Transformer fault diagnosis method based on ICDO-SVM

Abstract

The invention provides a transformer fault diagnosis method based on ICDO-SVM, and the method comprises the steps: optimizing a support vector machine based on an improved cloud drift optimization algorithm, dividing a transformer dissolved gas data set into a training set and a test set, and carrying out the sample expansion of the gas data of the training set through employing a smote algorithm; then, Bernoulli chaotic mapping and an adaptive t distribution algorithm are fused into a cloud drift algorithm, key parameters of the SVM are optimized by using ICDO, and an ICDO-SVM transformer fault diagnosis model is established; and finally, taking the expanded DGA data as an input set of the model, and verifying the transformer fault diagnosis model. According to the transformer fault diagnosis method, the defects of a traditional CDO algorithm are overcome by adopting a multi-strategy improved cloud drift algorithm, the convergence speed and precision are further improved, and an efficient and high-precision transformer fault diagnosis model can be formed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of transformer fault diagnosis, and in particular to a transformer fault diagnosis method based on ICDO-SVM. BACKGROUND

[0002] As an important electrical equipment in the power system, the power transformer undertakes the important role of step-up or step-down. Once the transformer fails, it will seriously threaten the normal operation of the power system. Therefore, improving the fault diagnosis accuracy of the transformer has great significance for the stability and reliability of the operation of the power system.

[0003] At present, dissolved gas analysis in oil has become a feasible method widely used in transformer fault diagnosis in the power industry. According to the relevant principles of dissolved gas analysis (DGA), a series of methods for diagnosing transformer faults have been proposed by many scholars, such as code-free ratio, IEC three-ratio method, improved three-ratio method, etc. Although these ratio discrimination methods are simple, they often result in low fault diagnosis accuracy due to the absence of coding. Combined with mathematical models and theories, some scholars have proposed transformer fault diagnosis methods based on grey correlation theory, fuzzy theory, rough set theory, and Bayesian theory. However, these methods generally have the disadvantage of low classification accuracy.

[0004] In recent years, with the development of artificial intelligence, a series of intelligent diagnosis methods using dissolved gas concentration values in oil as characteristic quantities have been developed. Some methods use support vector machines (SVM) for fault diagnosis, which can effectively solve common problems such as small sample and non-linear classification. However, the selection of two parameters, i.e., the penalty factor and the kernel function, has a great influence on the classification accuracy of SVM.

[0005] Cloud drift optimization algorithm has the advantages of few parameters and fast global search speed, and can be used for optimization of SVM hyperparameters to optimize model performance. Although the DBO algorithm can successfully solve certain engineering problems, it also has certain limitations, such as poor global search ability and premature convergence to local optimal solution.

[0006] To solve these problems, the present application proposes a multi-strategy improved cloud drift algorithm to compensate for the shortcomings of the traditional CDO algorithm, thereby improving the convergence speed and accuracy.

[0007] In summary, the application proposes a transformer fault diagnosis model based on ICDO optimized SVM. First, the traditional CDO algorithm is improved by multiple strategies;Then the CDO is used to optimize the parameters of the SVM model;Finally, the fault characteristics of the transformer are extracted and classified by the SVM classifier, and finally a high-efficiency and high-precision transformer fault diagnosis model is formed. SUMMARY

[0008] The application proposes a transformer fault diagnosis method based on ICDO-SVM, which uses a multi-strategy improved cloud drift algorithm to make up for the shortcomings of the traditional CDO algorithm, thereby improving the convergence speed and accuracy, and forming a high-efficiency and high-precision transformer fault diagnosis model.

[0009] The application adopts the following technical solutions.

[0010] A transformer fault diagnosis method based on ICDO-SVM, the diagnosis method is based on improved cloud drift optimization algorithm (Clouddrift optimization CDO) to optimize support vector machine, first, the transformer dissolved gas data set is divided into training set and test set, smote algorithm is used to expand the sample of training set gas data;Then Bernoulli chaotic mapping and adaptive t distribution algorithm are integrated into the cloud drift algorithm, the key parameters of SVM are optimized by ICDO, and the ICDO-SVM transformer fault diagnosis model is established;Finally, the expanded DGA data is used as the input set of the model, and the transformer fault diagnosis model is verified.

