A transformer oil-paper insulation fault diagnosis method

Abstract

The application relates to a transformer oil paper insulation fault diagnosis method, and the specific steps comprise the following steps: establishing an initial feature space and calculating the intra-group redundancy and inter-group correlation of the initial feature space; adopting an incremental search optimization algorithm to calculate the feature quantity sorting of the maximum correlation and minimum redundancy under a standard function difference, eliminating the features within a preset range at the rear of the feature merit sorting, obtaining a first feature selection threshold m1 and a first feature space; evaluating the importance of the feature quantity of the first feature space through a pre-trained GBDT model, eliminating the feature quantity with low importance, obtaining a second feature selection threshold m2 and a second feature space.

Description

Technical Field

[0001] This invention relates to the field of electronic fault detection technology, specifically a method for diagnosing faults in transformer oil-paper insulation. Background Technology

[0002] More than 30% of power grid failures are caused by transmission and transformation equipment, and most of these are insulation failures caused by the deterioration of the internal insulation system of oil-paper insulated transformers.

[0003] Research on feature mining using the time-domain dielectric response method is quite advanced. However, due to interference in the testing environment, using a single feature to diagnose the insulation state of a transformer is subject to randomness, resulting in non-unique evaluation results. While a comprehensive evaluation system integrating multiple features can compensate for the ambiguity of single feature evaluation, it suffers from the problem of lacking a reasonable basis for selecting the optimal feature space. Therefore, this embodiment proposes a two-level time-domain feature optimization strategy based on the principle of maximum correlation and minimum redundancy, combined with gradient ascending tree importance estimation. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention proposes a method for diagnosing transformer oil-paper insulation faults.

[0005] The technical solution of the present invention is as follows:

[0006] On the one hand, this invention proposes a method for diagnosing transformer oil-paper insulation faults, the specific steps of which include:

[0007] An initial feature space is established based on the set of time-domain initial features affecting oil-paper insulation faults. And calculate the initial feature space. Intra-group redundancy and inter-group correlation;

[0008] An incremental search optimization algorithm is used to calculate the ranking of features with maximum relevance and minimum redundancy under the standard function difference. Features that rank low within a preset range are removed to obtain the first feature selection threshold m1 and the first feature space.

[0009] The initial feature space is processed using a pre-trained GBDT model. The importance of features is evaluated, and features with low importance are removed to obtain the second feature selection threshold m2 and the second feature space.

[0010] Determining the first characteristic space based on Kendall's coefficient. Second feature space The rank correlation is used to comprehensively compare the rank correlation between the two feature spaces and the importance of the feature quantities to select the optimal feature space;

[0011] Transformer oil-paper insulation fault diagnosis based on selected feature space.

[0012] As a preferred embodiment, the initial feature space The specific formulas for calculating within-group redundancy and between-group correlation are as follows:

[0013]

[0014]

[0015] In the formula, maxD(s, c), D is the intra-group redundancy, minR(S), R is the inter-group correlation, S is the feature space; |S| is the number of features; c is the column vector of aging categories in the sample set; I(x i c) represents the mutual information between feature quantity i and aging category c in the feature space; I(x) i ,x j Let x be a feature quantity in the feature space. i With characteristic quantity x j Mutual information between them.

[0016] In a preferred embodiment, the feature quantity x in the feature space i With characteristic quantity x j Mutual information between I(x) i ,x j The specific calculation formula is as follows:

[0017]

[0018] In the formula, the characteristic quantity x i With characteristic quantity x j These are the sample column vectors for feature quantity i and feature quantity j in the initial feature set, respectively.

[0019] As a preferred embodiment, the specific steps for calculating the ranking of features with maximum correlation and minimum redundancy under the standard function difference using the incremental search optimization algorithm are as follows:

[0020] Based on the initial feature space The difference between the within-group redundancy and the between-group correlation is obtained by using the standard function of mutual information, as shown in the following formula:

[0021] maxΦ(D,R),Φ=DR

[0022] The optimal fault feature is obtained using an incremental search optimization algorithm. One feature is selected and then removed, and the feature sorting is reset to complete the process. The selection criteria for the optimal fault feature are as follows:

[0023]

[0024] In the formula, X is the initial feature set, and S is the initial feature set. m-1 It is a feature subset containing m-1 selected features; m is the optimal fault feature;

[0025] As a preferred embodiment, the initial feature space is processed by a pre-trained GBDT model. The specific steps for evaluating the importance of features are as follows:

[0026] The formula for calculating the importance of feature j in a single decision tree is:

[0027]

[0028] In the formula, K is the number of aging state categories, which is the number of leaf nodes in the decision tree; K-1 is the number of non-leaf nodes in the tree; It is the reduction in squared loss after node w splits.

