CN110133146A - A kind of Diagnosis Method of Transformer Faults and system considering unbalanced data sample - Google Patents
A kind of Diagnosis Method of Transformer Faults and system considering unbalanced data sample Download PDFInfo
- Publication number
- CN110133146A CN110133146A CN201910451726.XA CN201910451726A CN110133146A CN 110133146 A CN110133146 A CN 110133146A CN 201910451726 A CN201910451726 A CN 201910451726A CN 110133146 A CN110133146 A CN 110133146A
- Authority
- CN
- China
- Prior art keywords
- transformer
- classifier
- level
- data
- state
- Prior art date
- Legal status (The legal status 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 status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N30/00—Investigating or analysing materials by separation into components using adsorption, absorption or similar phenomena or using ion-exchange, e.g. chromatography or field flow fractionation
- G01N30/02—Column chromatography
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
- G01R31/12—Testing 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
- G01R31/1227—Testing 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
- G01R31/1263—Testing 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 of solid or fluid materials, e.g. insulation films, bulk material; of semiconductors or LV electronic components or parts; of cable, line or wire insulation
- G01R31/1281—Testing 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 of solid or fluid materials, e.g. insulation films, bulk material; of semiconductors or LV electronic components or parts; of cable, line or wire insulation of liquids or gases
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/2431—Multiple classes
Landscapes
- Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Data Mining & Analysis (AREA)
- Analytical Chemistry (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Immunology (AREA)
- Pathology (AREA)
- Biochemistry (AREA)
- Chemical & Material Sciences (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Other Investigation Or Analysis Of Materials By Electrical Means (AREA)
Abstract
The invention discloses a kind of Diagnosis Method of Transformer Faults for considering unbalanced data sample, comprising steps of 100: acquiring the oil chromatography sample data of transformer, and pre-processed to obtain by pretreated oil chromatography sample data to it;200: first order classifier is constructed and train, wherein being trained using by pretreated oil chromatography sample data to first order classifier;State feature and oil chromatography sample data based on the output of first order classifier obtain Fusion Features vector;300: second level classifier being trained based on EasyEnsemble integrated learning approach, the output based on several second level sub-classifiers synthesizes second level classifier;400: the transformer oil chromatographic data for needing to diagnose being inputted into first order classifier, then export the transformer state of the transformer oil chromatographic data characterization from the output end of second level classifier.The Diagnosis Method of Transformer Faults obtains more acurrate and balance effect relative to conventional method.
Description
Technical field
The present invention relates in electric system fault diagnosis method and system more particularly to a kind of transformer fault diagnosis side
Method and system.
Background technique
Can power transformer carries important function in the power system, whether normal operation reliable power supply with power grid
It is closely related.Although China's electrical equipment research and development manufacturing technology has ranked among international forefront in recent years, insulate natural aging, environment
The excessively high a variety of undesirable elements of inclement condition, operating load still repeatedly induce Power Transformer Faults, cause Socie-economic loss,
It is serious also to will lead to the major accidents such as a wide range of power-off, electric network from collapsing.
Gases Dissolved in Transformer Oil analyzes (dissolved gas anal-ysis, abbreviation DGA) method, passes through insulation
The difference of the characteristic gas such as hydrogen, methane, ethylene, acetylene and ethylene constituent content and gas production rate under different operating statuses in oil
Different analysis provides important evidence for running state of transformer assessment, and has the advantages that support electrification on-line checking, in China
Transformer's Condition Monitoring and fault diagnosis field are widely applied.
Based on practical experience and theory analysis, researcher establishes three-ratio method, Rogers ratio method and big in the early stage
The simple basic methods system of the processes such as triangulation method is defended, but by limitations such as coding missing, the absolute, inspections of the scene of a crime of threshold value, it is existing
It has been increasingly becoming the supplementary means of transformer fault diagnosis.
As machine Learning Theory develops, based on artificial intelligence method for diagnosing faults with it to running state of transformer type
Classify higher accuracy rate, becomes the research direction of educational circles's hot topic.However due to failure generation in transformer actual motion
Frequency is lower, and sample size gap is larger between collected fault type, final collected oil chromatography data fault case
Library is the lesser unbalanced dataset of scale.In the case where data nonbalance, support vector machines, decision tree, Bayesian network,
Most of model compared with based on such as deepness belief network is easy in the object procedure for maximizing all accuracys rate, is biased to most
The parameter of class sample updates and ignores the correct classification of minority class sample.
Currently, lack sampling, oversampler method and integrated learning approach, respectively in data sampling level and algorithm optimization layer
Face is widely used in alleviating the training problem of unbalanced dataset.It is inadequate in oil gas data fault case library total number of samples itself
In situation abundant, the mode that lack sampling algorithm reduces most class samples will lose the potentially useful information of part, therefore can not
Individually to be used in transformer fault diagnosis problem.
Based on this, it is expected that obtaining a kind of new Diagnosis Method of Transformer Faults, it can reduce fault case library data not
The adverse effect of balance, the generalization ability of each fault type of General Promotion.The Diagnosis Method of Transformer Faults can promoted entirely
In the case where body sample accuracy rate, the discrimination of the discrimination and most class samples that can also guarantee minority class sample simultaneously is horizontal
It is close, more acurrate and balance effect can be obtained relative to conventional method.
Summary of the invention
An object of the present invention is to provide a kind of Diagnosis Method of Transformer Faults for considering unbalanced data sample, the change
Depressor method for diagnosing faults can reduce the adverse effect of fault case library data nonbalance, each fault type of General Promotion it is general
Change ability.The Diagnosis Method of Transformer Faults is in terms of hierarchical classification and integrated study two to the fault diagnosis model of transformer
It is constructed, it is all being promoted by the information between balance classification using relationship constrained each other between multistage classification and in training
In the case where sample accuracy rate, while guaranteeing that the discrimination of minority class sample is close with the discrimination level of most class samples, because
And more acurrate and balance effect is achieved relative to conventional method.
According to foregoing invention purpose, the present invention proposes a kind of transformer fault diagnosis side for considering unbalanced data sample
Method, comprising steps of
100: acquiring the oil chromatography sample data of transformer, and it is pre-processed to obtain by pretreated oil chromatography
Sample data;
200: first order classifier is constructed and train, wherein using pretreated oil chromatography sample data is passed through to nerve net
The first order classifier of network building is trained;
State feature and oil chromatography sample data based on the output of first order classifier obtain Fusion Features vector;
300: the second level classifier using neural network building being carried out based on EasyEnsemble integrated learning approach
Training, in which:
According to the distributed number of the oil chromatography sample data of characterization each state of transformer, Fusion Features vector is divided into minority
Class sample and most class samples;Lack sampling wherein is carried out to reduce the quantity of most class samples to most class samples;
It is balancedly assigned in several subsets by minority class sample and by most class samples of lack sampling;
Several subsets are respectively corresponded to input in several second level sub-classifiers and are trained;
Output based on several second level sub-classifiers synthesizes second level classifier;
400: the transformer oil chromatographic data for needing to diagnose being inputted into first order classifier, then from the defeated of second level classifier
Outlet exports the transformer state of the transformer oil chromatographic data characterization.