[0011] The diagnosis method comprises the following steps:

[0012] Step S1, based on the collected transformer unbalanced small sample oil dissolved gas fault data, the data is normalized, and then the normalized small sample data is expanded by SMOTE sample data, so that the number of samples of each fault category is consistent;

[0013] Step S2, an improved cloud drift optimization algorithm is constructed, Bernoulli chaotic mapping and adaptive t distribution algorithm are integrated into the cloud drift algorithm to optimize the cloud particle population position updating strategy;

[0014] Step S3, the improved cloud drift algorithm is used, the diagnostic accuracy rate of SVM is reduced by 1, the hyperparameters of SVM are optimized, and the transformer fault diagnosis model is established;

[0015] In step S1, for the selected power company, the collected oil gas fault sample data includes: hydrogen, methane, ethane, ethylene and acetylene gas content data.

[0016] In step S1, the oil-immersed transformer state corresponding to the oil-gas fault sample data in the oil includes: low-temperature heat release, high-temperature heat release, low-energy discharge, high-energy discharge, partial discharge and normal oil-immersed transformer state.

[0017] In step S1, the oil-immersed transformer state is defined according to DL / T722-2014 "Transformer Oil Dissolved Gas Analysis and Judgment Standard", that is, according to whether the transformer has a fault and the fault type, the transformer is divided into six fault states.

[0018] Step S1 includes the following steps:

[0019] Step i) normalizing the oil gas;

[0020] Step ii) selecting a few samples: selecting sample x i as the benchmark for generating new samples, the Euclidean distance between the sample and the remaining fault samples is calculated as the standard;

[0021] Step iii) neighbor sample selection: according to the sample imbalance ratio, set the sampling rate N, for each benchmark sample x i , randomly select N samples x ij

[0022] Step iv) new sample synthesis: generate new samples according to the following formula:

[0023] X new = x i +r×(x ij -x i ) (1)

[0024] In the formula: x i is the i-th sample in the minority class; x ij is the j-th neighbor sample of sample x i , j = 1, 2, …, k; X new is the new sample synthesized between sample x ij and x i ; r is a random number between [0, 1].

[0025] The cloud drift optimization algorithm is an innovative natural-inspired meta-heuristic method to solve complex optimization problems. The cloud drift optimization (CDO) algorithm mimics the dynamic behavior of cloud particles affected by atmospheric forces, achieving a delicate balance between exploration and exploitation. It has an adaptive weight adjustment mechanism that can change the real-time drifting behavior of the cloud, thus efficiently navigating the search space. Using a cloud-based drift strategy, the number of digits uses probabilistic motion to more effectively maneuver in the optimization environment.

[0026] In step two, the cloud drift optimization algorithm starts with N candidate solutions (clouds) initialized by Bernoulli chaos mapping in the search space, ensuring unbiased exploration of the entire space, specifically:

[0027]

[0028] Where U(0,1) represents a uniform random distribution in the range [0,1]; this step simulates the random initial distribution of cloud particles in the sky, where the initial position is unknown and dispersed;

[0029] Each cloud or particle is assigned a dynamic weight according to its fitness value relative to other particles in the group; the weight determines the degree to which the particle is influenced by the positions of other particles;

[0030] Each cloud or particle is assigned a dynamic weight w i,j ; the weight determines the degree to which the particle is influenced by the positions of other particles;

[0031] The weight of each cloud is updated based on the fitness function:

[0032]

[0033] Where f * is the best fitness value found so far. f max is the maximum fitness value; this weight adaptation allows the algorithm to focus on particles with lower fitness, allowing them to explore new areas.