[0029] The global importance of feature j is the average importance of feature j across all trees, and the specific calculation formula is as follows:

[0030]

[0031] In the formula, M represents the number of decision trees.

[0032] On the other hand, the present invention proposes a transformer oil-paper insulation fault diagnosis system, comprising:

[0033] The initial feature space establishment module establishes the initial feature space based on the set of time-domain initial feature quantities that affect oil-paper insulation faults. And calculate the initial feature space. Intra-group redundancy and inter-group correlation;

[0034] The first feature space calculation module uses an incremental search optimization algorithm to calculate the ranking of features with maximum correlation and minimum redundancy under the standard function difference, and removes features within a preset range that are ranked low in terms of feature quality, thereby obtaining the first feature selection threshold m1 and the first feature space.

[0035] The second feature space calculation module calculates the initial feature space using a pre-trained GBDT model. The importance of features is evaluated, and features with low importance are removed to obtain the second feature selection threshold m2 and the second feature space.

[0036] The optimal feature space selection module determines the first feature space based on Kendall's coefficients. Second feature space The rank correlation is used to comprehensively compare the rank correlation between the two feature spaces and the importance of the feature quantities to select the optimal feature space;

[0037] The fault diagnosis module performs fault diagnosis of transformer oil-paper insulation based on the selected feature space.

[0038] As a preferred embodiment, the initial feature space The specific formulas for calculating within-group redundancy and between-group correlation are as follows:

[0039]

[0040]

[0041] In the formula, maxD(s, c), D is the intra-group redundancy, minR(S), R is the inter-group correlation, S is the feature space; |S| is the number of features; c is the column vector of aging categories in the sample set; I(x i c) represents the mutual information between feature quantity i and aging category c in the feature space; I(x) i ,x j Let x be a feature quantity in the feature space. i With characteristic quantity x j Mutual information between them.

[0042] In a preferred embodiment, the feature quantity x in the feature space i With characteristic quantity x j Mutual information between I(x) i ,x j The specific calculation formula is as follows:

[0043]

[0044] In the formula, the characteristic quantity x i With characteristic quantity x j These are the sample column vectors for feature quantity i and feature quantity j in the initial feature set, respectively.

[0045] As a preferred embodiment, the specific steps for calculating the ranking of features with maximum correlation and minimum redundancy under the standard function difference using the incremental search optimization algorithm are as follows:

[0046] Based on the initial feature space The difference between the within-group redundancy and the between-group correlation is obtained by using the standard function of mutual information, as shown in the following formula:

[0047] maxΦ(D,R),Φ=DR

[0048] The optimal fault feature is obtained using an incremental search optimization algorithm. One feature is selected and then removed, and the feature sorting is reset to complete the process. The selection criteria for the optimal fault feature are as follows:

[0049]

[0050] In the formula, X is the initial feature set, and S is the initial feature set. m-1 It is a feature subset containing m-1 selected features; m is the optimal fault feature;

[0051] As a preferred embodiment, the initial feature space is processed by a pre-trained GBDT model. The specific steps for evaluating the importance of features are as follows:

[0052] The formula for calculating the importance of feature j in a single decision tree is:

[0053]

[0054] In the formula, K is the number of aging state categories, which is the number of leaf nodes in the decision tree; K-1 is the number of non-leaf nodes in the tree; It is the reduction in squared loss after node w splits.

[0055] The global importance of feature j is the average importance of feature j across all trees, and the specific calculation formula is as follows:

[0056]

[0057] In the formula, M represents the number of decision trees.

[0058] The present invention has the following beneficial effects:

[0059] 1. This invention ranks the features by their relative merits based on the difference between the standard function and the maximum correlation and minimum redundancy through incremental search.

[0060] 2. This invention performs a secondary screening of the initial feature space by evaluating the importance of the features in the initial feature space and eliminating features with low importance.

