Transformer bushing partial discharge mode identification method and system

Abstract

The invention provides a transformer bushing partial discharge mode identification method and system, and relates to the technical field of power system fault diagnosis. Comprising the steps of obtaining transformer bushing partial discharge signals and extracting original features; carrying out dimension reduction processing on the extracted original features to obtain features after dimension reduction; a radial basis function SVM model is constructed, and parameters of the radial basis function SVM model are optimized through a GTO algorithm optimized by the chimpanzee troops; and inputting the dimensionality-reduced features into a parameter-optimized radial basis function SVM model for classification to obtain a classification identification result. According to the method, the KPCA algorithm and the GTO-SVM algorithm are combined, KPCA dimension reduction processing is carried out on high-frequency signal features, data redundancy is reduced, SVM parameter selection is optimized through the GTO algorithm, the classification and recognition performance is remarkably improved, the high recognition rate is shown in partial discharge fault diagnosis, and particularly the method has remarkable advantages in the aspect of processing complex fault modes.

Description

Technical Field

[0001] The invention belongs to the technical field of power system fault diagnosis, and in particular relates to a transformer bushing partial discharge pattern recognition method and system. Background Art

[0002] Transformer bushings are critical insulation components in power systems, and their operational reliability is directly related to the safety and stability of the power grid. If a bushing fails, it may cause serious chain reactions and huge economic losses. Traditional detection methods mainly rely on dielectric loss and capacitance tests, which usually require the transformer to be shut down. Not only are the operations complicated, but they are also unable to effectively detect potential partial discharge faults.

[0003] In addition, although existing online monitoring technologies such as relative capacitance, dielectric loss test and end screen voltage monitoring can monitor bushing faults to a certain extent, they are still significantly insufficient in timely and accurate identification of all types of partial discharge faults. Especially for transformer bushings with complex structures, their failure modes are diverse and highly concealed, making it extremely challenging to establish a complete fault diagnosis model.

[0004] In view of this, it is an urgent need to propose an efficient and accurate transformer bushing partial discharge pattern recognition method.

[0005] In the prior art, when establishing a transformer bushing partial discharge pattern recognition model, the determination of model parameters still has the following technical problems:

[0006] 1. High data dimension and high redundancy:

[0007] Partial discharge signals often contain a large number of time domain and frequency domain features, but directly using these high-dimensional data for classification and identification not only increases the computational complexity, but also easily causes data redundancy, affecting classification performance. Although traditional dimensionality reduction methods such as PCA can reduce data dimensions, they have limited processing effects on nonlinear data and cannot fully capture the nonlinear characteristics in complex fault modes.

[0008] 2. Model parameter selection depends on experience:

[0009] In support vector machines (SVMs), the choice of penalty factor C and kernel function parameter g directly affects the classification performance of the model. Existing methods mostly rely on manual parameter adjustment or grid search, which is not only time-consuming but also difficult to obtain the global optimal solution. Especially when dealing with small samples and nonlinear data, it is easy to fall into the local optimum, resulting in a decrease in classification accuracy.

[0010] 3. The optimization algorithm converges slowly and has insufficient precision:

[0011] Although traditional optimization algorithms such as particle swarm optimization (PSO) and genetic algorithm (GA) can optimize SVM parameters to a certain extent, they often face problems such as slow convergence and insufficient accuracy when solving complex optimization problems, resulting in unstable model performance, especially in multi-fault mode recognition, which shows great volatility.

[0012] These deficiencies cause the existing transformer partial discharge pattern recognition models to exhibit high misrecognition and false alarm rates in practical applications. In particular, when faced with complex and diverse partial discharge fault modes, the recognition performance is obviously insufficient. Summary of the invention

[0013] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides a transformer bushing partial discharge pattern recognition method and system, which combines KPCA and GTO-SVM algorithms, reduces data redundancy by performing KPCA dimensionality reduction processing on high-frequency signal features, and uses the GTO algorithm to optimize SVM parameter selection, thereby significantly improving the classification and recognition performance. It shows a high recognition rate in partial discharge fault diagnosis, and has significant advantages in processing complex fault modes.

[0014] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0015] A first aspect of the present invention provides a method for identifying partial discharge patterns of transformer bushings.

[0016] The transformer bushing partial discharge pattern recognition method comprises the following steps:

[0017] Obtain transformer bushing partial discharge signal and extract original features;

[0018] Perform dimensionality reduction processing on the extracted original features to obtain the features after dimensionality reduction;

[0019] Construct a radial basis function SVM model and optimize the parameters of the radial basis function SVM model through the gorilla troop optimization GTO algorithm;

[0020] The features after dimensionality reduction are input into the radial basis function SVM model with optimized parameters for classification to obtain the classification and recognition results.

[0021] As an optional technical solution, the transformer bushing partial discharge signal is obtained and the original feature is extracted, wherein the original feature includes time domain features and frequency domain features:

[0022] The time domain characteristics include absolute mean, root mean square value, variance, rise time, fall time, pulse width, skewness and kurtosis;

[0023] The frequency domain features include absolute mean, root mean square value, variance, maximum maximum corresponding amplitude, maximum maximum corresponding frequency, spectrum energy, skewness and kurtosis.

[0024] As an optional technical solution, the kernel principal component analysis (KPCA) method is used to reduce the dimension of the extracted original features, including:

[0025] Map the extracted original features into high-dimensional feature space using nonlinear mapping relationships;

[0026] Compute the covariance matrix in high-dimensional feature space;

[0027] Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and eigenvectors;

[0028] The cumulative contribution rate of each eigenvector is calculated according to the eigenvalue, and the eigenvectors with cumulative contribution rates higher than the set value are selected as principal components. The original features are projected onto the selected principal components to achieve dimensionality reduction.

[0029] As an optional technical solution, a radial basis function SVM model is constructed, and the optimization problem of RBF SVM is expressed as:

[0030]

[0031] Where w is the normal vector of the decision plane, b is the bias term, ξ i is the slack variable, y i is the label of the i-th sample, is the feature vector after dimensionality reduction by KPCA; C is the penalty parameter; n is the number of samples.

[0032] As an optional technical solution, the parameters of the radial basis function SVM model are optimized by using the Gorilla Force Optimization GTO algorithm, including:

[0033] Initialization parameters: Set the gorilla population size N and the maximum number of iterations Itermax, set the upper and lower bounds of the SVM model penalty parameter C and the RBF kernel parameter g, and obtain the initial population. i , g i ) indicates the position of the gorilla;

[0034] Determine the fitness function;

[0035] Update the position of the gorilla and calculate the fitness function. Iterate the optimal solution of the problem according to the fitness value of the new population. If the new fitness value exceeds the previous fitness value, set the current fitness value to the new fitness value.

