CN107516108A - Grader creation method and partial discharge of transformer method of fault pattern recognition - Google Patents

Grader creation method and partial discharge of transformer method of fault pattern recognition Download PDF

Info

Publication number
CN107516108A
CN107516108A CN201710697454.2A CN201710697454A CN107516108A CN 107516108 A CN107516108 A CN 107516108A CN 201710697454 A CN201710697454 A CN 201710697454A CN 107516108 A CN107516108 A CN 107516108A
Authority
CN
China
Prior art keywords
attribute
training
classification
discharge
grader
Prior art date
Application number
CN201710697454.2A
Other languages
Chinese (zh)
Inventor
冯运
陈凌
甘德刚
张宗喜
蒋伟
Original Assignee
国网四川省电力公司电力科学研究院
国家电网公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 国网四川省电力公司电力科学研究院, 国家电网公司 filed Critical 国网四川省电力公司电力科学研究院
Priority to CN201710697454.2A priority Critical patent/CN107516108A/en
Publication of CN107516108A publication Critical patent/CN107516108A/en

Links

Classifications

    • GPHYSICS
    • G01MEASURING; TESTING
    • G01RMEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
    • G01R31/00Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
    • G01R31/12Testing dielectric strength or breakdown voltage ; Testing or monitoring effectiveness or level of insulation, e.g. of a cable or of an apparatus, for example using partial discharge measurements; Electrostatic testing
    • G01R31/1227Testing dielectric strength or breakdown voltage ; Testing or monitoring effectiveness or level of insulation, e.g. of a cable or of an apparatus, for example using partial discharge measurements; Electrostatic testing of components, parts or materials
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06KRECOGNITION OF DATA; PRESENTATION OF DATA; RECORD CARRIERS; HANDLING RECORD CARRIERS
    • G06K9/00Methods or arrangements for reading or recognising printed or written characters or for recognising patterns, e.g. fingerprints
    • G06K9/62Methods or arrangements for recognition using electronic means
    • G06K9/6217Design or setup of recognition systems and techniques; Extraction of features in feature space; Clustering techniques; Blind source separation
    • G06K9/6256Obtaining sets of training patterns; Bootstrap methods, e.g. bagging, boosting
    • G06K9/6257Obtaining sets of training patterns; Bootstrap methods, e.g. bagging, boosting characterised by the organisation or the structure of the process, e.g. boosting cascade
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06KRECOGNITION OF DATA; PRESENTATION OF DATA; RECORD CARRIERS; HANDLING RECORD CARRIERS
    • G06K9/00Methods or arrangements for reading or recognising printed or written characters or for recognising patterns, e.g. fingerprints
    • G06K9/62Methods or arrangements for recognition using electronic means
    • G06K9/6267Classification techniques
    • G06K9/6279Classification techniques relating to the number of classes
    • G06K9/628Multiple classes
    • G06K9/6281Piecewise classification, i.e. whereby each classification requires several discriminant rules
    • G06K9/6282Tree-organised sequential classifiers

Abstract

The invention discloses grader creation method and partial discharge of transformer method of fault pattern recognition, grader creation method comprises the following steps:Choose corona discharge, bubble-discharge, multigroup characteristic of the class shelf depreciation of creeping discharge three respectively includes N number of attribute as training set and test set, each characteristic, wherein, N is the natural number more than 1;Resampling is carried out to training set, randomly generates training set;Utilize decision tree corresponding to the generation of each training set, before attribute being selected in each non-leaf nodes, Split Attribute collection of the m attribute as present node is randomly selected from N number of attribute, and line splitting is entered to the node with classification accuracy highest divisional mode in this m attribute;Test set is tested using decision tree, obtains corresponding classification, using exported in decision tree most classifications as the test set belonging to classification.Using the grader to partial discharge of transformer Fault Pattern Recognition while accuracy rate is ensured, substantially reduce the time.