[0061] 3. This invention selects the optimal feature space by ranking the merits and importance of feature quantities, and then analyzes insulation faults. Attached Figure Description

[0062] Figure 1 This is a schematic diagram of the process of the present invention;

[0063] Figure 2 A redundancy matrix diagram showing the relationships between various features;

[0064] Figure 3 This is a schematic diagram of the three rounds of screening according to an embodiment of the present invention;

[0065] Figure 4A 7-dimensional spatial correlation matrix diagram for the second round of selection;

[0066] Figure 5 The 6-dimensional spatial high correlation matrix diagram selected for the third round. Detailed Implementation

[0067] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0068] It should be understood that the step numbers used in the text are for ease of description only and are not intended to limit the order in which the steps are performed.

[0069] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0070] The terms “comprising” and “including” indicate the presence of the described feature, whole, step, operation, element and / or component, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or collections thereof.

[0071] The term “and / or” refers to any combination of one or more of the associated listed items, as well as all possible combinations, and includes these combinations.

[0072] Example 1:

[0073] See Figure 1 A method for diagnosing transformer oil-paper insulation faults, comprising the following steps:

[0074] An initial feature space is established based on the set of time-domain initial features affecting oil-paper insulation faults. And calculate the initial feature space. Intra-group redundancy and inter-group correlation;

[0075] An incremental search optimization algorithm is used to calculate the ranking of features with maximum relevance and minimum redundancy under the standard function difference. Features that rank low within a preset range are removed to obtain the first feature selection threshold m1 and the first feature space.

[0076] The initial feature space is processed using a pre-trained GBDT model. The importance of features is evaluated, and features with low importance are removed to obtain the second feature selection threshold m2 and the second feature space.

[0077] Determining the first characteristic space based on Kendall's coefficient. Second feature space The rank correlation is used to comprehensively compare the rank correlation between the two feature spaces and the importance of the feature quantities to select the optimal feature space;

[0078] Transformer oil-paper insulation fault diagnosis based on selected feature space.

[0079] In this embodiment, the time-domain characteristics of the oil-paper insulated transformer are first analyzed:

[0080] The time-domain characteristics of oil-paper insulated transformers are divided into three categories: extended Debye equivalent model circuit characteristics, RVM characteristics, and PDC characteristics.

[0081] Existing research has proposed more than 30 time-domain feature quantities. This embodiment briefly introduces 25 commonly used time-domain feature quantities in the field of insulation diagnosis, and selects the best of these 25 feature quantities as the research object below. Their composition is shown in Table 1, including 8 ED equivalent model feature quantities, 5 RVM feature quantities, and 10 PDC feature quantities, forming the initial feature space Ω. Z .

[0082]

[0083] Table 1. 25 Selectable Time-Domain Dielectric Response Characteristics

[0084] In the extended Debye equivalent circuit model, the geometric capacitance C g Insulation resistance R g Number of geometric branches N, number of polarization small resistance branches l lowRpi and average time constant C g With R g The ratio between them is defined as the aging factor K. L The relaxation time constant is greater than K τ =τ max / τ min It can reflect the relative aging state of the oil and paper. As the aging degree of the insulating oil increases, K... τ The larger it is, the smaller it is.

[0085] By analyzing the characteristics of the recovery voltage, the peak value U of the recovery voltage can be extracted from the single recovery voltage measurement curve. r The principal time constant t dcom Peak duration of action tpeak The initial slope Sr is used as a characteristic quantity, and the half-peak period time t is used as a characteristic quantity that can describe the relaxation process in the second half. 1 / 2 Furthermore, it was demonstrated that the half-peak cycle time decreases as the oil-paper insulation system ages.

[0086] In the study of PDC characteristic quantities, the average depolarization charge... Average relaxation contribution and relaxation energy W c In addition, the extracted feature quantity, polarization intensity P r The specific formula is as follows:

[0087]

[0088] In the formula, S is i d The cross-sectional area of ​​the insulating material flowing through it generally depends on the geometric structure of the oil-paper insulation and hardly changes with insulation aging. Therefore, the polarization intensity and depolarization charge change according to the same pattern. The formulas for calculating the transformer relaxation energy and relaxation loss during polarization are as follows:

[0089]

[0090]

[0091] A comparison of the two equations above shows that during the charging time t c When the value is sufficiently large, there is a linear relationship between relaxation energy and relaxation loss.

[0092] In addition, the initial depolarization charge S is extracted as a feature quantity. oil Terminal depolarization charge S paper Energy spectrum peak w max Absorption ratio K, maximum polarization intercept b, and maximum peak time t max And the maximum polarization slope 'a', the first five characteristic quantities have relatively small values ​​when the insulation is good, while the latter two show the opposite trend.

[0093] In this embodiment, based on the performance and diagnostic capabilities of each feature, an optimal feature space is optimized by establishing appropriate evaluation metrics. Specifically, a two-level time-domain feature optimization scheme is implemented, combining the principle of maximum correlation and minimum redundancy with gradient ascending tree importance assessment. The aim is to find the optimal time-domain dielectric feature space and improve the accuracy of oil-paper insulation diagnosis.

[0094] The maximum correlation minimum redundancy algorithm model is a feature selection method based on mutual information. It selects features according to the principle of maximum statistical dependence, which can effectively eliminate features with excessive redundancy to achieve dimensionality reduction.

[0095] First, we take two distinct sample column vectors x from the initial feature set, feature i and feature j. i and x j Their individual probability densities and joint density are defined as ρ(x) i ), ρ(x) j ), ρ(x) i x j The mutual information I between the two features is given by the following formula:

[0096]

[0097] Feature selection under the maximum relevance and minimum redundancy criterion should possess the following characteristics: the target feature quantity has the maximum relevance to the aging category, and the internal redundancy between the target feature quantity and other feature quantities is minimized, as shown in the following formula:

[0098]

[0099]

[0100] In the formula, S is the feature space; |S| is the number of features; c is the column vector of aging categories in the sample set; I(x i c) represents the mutual information between feature quantity i and aging category c in the feature space; I(x) i ,x j Let x be a feature quantity in the feature space. i With characteristic quantity x j Mutual information between them.

[0101] Combining the calculation formulas for maximum relevance and minimum redundancy, the mutual information standard function difference form in the standard expression of the maximum relevance and minimum redundancy criterion used in this embodiment is as follows:

[0102] maxΦ(D,R),Φ=DR

[0103] Based on the standard function, an incremental search optimization algorithm is used to obtain the optimal fault features. Assume the original feature set is X, and a feature subset S containing m-1 features has been selected. m-1 Then feature selection is to select from the remaining feature set {XS} m-1 In this process, we select the m-th feature that maximizes the maxΦ in the above formula. This feature should satisfy the following equation:

[0104]

[0105] Based on the feature space set, an initial decision tree f0(x) = 0 is constructed. Insulation diagnosis combined with multiple features is a typical multi-classification problem. In this context, the log-likelihood loss function is chosen as the loss function, as shown in the following formula:

[0106]

[0107]

[0108] Where K is the number of aging state categories, p k This represents the probability of classifying the aging state of a sample set as class k under the given decision tree environment.

[0109] Combining the above two formulas, the negative gradient error of the aging category c corresponding to the i-th sample of the t-th tree can be calculated, as follows:

[0110]

[0111] The error here is the difference between the true probability of sample i corresponding to fault category c and the predicted probability from the previous round. Repeating the above process, based on the resulting pseudo-residual combination {(x1, r...}... t1 ), (x2, r t2 ), ..., (x i r ti The (t+1)th decision tree is fitted. The decision tree consists of nodes and directed edges. Each node is divided into leaf nodes and branch nodes.

[0112] In each iteration of the decision tree, the feature column vector is used as the branch node of the tree, and the feature importance is the squared improvement based on a feature being selected as the split node of this tree.

[0113] The importance of feature j in a single tree is defined by the following formula:

[0114]

[0115] Where K is the number of aging state categories, i.e., the number of leaf nodes in the decision tree, then K-1 is the number of non-leaf nodes in the tree. It is the reduction in squared loss after node w splits.

[0116] The global importance of feature j is measured by averaging the importance of feature j across all trees, as shown in the following formula:

[0117]

[0118] Example 2:

[0119] This embodiment collects field measurement data of the time-domain dielectric response method from over 50 transformers and 80 windings. The collected 80 transformer samples were categorized into three main classes based on their oil-paper insulation condition: good, average, and poor, according to the furfural content test specifications in the "Preventive Testing Procedures for Power Equipment" and the transformers' operating years and maintenance status. Twenty-five time-domain characteristic quantities were extracted from all 80+ transformer winding measurement data to form the initial characteristic space Ω. Z Due to the large data space, Table 2 gives the initial feature space Ω. Z Some data is for reference only.

[0120]

[0121]

[0122] Table 2 Initial Dataset Ω Z Partial data

[0123] This embodiment uses low-dimensional, small-sample data. In practical applications, feature optimization typically retains about 1 / 3 to 1 / 4 of the original matrix's feature count. Considering that most existing oil-paper insulation diagnostics studies use 7 to 10 time-domain dielectric features, the optimized feature matrix ultimately retains 6 to 8 features. In the first-level feature selection based on the mRMR principle, 9 features that do not meet the requirements are eliminated, leaving 16 features to form a suboptimal feature space for importance screening. That is, the first-level threshold m1 = 16 is set. At the same time, the second-level GBDT importance evaluation is used to screen features that contribute highly to the discrimination of the target category to establish the final feature optimization matrix.

[0124] The correlation between each characteristic quantity and the aging category was calculated based on the standard formula of mutual information function, as shown in Table 3.

[0125]

[0126]

[0127] Table 3. Correlation between various characteristic quantities and aging categories (I)

[0128] The redundancy between each feature is calculated using the intra-group redundancy formula. Due to the large matrix dimension, only a portion of the matrix is ​​shown here. Figure 2 As shown.

[0129] Incremental search in the form of standard function difference is performed by combining the feature quantity rules of the standard function. Features with high correlation and low redundancy in the initial feature space are screened, and the results are sorted from best to worst according to the numerical values ​​of the feature quantities in the form of standard function difference. The results are shown in Table 4 below.

[0130]

[0131] Table 4. Ranking of temporal features based on mRMR

[0132] It can be seen K T S oil S paper τ max t dcom t max Q, W g Features with low global relevance and high redundancy with other features under the mRMR principle are eliminated based on the threshold m1 selected in the first layer. The nine features with the lowest ranking are then removed, forming the first-level optimized feature space Ω. E1 As shown in Table 5:

[0133]

[0134] Table 5 Feature space Ω after first-level optimization E1

[0135] The characteristic space Ω E1 Substituting these features into the GBDT trainer for importance evaluation, the feature importance primarily reflects the overall utilization rate of each feature under this learner. From Figure 3 The results of the first round of selection show that there are features with extremely low contribution in the feature space, and these features are removed. The test is repeated until the threshold m2 of the second layer selection is reached, which is the feature space consisting of 6 to 8 features.

[0136] Depend on Figure 2 It can be seen that nine features were of low importance in the first round of screening and were all removed. The two feature spaces, Ω1 (7-dimensional) selected in the second round and Ω2 (6-dimensional) selected in the third round, both meet the threshold m2 requirement of the second round, and each feature has a significant contribution. Kendall's coefficient is then used to calculate the rank correlation in the two feature spaces Ω1 and Ω2. Kendall's coefficient reflects the degree of rank correlation in a set of feature spaces, ranging from -1 to 1. A value closer to 1 indicates a high degree of rank correlation between the two random variables, while a value closer to -1 indicates a lower degree of rank correlation. The high correlation matrix of Ω1 is shown below. Figure 4 As shown, the high correlation matrix of Ω2 is as follows Figure 5 As shown.

[0137] from Figure 4 It can be seen that in the characteristic space Ω1, t 1 / 2 With t peak The Kendall coefficient is greater than 0.8 and is related to the characteristic R.g The coefficient between them also reached 0.7117, therefore t is considered to be... 1 / 2 With t peak and R g As a feature of high rank correlation, from Figure 5 As can be seen, the Kendall coefficients of Ω2 are relatively good, and there are no excessively large coefficients among the features. Furthermore, this is the same as the feature space Ω2 selected in the third round of GBDT. Therefore, feature space Ω2 is chosen as the final feature optimization space Ω. E The final results are shown in Table 6:

[0138]

[0139] Table 6. Feature combination of the final feature optimization space Ω3

[0140] To verify the feasibility and effectiveness of the transformer time-domain feature optimization strategy proposed in this embodiment, new achievements in insulation diagnosis using multiple time-domain features in recent years were collected. This was done by comparing several proposed feature spaces with the optimal feature space Ω. Z Comparative tests were conducted. Furthermore, even when using the same feature space, the diagnostic accuracy of different insulation diagnostic methods can vary. Therefore, this embodiment uses different diagnostic methods for comparison.