In Diagnosis Method of Transformer Faults of the present invention, each oil chromatography sample data meeting when in order to reduce acquisition
There is numerical fluctuations influence, therefore, in step 100, oil chromatography sample data is pre-processed, is then classified by the first order
Device is trained to by pretreated oil chromatography sample data, and then the state feature based on the output of first order classifier
Fusion Features vector is obtained with oil chromatography sample data.In view of oil chromatography sample data is DAG structure, there are leaf nodes
Therefore the case where possessing multiple father nodes, using first order classifier and is based on EasyEnsemble integrated learning approach pair
The second level classifier constructed using neural network, so that can establish several second level in the classifier of the second level
Classifier, while can also be by first order classifier institute to undertake the decision strategies of all node labels of second level sub-classifier
Input of the Fusion Features vector as second level sub-classifier is obtained, to provide drawing for each node relationships in the sub-classifier of the second level
Information is led, the output synthesis second level classifier of several second level sub-classifiers is based ultimately upon.The transformer that needs are diagnosed
Oil colours modal data inputs first order classifier, then exports the transformer oil chromatographic data characterization from the output end of second level classifier
Transformer state.
It is based on EasyEnsemble integrated learning approach due to using in technical solutions according to the invention, using it to more
Several classes of samples carry out multiple subsets of lack sampling construction balance and the second level sub-classifier of integrated differentiation as strong classification
Device, sufficiently to excavate the potentially useful information of unbalanced data concentration.And due to passing through in technical solutions according to the invention
The subset that lack sampling generates multiple data balancings carries out parallel training, is completed in each second level sub-classifier to minority
The guidance of class sample, then the output of second level sub-classifier is synthesized into final second level classifier, so as to avoid owing to adopt
Sample loses the problem of data information.
In order to further explain how technical solutions according to the invention are based on EasyEnsemble integrated learning approach
Second level classifier using neural network building is trained, to give the multi-class classification problem of K class, remembers uneven number
According to collection classification C1,C2,...,CKMiddle number of samples it is least it is a kind of be minority class sample set P, number is | P |, remaining most class sample
This collection is N1,N2,...,NK-1。
So, all most class sample sets are carried out repeating to have the random independent lack sampling put back to T times, then each most classes
Sample set NiIt can produce several subset Nit, and meet: | Nit|=| P |.
Most class sample set N that each lack sampling is obtained againt, andMost class samples
This subset NtEach second level sub-classifier h is combined into minority class sample set P grouptSubset Dt, and Dt=Nt∪P.At this point, subset
DtEach class number is identical, is equilibrium data collection.
Respectively with subset DtAs input to each second level sub-classifier htCarrying out number using AdaBoost algorithm is si
Repetitive exercise, export shown in following formula:
Wherein, y is the true tag for inputting x, htdIt (x) is output of training the number of iterations when being d, εtFor Weak Classifier ht
Error function, k ∈ { 1,2 ... K } is C1,C2,...,CKSubscript.
It should be noted that the second level classifier based on EasyEnsemble integrated learning approach is to all second level
Sub-classifier htParameter integrated, rather than to the output h of second level sub-classifiert(x) result ballot is carried out to obtain finally
Decision, that is to say, that the output H (x) of second level classifier can be expressed as being shown below:
It can thus be seen that the Diagnosis Method of Transformer Faults of consideration unbalanced data sample of the present invention can subtract
The adverse effect of glitch case library data nonbalance, the generalization ability of each fault type of General Promotion.The transformer fault is examined
Disconnected method constructs the fault diagnosis model of transformer in terms of hierarchical classification and integrated study two, by utilizing multistage
Relationship constrained each other and the information between balance classification in training between classification, in the case where promoting all sample accuracys rate, together
When guarantee that the discrimination of minority class sample is close with the discrimination level of majority class samples, thus achieved relative to conventional method
More acurrate and balance effect.
Further, in Diagnosis Method of Transformer Faults of the present invention, the output of first order classifier, which characterizes, to be become
The rough segmentation state of depressor, the finely divided state of the output characterization transformer of second level classifier.
In above scheme, rough segmentation state can " Gases Dissolved in Transformer Oil is analyzed and judgement according to DL/T 722-2014
Directive/guide " Preliminary division is carried out to the state of transformer, it then can be according to the difference or hot spot temperature of such as discharge energy density
Degree it is different by rough segmentation state further division.
Further, in Diagnosis Method of Transformer Faults of the present invention, rough segmentation state is included at least: normal,
Discharge fault, overheating fault;Finely divided state includes at least: normal, shelf depreciation, low energy electric discharge, high-energy discharge, low energy electric discharge are simultaneous
Overheat, high-energy discharge and overheat, cryogenic overheating, medium temperature overheat, hyperthermia and superheating.
It should be noted that some compound failures in rough segmentation state its can have both two kinds of rough segmentation states, such as put
Electric and overheat composite type failure, then it is not only discharge fault under rough segmentation state, but also is overheating fault.
Further, in Diagnosis Method of Transformer Faults of the present invention, using convolutional neural networks building first
Grade classifier and/or second level classifier.
Further, in Diagnosis Method of Transformer Faults of the present invention, oil chromatography sample data is located in advance
Reason includes at least normalized.
Correspondingly, another object of the present invention is to provide a kind of transformer fault diagnosis for considering unbalanced data sample
System can obtain more acurrate and balance effect by the transformer fault diagnosis system relative to conventional method.
According to foregoing invention purpose, the present invention proposes a kind of transformer fault diagnosis system for considering unbalanced data sample
In, comprising:
Data acquisition device acquires the oil chromatography sample data of transformer;
Data processing equipment is configured to carry out operations described below:
Oil chromatography sample data is pre-processed;
It constructs and trains first order classifier, in which: use by pretreated oil chromatography sample data to neural network
The first order classifier of building is trained;It is then based on the state feature and oil chromatography sample data of the output of first order classifier
Obtain Fusion Features vector;
The second level classifier using neural network building is trained based on EasyEnsemble integrated learning approach,
Wherein: according to the distributed number of the oil chromatography sample data of characterization each state of transformer, Fusion Features vector being divided into minority class
Sample and most class samples;Lack sampling wherein is carried out to reduce the quantity of most class samples to most class samples;By minority class sample
Originally it and by most class samples of lack sampling is balancedly assigned in several subsets;If several subsets are respectively corresponded input
It is trained in dry second level sub-classifier;Output based on several second level sub-classifiers synthesizes second level classifier;
When needing the state to transformer to diagnose, the transformer oil chromatographic data for needing to diagnose are inputted into the first order
Classifier then exports the transformer state of the transformer oil chromatographic data characterization from the output end of second level classifier.
Further, in transformer fault diagnosis system of the present invention, the output of first order classifier, which characterizes, to be become
The rough segmentation state of depressor, the finely divided state of the output characterization transformer of second level classifier.
