Oil-immersed transformer fault diagnosis method fusing digital twinborn model

Abstract

The invention relates to a transformer fault diagnosis method. The invention discloses an oil-immersed transformer fault diagnosis method fusing a digital twinborn model, and the method specifically comprises the four steps: 1, achieving the equalization of a fault sample through employing the I NNOS; 2, extracting 9-dimensional candidate ratio signs, and performing fusion and dimension reduction on the ratio signs by using sparse principal component analysis (SPCA); 3, a kernel extreme learning machine (KELM) model is built, an AO algorithm is introduced to optimize parameters of the KELM model, and an AO-KELM optimal diagnosis model is built; and 4, establishing a digital twinborn five-dimensional model to achieve real-time interactive mapping between a physical entity and a virtual space and realize online diagnosis of the transformer fault. The transformer operation state can be simulated in real time, the fault diagnosis precision is improved, and intelligent application of transformer digital twinning is achieved.

Description

Technical Field

[0001] The present invention relates to the research field of transformer fault diagnosis, in particular to the technical field of transformer fault diagnosis. Background Art

[0002] As the hub of the power system, the health status of the transformer directly affects the stability and reliability of the power system operation. Therefore, accurately understanding the health status of the transformer is of great significance to ensure the safe and stable operation of the power system. When the insulation of the transformer ages, gases such as H2, CH4, C2H6, C2H4, C2H2, CO and CO2 will be dissolved in the insulating oil. The composition and concentration of the dissolved gas in the oil can reflect the current operating status of the transformer. Common analysis methods mainly include the IEC three-ratio method, Rogers ratio method, Doernenberg ratio method, etc.

[0003] With the development of machine learning theory, models such as Support Vector Machine (SVM), Extreme Learning Machine (ELM), Convolutional Neural Network (CNN), and Bidirectional Long Short Term Memory Network (BiLSTM) have achieved good application results in classification and recognition. However, the above methods require a large number of training samples to support them. In actual operation, the transformer failure rate is low, and the frequency of different types of faults also varies significantly. It is difficult to meet the training requirements of artificial intelligence diagnosis methods for unbalanced small samples, and it is easy to cause misjudgment and omission of minority class samples during the recognition process. At present, experts and scholars have conducted a lot of research to eliminate the imbalance of data sets. Undersampling and oversampling are used to solve the training problem of small sample unbalanced data sets. The undersampling algorithm mainly eliminates part of the majority class sample information, which is easy to cause the loss of effective information. Therefore, it is rarely used in transformer fault diagnosis problems. The oversampling algorithm generates minority samples to balance the number of samples of each fault category. The most widely used ones are Synthetic Minority Over-sampling Technique (SMOTE), Generative Adversarial Network, ADASYN (Adaptive Synthetic Sampling), etc., which can effectively solve the problems of misjudgment and omission of minority fault samples. However, the above methods cannot achieve real-time monitoring and diagnosis of transformers. Summary of the invention

[0004] The purpose of the present invention is to propose a fault diagnosis method for oil-immersed transformers integrating a digital twin model, so as to effectively improve the fault diagnosis accuracy of the transformer.

[0005] The present invention is implemented by the following technical solution: A fault diagnosis method for oil-immersed transformers integrating a digital twin model comprises the following steps:

[0006] Step 1: Iterative nearest neighbor oversampling is used to achieve fault sample balancing;

[0007] Step 2: Normalize and balance the sample set, extract 9-dimensional candidate ratio signs, use sparse principal component analysis to reduce the dimension of the ratio signs, and divide them into training set, test set and validation set;

[0008] Step 3: Build a kernel extreme learning machine model, introduce the Sky Eagle optimization algorithm to optimize its parameters, and establish the AO-KELM optimal diagnosis model;

[0009] Step 4: Establish a five-dimensional digital twin model to achieve real-time interactive mapping between physical entities and virtual space, and realize online diagnosis of transformer faults.