[0141] This embodiment, based on the aforementioned collected example data, excludes the initial feature space Ω. Z In addition, to verify that the preferred strategy proposed in this embodiment is actually effective, the 7-dimensional feature space Ω eliminated in the second screening of the GBDT importance selection will be used. test1 Only the six-dimensional feature space Ω with the first six priority elements under the mRMR principle is used. test2 The feature space Ω is selected by the grey relational analysis method. test3 Comparative verification was performed, as well as the feature space Ω. test4 The feature quantities contained in each feature space are shown in Table 7:

[0142]

[0143]

[0144] Table 7. Feature space contents of the control group

[0145] Table 7 shows the characteristic space Ω of the control group. test1～test3 The included time-domain features are the same as those in the preferred feature space Ω. EIt also includes three main categories of feature quantities: equivalent circuit, RMV, and PDC, ensuring the diversity of feature quantities and guaranteeing that the feature space can reflect the occurrence process of the time-domain dielectric response to the greatest extent. At the same time, the matrix dimension also meets the selection range of the threshold m2. However, according to the process of the preferred strategy in this embodiment section, it can be seen that:

[0146] (1)Ω test1～test3 Both contain features with high linear correlation, and Ω test3 The features Q and τ have excessive redundancy and low global correlation. max t dcom .

[0147] (2)Ω test4 Containing only RVM and equivalent circuit features, it is at a disadvantage in terms of feature richness compared to other feature spaces.

[0148] This embodiment intends to use clustering algorithms with significant differences in principle: Support Vector Machine, Random Forest, Gradient Ascending Tree, and K-Nearest Neighbors method for cross-validation to verify the feasibility of the optimal selection.

[0149] Support Vector Machines (SVMs) perform well in solving classification problems with small sample sizes, nonlinearity, and high dimensions. They solve the classifier construction problem by mapping space using kernel functions and constructing the optimal classification hyperplane. Different kernel functions result in different SVMs. This embodiment uses the radial basis function kernel for prediction and incorporates particle swarm optimization to optimize the kernel coefficients and penalty function. K-Nearest Neighbors (KNN) is a basic classification method that calculates the distance between features and uses majority voting based on the K value. Random Forests (RF), as parallel-trained classifiers, feature random sampling and feature selection, ensuring diversity and strong generalization ability. Gradient Ascending Trees (FAB) is a typical serial algorithm that continuously fits residuals, starting with a weak learner and adjusting the data based on the previous classification results. The hyperparameter sets for each algorithm are shown in Table 8.

[0150]

[0151] Table 8 Hyperparameters of each algorithm

[0152] During the comparative validation process, two-fold cross-validation was performed on different feature spaces. To minimize randomness during the experiment, the test and validation sets were randomly sampled multiple times while ensuring that the number of samples from the three aging states was approximately the same, and the experiment was repeated. The experimental results are shown in Table 9.

[0153]

[0154] Table 9 Comparison of Classification / Clustering Performance Results for Each Feature Space

[0155] As shown in Table 9, the optimal feature space Ω in this embodiment is [missing information]. E The diagnostic results were all over 90% accurate, with some even reaching 97%, indicating that the use of Ω... E It can perform effective and accurate insulation diagnosis, and it has good applicability among various diagnostic algorithms. Furthermore, by comparing the accuracy rate with that of the control group, the following conclusions can be drawn:

[0156] Comparison of the optimal feature space and the feature space screened out during the optimization process in this embodiment (Ω) E Ω test1 Ω test2 ):

[0157] Regarding the diagnostic accuracy of various algorithms, Ω E For Ω test1 Ω test2 It forms a complete enclosure, but the feature space Ω, which has been filtered by importance, can be seen. test1 Compared to the feature space Ω that only considers redundancy and correlation test2 It has significant advantages, as can be inferred from the diagnostic results: feature space filtered by importance can significantly improve the accuracy of the final aging state determination. According to... Figure 2 The magnitude of the values ​​of each feature quantity can indicate that the importance of the half-peak time period is relatively lower than that of other features quantities. Figure 2 Each feature has an importance exceeding 5%. Based on the conclusions drawn from the Kendall coefficient discriminant matrix, the optimal feature space Ω... E Compared to Ω test1 Each feature has higher importance and lower rank correlation, which also demonstrates the superiority of this optimal scheme from the experimental results.