Further, in transformer fault diagnosis system of the present invention, rough segmentation state is included at least: normal,
Discharge fault, overheating fault;Finely divided state includes at least: normal, shelf depreciation, low energy electric discharge, high-energy discharge, low energy electric discharge are simultaneous
Overheat, high-energy discharge and overheat, cryogenic overheating, medium temperature overheat, hyperthermia and superheating.
Further, in transformer fault diagnosis system of the present invention, using convolutional neural networks building first
Grade classifier and/or second level classifier.
Further, in transformer fault diagnosis system of the present invention, data processing equipment is to oil chromatography sample
Data carry out pretreatment and include at least normalized.
The Diagnosis Method of Transformer Faults and system of consideration unbalanced data sample of the present invention are compared to existing skill
Art have the advantages described below and the utility model has the advantages that
The Diagnosis Method of Transformer Faults of consideration unbalanced data sample of the present invention can reduce fault case library
The adverse effect of data nonbalance, the generalization ability of each fault type of General Promotion.The Diagnosis Method of Transformer Faults is from level
Two aspects of classification and integrated study construct the fault diagnosis model of transformer, by mutually making an appointment using between multistage classification
The relationship of beam and the information between balance classification in training, in the case where promoting all sample accuracys rate, while guaranteeing minority class
The discrimination of sample is close with the discrimination level of most class samples, thus achieves more acurrate relative to conventional method and balance
Effect.
In addition, transformer fault diagnosis system of the present invention similarly has the above advantages and beneficial effect.
Detailed description of the invention
Fig. 1 schematically shows that the Diagnosis Method of Transformer Faults of consideration unbalanced data sample of the present invention exists
The DAG structure of transformer state type in some embodiments.
Fig. 2 is the Diagnosis Method of Transformer Faults of consideration unbalanced data sample of the present invention in some embodiments
In classifier schematic illustration.
Fig. 3 is the Diagnosis Method of Transformer Faults of consideration unbalanced data sample of the present invention in some embodiments
In based on EasyEnsemble integrated learning approach to the original that is trained of second level classifier using neural network building
Manage schematic diagram.
Fig. 4 is the Diagnosis Method of Transformer Faults of consideration unbalanced data sample of the present invention in some embodiments
In workflow schematic diagram.
Fig. 5 is schematically showed using Diagnosis Method of Transformer Faults of the present invention and other prior arts
The diagnostic result situation that finally obtains of diagnostic method.
Specific embodiment
Below in conjunction with Figure of description and specific embodiment to consideration unbalanced data sample of the present invention
Diagnosis Method of Transformer Faults and system make further explanation, however the explanation and illustration is not to skill of the invention
The improper restriction of art forecast scheme configuration.
Fig. 1 schematically shows that the Diagnosis Method of Transformer Faults of consideration unbalanced data sample of the present invention exists
The DAG structure of transformer state type in some embodiments.
As shown in Figure 1, according to transformer in DL/T 722-2014 " directive/guide is analyzed and judged to Gases Dissolved in Transformer Oil "
The category difference of operating status, in technical solutions according to the invention, the output of first order classifier characterizes the thick of transformer
Isloation state includes: normal, discharge fault, overheating fault, then according to the different further division discharge faults of discharge energy density,
And the height further division overheating fault according to hot(test)-spot temperature, the subdivision of the output characterization transformer of final second level classifier
State includes: normal, the electric discharge of shelf depreciation, low energy, high-energy discharge, the simultaneous overheat of low energy electric discharge, the simultaneous overheat of high-energy discharge, low temperature mistake
Heat, medium temperature overheat, hyperthermia and superheating.
From figure 1 it appears that discharge and overheat composite type failure (such as low energy electric discharge and overheat and high-energy discharge it is simultaneous
Overheat) two kinds of rough segmentation states are then had both, it is not only discharge fault under rough segmentation state, but also be overheating fault.And this point and biography
The plane classification of system is different, and in traditional plane classification, each sample is pertaining only to a classification, and usually assumes that class
Any connection is not present between not, plane classifier needs classification belonging to disposable decision sample.However in actual classification
In problem, often it can continue to be subdivided into multiple subclasses in the presence of one or more classifications, or a father can be combined into
The case where class, the hierarchical relationship between this type are referred to as class hierarchy, and thus each sample can be each according to class hierarchy
Grade level has one or more class labels simultaneously, and first order classifier and second level classifier used by this case
Classification closer to transformer actual conditions.
In addition, it can be seen from the DAG structure of the transformer state type of Fig. 1 its there are leaf nodes to possess multiple fathers
The case where node, therefore, using first order classifier and based on EasyEnsemble integrated learning approach to using nerve net
The second level classifier of network building, so that can establish several second level sub-classifiers in the classifier of the second level to hold
The decision strategy of all node labels of second level sub-classifier is carried on a shoulder pole, while feature obtained by first order classifier can also be melted
Input of the resultant vector as second level sub-classifier, to provide the guidance information of each node relationships in the sub-classifier of the second level, most
The output based on several second level sub-classifiers synthesizes second level classifier eventually.The transformer oil chromatographic data that needs are diagnosed
First order classifier is inputted, then exports the transformer shape of the transformer oil chromatographic data characterization from the output end of second level classifier
State.
It is illustrated in conjunction with principle of the Fig. 2 to above-mentioned classifier.Fig. 2 is consideration unbalanced data sample of the present invention
Diagnosis Method of Transformer Faults classifier schematic illustration in some embodiments.
As shown in Fig. 2, obtaining collected oil chromatography sample data after pretreatment by pretreated oil chromatography
Sample data X, the first order classifier H that neural network will be constructed by pretreated oil chromatography sample data X1It is instructed
Practice, first order classifier H1The state feature of outputThe rough segmentation state for characterizing transformer is exported based on first order classifier H1
State featureFusion Features vector is obtained with oil chromatography sample data X, is then based on the integrated study side EasyEnsemble
Method is to the second level classifier H using neural network building2It is trained, second level classifier H2The state feature of outputTable
Levy the finely divided state of transformer.
The second level classifier using neural network building is trained based on EasyEnsemble integrated learning approach
Process can be illustrated in conjunction with Fig. 3.Fig. 3 is the transformer fault diagnosis of consideration unbalanced data sample of the present invention
The second level that use neural network is constructed based on EasyEnsemble integrated learning approach of method in some embodiments
The schematic illustration that classifier is trained.