Description

Grader creation method and partial discharge of transformer method of fault pattern recognition
Technical field
The present invention relates to status of electric power diagnostic field, and in particular to a kind of grader wound based on random forests algorithm Construction method and partial discharge of transformer method of fault pattern recognition.
Background technology
Partial discharge of transformer is to cause Transformer Insulation Aging, one of damage, the main reason for causing electric power accident.Cause This, the diagnostic method of partial discharge of transformer in perfect operation, has to improving power transformer reliability of operation and security It is significant.
At present, the fault diagnosis of partial discharge of transformer is mainly based on threshold diagnostic, i.e., when discharge capacity exceedes a certain set Early warning is then carried out by system during fixed minimum discharge capacity early warning value, then carries out judgement post processing by operation test personnel.But threshold The information content that value method is relatively single, can provide is few, can not provide the more rich information such as shelf depreciation discharge characteristics, type. Therefore, sight has been placed in the identification of shelf depreciation discharge mode by substantial amounts of scientific researchers, and achieves certain to enter Exhibition.The main method of PD Pattern Recognition is that electric discharge is divided into several known types first at present, then to every species The shelf depreciation of type is largely tested, and therefrom extracts the characteristic parameter for representing such discharge mode, and the spy to obtaining Sign parameter establishes spectrum library, and finally obtained a few class spectrum datas are trained using intelligent algorithm, are finally reached classification Effect.The intelligent algorithm being commonly used during classification includes neutral net, SVMs, genetic algorithm etc..These Algorithm can obtain relatively good effect, it will be appreciated, however, that the pattern-recognition of shelf depreciation is often in accuracy rate and time On be difficult to take into account simultaneously.Want the relatively good training effect of acquirement and generally require substantial amounts of training data and training time.
The content of the invention
In order to solve the above-mentioned technical problem the present invention provides a kind of grader creation method and partial discharge of transformer failure Mode identification method, it substantially reduces the time while accuracy rate is ensured.
The present invention is achieved through the following technical solutions:
Grader creation method, comprises the following steps:
Corona discharge, bubble-discharge, multigroup characteristic of the class shelf depreciation of creeping discharge three are chosen respectively as training Collection and test set, each characteristic include N number of attribute, wherein, N is the natural number more than 1;
Resampling is carried out using Bootstrap methods to training set, randomly generates training set;
Using decision tree corresponding to the generation of each training set, before selecting attribute in each non-leaf nodes, from N number of attribute In randomly select Split Attribute collection of the m attribute as present node, and divided with classification accuracy highest in this m attribute Mode enters line splitting to the node, wherein, m is less than N;
Test set is tested using decision tree, obtains corresponding classification, most classifications will be exported in decision tree and is made For the classification belonging to the test set.
Random forest is all fully nonlinear water wave during each decision tree classification, although each in random forest Number is all very weak, but everybody combines output classification of the ballot selection mode as test set, allows for classification accuracy significantly Improve.Appropriate data can produce more decision tree in training, realize the accuracy rate of its classification.Random forest can be direct Produce multi-class classification results and each tree is all independent growths, therefore parallelization acceleration can be carried out, it is not necessary to too many Training data and training time.
Preferably, each characteristic includes 24 attributes, 24 attributes includeWithDegree of asymmetry Asy and cross-correlation coefficient Cc,With's Degree of skewness Sk, steepness Ku, peak value Peak.Wherein,Mean discharge magnitude phase distribution two dimension spectrogram is represented, Maximum pd quantity phase distribution two dimension spectrogram is represented,Represent discharge time phase distribution spectrogram.WithPoint Positive half cycle mean discharge magnitude phase distribution two dimension spectrogram, negative half period mean discharge magnitude phase distribution two dimension spectrogram are not represented.Positive half cycle maximum pd quantity phase distribution two dimension spectrogram is represented,Represent positive-negative half-cycle maximum pd quantity phase point Cloth two dimension spectrogram,Positive half cycle discharge time phase distribution spectrogram is represented,Represent negative half period discharge time phase point Cloth spectrogram.