[0034] The values 0.3 and 0.7 in the adaptive weight update formula are determined based on extensive empirical adjustments. They allow the scaling factor to vary controllably within the range [0.3, 1.0], ensuring effective diversity while adaptively balancing exploration and exploitation dynamics;

[0035] The cloud is randomly reinitialized during the process by calculating the fitness to prevent premature convergence and introduce exploration in the later stages of the algorithm; it occurs with a probability z that decreases as the algorithm progresses:

[0036]

[0037] Where t is the current iteration number and T is the maximum iteration number; by randomizing, it simulates unpredictable changes in atmospheric conditions to change the path of the drifting cloud, ensuring that the search continues even in the final stages; when the random number is less than z, an adaptive t-distribution is used for random disturbance, with boundary handling as follows

[0038]

[0039] The position is updated using exploration or exploitation when the random number is greater than z;

[0040] In the exploitation phase, the cloud moves towards the best solution found so far (i.e. the global best position);

[0041] The position update equation is:

[0042]

[0043] where X * (j) is the best known solution at iteration t; v b (j) ~ U(-0.2a, 0.2a) represents a small random adjustment affected by a factor a = atanh(-t / T + 1) that controls the degree of exploitation.

[0044] The cloud gradually refines its shape after the initial movement by refining its behavior, adjusting its position according to the best nearby solution;

[0045] In the exploration phase, the algorithm ensures the diversity of the search by allowing the cloud to move randomly within the search space, which is expressed in the formula as:

[0046]

[0047] where v c (j) ~ U(-0.2b, 0.2b) is another random factor that affects the movement of the particle, b = 1 - t / T, whose parameter decreases over time, causing the algorithm to shift from exploration to exploitation.

[0048] In step S2, the cloud particle population position update strategy is optimized by incorporating the Bernoulli chaotic mapping and adaptive t-distribution algorithm into the cloud drift algorithm;

[0049] In the initialization phase, the traditional cloud drift algorithm uses a random number generation method to initialize the population position, which will result in uneven initial cloud positions and cannot traverse all positions in the environment, resulting in poor optimization effect and low convergence speed;

[0050] Bernoulli chaotic mapping is introduced in the population initialization process, and its expression is

[0051]

[0052] In the formula: U t is the current value of the chaotic sequence at the tth iteration;

[0053] ρ is the control coefficient, and when ρ = 0.5, the ergodicity is best;

[0054] When the randomization simulates unpredictable changes in atmospheric conditions, the cloud particles are relocated, and the adaptive t-distributed perturbation is performed on the perturbed behavior; the t-distributed perturbation makes the algorithm jump out of the local optimum, and improves the optimization accuracy of the algorithm, and the probability density function thereof is:

[0055]

[0056] When the degree of freedom m = 1, the t-distribution is Cauchy distribution; when the degree of freedom m tends to infinity, the t-distribution is similar to Gaussian distribution

[0057]

[0058] wherein: represents the position of the optimal solution after adaptive t-distributed variation perturbation; represents the position of the optimal solution before adaptive t-distributed variation perturbation;

[0059] R(m) represents a random number obtained by sampling from the t-distribution with degree of freedom m

[0060]

[0061] At the beginning of the algorithm, the iteration number is small, and the t-distributed variation is similar to the Cauchy variation, so that the algorithm has good global search ability;

[0062] As the number of iterations increases, the t-distribution gradually tends to be Gaussian distribution, which can help the algorithm converge quickly.

[0063] In step S3, the support vector machine is a machine learning algorithm commonly used to solve linear binary classification problems, and it performs well in avoiding overfitting and processing high-dimensional data. In processing nonlinear classification tasks such as transformer fault diagnosis, the support vector machine improves its performance by introducing kernel functions and soft intervals. When solving classification problems, the optimal hyperplane problem of SVM is expressed as a maximum interval classification problem:

[0064]

[0065] In the formula, ω is the normal vector of the hyperplane, c is the penalty factor, ξ i is the relaxation variable, y i is the corresponding fault category, i x is the feature function for mapping x i to a high-dimensional feature space, b is the bias, and l is the number of samples.

[0066] Using dual optimization and Lagrange multiplier method, formula (2) can be converted to formula (3):

[0067]

[0068] wherein: a i is the Lagrange multiplier. The classification effect of SVM is highly dependent on the selection of kernel function. The radial basis kernel function performs outstandingly in classification problems, and its expression is:

[0069] K(x i ,x j )=exp(-g||x i -x j ||)g≥0 (12)

[0070] wherein: g is the kernel function parameter. Finally, the decision function is obtained as:

[0071]

[0072] The classification effect of support vector machine is closely related to the penalty factor c and the kernel function parameter g, and the values of the two parameters are very important for the accuracy of transformer fault diagnosis;

[0073] In step S3, the ICDO algorithm is selected to optimize the two parameters c and g of SVM, so as to improve the diagnosis accuracy.

