CN103745119A - Oil-immersed transformer fault diagnosis method based on fault probability distribution model - Google Patents

Abstract

The invention discloses an oil-immersed transformer fault diagnosis method based on a fault probability distribution model, aiming to provide the probability of various faults of the oil-immersed transformer according to analysis of dissolved gas in oil. The oil-immersed transformer fault diagnosis method includes the steps of collecting massive analysis data of the dissolved gas in the oil of the transformer and normalizing the data, utilizing a Sigmoid function to realize probability output on the basis of a support vector machine, computing according to the collected massive analysis data of the dissolved gas in the oil to acquire optimal parameters of the support vector machine and the Sigmoid function, fusing the probability output of multiple support vector machines into probability output of multi-class problems, and obtaining a diagnosis result by referring to the maximum probability classification and standard deviation. The oil-immersed transformer fault diagnosis method based on the fault probability distribution model has high accuracy without error diagnosis, can be used for processing multiple fault conditions effectively and has high practicability.

Description

A kind of oil-immersed transformer fault diagnosis method based on probability of malfunction model

Technical field

The invention belongs to Fault Diagnosis for Electrical Equipment technical field, be specifically related to a kind of oil-immersed transformer fault diagnosis method based on probability of malfunction model.

Background technology

In electric system, power transformer is being undertaken the function of change in voltage between electrical network, electric energy conversion, is one of most important equipment in electric system.The fault of transformer in operational process can cause damage or damage equipment itself on the one hand, and be with user is caused to power failure; On the other hand also may initiating system accident, its harm is larger.Along with the development of computer technology, intellectual technology and sensor technology, various Intelligent Fault Diagnosis Techniques have started to be applied in the fault diagnosis of power transformer, become convenient, the effective important means of one of transformer fault diagnosis.Utilize network technology realize technical supervision system, by the electrical equipments such as power transformer are carried out to fault detect and analyzing and diagnosing, thereby accurately, find reliably fault potential in these equipment, can effectively prevent the great electric power accident causing thus, realization changes to Mode of condition-oriented overhaul from existing preventative maintenance mode, and the safe reliability to Operation of Electric Systems and economy all tool are of great significance.

Oil-filled transformer is the key equipment in electric system, and its safe and stable operation is significant.The insulation system of oil-filled transformer is mainly comprised of insulating oil and insulating paper.In transformer operational process, insulating oil and insulating paper can be under heat and electric effect degradation and decomposition gradually, its dielectric strength can reduce gradually.Utilize various chemical analysis technologies the moisture that ageing process produces, CO, CO can be detected ₂, H ₂, the material such as hydro carbons and furans.The content of all kinds of gases can show larger difference because of the difference of fault type and the difference of the order of severity.In numerous physical detection means and chemical detection means, dissolved gas analysis (dissolved and free gas analysis, DGA) by component and the content of analytical gas, carry out checkout equipment state, can find in time inner latent fault and the order of severity thereof, because it becomes immersed electric apparatus oil fault finding and the diagnostic means of widespread use without affecting the normal operation of transformer.Dissolved gas analysis is the core content of IEC 60599 standards of the current international practice, and this standard is translated and is converted into standard GB/T/T7252.

Traditional Diagnosis Method of Transformer Faults has Rogers method, Dornerburg method, Duval triangulation method etc.On the basis of these methods, people also introduced various intellectual technologies in transformer fault diagnosis in the last few years, and for example expert system, fuzzy algorithm, artificial neural network, evidence theory, support vector machine theory etc., obtained certain achievement.But these Diagnosis Method of Transformer Faults provide clear and definite fault type, lay particular emphasis on the accuracy rate of diagnosis more, the phenomenon of wrong diagnosis cannot be avoided, the result of wrong diagnosis to field staff without any directive significance.

Summary of the invention

For the existing above-mentioned technical matters of prior art, the invention provides a kind of oil-immersed transformer fault diagnosis method based on probability of malfunction model, in feature mode matching process, introduced probability, make diagnostic result reflect oil-immersed type transformer inner case more comprehensively, effectively solve the drawback existing in existing same class methods, reach better effect.