[0036] When the number of iterations is greater than Itermax, C and g corresponding to the optimal solution are obtained as parameters of the SVM model.

[0037] As an optional technical solution, the gorilla's position is updated, including:

[0038] Entering the exploration phase, simulating the population migration behavior of gorillas and updating the individual positions of gorillas, the population migration behavior includes three migration positions, namely known position, unknown position and companion position;

[0039] Enter the development phase and obtain the parameter C of the exploration phase:

[0040] If the parameter C is greater than or equal to the given value, the population is used to follow the movement of the silverback gorilla to update the migration position;

[0041] If the parameter C is less than a given value, the migration position is updated using the male gorilla's courtship behavior.

[0042] As an optional technical solution, the reduced-dimensional features are input into the parameter-optimized radial basis function SVM model for classification to obtain the classification and recognition results. The specific formula is:

[0043]

[0044] Among them, f(x new ) is the predicted category label; α i and is the Lagrange multiplier, k(x i , x new ) is the kernel function RBF, b is the bias term, x i is the original feature obtained, x new is the new sample point;

[0045] If f(x new )>0, then x new are classified as positive, otherwise they are classified as negative.

[0046] A second aspect of the present invention provides a transformer bushing partial discharge pattern recognition system.

[0047] Transformer bushing partial discharge pattern recognition system, including:

[0048] The feature extraction module is configured to: obtain the transformer bushing partial discharge signal and perform original feature extraction;

[0049] The dimension reduction module is configured to: perform dimension reduction processing on the extracted original features to obtain features after dimension reduction;

[0050] The model optimization module is configured to: construct a radial basis function SVM model, and optimize the parameters of the radial basis function SVM model by using the gorilla troop optimization GTO algorithm;

[0051] The classification module is configured to: input the dimension-reduced features into a parameter-optimized radial basis function SVM model for classification to obtain a classification recognition result.

[0052] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the transformer bushing partial discharge pattern recognition method as described in the first aspect of the present invention.

[0053] The fourth aspect of the present invention provides an electronic device, comprising a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the transformer bushing partial discharge pattern recognition method as described in the first aspect of the present invention are implemented.

[0054] One or more of the above technical solutions have the following beneficial effects:

[0055] 1. The present invention provides a transformer bushing partial discharge pattern recognition method and system, which uses KPCA to reduce the dimension of the high-frequency signal features, and then optimizes the SVM parameter selection through the GTO algorithm. The GTO algorithm has greatly improved both the solution accuracy and the iteration speed, and has relatively good optimization calculation performance. The GTO algorithm can obtain the global optimal solution, avoid falling into the local optimum, resulting in low classification accuracy, thereby improving the classification recognition performance; at the same time, for the complex optimization problem of this application, it can still ensure the convergence speed, improve the model stability, and suppress fluctuations.

[0056] 2. The GTO algorithm simulates the migration and courtship behavior of gorillas, and optimizes the parameters in the SVM through population initialization, exploration phase, and development phase. When simulating the GTO algorithm:

[0057] Entering the exploration phase, the gorilla population migration behavior is simulated and the individual positions of the gorillas are updated. The algorithm simulates three migration positions: known position, unknown position and companion position. The migration behavior corresponds to the exploration phase of the bionic computing algorithm.

[0058] Enter the development phase, imitate the courtship behavior of adolescent male gorillas competing for females and the behavior of the group following the movement of the silverback gorilla. Get the parameter C of the exploration phase:

[0059] If the parameter C is greater than or equal to the given value, the population is used to follow the movement of the silverback gorilla to update the migration position;

[0060] If the parameter C is less than a given value, the migration position is updated using the male gorilla's courtship behavior.

[0061] 3. The original features of the transformer bushing partial discharge signal include time domain features and frequency domain features: time domain features include absolute mean, root mean square value, variance, rise time, fall time, pulse width, skewness and kurtosis; frequency domain features include absolute mean, root mean square value, variance, maximum maximum value corresponding amplitude, maximum maximum value corresponding frequency, spectrum energy, skewness and kurtosis. The features of the transformer bushing partial discharge signal can be fully extracted through the above time and frequency domain feature extraction.

[0062] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0064] Figure 1 This is a flow chart of the method of the first embodiment.

[0065] Figure 2 This is the KPCA algorithm framework diagram.

[0066] Figure 3 This is the solution graph of the test function by the GTO algorithm, PSO algorithm and SSA algorithm.

[0067] Figure 4 This is the layout diagram of the bushing partial discharge test.

[0068] Figure 5 Set up diagram for bushing partial discharge fault. DETAILED DESCRIPTION

[0069] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0070] It should be noted that the terms used herein are for describing specific embodiments only and are not intended to be limiting of exemplary embodiments according to the present invention.

[0071] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0072] Embodiment 1

[0073] The embodiment of the present invention proposes a transformer bushing partial discharge pattern recognition method based on KPCA and GTO-SVM algorithms. By performing KPCA dimensionality reduction processing on high-frequency signal features, data redundancy is reduced, and the SVM parameter selection is optimized using the GTO algorithm, which significantly improves the classification and recognition performance of transformer bushing partial discharge. Experimental results show that the method of this embodiment shows a high recognition rate in partial discharge fault diagnosis, especially in dealing with complex fault modes. Compared with traditional methods, the present invention not only improves the accuracy of diagnosis, but also has strong engineering application value, providing an effective solution for early fault detection of transformer bushings.

[0074] like Figure 1 As shown, the transformer bushing partial discharge pattern recognition method comprises the following steps:

[0075] Obtain transformer bushing partial discharge signal and extract original features;

[0076] Perform dimensionality reduction processing on the extracted original features to obtain the features after dimensionality reduction;

[0077] Construct a radial basis function SVM model and optimize the parameters of the radial basis function SVM model through the gorilla troop optimization GTO algorithm;

[0078] The features after dimension reduction are input into the radial basis function SVM model with optimized parameters for classification to obtain the classification and recognition results.

[0079] The specific implementation process of this embodiment will be described in detail below, including the following steps:

[0080] (1) Signal feature extraction: By extracting the features of the PD signal, the signal can be converted into a set of simple feature vectors. This paper extracts 16 time-frequency domain feature parameters and uses the KPCA method to reduce the dimension of the signal features, retaining the feature vectors with a cumulative contribution rate D greater than 85%.

