CN105629109A - ARTI-neural network-based oil-immersed transformer fault diagnosis method - Google Patents

Abstract

The invention discloses an ARTI-neural network-based oil-immersed transformer fault diagnosis method. The method comprises: step one, an ARTI-neural network-based oil-immersed transformer fault diagnosis model is constructed by using a specific ARTI neural network algorithm; step two, on the basis of a four-ratio method, input and output quantities of the constructed ARTI-neural network-based oil-immersed transformer fault diagnosis model obtained by the step one are determined; step three, after the step two, a parameter of the ARTI-neural network-based oil-immersed transformer fault diagnosis model is set; step four, learning training is carried out on a fault sample by using the ARTI-neural network-based oil-immersed transformer fault diagnosis model; and step five, an identification diagnosis is carried out on an actual fault data type by using the ARTI-neural network-based oil-immersed transformer fault diagnosis model. According to the invention, with the method, real-time on-line fault diagnoses on different types of overheating faults of the oil-immersed transformer can be carried out correctly.

Description

Based on the oil-immersed transformer fault diagnosis method of ART1 neural network

Technical field

The invention belongs to the comprehensive fault monitoring method technical field of power transformer, it is specifically related to a kind of oil-immersed transformer fault diagnosis method based on ART1 neural network.

Background technology

Power transformer is electric device equipment the most important in power system, is also one of electric installation of causing power system accident maximum. Its running status directly affects the security and stability level of power supply system. The potentiality fault of Timeliness coverage power transformer, ensures that transformer safety is effectively run, thus improves the reliability of transformer stable power-supplying, is the extremely important problem that power department is paid close attention to. Therefore, study diagnosing fault of power transformer technology, it is to increase the operation and maintenance level of transformer, has important practical significance.

Current diagnosing fault of power transformer technology, although having developed multiple effective method for diagnosing faults, but major part is superposition to a certain extent piles up, and most of method for diagnosing faults itself still lacks the concept system of perfect theoretical basis and systematize. In addition, most of method for diagnosing faults is all the fault that the characteristic signal utilizing object to show carrys out diagnostic characteristic type, often needs a large amount of fault learning sample, and large-scale effectively fault sample data are difficult to obtain in practice. Power transformer because of voltage range, insulation system different with fault degree, transformer oil dissolved gas content is all made to have bigger randomness, and there is certain redundant data, this just requires that method for diagnosing faults has very strong fault-tolerance, and the research of current this respect is also not nearly enough deep.

On the various failure condition of overall understanding oil-immersed power transformer existence and the basis of trouble diagnosis correlation technique, repair achievement in research and the practical experience in field in conjunction with home and abroad related scientific research mechanism in power transformer on-line monitoring, intelligence diagnosis and state, develop a kind of method mainly for research diagnosing fault of power transformer particularly important.

Summary of the invention

It is an object of the invention to provide a kind of oil-immersed transformer fault diagnosis method based on ART1 neural network, it is possible to correctly carry out real-time online trouble diagnosis for the Superheated steam drier that oil-filled transformer is dissimilar.

The present invention is adopted technical scheme to be, based on the oil-immersed transformer fault diagnosis method of ART1 neural network, specifically implement according to following step:

Step 1, utilize ART1 neural network specific algorithm, build the oil-filled transformer fault diagnosis model based on ART1 neural network;

Step 2, four ratioing technigues are utilized to determine to build the input and output amount of oil-filled transformer fault diagnosis model based on ART1 neural network through what step 1 obtained;

Step 3, after step 2, the parameter of oil-filled transformer trouble diagnosis based on ART1 neural network is set;

Fault sample is carried out learning training based on the oil-filled transformer fault diagnosis model of ART1 neural network by step 4, utilization;

Actual fault data type is carried out identifying and diagnosing based on the oil-filled transformer fault diagnosis model of ART1 neural network by step 5, utilization.

