Transformer fault diagnosis method for optimizing support vector machine based on improved dung beetle algorithm

Abstract

The invention belongs to the field of transformer fault diagnosis, and relates to an improved dung beetle algorithm optimized support vector machine-based transformer fault diagnosis method, which comprises the following steps of: searching DGA samples of five fault types of high-temperature overheating, medium-temperature overheating, high-energy discharge, low-energy discharge and partial discharge in a fault case; carrying out normalization processing on the collected DGA samples, and enabling features of different dimensions to have comparability; performing sample division on the normalized data according to the ratio of a training set to a test set being 8: 2; a DBO algorithm is utilized to obtain an optimal kernel parameter g, a penalty factor c and an optimal fitness value, and global search and local development are both considered; and substituting the test set into the HDBO-SVM fault classification model under the optimal kernel parameter g and penalty factor c for testing to obtain a fault diagnosis result under the data set. The method promotes the global search in the initial stage of iteration, accelerates the convergence speed, improves the diagnosis precision, and accelerates the diagnosis speed.

Description

Technical Field

[0001] The invention belongs to the technical field of transformer fault diagnosis, and in particular relates to a transformer fault diagnosis method based on an improved dung beetle algorithm to optimize a support vector machine. Background Art

[0002] As a key equipment in the power transmission and distribution link of the power grid, power transformers are responsible for the important tasks of changing voltage levels and converting electrical energy. Once a fault occurs, it will affect the safe and stable operation of the entire power system. Therefore, it is very important to promptly and accurately determine the location and type of fault and take corresponding intervention measures to reduce unnecessary losses caused by transformer failures.

[0003] Defects and shortcomings of existing transformer fault diagnosis technology:

[0004] Dissolved Gases Analysis (DGA) is recognized by the industry as an effective method for transformer fault analysis. It determines transformer faults based on the different contents and rates of the main gases such as hydrogen (H2), methane (CH4), ethane (C2H6), ethylene (C2H4), and acetylene (C2H2) produced under different faults.

[0005] The transformer fault diagnosis technology based on DGA has gone through the traditional stage and the intelligent diagnosis stage. The characteristic gas method in the traditional stage is simple and easy to implement, but its accuracy is difficult to meet the needs of practical applications. The gas content ratio rule is realized by mapping different ratios of dissolved gas in oil according to different fault types. Compared with the characteristic gas method, the accuracy of the ratio method diagnosis has been improved to a certain extent, but there are still certain problems in practical applications, which makes it difficult to meet the requirements for the accuracy of diagnosis. Therefore, the traditional diagnosis based on DGA technology has great limitations. Neural networks are the most widely used in the practical application of transformer fault diagnosis, but neural networks have high requirements for the number of samples. Small samples will cause insufficient model training and thus cause large errors, which conflicts with the characteristics of transformer fault samples being small samples. Compared with other algorithms, support vector machines (SVM) have better processing of local extreme values ​​and stronger generalization ability. Therefore, the use of support vector machines for transformer fault diagnosis can not only solve the problem of small sample characteristics in transformer fault data, but also more effectively deal with the ambiguity and randomness of fault causes and mechanisms on the basis of traditional diagnosis.

[0006] The values ​​of kernel parameter g and penalty factor c greatly affect the classification performance of support vector machine. Metaheuristic algorithm, as a combination of random algorithm and local search algorithm, is not limited to the unique conditions of a certain problem. It has the advantages of easy implementation, strong self-learning ability and simple framework. Compared with grid search and heuristic strategy, it has relative superiority and wide application in parameter optimization. However, it also has different application defects, such as particle swarm optimization is prone to premature convergence, gray wolf algorithm converges slowly in the later stage, and they are prone to fall into local optimal solution. Compared with other algorithms, dung beetle optimizer (DBO) takes into account global search and local development, and has the characteristics of fast convergence speed and better solution accuracy. However, dung beetle optimizer also has certain shortcomings in population initial distribution and later convergence optimization. Summary of the invention

[0007] The purpose of the present invention is to provide a transformer fault diagnosis method based on an improved dung beetle algorithm to optimize a support vector machine, so as to solve the problem of insufficient accuracy and reliability caused by small sample characteristics when processing dissolved gas fault data in transformer oil in the prior art, and to introduce Bernoulli mapping to optimize the uniformity of the population initialization stage, and to introduce Levy flight strategy and dynamic weight coefficient in the optimization stage to improve the position update of the dung beetle algorithm, thereby improving the global and local search capabilities and accelerating the convergence speed.