Based on an oil-immersed transformer fault diagnosis method for probability of malfunction model, comprise the steps:

(1) by collecting the historical data information of Oil Dissolved Gases Concentration under transformer fault state, to obtain multiple features training samples of the corresponding various fault types of transformer; Described fault type comprises that cryogenic overheating, middle temperature are overheated, six kinds of hyperthermia and superheating, shelf depreciation, low energy electric discharge and high-energy discharges;

(2) for any two class fault type E1 and E2 combination, according to the features training sample calculation of this two classes fault type, go out the decision function of two category support vector machines of this two classes fault type combination, and then it is as follows according to described decision function, to set up two classification probability of malfunction models of this two classes fault type combination:

$P(y=1|x)=\frac{1}{1+\exp (\mathrm{Af}\left(x\right)+B)}$

P(y=0|x)=1-P(y=1|x)

Wherein: the actual characteristic sample that x is failure transformer, f (x) is described decision function, y is that the label value of fault type and fault type E1 and label value corresponding to E2 are respectively 1 and 0, P (y=1|x) and P (y=0|x) are respectively actual characteristic sample x are inputed to the corresponding fault type E1 that obtains in above two classification probability of malfunction models and the probable value of E2, and A and B are model parameter;

(3) according to step (2), travel through all 15 kinds of fault types combinations, correspondence obtains 15 kind of two classification probability of malfunction model; By detecting the content of current failure transformer oil dissolved gas with the actual characteristic sample of acquisition failure transformer, and then by 15 kind of two classification probability of malfunction model, calculate 15 groups of probable values according to described actual characteristic sample;

(4) according to 15 groups of probable values that calculate, following many classification probability of malfunction models are minimized and solved, obtain fault type probability P, and then fault type probability P is analyzed to determine the fault type of current failure transformer;

$\underset{P}{\min }\underset{i=1}{\overset{6}{\mathrm{\Σ}}}\underset{j=1}{\overset{6}{\mathrm{\Σ}}}{(r_{i+j}^j{p}_i-r_{i+j}^i{p}_j)}^2,s.t.\underset{i=1}{\overset{6}{\mathrm{\Σ}}}{p}_i=1p_i\≥0i\≠j$

P=[p ₁?p ₂?p ₃?p ₄?p ₅?p ₆]

Wherein: with

be respectively the probable value that described actual characteristic sample is inputed to corresponding this two classes fault type obtaining in the two classification probability of malfunction models that combined by i class fault type and j class fault type, p _iand p _jbe respectively the probability that current failure transformer is attributed to i class fault type and is attributed to j class fault type; I and j are natural number and 1≤i≤6,1≤j≤6.

Described features training sample and actual characteristic sample standard deviation be by oil under transformer fault state about H ₂, CH ₄, C ₂h ₆, C ₂h ₄and C ₂h ₂five dimensional vectors of five class dissolved gas content compositions.

Described model parameter A and B obtain by following cross entropy error function is minimized to solve:

$\underset{A,B}{\min }\{-\underset{k=1}{\overset{k}{\mathrm{\Σ}}}[v_k\log P_k+(1-v_k)\log (1-P_k)\left]\right\}$

$P_k=\frac{1}{1+\exp (\mathrm{Af}\left(x_k\right)+B)}$

$v_k=\left\{\begin{array}{ccc}\frac{{\mathrm{<figure-callout\; id="1"\; label="N"\; filenames="HDA0000460183640000011.png"\; state="\{\{state\}\}">N</figure-callout>}}_1+1}{{\mathrm{<figure-callout\; id="1"\; label="N"\; filenames="HDA0000460183640000011.png"\; state="\{\{state\}\}">N</figure-callout>}}_1+2}& \mathrm{if}& y_k=1\\ \frac{1}{N_0+2}& \mathrm{if}& y_k=0\end{array}\right.$

Wherein: x _kfor belonging to k features training sample in the features training sample set of fault type E1 and E2, k is natural number and 1≤i≤K, and K is the features training total sample number in the features training sample set of corresponding fault type E1 and E2, f (x _k) for features training sample xk is inputed to the functional value obtaining in decision function f (x), N ₁and N ₀be respectively in described features training sample set, belong to label value be 1 correspondence fault type E1 features training number of samples and belong to the features training number of samples that label value is the fault type E2 of 0 correspondence, y _kfor features training sample x _kthe label value of corresponding fault type.

