CN110161389A - A kind of Electric Power Equipment Insulation defect identification method and AEVB self-encoding encoder - Google Patents

Abstract

The invention discloses a kind of Electric Power Equipment Insulation defect identification methods comprising step: (1) constructing AEVB self-encoding encoder, AEVB self-encoding encoder includes probability encoding device and probability decoder；(2) AEVB self-encoding encoder is trained using the history Partial Discharge Data of power equipment, so that: characteristic value of the history Partial Discharge Data of input in the output end output Partial Discharge Data of probability encoding device, the characteristic value input probability decoder of the Partial Discharge Data, to export corresponding insulation defect type in the output end of probability decoder；(3) by the local discharge signal input AEVB self-encoding encoder of power equipment to be identified, AEVB self-encoding encoder exports insulation defect type.In addition, the invention also discloses a kind of AEVB self-encoding encoder for Electric Power Equipment Insulation defect recognition, AEVB self-encoding encoder includes probability encoding device and probability decoder.

Description

A Method for Identifying Insulation Defects of Power Equipment and AEVB Autoencoder

technical field

The invention relates to an identification method and its neural network, in particular to an identification method and its neural network for electrical equipment insulation defects.

Background technique

At present, the commonly used methods for diagnosis of insulation defects are feature extraction and pattern recognition for partial discharge maps constructed from partial discharge information. Quickly and accurately identifying the obtained partial discharge spectrum is helpful for the diagnosis of insulation defects, and then grasps the insulation status of power equipment. When there is an insulation defect, it can be repaired at the first time. However, the current partial discharge spectrum identification technology still has the problem of low accuracy, which may easily lead to misjudgment of faults.

Based on this, it is expected to obtain a method for identifying insulation defects of power equipment, which can significantly improve the recognition accuracy of partial discharge maps, so as to better evaluate the status of power equipment, and help to grasp the insulation status of power equipment.

Contents of the invention

One of the objectives of the present invention is to provide a method for identifying insulation defects in electric equipment, which can effectively improve the recognition accuracy of partial discharge atlases, thereby better evaluating the state of electric equipment, so as to facilitate grasping Insulation state of electrical equipment.

Based on the above purpose, the present invention proposes a method for identifying electrical equipment insulation defects, which includes the steps of:

(1) Build AEVB self-encoder, AEVB self-encoder includes probability encoder and probability decoder;

(2) Use the historical partial discharge data of electric equipment to train the AEVB autoencoder, so that: the input historical partial discharge data outputs the eigenvalue of the partial discharge data at the output end of the probability encoder, and the eigenvalue of the partial discharge data input into a probabilistic decoder to output the corresponding insulation defect type at the output of the probabilistic decoder;

(3) Input the partial discharge signal of the power equipment to be identified into the AEVB self-encoder, and the AEVB self-encoder outputs the type of insulation defect.

In the method for identifying insulation defects of power equipment described in the present invention, in order to improve the recognition accuracy of partial discharge maps, the inventors of this case proposed the method for identifying insulation defects of power equipment in this case based on the deep autoencoder matching algorithm. Decibelsian autoencoder (hereinafter referred to as AEVB autoencoder), the AEVB autoencoder includes a probabilistic encoder and a probabilistic decoder, and then uses the historical partial discharge data of power equipment to train the AEVB autoencoder, so that : The input historical partial discharge data outputs the characteristic value of the partial discharge data at the output terminal of the probability encoder, and the characteristic value of the partial discharge data is input into the probability decoder to output the corresponding insulation defect type at the output terminal of the probability decoder. The historical partial discharge data can be obtained from the case information database. Finally, the partial discharge signal of the power equipment to be identified is input into the AEVB self-encoder, and the insulation defect type is output through the AEVB self-encoder.

As a result, the method for identifying electrical equipment insulation defects in this case can accurately and effectively evaluate and understand the operation of electrical equipment, improve the recognition accuracy of partial discharge maps, and thus better evaluate the status of electrical equipment, in order to benefit Master the insulation status of electrical equipment.