As shown in figure 3, since the oil chromatography sample data of characterization each state of transformer is unbalanced dataset, root
According to the distributed number of the oil chromatography sample data of characterization each state of transformer, Fusion Features vector is divided into minority class sample and more
Several classes of samples, wherein carry out lack sampling to most class samples to reduce the quantity of most class samples, then by minority class sample and
It is balancedly assigned in several subsets by most class samples of lack sampling, several subsets is respectively corresponded and input several
It is trained in the sub-classifier of the second level;Output based on several second level sub-classifiers synthesizes second level classifier H2。
In some more preferably embodiments, when transformer fault diagnosis system is worked, data acquisition device,
Its oil chromatography sample data for acquiring transformer;And data processing equipment, it is configured to carry out operations described below: to oil chromatography sample
Notebook data is pre-processed;It constructs and trains first order classifier, in which: using by pretreated oil chromatography sample data pair
The first order classifier of neural network building is trained;It is then based on the state feature and oil chromatography of the output of first order classifier
Sample data obtains Fusion Features vector;Based on EasyEnsemble integrated learning approach to using the second of neural network building
Grade classifier is trained, in which: according to the distributed number of the oil chromatography sample data of characterization each state of transformer, feature is melted
Resultant vector is divided into minority class sample and most class samples;Lack sampling wherein is carried out to reduce most class samples to most class samples
Quantity;It is balancedly assigned in several subsets by minority class sample and by most class samples of lack sampling;By several height
Collection is respectively corresponded to input in several second level sub-classifiers and is trained;Output based on several second level sub-classifiers is closed
At second level classifier;It is when needing the state to transformer to diagnose, the transformer oil chromatographic data for needing to diagnose are defeated
Enter first order classifier, then exports the transformer shape of the transformer oil chromatographic data characterization from the output end of second level classifier
State.
Certainly it should be noted that data acquisition can be obtained by the data in existing case library, can also pass through
Data collector acquisition obtains.
In some more preferably embodiments, transformer fault diagnosis is carried out based on dissolved gas analysis, is made
Use this 5 kinds of gases of hydrogen, methane, ethylene, acetylene, ethane as characteristic gas.In order to reduce each characteristic gas concentration in difference
The influence that absolute figure fluctuates in case, using the method for normalizing of following formula to each original gas sampleIt is handled.
Wherein, j ∈ { 1,2,3,4,5 } respectively indicates 5 kinds of gases, and i=1,2 ..., M indicate sample serial number.After normalizing
Gas data be denoted as:
According to the taxonomic structure of the transformer state type indicated in Fig. 1, the first order is carried out according to this class hierarchy respectively
Rough segmentation state in classifier includes 3 kinds of first order coding Y of normal, discharge fault and overheating fault1And second level classifier
In finely divided state include 9 kinds subdivision the second level coding Y2, as shown in table 1.
Table 1.
Status Type | First order coding | Second level coding |
Normally | 1,0,0 | 0:1,0,0,0,0,0,0,0,0 |
Shelf depreciation | 0,1,0 | 1:0,1,0,0,0,0,0,0,0 |
Low energy electric discharge | 0,1,0 | 2:0,0,1,0,0,0,0,0,0 |
High-energy discharge | 0,1,0 | 3:0,0,0,1,0,0,0,0,0 |
Cryogenic overheating | 0,0,1 | 4:0,0,0,0,1,0,0,0,0 |
Medium temperature overheat | 0,0,1 | 5:0,0,0,0,0,1,0,0,0 |
Hyperthermia and superheating | 0,0,1 | 6:0,0,0,0,0,0,1,0,0 |
Low energy electric discharge and overheat | 0,1,1 | 7:0,0,0,0,0,0,0,1,0 |
High-energy discharge and overheat | 0,1,1 | 8:0,0,0,0,0,0,0,0,1 |
On the disaggregated model of running state of transformer as shown in Figure 1, establishes and be based on hierarchical classification and EasyEnsemble
The Fault Diagnosis Model for Power Transformer of integrated study establishes corresponding classifier with the uneven grading of every grade of classification sample
(i.e. first order classifier and second level classifier used by the Diagnosis Method of Transformer Faults of this case), deeply carries out step by step
Diagnosis.
In first order classifier, normal, discharge fault and overheating fault three classes number of samples are than for 652:973:768's
Sample size, sample size relative equilibrium.Since complex class failure exists simultaneously overheating fault in first order classifier and puts
Two labels of electric fault, therefore, the reality of first order classifier processing are multi-tag classification problem.When considering to shorten model training
Between, first order classifier can solve multi-tag classification problem using neural network, and take the mode of Fusion Features
Reinforce the guidance of first order label in secondary classifier.
In first order classifier, first order classifier H is used in order to realize1It realizes multi-tag classification problem, uses
Sigmoid activation primitive is directed to network output valve H1(X) carry out Interval Maps, then to label position by rank get over function progress about
The size of threshold value ρ compares, and judges whether it belongs to Status Type.The final output label of first order classifierFor following formula
In formula,ρ is default classification thresholds.
And gradient is carried out using the loss function under formula and calculates update first order classifier H1Associated weight bias variable:
Loss1=-(Y1logσ1(H1(X))+(1-Y1)log(1-σ1(H1(X))
For the final label for using first order classifier to exportTo second in more careful second level classifier
The supervision guiding function of grade labeling by the output feature F of the first order and is originally inputted oil chromatography gas data X progress feature
Fusion, together as the input of second level classifier.
X2=[X, F]
Wherein, in the training process, to first order classifier output labelFrom a small number of different samples of physical tags, use
Y1As feature, to reduce the input deviation of second level classifier.Remaining output label situation identical as physical tags uses
Sigmoid is that the network of activation primitive exports, such as above formulafiShown in ∈ F.In test process
In, first order classification error does not use physical tags to substitute, and is used uniformly the network that sigmoid is activation primitive and exports.
In the classifier of the second level, when using EasyEnsemble integrated learning approach, X is inputted to the second level first2Structure
The training subset for making data balancing, using sample size it is least as minority class sample (such as with high-energy discharge and mistake in table 2
Heat is minority class sample), remaining classification carries out lack sampling as most class samples and obtains T subset
T second level sub-classifier h of parallel training2,t, the more of transformer state are completed using softmax as activation primitive
Class probability calculates, the Status Type with the class label k ∈ { 0,1,2,3,4,5,6,7,8 } of maximum probability as diagnosis.And
Loss function, which is calculated, with cross entropy updates sub-classifier parameter:
In formula,
It adjusts second level sub-classifier weight and offset parameter after the completion of joining according to default αtWeighting merges into final the
Secondary classifier H2, final output are as follows:
In order to measure performance of the classifier in notebook data imbalance example comprehensively, a variety of evaluations are used in subsequent analysis
Mode: the accurate rate of each classification and recall rate in test set are calculated and is being lacked from the angle estimator classifier of precision ratio and recall ratio
Generalization Capability on several classes of samples;Calculate test set entirety sample accuracy rate and Receiver Operating Characteristics (receiver
Operating characteristic, ROC) curve and ROC curve surround the area (Area Under ROC Curve, AUC)
It is assessed in terms of the discrimination and overall generalisation properties on all samples.
Fig. 4 illustrates the Diagnosis Method of Transformer Faults of discussed above unbalanced data sample in some embodiments
Workflow schematic diagram.
In order to verify this case transformer fault diagnosis mode implementation result, comprehensive grid company fault case library data
Data in literature composition data collection has been delivered, 80% and 20% sample is used to train and be tested respectively, each fault case
Sample size is as shown in table 2.Table 2 shows sample distribution amount.
Table 2.