[0081] (2) GTO-SVM algorithm: The GTO algorithm simulates the population migration and courtship behavior of gorillas and optimizes the parameters in the SVM through population initialization, exploration phase, and development phase. Experimental results show that the GTO algorithm is superior to the PSO and SSA algorithms in terms of solution accuracy and iteration speed.

[0082] (3) Experimental verification: In the 110kV bushing partial discharge experiment, four typical faults were set: metal tip discharge, voltage equalizing cover suspension discharge, lower porcelain sleeve surface discharge, and capacitor screen notch defect discharge. The experimental results show that the total recognition rate of the GTO-SVM algorithm for the four faults is 95.83%, which is significantly better than the traditional SVM and BP neural network.

[0083] The specific method of step (1) is:

[0084] In transformer bushing partial discharge pattern recognition, signal feature extraction is a key step to achieve high-precision fault diagnosis. Partial discharge signals are usually high-frequency pulse signals. By extracting the time domain and frequency domain features of these signals, the original signals can be converted into a set of representative feature vectors, thus providing an effective data basis for subsequent pattern recognition.

[0085] First, 16 time-frequency domain characteristic parameters are extracted from the partial discharge signal recorded by the high-frequency sensor. These characteristic parameters include:

[0086] Time domain characteristic parameters: absolute mean, root mean square value, variance, rise time, fall time, pulse width, skewness, and kurtosis.

[0087] Frequency domain characteristic parameters: absolute mean, root mean square value, variance, amplitude corresponding to the maximum maximum, frequency corresponding to the maximum maximum, spectrum energy, skewness, and kurtosis.

[0088] After extracting these characteristic parameters, directly using high-dimensional characteristic parameters for classification and recognition may lead to high computational complexity and slow recognition speed, and high-dimensional data may have large redundancy, affecting classification accuracy. Therefore, this embodiment uses the Kernel Principal Components Analysis (KPCA) method to reduce the dimensionality of high-dimensional characteristic parameters. KPCA is an algorithm for nonlinear data dimensionality reduction, which can remove useless information and reduce data dimensions. The specific steps are as follows:

[0089] 1) Data Mapping

[0090] Using the nonlinear mapping relationship φ, the data sample X in the input space is transformed into {x i |i=1,2,...,n;x i ∈R} is mapped to the high-dimensional feature space G, that is, x i →φ(x i ), and perform principal component analysis in G to extract nonlinear structural information. The basic framework of the KPCA algorithm is as follows Figure 2 As shown. Among them, x i The original features obtained.

[0091] 2) Covariance matrix calculation

[0092] Assume that the number of samples is n, and each sample has m-dimensional features, then the original data is the following matrix X:

[0093]

[0094] The nonlinear function φ(x) is used to map the original data to a high-dimensional space, that is, φ(x i )=[φ(x1),φ(x2),…,φ(x n )] The covariance matrix in the high-dimensional feature space is obtained as follows:

[0095]

[0096] 3) Feature vector analysis

[0097] right Perform eigenvector analysis, assuming its eigenvalue is λ and eigenvector is v, then:

[0098]

[0099] 4) Linear combination of eigenvectors

[0100] Because each vector in the space can be composed of a linear combination of other vectors in the space, all samples are used to linearly construct the feature vector, that is:

[0101]

[0102] In formula (3), α i (i=1,2,…,n) is the coefficient corresponding to v. Substituting v into equation (3), we can obtain the following by replacing the eigenvector:

[0103]

[0104] 5) Introduction of kernel function

[0105] The kernel function K(x i , x j )=φ(x i )×φ(x j ) and the kernel matrix K=(K ij ), K ij =K(x i , x j ) is introduced into formula (5) and can be simplified as:

[0106] Kα=λ i α (6)

[0107] In formula (6), α=(α1, α2,…,a n ) T is the kernel function inner product matrix K ij The corresponding eigenvector, matrix K ij The expression is:

[0108]

[0109] 6) New sample conversion

[0110] Let x i is a new sample, x i After KPCA transformation, the j-th dimension coordinate z j The expression is:

[0111]

[0112] In the judgment process, a component z j The overall importance is reflected by the contribution rate. The component with a high cumulative contribution rate D is regarded as the principal component, which plays a major role in making judgments on the system. In order to retain the characteristic information of the fault as much as possible, the cumulative contribution rate D is usually selected to be higher than a certain set value as the selection standard for the number of principal components. Generally, D is taken as ≥ 85%. At this time, the first n feature vectors can relatively completely represent the retained feature information.

[0113] The specific method of step (2) is:

[0114] In order to improve the accuracy of partial discharge pattern recognition of transformer bushing, a method for optimizing support vector machine (SVM) parameters based on Gorilla Troops Optimizer (GTO) algorithm was proposed.

[0115] The GTO algorithm simulates the population migration and courtship behavior of gorillas. Population migration refers to the process of a population moving in the same direction. The algorithm simulates three migration positions, namely known positions, unknown positions, and companion positions. The migration behavior corresponds to the exploration stage of the bionic computing algorithm. Courtship behavior refers to the behavior of adolescent male gorillas competing for females, which corresponds to the development stage of the bionic computing algorithm. It is worth mentioning that in order to improve the convergence speed, the algorithm also simulates the movement of the silverback gorilla, the leader of the population. When the population migrates, the silverback gorilla will make decisions and lead other gorillas to migrate. This behavior corresponds to the development stage of the bionic computing algorithm. The bionic optimization algorithm is used to perform global search and local optimization of the SVM parameters to improve the performance of the classifier. The steps of the GTO algorithm are as follows:

[0116] 1) Population initialization

[0117] The mathematical expression for initializing the position of the gorilla population is as follows:

[0118] X(i, j)=rand×(UB-LB)+LB (10)

[0119] Where UB and LB represent the upper and lower bounds of the search space, respectively, and rand is a random number in [0,1].

[0120] 2) Exploration stage

[0121] The GTO algorithm has three exploration mechanisms, corresponding to the three migration positions of gorillas. When rand < p(0.03), it means that the gorilla migrates to an unknown location; when rand < 0.5, it means migrating to a known location; when rand ≥ 0.5, it means migrating to a companion. The mathematical expressions are as follows:

[0122] where r1, r2, r3 are random numbers in [0,1], Xr represents the optimal position of any gorilla in the previous generation population, X represents the optimal position of the previous generation of gorillas, GX represents the position of the current gorilla, GXr represents the position of any gorilla in the current population, t represents the gorilla individual, and the parameters C, L, H are as shown in formulas (12), (13), and (14).