[0074] In order to solve the problem that the accuracy of the transformer fault diagnosis model is not high enough, the present application provides a transformer fault diagnosis model based on ICDO-SVM, and the present application has the following advantages:

[0075] i) By introducing Bernoulli chaotic mapping and adaptive t-distribution algorithm, the CDO optimization algorithm is improved, so that the convergence performance and local search ability of the algorithm can be effectively improved, and the ICDO has better optimization performance.

[0076] ii) The ICDO is used to optimize the kernel function parameters g and c of SVM, so that the generalization ability and classification accuracy of SVM can be effectively improved. According to the fault diagnosis and stability comparison experiments, the comprehensive accuracy of the ICDO-SVM model for fault diagnosis can reach 93.07%, which is increased by 7.92%, 4.46% and 0.99% respectively compared with the accuracy of SVM, PSO-SVM and CDO-SVM models. The ICDO-SVM model has higher accuracy and stability when used for transformer fault diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0077] The present application will be further described in detail below in combination with the drawings and specific embodiments:

[0078] Figure 1 is a flowchart of a transformer fault diagnosis method based on ICDO-SVM according to an embodiment of the present application; Figure 1 Figure 1 is a flowchart of a transformer fault diagnosis method based on ICDO-SVM according to an embodiment of the present application;

[0079] Figure 2 ​A schematic diagram of an ICDO-SVM result confusion matrix in an embodiment of the present application. DETAILED DESCRIPTION

[0080] The present application is further described below in conjunction with the accompanying drawings and embodiments.

[0081] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0082] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments in accordance with the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, steps, operations, elements, components, and / or groups thereof, but do not preclude the presence or addition of one or more other features, steps, operations, elements, components, and / or groups thereof.

[0083] An ICDO-SVM-based transformer fault diagnosis method, the diagnosis method is based on an improved cloud drift optimization algorithm (Clouddrift optimization CDO) to optimize a support vector machine. First, a transformer dissolved gas data set is divided into a training set and a test set, and a smote algorithm is used to expand the sample of the training set gas data. Then, Bernoulli chaotic mapping and adaptive t-distribution algorithm are integrated into the cloud drift algorithm, and the key parameters of SVM are optimized by ICDO to establish an ICDO-SVM transformer fault diagnosis model. Finally, the expanded DGA data is used as the input set of the model to verify the transformer fault diagnosis model.

[0084] As shown in Figure 1 the diagnosis method comprises the following steps:

[0085] Step S1, based on the collected transformer unbalanced small sample oil dissolved gas fault data, the data is normalized, and then the normalized small sample data is expanded by SMOTE sample data, so that the number of samples of each fault category is consistent;

[0086] Step S2, an improved cloud drift optimization algorithm is constructed, and Bernoulli chaotic mapping and adaptive t-distribution algorithm are integrated into the cloud drift algorithm to optimize the cloud particle population position updating strategy;

[0087] Step S3, the improved cloud drift algorithm is used to optimize the hyperparameters of SVM with the diagnostic accuracy rate of SVM minus 1 as the fitness value, and a transformer fault diagnosis model is established;

[0088] In step S1, the collected oil gas fault sample data of the selected power company includes gas content data of hydrogen, methane, ethane, ethylene and acetylene.

[0089] In step S1, the oil-immersed transformer state corresponding to the oil gas fault sample data includes low-temperature heat release, high-temperature heat release, low-energy discharge, high-energy discharge, partial discharge and normal oil-immersed transformer state.

[0090] In step S1, the oil-immersed transformer state is defined according to DL / T722-2014 "Transformer Oil Dissolved Gas Analysis and Judgment Standard", i.e. according to whether the transformer has a fault and the fault type, the transformer is divided into six fault states.