Status Type | Total number of samples | Number of training | Test sample number |
Normally | 652 | 519 | 133 |
Shelf depreciation | 161 | 129 | 32 |
Low energy electric discharge | 216 | 173 | 43 |
High-energy discharge | 460 | 366 | 94 |
Cryogenic overheating | 119 | 95 | 24 |
Medium temperature overheat | 203 | 162 | 41 |
Hyperthermia and superheating | 310 | 247 | 63 |
Low energy electric discharge and overheat | 90 | 72 | 18 |
High-energy discharge and overheat | 46 | 37 | 9 |
It amounts to | 2257 | 1800 | 457 |
As can be seen from Table 2, high-energy discharge fault sample number is up to the 10 of high-energy discharge and overheating fault number of samples
Times, low energy electric discharge is simultaneous to be overheated, and the type faults such as cryogenic overheating sample size is also more rare.That is, further to event
When hindering type subdivision, second level label has serious disequilibrium.Therefore it uses and is based in second level classifier
The method of EasyEnsemble integrated study carries out more classification, is carried out by the training subset that lack sampling generates multiple data balancings
Parallel training, to alleviate the weak problem of data nonbalance bring minority class sample classification.
It is analyzed using the transformer state type sample data in table 2 as example, using 5 kinds of gas concentrations as defeated
Enter, 9 in the classifier of the second level kind Status Type is as the tag along sort finally judged.
When carrying out first order classification, using the neural network of three layers of hidden layer as classifier, input layer, hidden layer,
Output layer neuron number is respectively 5, [15,30,15], 3, and hidden layer uses tanh activation primitive, learning rate 0.001, training
Period 50000 times.In first order classification, normal, discharge fault, overheating fault accuracy rate and classification preset threshold ρ relationship such as table 3
It is shown.
Table 3 shows the first order classification accuracy under different ρ.
Table 3.
ρ | 0.5 | 0.55 | 0.6 | 0.65 | 0.7 | 0.75 |
Training set | 97. | 98.0 | 97.8 | 97.6 | 97.6 | 97.5 |
Test set | 97. | 97.3 | 97.5 | 96.9 | 97.1 | 96.5 |
Since the incorrect sample of first order classification can be according to formula in the training process
fiThe case where ∈ F is substituted by physical tags, can preferably use test set accuracy rate highest, i.e. setting ρ=0.6 carries out feature
Fusion, then carry out second level classifier training.
Possess the layer-by-layer progressive effect deeply classified to verify hierarchy model relative to plane disaggregated model: i.e. the
Guiding function is played in the classification of the second level in Fusion Features part in first-level class device, by the method for this case and omits first order classification
Device directly uses the model (hereinafter referred to as reference method 1) of EasyEnsemble as control;Increase first order classifier
BPNN (hereinafter referred to as reference method 2) is used as with simple BP neural network (hereinafter referred to as reference method 3) and compares, referring in particular to
Model setting is tested as shown in table 4.It should be noted that wherein BPNN is using the second level sub-classifier phase with this case
Same neuron number is that the hidden layer of [30,50,30] is arranged, learning rate 0.001, cycle of training 50000 times, second level point
Class device radix T is set as 25.Diagnostic result ROC curve, accuracy rate and the AUC value such as Fig. 5 of 2 groups of control classifiers to test set
With shown in table 4.
Table 4 shows classification accuracy and AUC value under distinct methods
Table 4.
This case method | Reference method 1 | Reference method 2 | Reference method 3 | |
First order classifier | BPNN | Nothing | BPNN | Nothing |
Second level classifier | EasyEnsemble | EasyEnsemble | BPNN | BPNN |
Test set accuracy rate % | 90.37 | 88.62 | 87.31 | 83.59 |
Test set AUC value | 0.9612 | 0.9555 | 0.9281 | 0.8742 |
By result of the control group (this case method and reference method 1, reference method 2 and reference method 3) in Fig. 5 it is found that
By first order classifier Fusion Features, the area that the ROC curve of two groups of second level classifiers surrounds is increased, wherein with
BPNN has reached 0.05 amplification as the reference method AUC value of second level classifier, and accuracy rate also has nearly 4% growth.Though
0.005 is only so improved using the AUC value of EasyEnsemble as the reference method of second level classifier, but its method is in standard
1.7% is increased in true rate, as shown in table 4.In conclusion first order classifier output label and input gas are carried out feature
After fusion, subsequent classification can be played on class relations using the next stage stratigraphic classification information contained in its broad sense level
Effect of contraction enables fault diagnosis effect obtain promotion with this.
Fig. 5 is schematically showed using Diagnosis Method of Transformer Faults of the present invention and other prior arts
The diagnostic result situation that finally obtains of diagnostic method.
To examine sub-classifier number T and sub-classifier foundation structure neural network number of plies L to the shadow of model accuracy
It rings, chooses several groups of empirical parameters and fault diagnosis result accuracy rate is compared, comparing result is as shown in table 5, with T=25, L
=3 be the parameter setting of final mask.Table 5 shows classification accuracy under different parameters.
Table 5.
Accuracy rate | L=2 | L=3 | L=4 | L=5 | L=6 |
Τ=10 | 71.77 | 74.40 | 76.59 | 78.12 | 76.37 |
T=25 | 87.31 | 90.37 | 89.28 | 88.18 | 88.62 |
T=40 | 87.09 | 89.72 | 86.43 | 87.75 | 87.75 |
The validity of accuracy rate is promoted to verify EasyEnsemble integrated study on processing unbalanced data, ditto
2 groups of control groups of hierarchical classification verification setting (this case method and reference method 2) and (reference method 1 and reference method 3) are stated, respectively
Using EasyEnsemble and isostructural BPNN as second level classifier.
It can be obtained according to the ROC curve in Fig. 5, using BPNN as the reference method 1 of second level classifier and reference method 3
Real rate be respectively sharply to decline near 0.12 and 0.25 in false positive rate, the spy with the perfect high real rate of classifier low false positive rate
Property deviates from.And still have 0.75 or more in the case where 0.02 low false positive rate using EasyEnsemble as the model of second level classifier
Real rate, more approach perfect classifier in (0,1) ideal point of ROC curve, its opposite control group is in AUC value and accuracy rate
On also have growth higher than 0.3 and 3% respectively, there are higher test set generalisation properties.
It should be noted that the Diagnosis Method of Transformer Faults in Fig. 5, shown in curve I to take this case invention described
Resulting ROC curve takes ginseng shown in curve III to take with reference to the resulting ROC curve of method 1 shown in curve II
The resulting ROC curve of test method 2 is taken with reference to the resulting ROC curve of method 3 shown in curve IV.
Choose BPNN, support vector machines (SVM), 3 kinds of common methods of random forest and transformer fault of the present invention
Diagnostic method compares, effect of the more different models on transformer fault diagnosis.Wherein BPNN and aforementioned setting are consistent;
SVM is using the radial base core letter (RBF) of selection;Random forest decision tree number is set as 5000;SVM and random forest exist simultaneously
Weighed value adjusting uneven between class is used when training.