[0008] The technical solution adopted by the present invention is a transformer fault diagnosis method based on an improved dung beetle algorithm to optimize a support vector machine, comprising the following steps:

[0009] S1: Search for DGA samples of five fault categories: high temperature overheating, medium temperature overheating, high energy discharge, low energy discharge and partial discharge in the fault cases provided by State Grid Corporation and transformer manufacturers.

[0010] S2: To ensure the accuracy and reliability of the data, the collected DGA samples are normalized to make the features of different dimensions comparable.

[0011] S3: The normalized data is divided into samples according to the ratio of 8:2 between the training set and the test set.

[0012] S4: Use the DBO algorithm to obtain the optimal kernel parameter g, penalty factor c, and optimal fitness value, taking into account both global search and local development.

[0013] S5: The test set is brought into the HDBO-SVM fault classification model with the best kernel parameter g and penalty factor c for testing, and the fault diagnosis result under the data set is obtained.

[0014] Furthermore, in S1, DGA samples of five fault categories, namely, high temperature overheating, medium temperature overheating, high energy discharge, low energy discharge and partial discharge, are searched. The five fault categories cover the main types of transformer faults and have clear DGA features, which facilitate accurate fault diagnosis. The specific method of step S1 is as follows:

[0015] The collected DGA samples selected the unit contents of five gases, H2, CH4, C2H6, C2H4 and C2H2, as well as the ratios between multiple gases, including CH4 / H2, C2H4 / C2H2, C2H4 / C2H6, C2H2 / (C x H x )、H2 / (H2+C x H x ) and C2H4 / (C x H x ), as the feature quantity.

[0016] Furthermore, the specific method of step S2 is as follows:

[0017] Using the max-min normalization method, all eigenvalues ​​are scaled to the interval [0,1], and the formula is as follows:

[0018]

[0019] In the formula, x i is the original data, x min is the minimum value in the data set, x max is the maximum value in the data set, and x is the normalized data.

[0020] Furthermore, in step S3, the sample division method adopts a five-fold cross-validation method to ensure that the model learns enough information on the training set, and at the same time verifies the generalization ability and diagnostic accuracy of the model through the test set.

[0021] Furthermore, the specific method of step S4 includes:

[0022] S4.1: In the population initialization stage, Bernoulli mapping is introduced to optimize the population distribution to improve the global search ability of the algorithm and prevent the algorithm from converging to the local optimal solution too early. Its mathematical expression is:

[0023]

[0024] Z K is the current value of the Kth generation chaotic sequence, ρ is the control parameter and its value is around 0.5 to obtain good ergodicity. After traversing the entire two-dimensional chaotic sequence, the chaotic particles are mapped to the modeling search range:

[0025]

[0026] Where X min and X max is the maximum value of the horizontal axis, is the position of the Kth generation of dung beetles.

[0027] S4.2: To prevent falling into the local optimum at the early stage of iteration, a dynamic weight coefficient is added to the original position update formula of the thief dung beetle to promote global search at the early stage of iteration. It becomes smaller adaptively in the later stage of iteration to promote local search and accelerate convergence. The weight factor and the position update formula of the thief dung beetle are expressed as:

[0028]

[0029] Where ω is the dynamic weight coefficient and t is the current iteration coefficient.