[0010] Preferably, the specific steps of step 1 are as follows:

[0011] Step 1: Define the imbalance of the sample set and use the standard deviation var(r) to measure the discreteness of the sample set;

[0012] Step 2: Based on the greedy search idea, find the minority class sample r j ;

[0013] Step 3: Continuously update the imbalance until it reaches the preset value and the iteration stops.

[0014] Preferably, the specific steps of step 2 are as follows:

[0015] Step 1: The equalized samples are recorded as m is the number of samples collected, and n is the number of characteristic gas types collected;

[0016] Specifically, the dissolved gases in the oil mainly include H2, CH4, C2H4, C2H6, and C2H2. Depending on whether the transformer fails and what kind of fault exists, the transformer can be divided into normal, high temperature overheating, high energy discharge, medium and low temperature overheating, low energy discharge, partial discharge, etc.

[0017] Step 2: Use z-score normalization on the equalized sample set, denoted as

[0018]

[0019] Step 3: Use the uncoded ratio method to construct 9-dimensional candidate ratio signs;

[0020] Specifically, the 9-dimensional candidate ratios include: CH4 / H2, C2H2 / C2H4, C2H4 / C2H6, CH4 / THC, C2H4 / THC, C2H6 / THC, C2H2 / THC, (CH4+C2H4) / THC and H2 / ALL

[0021] Step 4: Decompose the 9-dimensional ratio feature, calculate the principal component matrix T, and project the ratio feature to obtain a new matrix containing the information of the corresponding principal component.

[0022] Step 5: Introduce the LASSO penalty term to obtain a sparse solution and calculate the cumulative explained variance contribution CPV;

[0023] Step 6: Divide the reduced-dimensional data set into training set, test set, and validation set in proportion.

[0024] Preferably, the specific steps of step 3 are as follows:

[0025] Step 1: Construct the ELM model and use the least squares method to obtain the output weight value β * ;

[0026] Step 2: Introduce the kernel function to obtain the KELM weight output;

[0027] Step 3: Calculate the KELM model output function;

[0028] Step 4: Introduce the Sky Eagle optimization algorithm to optimize the regularization parameter C and the kernel function σ parameter, and establish the AO-KELM optimal diagnosis model.

[0029] Preferably, the specific steps of step 4 are as follows:

[0030] Step 1: Establish a twin mechanism model based on geometric model, physical model, behavioral model and rule model;

[0031] Step 2: Use the state perception device to collect multi-source heterogeneous data in real time to drive the twin model, so as to achieve real-time interactive mapping between physical entities and virtual space;

[0032] Step 3: Build a transformer twin database to achieve real-time data transmission between the physical space and the digital twin;

[0033] Step 4: Drive the monitored DGA data to realize online monitoring and fault identification and diagnosis of transformers.

[0034] Beneficial effects of the present invention: The present invention can not only enhance the learning ability of unbalanced samples and thus improve the recognition accuracy of minority fault samples, but also monitor and diagnose the transformer entity through the transformer digital twin. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a flowchart of the diagnostic method involved in the present invention;

[0036] Figure 2 is the Pareto diagram of the cumulative contribution of explained variance of SPCA in the embodiment;

[0037] Figure 3 is the fault diagnosis result in the embodiment;

[0038] Figure 4 is a comparison diagram of the diagnosis results under different feature extraction methods in the embodiment;

[0039] Figure 5 4 is a comparison chart of the diagnosis results of different diagnosis models in the embodiment. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solution and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.

