基于证据推理的电力变压器故障诊断策略.pdf

《基于证据推理的电力变压器故障诊断策略.pdf》由会员分享，可在线阅读，更多相关《基于证据推理的电力变压器故障诊断策略.pdf（9页珍藏版）》请在淘文阁 - 分享文档赚钱的网站上搜索。

1、第 26 卷 第 1 期 中 国 电 机 工 程 学 报 Vol.26 No.1 Jan.2006 2006 年 1 月 Proceedings of the CSEE 2006 Chin.Soc.for Elec.Eng.文章编号：0258-8013（2006）01-0106-09 中图分类号：TM42 文献标识码：A 学科分类号：47040 基于证据推理的电力变压器故障诊断策略 董 明1，严 璋1，杨 莉2，M.D.Judd3（1西安交通大学电气工程学院，陕西省 西安市 710049；2卡里多尼亚大学高电压绝缘诊断研究所，英国 格拉斯哥；3斯特拉思克莱德大学能源与环境学院，英国 格拉斯哥）

2、An Evidential Reasoning Approach to Transformer Fault Diagnosis DONG Ming1,YAN Zhang1,YANG Li2,M.D.Judd3（1.School of Electrical Engineering,Xian Jiaotong University,Xian 710049,Shaanxi Province,China;2.High Voltage Insulation Diagnostics Group,Glasgow Caledonian University,Glasgow,Scotland,UK;3.Inst

3、itute for Energy and Environment,University of Strathclyde,Glasgow,Scotland,UK）ABSTRACT:Methods used to assess the insulation status of power transformers before they deteriorate to a critical state include dissolved gas analysis(DGA),partial discharge(PD)detection and transfer function techniques,e

4、tc.All of these approaches require experience in order to correctly interpret the observations.Artificial Intelligence(AI)is increasingly used to improve interpretation of the individual data sets.However,a satisfactory diagnosis may not be obtained if only one technique is used.For example,the exac

5、t location of PD cannot be predicted if only DGA is performed.However,using diverse methods may result in different diagnosis solutions,a problem that is addressed through the introduction of a fuzzy information infusion model.An inference scheme is proposed that yields consistent conclusions and ma

6、nages the inherent uncertainty in the various methods.With the aid of information fusion,a framework is established that allows different diagnostic tools to be combined in a systematic way.The application of information fusion technique for insulation diagnostics of transformer is effective by mean

7、s of examples.KEY WORDS:Power transformers;Condition monitoring;Dissolved gas analysis(DGA);Information fusion;Insulation diagnostics 摘要：在变压器绝缘劣化之前，可以进行油中溶解气体分析、局部放电检测、传递函数测量等试验方法对其状态进行评估。所有这些试验现象需要很多实际经验才能正确解释。因此人工智能技术逐渐被应用于提高单一试验数据的分析中。但是，仅使用一种方法，可能难以得到满意的诊断结果，如 基金项目：国家自然科学基金项目（59637200）。Project S

8、upported by National Natural Science Foundation of China(59637200).油中溶解气体分析是不能准确对局部放电进行定位。然而，应用不同的方法可能产生各异的诊断结果，因此文中引入模糊信息融合系统来解决此问题，提出了产生一致性结论和处理不同方法中不确定性的证据推理策略。并在信息融合的帮助下，建立了有机组合多种诊断方法系统框架。通过实例证明，基于信息融合的变压器绝缘故障诊断方法是有效的。关键词：电力变压器；状态监测；油中溶解气体分析；信息融合；绝缘诊断学 1 INTRODUCTION Power transformers are key c

9、omponents in electricity transmission and distribution systems.Many different techniques can be used to detect malfunctions of power transformer insulation.These methods all have a role to play in establishing the insulation condition of power transformers.However,few of these methods can,in isolati

10、on,provide all of the information that the transformer operator might require to decide upon a course of action.Moreover,not all of the methods can be applied universally.From practical standpoint,after we recognize the existence of a fault and assess its severity,we are likely to know its location