Described model parameter A and B obtain by adopting Newton method to minimize to solve to cross entropy error function.

The expression formula of described decision function is as follows:

$f\left(x\right)=\underset{k=1}{\overset{k}{\mathrm{\&Sigma;}}}{y}_k{a}_{k}Z(x_k,x)+b$

Wherein: x _kfor belonging to k features training sample in the features training sample set of fault type E1 and E2, k is natural number and 1≤i≤K, and K is the features training total sample number in the features training sample set of corresponding fault type E1 and E2, y _kfor features training sample x _kthe label value of corresponding fault type, a _kfor features training sample x _kcorresponding Lagrange multiplier, Z (x _k, be x) about x _kwith the kernel function of x, b is the intercept of classification lineoid.

The present invention adopts grid search to carry out cross validation to described features training sample set, select punishment parameter that wherein accuracy rate the is the highest parameter as two category support vector machines kernel functions, and then determine the each parameter in two category support vector machines decision functions.

The present invention adopts process of iteration that many classification probability of malfunction model is minimized and solved, and obtains fault type probability P.

In the last few years, intelligent algorithm has obtained more application in oil-filled transformer fault diagnosis field, wherein core concept is the feature mode that extracts Oil Dissolved Gases Concentration, itself and existing standard or comparable data is compared, thereby obtain contingent fault.The present invention, after fully having studied existing diagnostic method, has introduced probability in feature mode matching process, makes diagnostic result reflect oil-immersed type transformer inner case more comprehensively, effectively solves the drawback existing in existing same class methods, reaches better effect.

Whether the inventive method can be distinguished the fault signature of sample by maxP, standard deviation sigma clear and definite; For the clear and definite sample of fault signature, the inventive method and traditional SVM(support vector machine) diagnostic method all can be judged accurately; For the indefinite sample of fault signature, only merely from DGA(dissolved gas analysis) data are difficult to provide the single failure conclusion of high confidence level, often there is wrong diagnosis situation in tradition SVM method, cannot take correct corrective action, and can waste valuable Deal with Time and cannot deal with problems.And the inventive method can provide the probability that every class fault occurs, for the wrong diagnosis sample that may exist in prior art, also can provide the fault that may exist, and can effectively process the situation of multiple failure, make up the deficiency that traditional SVM method exists wrong diagnosis, and then for the fault that may exist, arrange corresponding disposal options according to diagnostic result, rational allocation maintenance resource, as much as possible guarantee the effective of corrective action, can finely meet the requirement of rig-site utilization, for staff provides more help, there is better practicality.

Accompanying drawing explanation

Fig. 1 is that Sigmoid function parameter solves process flow diagram.

Fig. 2 is the cross validation accuracy rate calculation flow chart of selecting for parameter.

Fig. 3 is the diagnostic result scatter diagram of the inventive method.

Embodiment

In order more specifically to describe the present invention, below in conjunction with the drawings and the specific embodiments, technical scheme of the present invention is elaborated.

Based on an oil-immersed transformer fault diagnosis method for probability of malfunction model, comprise the steps:

Step (1): collect the data message of the Gases Dissolved in Transformer Oil of a large amount of known fault, select the H dissolving in oil ₂, CH ₄, C ₂h ₆, C ₂h ₄, C ₂h ₂gas content is normalized as characteristic parameter, forms a large amount of data samples consisting of characteristic parameter and fault type.With reference to IEC 60599, transformer fault type can be divided into six kinds: cryogenic overheating T1 (lower than 300 ° of C), the overheated T2 of middle temperature (higher than 300 ° of C and lower than 700 ° of C), hyperthermia and superheating T3 (higher than 700 ° of C), shelf depreciation PD, low energy electric discharge D1, high-energy discharge D2.The proper vector x of one of them sample (x, y) is the vector (x that the content of this 5 class gas forms ₁, x ₂, x ₃, x ₄, x ₅), the fault type that y is sample.