Specific diagnostic result is as shown in table 6, and P (precision) indicates that rate of precision, R (recall) indicate recall rate in table,
0-8 is respectively normal, the electric discharge of shelf depreciation, low energy, high-energy discharge, cryogenic overheating, medium temperature overheat, hyperthermia and superheating, low energy electric discharge
And it overheats and high-energy discharge and superheat state type coding.Table 6 shows test set diagnostic result.
Table 6.
As can be seen from Table 6, in terms of holistic diagnosis accuracy rate, this case method has highest performance in 4 kinds of methods.It is instructing
During white silk, BPNN, SVM and 3 kinds of models of random forest are only exercised supervision study with 9 kinds of labels of careful classification, and this case exists
First order label is added on the basis of the label of the second level to be trained, is known in accuracy rate 7% promotion better than the second height is possessed
The not BPNN classifier of rate.As it can be seen that this case method extracts first hidden in the label of the second level by establishing hierarchy model
Grade tag along sort information and the input of original gas concentration carry out Fusion Features, and it is deeper to advance the second level for guidance to a certain extent
Enter and careful classification.
From minority class sample generalization, the distributed area of this case method rate of precision and recall rate on 9 kinds of Status Types
Between it is relatively stable, be not less than 60%, achieve relatively stable table looking into standard and looking into other opposite 3 kinds of models of complete two aspect
It is existing.Especially on minority class sample type high-energy discharge and superheat state, BPNN is not due to having using any for imbalance
The technology of data training, it is rebasing with the recall rate of 21.1% rate of precision and 44.4%.SVM uses uneven weight tune between class
It is whole, the otherness showed between most class samples and minority class sample class is reduced slightly, but its effect is still undesirable.Although
Random forest is because uneven between the multifarious method of parameter and class between itself Bagging integrated study being used to utilize Weak Classifier
Weighed value adjusting, so that minority class generalisation properties are enhanced, but the training set still pole that its Bagging random sampling mechanism obtains can
It can be unbalanced dataset, it is even more uneven, so that effect is weaker than the EasyEnsemble collection for guaranteeing training subset balance
At learning method.
In summary as can be seen that the Diagnosis Method of Transformer Faults of consideration unbalanced data sample of the present invention can
To reduce the adverse effect of fault case library data nonbalance, the generalization ability of each fault type of General Promotion.Transformer event
Barrier diagnostic method constructs the fault diagnosis model of transformer in terms of hierarchical classification and integrated study two, passes through utilization
Relationship constrained each other and the information between balance classification in training between multistage classification, the case where promoting all sample accuracys rate
Under, while guaranteeing that the discrimination of minority class sample is close with the discrimination level of most class samples, thus relative to conventional method
Achieve more acurrate and balance effect.
In addition, transformer fault diagnosis system of the present invention similarly has the above advantages and beneficial effect.
It should be noted that prior art part is not limited to given by present specification in protection scope of the present invention
Embodiment, all prior arts not contradicted with the solution of the present invention, including but not limited to first patent document, formerly
Public publication, formerly openly use etc., it can all be included in protection scope of the present invention.
In addition, in this case in the combination of each technical characteristic and unlimited this case claim documented combination or
It is combination documented by specific embodiment, all technical characteristics that this case is recorded can be freely combined in any way
Or combine, unless generating contradiction between each other.
It is also to be noted that embodiment enumerated above is only specific embodiments of the present invention.The obvious present invention is not
Above embodiments are confined to, the similar variation or deformation made therewith are that those skilled in the art can be from present disclosure
It immediately arrives at or is easy to just to associate, be within the scope of protection of the invention.
Claims (10)
1. a kind of Diagnosis Method of Transformer Faults for considering unbalanced data sample, which is characterized in that comprising steps of
100: acquiring the oil chromatography sample data of transformer, and it is pre-processed to obtain by pretreated oil chromatography sample
Data;
200: first order classifier is constructed and train, wherein using pretreated oil chromatography sample data is passed through to neural network structure
The first order classifier built is trained;
State feature and oil chromatography sample data based on the output of first order classifier obtain Fusion Features vector;
300: the second level classifier using neural network building is trained based on EasyEnsemble integrated learning approach,
Wherein:
According to the distributed number of the oil chromatography sample data of characterization each state of transformer, Fusion Features vector is divided into minority class sample
Sheet and most class samples;Lack sampling wherein is carried out to reduce the quantity of most class samples to most class samples;
It is balancedly assigned in several subsets by minority class sample and by most class samples of lack sampling;
Several subsets are respectively corresponded to input in several second level sub-classifiers and are trained;
Output based on several second level sub-classifiers synthesizes second level classifier;
400: the transformer oil chromatographic data for needing to diagnose being inputted into first order classifier, then from the output end of second level classifier
Export the transformer state of the transformer oil chromatographic data characterization.
2. Diagnosis Method of Transformer Faults as described in claim 1, which is characterized in that the output table of the first order classifier
Levy the rough segmentation state of transformer, the finely divided state of the output characterization transformer of the second level classifier.
3. Diagnosis Method of Transformer Faults as claimed in claim 2, which is characterized in that the rough segmentation state includes at least: just
Often, discharge fault, overheating fault;The finely divided state includes at least: normal, shelf depreciation, low energy electric discharge, high-energy discharge, low
It can electric discharge and overheat, high-energy discharge and overheat, cryogenic overheating, medium temperature overheat, hyperthermia and superheating.
4. Diagnosis Method of Transformer Faults as described in claim 1, which is characterized in that using convolutional neural networks building first
Grade classifier and/or second level classifier.
5. Diagnosis Method of Transformer Faults as described in claim 1, which is characterized in that located in advance to oil chromatography sample data
Reason includes at least normalized.
6. a kind of transformer fault diagnosis system for considering unbalanced data sample characterized by comprising
Data acquisition device acquires the oil chromatography sample data of transformer;
Data processing equipment is configured to carry out operations described below:
Oil chromatography sample data is pre-processed;
It constructs and trains first order classifier, in which: neural network is constructed using by pretreated oil chromatography sample data
First order classifier be trained;The state feature for being then based on the output of first order classifier is obtained with oil chromatography sample data
Fusion Features vector;
The second level classifier using neural network building is trained based on EasyEnsemble integrated learning approach,
In: according to the distributed number of the oil chromatography sample data of characterization each state of transformer, Fusion Features vector is divided into minority class sample
Sheet and most class samples;Lack sampling wherein is carried out to reduce the quantity of most class samples to most class samples;By minority class sample
It is balancedly assigned in several subsets with by most class samples of lack sampling;It is several that several subsets are respectively corresponded into input
It is trained in a second level sub-classifier;Output based on several second level sub-classifiers synthesizes second level classifier;
When needing the state to transformer to diagnose, the transformer oil chromatographic data input first order that will need to diagnose is classified
Device then exports the transformer state of the transformer oil chromatographic data characterization from the output end of second level classifier.