[0030] x i (t+1)=X b +Sg(|x i (t)-X * |+|x i (t)-X b |)+ω(|x i (t)-X * |+|x i (t)-X b |)·rand (5)

[0031] Where x i (t) is the position of the ith thief at the tth iteration, X b is the optimal position of the thief dung beetle in the current iteration, X * is the global optimal position, that is, the optimal solution found in the entire population so far, and rand is a random number between 0 and 1, which is used to increase the randomness of position updates, help the algorithm jump out of the local optimal solution, and enhance the global search capability.

[0032] S4.3: Introduce the Levy flight strategy to generate random step sizes and improve the position update mechanism of the dung beetle algorithm. The Levy flight strategy is a unique random walk strategy that obeys the Levy distribution and is represented by a power law distribution:

[0033] L(S)~|S| -1-β (0<β<2) (6)

[0034] Where S is the step length, and L(S) is the probability of moving step length S. Due to the complexity of Levy distribution, the Mantegna algorithm is used to simulate it:

[0035]

[0036] Where θ and ω follow a normal distribution:

[0037]

[0038] Change the position update formula of the thief dung beetle in equation (6) to:

[0039] x i (t+1)=X b +Sg(|x i (t)-X * |+|x i (t)-X b |)+L(S)ξ (9)

[0040] Where ξ is the step size parameter.

[0041] S4.4: In each iteration, the new position of the thief dung beetle is calculated according to the above update formula, and its fitness is calculated. The current fitness is compared with the best fitness of the current Kth generation, and the optimal value is selected for position update.

[0042] S4.5: After the local search is completed, the global optimal kernel parameter g and penalty factor c and the best fitness value are output.

[0043] Furthermore, the SVM model in step S5 is as follows:

[0044] When the support vector machine solves general classification, its optimal hyperplane is expressed as the maximum margin classification problem, which is converted into a quadratic programming problem based on the maximization of the classification interval:

[0045]

[0046] Where w represents the normal vector of the optimal hyperplane, ||w|| 2 represents the complexity of the hyperplane, x i is the feature vector of the i-th sample point, y i is the category label of the i-th sample point, which takes values ​​of +1 or -1, indicating two different categories, and b is the bias term, which allows the optimal hyperplane not to pass through the origin.

[0047] Since there are noise points in the data, when there are a small number of samples that cannot be separated by linear methods, a slack variable ξ is introduced i , its purpose is to facilitate the solution in a larger feasible domain, and the constraints are transformed into:

[0048] y i (wx i +b)≥1-ξ i ,ξ i ≥0,i=1,2,...,n (11)

[0049] After relaxation, the objective function is transformed into the following optimization problem:

[0050]

[0051] The kernel function K(xi,xj) is introduced to perform nonlinear transformation to solve the nonlinearity in the actual problem of transformer fault, and its dual optimization representation is obtained:

[0052]

[0053] a i is the Lagrange multiplier, and the corresponding optimal classification function is:

[0054]

[0055] The function of the kernel function is to realize the mapping of data from low-dimensional space to high-dimensional space, making the construction of classification decision surface simpler. Different kernel functions have different reflection methods. Common kernel functions include linear kernel function, polynomial kernel function, radial basis kernel function, and Sigmoid kernel function. At the same time, the selection of different kernel functions and kernel parameters has a great influence on the classification results of vector machines. Among them, the RBF kernel function has superior generalization ability and requires fewer parameters, so the radial basis kernel function is selected, which can be expressed as:

[0056]

[0057] In the formula That is, it represents the kernel parameter g. The larger the g, the smaller the influence range of a single training sample and the more complex the decision boundary. The smaller the g, the larger the influence range of a single training sample and the smoother the decision boundary.

[0058] Beneficial effects of the present invention:

[0059] (1) The present invention optimizes the population distribution by introducing Bernoulli mapping in the population initialization stage of the dung beetle algorithm, so that the generated initial population fully covers the entire space, solves the problem of premature convergence caused by the random distribution method used in the population initialization of the dung beetle algorithm, and promotes global optimization.

[0060] (2) A dynamic weight coefficient is added to the position update formula of the dung beetle algorithm to promote global search in the early stage of iteration and accelerate the convergence speed. The Levy flight strategy is introduced to promote the diversity of individual populations and take into account both local search and global optimization capabilities to avoid falling into local optimality too early.