[0091] Step S1 includes the following steps:

[0092] Step i) normalizing the oil gas;

[0093] Step ii) selecting a few class samples: selecting a sample x i As a benchmark for generating new samples, the Euclidean distance between the sample and the remaining fault samples is calculated as the standard;

[0094] Step iii) neighbor sample selection: according to the sample imbalance ratio, set the sampling rate N, for each benchmark sample x i , randomly select N samples x ij

[0095] Step iv) new sample synthesis: generate a new sample according to the following formula:

[0096] X new = x i +r×(x ij -x i ) (1) In the formula: x i is the i-th sample in the minority class; x ij is the j-th neighbor sample of sample x i , j = 1, 2, …, k; X new is a new sample synthesized by interpolating between samples x ij and x i ; r is a random number between 0 and 1.

[0097] The cloud drift optimization algorithm, an innovative nature-inspired metaheuristic approach to solve complex optimization problems. Cloud drift optimization (CDO) mimics the dynamic behavior of cloud particles influenced by atmospheric forces, striking a delicate balance between exploration and exploitation. It has an adaptive weight adjustment mechanism that can change the real-time drift behavior of clouds, thus navigating the search space efficiently. Using a cloud-based drift strategy, the number of bits utilizes probabilistic movement to maneuver more effectively in the optimization environment.

[0098] In step two, the cloud drift optimization algorithm starts with N candidate solutions (clouds) initialized in the search space by Bernoulli chaotic mapping, ensuring unbiased exploration of the entire space, specifically:

[0099]

[0100] Where U(0,1) represents a uniform random distribution over the range [0,1]; this step simulates the random initial distribution of cloud particles in the sky, where the initial positions are unknown and dispersed;

[0101] Each cloud or particle is assigned a dynamic weight based on its fitness value relative to other particles in the population; the weight determines the extent to which a particle is influenced by the positions of other particles;

[0102] Each cloud or particle is assigned a dynamic weight w i,j based on its fitness value relative to other particles in the population; the weight determines the extent to which a particle is influenced by the positions of other particles;

[0103] The weight of each cloud is updated based on the fitness function:

[0104]

[0105] Where f * is the best fitness value found so far. f max is the maximum fitness value; this weight adaptation allows the algorithm to focus on lower fitness particles, allowing lower fitness particles to explore new areas.

[0106] The values 0.3 and 0.7 in the adaptive weight update formula are determined based on extensive empirical adjustments. They make the scaling factor controllably vary in the range [0.3, 1.0], ensuring adaptive balance between exploration and exploitation dynamics while ensuring effective diversity;

[0107] The cloud random re-initialization process calculates the fitness to prevent premature convergence and introduce exploration in the later stages of the algorithm; it occurs with probability z, which decreases as the algorithm progresses:

[0108]

[0109] where t is the current iteration number, T is the maximum iteration number, the randomization simulates unpredictable changes in atmospheric conditions to change the path of the drifting cloud, ensuring that the search continues even in the last stage; when the random number is less than z, a random disturbance is used using adaptive t-distribution, the position is boundary processing, the formula is as follows

[0110]

[0111] When the random number is greater than z, the position is updated using exploration or development;

[0112] In the development stage, the cloud moves to the best solution found so far (i.e. the global best position);

[0113] The position update equation is:

[0114]

[0115] where X * (j) is the best known solution at iteration t; v b (j) ~ U(-0.2a, 0.2a) represents a small random adjustment affected by the factor a = atanh(-t / T + 1), which controls the degree of development.

[0116] By refining the behavior, the cloud simulates how it gradually refines its shape after the initial movement, and adjusts its position according to the best nearby solution;

[0117] In the exploration stage, the algorithm ensures the diversity of the search by allowing the cloud to move randomly within the search space, which is expressed in the formula as:

[0118]

[0119] where v c (j) ~ U(-0.2b, 0.2b) is another random factor affecting the motion of the particle, b = 1-t / T, the parameter decreases with time, making the algorithm shift from exploration to development.