11、within the transformer,which is useful for guiding further maintenance and repair.As more methods are introduced into the fault diagnosis process,more detail and more reliable results should be expected.Hence,in accordance with the basic framework of information fusion,a PDF 文件使用 pdfFactory Pro 试用版本

12、创建 第 1 期 董 明等：基于证据推理的电力变压器故障诊断策略 107 multi-level comprehensive fault decision model is proposed.The DGA and the results of other electrical tests of power transformers are combined efficiently in the model.In addition,the model takes into account on-site experiences of operation and maintenance in r

13、eaching its diagnosis.2 TRANSFORMER DIAGNOSTIC MET-HODS DGA is a simple,effective,cheap and non-intrusive method that has been widely applied to the fault diagnosis of oil-immersed transformers.Tab.1 shows IEC/IEEE criteria in which overheating and discharging faults are categorized1.With the develo

14、pment of computer science,artificial intelli-gence(AI)is increasingly introduced to diagnose faulty transformers by analyzing the gases.Techni-ques available include expert systems 2,fuzzy logic 3,evolutionary algorithm(EA)4 and the artificial neural network(ANN)5.Furthermore,bodies such as utilitie

15、s,transformer manufacturers,testing laboratories and consultants have developed additional,more advanced testing methods.For example,furans can be used as indicator for the aging status of the solid insulation 6;acoustic and UHF sensors are employed to detect and subsequently locate PD 7;frequency r

16、esponse measurement 8 and recovery voltage measurement 9 are applied to detect damaged windings.Each technique has its strengths and weaknesses.A shortcoming of the plethora of techniques available is that although they were developed to give more specific results,the diagnosis obtained is sometimes

17、 not comparable 表 1 IEC 溶解气体比值法 Tab.1 IEC/IEEE codes for interpretation of DGA method No Classification of Fault Type 2224(C H)(C H)42(CH)(H)2226(C H)(C H)0 No fault 0.1 0.11.0 1.0 1 Low energy partial discharges 0.1 0.1 1.0 2 High energy partial discharges 0.13.0 0.1 3 0.11.0 3 4 High energy discha

18、rges 0.13 0.11.0 3.0 5 150C thermal fault 0.1 0.11.0 1.03.0 6 150-300C thermal fault 1.0 1.0 7 300-700C thermal fault 1.0 1.03.0 8 700C thermal fault 1.0 3.0 across techniques,e.g.,DGA is at this time not a science,but an art subject to variability 1.Nevertheless,when an individual method indicates

19、a problem,the evidence may often be confirmed or enhanced by other methods.For example,when DGA indicates PD,acoustic or UHF sensors could be used to locate the PD source.In this paper,we show how various methods can be integrated to provide coherent,unified analysis.3 MUTI-LEVEL DECISION FUSION OF

20、FAULT INFORMATION FOR TRANSFOR-MERS 3.1 Model Establishment Generally,in the on-site procedure of fault diagnosis,it is highly preferable not to interrupt the supply of power.Where an outage cannot be avoided,it is better to take the suspect transformer off-line for test rather than disassemble it f

21、or internal inspection.By obtaining a more accurate diagnosis from the range of techniques available,we can minimize cost and disruption while repairs take place.Information fusion offers a way of coping with the variance in accuracy and effectiveness among off-line tests and on-line monitoring,on s

22、ite tests and lab analysis,yielding both economic and practical benefits.The rationale behind bringing multiple input information sources is that information in any individual source is either partial or corrupted,that is,it is uncertain and/or imprecise.Pattern classifiers,scene analysis systems,im

23、age processing systems,and computer vision systems all must be capable of integrating knowledge from various sources.In general,information fusion comprises three levels,namely:Data fusion,Feature fusion and Decision fusion.This hierarchical structure is shown in Fig.110.In this section,a more flexi