[0061] (3) The transformer fault diagnosis model based on the improved dung beetle algorithm and support vector machine is used for fault diagnosis to improve the diagnosis accuracy and speed up the diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0063] Figure 1 It is a schematic diagram of the HDBO-SVM diagnostic process of the method of the present invention;

[0064] Figure 2 A comparison chart of the fitness evolution curves of the HDBO algorithm of the present invention and other commonly used algorithms;

[0065] Figure 3 This is a comparison chart of the diagnosis results between the HDBO-SVM method of the present invention and different SVM optimization methods. DETAILED DESCRIPTION

[0066] The technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments.

[0067] like Figure 1 As shown, a transformer fault diagnosis method based on the improved dung beetle algorithm to optimize the support vector machine is implemented in the following steps:

[0068] S1: Search for DGA samples of five fault categories: high temperature overheating, medium temperature overheating, high energy discharge, low energy discharge and partial discharge. Some sample data and fault types are shown in Table 1, and the feature quantities are determined. The specific features include:

[0069] The collected DGA samples selected the unit contents of five main gases, H2, CH4, C2H6, C2H4, and C2H2, as well as the ratios between multiple gases, including CH4 / H2, C2H4 / C2H2, C2H4 / C2H6, C2H2 / (C x H x )、H2 / (H2+C x H x )、C2H4 / (C x H x ), as the feature quantity.

[0070] Table 1 Partial sample data and fault types

[0071]

[0072]

[0073] S2: To ensure the accuracy and reliability of the data, the collected DGA samples are normalized to make the features of different dimensions comparable. The specific method is as follows:

[0074] Using the max-min normalization method, all eigenvalues ​​are scaled to the interval [0,1], and the formula is as follows:

[0075]

[0076] In the formula, x i is the original data, x min is the minimum value in the data set, x max is the maximum value in the data set, and x is the normalized data.

[0077] S3: The data is divided into samples according to the ratio of 8:2 between the training set and the test set. The 180 groups of transformer fault sample instances collected on site include 30 cases without fault, 30 cases of high temperature overheating, 30 cases of medium temperature overheating, 30 cases of high energy discharge, 30 cases of low energy discharge, and 30 cases of partial discharge. The sample division method adopts the five-fold cross-validation method widely used in machine learning to evaluate model performance, to ensure that the model learns enough information on the training set, and at the same time verify the generalization ability and diagnostic accuracy of the model through the test set.

[0078] S4: Use the DBO algorithm to obtain the best kernel parameter g, penalty factor c, and optimal fitness value, taking into account both global search and local development. The specific steps include:

[0079] S4.1: In the population initialization stage, Bernoulli mapping is introduced to optimize the population distribution to improve the global search ability of the algorithm and prevent the algorithm from converging to a local optimal solution too early. Its mathematical expression can be expressed as:

[0080]

[0081] Z K is the Kth generation chaotic sequence, Z0 is a random number between (0,1], ρ is the control parameter and its value is around 0.5 to obtain good ergodicity. After traversing the entire two-dimensional chaotic sequence, the chaotic particles are mapped to the modeling search range:

[0082]

[0083] Where X min and X max is the maximum value of the horizontal axis, is the position of the Kth generation of dung beetles.

[0084] S4.2: Rolling ball: The position update of the dung beetle during the rolling process is expressed as:

[0085] x i (t+1)=x i (t)+akx i (t+1)+bΔx

[0086] Δx=|x i (t)-X w |

[0087] Where t represents the current iteration number, Δx is used to simulate the change of light intensity, and k is the deflection coefficient and b is the natural coefficient, which are constants between 0 and 1.

