CN101701940A - On-line transformer fault diagnosis method based on SVM and DGA - Google Patents

Abstract

The invention discloses an on-line transformer fault diagnosis method based on support vector machines (SVM) and dissolved gas analysis (DGA), which belongs to the field of transformer fault diagnosis. The method comprises the following steps of: acquiring the concentration of fault feature gases including H2, CH4, C2H2, C2H4 and C2H6 in transformer insulating oil by a gas chromatographic analysis method, normalizing the data through a preprocessing system, sending the data to a classification diagnosis system assembled by an integrative learning method and formed by six SVM classifiers according to a decision process, classifying the measurement data through calculation, judging the running state of the transformer, and at last outputting the diagnosis results. In the method, the support vector machines (SVM) of the artificial intelligence technology are adopted to analyze the gases in the oil, and the relation between gas components and the running state of the transformer can be objectively and essentially reflected, thereby the accuracy of fault diagnosis is effectively improved.

Description

Transformer on-line fault diagnosis method based on SVM and DGA

Technical field

The present invention relates to a kind of transformer fault diagnosis method, relate in particular to a kind of transformer on-line fault diagnosis method, belong to the transformer fault diagnosis field based on SVM (Support Vector Machine) and DGA (Dissolved Gas Analysis).

Background technology

Safety, stable electric power system are the important foundations of the national economic development, are the necessary conditions of modern industrial society's harmonious development.Transformer is monitored the reliability that can guarantee electric power system effectively as the hinge equipment of electric power system in real time to it.

At present, power transformer interior fault diagnose effective method be the dissolved gas analysis method (Dissolved Gas Analysis, DGA).Most of power transformers adopt insulating oil that built-in system is dispelled the heat and insulate in the domestic electric system, and transformer is when being in different running status, its insulation oil dissolved gas component and concentration have nothing in common with each other, oil dissolved gas is sampled and analyzed, can infer the running status of transformer, can carry out fault diagnosis to transformer effectively.Yet, the transformer fault type is various, and has the nonlinear relationship of complexity between fault signature gas and the fault, and traditional IEC three-ratio method and nothing coding ratioing technigue are judged too simplification, and there is the ratio blind zone problem, often can not diagnoses accurately fault.Therefore, artificial intelligence technology is extensively introduced DGA, and the transformer fault type is discerned, and wherein support vector machine has caused concern widely.

(Support Vector Machine SVM) is a kind of intelligent classification device based on the minimum principle of structure risk to support vector machine.Compare artificial neural network (Artificial Neural Network, traditional intelligence Processing Algorithm such as ANN), this algorithm can be set up out two sorters with outstanding extensive performance under the situation of small sample amount, avoided local minimum and dimension disaster problem simultaneously.Therefore, this algorithm has caused widely to be paid close attention to, and successfully uses to transformer off-line fault diagnosis field.

Although there are a large amount of researchs of adopting support vector machine to transformer fault diagnosis at present, still there are the following problems in its practice: 1) do not set up the real-time diagnosis system based on support vector machine: most researchs have provided the thinking of fault diagnosis and have adopted data with existing to carry out system and carry out the off-line diagnostic test, but seldom build system architecture concrete, can real-time diagnosis; 2) lack rational data pre-service: the data that oil dissolved gas is sampled and obtained, must could effectively describe the transformer operation characteristic through rational pre-service, some researchs lack the pre-service to raw data, and perhaps preprocess method is unreasonable; 3) the diagnosis decision process is unreasonable: support vector machine is from being a kind of two sorters in essence, and transformer fault is eurypalynous, therefore can not all faulty sections be separated by SVM fault grader only, traditional diagnosis decision process comes with some shortcomings, exist blind area, training sample amount asymmetric etc. as division, influenced the reliability of diagnosis.

Summary of the invention

The present invention is directed to and adopt SVM that transformer is carried out the defective that fault diagnosis technology exists in the background technology, and propose the high transformer on-line fault diagnosis method of a kind of accuracy rate of diagnosis based on SVM and DGA.

Transformer on-line fault diagnosis method based on SVM and DGA of the present invention comprises the steps:

(1) data acquisition

Adopt gas chromatography, with the fixed sampling frequency to the failure gas H in the transformer insulation oil ₂, CH ₄, C ₂H ₂, C ₂H ₄, C ₂H ₆Carry out concentration and detect in real time, then institute's survey data are delivered to PC and carry out pre-service;

(2) data pre-service

Step (1) gained is gathered sample raw data X={x ₁, x ₂, x ₃, x ₄, x ₅, order

Carry out the data pre-service by following formula:

$X^{\&prime;}=\{x_1/{\mathrm{<figure-callout\; id="1"\; label="c"\; filenames="H200910233643XE0000012.png,H200910233643XE0000021.png"\; state="\{\{state\}\}">c</figure-callout>}}_1,x_2/{\mathrm{<figure-callout\; id="1"\; label="c"\; filenames="H200910233643XE0000012.png,H200910233643XE0000021.png"\; state="\{\{state\}\}">c</figure-callout>}}_1,x_3/{\mathrm{<figure-callout\; id="1"\; label="c"\; filenames="H200910233643XE0000012.png,H200910233643XE0000021.png"\; state="\{\{state\}\}">c</figure-callout>}}_1,x_4/{\mathrm{<figure-callout\; id="1"\; label="c"\; filenames="H200910233643XE0000012.png,H200910233643XE0000021.png"\; state="\{\{state\}\}">c</figure-callout>}}_1,x_5/{\mathrm{<figure-callout\; id="1"\; label="c"\; filenames="H200910233643XE0000012.png,H200910233643XE0000021.png"\; state="\{\{state\}\}">\; <figure-callout\; id="1"\; label="c"\; filenames="H200910233643XE0000012.png,H200910233643XE0000021.png"\; state="\{\{state\}\}">c</figure-callout>\; </figure-callout>}}_1,\frac{1}{6}{\log }_{10}\left(c_2\right)\}=\{\frac{\mathrm{<figure-callout\; id="1"\; label="X\; c"\; filenames="H200910233643XE0000012.png,H200910233643XE0000021.png"\; state="\{\{state\}\}">\; <figure-callout\; id="1"\; label="X\; c"\; filenames="H200910233643XE0000012.png,H200910233643XE0000021.png"\; state="\{\{state\}\}">X</figure-callout>\; </figure-callout>}}{c_{}{}_{}},\frac{1}{6}{\log }_{10}\left(c_2\right)\}$

Wherein: x ₁, x ₂, x ₃, x ₄, x ₅Be respectively failure gas H ₂, CH ₄, C ₂H ₂, C ₂H ₄, C ₂H ₆Absolute concentration in insulating oil; c ₁, c ₂Be respectively the total concentration and the total hydrocarbon absolute concentration of the failure gas that dissolves in the insulating oil; X ' is gained data after the pre-service; I is the numbering of failure gas absolute concentration in insulating oil;

(3) data diagnosis

Adopt six svm classifier device: SVM1 to be used to judge the normal condition and the low energy discharge condition of transformer, SVM2 is used to judge the normal condition and the high-energy discharge state of transformer, SVM3 is used to judge the normal condition and the superheat state of transformer, SVM4 is used to judge the superheat state and the low energy discharge condition of transformer, SVM5 is used to judge the superheat state and the high-energy discharge state of transformer, SVM6 is used to judge the low energy discharge condition and the high-energy discharge state of transformer, six svm classifier devices adopt the integrated study method to assemble, wherein: SVM1～SVM6 represents six svm classifier devices respectively, down together;

Diagnostic process divides following 8 paths:

Path 1:

Step (2) gained data are delivered to SVM1,, then data are further delivered to SVM2 when the SVM1 diagnostic result is " normally "; When the SVM2 diagnostic result is " normally ", then data are further delivered to SVM3 again; When the SVM3 diagnostic result is " normally ", then last diagnostic result is " normally ";

Path 2:

Step (2) gained data are delivered to SVM1,, then data are further delivered to SVM2 when the SVM1 diagnostic result is " normally "; When the SVM2 diagnostic result is " normally ", then data are further delivered to SVM3 again; When the SVM3 diagnostic result is " overheated ", then last diagnostic result is " overheated ";

Path 3:

Step (2) gained data are delivered to SVM1,, then data are further delivered to SVM2 when the SVM1 diagnostic result is " normally "; When the SVM2 diagnostic result is " high-energy discharge ", then data are further delivered to SVM5 again; When the SVM5 diagnostic result is " high-energy discharge ", then last diagnostic result is " high-energy discharge ";

Path 4:

Step (2) gained data are delivered to SVM1,, then data are further delivered to SVM2 when the SVM1 diagnostic result is " normally "; When the SVM2 diagnostic result is " high-energy discharge ", then data are further delivered to SVM5 again; When the SVM5 diagnostic result is " overheated ", then last diagnostic result is " overheated ";

Path 5:

Step (2) gained data are delivered to SVM1,, then data are further delivered to SVM4 when the SVM1 diagnostic result is " low energy discharge "; When the SVM4 diagnostic result is " low energy discharge ", then data are further delivered to SVM6 again; When the SVM6 diagnostic result is " low energy discharge ", then last diagnostic result is " low energy discharge ";

Path 6:

Step (2) gained data are delivered to SVM1,, then data are further delivered to SVM4 when the SVM1 diagnostic result is " low energy discharge "; When the SVM4 diagnostic result is " low energy discharge ", then data are further delivered to SVM6 again; When the SVM6 diagnostic result is " high-energy discharge ", then last diagnostic result is " high-energy discharge ";

Path 7:

Step (2) gained data are delivered to SVM1,, then data are further delivered to SVM4 when the SVM1 diagnostic result is " low energy discharge "; When the SVM4 diagnostic result is " overheated ", then data are further delivered to SVM5 again; When the SVM5 diagnostic result is " overheated ", then last diagnostic result is " overheated ";

Path 8:

Step (2) gained data are delivered to SVM1,, then data are further delivered to SVM4 when the SVM1 diagnostic result is " low energy discharge "; When the SVM4 diagnostic result is " overheated ", then data are further delivered to SVM5 again; When the SVM5 diagnostic result is " high-energy discharge ", then last diagnostic result is " high-energy discharge ";

(4) output shows the diagnostic data result.

The present invention has following beneficial effect:

1. consider that fault type is relevant with the ratio and the total hydrocarbon concentration of each characteristic gas, the present invention at first carries out pre-service to the failure gas concentration data of gathering, and the data after the processing more can embody its essential characteristic than classic method;

2. at traditional multilayer SVM diagnostic process, the present invention adopts the acyclic figure method of decision-directed that diagnostic process is optimized, and this flow process has overcome to exist in the traditional process divides blind zone problem;

3. traditional employing SVM is applied to the off-line fault diagnosis to the research that transformer carries out fault diagnosis more, and the present invention has set up the model of online system failure diagnosis, has made things convenient for the real time intelligent failure diagnosis of further research transformer;

To sum up, the present invention can analyze the transformer running status and diagnose in real time more accurately.

Description of drawings

Fig. 1 is a transformer on-line fault diagnosis method flow structural drawing of the present invention.

Fig. 2 is support vector machine fault data principle of classification figure.

Fig. 3 is based on the failure modes diagnosis synoptic diagram of the acyclic figure method of decision-directed.

Fig. 4 adopts the svm classifier device to carry out failure modes diagnosis synoptic diagram among the present invention, among the figure: the 1-normal condition; The 2-superheat state; 3-high-energy discharge state; 4-low energy discharge condition.

Embodiment

Be transformer on-line fault diagnosis method flow structural drawing of the present invention as shown in Figure 1, it is mainly by forming as the lower part:

1. part of data acquisition

Transformer running status and failure gas H ₂, CH ₄, C ₂H ₂, C ₂H ₄, C ₂H ₆Concentration in insulating oil is complicated nonlinear relationship, adopt gas chromatography these five kinds of fault signature gases in the transformer insulation oil to be carried out concentration and detect, then the gas concentration data in real time is delivered to PC data preprocessing part with the fixed sampling frequency.

2. data preprocessing part

The situation of change of the relative concentration of the failure gas that dissolves in the insulating oil and total hydrocarbon absolute concentration is that the direct dominance of different faults type is expressed.And directly adopt raw data to carry out sorter when training, and can have different dimensions because of gas with various, cause training gained sorter performance not good; And raw data is directly carried out normalization, though can eliminate the influence of dimension, the absolute concentration information of having lost failure gas to calculating.

At as above problem, this method has designed effective normalization algorithm according to the transformer fault data characteristics, and this algorithm has kept the relative concentration and the absolute concentration information of fault signature gas when eliminating the dimension influence, and computation process is as follows:

The sample raw data X={x that part of data acquisition is gathered ₁, x ₂, x ₃, x ₄, x ₅, order

Carry out the data pre-service by following formula:

$X^{\&prime;}=\{x_1/{\mathrm{<figure-callout\; id="1"\; label="c"\; filenames="H200910233643XE0000012.png,H200910233643XE0000021.png"\; state="\{\{state\}\}">c</figure-callout>}}_1,x_2/{\mathrm{<figure-callout\; id="1"\; label="c"\; filenames="H200910233643XE0000012.png,H200910233643XE0000021.png"\; state="\{\{state\}\}">c</figure-callout>}}_1,x_3/{\mathrm{<figure-callout\; id="1"\; label="c"\; filenames="H200910233643XE0000012.png,H200910233643XE0000021.png"\; state="\{\{state\}\}">c</figure-callout>}}_1,x_4/{\mathrm{<figure-callout\; id="1"\; label="c"\; filenames="H200910233643XE0000012.png,H200910233643XE0000021.png"\; state="\{\{state\}\}">c</figure-callout>}}_1,x_5/{\mathrm{<figure-callout\; id="1"\; label="c"\; filenames="H200910233643XE0000012.png,H200910233643XE0000021.png"\; state="\{\{state\}\}">\; <figure-callout\; id="1"\; label="c"\; filenames="H200910233643XE0000012.png,H200910233643XE0000021.png"\; state="\{\{state\}\}">c</figure-callout>\; </figure-callout>}}_1,\frac{1}{6}{\log }_{10}\left(c_2\right)\}=\{\frac{\mathrm{<figure-callout\; id="1"\; label="X\; c"\; filenames="H200910233643XE0000012.png,H200910233643XE0000021.png"\; state="\{\{state\}\}">\; <figure-callout\; id="1"\; label="X\; c"\; filenames="H200910233643XE0000012.png,H200910233643XE0000021.png"\; state="\{\{state\}\}">X</figure-callout>\; </figure-callout>}}{c_{}{}_{}},\frac{1}{6}{\log }_{10}\left(c_2\right)\}---\left(1\right)$

Wherein: x ₁, x ₂, x ₃, x ₄, x ₅Be respectively failure gas H ₂, CH ₄, C ₂H ₂, C ₂H ₄, C ₂H ₆Absolute concentration in insulating oil; c ₁, c ₂Be respectively the total concentration and the total hydrocarbon absolute concentration of the failure gas that dissolves in the insulating oil; X ' is gained data after the pre-service; I is the numbering of failure gas absolute concentration in insulating oil;

In formula (1), relative concentration

The component ratio that can reflect oil dissolved gas well, and total hydrocarbon absolute concentration c ₂Span generally is between (17.5,299000), by

Its span can be limited to (0.2072,0.9126).

3.SVM sorter part

Support vector machine (SVM) is two sorters, and promptly each support vector machine can only be judged specific two kinds of running status A, B, test sample book is divided in the higher class of likelihood of failure goes.And the transformer fault type is various (as A, B, C), so system must make up all running statuses (A-B, A-C, B-C) in twos, sets up a plurality of SVM fault graders.This method comprises normal condition and the low energy discharge condition that six svm classifier device: SVM1 are used to judge transformer altogether, SVM2 is used to judge the normal condition and the high-energy discharge state of transformer, SVM3 is used to judge the normal condition and the superheat state of transformer, SVM4 is used to judge the superheat state and the low energy discharge condition of transformer, SVM5 is used to judge the superheat state and the high-energy discharge state of transformer, SVM6 is used to judge the low energy discharge condition and the high-energy discharge state of transformer, wherein: SVM1～SVM6 represents six svm classifier devices respectively, down together.

Each svm classifier device all is to adopt the SVM algorithm that data in the existing expert knowledge library are trained gained.This method training data derives from (Electric Power Systems Research, 2005, related works 75:9-15) such as Ganyun Lv.

Svm classifier device data qualification principle as shown in Figure 2.SVM employing kernel function to higher dimensional space, forms the data set of a linear separability with the inseparable data projection of low-dimensional, and by making up largest interval classification lineoid data is classified.If given sample { (x _{I '}, y _{I '}) _{I '=1} ^l, x _{I '}∈ R ^m, y _{I '}{ ± 1}, SVM are at first by Nonlinear Mapping φ: R for ∈ ^m→ R ⁿTo the higher-dimension separable space, m, n are the sample dimension with data projection, construct largest interval classification lineoid w φ (x)+b=0 then, and the procedural representation of seeking largest interval classification lineoid is following quadratic programming problem:

$\mathrm{Minimize}:w\&CenterDot;w+C\underset{i^{\&prime;}=1}{\overset{l}{\mathrm{\&Sigma;}}}{\mathrm{\&xi;}}_{i^{\&prime;}}^2$

S.t.：y _i′(<w·x _i′>+b)≥1-ξ _i′??i′＝1，...，l????(2)

ξ _i′≥0??i′＝1，...，l

Wherein: w is the lineoid weight vectors; B is biasing; ξ is a slack variable; L is a number of training; C is a penalty factor; I ' is the training sample numbering; Down together;

Adopt method of Lagrange multipliers to find the solution and can get its dual problem:

$\mathrm{Maximize}:W\left(\mathrm{\&alpha;}\right)=\underset{i^{\&prime;}=1}{\overset{l}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_{i^{\&prime;}}-\frac{1}{2}\underset{i^{\&prime;},j=1}{\overset{l}{\mathrm{\&Sigma;}}}{y}_{i^{\&prime;}}{y}_j{\mathrm{\&alpha;}}_{i^{\&prime;}}{\mathrm{\&alpha;}}_j(K(x_{i^{\&prime;}},x_j)+\frac{1}{C}{\mathrm{\&delta;}}_{i^{\&prime;},j}),i^{\&prime;}=1,...,l;j=1,...,l$

$s.t.:\underset{i^{\&prime;}=1}{\overset{l}{\mathrm{\&Sigma;}}}{y}_{i^{\&prime;}}{\mathrm{\&alpha;}}_{i^{\&prime;}}=0---\left(3\right)$

α _i′≥0??i′＝1，...，l

Wherein: α is the Lagrange multiplier of sample; δ _{I ' j}Be Kronecker δ, be defined as 1 when i '=j, all the other are 0; K () is a kernel function; J also is the training sample numbering; Down together;

The Karush-Kuhn-Tucker complementarity condition of this problem is:

α _i′[y _i′(<x _i′·w>+b)-1+ξ _i′]＝0??i′＝1，...，l????(4)

When obtaining optimum solution a ^*, b ^*After, substitution lineoid equation obtains decision function

Last classification results is provided by decision rule sgn (f (x)), and sgn is the number of getting function.

During training svm classifier device, penalty factor C span is [2,1000], and kernel function is chosen as the RBF kernel function, promptly

The σ value is [0.1,2]; Penalty factor is got C=100 among this method embodiment, σ in the kernel function=1, and the svmtrain function that the LibSVM software package of svm classifier device employing woods intelligence benevolence exploitation or MATLAB carry is realized.

Fig. 3 is the failure modes diagnosis synoptic diagram based on the acyclic figure method of decision-directed, this decision process by three layers totally six nodes form, wherein each node is represented the sorter of a specific fault type, and the ground floor node is called root node, and all the other nodes are called child node or leaf node.For test sample book, at first adopt root node that it is divided, divide the result according to root node then, this test sample book is delivered in the position of second layer leaf node in decision, and then learn the position of this test sample book at the 3rd layer of leaf node by the division result of second layer leaf node, by the division of the 3rd layer of leaf node, determine the fault type of test sample book at last.In this method, the division result of upper strata node will instruct the position of test sample book at lower floor's node.

Given test sample book X, its true classification is 2 classes.This test sample book is at first delivered to root node 1-a-4, and this sorter is used to assess the deflection of sample for 1,4 classifications; In the root sorter, suppose that X is divided to 1 class, then decision system is delivered to this sample among the second layer lobus lateralis sinister node 1-a-3, and the assessment test sample book is for the deflection of 1,3 classifications; If this leaf node is divided to it 3 classes at random, among the leaf node 2-a-3, this node was used to assess the deflection of test sample book for 2,3 classifications in the middle of then test sample book will be sent to the 3rd layer; At this moment, test sample book can correctly be divided to 2 classes.

For the test sample book of any classification, this structure can be divided the back at three times and be judged its fault type, is one of fault diagnosis flow scheme the most efficiently at present.Yet still there is a problem in this method: divide preference, promptly the leaf node place divides accuracy usually a little more than the root node place, is analyzed as follows:

With Fig. 3 is example, supposes that dividing accuracy for every layer is p, the division accuracy r (1) of 1 class and 4 classes=r (4)=p*p*p then, and 2 classes and 3 classes are divided accuracy and are

Order

Y ' (p)=2p-1, when The time, derivative y '＞0, and y (1)=-0.5 can get: y (p)＜0, p ∈ (0.25,1], ∵ p＞0, ∴ r (1)-r (2)=y (p) * p＜0, r (2)＞r (1), p ∈ (0.25,1].Promptly for four classification problems, this decision process helps 2 classes and 3 classes.

There is the problem of finishing of dividing in the acyclic figure method of decision-directed, and there is influence in this reliability that makes the numbering of fault type diagnose reality.As a rule, therefore the diagnostic reliability at leaf node place, takes following principle a little more than the root node place when the present invention designs: for important fault type, allow it be numbered 2,3 types as far as possible.

Following table is one group of experimental data at domestic transformer fault, and this table is added up the fault type of 359 failure transformers.

The transformer fault type statistics

Fault type	Platform	Ratio (%)
			Overheated	??226	??63
High-energy discharge	??65	??18.1
			The low energy discharge	??25	??10
Overheated double high-energy discharge	??36	??7
			Other	??7	??1.9

As above shown in the table, overheating fault is the highest fault type, and the high-energy discharge fault is taken second place.Therefore at transformer fault diagnosis, running status is numbered in the inventive method: the 1-normal condition; The 2-superheat state; 3-high-energy discharge state; 4-low energy discharge condition.

Be illustrated in figure 4 as and adopt the svm classifier device to carry out failure modes diagnosis synoptic diagram among the present invention.The svm classifier device is assembled with the integrated study method, and diagnostic procedure divides following 8 paths:

Path 1:

Gained data after the pre-service are delivered to SVM1,, then data are further delivered to SVM2 when the SVM1 diagnostic result is " normally "; When the SVM2 diagnostic result is " normally ", then data are further delivered to SVM3 again; When the SVM3 diagnostic result is " normally ", then last diagnostic result is " normally ";

Path 2:

Gained data after the pre-service are delivered to SVM1,, then data are further delivered to SVM2 when the SVM1 diagnostic result is " normally "; When the SVM2 diagnostic result is " normally ", then data are further delivered to SVM3 again; When the SVM3 diagnostic result is " overheated ", then last diagnostic result is " overheated ";

Path 3:

Gained data after the pre-service are delivered to SVM1,, then data are further delivered to SVM2 when the SVM1 diagnostic result is " normally "; When the SVM2 diagnostic result is " high-energy discharge ", then data are further delivered to SVM5 again; When the SVM5 diagnostic result is " high-energy discharge ", then last diagnostic result is " high-energy discharge ";

Path 4:

Gained data after the pre-service are delivered to SVM1,, then data are further delivered to SVM2 when the SVM1 diagnostic result is " normally "; When the SVM2 diagnostic result is " high-energy discharge ", then data are further delivered to SVM5 again; When the SVM5 diagnostic result is " overheated ", then last diagnostic result is " overheated ";

Path 5:

Gained data after the pre-service are delivered to SVM1,, then data are further delivered to SVM4 when the SVM1 diagnostic result is " low energy discharge "; When the SVM4 diagnostic result is " low energy discharge ", then data are further delivered to SVM6 again; When the SVM6 diagnostic result is " low energy discharge ", then last diagnostic result is " low energy discharge ";

Path 6:

Gained data after the pre-service are delivered to SVM1,, then data are further delivered to SVM4 when the SVM1 diagnostic result is " low energy discharge "; When the SVM4 diagnostic result is " low energy discharge ", then data are further delivered to SVM6 again; When the SVM6 diagnostic result is " high-energy discharge ", then last diagnostic result is " high-energy discharge ";

Path 7:

Gained data after the pre-service are delivered to SVM1,, then data are further delivered to SVM4 when the SVM1 diagnostic result is " low energy discharge "; When the SVM4 diagnostic result is " overheated ", then data are further delivered to SVM5 again; When the SVM5 diagnostic result is " overheated ", then last diagnostic result is " overheated ";

Path 8:

Gained data after the pre-service are delivered to SVM1,, then data are further delivered to SVM4 when the SVM1 diagnostic result is " low energy discharge "; When the SVM4 diagnostic result is " overheated ", then data are further delivered to SVM5 again; When the SVM5 diagnostic result is " high-energy discharge ", then last diagnostic result is " high-energy discharge ";

This flow process can effectively improve the accuracy of in-circuit diagnostic system to fault diagnosis occurred frequently.

4. output as a result

Last computing machine shows output with diagnostic result, finishes this diagnostic method flow process.

Claims (2)

1. the transformer on-line fault diagnosis method based on SVM and DGA is characterized in that comprising the steps:

(1) data acquisition

Adopt gas chromatography, with the fixed sampling frequency to the failure gas H in the transformer insulation oil ₂, CH ₄, C ₂H ₂, C ₂H ₄, C ₂H ₆Carry out concentration and detect in real time, then institute's survey data are delivered to PC and carry out pre-service;

(2) data pre-service

Step (1) gained is gathered sample raw data X={x ₁, x ₂, x ₃, x ₄, x ₅, order

Carry out the data pre-service by following formula:

$X^{\&prime;}=\{x_1/c_1,x_2/c_1,x_3/c_1,x_4/c_1,x_5/c_1,\frac{1}{6}{\log }_{10}\left(c_2\right)\}=\{\frac{X}{c_1},\frac{1}{6}{\log }_{10}\left(c_2\right)\}$

Wherein: x ₁, x ₂, x ₃, x ₄, x ₅Be respectively failure gas H ₂, CH ₄, C ₂H ₂, C ₂H ₄, C ₂H ₆Absolute concentration in insulating oil; c ₁, c ₂Be respectively the total concentration and the total hydrocarbon absolute concentration of the failure gas that dissolves in the insulating oil; X ' is gained data after the pre-service; I is the numbering of failure gas absolute concentration in insulating oil;