24、ble and synthetic model based on information fusion is proposed,utilizing available information for interpreting transformer faults.Using this model,fault diagnosis can be conducted step by step,converging on the correct deduction so that reasoning efficiency is enhanced.The fundamental architecture

25、 of this model is shown in Fig.2.PDF 文件使用 pdfFactory Pro 试用版本创建 108 中 国 电 机 工 程 学 报 第 26 卷 Decision Fusion Feature Fusion Feature Fusion Interactive system information Information selection World model Data transformation X1,n X1,2,3 X1,2 X1 X2 X3 Xn Environment 1 2 3 n Information Level of represen

26、tation Low High SYSTEM 图 1 多源信息的等级结构图 Fig.1 The hierarchical structure of multi-source information Data of dissolved gases in oil IEC ratio code Rogers ratio code Improved 3-ratio code Gas generation rate Key gas Evidences Evidence reasoning F0 F1 F2 Discharge faulat Overheating faulat No fault Serv

27、ice history Available off-line tests The content of CO,CO2 or the ratio of CO2/CO Evidences G2 G3 G5 G1 G4 Overheating faults in magnetic circuit Overheating faults in electrical circuit Evidence reasoning Discharge faults related to solid insulation Discharge faults unrelated to solid insulation Mo

28、istened or unnormal aged,etc Level-2 Level-1 图 2 电力变压器故障诊断的多层决策模型 Fig.2 Multilevel decision model for power transformer fault diagnosis 3.2 Level-1 Fault Information Fusion An important feature of this level is that unnecessary outages are avoided if possible.While conventional test of insulating oi

29、l can be conducted online,DGA is usually considered to be the first step of fault monitoring and diagnosis.The goal of level-1 fault information fusion as shown in Fig.2 is to obtain a rough assessment of the technical status of transformers.Indications from DGA,such as,the IEC ratio code method,Rog

30、ers ratio code method,the improved 3-ratio method,key gas method and gas generation rate are synthesized at this level.The recognition frame in level-1 is defined as=F0,F1,F2,where F0,F1,F2 correspond to“no fault”,“discharge fault”,and“overheating fault”,respectively.3.3 Level-2 Fault Information Fu

31、sion In order to obtain more specific evaluation of the fault,building on the decision of level-1,the conclusion is more detailed in relation to fault spot in level-2.Here,some available off-line tests are required.This corresponds to refinement of the recognition frame in evidence theory.Therefore,

32、the above mentioned frame=F1,F2 is refined to a sub-frame as expanded=G1,G2,G3,G4,G5,where G1,G2,G3,G4,G5 correspond to“overheating fault in magnetic circuit”,“overheating fault in electrical circuit”,“discharge fault related to solid insulation”,“discharge fault unrelated to solid insulation”,“mois

33、ture or abnormal ageing”.4 MODIFED COMBINATION RULE OF EVIDENCE THEORY Dempster-Shafer Theory(DST)is a mathematical theory of evidence often interpreted as an extension of Bayesian probability theory 11.The basic concept is that two uncertainty measures,which include supportability(the degree of bel

34、ief that directly supports a specific element of the power set)and plausibility(the degree of belief that does not directly contradict a specific element of the power set),are computed using the unified probability measures.Measures of supportability and plausibility can be expressed in terms of a B

35、asic Probability Assignment(BPA)function.A probability mass m(A)is defined as the degree of evidence supporting the claim that the truth is in the observed object(A).In absence of further evidence no more specific statement can be made.An important aspect of this theory is to combine the evidence ob

36、tained from multiple sources and model the conflict between them.Dempsters rule of combination produces a unified probability measure.So this rule has the effect of completely ignoring conflict and attributing any probability mass associated with conflict to the null set 12.Consequently,this operati

37、on will yield counterin-tuitive results in the face of significant conflict in certain contexts.Therefore,we PDF 文件使用 pdfFactory Pro 试用版本创建 第 1 期 董 明等：基于证据推理的电力变压器故障诊断策略 109 propose algorithms that integrate consistent and conflicting evidence for multivariate information fusion problems.There are t