[0041] Example 1

[0042] The oil-immersed transformer fault diagnosis method integrated with the digital twin model proposed in this embodiment includes the following steps:

[0043] Step 1: Use iterative neighbor oversampling to achieve fault sample balance:

[0044] Step 1: Define the imbalance of the sample set and use the standard deviation var(r) to measure the discreteness of the sample set;

[0045] Specifically, the imbalance IR:

[0046]

[0047] In the formula,

[0048] Step 2: Based on the greedy search idea, find the minority class sample r j ;

[0049] Specifically, the search process is as follows:

[0050]

[0051] In the formula, x j Represents the label data in category j. If x maxk To be the classification boundary, remove and reselect the next nearest neighbor point. Then, mark it as category j and remove it from the unlabeled dataset X U Removed from and added to the label dataset X L , and let r j =r j +1.

[0052] Step 3: Continuously update the imbalance until it reaches the preset value and the iteration stops.

[0053] Step 2: Normalize and balance the sample set, extract 9-dimensional candidate ratio signs, use sparse principal component analysis to reduce the dimension of the ratio signs, and divide them into training set, test set and validation set;

[0054] Step 1: The equalized samples are recorded as m represents the number of samples collected, and n represents the number of characteristic gas types collected;

[0055] Specifically, the dissolved gases in the oil mainly include H2, CH4, C2H4, C2H6, C2H2, CO, CO2, etc. Depending on whether the transformer has a fault and what kind of fault exists, the transformer can be divided into normal, high temperature overheating, high energy discharge, medium and low temperature overheating, low energy discharge, partial discharge, etc.

[0056] Step 2: Use z-score normalization on the equalized sample set, denoted as

[0057]

[0058] Step 3: Use the uncoded ratio method to construct 9-dimensional candidate ratio signs;

[0059] Specifically, the 9-dimensional candidate ratios include: CH4 / H2, C2H2 / C2H4, C2H4 / C2H6, CH4 / THC, C2H4 / THC, C2H6 / THC, C2H2 / THC, (CH4+C2H4) / THC, and H2 / ALL;

[0060] Step 4: Decompose the 9-dimensional ratio feature, calculate the principal component matrix T, and project the ratio feature to obtain a new matrix containing the information of the corresponding principal component.

[0061] Specifically, the formula for eigendecomposition is as follows:

[0062]

[0063] In the formula, Λ∈R m×mRepresented as a diagonal matrix of eigenvalues, arranged in descending order. Λ∈R m×m is a unitary matrix, and the column vector is the loading vector of the principal element.

[0064] Project T onto X to obtain a new matrix containing the information of the corresponding pivot element. The difference with X is set to E, as shown in formula (15-16):

[0065]

[0066]

[0067] Then we get:

[0068] X=TP T +E=XPP T +E

[0069] Ensure that the pivot is as close to the original data as possible, that is, require E to be as small as possible. Therefore, the pivot solution is as follows:

[0070]

[0071] In the formula, is the solution to the minimum value of the pivot matrix P.

[0072] Step 5: Introduce the LASSO penalty term to obtain a sparse solution and calculate the cumulative explained variance contribution CPV;

[0073] Specifically, the solution formula for sparse principal element is as follows:

[0074]

[0075] Where: Matrix A represents the load matrix expected to be obtained, and Matrix B is the load matrix expected to be obtained under the regression problem. A and B are m×k matrices, and Then we need to find the minimum value of matrix A and B, and b j ∝P j , λ and λ 1,j is the penalty term coefficient, and λ>0. The adjusted explained variance is as follows:

[0076]

[0077] Where SΛ is the diagonal matrix of explained variance after coefficients, is the load matrix after the coefficient. When the model contribution is greater than 85%, the principal component is extracted, as shown in the formula below:

[0078]

[0079] Step 6: Divide the reduced-dimensional data set into training set, test set, and validation set according to the ratio of 6:2:2.

[0080] Step 3: Build a kernel extreme learning machine model, introduce the Sky Eagle optimization algorithm to optimize its parameters, and establish the AO-KELM optimal diagnosis model;

[0081] Step 1: Construct the ELM model and use the least squares method to obtain the output weight value β * ;

[0082] Specifically, the formula is as follows:

[0083]

[0084] Where: the activation function is g(x), the number of hidden nodes is L, β j =[β j1 ,β j2 ,…,β jL ] T (j=1,2,…,L) is the output weight value connecting the jth hidden layer node and the output layer node, H={h ij}(i=1,2,…,N;j=1,2,…,L) is the hidden layer output matrix, H represents the input value corresponding to the jth column of the jth hidden layer node, and the output vector corresponding to the jth row input in H.

[0085] The output weight value is obtained using the least squares method as shown below:

[0086] β * =H'T

[0087] Where H' represents the generalized inverse matrix of the hidden layer output matrix H.

[0088] Step 2: Introduce the kernel function to obtain the KELM weight output;

[0089] Specifically, the KELM weight output is as follows:

[0090]

[0091] The KELM output function is as follows:

[0092]

[0093] When h(x) is unknown, the kernel function matrix is ​​as follows:

[0094] Ω ELM =HH T ;Ω ELMi,j =h(x i )h(x j )=K(xi ,x j )

[0095] In the formula, K(x i ,x j ) is the kernel function, expressed as:

[0096]

[0097] Step 3: Calculate the KELM model output function;

[0098]

[0099] Step 4: Introduce the Sky Eagle optimization algorithm to optimize the regularization parameter C and the kernel function σ parameter, and establish the AO-KELM optimal diagnosis model.

[0100] Specifically, the diagnostic model is trained on the training set; the accuracy of each iterative model is verified with the validation set. When the accuracy improvement is less than 0.001 after two consecutive trainings, the model training is completed and the model parameters are saved. Otherwise, the model training is repeated until the conditions are met; finally, the test set is sent to the trained AO-KELM model to finally obtain the transformer fault diagnosis result.

[0101] Step 4: Establish a five-dimensional digital twin model to achieve real-time interactive mapping between physical entities and virtual space, and realize online diagnosis of transformer faults;

[0102] Step 1: Establish a twin mechanism model based on geometric model, physical model, behavioral model and rule model;

[0103] Specifically, the expression is:

[0104] VE={Gv,Pv,Bv,Rv}

[0105] Where: Gv represents the geometric model, Pv represents the physical model, Bv represents the behavioral model, and Rv represents the rule model, which depicts the entity with high precision in the virtual space.

[0106] Specifically, the geometric model is a comprehensive description of the entities such as transformer box, windings, oil storage cabinets, high and low voltage bushings, etc. in terms of geometric dimensions, material properties, assembly relationships, etc.; the physical model is based on prior knowledge, physical properties and operation mechanism, and conducts modeling analysis and experiments on the magnetic field, structure, etc. of Gv to give the transformer physical properties; the behavioral model mainly describes the changes in the operating behavior of the transformer entity, and combines artificial intelligence algorithms to realize the creation of Bv; the rule model includes transformer industry standards and guidelines and transformer manuals, and performs rule reasoning and deduction on historical operation and maintenance data, and collects relevant parameters and characteristics of PE to make it consistent with real-time updates of VE.

[0107] Step 2: Use the state perception device to collect multi-source heterogeneous data in real time to drive the twin model, so as to achieve real-time interactive mapping between physical entities and virtual space;

[0108] Specifically, the status device mainly includes: dissolved gas sensor in oil, electric field sensor in oil, oil temperature / oil level sensor, SF6 gas density and temperature and humidity sensor, etc. Through sensors arranged in various parts, status data is collected in real time, and corresponding detection points are set in the virtual transformer.

[0109] Specifically, the working principle of the dissolved gas sensor in oil is as follows: after the dissolved gas in the transformer oil is sucked into the oil sample collection module through the oil pump, the oil and gas are quickly separated through the oil-gas separation module, and the dissolved gas in the oil is pushed into the chromatographic column by the carrier gas to separate each type of gas. The chromatographic peak diagram is obtained by the A / D converter, and the characteristic gas concentration value is converted into an electrical signal according to the peak diagram, transmitted to the database for conversion, processing and storage, and finally sent to the twin behavior model for comprehensive analysis and judgment, so as to realize online monitoring and diagnostic analysis of the transformer.

[0110] Furthermore, the collected multi-source heterogeneous data information will drive the twin model in real time to achieve real-time interactive mapping between physical entities and virtual space.

[0111] Step 3: Establish a twin database as a MySQL database, and upload and store the current status monitoring data of the transformer into the database through the 5G network;

[0112] Specifically, the database contains the entire life cycle data of the transformer from production to scrapping, real-time status monitoring data, historical maintenance management data, and data derived from intelligent algorithm calculations, etc., to achieve real-time data interaction and build a data transmission channel between the physical space and the digital twin.

[0113] Step 4: Drive the monitored DGA data to realize online monitoring and fault identification and diagnosis of transformers.

[0114] Example 2

[0115] 336 transformer fault samples were collected, representing the number of rows in the DGA sample matrix. Among them, 102 were normal, 70 were high-temperature overheating, 45 were medium-low temperature overheating, 64 were high-energy discharge, 20 were low-energy discharge, and 36 were partial discharge. The dissolved gases in the oil include H2, CH4, C2H4, C2H2, and C2H6, representing the number of columns in the DGA sample matrix. Depending on whether the transformer has a fault and what kind of fault exists, the transformer can be divided into normal, high-temperature overheating, medium-low temperature overheating, high-energy discharge, low-energy discharge, and partial discharge, and labeled 1 to 6.

[0116] First, the iterative nearest neighbor oversampling algorithm (INNO) is used to balance the sample data set. The specific distribution of the number of samples before and after processing is shown in Table 1.

[0117] Table 1 Comparison of training samples before and after preprocessing

[0118]

[0119] Then, the SPCA is used to fuse the equalized data, and the first five principal components with a cumulative contribution rate greater than 90% are selected as fusion features. The SPCA cumulative variance contribution rate is as follows: Figure 2 As shown. To verify the superiority of the feature fusion method based on SPCA, this paper qualitatively observes and quantitatively analyzes the performance of the SPCA method. PCA and SPCA are used to extract features from the preprocessed data respectively. The contribution rate of the SPCA principal component is slightly lower than that of PCA, indicating that the introduction of the norm makes the matrix sparse, loses part of the variance contribution rate, clarifies the relationship between variables, better removes redundant information in the ratio feature, and provides a more effective data basis for fault diagnosis.

[0120] Table 2 Comparison of cumulative variance contribution rates between SPCA and PCA

[0121]

[0122] The fusion features extracted from SPCA are divided into training set, test set and validation set according to the ratio of 6:2:2. The ACO optimization algorithm is used to optimize C and σ, and the ACO-HKELM diagnosis model is established. The diagnosis results are shown in Figure 3 shown.

[0123] To illustrate the performance of various feature extraction methods, SPCA and PCA fusion features, 9-dimensional joint features, and IEC three ratios were input into the diagnosis model. The results are shown in Figure 4 The recall rate (R), precision rate (P) and F1-score are introduced as the evaluation index system of the diagnosis model, and the evaluation results are shown in Table 3.

[0124] Table 3 Model evaluation results

[0125]

[0126] In order to verify the effectiveness of the new samples synthesized based on the INNO algorithm in improving the accuracy of transformer fault diagnosis, the samples were expanded using unbalanced number sets, random oversampling, SMOTE and ADASYN oversampling algorithms. The diagnosis results are shown in Figure 2. Figure 5 shown.

[0127] In order to verify the effectiveness of the ACO-HKELM diagnostic model, three diagnostic models, ELM, KELM, HKELM and ACO-HKELM, were selected. Ten-fold cross validation was used for each model. The diagnostic results are shown in Tables 4-5.

