combining offline statistical learning with online semantic rules. It solves the problem that it is difficult to divide the sentence components of the defective text, and the digital quantity can not be extracted accurately, and realizes the semi-automatic construction technology of ontology dictionary .3. In view of the multi-source and multi-type characteristics of equipment health index (Health Index), an improved kNN algorithm is proposed, and a power defect short text classification model with autonomous interval optimization ability is proposed. The model realizes the self-learning mapping from fault defect text to HI. It can mine the similar information of the same equipment to the maximum extent and fuse the historical information of the same equipment, making the HI more reasonable and accurate. To a certain extent, it changes the way of subjective evaluation of fault / defect grade by operation and maintenance personnel, and meets the development requirement of fine management of power grid enterprises. In order to make full use of the sub-health state information mined from the whole life data and to deal with multiple recurrence events at the same time, a prediction model of equipment failure rate based on martingale process and non-parametric theory is proposed. The idea is completely different from the failure rate model of single variable (the fixed parameter distribution function based on time or state quantity). The results of example, martingale residuals test and sensitivity analysis show that the model can be used quantitatively. Describe the short-term failure rate accurately and provide basic reliability index. 5. 5 for residual life prediction and condition-based maintenance optimization. Aiming at the new situation of intelligent substation automation system (Smart Substation Automation System,SSAS), which lacks the operational data, a generalized reliability evaluation model based on the analytic hierarchy process (AHP) is proposed. Referring to IEC 61850 standard, design documents and related documents, the SSAS ontology platform is constructed, including physical ontology, logical ontology and information flow ontology. Correspondingly, the reliability of the system is taken as the total index, which covers physical reliability, logic reliability, information flow reliability and system economy. The four indicators contain a number of quantitative and qualitative indicators. The results of numerical examples and sensitivity analysis show that the model can be used to evaluate the reliability of the new system without operational and maintenance data and provide a new idea for the selection of design scheme.
【学位授予单位】：浙江大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TM732;TP391.1

【相似文献】