The feature of the present invention is also:

Step 1 is specifically implemented according to following step:

Step 1.1, the connection weights between ART1 neural network model input layer and output layer being carried out initialization process, concrete grammar is as follows:

The input layer of setting ART1 neural network model has N number of neurone, and output layer has M neurone;

Two-value fault input vector pattern A_kWith output vector B_kAs follows respectively:

$A_k=(a_1^k,a_2^k,...,a_N^k);$

$B_k=(b_1^k,b_2^k,...,b_M^k);$

Wherein, k=1,2 ..., p, p are the number of input mode of learning;

Make t_ij(0)=1,I=1,2 ..., N, j=1,2 ..., M;

Wherein, Vigilance parameter 0 < �ѡ�1, t_ijFor network model feedback link weights, W_ijFor network model feedforward connects weights;

Step 1.2, after step 1.1, by two-value fault input vector patternIt is supplied to the input layer of network;

Step 1.3, after step 1.2, calculate each neuronic weighted input of ART1 neural network model output layer and, specifically implement according to following algorithm:

$S_j=\underset{i=1}{\overset{N}{\&Sigma;}}{w}_{ij}{a}_i^k,j=1,2,...,M;$

Wherein, W_ijFor network model feedforward connects weights;

For specifically forming two-value fault input vector pattern A_kValue, be specially 0 or 1;

S_jFor each neuronic weighted input of ART1 neural network model output layer and;

Step 1.4, after step 1.3, select input pattern network model optimal classification result, this result according to following algorithm through calculate obtain:

$S_g=\underset{j=1,2,...,M}{max}{S}_j;$

In formula, the output making neurone g is 1, S_jFor each neuronic weighted input of ART1 neural network model output layer and;

Step 1.5, following three formulas are calculated, and calculation result is judged, judge whether final calculation result meets the Vigilance parameter �� (0 < �ѡ�1) of ART1 neural network model setting, namely judge that can ART1 neural network model accept this recognition result:

$\left|A_k\right|=\underset{i=1}{\overset{N}{\&Sigma;}}{a}_i^k---\left(1\right);$

$\left|T_g\&CenterDot;A_k\right|=\underset{i=1}{\overset{N}{\&Sigma;}}{t}_{gi}{a}_i^k---\left(2\right);$

$\frac{\left|T_g\&CenterDot;A_k\right|}{\left|A_k\right|}>\mathrm{\&rho;}---\left(3\right);$

If the last same form is set up, then enter step 1.7;

If the last same form is false, then enter step 1.6;

Step 1.6, after step 1.5, cancel recognition result, concrete grammar is:

First output value by output layer neurone g is reset to 0, and is got rid of outside the scope identified next time by this neurone, then returns step 1.5;

When if the neurone utilized all cannot meet the last same form in step 1.5, then select a new neurone as classification results, then proceed to step 1.7;

Step 1.7, accept recognition result, and connect weights according to the adjustment of following algorithm:

$w_{ig}(t+1)=\frac{t_{gi}\left(t\right)a_i^k}{0.5+\underset{i=1}{\overset{N}{\&Sigma;}}{t}_{gi}\left(t\right)a_i^k};$

$t_{gi}(t+1)=t_{gi}\left(t\right)a_i^k;$

Wherein, i=1,2 ..., N;

T is current time;

For specifically forming the value of two-value fault input vector pattern Ak, it is specially: 0 or 1;

t_giFor network model feedback link weights;

w_igFor network model feedforward connects weights;

Step 1.8, all neurones resetted in step 1.6 are rejoined in identification range, return step 1.2 and carry out identifying to next pattern and store;

The oil-filled transformer fault diagnosis model based on ART1 neural network has just been constructed through step 1.1��step 1.8.

Step 2 is concrete to be implemented in accordance with the following methods:

First, failure transformer is carried out the extraction of fault characteristic gases component;

Secondly, oil dissolved gas CH is measured₄��C₂H₂��C₂H₆��C₂H₄��H₂Concentration of component content;

Finally, four ratioing technigues are utilized to obtain fault feature vector T:

If two concentration of component ratios are greater than 1, then represent with 1;

If two concentration of component ratios are less than 1, then represent with 0;

About 1, then representing the intermediate change process of nature of trouble, namely nature of trouble exposes not too obvious;

Ratio is more big, then the display of nature of trouble is more obvious.

Step 2 relates to utilize four ratioing technigues to obtain the method for fault feature vector T specific as follows:

Utilize five kinds of main characteristic gas CH₄��C₂H₂��C₂H₆��C₂H₄��H₂Form four correlative values, that is: getIts ratio is carried out binary coding, obtains fault feature vector T.

Step 3 is specifically implemented according to following step:

Step 3.1, according to the environment of oil-immersed power transformer actual motion and condition, for the different accuracy of detection of transformer fault diagnosis, the identification rapidly and accurately to transformer fault type judges to realize to arrange the concrete Vigilance parameter �� (0 < �ѡ�1) of ART1 neural network;

Step 3.2, after step 3.1, then ART1 neural network is carried out data initialize, wait the input of fault feature vector T, then realize ART1 neural network model to the diagnosis and distinguish of transformer fault.

Step 4 is concrete to be implemented in accordance with the following methods:

Utilize fault sample data, ART1 neural network is carried out the study of fault sample, it is achieved to the classification of transformer fault type; The binary coding fault feature vector T obtained in conjunction with four ratioing technigues, as the input of ART1 neural network, have selected the sampled data of the 20 groups of known fault types comprising nine kinds of transformer fault types and ART1 neural network carried out learning training.

Step 5 is concrete to be implemented in accordance with the following methods:

The fault feature vector T of the fault data of the oil-filled transformer under step 2 obtains actual working environment, inputs the oil-filled transformer fault diagnosis model based on ART1 neural network by actual fault feature vector T and carries out trouble analysis, export fault type.

The useful effect of the present invention is:

(1) the present invention is based on the oil-immersed transformer fault diagnosis method of ART1 neural network, it is possible to realizes the real-time online study of transformer fault data and accurately identifies transformer fault type.

(2) the present invention is based on the oil-immersed transformer fault diagnosis method of ART1 neural network, it is not necessary to a large amount of transformer fault data sample training, need not carry out yojan to the fault data sample of redundancy.

(3) in the oil-immersed transformer fault diagnosis method of the present invention based on ART1 neural network, network can self-organization, adaptive learning to transformer fault data, moreover it is possible to can carry out the transformer fault data type of the unknown from returning a classification.

(4) relatively in ART2 and ART3 neural network, the present invention is simple and easy to realize based on the ART1 neural network algorithm adopted in the oil-immersed transformer fault diagnosis method of ART1 neural network, it is not necessary to the artificial objectivity setting too much parameter value and causing trouble is diagnosed reduces.

(5) the present invention is based on the oil-immersed transformer fault diagnosis method of ART1 neural network, by the value of adjustment Vigilance parameter �� (0 < �ѡ�1), oil-filled transformer fault when particular job can be carried out the diagnosis of technique of different accuracy, it is achieved maximumization of trouble diagnosibility.

(6) the present invention is based on the oil-immersed transformer fault diagnosis method of ART1 neural network, its principle is: combine in tradition DGA technical Analysis four ratioing technigues used, different Superheated steam drier types in being worked by oil-filled transformer carry out binary coding and form fault feature vector, and ART1 neural network is trained and learns by the fault feature vector that recycling obtains; Shown by theoretical analysis and emulation experiment data results: the present invention correctly can carry out real-time online trouble diagnosis for the Superheated steam drier that oil-filled transformer is dissimilar based on the oil-immersed transformer fault diagnosis method of ART1 neural network, and it has broad application prospects.

Accompanying drawing explanation

Fig. 1 is the schema of the present invention based on the oil-immersed transformer fault diagnosis method of ART1 neural network.

Embodiment

Below in conjunction with the drawings and specific embodiments, the present invention is described in detail.

The present invention, based on the oil-immersed transformer fault diagnosis method of ART1 neural network, as shown in Figure 1, specifically implements according to following step:

Step 1, utilize ART1 neural network specific algorithm, build the oil-filled transformer fault diagnosis model based on ART1 neural network, specifically implement according to following step:

Step 1.1, the connection weights between ART1 neural network model input layer and output layer being carried out initialization process, concrete grammar is as follows:

The input layer of setting ART1 neural network model has N number of neurone, and output layer has M neurone;

Two-value fault input vector pattern A_kWith output vector B_kAs follows respectively:

$A_k=(a_1^k,a_2^k,...,a_N^k);$

$B_k=(b_1^k,b_2^k,...,b_M^k);$

Wherein, k=1,2 ..., p, p are the number of input mode of learning;

Make t_ij(0)=1,I=1,2 ..., N, j=1,2 ..., M;

Wherein, Vigilance parameter 0 < �ѡ�1, t_ijFor network model feedback link weights, W_ijFor network model feedforward connects weights;

Step 1.2, after step 1.1, by two-value fault input vector patternIt is supplied to the input layer of network;

Step 1.3, after step 1.2, calculate each neuronic weighted input of ART1 neural network model output layer and, specifically implement according to following algorithm:

$S_j=\underset{i=1}{\overset{N}{\&Sigma;}}{w}_{ij}{a}_i^k,j=1,2,...,M;$

Wherein, W_ijFor network model feedforward connects weights;

For specifically forming two-value fault input vector pattern A_kValue (0 or 1);

S_jFor each neuronic weighted input of ART1 neural network model output layer and.

Step 1.4, after step 1.3, select input pattern network model optimal classification result, this result according to following algorithm through calculate obtain:

$S_g=\underset{j=1,2,...,M}{max}{S}_j;$

In formula, the output making neurone g is 1, S_jFor each neuronic weighted input of ART1 neural network model output layer and;

Step 1.5, following three formulas are calculated, and calculation result is judged, judge whether final calculation result meets the Vigilance parameter �� (0 < �ѡ�1) of ART1 neural network model setting, namely judge that can ART1 neural network model accept this recognition result:

$\left|A_k\right|=\underset{i=1}{\overset{N}{\&Sigma;}}{a}_i^k---\left(1\right);$

$\left|T_g\&CenterDot;A_k\right|=\underset{i=1}{\overset{N}{\&Sigma;}}{t}_{gi}{a}_i^k---\left(2\right);$

$\frac{\left|T_g\&CenterDot;A_k\right|}{\left|A_k\right|}>\mathrm{\&rho;}---\left(3\right);$

If the last same form is set up, then enter step 1.7;

If the last same form is false, then enter step 1.6;

Step 1.6, after step 1.5, cancel recognition result, concrete grammar is:

First output value by output layer neurone g is reset to 0, and is got rid of outside the scope identified next time by this neurone, then returns step 1.5;

When if the neurone utilized all cannot meet the last same form in step 1.5, then select a new neurone as classification results, then proceed to step 1.7;

Step 1.7, accept recognition result, and connect weights according to the adjustment of following algorithm:

$w_{ig}(t+1)=\frac{t_{gi}\left(t\right)a_i^k}{0.5+\underset{i=1}{\overset{N}{\&Sigma;}}{t}_{gi}\left(t\right)a_i^k};$

$t_{gi}(t+1)=t_{gi}\left(t\right)a_i^k;$

Wherein, i=1,2 ..., N; T is current time;

For specifically forming two-value fault input vector pattern A_kValue (0 or 1);

t_giFor network model feedback link weights;

w_igFor network model feedforward connects weights;

Step 1.8, all neurones resetted in step 1.6 are rejoined in identification range, return step 1.2 and carry out identifying to next pattern and store;

The oil-filled transformer fault diagnosis model based on ART1 neural network has just been constructed through step 1.1��step 1.8.

Step 2, utilize four ratioing technigues to determine to build the input and output amount of oil-filled transformer fault diagnosis model based on ART1 neural network through what step 1 obtained, specifically implement in accordance with the following methods:

First failure transformer is carried out the extraction of fault characteristic gases component, measures oil dissolved gas CH afterwards₄��C₂H₂��C₂H₆��C₂H₄��CO��CO₂And H₂Concentration of component content;

In the inventive solutions, it does not have using seven kinds of main characteristic gas in above transformer oil directly as network input, but have employed four ratioing technigues and obtain fault feature vector, only make use of the main characteristic gas CH of five wherein kind₄��C₂H₂��C₂H₆��C₂H₄��H₂;

Four ratioing technigues make full use of five kinds of main characteristic gas CH₄��C₂H₂��C₂H₆��C₂H₄��H₂Form four correlative values, that is: getIts ratio is carried out binary coding, obtains fault feature vector T:

If two concentration of component ratios are greater than 1, then represent with 1;

If two concentration of component ratios are less than 1, then represent with 0;

About 1, then representing the intermediate change process of nature of trouble, namely nature of trouble exposes not too obvious;

Ratio is more big, then the display of nature of trouble is more obvious;

If the fault having two kinds of character exists simultaneously, such as: 1011, then may be interpreted as continuous electrical spark and overheated.

Four ratioing technigues are utilized to obtain fault feature vector T, often kind of corresponding a kind of fault characteristic type of fault feature vector T; Then fault feature vector T is the input vector of ART1 neural network, and the output of network is the fault characteristic type corresponding to fault feature vector T.

Step 3, after step 2, the parameter of oil-filled transformer trouble diagnosis based on ART1 neural network is set, specifically implements according to following step:

Step 3.1, according to the environment of oil-immersed power transformer actual motion and condition, for the different accuracy of detection of transformer fault diagnosis, the identification rapidly and accurately to transformer fault type judges to realize to arrange the concrete Vigilance parameter �� (0 < �ѡ�1) of ART1 neural network;

Step 3.2, after step 3.1, then ART1 neural network is carried out data initialize, wait the input of fault feature vector T, then realize ART1 neural network model to the diagnosis and distinguish of transformer fault.

Fault sample is carried out learning training based on the oil-filled transformer fault diagnosis model of ART1 neural network by step 4, utilization, specifically implements in accordance with the following methods:

Utilize fault sample data, ART1 neural network is carried out the study of fault sample, thus realize the classification to transformer fault type; The binary coding fault feature vector T obtained in conjunction with four ratioing technigues, as the input of ART1 neural network, have selected the sampled data of the 20 groups of known fault types comprising nine kinds of transformer fault types and ART1 neural network carried out learning training;

In order to accurate objective the trouble diagnosibility embodying ART1 neural network, input sample of data is learnt 2 and takes turns training network, then check ART1 neural network to the accuracy rate of transformer fault diagnosis by test data sample.

Actual fault data type is carried out identifying and diagnosing based on the oil-filled transformer fault diagnosis model of ART1 neural network by step 5, utilization, specifically implements in accordance with the following methods:

The fault feature vector T of the fault data of the oil-filled transformer under step 2 obtains actual working environment, inputs the oil-filled transformer fault diagnosis model based on ART1 neural network by actual fault feature vector T and carries out trouble analysis, export fault type.

The input of 10 groups of fault data since running that the transformer of certain substation 180MVA is put into production as ART1 neural network, and diagnose and also ART1 neural network and conventional three-ratio method are contrasted, contrast ART1 neural network to the diagnosis capability of this transformer fault type with this. Showing by comparing, transformer fault type is had diagnositc analysis ability more accurately by the oil-filled transformer trouble diagnosis based on ART1 neural network, and its accuracy rate is 90%, and conventional three-ratio method accuracy rate only has 70%.

The oil-immersed transformer fault diagnosis method of the present invention based on ART1 neural network has been applied to ART1 neural network, ART1 neural network can work on study limit, limit, to transformer fault data can self-organization, from study, the transformer fault type of the unknown can be carried out from returning a classification and remember. When it being tested, the sample of mnemonic learning can be identified and respond out transformer fault type fast by ART1 neural network automatically.

The present invention has following advantage based on the oil-immersed transformer fault diagnosis method of ART1 neural network:

(1) the transformer fault diagnosis model based on ART1 neural network related in the oil-immersed transformer fault diagnosis method of the present invention based on ART1 neural network is without the need to a large amount of fault data samples, do not require fault data are carried out yojan, avoid the inconvenience that redundant data is brought with the big data problem of process.

(2) in the oil-immersed transformer fault diagnosis method of the present invention based on ART1 neural network, four ratioing technigues are utilized to obtain transformer fault feature scale-of-two input vector, fault data characteristics is inputted vector and carries out dimension-reduction treatment, reduce the complicacy of ART1 neural network, making ART1 neural network diagnosis model calculation speed faster, its diagnostic accuracy is also higher.

(3) the present invention is based on the diagnosis model of the oil-filled transformer trouble diagnosis based on ART1 neural network built in the oil-immersed transformer fault diagnosis method of ART1 neural network, according to the transformer fault diagnosis example data collected, this fault diagnosis model can be trained and emulate.

The present invention, based on the oil-immersed transformer fault diagnosis method of ART1 neural network, is shown by theoretical analysis and emulation result, and the oil-immersed transformer fault diagnosis method based on ART1 neural network has higher transformer fault diagnosis ability; The more important thing is this strategy can self-organization, from study, can carry out remembering from normalization to the unknown fault type of oil-filled transformer, it is not necessary to a large amount of learning training fault data sample, it is possible to realize the diagnosis of real-time online transformer fault; It is also better than ordinary method and other traditional neural net method in oil-filled transformer trouble diagnosis performance.

Claims (7)

1. based on the oil-immersed transformer fault diagnosis method of ART1 neural network, it is characterised in that, specifically implement according to following step:

Step 1, utilize ART1 neural network specific algorithm, build the oil-filled transformer fault diagnosis model based on ART1 neural network;

Step 2, four ratioing technigues are utilized to determine to build the input and output amount of oil-filled transformer fault diagnosis model based on ART1 neural network through what step 1 obtained;

Step 3, after step 2, the parameter of oil-filled transformer trouble diagnosis based on ART1 neural network is set;

Fault sample is carried out learning training based on the oil-filled transformer fault diagnosis model of ART1 neural network by step 4, utilization;

Actual fault data type is carried out identifying and diagnosing based on the oil-filled transformer fault diagnosis model of ART1 neural network by step 5, utilization.

2. the oil-immersed transformer fault diagnosis method based on ART1 neural network according to claim 1, it is characterised in that, described step 1 is specifically implemented according to following step:

Step 1.1, the connection weights between ART1 neural network model input layer and output layer being carried out initialization process, concrete grammar is as follows:

The input layer of setting ART1 neural network model has N number of neurone, and output layer has M neurone;

Two-value fault input vector pattern A_kWith output vector B_kAs follows respectively:

$A_k=(a_1^k,a_2^k,...,a_N^k);$

$B_k=(b_1^k,b_2^k,...,b_M^k);$

Wherein, k=1,2 ..., p, p are the number of input mode of learning;

Make t_ij(0)=1, $W_{ij}\left(0\right)=\frac{1}{N+1},i=1,2,...,N,j=1,2,...,M;$

Wherein, Vigilance parameter 0 < �ѡ�1, t_ijFor network model feedback link weights, W_ijFor network model feedforward connects weights;

Step 1.2, after step 1.1, by two-value fault input vector patternIt is supplied to the input layer of network;

Step 1.3, after step 1.2, calculate each neuronic weighted input of ART1 neural network model output layer and, specifically implement according to following algorithm:

$S_j=\underset{i=1}{\overset{N}{\&Sigma;}}{w}_{ij}^k{a}_i^k,j=1,2,...,M;$

Wherein, W_ijFor network model feedforward connects weights;

For specifically forming two-value fault input vector pattern A_kValue, be specially: 0 or 1;

S_jFor each neuronic weighted input of ART1 neural network model output layer and;

Step 1.4, after step 1.3, select input pattern network model optimal classification result, this result according to following algorithm through calculate obtain:

$S_g=\underset{j=1,2,...,M}{max}{S}_j;$

In formula, the output making neurone g is 1, S_jFor each neuronic weighted input of ART1 neural network model output layer and;

Step 1.5, following three formulas are calculated, and calculation result is judged, judge whether final calculation result meets the Vigilance parameter �� (0 < �ѡ�1) of ART1 neural network model setting, namely judge that can ART1 neural network model accept this recognition result:

$\left|A_k\right|=\underset{i=1}{\overset{N}{\&Sigma;}}{a}_i^k---\left(1\right);$

$\left|T_g\&CenterDot;A_k\right|=\underset{i=1}{\overset{N}{\&Sigma;}}{t}_{gi}{a}_i^k---\left(2\right);$

$\frac{\left|T_g\&CenterDot;A_k\right|}{\left|A_k\right|}>\mathrm{\&rho;}---\left(3\right);$

If the last same form is set up, then enter step 1.7;

If the last same form is false, then enter step 1.6;

Step 1.6, after step 1.5, cancel recognition result, concrete grammar is:

First output value by output layer neurone g is reset to 0, and is got rid of outside the scope identified next time by this neurone, then returns step 1.5;

When if the neurone utilized all cannot meet the last same form in step 1.5, then select a new neurone as classification results, then proceed to step 1.7;

Step 1.7, accept recognition result, and connect weights according to the adjustment of following algorithm:

$w_{ig}(t+1)=\frac{t_{gi}\left(t\right)a_i^k}{0.5+\underset{i=1}{\overset{N}{\&Sigma;}}{t}_{gi}\left(t\right)a_i^k};$

$t_{gi}(t+1)=t_{gi}\left(t\right)a_i^k;$

Wherein, i=1,2 ..., N;

T is current time;

For specifically forming two-value fault input vector pattern A_kValue, be specially: 0 or 1;

t_giFor network model feedback link weights;

w_igFor network model feedforward connects weights;

Step 1.8, all neurones resetted in step 1.6 are rejoined in identification range, return step 1.2 and carry out identifying to next pattern and store;

The oil-filled transformer fault diagnosis model based on ART1 neural network has just been constructed through step 1.1��step 1.8.

3. the oil-immersed transformer fault diagnosis method based on ART1 neural network according to claim 1, it is characterised in that, described step 2 is concrete to be implemented in accordance with the following methods:

First, failure transformer is carried out the extraction of fault characteristic gases component;

Secondly, oil dissolved gas CH is measured₄��C₂H₂��C₂H₆��C₂H₄��H₂Concentration of component content;

Finally, four ratioing technigues are utilized to obtain fault feature vector T:

If two concentration of component ratios are greater than 1, then represent with 1;

If two concentration of component ratios are less than 1, then represent with 0;

About 1, then representing the intermediate change process of nature of trouble, namely nature of trouble exposes not too obvious;

Ratio is more big, then the display of nature of trouble is more obvious.

4. the oil-immersed transformer fault diagnosis method based on ART1 neural network according to claim 3, it is characterised in that, described step 2 relates to utilize four ratioing technigues to obtain the method for fault feature vector T specific as follows:

Utilize five kinds of main characteristic gas CH₄��C₂H₂��C₂H₆��C₂H₄��H₂Form four correlative values, that is: getIts ratio is carried out binary coding, obtains fault feature vector T.

5. the oil-immersed transformer fault diagnosis method based on ART1 neural network according to claim 1, it is characterised in that, described step 3 is specifically implemented according to following step:

Step 3.1, according to the environment of oil-immersed power transformer actual motion and condition, for the different accuracy of detection of transformer fault diagnosis, the identification rapidly and accurately to transformer fault type judges to realize to arrange the concrete Vigilance parameter �� (0 < �ѡ�1) of ART1 neural network;

Step 3.2, after step 3.1, then ART1 neural network is carried out data initialize, wait the input of fault feature vector T, then realize ART1 neural network model to the diagnosis and distinguish of transformer fault.

6. the oil-immersed transformer fault diagnosis method based on ART1 neural network according to claim 1, it is characterised in that, described step 4 is concrete to be implemented in accordance with the following methods:

Utilize fault sample data, ART1 neural network is carried out the study of fault sample, it is achieved to the classification of transformer fault type; The binary coding fault feature vector T obtained in conjunction with four ratioing technigues, as the input of ART1 neural network, have selected the sampled data of the 20 groups of known fault types comprising nine kinds of transformer fault types and ART1 neural network carried out learning training.

7. the oil-immersed transformer fault diagnosis method based on ART1 neural network according to claim 1, it is characterised in that, described step 5 is concrete to be implemented in accordance with the following methods:

The fault feature vector T of the fault data of the oil-filled transformer under step 2 obtains actual working environment, inputs the oil-filled transformer fault diagnosis model based on ART1 neural network by actual fault feature vector T and carries out trouble analysis, export fault type.