Further, in the electrical equipment insulation defect identification method of the present invention, the AEVB self-encoder has an input layer, an output layer, a hidden variable layer and four intermediate layers; wherein one input layer, two intermediate layers and A hidden variable layer forms the probability encoder, and the hidden variable layer outputs the characteristic value of partial discharge data; two intermediate layers and an output layer form the probability decoder, and the output layer outputs the type of insulation defect.

Furthermore, in the method for identifying insulation defects of electric equipment according to the present invention, the parameters of the probabilistic encoder and the probabilistic decoder are optimized by using the stochastic gradient descent method.

Further, in the method for identifying electrical equipment insulation defects according to the present invention, the probabilistic encoder is characterized by the following formula:

In the formula, z represents the output of the hidden variable layer, which is the eigenvalue of the partial discharge data; σ _enc represents the proportional parameter; ε represents the random parameter satisfying the N(0,1) distribution; N(0,I) is the standard normal distribution ; f represents the activation function; Indicates the proportional parameter to be solved; Indicates the translation amount to be solved; Indicates the proportional parameter to be solved; Indicates the translation amount to be solved; Indicates the scale parameter of h _l to be solved; Indicates the translation amount of h _l to be solved; x is the input partial discharge data.

Further, in the method for identifying electrical equipment insulation defects according to the present invention, the probabilistic decoder is characterized by the following formula:

In the formula, x represents the input partial discharge data; z represents the eigenvalue of the partial discharge data; p(x|z) represents the conditional probability distribution function; N represents the normal distribution function; μ _dec represents the probability likelihood parameter, σ _dec represents Scale parameter; I means offset; Indicates the proportional parameter to be solved; Indicates the translation amount to be solved; Indicates the proportional parameter to be solved; Indicates the translation amount to be solved; Indicates the scale parameter of h2 to be solved _; Indicates the translation of _h2 to be solved.

Correspondingly, another object of the present invention is to provide an AEVB self-encoder for identifying insulation defects in electric equipment, through which the AEVB self-encoder can effectively improve the recognition accuracy of partial discharge patterns, thereby better identifying electric equipment The state is evaluated to help grasp the insulation state of the power equipment.

Based on the above-mentioned purpose, the present invention also proposes a kind of AEVB self-encoder for electric equipment insulation defect identification, AEVB self-encoder comprises probability encoder and probability decoder; Wherein the probability encoder is structured as: input electric power to it The partial discharge signal of the equipment, then it outputs the eigenvalue of the partial discharge data; the said probabilistic decoder is structured as: its input terminal inputs the output of the probability encoder, and its output is the type of insulation defect.

Further, in the AEVB self-encoder of the present invention, the AEVB self-encoder has an input layer, an output layer, a hidden variable layer and four intermediate layers; wherein an input layer, two intermediate layers and a hidden variable layer The variable layer constitutes the probability encoder, and the hidden variable layer outputs the characteristic value of the partial discharge data; two intermediate layers and an output layer constitute the probability decoder, and the output layer outputs the type of insulation defect.

Furthermore, in the AEVB self-encoder described in the present invention, the parameters of the probabilistic encoder and the probabilistic decoder are optimized using the stochastic gradient descent method.

Further, in the AEVB self-encoder of the present invention, the probabilistic encoder is characterized by the following formula:

In the formula, z represents the output of the hidden variable layer, which is the eigenvalue of the partial discharge data; σ _enc represents the proportional parameter; ε represents the random parameter satisfying the N(0,1) distribution; N(0,I) is the standard normal distribution ; f represents the activation function; Indicates the proportional parameter to be solved; Indicates the translation amount to be solved; Indicates the proportional parameter to be solved; Indicates the translation amount to be solved; Indicates the proportional parameter to be solved; Indicates the translation amount to be solved; x is the input partial discharge data.

Further, in the AEVB self-encoder of the present invention, the probabilistic decoder is characterized by the following formula:

In the formula, x represents the input partial discharge data; z represents the eigenvalue of the partial discharge data; p(x|z) represents the conditional probability distribution function; N represents the normal distribution function; μ _dec represents the probability likelihood parameter, σ _dec represents Scale parameter; I means offset; Indicates the proportional parameter to be solved; Indicates the translation amount to be solved; Indicates the proportional parameter to be solved; Indicates the translation amount to be solved; Indicates the scale parameter of h2 to be solved _; Indicates the translation of _h2 to be solved.