Preferably, the confirmation method of the classification accuracy highest divisional mode is:
Be n classes by m Attribute transposition, the ratio Pi of each of which class is number/m of the classes of Pi=i-th, i=1,2, 3、……、n;
The entropy Info (D) of training set is calculated,
Respectively calculate m Attribute transposition sample set after training set entropy Infoy(D), Wherein, V be each attribute by training set divide class quantity, y=1,2,3 ..., m;
Calculate information gain Gain (A):Gain (A)=Info (D)-InfoA(D),
Classification accuracy highest divisional mode is judged according to information gain Gain (A).
Partial discharge of transformer method of fault pattern recognition, comprises the following steps:
Characteristic is input in the grader for using and being created by the above method, to export identification classification.
Because using the grader of above method establishment, its decision tree nodes uses classification accuracy highest divisional mode Enter line splitting, the accuracy of classification can be effectively improved;And grader can directly produce multiclass using random forests algorithm Other classification results, and each tree in random forest is all independent growths, therefore parallelization acceleration can be carried out, greatly improve Classification and Identification speed.
The present invention compared with prior art, has the following advantages and advantages:
1st, the present invention is created using random forests algorithm to grader, the establishment of its classification tree with classification accuracy most High divisional mode enters line splitting, and the independent growth of classification tree, can run parallel, it is while accuracy rate is ensured, significantly Shorten recognition time.
Brief description of the drawings
Accompanying drawing described herein is used for providing further understanding the embodiment of the present invention, forms one of the application Point, do not form the restriction to the embodiment of the present invention.In the accompanying drawings:
Fig. 1 is the operation result that the characteristic of discharge fault pattern known to 300 groups is inputted in grader.
Embodiment
For the object, technical solutions and advantages of the present invention are more clearly understood, with reference to embodiment and accompanying drawing, to this Invention is described in further detail, and exemplary embodiment of the invention and its explanation are only used for explaining the present invention, do not make For limitation of the invention.
Embodiment 1
Grader creation method, comprises the following steps:
Corona discharge, bubble-discharge, multigroup characteristic of the class shelf depreciation of creeping discharge three are chosen respectively as training Collection and test set, each characteristic include N number of attribute, wherein, N is the natural number more than 1;
Resampling is carried out using Bootstrap methods to training set, randomly generates training set;
Using decision tree corresponding to the generation of each training set, before selecting attribute in each non-leaf nodes, from N number of attribute In randomly select Split Attribute collection of the m attribute as present node, and divided with classification accuracy highest in this m attribute Mode enters line splitting to the node, wherein, m is less than N;Choose m attribute and come to instruction in divisional mode the inside exactly from N number of attribute Practice collection to be classified, wherein m value, and choose be any m attribute combination, be exactly different divisional modes, the step Selection sort accuracy rate highest divisional mode in rapid;
The confirmation method of classification accuracy highest divisional mode is:
Be n classes by m Attribute transposition, the ratio Pi of each of which class is number/m of the classes of Pi=i-th, i=1,2, 3、……、n;
The entropy Info (D) of training set is calculated,
Respectively calculate m Attribute transposition sample set after training set entropy Infoy(D), Wherein, V be each attribute by training set divide class quantity, y=1,2,3 ..., m;For example, attribute 1 divides the sample in training set This, training set is divided into v different classes by attribute 1, and after division, the entropy of training set is: Wherein, V is the quantity that each attribute divides training set class;
Calculate information gain Gain (A):Gain (A)=Info (D)-InfoA(D),
Classification accuracy highest divisional mode, the maximum attribute institute of information gain are judged according to information gain Gain (A) Corresponding divisional mode is best divisional mode.