[0088] S4.3: Oviposition: A boundary selection strategy is used to model the oviposition area of ​​female dung beetles, defined as:

[0089] Lb * =max(X * (1-R),Lb)

[0090] Ub * =max(X * (1+R),Ub)

[0091] Ub * and Lb * Respectively represent the upper and lower boundaries of the spawning area, X * Represents the current local optimal position. Ub and Lb represent the upper and lower bounds of the optimization problem, respectively. The spawning position update is expressed as:

[0092] x i (t+1)=x i (t)+random(Ub * ,Lb * )

[0093] In the formula, random(Ub * ,Lb * ) means randomly choosing a value between the two.

[0094] S4.4: Foraging: The foraging process of dung beetles in nature, the boundaries of their optimal foraging area are as follows:

[0095] Lb b =max(X b (1-R),Lb)

[0096] Ub b =max(X b (1+R),Ub)

[0097] X b is the global optimal position, Lb b ,Ub b are the lower and upper bounds of the optimal foraging area, respectively. The position of the foraging dung beetle is updated as follows:

[0098] x i (t+1)=x i(t)+C1(x i (t)-Lb b )+C2(x i (t)-Ub b )

[0099] S4.5: Stealing: The position of the thief is constantly updated during the iteration process, which can be described as:

[0100] x i (t+1)=X b +Sη(|x i (t)-X * |+|x i (t)-X b |)

[0101] where x i (t) represents the location information of the i-th thief at the t-th iteration, η is a random vector of size 1×D that obeys the normal distribution, and S represents a constant value.

[0102] S4.5.1: To prevent falling into the local optimum at the beginning of the iteration, a dynamic weight coefficient is added to the original position update formula of the thief dung beetle to promote the global search at the beginning of the iteration. It becomes smaller adaptively in the later stage of the iteration to promote local search and accelerate the convergence speed. The weight factor and the position update formula of the thief dung beetle are expressed as follows:

[0103]

[0104] Where ω is the dynamic weight coefficient and t is the current iteration coefficient. max Indicates the maximum number of iterations, which is 100 here.

[0105] x i (t+1)=X b +Sg(|x i (t)-X * |+|x i (t)-X b |)+ω(|x i (t)-X * |+|x i (t)-X b |)·rand

[0106] Where x i (t) is the position of the ith thief at the tth iteration, X b is the optimal position of the thief dung beetle in the current iteration, X *is the global optimal position, that is, the optimal solution found in the entire population so far, and rand is a random number between 0 and 1, which is used to increase the randomness of position updates, help the algorithm jump out of the local optimal solution, and enhance the global search capability.

[0107] S4.5.2: Introduce the Levy flight strategy to generate random step sizes and improve the position update mechanism of the dung beetle algorithm. The Levy flight strategy is a unique random walk strategy that obeys the Levy distribution and is represented by a power law distribution:

[0108] L(S)~|S| -1-β (0<β<2)

[0109] Where S is the step length, and L(S) is the probability of moving step length S. Due to the complexity of Levy distribution, the Mantegna algorithm is used to simulate it:

[0110]

[0111] Where θ and ω follow a normal distribution:

[0112]

[0113] Change the position update formula of the thief dung beetle in equation (6) to:

[0114] x i (t+1)=X b +Sg(|x i (t)-X * |+|x i (t)-X b |)+L(S)ξ

[0115] Where ξ is the step size parameter.

[0116] S4.6: In each iteration, the new position of the thief dung beetle is calculated according to the above update formula, and its fitness is calculated. The current fitness is compared with the best fitness of the current Kth generation, and the optimal value is selected for position update.

[0117] S4.7: After the local search is completed, the global optimal kernel parameter g and penalty factor c and the best fitness value are output.

[0118] S5: The test set is brought into the HDBO-SVM fault classification model under the optimal parameters c and g for testing, and the fault diagnosis results under the data set are obtained.

[0119] Experiment 1: Convergence and convergence speed comparison of HDBO algorithm, DBO algorithm, sparrow search algorithm and grey wolf algorithm for four classic functions: Sphere Function, Schwefel's Problem 2.22, Schwefel's Problem 2.21 and QuarticFunction ieNoise.