The electrical equipment insulation defect identification method and the AEVB self-encoder of the present invention have the following advantages and beneficial effects:

The method for identifying insulation defects of electric equipment according to the present invention can effectively improve the recognition accuracy of partial discharge atlas, thereby better evaluating the state of electric equipment, so as to facilitate grasping the insulation state of electric equipment.

In addition, the AEVB self-encoder of the present invention also has the above-mentioned advantages and beneficial effects.

Description of drawings

Fig. 1 is a schematic flow chart of an embodiment of the method for identifying insulation defects of electric equipment according to the present invention.

Fig. 2 schematically shows the structure of the AEVB autoencoder in an embodiment of the present invention.

FIG. 3 to FIG. 6 illustrate the detection data detected in different implementations of the method for identifying insulation defects of electrical equipment according to the present invention.

Detailed ways

The method for identifying electrical equipment insulation defects and the AEVB self-encoder according to the present invention will be further described below based on specific embodiments and the accompanying drawings, but this description does not constitute an improper limitation to the technical solution of the present invention.

Fig. 1 is a schematic flow chart of an embodiment of the method for identifying insulation defects of electric equipment according to the present invention.

As shown in FIG. 1, in this embodiment, the method for identifying insulation defects of electric equipment includes steps:

Step 100: construct AEVB self-encoder, described AEVB self-encoder comprises probabilistic encoder and probabilistic decoder;

Step 200: Use the historical partial discharge data of the electric equipment to train the AEVB autoencoder, so that: the input historical partial discharge data outputs the eigenvalue of the partial discharge data at the output end of the probability encoder, and the eigenvalue of the partial discharge data input into a probabilistic decoder to output the corresponding insulation defect type at the output of the probabilistic decoder;

Step 300: Input the partial discharge signal of the power equipment to be identified into the AEVB self-encoder, and the AEVB self-encoder outputs the type of insulation defect.

It should be noted that the AEVB autoencoder has an input layer, an output layer, a hidden variable layer and four intermediate layers; wherein an input layer, two intermediate layers and a hidden variable layer form the probability encoder, The hidden variable layer outputs the characteristic value of the partial discharge data; the probability decoder is composed of two intermediate layers and an output layer, and the output layer outputs the insulation defect type.

Fig. 2 schematically shows the structure of the AEVB autoencoder in an embodiment of the present invention.

As shown in Figure 2, the parameters of the probabilistic encoder and the probabilistic decoder are optimized by stochastic gradient descent. Among them, the probabilistic encoder is characterized by the following formula:

In the formula, z represents the output of the hidden variable layer, which is the eigenvalue of the partial discharge data; σ _enc represents the proportional parameter; ε represents the random parameter satisfying the N(0,1) distribution; N(0,I) is the standard normal distribution ; f represents the activation function; Indicates the proportional parameter to be solved; Indicates the translation amount to be solved; Indicates the proportional parameter to be solved; Indicates the translation amount to be solved; Indicates the proportional parameter to be solved; Indicates the translation amount to be solved; x is the input partial discharge data.

The probabilistic decoder is characterized by the following formula:

In the formula, x represents the input partial discharge data; z represents the eigenvalue of the partial discharge data; p(x|z) represents the conditional probability distribution function; N represents the normal distribution function; μ _dec represents the probability likelihood parameter, σ _dec represents Scale parameter; I means offset; Indicates the proportional parameter to be solved; Indicates the translation amount to be solved; Indicates the proportional parameter to be solved; Indicates the translation amount to be solved; Indicates the scale parameter of h2 to be solved _; Indicates the translation of _h2 to be solved.

In order to better illustrate the recognition effect of the method for identifying insulation defects of electric equipment in this case, Fig. 3 to Fig. 5 illustrate the detection data detected in different implementations of the method for identifying insulation defects of electric equipment according to the present invention.

Table 1 lists the description of the relationship between the type of discharge detected by different partial discharge signals and the location of the discharge.