[0128] Table 4 Fault diagnosis results under different diagnosis models

[0129]

[0130] Table 5 Average diagnostic accuracy of different diagnostic models

[0131]

[0132] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, it should be understood by those skilled in the art that the technical solutions described in the aforementioned embodiments can still be modified, or some or all of the technical features therein can be replaced by equivalents. Therefore, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A fault diagnosis method for oil-immersed transformers integrating a digital twin model, characterized in that: The following steps are involved: Step 1: Iterative neighbor oversampling is used to achieve fault sample balancing; Step 2: Normalize and balance the sample set, extract 9-dimensional candidate ratio signs, use sparse principal component analysis to reduce the dimension of the ratio signs, and divide them into training set, test set and validation set; Step 3: Build a kernel extreme learning machine model, introduce the Sky Eagle optimization algorithm to optimize its parameters, and establish the AO-KELM optimal diagnosis model; Step 4: Establish a five-dimensional digital twin model to achieve real-time interactive mapping between physical entities and virtual space, and realize online diagnosis of transformer faults.

2. According to the oil-immersed transformer fault diagnosis method integrating the digital twin model according to claim 1, it is characterized in that: The specific steps of step 1 are as follows: Step 1: Define the imbalance of the sample set and use the standard deviation var(r) to measure the discreteness of the sample set; Step 2: Based on the greedy search idea, find the minority class sample r j ; Step 3: Continuously update the imbalance until it reaches the preset value and the iteration stops.

3. The oil-immersed transformer fault diagnosis method integrating a digital twin model according to claim 1 is characterized in that: The specific steps of step 2 are as follows: Step 1: The equalized samples are recorded as m is the number of samples collected, and n is the number of characteristic gas types collected; Specifically, the dissolved gases in the oil mainly include H2, CH4, C2H6, C2H4, C2H2, CO and CO2, etc. Depending on whether the transformer fails and what kind of failure exists, the transformer can be divided into normal, high temperature overheating, high energy discharge, medium and low temperature overheating, low energy discharge, and partial discharge; Step 2: Use z-score normalization on the equalized sample set, denoted as Step 3: Use the uncoded ratio method to construct 9-dimensional candidate ratio signs; Specifically, the 9-dimensional candidate ratios include: CH4 / H2, C2H2 / C2H4, C2H4 / C2H6, CH4 / THC, C2H4 / THC, C2H6 / THC, C2H2 / THC, (CH4+C2H4) / THC and H2 / ALL Step 4: Decompose the 9-dimensional ratio feature, calculate the principal component matrix T, and project the ratio feature to obtain a new matrix containing the information of the corresponding principal component. Step 5: Introduce the LASSO penalty term to obtain a sparse solution and calculate the cumulative explained variance contribution CPV; Step 6: Divide the reduced-dimensional dataset into training set, test set, and validation set in proportion.

4. The oil-immersed transformer fault diagnosis method integrating a digital twin model according to claim 1 is characterized in that: The specific steps of step 3 are as follows: Step 1: Construct the ELM model and use the least squares method to obtain the output weight value β * ; Step 2: Introduce the kernel function to obtain the KELM weight output; Step 3: Calculate the KELM model output function; Step 4: Introduce the Tianying optimization algorithm to optimize the regularization parameter C and the kernel function σ parameter, and establish the AO-KELM optimal diagnosis model.

5. The oil-immersed transformer fault diagnosis method integrating a digital twin model according to claim 1 is characterized in that: The specific steps of step 4 are as follows: Step 1: Establish a twin mechanism model based on geometric model, physical model, behavioral model and rule model; Step 2: Use the state perception device to collect multi-source heterogeneous data in real time to drive the twin model, so as to achieve real-time interactive mapping between physical entities and virtual space; Step 3: Build a transformer twin database to realize real-time data transmission between physical space and digital twin; Step 4: Drive the monitored DGA data to realize online monitoring and fault identification and diagnosis of transformers.

Images