[0120] according to Figure 2 As shown in the figure, the improved dung beetle algorithm is far superior to the other three algorithms in terms of convergence speed and accuracy. HDBO can converge to the theoretical optimal solution for the three test functions 1, 2, and 4. This shows that HDBO has better convergence performance and higher accuracy for both single-peak and multi-peak functions, further proving the reliability and engineering value of HDBO.

[0121] Experiment 2: Simulation test of fault diagnosis model under different algorithm optimization.

[0122] according to Figure 3 As shown in the figure, the fault diagnosis accuracy of the HDBO-SVM model is about 97.2%, the fault diagnosis accuracy of the GWO-SVM model is about 88.9%, the fault diagnosis accuracy of the SSA-SVM model is about 91.7%, and the fault diagnosis accuracy of the DBO-SVM model is about 94.44%. The fault diagnosis accuracy of the support vector machine model optimized by the improved dung beetle algorithm is significantly higher than that of other algorithm diagnosis models. The dung beetle algorithm has better global and local search capabilities than the gray wolf algorithm and the sparrow search algorithm. The setting of the boundary range indicates the size of the search space, so the parameter boundary range in the dung beetle algorithm is set wider. At the same time, the performance of the improved dung beetle optimization algorithm in solving the optimization of model parameters is significantly better than that of the gray wolf optimization algorithm and the sparrow optimization algorithm. This method not only improves the diagnosis accuracy of transformer fault samples, but also completes the diagnosis faster than the other three algorithms.

[0123] As shown in Table 2, in terms of training speed, the iteration time of the Grey Wolf Algorithm is relatively long, about 5.339 seconds, the iteration time of the Sparrow Algorithm and the Dung Beetle Algorithm is relatively short, about 4.947 seconds and 4.437 seconds respectively, and the improved Dung Beetle Algorithm has the shortest iteration time of 3.722 seconds. Its application in engineering can effectively improve the efficiency of transformer fault diagnosis.

[0124] Table 2 Diagnostic indicators of algorithm optimization model

[0125]

[0126] In the embodiments of the present invention, technical features that are not described in detail are all existing technologies or conventional technical means and will not be described in detail here.

[0127] Finally, it should be noted that the above embodiments are only specific implementation methods of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them, and the protection scope of the present invention is not limited thereto. Those skilled in the art should understand that any technician familiar with the technical field can modify the technical solutions recorded in the above embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be covered within the protection scope of the present invention.

Claims

1. A transformer fault diagnosis method based on an improved dung beetle algorithm to optimize support vector machine, characterized in that: The following steps are involved: S1: Search for DGA samples of five fault categories, including high temperature overheating, medium temperature overheating, high energy discharge, low energy discharge and partial discharge, in the fault cases; S2: Normalize the collected DGA samples to make the features of different dimensions comparable; S3: Divide the normalized data into samples according to the ratio of 8:2 between the training set and the test set; S4: Use the DBO algorithm to obtain the optimal kernel parameter g, penalty factor c, and optimal fitness value, taking into account both global search and local development; S5: The test set is brought into the HDBO-SVM fault classification model with the best kernel parameter g and penalty factor c for testing, and the fault diagnosis result under the data set is obtained.

2. The transformer fault diagnosis method based on the improved dung beetle algorithm optimized support vector machine according to claim 1 is characterized in that: The specific method of step S1 is as follows: The collected DGA samples selected the unit contents of five gases, H2, CH4, C2H6, C2H4 and C2H2, as well as the ratios between multiple gases, including CH4 / H2, C2H4 / C2H2, C2H4 / C2H6, C2H2 / (C x H x )、H2 / (H2+C x H x ) and C2H4 / (C x H x ), as the feature quantity.

3. The transformer fault diagnosis method based on the improved dung beetle algorithm optimized support vector machine according to claim 1 is characterized in that: The specific method of step S2 is as follows: Using the max-min normalization method, all eigenvalues ​​are scaled to the interval [0,1], and the formula is as follows: In the formula, x i is the original data, x min is the minimum value in the data set, x max is the maximum value in the data set, is the normalized data.