Table 1.

case number discharge type discharge site 1 Levitation discharge Isolate the connecting part of the insulating pull rod of the knife switch and the transmission mechanism 2 Levitation discharge Isolate the connecting part of the insulating pull rod of the knife switch and the transmission mechanism 3 Levitation discharge Built-in sensor connector 4 Insulation type discharge GIS cable terminal compartment The cable terminal connector is damaged

In addition, it should be noted that the case information is shown in Table 1. Among them, Case 1 and Case 2 are the discharge cases detected by the same equipment manufacturer, the same type of equipment, and the same location. Case 3 has the same discharge type as Case 1 and Case 2, but the equipment manufacturers and discharge locations are different. Case 4 is a comparative case of different discharge types. The partial discharge data detected in the above four cases are shown in Fig. 3 to Fig. 6 . By extracting various eigenvalues of the partial discharge data shown in Figure 3 to Figure 6, and combining the cosine algorithm to calculate the matching degree between each other, the matching results of different eigenvalues are obtained as shown in Table 2 below.

Table 2.

Note: The statistical eigenvalues in Table 2 can be obtained in the following ways: the statistical eigenvalues are determined by the partial discharge amplitude and frequency in the entire power frequency cycle and the skewness Sk, steepness Ku, and asymmetry Q of the positive and negative half cycles of the power frequency and the cross-correlation coefficient Cc and other 16 characteristic parameters;

The DBN eigenvalues are obtained in the following way: the deep belief network is a network structure model formed by superposition and expansion of multi-layer restricted Boltzmann machines. The DBN network used for comparison in this paper is 6 layers, and the number of nodes is 3600, 1000, 500, 100, 10, 4. The output of the fifth layer is used as the extracted feature value;

CNN eigenvalues are obtained through the following methods: deep convolutional network is a network structure model that uses multi-layer convolution and pooling operations to obtain deep features of data, and can achieve translation, scaling, and distortion invariance to data. The CNN network input layer used in this paper is 50×72, and the two convolutional layers are 6 3×3 convolution kernels and 36 3×3 convolution kernels, and the corresponding pooling layer is 1×2, 1 ×11. The two fully connected layers are 500, 10, and the output layer is 4. The output of the second fully connected layer is used as the extracted feature value;

Principal component analysis (PCA for short) + linear discriminant analysis (LDA for short) eigenvalues are obtained by the following method: firstly, PCA is used to reduce the dimension of the sample, and the redundancy of the sample is eliminated, so as to ensure the dispersion Non-singularity of matrices. Then use LDA to solve the optimal transformation feature information.

It should be noted that the numbers in Table 2 refer to extracting the feature values of the two case numbers in Table 1, and combining the cosine algorithm to calculate the matching degree between the two. Further, Case 1-2 refers to the use of extraction case 1 and various eigenvalues of Case 2, combined with the cosine algorithm to calculate the matching degree of the two. Similarly, Case 2-3 refers to the use of various eigenvalues of Case 2 and Case 3, combined with the cosine algorithm to calculate the matching degree of the two .

It can be seen from Table 2 that among the matching results obtained by the AEVB autoencoder, Case 1-2 has a higher matching degree than other case combinations, and the matching degree is 23.09% higher than Case 1-3. Compared with Cases 1-4, the matching degree is 89.94% higher. As a comparison, the matching degree calculated by the statistical eigenvalues is difficult to draw obvious rules. The matching degree between the data of the same discharge type is relatively close, and the matching degree between data of different discharge types is slightly lower, but it is better than that obtained by the AEVB autoencoder. The results are less effective. It can be seen from the results of DBN model, CNN model and PCA+LDA model that the matching degree between Case 1-4 and Case 2-4 is low, so it can achieve better recognition effect on data of different discharge types. However, for Case 1-2, Case 1-3, and Case 2-3, the obtained matching degrees are not much different, and similar cases cannot be distinguished, so the matching application effect is poor.