Step (2): a large amount of fault data samples of collecting are processed, calculated the mathematics submodel for oil-immersed type transformer fault diagnosis.All fault data sample evidence fault types can be divided into six classes, for any two class fault samples, all can utilize support vector machine method to calculate a kind of mathematical differentiation (being designated hereinafter simply as two classification problems), final two category support vector machines that form 15 support probability outputs, wherein support the mathematical description of two category support vector machines of probability output to be for 1:

The decision function form of support vector machine is:

$f\left(x\right)=\underset{k=1}{\overset{k}{\mathrm{\&Sigma;}}}{y}_k{a}_{k}Z(x_k,x)+b---\left(1\right)$

Wherein: Z (x _k, x) be kernel function, a _klagrange multiplier, x _kfor sample data, y _kfor sample label value, b be classification lineoid intercept and be a constant, the general type of lineoid is ω ^tx+b=0, wherein ω ^tfor the transposition of plane normal vector, ω and b have determined the position of lineoid jointly.

For any one group of sample (x _k, y _k), SVM is output as a definite numerical value.Adopt John C.Platt at document Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods (Advances in large margin classifiers, 1999,10 (3): 61-74) method proposing in, by a Sigmoid function with parameter, SVM decision function f (x) is mapped to interval [0,1], thereby realize probability output, suc as formula (2):

$P(y=1|x)=\frac{1}{1+\exp (\mathrm{Af}\left(x\right)+B)}---\left(2\right)$

P(y=0|x)=1-P(y=1|x)

Wherein: A, B are the undetermined parameters of Sigmoid function, A<0 is used for guaranteeing Sigmoid function monotone increasing; A, B utilize training set (f _k, y _k) carry out maximum likelihood estimate obtain, f _k=f (x _k).

In present embodiment, the formation of training set is as follows: the sample set (x of two classification problems _k, y _k) need to train parameter A, the B of a SVM and a Sigmoid function simultaneously.Training sample (the x of each SVM _k, y _k) there is a correspondence (f _k, y _k) as the training set of Sigmoid function, wherein f _kthe value calculating for SVM decision function.The SVM that sample set trains calculates decision function output fk and is not directly used in the training set (f that forms Sigmoid function _k, y _k), adopt the method for cross validation to form training set (f _k, y _k).By sample set (x _k, y _k) be divided into 5 groups, utilize wherein 4 groups of training to obtain SVM and also with SVM decision function, calculate all the other f of one group _k, so process and get final product to obtain the f of all samples for 5 times _k, and then form Sigmoid parameter training collection.

Parameter A, B obtain by minimizing cross entropy error function, note z=[A, B], solving model is as (3):

$\underset{A,B}{\min }\{-\underset{k=1}{\overset{k}{\mathrm{\&Sigma;}}}[v_k\log P_k+(1-v_k)\log (1-P_k)\left]\right\}---\left(3\right)$

$P_k=\frac{1}{1+\exp (\mathrm{Af}\left(x_k\right)+B)}$

$v_k=\left\{\begin{array}{ccc}\frac{{\mathrm{<figure-callout\; id="1"\; label="N"\; filenames="HDA0000460183640000011.png"\; state="\{\{state\}\}">N</figure-callout>}}_1+1}{{\mathrm{<figure-callout\; id="1"\; label="N"\; filenames="HDA0000460183640000011.png"\; state="\{\{state\}\}">N</figure-callout>}}_1+2}& \mathrm{if}& y_k=1\\ \frac{1}{N_0+2}& \mathrm{if}& y_k=0\end{array}\right.$

Wherein: the length that K is sample set, N ₁label y in whole samples _k=1 number of samples, N ₀label y in whole samples _k=0 number.

The optimum solution of this model is tried to achieve by Newton method, obtains optimum solution when the gradient matrix of F (z) equals 0.F (z) gradient matrix is formula (4), and Hessian matrix is formula (5), z=[A, B]; Model solve flow process as shown in Figure 1.