Test set is tested using decision tree, obtains corresponding classification, most classifications will be exported in decision tree and is made For the classification belonging to the test set.
Wherein, attribute has 24, including degree of asymmetry Asy, degree of skewness Sk, steepness Ku, cross-correlation coefficient Cc, peak value Peak, it is specific as shown in table 1.
The statistical nature parameter of the shelf depreciation collection of illustrative plates of table 1
Now using 1500 groups of characteristics as training set, 300 groups of characteristics create as exemplified by test set to grader Method is described in detail.This method is created based on random forests algorithm.
Training data can make training effect bad very little, can then cause the training time long too much.The data set of this programme Quantity and attribute N and training set data itself the characteristics of have relation, for partial discharge of transformer Fault Pattern Recognition, use When the classification of shelf depreciation, the quantity of training set, which is chosen for 600 to 1500 groups, to be advisable, and test set is then desirably no more than 500 groups.
The selection of training set, test set:Corona discharge, bubble-discharge, creeping discharge three class shelf depreciation are chosen respectively 500 groups of characteristics are as training set, then choose corona discharge, bubble-discharge, the 100 of the class shelf depreciation of creeping discharge three respectively Group characteristic is as test set.
The establishment of random forest grader:Bootstrap method resamplings are utilized to 1500 groups of training set datas, random production Raw 1500 training set S1、S2、S3、…、S1500
Utilize decision tree C corresponding to the generation of each training set1、C2、C3、…、C1500, selected in each non-leaf nodes Before attribute, randomly select Split Attribute collection of the m attribute as present node from 24 attributes, and with this m attribute most Good divisional mode enters line splitting to the node, and each tree is all completely grown up, and without beta pruning, m<24;
For test set data, tested using each decision tree, obtain corresponding classification C1(X)、C2(X)、C3 (X)、…、C1500(X), using the method for ballot, most classifications will be exported in 1500 decision trees as test set sample X institutes The classification of category.
Embodiment 2
Partial discharge of transformer method of fault pattern recognition, comprises the following steps:Characteristic is input to using by reality In the grader of method establishment for applying example 1, to export identification classification.
When partial discharge of transformer fault mode is identified the grader created using method in embodiment 1, its is defeated The neuron entered is 24, represents 24 attributes of shelf depreciation spectrogram respectively, the neuron number of output is 3, is respectively Corona discharge, creeping discharge, bubble-discharge.
In order to verify the accuracy rate of this programme, by the characteristic input grader of discharge fault pattern known to 300 groups. The classification performance of random forest is analyzed using three-dimensional diagrammatic representation.It can be seen from above step, random forest pair is being used When training set is trained, generate 1500 decision trees altogether, and the output classification of decision tree only has three kinds, i.e., corona discharge, Creeping discharge, bubble-discharge, represented respectively with x, y, z in graphics, it meets x+y+z=1500.Therefore, for a certain For individual sample, it trains coordinate P (x, y, z) of the output result on graphics all the time flat represented by x+y+z=1500 On face.Wherein x, y, z is all higher than 0 and is integer.
Obtained operation result is as shown in figure 1, the actually Green triangle shape is equilateral triangle, three perpendicular bisectors Green triangle shape is divided into equal three parts, the region where each section corresponds to a kind of result.Operation result shows, Have that 93 groups of tests are accurate in the corona discharge of 95 groups of tests, accuracy rate 97.8947%;There are 97 groups of tests in 97 groups of creeping discharge Accurately, accuracy rate 100%;There are 106 groups of tests accurate in 107 groups of bubble-discharges, accuracy rate 99.0741%, averagely make a definite diagnosis Rate is 99%.In addition, it is only 5-6s to use the time required to this method classification, Diagnostic Time is substantially reduced.And use existing side Method carries out grader and generally requires 15s or so, or even longer.
Above-described embodiment, the purpose of the present invention, technical scheme and beneficial effect are carried out further Describe in detail, should be understood that the embodiment that the foregoing is only the present invention, be not intended to limit the present invention Protection domain, within the spirit and principles of the invention, any modification, equivalent substitution and improvements done etc., all should include Within protection scope of the present invention.