(3) data diagnosis

Adopt six svm classifier device: SVM1 to be used to judge the normal condition and the low energy discharge condition of transformer, SVM2 is used to judge the normal condition and the high-energy discharge state of transformer, SVM3 is used to judge the normal condition and the superheat state of transformer, SVM4 is used to judge the superheat state and the low energy discharge condition of transformer, SVM5 is used to judge the superheat state and the high-energy discharge state of transformer, SVM6 is used to judge the low energy discharge condition and the high-energy discharge state of transformer, six svm classifier devices adopt the integrated study method to assemble, wherein: SVM1～SVM6 represents six svm classifier devices respectively, down together;

Diagnostic process divides following 8 paths:

Path 1:

Step (2) gained data are delivered to SVM1,, then data are further delivered to SVM2 when the SVM1 diagnostic result is " normally "; When the SVM2 diagnostic result is " normally ", then data are further delivered to SVM3 again; When the SVM3 diagnostic result is " normally ", then last diagnostic result is " normally ";

Path 2:

Step (2) gained data are delivered to SVM1,, then data are further delivered to SVM2 when the SVM1 diagnostic result is " normally "; When the SVM2 diagnostic result is " normally ", then data are further delivered to SVM3 again; When the SVM3 diagnostic result is " overheated ", then last diagnostic result is " overheated ";

Path 3:

Step (2) gained data are delivered to SVM1,, then data are further delivered to SVM2 when the SVM1 diagnostic result is " normally "; When the SVM2 diagnostic result is " high-energy discharge ", then data are further delivered to SVM5 again; When the SVM5 diagnostic result is " high-energy discharge ", then last diagnostic result is " high-energy discharge ";

Path 4:

Step (2) gained data are delivered to SVM1,, then data are further delivered to SVM2 when the SVM1 diagnostic result is " normally "; When the SVM2 diagnostic result is " high-energy discharge ", then data are further delivered to SVM5 again; When the SVM5 diagnostic result is " overheated ", then last diagnostic result is " overheated ";

Path 5:

Step (2) gained data are delivered to SVM1,, then data are further delivered to SVM4 when the SVM1 diagnostic result is " low energy discharge "; When the SVM4 diagnostic result is " low energy discharge ", then data are further delivered to SVM6 again; When the SVM6 diagnostic result is " low energy discharge ", then last diagnostic result is " low energy discharge ";

Path 6:

Step (2) gained data are delivered to SVM1,, then data are further delivered to SVM4 when the SVM1 diagnostic result is " low energy discharge "; When the SVM4 diagnostic result is " low energy discharge ", then data are further delivered to SVM6 again; When the SVM6 diagnostic result is " high-energy discharge ", then last diagnostic result is " high-energy discharge ";

Path 7:

Step (2) gained data are delivered to SVM1,, then data are further delivered to SVM4 when the SVM1 diagnostic result is " low energy discharge "; When the SVM4 diagnostic result is " overheated ", then data are further delivered to SVM5 again; When the SVM5 diagnostic result is " overheated ", then last diagnostic result is " overheated ";

Path 8:

Step (2) gained data are delivered to SVM1,, then data are further delivered to SVM4 when the SVM1 diagnostic result is " low energy discharge "; When the SVM4 diagnostic result is " overheated ", then data are further delivered to SVM5 again; When the SVM5 diagnostic result is " high-energy discharge ", then last diagnostic result is " high-energy discharge ";

(4) output shows the diagnostic data result.

2. the transformer on-line fault diagnosis method based on SVM and DGA according to claim 1 is characterized in that: described svm classifier device is to adopt the SVM algorithm that the expertise database data is trained to obtain, and training parameter is selected as follows:

Penalty factor C value is [2,1000]; Kernel function is chosen as the RBF kernel function, promptly

The σ value is [0.1,2].

Images