$\&dtri;\left(z\right)=\left[\begin{array}{c}{\mathrm{\&Sigma;}}_{k=1}^K{f}_k(v_k-P_k)\\ {\mathrm{\&Sigma;}}_{k=1}^K(v_k-P_k)\end{array}\right]---\left(4\right)$

$H\left(z\right)=\left[\begin{array}{cc}{\mathrm{\&Sigma;}}_{k=1}^K{f}_k^2{P}_k(1-P_k)& {\mathrm{\&Sigma;}}_{k=1}^K{f}_k{P}_k(1-P_k)\\ {\mathrm{\&Sigma;}}_{k=1}^K{f}_k{P}_k(1-P_k)& {\mathrm{\&Sigma;}}_{k=1}^K{P}_k(1-P_k)\end{array}\right]---\left(5\right)$

After having obtained parameter A, B, can the output f (x) of SVM decision function be mapped to probability output by Sigmoid function.At a sample x, belong to classification i or belong in two classification problems of classification j, utilizing formula (2) can belong in the hope of x the probability estimate r of a certain class _ij≈ P (y=i|y=iorj), and have r _ij+ r _ji=1 relation.

In the present embodiment, the selection of support vector machines method Kernel Function and punishment parameters C is larger on the accuracy impact of transformer fault diagnosis.The kernel function of model adopts radial basis function, shown in (6).

$K(x_i,x_j)=\exp (-\mathrm{\&gamma;}||x_i-x_j{\left|\right|}^2)---\left(6\right)$

The punishment parameters C of parameter γ in kernel function and SVM need to be by further determining.Present embodiment has adopted grid search, and training sample set is carried out to cross validation, selects one group (C, γ) that wherein accuracy rate is the highest parameter as model.The span of punishment parameters C is [10 ^-6, 10 ^-5..., 10 ⁵, 10 ⁶], the span of the parameter γ of radial basis function is [10 ^-5, 10 ^-4..., 10 ⁴, 10 ⁵], come to parameter combinations in 143; Fig. 2 is the calculation process of the accuracy rate of one group of parameter.

Step (3): for an oil-immersed type transformer for the treatment of tracing trouble, obtain its oil dissolved gas situation, obtain H ₂, CH ₄, C ₂h ₆, C ₂h ₄, C ₂h ₂gas content, utilizes in step (2), to have obtained 15 two disaggregated models and calculate whole r _ij, r _ji.Utilize following mathematics submodel 2, by its probability estimate r _ijcomprehensive is polytypic probability output, obtains belonging to the Probability p of each class _i, that is:

p _i=P(y=i|x),i=1,...,6

$\underset{i=1}{\overset{6}{\mathrm{\&Sigma;}}}{p}_i=1$

Present embodiment has been used by people such as WU T.F, LIN C.J and WENG R.C at document Probability estimates for multi-class classification by pairwise coupling (The Journal of Machine Learning Research, 2004, what 5:975-1005.), propose obtains the method for many class probabilities from two class probabilities, and mathematical description is as follows:

Consider to have r _ij+ r _ji=1, r _ij≈ p _i/ (p _i+ p _j), obtain:

$\frac{r_{\mathrm{ij}}}{r_{\mathrm{ij}}}\&ap;\frac{p_i}{p_j}$

It is carried out simple transformation and sues for peace obtaining formula (7):

${\mathrm{\&Sigma;}}_{j:j\&NotEqual;i}{r}_{\mathrm{ji}}{p}_i\&ap;{\mathrm{\&Sigma;}}_{j:j\&NotEqual;i}{r}_{\mathrm{ij}}{p}_j---\left(7\right)$

Utilize r in formula (7) _ijand p _ithe relation existing, solves with drag and can obtain p _i:

$\min \underset{i=1}{\overset{6}{\mathrm{\&Sigma;}}}\underset{j:j\&NotEqual;i}{\mathrm{\&Sigma;}}{(r_{\mathrm{ji}}{p}_i-r_{\mathrm{ij}}{p}_j)}^2$

$s.t.\underset{i=1}{\overset{6}{\mathrm{\&Sigma;}}}{p}_i=1,p_i\&GreaterEqual;0,\&ForAll;i.$

Wherein: p _i>=0 constraint is redundancy, note p ^t=[p ₁, p ₂..., p ₆], model conversation is (8), this is a convex quadratic programming problem, when meeting formula (9), obtains optimum solution.