4. The transformer fault diagnosis method based on the improved dung beetle algorithm optimized support vector machine according to claim 1 is characterized in that: In step S3, the sample division method adopts the five-fold cross-validation method to ensure that the model learns enough information on the training set, and at the same time verifies the generalization ability and diagnostic accuracy of the model through the test set.

5. The transformer fault diagnosis method based on the improved dung beetle algorithm optimized support vector machine according to claim 1 is characterized in that: The specific method of step S4 includes: S4.1: In the population initialization stage, Bernoulli mapping is introduced to optimize the population distribution. The mathematical expression is: Z K is the current value of the K-th generation chaotic sequence, ρ is the control parameter and its value is around 0.5 to obtain good ergodicity. After traversing the entire two-dimensional chaotic sequence, the chaotic particles are mapped to the modeling search range: Where X min and X max is the maximum value of the horizontal axis, is the position of the Kth generation dung beetle; S4.2: A dynamic weight coefficient is added to the original position update formula of the thief dung beetle to promote global search in the early stage of iteration. It is adaptively reduced in the later stage of iteration to promote local search and accelerate convergence. The weight factor and the position update formula of the thief dung beetle are expressed as: Where ω is the dynamic weight coefficient, t is the current iteration coefficient; x i (t+1)=X b +Sg(|x i (t)-X * |+|x i (t)-X b |)+ω(|x i (t)-X * |+|x i (t)-X b |)·rand (5) Where x i (t) is the position of the ith thief at the tth iteration, X b is the optimal position of the thief dung beetle in the current iteration, X * is the global optimal position, that is, the optimal solution found in the entire population so far, and rand is a random number between 0 and 1; S4.3: Introduce the Levy flight strategy to generate random step sizes and improve the position update mechanism of the dung beetle algorithm, which is represented by a power law distribution: L(S)~|S| -1-β (0<β<2) (6) Where S is the step length, L(S) is the probability of moving step length S, and the Mantegna algorithm is used to simulate it: Where θ and ω follow a normal distribution: Change the position update formula of the thief dung beetle in equation (6) to: x i (t+1)=X b +Sg(|x i (t)-X * |+|x i (t)-X b |)+L(S)ξ (9) Where ξ is the step size parameter; S4.4: In each iteration, the new position of the thief dung beetle is calculated according to the above update formula (9), and its fitness is calculated, and the current fitness is compared with the best fitness of the current Kth generation, and the optimal value is selected to update the position; S4.5: After the local search is completed, the global optimal kernel parameter g and penalty factor c and the best fitness value are output.

6. The transformer fault diagnosis method based on the improved dung beetle algorithm optimized support vector machine according to claim 1 is characterized in that: The SVM model in step S5 is as follows: When the support vector machine solves general classification, its optimal hyperplane is expressed as the maximum margin classification problem, which is converted into a quadratic programming problem based on the maximization of the classification interval: Where w represents the normal vector of the optimal hyperplane, ||w|| 2 represents the complexity of the hyperplane, x i is the feature vector of the i-th sample point, y i is the category label of the i-th sample point, with a value of +1 or -1, indicating two different categories, and b is a bias term, which allows the optimal hyperplane not to pass through the origin; Since there are noise points in the data, when there are a small number of samples that cannot be separated by linear methods, a slack variable ξ is introduced i , the constraints are transformed into: y i (wx i +b)≥1-ξ i ,ξ i ≥0,i=1,2,…,n (11) After relaxation, the objective function is transformed into the following optimization problem: The kernel function K(xi,xj) is introduced to perform nonlinear transformation to solve the nonlinearity in the actual problem of transformer fault, and its dual optimization representation is obtained: a i is the Lagrange multiplier, and the corresponding optimal classification function is: The radial basis kernel function is selected, which is expressed as: In the formula That is, it represents the kernel parameter g. The larger the kernel parameter g, the smaller the influence range of a single training sample and the more complex the decision boundary. The smaller the kernel parameter g, the larger the influence range of a single training sample and the smoother the decision boundary.

Images