[0123]

[0124] L = C × l (13)

[0125] H = Z × X(t) (14)

[0126] F = cos(2r4) + 1 (15)

[0127] Z ∈ [-C, C] (16)

[0128] where r4 is a random number in [0,1], It is the current iteration value, MaxIt is the maximum iteration value, the value range of l is from -1 to 1, and Z is a random value in the range from -C to C.

[0129] 3) Exploitation stage

[0130] When C ≥ W, it means that the population follows the silverback gorilla to migrate. The mathematical expression is:

[0131] GX(t) = L·M·(X(t) - X b ) + X(t) (17)

[0132]

[0133] g = 2 L (19)

[0134] where N is the population size, Xb is the position of the silverback gorilla (optimal solution), W is a given parameter with a value of 0.8.

[0135] When C < W, it means the courtship behavior of male gorillas. The mathematical expression is:

[0136] GX(t) = Xb -A(X b ·QX(t)·Q) (20)

[0137] The parameter Q is used to simulate the impact force, and the parameter A is used to simulate the degree of violence. The mathematical expression is as follows:

[0138] Q=2r5-1 (21)

[0139] A=β·E (22)

[0140]

[0141] Where r5 is a random number in [0,1], β is a given parameter with a value of 3, randn(1,d) is a function that generates d standard normal distribution random numbers, E is used to simulate the impact of violence levels in different dimensions, and d represents the population dimension.

[0142] In order to verify the effectiveness and superiority of the proposed GTO algorithm, the test function shown in formula (24) was solved 50 times using the GTO algorithm, PSO algorithm and SSA algorithm respectively.

[0143]

[0144] The algorithm parameter settings are given as follows: the total number of individuals in the three algorithm populations is set to 100, and the maximum number of iterations is 500. The calculation results are given as follows Figure 3 shown.

[0145] Figure 3 From the calculation results of the three algorithms for two classic functions, it can be seen that the GTO algorithm has greatly improved both in solution accuracy and iteration speed. This shows that the GTO algorithm has relatively good optimization calculation performance.

[0146] This embodiment aims at the diagnosis problem of partial discharge types with small samples and nonlinear characteristics, and establishes a multi-classification SVM diagnosis and recognition model based on the principle of small structural risk. In the SVM model, C is a penalty factor that controls the degree of model fitting, and g is a radial basis width parameter; and the classification and recognition effect of the SVM algorithm is greatly affected by the C and g parameters. Therefore, selecting a set of appropriate parameters plays an important role in improving the classification and recognition accuracy of the SVM algorithm. In view of this, a SVM parameter optimization selection method based on the GTO algorithm is proposed. The specific process of this method is:

[0147] 1) Initialization parameters

[0148] First, set the population size N and the maximum number of iterations Itermax of the algorithm, set the upper and lower bounds of the SVM parameters C and g, and obtain the initial population. i , gi ) to indicate the location of the gorilla.

[0149] 2) Calculate the fitness function

[0150] According to each individual's (C i , g i ), classify the training set based on SVM and obtain the classification result. According to the fitness function Calculate individual fitness, where y t Indicates the number of correct classifications, y f Indicates the number of classification errors. The smaller the Fitness is, the better the classification effect is.

[0151] 3) Population update

[0152] According to equations (11)-(23), the gorilla position is updated and the fitness function is calculated, and the optimal solution of the problem is iterated based on the fitness value of the new population.

[0153] 4) Output results

[0154] When the number of iterations is greater than Itermax, C and g corresponding to the optimal solution at this time are output as parameters of the SVM algorithm.

[0155] The specific method of step (3) is:

[0156] 1) Experimental platform construction

[0157] like Figure 4 As shown in the figure, the experimental platform consists of a booster transformer and a bushing to form the main circuit. Among them, booster transformer is a booster transformer; HFCT is a high-frequency current transformer; HV probe is a high-voltage probe.

[0158] A voltage-dividing capacitor is connected in parallel to the outside of the bushing to collect the reference voltage signal in real time. The step-up transformer is connected to the top of the bushing through a bellows and a voltage-equalizing cover. Four high-frequency sensors are arranged at different positions 3m away from the bushing to improve the discrimination of the corresponding time of the high-frequency signal received by the sensor. A partial discharge instrument is connected to the grounding wire of the bushing flange to determine the partial discharge starting voltage and discharge amount.

[0159] 2) Fault type setting

[0160] like Figure 5 As shown in Figure 1, four different types of typical faults are set in the 110kV bushing partial discharge experiment.

[0161] Fault a: Metal tip discharge. In the simulation of actual working conditions, the top surface of the bushing has metal protrusions, which causes partial discharge. The metal tip is attached to the high-voltage end surface of the bushing top, and the tip extends 5 cm.

[0162] Fault b: Suspended discharge of the voltage-equalizing cover. Simulate the fault under actual working conditions where the voltage-equalizing cover falls off and hangs on the lead-out line. Use cloth tape to suspend the voltage-equalizing cover so that the distance between the voltage-equalizing cover and the high-voltage end of the bushing is about 3 cm.

[0163] Fault c: Discharge on the surface of the lower porcelain sleeve. During the long-term operation of the simulated bushing, due to the dirt or roughness on the surface of the lower porcelain sleeve, discharge occurs on the surface when the voltage increases. Use a cloth tape to fix the hard copper wire on the surface of the lower porcelain sleeve to keep it in good contact with the voltage equalizing cover at the high-voltage end.

[0164] Fault d: Discharge due to a gap defect in the capacitor screen. This simulates an artificial gap defect at the bottom of the capacitor screen (end screen), which causes an extremely uneven electric field and thus discharges. This fault is located inside the casing body and has no direct contact with the outside. The high-frequency signal generated by partial discharge leaks from the inside of the casing to the external space.

[0165] 3) Identification results and analysis

[0166] In order to verify the effectiveness and superiority of the proposed partial discharge type diagnosis method, 150 signal samples were sampled for each partial discharge fault type and divided into training set and test set in a ratio of 4:1. The collected high-frequency signals were feature extracted, and the KPCA algorithm was used to reduce the dimension of the signal features, and the feature vectors with a cumulative contribution rate D ≥ 85% were selected. GTO-SVM, SVM and BP neural network methods were used to classify and identify various types of partial discharge signals.

[0167] The transformer bushing partial discharge pattern recognition method based on KPCA and GTO-SVM algorithm in the embodiment of the present invention first uses KPCA to reduce the dimension of the high-frequency signal features, and then optimizes the SVM parameter selection through the GTO algorithm to improve the classification and recognition performance. The experimental results show that the algorithm has a high recognition rate in partial discharge fault diagnosis. Compared with the traditional SVM and BP neural network algorithms, the total recognition rates are increased by 5% and 5.83% respectively, verifying its effectiveness and superiority in transformer bushing fault diagnosis.

[0168] The specific implementation process is:

[0169] 1) Original signal acquisition and data preprocessing:

[0170] The high-frequency sensor collects the original partial discharge signal of the transformer bushing. For each typical fault, 150 valid signals are collected. During the collection period, signals with too small amplitude are actively discarded. A total of 600 signals are collected for the four defects. Pre-processing is performed depending on the noise situation, such as using wavelet noise reduction.

[0171] 2) Feature extraction:

[0172] 16 time-frequency domain feature parameters of all 600 signals are extracted, including time domain features (such as absolute mean, root mean square value, variance, etc.) and frequency domain features (such as spectrum energy, corresponding amplitude of maximum maximum value, etc.).

[0173] 3) Kernel function selection and data mapping:

[0174] Select Gaussian radial basis function (RBF) to map data into high-dimensional space. Function formula: x i and x j is the data point in the original space, and σ is the bandwidth parameter of the RBF kernel. The kernel function is used to map the original data to a high-dimensional feature space. φ(x)∈G, where φ is a nonlinear mapping function and G is a high-dimensional feature space.

[0175] 4) Covariance matrix calculation:

[0176] The nonlinear function φ(x) maps the original data to a high-dimensional space to obtain the covariance matrix in the high-dimensional feature space.

[0177] 5) Eigenvalue decomposition:

[0178] Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and eigenvectors. Kv = λv, where λ is the eigenvalue and v is the corresponding eigenvector.

[0179] 6) Select the principal component for dimensionality reduction:

[0180] The cumulative contribution rate of each eigenvector is calculated according to the eigenvalue, and the eigenvectors with cumulative contribution rates higher than 85% are selected as principal components. The original data are projected onto the selected principal components to achieve dimensionality reduction.

[0181] 7) Start classification:

[0182] The data after KPCA dimension reduction are used as the input features of radial basis function (RBF) SVM, and its penalty parameter C and RBF kernel parameter g are established by gorilla troop optimization (GTO) algorithm.

[0183] 8) Build SVM model:

[0184] Use the training data to build the SVM model. The optimization problem of RBF SVM can be expressed as:

[0185] Where w is the normal vector of the decision plane, b is the bias term, ξ i is the slack variable, y i is the label of the i-th sample, It is the feature vector after dimensionality reduction by KPCA.

[0186] 9) Train SVM and evaluate the model:

[0187] Use the training data set to train the SVM and find the optimal decision boundary. Randomly select 120 data for each defect as the training set, and the total number of defects is 480. Evaluate the performance of the SVM model on the remaining 120 data test sets.

[0188] 10) Classification decision:

[0189] For the test set of 120 data, the SVM model is used for classification. α i and is the Lagrange multiplier, k(x i , x new ) is the kernel function RBF, and b is the bias term.

[0190] 11) Predict categories and compare:

[0191] The category of the test set is determined according to the value of the decision function. If f(x new )>0, then x new It is classified into positive class (Positive class, the result that we hope the model can identify and predict), otherwise it is classified into negative class (Negative class, prediction failure). Finally, the result is output.

[0192] 12) The three classifiers are compared and the recognition results are shown in Table 1.

[0193] Table 1 Recognition results of each method

[0194]

[0195] According to Table 1, the total recognition rate of the GTO-SVM algorithm in this paper for the four defects is relatively high, which is 95.83%. Compared with the SVM method and the BP neural network, the total recognition rates of the algorithm in this paper are increased by 5% and 5.83% respectively, which proves that the algorithm of this embodiment is effective and feasible in the partial discharge fault diagnosis of transformer bushings, and provides a new idea for the identification of partial discharge.

[0196] Embodiment 2

[0197] This embodiment discloses a transformer bushing partial discharge pattern recognition system.

[0198] Transformer bushing partial discharge pattern recognition system, including:

[0199] The feature extraction module is configured to: obtain the transformer bushing partial discharge signal and perform original feature extraction;

[0200] The dimension reduction module is configured to: perform dimension reduction processing on the extracted original features to obtain features after dimension reduction;

[0201] The model optimization module is configured to: construct a radial basis function SVM model, and optimize the parameters of the radial basis function SVM model by using the gorilla troop optimization GTO algorithm;

[0202] The classification module is configured to: input the dimension-reduced features into a parameter-optimized radial basis function SVM model for classification to obtain a classification recognition result.

[0203] The above-mentioned feature extraction module, dimension reduction module, model optimization module and classification module are specifically configured to perform the following process:

[0204] The specific method of step (1) is:

[0205] In transformer bushing partial discharge pattern recognition, signal feature extraction is a key step to achieve high-precision fault diagnosis. Partial discharge signals are usually high-frequency pulse signals. By extracting the time domain and frequency domain features of these signals, the original signals can be converted into a set of representative feature vectors, thus providing an effective data basis for subsequent pattern recognition.

[0206] First, 16 time-frequency domain characteristic parameters are extracted from the partial discharge signal recorded by the high-frequency sensor. These characteristic parameters include:

[0207] Time domain characteristic parameters: absolute mean, root mean square value, variance, rise time, fall time, pulse width, skewness, and kurtosis.

[0208] Frequency domain characteristic parameters: absolute mean, root mean square value, variance, amplitude corresponding to the maximum maximum, frequency corresponding to the maximum maximum, spectrum energy, skewness, and kurtosis.

[0209] After extracting these characteristic parameters, directly using high-dimensional characteristic parameters for classification and recognition may lead to high computational complexity and slow recognition speed, and high-dimensional data may have large redundancy, affecting classification accuracy. Therefore, this embodiment uses the Kernel Principal Components Analysis (KPCA) method to reduce the dimensionality of high-dimensional characteristic parameters. KPCA is an algorithm for nonlinear data dimensionality reduction, which can remove useless information and reduce data dimensions. The specific steps are as follows:

[0210] 7) Data Mapping

[0211] Using the nonlinear mapping relationship φ, the data sample X in the input space is transformed into {x i |i=1,2,…,n;x i∈R} is mapped to the high-dimensional feature space G, that is, x i →φ(x i ), and perform principal component analysis in G to extract nonlinear structural information. The basic framework of the KPCA algorithm is as follows Figure 2 As shown. Among them, x i The original features obtained.

[0212] 8) Covariance matrix calculation

[0213] Assume that the number of samples is n, and each sample has m-dimensional features, then the original data is the following matrix X:

[0214]

[0215] The nonlinear function φ(x) is used to map the original data to a high-dimensional space, that is, φ(xi) = [φ(x1), φ(x2), …, φ(x n )] The covariance matrix in the high-dimensional feature space is obtained as follows:

[0216]

[0217] 9) Feature vector analysis

[0218] right Perform eigenvector analysis, assuming its eigenvalue is λ and eigenvector is v, then:

[0219]

[0220] 10) Linear combination of eigenvectors

[0221] Because each vector in the space can be composed of a linear combination of other vectors in the space, all samples are used to linearly construct the feature vector, that is:

[0222]

[0223] In formula (3), α i (i=1,2,…,n) is the coefficient corresponding to v. Substituting v into equation (3), we can obtain the following by replacing the eigenvector:

[0224]

[0225] 11) Introduction of kernel function

[0226] The kernel function K(x i , x j )=φ(x i )×φ(x j ) and the kernel matrix K=(K ij ), K ij =K(xi , x j ) is introduced into formula (5) and can be simplified as:

[0227] Kα=λ i α (6)

[0228] In formula (6), α = (α1, α2, ..., α n ) T is the kernel function inner product matrix K ij The corresponding eigenvector, matrix K ij The expression is:

[0229]

[0230] 12) New sample conversion

[0231] Let x i is a new sample, x i After KPCA transformation, the j-th dimension coordinate z j The expression is:

[0232]

[0233] In the judgment process, a component z j The overall importance is reflected by the contribution rate. The component with a high cumulative contribution rate D is regarded as the principal component, which plays a major role in making judgments on the system. In order to retain the characteristic information of the fault as much as possible, the cumulative contribution rate D is usually selected to be higher than a certain set value as the selection standard for the number of principal components. Generally, D is taken as ≥ 85%. At this time, the first n feature vectors can relatively completely represent the retained feature information.

[0234] The specific method of step (2) is:

[0235] In order to improve the accuracy of partial discharge pattern recognition of transformer bushing, a method for optimizing support vector machine (SVM) parameters based on Gorilla Troops Optimizer (GTO) algorithm was proposed.

[0236] The GTO algorithm simulates the population migration and courtship behavior of gorillas. Population migration refers to the process in which a population moves in the same direction. The algorithm simulates three migration positions, namely known positions, unknown positions, and companion positions. The migration behavior corresponds to the exploration stage of the bionic computing algorithm. Courtship behavior refers to the behavior of male gorillas in adolescence competing for females, which corresponds to the exploitation stage of the bionic computing algorithm. It is worth mentioning that to improve the convergence speed, the algorithm also simulates the movement of the silverback gorilla, the leader of the population. When the population migrates, the silverback gorilla makes decisions and leads other gorillas to migrate, and this behavior corresponds to the exploitation stage of the bionic computing algorithm. The parameters of the SVM are globally searched and locally optimized through the bionic optimization algorithm, thereby improving the performance of the classifier. The steps of the GTO algorithm are as follows:

[0237] 4) Population initialization

[0238] The mathematical expression for initializing the positions of the gorilla population is as follows:

[0239] X(i, j) = rand × (UB - LB) + LB (10)

[0240] where UB and LB represent the upper and lower bounds of the search space respectively, and rand is a random number in [0, 1].

[0241] 5) Exploration stage

[0242] The GTO algorithm has three exploration mechanisms, corresponding to the three migration positions of gorillas. When rand < p (0.03), it means that the gorilla migrates to an unknown location; when rand < 0.5, it means migrating to a known location; when rand ≥ 0.5, it means migrating to a companion. The mathematical expressions are as follows:

[0243]

[0244] where r1, r2, r3 are random numbers in [0, 1], Xr represents the optimal position of any gorilla in the previous generation population, X represents the optimal position of the previous generation of gorillas, GX represents the position of the current gorilla, GXr represents the position of any gorilla in the current population, t represents the gorilla individual, and the parameters C, L, H are as shown in formulas (12), (13), and (14).

[0245]

[0246] L = C × l (13)

[0247] H = Z × X(t) (14)

[0248] F = cos(2r4) + 1 (15)

[0249] Z ∈ [-C, C] (16)

[0250] where r4 is a random number in [0, 1], It is the current iteration value, MaxIt is the maximum iteration value, the value range of l is from -1 to 1, and Z is a random value in the range of -C to C.

[0251] 6) Development stage

[0252] When C ≥ W, it means the population follows the silverback gorilla to migrate. The mathematical expression is:

[0253] GX(t) = L·M·(X(t) - X b ) + X(t) (17)

[0254]

[0255] g = 2 L (19)

[0256] where N is the population size, Xb is the position of the silverback gorilla (optimal solution), W is a given parameter with a value of 0.8.

[0257] When C < W, it means the courtship behavior of male gorillas. The mathematical expression is:

[0258] GX(t) = X b - A(X b ·Q - X(t)·Q) (20)

[0259] where the parameter Q is used to simulate the impact force, and the parameter A is used to simulate the degree of violence. Their mathematical expressions are as follows:

[0260] Q = 2r5 - 1 (21)

[0261] A = β·E (22)

[0262]

[0263] where r5 is a random number in [0, 1], β is a given parameter with a value of 3, randn(1, d) is a function that generates d standard normal distribution random numbers, E is used to simulate the influence of the degree of violence in different dimensions, and d represents the population dimension.

[0264] To verify the effectiveness and superiority of the proposed GTO algorithm, the GTO algorithm, PSO algorithm, and SSA algorithm were used to solve the test function shown in Equation (24) 50 times respectively.

[0265]

[0266] The algorithm parameter settings are given as follows: the total number of individuals in the three algorithm populations is set to 100, and the maximum number of iterations is 500. The calculation results are given as follows Figure 3 shown.

[0267] Figure 3 From the calculation results of the three algorithms for two classic functions, it can be seen that the GTO algorithm has greatly improved both in solution accuracy and iteration speed. This shows that the GTO algorithm has relatively good optimization calculation performance.

[0268] This embodiment aims at the diagnosis problem of partial discharge types with small samples and nonlinear characteristics, and establishes a multi-classification SVM diagnosis and recognition model based on the principle of small structural risk. In the SVM model, C is a penalty factor that controls the degree of model fitting, and g is a radial basis width parameter; and the classification and recognition effect of the SVM algorithm is greatly affected by the C and g parameters. Therefore, selecting a set of appropriate parameters plays an important role in improving the classification and recognition accuracy of the SVM algorithm. In view of this, a SVM parameter optimization selection method based on the GTO algorithm is proposed. The specific process of this method is:

[0269] 1) Initialization parameters

[0270] First, set the population size N and the maximum number of iterations Itermax of the algorithm, set the upper and lower bounds of the SVM parameters C and g, and obtain the initial population. i , g i ) to indicate the location of the gorilla.