In order to verify the recognition effect of the AEVB autoencoder in this case, the feature value is extracted based on the AEVB autoencoder, and then the matching degree between the case data is calculated by using the cosine algorithm, the Euclidean distance and the best entropy. The results are shown in Table 3.

table 3.

case number cosine algorithm Euclidean distance optimal entropy Case 1-2 96.87% 93.03% 98.62% Case 1-3 73.78% 75.23% 74.44% Case 1-4 6.93% 8.31% 5.00% Case 2-3 61.75% 60.84% 77.43% Case 2-4 1.92% 6.80% 3.34% Case 3-4 9.51% 7.18% 7.15%

Note: The cosine algorithm in Table 3 can be obtained as follows: the cosine algorithm of the following formula calculates the distance between the partial discharge data, and then the matching degree MR of the partial discharge data can be obtained.

In the formula, V _a and V _b are the feature vectors extracted from two partial discharge data respectively, and ||·|| represents the modulus of the vector;

The matching degree of the Euclidean distance is obtained as follows: The matching degree thereof is obtained based on the Euclidean distance of two sets of vectors. The problem with matching based on Euclidean distance is that it is difficult to determine a suitable measure and thus difficult to normalize. In this paper, the maximum distance among all sample data is selected as the standard, and the matching degree is calculated according to the following formula:

The matching degree based on the best entropy is obtained as follows: the entropy value of the signal is mainly calculated, and the calculation parameters are few, which can reduce the error caused by time asynchrony to a certain extent.

It can be seen from Table 3 that there is little difference between using the Euclidean distance and the best entropy to calculate the matching degree and the cosine algorithm for the above cases.

Further calculate the matching degree of data in 200 million cases, use AEVB autoencoder to extract eigenvalues, and use cosine algorithm, Euclidean distance and optimal entropy to calculate the matching degree between pairs. The matching degree is higher than 80% in similar cases and lower than 20% in dissimilar cases as correct matching, and the statistical information shown in Table 4 is obtained by calculating the matching correct rate.

Table 4.

discharge type Levitation discharge insulation discharge cosine algorithm 82.6% 85.7% Euclidean distance 63.6% 75.3% optimal entropy 47.5% 72.4%

It can be seen from Table 4 that for a large number of cases, the matching accuracy rate based on Euclidean distance and optimal entropy is lower than that of the cosine algorithm, because in the matching degree calculation of Euclidean distance, all samples are compared with the fixed maximum distance , so it is prone to singular values, resulting in poor overall effect. The calculation of matching degree based on the best entropy has a faster calculation speed, but because the best entropy is related to the degree of dispersion of the sample, it is difficult to measure partial discharge data under complex data sources. Similarity works poorly.

In summary, it can be seen that the method for identifying insulation defects of electrical equipment according to the present invention can effectively improve the identification accuracy of partial discharge atlases, thereby better evaluating the status of electrical equipment, so as to help grasp the insulation status of electrical equipment .

In addition, the AEVB self-encoder of the present invention also has the above-mentioned advantages and beneficial effects.

It should be noted that the prior art part in the scope of protection of the present invention is not limited to the embodiments given in the application documents, and all prior art that does not contradict the solution of the present invention, including but not limited to the prior art Patent documents, prior publications, prior public use, etc., can all be included in the scope of protection of the present invention.

In addition, it should be noted that the combination of the technical features in this case is not limited to the combination described in the claims of this case or the combination described in the specific examples, all the technical features recorded in this case can be used in any way Free combination or combination, unless contradictory to each other.

The above-mentioned embodiment is a preferred embodiment of the present invention, but the embodiment of the present invention is not limited by the above-mentioned embodiment, and any other changes, modifications, substitutions, combinations, Simplifications should be equivalent replacement methods, and all are included in the protection scope of the present invention.