Claims (5)

1. grader creation method, it is characterised in that comprise the following steps:
Choose respectively corona discharge, bubble-discharge, multigroup characteristic of the class shelf depreciation of creeping discharge three as training set and Test set, each characteristic include N number of attribute, wherein, N is the natural number more than 1;
Resampling is carried out using Bootstrap methods to training set, randomly generates training set;
Using each training set generation corresponding to decision tree, in each non-leaf nodes select attribute before, from N number of attribute with Machine extracts Split Attribute collection of the m attribute as present node, and with classification accuracy highest divisional mode in this m attribute Line splitting is entered to the node, wherein, m is less than N;
Test set is tested using decision tree, obtains corresponding classification, most classifications will be exported in decision tree and is used as this Classification belonging to test set.
2. grader creation method according to claim 1, it is characterised in that each characteristic includes 24 attributes.
3. grader creation method according to claim 2, it is characterised in that 24 attributes include WithDegree of asymmetry Asy and cross-correlation coefficient Cc, WithDegree of skewness Sk, steepness Ku, peak value Peak.
4. grader creation method according to claim 1, it is characterised in that the classification accuracy highest division side The confirmation method of formula is:
Be n classes by m Attribute transposition, the ratio Pi of each of which class is number/m of the classes of Pi=i-th, i=1,2,3 ..., n;
The entropy Info (D) of training set is calculated,
Respectively calculate m Attribute transposition sample set after training set entropy Infoy(D),Its In, V be each attribute by training set divide class quantity, y=1,2,3 ..., m;
Calculate information gain Gain (A):Gain (A)=Info (D)-InfoA(D),
Classification accuracy highest divisional mode is judged according to information gain Gain (A).
5. partial discharge of transformer method of fault pattern recognition, it is characterised in that comprise the following steps:
Characteristic is input in the grader using the method establishment by Claims 1-4, to export identification classification.
CN201710697454.2A 2017-08-15 2017-08-15 Grader creation method and partial discharge of transformer method of fault pattern recognition CN107516108A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201710697454.2A CN107516108A (en) 2017-08-15 2017-08-15 Grader creation method and partial discharge of transformer method of fault pattern recognition

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201710697454.2A CN107516108A (en) 2017-08-15 2017-08-15 Grader creation method and partial discharge of transformer method of fault pattern recognition

Publications (1)

Publication Number Publication Date
CN107516108A true CN107516108A (en) 2017-12-26

Family

ID=60722831

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201710697454.2A CN107516108A (en) 2017-08-15 2017-08-15 Grader creation method and partial discharge of transformer method of fault pattern recognition

Country Status (1)

Country Link
CN (1) CN107516108A (en)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109613351A (en) * 2018-11-21 2019-04-12 北京国网富达科技发展有限责任公司 A kind of method for diagnosing faults of transformer, equipment and system
CN109657720A (en) * 2018-12-20 2019-04-19 浙江大学 A kind of inline diagnosis method of power transformer shorted-turn fault
CN110068741A (en) * 2019-05-29 2019-07-30 国网河北省电力有限公司石家庄供电分公司 A method of the transformer fault diagnosis based on categorised decision tree
CN110161387A (en) * 2019-06-03 2019-08-23 河海大学常州校区 A kind of power equipment partial discharge amount prediction technique based on improvement gradient boosted tree
CN111693812A (en) * 2020-06-15 2020-09-22 中国科学技术大学 Large transformer fault detection method based on sound characteristics

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7007035B2 (en) * 2001-06-08 2006-02-28 The Regents Of The University Of California Parallel object-oriented decision tree system
CN103245861A (en) * 2013-05-03 2013-08-14 云南电力试验研究院(集团)有限公司电力研究院 Transformer fault diagnosis method based on Bayesian network
CN103366160A (en) * 2013-06-28 2013-10-23 西安交通大学 Objectionable image distinguishing method integrating skin color, face and sensitive position detection
CN105894177A (en) * 2016-03-25 2016-08-24 国家电网公司 Decision-making-tree-algorithm-based analysis and evaluation method for operation risk of power equipment
CN106202886A (en) * 2016-06-29 2016-12-07 中国铁路总公司 Track circuit red band Fault Locating Method based on fuzzy coarse central Yu decision tree