[0271] 2) Calculate the fitness function

[0272] According to each individual's (C i , g i ), classify the training set based on SVM and obtain the classification result. According to the fitness function Calculate individual fitness, where y t Indicates the number of correct classifications, y f Indicates the number of classification errors. The smaller the Fitness is, the better the classification effect is.

[0273] 3) Population update

[0274] According to equations (11)-(23), the gorilla position is updated and the fitness function is calculated, and the optimal solution of the problem is iterated based on the fitness value of the new population.

[0275] 4) Output results

[0276] When the number of iterations is greater than Itermax, C and g corresponding to the optimal solution at this time are output as parameters of the SVM algorithm.

[0277] The specific method of step (3) is:

[0278] 4) Experimental platform construction

[0279] like Figure 4 As shown in the figure, the experimental platform consists of a booster transformer and a bushing to form the main circuit. Among them, booster transformer is a booster transformer; HFCT is a high-frequency current transformer; HV probe is a high-voltage probe.

[0280] A voltage-dividing capacitor is connected in parallel to the outside of the bushing to collect the reference voltage signal in real time. The step-up transformer is connected to the top of the bushing through a bellows and a voltage-equalizing cover. Four high-frequency sensors are arranged at different positions 3m away from the bushing to improve the discrimination of the corresponding time of the high-frequency signal received by the sensor. A partial discharge instrument is connected to the grounding wire of the bushing flange to determine the partial discharge starting voltage and discharge amount.

[0281] 5) Fault type setting

[0282] like Figure 5 As shown in Figure 1, four different types of typical faults are set in the 110kV bushing partial discharge experiment.

[0283] Fault a: Metal tip discharge. In the simulation of actual working conditions, the top surface of the bushing has metal protrusions, which causes partial discharge. The metal tip is attached to the high-voltage end surface of the bushing top, and the tip extends 5 cm.

[0284] Fault b: Suspended discharge of the voltage-equalizing cover. Simulate the fault under actual working conditions where the voltage-equalizing cover falls off and hangs on the lead-out line. Use cloth tape to suspend the voltage-equalizing cover so that the distance between the voltage-equalizing cover and the high-voltage end of the bushing is about 3 cm.

[0285] Fault c: Discharge on the surface of the lower porcelain sleeve. During the long-term operation of the simulated bushing, due to the dirt or roughness on the surface of the lower porcelain sleeve, discharge occurs on the surface when the voltage increases. Use a cloth tape to fix the hard copper wire on the surface of the lower porcelain sleeve to keep it in good contact with the voltage equalizing cover at the high-voltage end.

[0286] Fault d: Discharge due to a gap defect in the capacitor screen. This simulates an artificial gap defect at the bottom of the capacitor screen (end screen), which causes an extremely uneven electric field and thus discharges. This fault is located inside the casing body and has no direct contact with the outside. The high-frequency signal generated by partial discharge leaks from the inside of the casing to the external space.

[0287] 6) Identification results and analysis

[0288] In order to verify the effectiveness and superiority of the proposed partial discharge type diagnosis method, 150 signal samples were sampled for each partial discharge fault type and divided into training set and test set in a ratio of 4:1. The collected high-frequency signals were feature extracted, and the KPCA algorithm was used to reduce the dimension of the signal features, and the feature vectors with a cumulative contribution rate D ≥ 85% were selected. GTO-SVM, SVM and BP neural network methods were used to classify and identify various types of partial discharge signals.

[0289] The transformer bushing partial discharge pattern recognition method based on KPCA and GTO-SVM algorithm in the embodiment of the present invention first uses KPCA to reduce the dimension of the high-frequency signal features, and then optimizes the SVM parameter selection through the GTO algorithm to improve the classification and recognition performance. The experimental results show that the algorithm has a high recognition rate in partial discharge fault diagnosis. Compared with the traditional SVM and BP neural network algorithms, the total recognition rates are increased by 5% and 5.83% respectively, verifying its effectiveness and superiority in transformer bushing fault diagnosis.

[0290] The specific implementation process is:

[0291] 1) Original signal acquisition and data preprocessing:

[0292] The high-frequency sensor collects the original partial discharge signal of the transformer bushing. For each typical fault, 150 valid signals are collected. During the collection period, signals with too small amplitude are actively discarded. A total of 600 signals are collected for the four defects. Pre-processing is performed depending on the noise situation, such as using wavelet noise reduction.

[0293] 2) Feature extraction:

[0294] 16 time-frequency domain feature parameters of all 600 signals are extracted, including time domain features (such as absolute mean, root mean square value, variance, etc.) and frequency domain features (such as spectrum energy, corresponding amplitude of maximum maximum value, etc.).

[0295] 3) Kernel function selection and data mapping:

[0296] Select Gaussian radial basis function (RBF) to map data into high-dimensional space. Function formula: x i and x j is the data point in the original space, and σ is the bandwidth parameter of the RBF kernel. The kernel function is used to map the original data to a high-dimensional feature space. φ(x)∈G, where φ is a nonlinear mapping function and G is a high-dimensional feature space.

[0297] 4) Covariance matrix calculation:

[0298] The nonlinear function φ(x) maps the original data to a high-dimensional space to obtain the covariance matrix in the high-dimensional feature space.

[0299] 5) Eigenvalue decomposition:

[0300] Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and eigenvectors. Kv = λv, where λ is the eigenvalue and v is the corresponding eigenvector.

[0301] 6) Select the principal component for dimensionality reduction:

[0302] The cumulative contribution rate of each eigenvector is calculated according to the eigenvalue, and the eigenvectors with cumulative contribution rates higher than 85% are selected as principal components. The original data are projected onto the selected principal components to achieve dimensionality reduction.

[0303] 7) Start classification:

[0304] The data after KPCA dimension reduction are used as the input features of radial basis function (RBF) SVM, and its penalty parameter C and RBF kernel parameter g are established by gorilla troop optimization (GTO) algorithm.

[0305] 8) Build SVM model:

[0306] Use the training data to build the SVM model. The optimization problem of RBF SVM can be expressed as:

[0307] Where w is the normal vector of the decision plane, b is the bias term, ξ i is the slack variable, y i is the label of the i-th sample, It is the feature vector after dimensionality reduction by KPCA.