[0158] Comparison of the optimal feature space with other feature spaces in this embodiment (Ω) E Ω test3 Ω test4 ):

[0159] First Ω E Compared to the feature space Ω test3 Ω test4 The accuracy in identifying aging conditions has been significantly improved because the selection of the feature space in traditional insulation assessment methods lacks a reasonable basis. test3 Ω test4 The features contained in the algorithm are too redundant and have low contribution to the algorithm, resulting in poor generalization ability of the subsequent diagnostic algorithm and unsatisfactory diagnostic effect.

[0160] A comparison of the feature spaces filtered out during the optimization process in this embodiment with other feature spaces (Ω) test1 Ω test2 Ω test3 Ω test4 ):

[0161] It can be seen that the feature space Ω after evaluation based on redundancy, relevance, and importance... test1 Ω test2 Compared to the feature space Ω without selection criteria test3 Ω test4 The performance of each algorithm has improved relatively, proving that if there are too many cases in the feature space that have low correlation with aging categories, strong linear relationships between features, or incomplete feature types, the error rate of transformer aging judgment will increase.

[0162] The above comparative analysis proves that the optimal time-domain dielectric characteristic space Ω proposed in this embodiment is valid. E Compared to traditional insulation diagnostic methods, the feature space can carry more effective information, significantly improving the accuracy of insulation diagnosis.

[0163] Furthermore, based on the diagnostic results of each algorithm shown in Table 9, the average accuracy of each feature space and the Ω were calculated. Z The five evaluation indicators of accuracy—mean relative error, accuracy variation width, minimum error, and maximum error—were compared globally, and the calculation results are shown in Table 10.

[0164]

[0165] Table 10 Results of evaluation indicators for each feature space

[0166] As shown in Table 10, the optimal feature space Ω based on the preferred feature space theory proposed in this embodiment is... E The corresponding average accuracy was higher than that of the control group across all four feature spaces. When compared to the initial space, the average relative error of the accuracy was 7.575% higher, although Ω... test4 The optimal feature space Ω with varying width interval and variance ratio E However, it is smaller, but its average accuracy is nearly 12 points lower, and the variation width and variance of the optimal feature space are significantly smaller than those of the other remaining feature spaces. In summary, Ω E The diagnostic accuracy is greatly improved for different diagnostic methods, and the diagnostic results have the characteristics of small fluctuations, higher stability, and stronger generalization ability.

[0167] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for diagnosing faults in transformer oil-paper insulation, characterized in that, The specific steps include: An initial feature space is established based on the set of time-domain initial features affecting oil-paper insulation faults. And calculate the initial feature space. Intra-group redundancy and inter-group correlation; The initial feature space The specific formulas for calculating within-group redundancy and between-group correlation are as follows: In the formula, For intra-group redundancy, For inter-group correlation, For feature space; | | represents the number of features; c represents the column vector of aging categories in the sample set; I(x) i c) represents the mutual information between feature i in the feature space and the column vector c of the aging category in the sample set; I(x i ,x j Let x be a feature quantity in the feature space. i With characteristic quantity x j Mutual information between them; The feature quantity x in the feature space i With characteristic quantity x j Mutual information between I(x) i ,x j The specific calculation formula is as follows: In the formula, the characteristic quantity x i With characteristic quantity x j These are the sample column vectors representing feature i and feature j in the initial feature set, respectively. An incremental search optimization algorithm is used to calculate the ranking of features with maximum relevance and minimum redundancy under the standard function difference. Features that rank low within a preset range are removed to obtain the first feature selection threshold m1 and the first feature space. ; The specific steps for using the incremental search optimization algorithm to calculate the ranking of features with maximum correlation and minimum redundancy under the standard function difference are as follows: Based on the initial feature space The difference between the within-group redundancy and the between-group correlation is obtained by using the standard function of mutual information, as shown in the following formula: In the formula, For intra-group redundancy, Inter-group correlation; The optimal fault feature is obtained using an incremental search optimization algorithm. One feature is selected and then removed, and the feature sorting is reset to complete the process. The selection criteria for the optimal fault feature are as follows: In the formula, X is the initial feature set, and S is the initial feature set. m-1 It is a feature subset containing m-1 selected features; m is the optimal fault feature; The initial feature space is processed using a pre-trained GBDT model. The importance of features is evaluated, and features with low importance are removed to obtain the second feature selection threshold m2 and the second feature space. ; The initial feature space is obtained through a pre-trained GBDT model. The specific steps for evaluating the importance of features are as follows: The formula for calculating the importance of feature j in a single decision tree is: In the formula, K This represents the number of aging state categories, which is the number of leaf nodes in the decision tree. K- 1 represents the number of non-leaf nodes in the tree; It is a node w The reduction in squared loss after splitting; The global importance of feature j is the average importance of feature j across all trees, and the specific calculation formula is as follows: In the formula, M represents the number of decision trees; Determining the first characteristic space based on Kendall's coefficient. Second feature space The rank correlation is used to comprehensively compare the rank correlation between the two feature spaces and the importance of the feature quantities to select the optimal feature space; Transformer oil-paper insulation fault diagnosis based on selected feature space.