$\underset{P}{\min 2P^T}\mathrm{QP}\&equiv;\underset{P}{\mathrm{<figure-callout\; id="1"\; label="min\; P"\; filenames="HDA0000460183640000011.png"\; state="\{\{state\}\}">min</figure-callout>}}\frac{1}{2}{P}^T\mathrm{QP}$

$Q_{\mathrm{ij}}=\left\{\begin{array}{cc}{\mathrm{\&Sigma;}}_{s:s\&NotEqual;i}{r}_{\mathrm{<figure-callout\; id="2"\; label="si"\; filenames="HDA0000460183640000031.png"\; state="\{\{state\}\}">si</figure-callout>}}^2,& i=j\\ {-r}_{\mathrm{ji}}{r}_{\mathrm{ij}},& i\&NotEqual;j\end{array}\right.$

$\left[\begin{array}{cc}Q& e\\ e^T& 0\end{array}\right]\left[\begin{array}{c}P\\ b\end{array}\right]=\left[\begin{array}{c}0\\ 1\end{array}\right]---\left(9\right)$

Optimum solution solves flow process:

Iteration: make t=1 ..., 6,1 ..., 6,1

1) utilize formula (10) to p _tupgrade.

2) p is normalized.

3) whether checking p meets formula (9), if meet, stops iteration, obtains polytypic Probability p.

$P_t\&LeftArrow;\frac{1}{Q_{\mathrm{tt}}}[-\underset{j:j\&NotEqual;t}{\mathrm{\&Sigma;}}{Q}_{\mathrm{tj}}{p}_j+p^T\mathrm{QP}]---\left(10\right)$

The p finally obtaining is the probable value that various faults occur oil-immersed type transformer, and then, p is analyzed and obtains present embodiment diagnosis.

To the diagnostic result of each sample, the probability of fault T1, T2, T3, PD, D1, D2 is designated as to p ₁, p ₂, p ₃, p ₄, p ₅, p ₆, take out its maximum probability maxP=max{p ₁, p ₂, p ₃, p ₄, p ₅, p ₆, calculate (p simultaneously ₁, p ₂, p ₃, p ₄, p ₅, p ₆) standard deviation sigma.All larger samples of maxP, σ in diagnostic result, what select maximum probability is transformer fault.The all less sample of maxP, σ in diagnostic result, selects larger two to three of probability to investigate for possible transformer fault carries out scene successively.The categorised demarcation line of two kinds of selection modes is maxP=0.6 substantially, σ=0.06, and standard can be adjusted according to actual conditions, and Fig. 3 has provided the scatter diagram of some typical diagnostic results about maxP and σ.

Claims (7)

1. the oil-immersed transformer fault diagnosis method based on probability of malfunction model, comprises the steps:

(1) by collecting the historical data information of Oil Dissolved Gases Concentration under transformer fault state, to obtain multiple features training samples of the corresponding various fault types of transformer; Described fault type comprises that cryogenic overheating, middle temperature are overheated, six kinds of hyperthermia and superheating, shelf depreciation, low energy electric discharge and high-energy discharges;

(2) for any two class fault type E1 and E2 combination, according to the features training sample calculation of this two classes fault type, go out the decision function of two category support vector machines of this two classes fault type combination, and then it is as follows according to described decision function, to set up two classification probability of malfunction models of this two classes fault type combination:

$P(y=1|x)=\frac{1}{1+\exp (\mathrm{Af}\left(x\right)+B)}$

P(y=0|x)=1-P(y=1|x)

Wherein: the actual characteristic sample that x is failure transformer, f (x) is described decision function, y is that the label value of fault type and fault type E1 and label value corresponding to E2 are respectively 1 and 0, P (y=1|x) and P (y=0|x) are respectively actual characteristic sample x are inputed to the corresponding fault type E1 that obtains in above two classification probability of malfunction models and the probable value of E2, and A and B are model parameter;

(3) according to step (2), travel through all 15 kinds of fault types combinations, correspondence obtains 15 kind of two classification probability of malfunction model; By detecting the content of current failure transformer oil dissolved gas with the actual characteristic sample of acquisition failure transformer, and then by 15 kind of two classification probability of malfunction model, calculate 15 groups of probable values according to described actual characteristic sample;