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7007035B2 (en) * 2001-06-08 2006-02-28 The Regents Of The University Of California Parallel object-oriented decision tree system
CN103245861A (en) * 2013-05-03 2013-08-14 云南电力试验研究院(集团)有限公司电力研究院 Transformer fault diagnosis method based on Bayesian network
CN103366160A (en) * 2013-06-28 2013-10-23 西安交通大学 Objectionable image distinguishing method integrating skin color, face and sensitive position detection
CN105894177A (en) * 2016-03-25 2016-08-24 国家电网公司 Decision-making-tree-algorithm-based analysis and evaluation method for operation risk of power equipment
CN106202886A (en) * 2016-06-29 2016-12-07 中国铁路总公司 Track circuit red band Fault Locating Method based on fuzzy coarse central Yu decision tree

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
陈伟根 等: "不同气隙尺寸的油纸绝缘气隙放电特征及发展阶段识别", 《电工技术学报》 *
陈新岗 等: "信息融合在变压器油纸绝缘局部放电识别中的应用", 《高电压技术》 *

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109613351A (en) * 2018-11-21 2019-04-12 北京国网富达科技发展有限责任公司 A kind of method for diagnosing faults of transformer, equipment and system
CN109657720A (en) * 2018-12-20 2019-04-19 浙江大学 A kind of inline diagnosis method of power transformer shorted-turn fault
CN110068741A (en) * 2019-05-29 2019-07-30 国网河北省电力有限公司石家庄供电分公司 A method of the transformer fault diagnosis based on categorised decision tree
CN110161387A (en) * 2019-06-03 2019-08-23 河海大学常州校区 A kind of power equipment partial discharge amount prediction technique based on improvement gradient boosted tree
CN111693812A (en) * 2020-06-15 2020-09-22 中国科学技术大学 Large transformer fault detection method based on sound characteristics

Similar Documents

Publication Publication Date Title
Van Gassen et al. FlowSOM: Using self‐organizing maps for visualization and interpretation of cytometry data
CN102521656B (en) Integrated transfer learning method for classification of unbalance samples
Garg et al. A probabilistic language formalism for stochastic discrete-event systems
DE69907498T2 (en) METHOD FOR QUICKLY DETECTING THE TONE HEIGHT
Listerman Framing of science issues in opinion-leading news: international comparison of biotechnology issue coverage
CN106845717B (en) Energy efficiency evaluation method based on multi-model fusion strategy
CN101383748B (en) Community division method in complex network
DE10157854A1 (en) Method and apparatus for generating a transaction-based stimulus for a simulation of VLSI circuits using an event coverage analysis
DE10014492A1 (en) Method for testing semiconductor wafers identifies signs of failure in semiconductor wafers by splitting test data into sub-sets of test data belonging to one failure characteristic and not belonging to this characteristic.
CN104318482A (en) Comprehensive assessment system and method of smart distribution network
CN101625755B (en) Image division method based on watershed-quantum evolution clustering algorithm
CN104777418B (en) A kind of analog-circuit fault diagnosis method based on depth Boltzmann machine
CN103914353A (en) Combined test method for combining software reliability tests with hardware reliability tests
CN109444656B (en) Online diagnosis method for deformation position of transformer winding
CN107657288A (en) A kind of power scheduling flow data method for detecting abnormality based on isolated forest algorithm
CN105866725A (en) Method for fault classification of smart electric meter based on cluster analysis and cloud model
CN104715318A (en) Multi-dimensional operational risk evaluating method for communication network
CN104462184B (en) A kind of large-scale data abnormality recognition method based on two-way sampling combination
CN104020401A (en) Cloud-model-theory-based method for evaluating insulation thermal ageing states of transformer
CN104392056B (en) A kind of directly driven wind-powered unit modeling method and device
CN109378010A (en) Training method, the speech de-noising method and device of neural network model
CN107292335A (en) A kind of transmission line of electricity cloud data automatic classification method based on Random Forest model
CN103778262B (en) Information retrieval method and device based on thesaurus
CN104793124A (en) Switched circuit fault diagnosis method based on wavelet transformation and ICA (independent component analysis) feature extraction
CN108680348A (en) A kind of breaker mechanical fault diagnosis method and system based on random forest

Legal Events

Date Code Title Description
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination