CN103207950A - Intelligent transformer fault diagnostic method based on RBF (radial basis function) neural network - Google Patents

Abstract

The invention discloses an intelligent transformer fault diagnostic method based on an RBF (radial basis function) neural network. The intelligent method includes (1), acquiring three ratios of five gases C2H2/C2H4, CH4/H2, C2H4/C2H6 as training sample data by the utilizing IEC (international electrotechnical commission) three ratio method; (2) performing fuzzily processing on the three ratios by utilizing a membership function; (3), coding fault types; (4), training the RBF neural network according to the training sample data until the RBF neural network meets precision requirements; (5), inputting to-be-diagnosed samples after being fuzzily processed; and (6), outputting diagnosed results. The intelligent transformer fault diagnostic method has good reasoning ability and high diagnosed precision, overcomes the defects of the IEC three ratio method, and can precisely display all transformer fault problems.

Description

A kind of intelligent method of the transformer fault diagnosis based on the RBF neural network

Technical field

The present invention relates to the fault diagnosis field of transformer, relate in particular to a kind of intelligent method of the transformer fault diagnosis based on the RBF neural network.

Background technology

At present, transformer is one of important pivot equipment in the electric system, and it is bearing voltage of transformation, distributes the important task with electric energy transmitting, economy transmission, flexible allocation and safe handling to electric energy have great importance, and its running status directly influences the safe operation of whole electric system with stable.

The method of transformer fault diagnosis is a lot, wherein most methods all are this character of gas dissolved in oil of power trans-formers of utilizing dissimilar transformer faults corresponding different, simultaneously the concentration analysis of various faults characteristic gas is found the incipient fault of transformer and the type of fault, these means are not subjected to the influence of external electromagnetic field, can regularly carry out fault diagnosis to the transformer inside under the running status.Main method has: method, characteristic gas diagnosis grind in IEC three-ratio method, electricity association.

The IEC three-ratio method is a kind of common method of present stage transformer fault being diagnosed.The ultimate principle of three-ratio method is: cracking under fault produces the relative concentration of gas composition content and the relation of interdependence of temperature according to TRANSFORMER INSULATING MATERIAL, extracts 5 kinds of characteristic gas (H from transformer oil ₂, CH ₄, C ₂H ₆, C ₂H ₄And C ₂H ₂, calculate corresponding three correlative value C according to the component content of all gases ₂H ₂/ C ₂H ₄, CH ₄/ H ₂, C ₂H ₄/ C ₂H ₆, give corresponding coding to every correlative value then, obtain a coding schedule according to coding, judge according to the diagnostic criteria that provides in the coding schedule whether transformer breaks down and the type of fault (seeing Table 2-1, table 2-2).

The coding rule of table 2-1 three-ratio method

Table 2-2 three-ratio method is judged the diagnostic criteria of transformer fault

Scope in three correlative values of characteristic gas and the coding schedule is compared, be easy to just can judge whether transformer breaks down and the type of fault.But there is following deficiency in the IEC three-ratio method: 1. because power transformer interior fault is very complicated, the fault of coded combination correspondence in the fault type determination methods coding schedule scope usually appears being not included in coded combination that the three-ratio method that is obtained by typical Accident Statistical Analysis is recommended in actual applications.2. have only in the oil that each component concentration of gas is enough high or ultra crosses demand value, and determine that through analysis-by-synthesis there is fault in transformer inside, could further judge the character of its fault with three-ratio method.No matter if whether transformer exists fault, use three-ratio method without exception, just might cause wrong diagnosis to normal transformer.3. in actual applications, when the various faults synergy, may in coded combination and fault type mapping table, can not find corresponding ratio combination; Simultaneously, the fault in the ratio interval of three ratios coding obscurity boundary, often erroneous judgement easily.Because there is ambiguity in failure modes, malfunction may cause the various faults feature, and a kind of fault signature also can reflect the various faults state in varying degrees, so three-ratio method can not reflect fault state comprehensively.There is ambiguity fault in itself, also has ambiguity between each group coding and the fault type, and three ratios fail to comprise and reflect all forms of power transformer interior fault.

Summary of the invention

The purpose of this invention is to provide a kind of intelligent method of the transformer fault diagnosis based on the RBF neural network, higher inferential capability and diagnostic accuracy are not only arranged, also can accurately reflect all forms of transformer fault.

The present invention adopts following technical proposals: a kind of intelligent method of the transformer fault diagnosis based on the RBF neural network may further comprise the steps: (1), utilize the IEC three-ratio method to draw three ratios of five kinds of gases: C ₂H ₂/ C ₂H ₄, CH ₄/ H ₂, C ₂H ₄/ C ₂H ₆As the training sample data; (2), utilize subordinate function that three ratios are carried out Fuzzy processing; (3), fault type is encoded; (4), according to training sample data training RBF neural network, satisfy accuracy requirement up to the RBF network, the sample to be diagnosed after (5), the input Fuzzy processing; (6), output diagnostic result.

The subordinate function that adopts in the described step (2) is normal distyribution function, establishes χ ₁=C ₂H ₂/ C ₂H ₄, χ ₂=CH ₄/ H ₂, χ ₃=C ₂H ₄/ C ₂H ₆, a ₁, a ₂, a ₃The actual input quantity of the neural network input layer of crossing for Fuzzy processing, then subordinate function is:

$a_1=\left\{\begin{array}{cc}0,& ,0\&le;{\mathrm{\&chi;}}_1<0.1\\ 1-e^{-0.5{({\mathrm{\&chi;}}_1-0.1)}^2},& 0.1\&le;{\mathrm{\&chi;}}_1<3\\ 1,& {\mathrm{\&chi;}}_1\&GreaterEqual;3\end{array}\right\}$

$a_2=\left\{\begin{array}{cc}0.5,& 0\&le;{\mathrm{\&chi;}}_2<0.1\\ 0,& 0.1\&le;{\mathrm{\&chi;}}_2<1\\ 1,& {\mathrm{\&chi;}}_2\&GreaterEqual;1\end{array}\right\}$

$a_3=\left\{\begin{array}{cc}0,& ,0\&le;{\mathrm{\&chi;}}_3<1\\ 1-e^{-{({\mathrm{\&chi;}}_3-1)}^2},& 1\&le;{\mathrm{\&chi;}}_3<3\\ 1,& {\mathrm{\&chi;}}_3\&GreaterEqual;3\end{array}\right\}.$

Fault type coding in the described step (3) comprises: if be encoded to 100000, then fault type corresponds to the low energy discharge; If be encoded to 010000, temperature was overheated during then fault type corresponded to; If be encoded to 001000, then fault type corresponds to high-energy discharge; If be encoded to 000100, then fault type corresponds to hyperthermia and superheating; If be encoded to 000010, then to correspond to ground temperature overheated for fault type; If be encoded to 000001, then fault type corresponds to shelf depreciation.

Training the RBF neural network according to sample data in the described step (4) is with the training of MATLAB software, three input nodes are three ratios after the obfuscation, six output nodes are six kinds of fault type codings: it is 0.05 that training precision is set, and dispersion constant is 5.

Also include the step with RBF neural network output data obfuscation, the data greater than 0.5 be defined as 1, other be 0.

The present invention has following beneficial effect:

1. utilize subordinate function that the input data are carried out Fuzzy Processing, enlarged the diagnostic area of neural network, improved the degree of accuracy of neural network diagnosis;

2. adopt the C after bluring ₂H ₂/ C ₂H ₄, CH ₄/ H ₂, C ₂H ₄/ C ₂H ₆Three correlative values are as the input node of neural network, and the input dimension of having simplified neural network has been accelerated the speed of training;

3. creating the RBF neural network diagnoses transformer fault, utilize the neural network hidden layer that input vector is carried out conversion, the pattern input data of low level are transformed in the high bit space, the linear inseparable problem in the lower dimensional space that makes is converted at the higher dimensional space linear separability, has improved inferential capability and the diagnostic accuracy of network.

Description of drawings

Fig. 1 is method flow diagram of the present invention;

Fig. 2 is the structural drawing of RBF neural network among the present invention;

Error variation diagram one when Fig. 3 is the RBF neural metwork training;

Fig. 4 is the actual diagnostic result figure one of RBF neural network;

Error variation diagram two when Fig. 5 is the RBF neural metwork training;

Fig. 6 is the actual diagnostic result two of RBF neural network.

Embodiment

The present invention proposes a kind of intelligent method of the transformer fault diagnosis based on the RBF neural network, as shown in Figure 1, specifically may further comprise the steps:

(1), utilize the IEC three-ratio method to draw three ratios of five kinds of gases: C ₂H ₂/ C ₂H ₄, CH ₄/ H ₂, C ₂H ₄/ C ₂H ₆As the training sample data;

(2), utilize subordinate function that three ratios are carried out Fuzzy processing;

(3), fault type is encoded;

(4), according to the corresponding relation of three ratio data after the obfuscation and fault type as sample data, training RBF neural network enters next step after the RBF network satisfies accuracy requirement;

(5), the sample to be tested after the input Fuzzy processing;

(6), output diagnostic result.

Radially base net network neural network described in the present invention (being called for short the RBF network) is three layers of feedforward network, is made up of input layer, hidden layer, output layer, and concrete structure as shown in Figure 2.

Ground floor is input layer, and each node of this layer directly and each component χ of input vector _iConnect, it plays a part the input data are sent to down one deck, and the node number is n.

The second layer is hidden layer, and each latent node all is a RBF node, represents an independent radial basis function relevant with the expansion constant with the center, handles the input data by radial basis function as transport function.Radial basis function is Gaussian function, realizes that by the Euclidean distance that calculates between input vector (χ) and the radial basis function center data are in the non-linear transmission of hidden layer.As formula (1):

$h_j\left(\mathrm{\&chi;}\right)=\exp (-\frac{{\left|\right|\mathrm{\&chi;}-c_j\left|\right|}^2}{r_j^2})---\left(1\right)$

H wherein _j(χ) be j RBF node output, c _jAnd r _jBe respectively central value and the expansion constant of j RBF node.

The 3rd layer is output layer, is linear unit, realizes network output, represents as formula (2).

$y_k\left(\mathrm{\&chi;}\right)=\underset{j=1}{\overset{m}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_{\mathrm{kj}}{h}_j\left(\mathrm{\&chi;}\right)+b_k;$ j＝1，2，3，...，m (2)

y _k(χ) be network for κ output of input vector (χ), m is latent node number, ω _KjBe the connection weight of κ output node and j latent node, b _kIt is base.

" base " of hidden unit is radial basis function, it constitutes the hidden layer space, input vector is carried out conversion by hidden layer, the pattern of low-dimensional is imported data transform in the higher dimensional space, makes that the linear inseparable problem in the lower dimensional space is converted at the higher dimensional space linear separability.

From the characteristic gas sample data, randomly draw 100 groups of data as the neural network training data sample, make C ₂H ₂/ C ₂H ₄, CH ₄/ H ₂, C ₂H ₄/ C ₂H ₆Three ratios (partial data is shown in table 3-1).

Table 3-1 fault signature gas sample and ratio

Fault type is encoded, as table 3-2.

Table 3-2 fault type coding schedule

Fault type	Coding
		The low energy discharge	1 0 0 0 0 0
Middle temperature is overheated	0 1 0 0 0 0
		High-energy discharge	0 0 1 0 0 0
Hyperthermia and superheating	0 0 0 1 0 0
		Cryogenic overheating	0 0 0 0 1 0
Shelf depreciation	0 0 0 0 0 1

For embodying the uncertain factor in the transformer fault diagnosis, use fuzzy mathematics theory to finish data pre-service work.Set up C respectively ₂H ₂/ C ₂H ₄, CH ₄/ H ₂, C ₂H ₄/ C ₂H ₆The subordinate function of three ratios is asked for the actual input quantity of neural network.

If χ ₁=C ₂H ₂/ C ₂H ₄, χ ₂=CH ₄/ H ₂, χ ₃=C ₂H ₄/ C ₂H ₆, a ₁, a ₂, a ₃Be the actual input quantity of neural network input layer, then subordinate function is:

$a_1=\left\{\begin{array}{cc}0,& ,0\&le;{\mathrm{\&chi;}}_1<0.1\\ 1-e^{-0.5{({\mathrm{\&chi;}}_1-0.1)}^2},& 0.1\&le;{\mathrm{\&chi;}}_1<3\\ 1,& {\mathrm{\&chi;}}_1\&GreaterEqual;3\end{array}\right\}$

$a_2=\left\{\begin{array}{cc}0.5,& 0\&le;{\mathrm{\&chi;}}_2<0.1\\ 0,& 0.1\&le;{\mathrm{\&chi;}}_2<1\\ 1,& {\mathrm{\&chi;}}_2\&GreaterEqual;1\end{array}\right\}$

$a_3=\left\{\begin{array}{cc}0,& ,0\&le;{\mathrm{\&chi;}}_3<1\\ 1-e^{-{({\mathrm{\&chi;}}_3-1)}^2},& 1\&le;{\mathrm{\&chi;}}_3<3\\ 1,& {\mathrm{\&chi;}}_3\&GreaterEqual;3\end{array}\right\}.$

By subordinate function to C ₂H ₂/ C ₂H ₄, CH ₄/ H ₂, C ₂H ₄/ C ₂H ₆Three ratios carry out Fuzzy Processing, and result (partial data is shown in table 3-3) obtains a ₁, a ₂, a ₃Input sample of data as the RBF neural network.