[0120] In step S2, the Bernoulli chaotic mapping and adaptive t-distribution algorithm is integrated into the cloud drift algorithm to optimize the cloud particle population position update strategy;

[0121] In the initialization stage, the traditional cloud drift algorithm generates random numbers when initializing the population position, which will result in uneven initial cloud positions and cannot traverse all positions in the environment, resulting in poor optimization effect and low convergence speed;

[0122] Bernoulli chaotic mapping is introduced in the population initialization process, and its expression is

[0123]

[0124] In the formula: U t is the current value of the chaotic sequence at the tth iteration;

[0125] ρ is a control coefficient, when ρ = 0.5, the ergodicity is the best;

[0126] When the randomization simulates the unpredictable changes in atmospheric conditions, the cloud particles are relocated, and the adaptive t-distributed disturbance is performed on the disturbance behavior; the t-distributed disturbance makes the algorithm jump out of the local optimum, and improves the optimization accuracy of the algorithm, and its probability density function is:

[0127]

[0128] When the degree of freedom m = 1, the t-distribution is Cauchy distribution; when the degree of freedom m tends to infinity, the t-distribution is similar to Gaussian distribution

[0129]

[0130] Wherein: represents the position of the optimal solution after adaptive t-distributed variation disturbance; represents the position of the optimal solution before adaptive t-distributed variation disturbance;

[0131] R(m) represents a random number obtained by sampling from the t-distribution with degree of freedom m

[0132]

[0133] When the algorithm starts, the iteration number is small, the t-distributed variation is similar to the Cauchy variation, so that the algorithm has good global search ability;

[0134] As the number of iterations increases, the t-distribution gradually tends to be Gaussian distribution, which can help the algorithm converge quickly.

[0135] In step S3, the support vector machine is a machine learning algorithm commonly used to solve linear binary classification problems, and it performs well in avoiding overfitting and processing high-dimensional data. In processing nonlinear classification tasks such as transformer fault diagnosis, support vector machine improves its performance through the introduction of kernel function and soft interval. When solving classification problems, the optimal hyperplane problem of SVM is expressed as a maximum interval classification problem:

[0136]

[0137] In the formula: ω is the hyperplane normal vector, c is the penalty factor, ξ i is the relaxation variable, y i is x i corresponding fault category, To map x i to high-dimensional feature space, b is bias, and l is the number of samples. Using dual optimization and Lagrange multiplier method, formula (2) can be converted to formula (3):

[0138]

[0139] In the formula: α i is the Lagrange multiplier. The classification effect of SVM is highly dependent on the selection of kernel function. The radial basis kernel function performs outstandingly in classification problems, and its expression is:

[0140] K(x i ,x j )=exp(-g||x i -x j ||)g≥0 (12)

[0141] In the formula: g is the kernel function parameter. Finally, the decision function is obtained as:

[0142]

[0143] The classification effect of support vector machine is closely related to the penalty factor c and the kernel function parameter g. The values of these two parameters are very important for the accuracy of transformer fault diagnosis;

[0144] In step S3, the ICDO algorithm is selected to optimize the two parameters c and g of SVM to improve the diagnosis accuracy.

[0145] Embodiment:

[0146] In this example, an ICDO-SVM fault diagnosis method based on small sample data expansion is proposed, which includes two steps of data expansion and transformer fault diagnosis. First, the original data is divided into training set and test set, and the training set is expanded and sampled by Smote. Then, the penalty factor and kernel function parameter of SVM are optimized by IBKA algorithm, the optimal transformer diagnosis model is constructed, and the test set is diagnosed for fault. The transformer fault diagnosis based on data expansion and ICDO-SVM is shown in the figure.

[0147] According to DL / T722-2014 "Analysis and Judgment Standard of Dissolved Gases in Transformer Oil", the transformer is divided into six categories according to whether the transformer fails and the type of failure. They are represented by labels 1-6, including low-energy discharge (D1), high-energy discharge (D2), low-temperature heat release (T1&T2), high-temperature heat release (T2), partial discharge (PD) and normal (N) Research data from the fault records of 500kV oil-immersed transformers in a certain area of China, including various gases dissolved in the insulating oil of the failed transformer (H2, CH4, C2H4, C2H6, C2H2); At the same time, the state category of the transformer can be divided into 6 kinds, and the sample expansion table is shown in Table 1.

[0148] Research data from the fault records of 500kV oil-immersed transformers in a certain area of China, including various gases dissolved in the insulating oil of the failed transformer (H2, CH4, C2H4, C2H6, C2H2); At the same time, the state category of the transformer can be divided into 6 kinds, and the sample expansion table is shown in Table 1.

[0149] Tab.1 Comparison of samples before and after augmentation

[0150]

[0151] Before the training set is augmented, the accuracy of the fault diagnosis test set is 89.62%, and after the training set is augmented, the accuracy of the test set is improved to 93.07%

[0152] Experimental results

[0153] In order to effectively evaluate the performance of the ICDO-SVM model, the accuracy, precision, recall, F1_score and kappa coefficient are introduced for the fault diagnosis results for further description. According to the results of each index shown in Table 2, it can be concluded that the model has strong classification ability and high accuracy and reliability in fault diagnosis.

[0154] Tab.2 Index results of different diagnostic models

[0155]

[0156] As shown in Table 2, the fault accuracy of ICDO-SVM is the highest, with an accuracy of 93.07%. In addition, the fault diagnosis accuracy of SVM, PSO-SVM and CDO-SVM is 85.15%, 88.61% and 92.08% respectively, and the fault diagnosis accuracy of ICDO-SVM is improved by 7.92%, 4.46% and 0.99% compared with the above three methods.

Claims

1. A transformer fault diagnosis method based on ICDO-SVM, characterized in that: The diagnostic method is based on an improved cloud drift optimization algorithm to optimize a support vector machine, first divides a transformer dissolved gas data set into a training set and a test set, and uses an smote algorithm to expand the sample of the training set gas data; then Bernoulli chaotic mapping and an adaptive t distribution algorithm are integrated into the cloud drift algorithm, the key parameters of the SVM are optimized by using ICDO, and an ICDO-SVM transformer fault diagnosis model is established; finally, the expanded DGA data is used as the input set of the model to verify the transformer fault diagnosis model.

2. The method of claim 1, wherein the method is based on ICDO-SVM for transformer fault diagnosis. The diagnostic method comprises the following steps: Step S1, based on the collected transformer unbalanced small sample oil dissolved gas fault data, the data is normalized, and then the normalized small sample data is expanded by SMOTE sample data, so that the number of samples of each fault category is consistent; Step S2, an improved cloud drift optimization algorithm is constructed, Bernoulli chaotic mapping and adaptive t distribution algorithm are integrated into the cloud drift algorithm to optimize the cloud particle population position updating strategy; Step S3, the improved cloud drift algorithm is used to optimize the hyperparameters of the SVM, and a transformer fault diagnosis model is established.

3. The method of claim 2, wherein the method is based on ICDO-SVM. In step S1, the collected oil gas fault sample data of the selected power company includes the gas content data of hydrogen, methane, ethane, ethylene and acetylene.

4. The method of claim 3, wherein the method is based on ICDO-SVM. In step S1, the oil gas fault sample data corresponds to the state of the oil-immersed transformer, including medium-low temperature heat release, high temperature heat release, low energy discharge, high energy discharge, partial discharge and normal oil-immersed transformer state.

5. The ICDO-SVM based transformer fault diagnosis method of claim 4, wherein: In step S1, the state of the oil-immersed transformer is limited according to DL / T722-2014 "Transformer Oil Dissolved Gas Analysis and Judgment Standard".

6. The ICDO-SVM based transformer fault diagnosis method of claim 2, wherein: Step S1 comprises the following steps: Step i) normalizing the oil gas; Step ii) selecting a minority class sample: selecting a sample x in the minority class sample i As a benchmark for generating new samples, the Euclidean distance between the sample and the remaining fault samples is calculated as the standard. Step iii) Near neighbors selection: Set sampling rate N according to the sample imbalance ratio, for each reference sample x i Randomly select N samples x from its k near neighbors ij Step iv) new sample synthesis: new samples are generated according to the following formula: X new = x i + r x (x ij - x i ) (1) where: x i is the i-th sample in the minority class; x ij is the j-th nearest neighbor sample of x i , j = 1, 2,..., k; X new is the new sample synthesized by interpolating between x ij and x i ; r is a random number between [0, 1].