Claims (10)

1. A method for identifying electrical equipment insulation defects, characterized in that it comprises the steps of: (1) build AEVB self-encoder, described AEVB self-encoder comprises probability encoder and probability decoder; (2) Use the historical partial discharge data of electric equipment to train the AEVB autoencoder, so that: the input historical partial discharge data outputs the eigenvalue of the partial discharge data at the output end of the probability encoder, and the eigenvalue of the partial discharge data input into a probabilistic decoder to output the corresponding insulation defect type at the output of the probabilistic decoder; (3) Input the partial discharge signal of the power equipment to be identified into the AEVB self-encoder, and the AEVB self-encoder outputs the type of insulation defect. 2. the electrical equipment insulation defect identification method as claimed in claim 1, is characterized in that, described AEVB self-encoder has an input layer, an output layer, a latent variable layer and four intermediate layers; Wherein an input layer, Two intermediate layers and a hidden variable layer form the probability encoder, and the hidden variable layer outputs the eigenvalues of the partial discharge data; two intermediate layers and an output layer form the probability decoder, and the output layer Outputs the type of insulation defect. 3. The method for identifying insulation defects in power equipment according to claim 1, wherein the parameters of the probabilistic encoder and the probabilistic decoder are optimized using a stochastic gradient descent method. 4. The electrical equipment insulation defect identification method as claimed in claim 1, wherein the probability encoder is characterized by the following formula: In the formula, z represents the output of the hidden variable layer, which is the eigenvalue of the partial discharge data; σ _enc represents the proportional parameter; ε represents the random parameter satisfying the N(0,1) distribution; N(0,I) is the standard normal distribution ; f represents the activation function; Indicates the proportional parameter to be solved; Indicates the translation amount to be solved; Indicates the proportional parameter to be solved; Indicates the translation amount to be solved; Indicates the proportional parameter to be solved; Indicates the translation amount to be solved; x is the input partial discharge data. 5. The electrical equipment insulation defect identification method as claimed in claim 4, wherein the probabilistic decoder is characterized by the following formula: In the formula, x represents the input partial discharge data; z represents the eigenvalue of the partial discharge data; p(x|z) represents the conditional probability distribution function; N represents the normal distribution function; μ _dec represents the probability likelihood parameter, σ _dec represents Scale parameter; I means offset; Indicates the proportional parameter to be solved; Indicates the translation amount to be solved; Indicates the proportional parameter to be solved; Indicates the translation amount to be solved; Indicates the scale parameter of h2 to be solved _; Indicates the translation of _h2 to be solved. 6. An AEVB self-encoder for identification of electrical equipment insulation defects, characterized in that, the AEVB self-encoder comprises a probability encoder and a probability decoder; wherein the probability encoder is structured to: input the Partial discharge signal, then it outputs the eigenvalue of partial discharge data; the probabilistic decoder is constructed as follows: its input terminal inputs the output of the probability encoder, and its output is the type of insulation defect. 7. AEVB self-encoder as claimed in claim 6, is characterized in that, described AEVB self-encoder has an input layer, an output layer, a latent variable layer and four intermediate layers; Wherein one input layer, two The middle layer and a hidden variable layer form the probability encoder, and the hidden variable layer outputs the characteristic value of the partial discharge data; two middle layers and an output layer form the probability decoder, and the output layer outputs the isolated defect type. 8. AEVB self-encoder as claimed in claim 6, is characterized in that, the parameter of described probability coder and probability decoder adopts stochastic gradient descent method to optimize. 9. AEVB self-encoder as claimed in claim 6, is characterized in that, described probabilistic encoder adopts following formula to represent: In the formula, z represents the output of the hidden variable layer, which is the eigenvalue of the partial discharge data; σ _enc represents the proportional parameter; ε represents the random parameter satisfying the N(0,1) distribution; N(0,I) is the standard normal distribution ; f represents the activation function; Indicates the proportional parameter to be solved; Indicates the translation amount to be solved; Indicates the proportional parameter to be solved; Indicates the translation amount to be solved; Indicates the proportional parameter to be solved; Indicates the translation amount to be solved; x is the input partial discharge data. 10. AEVB self-encoder as claimed in claim 9, is characterized in that, described probability decoder adopts following formula to represent: In the formula, x represents the input partial discharge data; z represents the eigenvalue of the partial discharge data; p(x|z) represents the conditional probability distribution function; N represents the normal distribution function; μ _dec represents the probability likelihood parameter, σ _dec represents Scale parameter; I means offset; Indicates the proportional parameter to be solved; Indicates the translation amount to be solved; Indicates the proportional parameter to be solved; Indicates the translation amount to be solved; Indicates the scale parameter of h2 to be solved _; Indicates the translation of _h2 to be solved.