Table 3-3 Fuzzy Processing data

Choose 60 groups of data as the data to be tested sample from former characteristic gas sample data, partial data is shown in table 3-4.

Table 3-4 data to be tested sample

Utilize subordinate function to above-mentioned data to be tested sample Fuzzy processing, obtain the fuzzy numerical value a1 of data to be tested sample, a2, a3, partial data is shown in table 3-5.

The fuzzy numerical value of table 3-5 data to be tested sample

The training of neural network: according to the data sample input/output relation shown in the table 3-1, train the RBF neural network with the newrb function.According to definite principle of RBF neural network model structure, determine that network has a ₁, a ₂, a ₃Three input nodes, three correlative values of corresponding 5 kinds of characteristic gas, 6 output nodes (malfunction coding in the corresponding tables two), calling the RBF neural network procedure in MATLAB trains selected sample data, it is 0.06 that training precision is set, dispersion constant is 10 o'clock, and in MATLAB software training process, the error variation diagram as shown in Figure 3.The actual diagnostic result of RBF as shown in Figure 4.

The result carries out Fuzzy processing to output, and regulation output data are 1 greater than 0.5 data, less than 0.5 be 0, diagnostic result is compared with the actual sample data, have 48 groups of diagnosis correct in 60 groups of data, accuracy reaches 70%.

It is 0.05 that training precision is set, and dispersion constant is 5 o'clock, and in the MATLAB training process, the error variation diagram as shown in Figure 5.The actual diagnostic result of RBF as shown in Figure 6.

The result carries out Fuzzy processing to output, and regulation output data are 1 greater than 0.5 data, less than 0.5 be 0, diagnostic result is compared with the actual sample data, in 60 groups of data 53 groups of diagnosis correct, accuracy reaches 88.3%.

Utilize the IEC three-ratio method, according to the diagnostic criteria of table 2-2 judgement transformer fault, 60 stack features gas data samples are carried out fault judge, the result is as table 3-6.

Table 3-6

3-6 can draw the IEC three-ratio method to the result of transformer fault diagnosis by table, in the 60 stack features gas sample data, have 25 groups because of what ratio exceeded that coding range can't judge, wrongheaded have 12 groups, and final IEC three-ratio method is about 38.3% to the accuracy of transformer fault diagnosis.And utilize fuzzy neural network that transformer fault is carried out accurate rate of diagnosis up to 88.3%, be higher than far away and utilize the IEC three-ratio method that transformer fault is carried out accurate rate of diagnosis.

Claims (5)

1. intelligent method based on the transformer fault diagnosis of RBF neural network is characterized in that: may further comprise the steps: (1), utilize the IEC three-ratio method to draw three ratios of five kinds of gases: C ₂H ₂/ C ₂H ₄, CH ₄/ H ₂, C ₂H ₄/ C ₂H ₆As the training sample data; (2), utilize subordinate function that three ratios are carried out Fuzzy processing; (3), fault type is encoded; (4), according to training sample data training RBF neural network, satisfy accuracy requirement up to the RBF network, the sample to be diagnosed after (5), the input Fuzzy processing; (6), output diagnostic result.

2. the intelligent method of the transformer fault diagnosis based on the RBF neural network according to claim 1, it is characterized in that: the subordinate function that adopts in the described step (2) is normal distyribution function, establishes

_,

,

,

,

, The actual input quantity of the neural network input layer of crossing for Fuzzy processing, then subordinate function is:

。

3. the intelligent method of the transformer fault diagnosis based on the RBF neural network according to claim 1, it is characterized in that: the fault type coding in the described step (3) comprises: if be encoded to 100000, then fault type corresponds to the low energy discharge; If be encoded to 010000, temperature was overheated during then fault type corresponded to; If be encoded to 001000, then fault type corresponds to high-energy discharge; If be encoded to 000100, then fault type corresponds to hyperthermia and superheating; If be encoded to 000010, then to correspond to ground temperature overheated for fault type; If be encoded to 000001, then fault type corresponds to shelf depreciation.

4. the intelligent method of the transformer fault diagnosis based on the RBF neural network according to claim 1, it is characterized in that: training the RBF neural network according to sample data in the described step (4) is with the training of MATLAB software, three input nodes are three ratios after the obfuscation, six output nodes are six kinds of fault type codings: it is 0.05 that training precision is set, and dispersion constant is 5.

5. the intelligent method of the transformer fault diagnosis based on the RBF neural network according to claim 4 is characterized in that: also include the step with RBF neural network output data obfuscation, the data greater than 0.5 be defined as 1, other be 0.

Images