38、wo motivations behind this approach.First,each feature value of an unknown observation provides some evidence about the decision-making.Hence,the aggregation of such evidences using the Dempster-Shafer framework is expected to result in a good performance of fusion.Second,it enables us to extend the

39、 combination algorithm for dealing with multi-level information.For each feature we find possible information and then define a feature-specific BPA.These BPAs are then combined using our modified Dempsters rule of aggregation,which will now be described.This algorithm maintains the basic architectu

40、re of Dempsters rule but differs in its dealing with conflicting evidence.Let M be a BPA set and F be its corresponding evidence set.The local conflict Cij between Fi and Fj is formalized as,()()ijijjjAAF AFijijAiCm A mA=(1)In order to measure the degree of consistency and reliability of evidence on

41、 the same frame of discernment,the distance between two bodies of evidence 13 is 221212121(,)(2,)2BPAdm mmmm m=+(2)Here,is the scalar product,which means 22121211,()()NNijijijijAAm mm A mAAA=,()ijA AP where|A|denotes the cardinality of A.Since the N objects of are indiscernible,the only measure dist

42、inguishing any two subsets is their cardinality.So the similarity measure Fsim(mi,mj)between mi and mj may be obtained as sim(,)1(,)ijBPAijFm mdm m=(3)If the distance is smaller,the degree of similarity will be larger.The supporting degree of evidence mi may be obtained as supsim1()(,)niijjj iFmFm m

43、=(4)Here,n is the number of pieces of evidence.As with the mass function,the degree of belief for mi may be defined by the following formula crdsupsup1()()/()niiiiFmFmFm=(5)where Fsup(mi)represents the degree to which other evidence supports mi,that is,it is function of similarity.Global conflict is

44、 described by(1)/2()()glo1n nssijijsCw c=(6)where(1)/2()()()1/n nsssijijijswpp=and()crdcrd()()/2sijijpFmFm=+The validity coefficient for conflicting evidence,which describes the total degree of conflict,is defined by gloeC=(7)The modified Dempster-Shafer rule of combination then becomes()()(),()()()

45、(1)()0m Ap Ak q AAmp Xk q Xkm=+=+=(8)where the factor k is the estimated value of conflicting evidence,(0,1)kand 11122()()()()iniinnA FAAp Am A mAmA=11()()niiq Am An=Associated with m are a belief(or credibility)function:201Bel，and a plausibility function:201Pl，defined for all B as()()BABel Bm A=(9)

46、()()BAPl Bm A=(10)The quantity Bel(B)can be interpreted as a global measure of belief that hypothesis B is true,while Pl(B)may be viewed as the level of belief that could potentially be placed in B if further information became available.PDF 文件使用 pdfFactory Pro 试用版本创建 110 中 国 电 机 工 程 学 报 第 26 卷 To s

47、ummarize,the important attribute of this rule of combination is that conflicting evidence also contains much information to assist decision-making.A global conflict is first calculated as the weighted average of the local conflict.A validity coefficient is then defined to show the effect of the conf

48、lict evidence on the combined results.Thus mass associated with conflict is interpreted as the degree of association instead of allocating conflict to the universal set()or the null set().This proposed rule considers not only the consistency of pieces of evidence,but also the conflict between them.I

49、n this way,the reliability and rationality of the combined results is strengthened.5 THE PROPOSED INFERENCE SCHEME 5.1 Diagnostic Knowledge Representation In many situations,it is difficult to capture and properly represent real-world knowledge in a precise form.Some structures that have been used i

50、nclude frames,script and production rules.In this paper,fuzzy production rules will be used for diagnostic knowledge representation.These are derived from production rules by allowing the antecedent portion of the rule to contain some fuzzy quantifiers 14(i.e.strong,weak,etc.).Let Di represent the a