7. transformer fault diagnosis system as claimed in claim 6, which is characterized in that the output table of the first order classifier
Levy the rough segmentation state of transformer, the finely divided state of the output characterization transformer of the second level classifier.
8. transformer fault diagnosis system as claimed in claim 7, which is characterized in that the rough segmentation state includes at least: just
Often, discharge fault, overheating fault;The finely divided state includes at least: normal, shelf depreciation, low energy electric discharge, high-energy discharge, low
It can electric discharge and overheat, high-energy discharge and overheat, cryogenic overheating, medium temperature overheat, hyperthermia and superheating.
9. transformer fault diagnosis system as claimed in claim 6, which is characterized in that using convolutional neural networks building first
Grade classifier and/or second level classifier.
10. transformer fault diagnosis system as claimed in claim 6, which is characterized in that the data processing equipment is to oil colours
Spectrum sample data carries out pretreatment and includes at least normalized.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910451726.XA CN110133146A (en) | 2019-05-28 | 2019-05-28 | A kind of Diagnosis Method of Transformer Faults and system considering unbalanced data sample |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910451726.XA CN110133146A (en) | 2019-05-28 | 2019-05-28 | A kind of Diagnosis Method of Transformer Faults and system considering unbalanced data sample |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110133146A true CN110133146A (en) | 2019-08-16 |
Family
ID=67582303
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910451726.XA Pending CN110133146A (en) | 2019-05-28 | 2019-05-28 | A kind of Diagnosis Method of Transformer Faults and system considering unbalanced data sample |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110133146A (en) |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110609210A (en) * | 2019-09-25 | 2019-12-24 | 广东电网有限责任公司 | Power grid fault diagnosis and alarm method based on multi-source data fusion |
CN111507504A (en) * | 2020-03-18 | 2020-08-07 | 中国南方电网有限责任公司 | Adaboost integrated learning power grid fault diagnosis system and method based on data resampling |
CN111638427A (en) * | 2020-06-03 | 2020-09-08 | 西南交通大学 | Transformer fault detection method based on nuclear capsule neuron coverage |
CN112364995A (en) * | 2020-11-10 | 2021-02-12 | 重庆大学 | Transformer fault rate evaluation method and system and readable storage medium |
CN112434743A (en) * | 2020-11-27 | 2021-03-02 | 国网江苏省电力有限公司检修分公司 | Fault identification method based on GIL metal particle partial discharge time domain waveform image |
CN112748330A (en) * | 2020-12-24 | 2021-05-04 | 国网江苏省电力有限公司电力科学研究院 | High-voltage circuit breaker operating mechanism fault diagnosis method and device based on distributed integrated neural network |
CN112842342A (en) * | 2021-01-25 | 2021-05-28 | 北京航空航天大学 | Electrocardiogram and magnetic signal classification method combining Hilbert curve and integrated learning |
CN112904157A (en) * | 2021-01-19 | 2021-06-04 | 重庆邮电大学 | Fault arc detection method based on integrated machine learning |
CN113066540A (en) * | 2021-03-19 | 2021-07-02 | 新疆大学 | Method for preprocessing non-equilibrium fault sample of oil-immersed transformer |
CN113702728A (en) * | 2021-07-12 | 2021-11-26 | 广东工业大学 | Transformer fault diagnosis method and system based on combined sampling and LightGBM |
CN115476734A (en) * | 2022-09-20 | 2022-12-16 | 中国汽车工程研究院股份有限公司 | Fuzzy reasoning based water management fault diagnosis method for fuel cell system |
CN115792729A (en) * | 2022-11-30 | 2023-03-14 | 广东粤电科试验检测技术有限公司 | Transformer composite fault diagnosis method, device, equipment and storage medium |
CN117493980A (en) * | 2023-12-29 | 2024-02-02 | 合肥工业大学 | Bearing fault diagnosis method integrating feature extraction and sequencing causal discovery |
Citations (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103116111A (en) * | 2013-01-21 | 2013-05-22 | 上海市电力公司 | Method for diagnosing power transformer winding working condition |
CN105353255A (en) * | 2015-11-27 | 2016-02-24 | 南京邮电大学 | Transformer fault diagnosis method based on neural network |
CN106093612A (en) * | 2016-05-26 | 2016-11-09 | 国网江苏省电力公司电力科学研究院 | A kind of method for diagnosing fault of power transformer |
CN107886135A (en) * | 2017-12-01 | 2018-04-06 | 江苏蓝深远望科技股份有限公司 | A kind of parallel random forests algorithm for handling uneven big data |
US10024904B2 (en) * | 2014-03-27 | 2018-07-17 | International Business Machines Corporation | Method of characterizing and modeling leakage statistics and threshold voltage for ensemble devices |
CN108345904A (en) * | 2018-01-26 | 2018-07-31 | 华南理工大学 | A kind of Ensemble Learning Algorithms of the unbalanced data based on the sampling of random susceptibility |
CN108540451A (en) * | 2018-03-13 | 2018-09-14 | 北京理工大学 | A method of classification and Detection being carried out to attack with machine learning techniques |
CN108596199A (en) * | 2017-12-29 | 2018-09-28 | 北京交通大学 | Unbalanced data classification method based on EasyEnsemble algorithms and SMOTE algorithms |
CN108985342A (en) * | 2018-06-22 | 2018-12-11 | 华南理工大学 | A kind of uneven classification method based on depth enhancing study |
CA2981036A1 (en) * | 2017-08-21 | 2019-02-21 | Delta-X Research Inc. | Method and system for assessment of fault severity, risk exposure, and gassing status for liquid-filled high-voltage apparatus |
WO2019067960A1 (en) * | 2017-09-28 | 2019-04-04 | D5Ai Llc | Aggressive development with cooperative generators |
-
2019
- 2019-05-28 CN CN201910451726.XA patent/CN110133146A/en active Pending
Patent Citations (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103116111A (en) * | 2013-01-21 | 2013-05-22 | 上海市电力公司 | Method for diagnosing power transformer winding working condition |
US10024904B2 (en) * | 2014-03-27 | 2018-07-17 | International Business Machines Corporation | Method of characterizing and modeling leakage statistics and threshold voltage for ensemble devices |
CN105353255A (en) * | 2015-11-27 | 2016-02-24 | 南京邮电大学 | Transformer fault diagnosis method based on neural network |
CN106093612A (en) * | 2016-05-26 | 2016-11-09 | 国网江苏省电力公司电力科学研究院 | A kind of method for diagnosing fault of power transformer |
CA2981036A1 (en) * | 2017-08-21 | 2019-02-21 | Delta-X Research Inc. | Method and system for assessment of fault severity, risk exposure, and gassing status for liquid-filled high-voltage apparatus |
WO2019067960A1 (en) * | 2017-09-28 | 2019-04-04 | D5Ai Llc | Aggressive development with cooperative generators |
CN107886135A (en) * | 2017-12-01 | 2018-04-06 | 江苏蓝深远望科技股份有限公司 | A kind of parallel random forests algorithm for handling uneven big data |