2. A transformer oil-paper insulation fault diagnosis system, characterized in that, include: The initial feature space establishment module establishes the initial feature space based on the set of time-domain initial feature quantities that affect oil-paper insulation faults. And calculate the initial feature space. Intra-group redundancy and inter-group correlation; The initial feature space The specific formulas for calculating within-group redundancy and between-group correlation are as follows: In the formula, For intra-group redundancy, For inter-group correlation, For feature space; | | represents the number of features; c represents the column vector of aging categories in the sample set; I(x) i c) represents the mutual information between feature i in the feature space and the column vector c of the aging category in the sample set; I(x i ,x j Let x be a feature quantity in the feature space. i With characteristic quantity x j Mutual information between them; The feature quantity x in the feature space i With characteristic quantity x j Mutual information between I(x) i ,x j The specific calculation formula is as follows: In the formula, the characteristic quantity x i With characteristic quantity x j These are the sample column vectors representing feature i and feature j in the initial feature set, respectively. The first feature space calculation module uses an incremental search optimization algorithm to calculate the ranking of features with maximum correlation and minimum redundancy under the standard function difference, and removes features within a preset range that are ranked low in terms of feature quality, thereby obtaining the first feature selection threshold m1 and the first feature space. ; The specific steps for using the incremental search optimization algorithm to calculate the ranking of features with maximum correlation and minimum redundancy under the standard function difference are as follows: Based on the initial feature space The difference between the within-group redundancy and the between-group correlation is obtained by using the standard function of mutual information, as shown in the following formula: In the formula, For intra-group redundancy, Inter-group correlation; The optimal fault feature is obtained using an incremental search optimization algorithm. One feature is selected and then removed, and the feature sorting is reset to complete the process. The selection criteria for the optimal fault feature are as follows: In the formula, X is the initial feature set, and S is the initial feature set. m-1 A subset of features containing the selected m-1 features; m represents the optimal fault characteristic; The second feature space calculation module calculates the initial feature space using a pre-trained GBDT model. The importance of features is evaluated, and features with low importance are removed to obtain the second feature selection threshold m2 and the second feature space. ; The initial feature space is obtained through a pre-trained GBDT model. The specific steps for evaluating the importance of features are as follows: The formula for calculating the importance of feature j in a single decision tree is: In the formula, K This represents the number of aging state categories, which is the number of leaf nodes in the decision tree. K- 1 represents the number of non-leaf nodes in the tree; It is a node w The reduction in squared loss after splitting; The global importance of feature j is the average importance of feature j across all trees, and the specific calculation formula is as follows: In the formula, M represents the number of decision trees; The optimal feature space selection module determines the first feature space based on Kendall's coefficients. Second feature space The rank correlation is used to comprehensively compare the rank correlation between the two feature spaces and the importance of the feature quantities to select the optimal feature space; The fault diagnosis module performs fault diagnosis of transformer oil-paper insulation based on the selected feature space.

Images