(4) according to 15 groups of probable values that calculate, following many classification probability of malfunction models are minimized and solved, obtain fault type probability P, and then fault type probability P is analyzed to determine the fault type of current failure transformer;

$\underset{P}{\min }\underset{i=1}{\overset{6}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{6}{\mathrm{\&Sigma;}}}{(r_{i+j}^j{p}_i-r_{i+j}^i{p}_j)}^2,s.t.\underset{i=1}{\overset{6}{\mathrm{\&Sigma;}}}{p}_i=1p_i\&GreaterEqual;0i\&NotEqual;j$

P=[p ₁?p ₂?p ₃?p ₄?p ₅?p ₆]

Wherein:

with

be respectively the probable value that described actual characteristic sample is inputed to corresponding this two classes fault type obtaining in the two classification probability of malfunction models that combined by i class fault type and j class fault type, p _iand p _jbe respectively the probability that current failure transformer is attributed to i class fault type and is attributed to j class fault type; I and j are natural number and 1≤i≤6,1≤j≤6.

2. oil-immersed transformer fault diagnosis method according to claim 1, is characterized in that: described features training sample and actual characteristic sample standard deviation for by oil under transformer fault state about H ₂, CH ₄, C ₂h ₆, C ₂h ₄and C ₂h ₂five dimensional vectors of five class dissolved gas content compositions.

3. oil-immersed transformer fault diagnosis method according to claim 1, is characterized in that: described model parameter A and B obtain by following cross entropy error function is minimized to solve:

$\underset{A,B}{\min }\{-\underset{k=1}{\overset{k}{\mathrm{\&Sigma;}}}[v_k\log P_k+(1-v_k)\log (1-P_k)\left]\right\}$

$P_k=\frac{1}{1+\exp (\mathrm{Af}\left(x_k\right)+B)}$

$v_k=\left\{\begin{array}{ccc}\frac{N_1+1}{N_1+2}& \mathrm{if}& y_k=1\\ \frac{1}{N_0+2}& \mathrm{if}& y_k=0\end{array}\right.$

Wherein: x _kfor belonging to k features training sample in the features training sample set of fault type E1 and E2, k is natural number and 1≤i≤K, and K is the features training total sample number in the features training sample set of corresponding fault type E1 and E2, f (x _k) for features training sample xk is inputed to the functional value obtaining in decision function f (x), N ₁and N ₀be respectively in described features training sample set, belong to label value be 1 correspondence fault type E1 features training number of samples and belong to the features training number of samples that label value is the fault type E2 of 0 correspondence, y _kfor features training sample x _kthe label value of corresponding fault type.

4. oil-immersed transformer fault diagnosis method according to claim 3, is characterized in that: described model parameter A and B obtain by adopting Newton method to minimize to solve to cross entropy error function.

5. according to the oil-immersed transformer fault diagnosis method described in claim 1 or 3, it is characterized in that: the expression formula of described decision function is as follows:

$f\left(x\right)=\underset{k=1}{\overset{k}{\mathrm{\&Sigma;}}}{y}_k{a}_{k}Z(x_k,x)+b$

Wherein: x _kfor belonging to k features training sample in the features training sample set of fault type E1 and E2, k is natural number and 1≤i≤K, and K is the features training total sample number in the features training sample set of corresponding fault type E1 and E2, y _kfor features training sample x _kthe label value of corresponding fault type, a _kfor features training sample x _kcorresponding Lagrange multiplier, Z (x _k, be x) about x _kwith the kernel function of x, b is the intercept of classification lineoid.

6. oil-immersed transformer fault diagnosis method according to claim 5, it is characterized in that: the present invention adopts grid search to carry out cross validation to described features training sample set, select punishment parameter that wherein accuracy rate the is the highest parameter as two category support vector machines kernel functions, and then determine the each parameter in two category support vector machines decision functions.

7. oil-immersed transformer fault diagnosis method according to claim 1, is characterized in that: in described step (4), adopt process of iteration that many classification probability of malfunction model is minimized and solved, obtain fault type probability P.

Images