CN108596199A (en) * | 2017-12-29 | 2018-09-28 | 北京交通大学 | Unbalanced data classification method based on EasyEnsemble algorithms and SMOTE algorithms |
CN108345904A (en) * | 2018-01-26 | 2018-07-31 | 华南理工大学 | A kind of Ensemble Learning Algorithms of the unbalanced data based on the sampling of random susceptibility |
CN108540451A (en) * | 2018-03-13 | 2018-09-14 | 北京理工大学 | A method of classification and Detection being carried out to attack with machine learning techniques |
CN108985342A (en) * | 2018-06-22 | 2018-12-11 | 华南理工大学 | A kind of uneven classification method based on depth enhancing study |
Non-Patent Citations (1)
Title |
---|
李倩倩等: "多类类别不平衡学习算法: EasyEnsemble.M*", 《模式识别与人工智能》 * |
Cited By (20)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110609210B (en) * | 2019-09-25 | 2020-09-25 | 广东电网有限责任公司 | Power grid fault diagnosis and alarm method based on multi-source data fusion |
CN110609210A (en) * | 2019-09-25 | 2019-12-24 | 广东电网有限责任公司 | Power grid fault diagnosis and alarm method based on multi-source data fusion |
CN111507504A (en) * | 2020-03-18 | 2020-08-07 | 中国南方电网有限责任公司 | Adaboost integrated learning power grid fault diagnosis system and method based on data resampling |
CN111638427A (en) * | 2020-06-03 | 2020-09-08 | 西南交通大学 | Transformer fault detection method based on nuclear capsule neuron coverage |
CN111638427B (en) * | 2020-06-03 | 2021-05-28 | 西南交通大学 | Transformer fault detection method based on nuclear capsule neuron coverage |
CN112364995A (en) * | 2020-11-10 | 2021-02-12 | 重庆大学 | Transformer fault rate evaluation method and system and readable storage medium |
CN112364995B (en) * | 2020-11-10 | 2024-05-03 | 重庆大学 | Transformer fault rate assessment method and system and readable storage medium |
CN112434743A (en) * | 2020-11-27 | 2021-03-02 | 国网江苏省电力有限公司检修分公司 | Fault identification method based on GIL metal particle partial discharge time domain waveform image |
CN112748330A (en) * | 2020-12-24 | 2021-05-04 | 国网江苏省电力有限公司电力科学研究院 | High-voltage circuit breaker operating mechanism fault diagnosis method and device based on distributed integrated neural network |
CN112904157A (en) * | 2021-01-19 | 2021-06-04 | 重庆邮电大学 | Fault arc detection method based on integrated machine learning |
CN112842342B (en) * | 2021-01-25 | 2022-03-29 | 北京航空航天大学 | Electrocardiogram and magnetic signal classification method combining Hilbert curve and integrated learning |
CN112842342A (en) * | 2021-01-25 | 2021-05-28 | 北京航空航天大学 | Electrocardiogram and magnetic signal classification method combining Hilbert curve and integrated learning |
CN113066540A (en) * | 2021-03-19 | 2021-07-02 | 新疆大学 | Method for preprocessing non-equilibrium fault sample of oil-immersed transformer |
CN113066540B (en) * | 2021-03-19 | 2023-04-11 | 新疆大学 | Method for preprocessing non-equilibrium fault sample of oil-immersed transformer |
CN113702728A (en) * | 2021-07-12 | 2021-11-26 | 广东工业大学 | Transformer fault diagnosis method and system based on combined sampling and LightGBM |
CN115476734A (en) * | 2022-09-20 | 2022-12-16 | 中国汽车工程研究院股份有限公司 | Fuzzy reasoning based water management fault diagnosis method for fuel cell system |
CN115476734B (en) * | 2022-09-20 | 2023-05-23 | 中国汽车工程研究院股份有限公司 | Fuzzy reasoning-based fuel cell system water management fault diagnosis method |
CN115792729A (en) * | 2022-11-30 | 2023-03-14 | 广东粤电科试验检测技术有限公司 | Transformer composite fault diagnosis method, device, equipment and storage medium |
CN117493980A (en) * | 2023-12-29 | 2024-02-02 | 合肥工业大学 | Bearing fault diagnosis method integrating feature extraction and sequencing causal discovery |
CN117493980B (en) * | 2023-12-29 | 2024-03-19 | 合肥工业大学 | Bearing fault diagnosis method integrating feature extraction and sequencing causal discovery |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110133146A (en) | A kind of Diagnosis Method of Transformer Faults and system considering unbalanced data sample | |
CN106093612B (en) | A kind of method for diagnosing fault of power transformer | |
CN102289682B (en) | Transformer fault diagnosis method based on integrated learning Bagging algorithm | |
CN105930901B (en) | A kind of Diagnosis Method of Transformer Faults based on RBPNN | |
CN110579709B (en) | Fault diagnosis method for proton exchange membrane fuel cell for tramcar | |
CN107992880A (en) | A kind of optimal lump classification method for diagnosing faults of power transformer | |
Malik et al. | Fuzzy reinforcement learning based intelligent classifier for power transformer faults | |
CN109102032A (en) | A kind of pumping plant unit diagnostic method based on depth forest and oneself coding | |
CN109142946A (en) | Transformer fault detection method based on ant group algorithm optimization random forest | |
CN105135591B (en) | Train air conditioning unit method for diagnosing faults based on more classification policys | |
CN104535865A (en) | Comprehensive diagnosing method for operation troubles of power transformer based on multiple parameters | |
CN110542819A (en) | transformer fault type diagnosis method based on semi-supervised DBNC | |
CN107958292A (en) | Transformer based on cost sensitive learning obscures careful reasoning method for diagnosing faults | |
CN108875783A (en) | A kind of extreme learning machine Diagnosis Method of Transformer Faults towards unbalanced dataset | |
CN107656152B (en) | One kind being based on GA-SVM-BP Diagnosis Method of Transformer Faults | |
CN110879373B (en) | Oil-immersed transformer fault diagnosis method with neural network and decision fusion | |
CN110059714A (en) | Diagnosis Method of Transformer Faults based on multi-category support vector machines | |
CN107153845A (en) | A kind of isolated island detection method of the adaptive grid-connected photovoltaic system based on machine learning | |
CN111275204B (en) | Transformer state identification method based on hybrid sampling and ensemble learning | |
CN107884663A (en) | A kind of Diagnosis Method of Transformer Faults based on combination core Method Using Relevance Vector Machine | |
CN111737907A (en) | Transformer fault diagnosis method and device based on deep learning and DGA | |
CN110569888A (en) | transformer fault diagnosis method and device based on directed acyclic graph support vector machine | |
CN114266297A (en) | Semantic knowledge base of thermal power equipment, construction method and zero sample fault diagnosis method | |
Jian et al. | Transformer fault diagnosis model with unbalanced samples based on SMOTE algorithm and focal loss | |
CN116561573A (en) | Training set circulation splitting training method for voltage stability evaluation |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
WD01 | Invention patent application deemed withdrawn after publication | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190816 |