[0308] 9) Train SVM and evaluate the model:

[0309] Use the training data set to train the SVM and find the optimal decision boundary. Randomly select 120 data for each defect as the training set, and the total number of defects is 480. Evaluate the performance of the SVM model on the remaining 120 data test sets.

[0310] 10) Classification decision:

[0311] For the test set of 120 data, the SVM model is used for classification. α i and is the Lagrange multiplier, k(x i , x new ) is the kernel function RBF, and b is the bias term.

[0312] 11) Predict categories and compare:

[0313] The category of the test set is determined according to the value of the decision function. If f(x new )>0, then x new It is classified into positive class (Positive class, the result that we hope the model can identify and predict), otherwise it is classified into negative class (Negative class, prediction failure). Finally, the result is output.

[0314] 12) The three classifiers are compared and the recognition results are shown in Table 1.

[0315] Table 1 Recognition results of each method

[0316]

[0317] According to Table 1, the total recognition rate of the GTO-SVM algorithm in this paper for the four defects is relatively high, which is 95.83%. Compared with the SVM method and the BP neural network, the total recognition rates of the algorithm in this paper are increased by 5% and 5.83% respectively, which proves that the algorithm of this embodiment is effective and feasible in the partial discharge fault diagnosis of transformer bushings, and provides a new idea for the identification of partial discharge.

[0318] Embodiment 3

[0319] The purpose of this embodiment is to provide a computer-readable storage medium.

[0320] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the transformer bushing partial discharge pattern recognition method as described in Embodiment 1 of the present disclosure.

[0321] Embodiment 4

[0322] The purpose of this embodiment is to provide an electronic device.

[0323] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps in the transformer bushing partial discharge pattern recognition method as described in Embodiment 1 of the present disclosure are implemented.

[0324] The steps involved in the apparatuses of the above embodiments 2, 3 and 4 correspond to the method embodiment 1, and the specific implementation methods can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0325] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0326] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A method for identifying partial discharge patterns of transformer bushings, characterized in that: The following steps are involved: Obtain transformer bushing partial discharge signal and extract original features; Perform dimensionality reduction processing on the extracted original features to obtain the features after dimensionality reduction; Construct a radial basis function SVM model and optimize the parameters of the radial basis function SVM model through the gorilla troop optimization GTO algorithm; The features after dimensionality reduction are input into the radial basis function SVM model with optimized parameters for classification to obtain the classification and recognition results.

2. The transformer bushing partial discharge pattern recognition method according to claim 1, characterized in that: The transformer bushing partial discharge signal is obtained and the original features are extracted, wherein the original features include time domain features and frequency domain features: The time domain characteristics include absolute mean, root mean square value, variance, rise time, fall time, pulse width, skewness and kurtosis; The frequency domain features include absolute mean, root mean square value, variance, maximum maximum corresponding amplitude, maximum maximum corresponding frequency, spectrum energy, skewness and kurtosis.

3. The transformer bushing partial discharge pattern recognition method according to claim 1, characterized in that: The kernel principal component analysis (KPCA) method is used to reduce the dimension of the extracted original features, including: Map the extracted original features into high-dimensional feature space using nonlinear mapping relationships; Compute the covariance matrix in high-dimensional feature space; Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and eigenvectors; The cumulative contribution rate of each eigenvector is calculated according to the eigenvalue, and the eigenvectors with cumulative contribution rates higher than the set value are selected as principal components. The original features are projected onto the selected principal components to achieve dimensionality reduction.

4. The transformer bushing partial discharge pattern recognition method according to claim 1, characterized in that: Construct a radial basis function SVM model and express the optimization problem of RBF SVM as: Where w is the normal vector of the decision plane, b is the bias term, ξ i is the slack variable, y i is the label of the i-th sample, is the feature vector after dimensionality reduction by KPCA; C is the penalty parameter; n is the number of samples.

5. The transformer bushing partial discharge pattern recognition method according to claim 1, characterized in that: The parameters of the radial basis function SVM model are optimized by using the Gorilla Force Optimization GTO algorithm, including: Initialization parameters: Set the gorilla population size N and the maximum number of iterations Itermax, set the upper and lower bounds of the SVM model penalty parameter C and the RBF kernel parameter g, and obtain the initial population. i , g i ) indicates the position of the gorilla; Determine the fitness function; Update the position of the gorilla and calculate the fitness function. Iterate the optimal solution of the problem according to the fitness value of the new population. If the new fitness value exceeds the previous fitness value, set the current fitness value to the new fitness value. When the number of iterations is greater than Itermax, C and g corresponding to the optimal solution are obtained as parameters of the SVM model.

6. The transformer bushing partial discharge pattern recognition method according to claim 5, characterized in that: Updates to the gorilla's location include: Entering the exploration phase, simulating the population migration behavior of gorillas and updating the individual positions of gorillas, the population migration behavior includes three migration positions, namely known position, unknown position and companion position; Enter the development phase and obtain the parameter C of the exploration phase: If the parameter C is greater than or equal to the given value, the population is used to follow the movement of the silverback gorilla to update the migration position; If the parameter C is less than a given value, the migration position is updated using the male gorilla's courtship behavior.

7. The transformer bushing partial discharge pattern recognition method according to claim 1, characterized in that: The reduced-dimensional features are input into the parameter-optimized radial basis function SVM model for classification to obtain the classification and recognition results. The specific formula is: Among them, f(x new ) is the predicted category label; α i and is the Lagrange multiplier, k(x i , x new ) is the kernel function RBF, b is the bias term, x i is the original feature obtained, x new is the new sample point; If f(x new )>0, then x new are classified as positive, otherwise they are classified as negative.

8. Transformer bushing partial discharge pattern recognition system, characterized in that: include: The feature extraction module is configured to: obtain the transformer bushing partial discharge signal and perform original feature extraction; The dimension reduction module is configured to: perform dimension reduction processing on the extracted original features to obtain features after dimension reduction; The model optimization module is configured to: construct a radial basis function SVM model, and optimize the parameters of the radial basis function SVM model by using the gorilla troop optimization GTO algorithm; The classification module is configured to: input the dimension-reduced features into a parameter-optimized radial basis function SVM model for classification to obtain a classification recognition result.

9. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps in the transformer bushing partial discharge pattern recognition method according to any one of claims 1 to 7 are implemented.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the transformer bushing partial discharge pattern recognition method according to any one of claims 1 to 7 are implemented.

Images