7. The ICDO-SVM based transformer fault diagnosis method of claim 2, wherein: In step two, the cloud drift optimization algorithm starts with N candidate solutions, i.e. N clouds, initialized by Bernoulli chaotic mapping in the search space, to ensure unbiased exploration of the entire space, specifically: Where U(0,1) represents a uniform random distribution in the range [0,1]; the random initial distribution of cloud particles in the sky is simulated, where the initial position is unknown and dispersed; Each cloud or particle is assigned a dynamic weight according to its fitness value relative to other particles in the group; the weight determines the degree to which the particle is affected by the position of other particles; Each cloud or particle is assigned a dynamic weight w according to its fitness value relative to other particles in the population i,j ; the weight determines the degree to which the particle is influenced by the positions of other particles; The weight of each cloud is updated based on the fitness function, which can be expressed as: where f * is the best fitness value found so far. f max is the maximum fitness value; During the random reinitialization of the cloud, the fitness is calculated to prevent premature convergence and introduce exploration in the later stages of the algorithm; Occur with probability z, which decreases as the algorithm progresses: Where t is the current iteration number, T is the maximum iteration number, the path of the drifting cloud is changed by randomization to simulate unpredictable changes in atmospheric conditions, ensuring that the search continues even in the last stage; When the random number is less than z, a random disturbance is performed using the adaptive t distribution, and the position is handled as a boundary, as follows The position is updated using exploration or exploitation when the random number is greater than z; In the development phase of the algorithm, the cloud moves to the best solution found so far, i.e. to the global best position; The position update equation is: where X * (j) is the best known solution at iteration t; v b (j) ~ U(-0.2a, 0.2a) represents a small random adjustment affected by a factor a = atanh(-t / T + 1) that controls the degree of exploitation. The cloud gradually refines its shape after the initial movement by adjusting its position according to the best nearby solution; In the exploration phase of the algorithm, the algorithm ensures the diversity of the search by allowing the cloud to move randomly within the search space, which is expressed as: where v c (j) ~ U(-0.2b, 0.2b) is another random factor affecting the motion of particles, b = 1 - t / T, whose parameter decreases with time, making the algorithm turn from exploration to exploitation.

8. The ICDO-SVM based transformer fault diagnosis method of claim 2, wherein: In step S2, in the initialization phase, the Bernoulli chaotic mapping is introduced in the population initialization process, and its expression is where: U t is the current value of the chaotic sequence at the tth iteration; ρ is the control coefficient, and when ρ = 0.5, the ergodicity is the best; When the randomization simulates the unpredictable changes in atmospheric conditions, the cloud particles are repositioned, and the perturbation behavior is adaptively disturbed by the t distribution; the t distribution perturbation makes the algorithm jump out of the local optimum, and improves the optimization accuracy of the algorithm, and its probability density function is: When the degree of freedom m = 1, the t distribution is Cauchy distribution; When the degree of freedom m tends to infinity, the t distribution is similar to the Gaussian distribution, which is expressed as wherein: represents the position of the optimal solution after adaptive t-distributed variant disturbances; represents the position of the optimal solution before adaptive t-distributed variant disturbances; R(m) represents the random number obtained by sampling from the t distribution with degree of freedom m As the number of iterations increases, the t distribution gradually tends to the Gaussian distribution, which can help the algorithm converge quickly.

9. The ICDO-SVM based transformer fault diagnosis method of claim 2, wherein: In step S3, when dealing with the nonlinear classification task of transformer fault diagnosis, the support vector machine introduces the kernel function and soft interval, and the optimal hyperplane problem of SVM in solving the classification problem is expressed as the maximum interval classification problem: where ω is the hyperplane normal vector, c is the penalty factor, ξ i is the slack variable, y i is x i corresponding to the failure class, is the feature function that maps x i to a high-dimensional feature space, b is the bias, and l is the number of samples. Using dual optimization and the Lagrange multiplier method, formula 2 can be converted to formula 3: wherein: a i is the Lagrange multiplier; the classification effect of SVM depends on the selection of kernel function. The expression of radial basis kernel function is: K(x i ,x j ) = exp(-g||x i -x j ||) g > 0 (12) In the formula: g is the kernel function parameter, and the decision function is finally solved as:

10. The ICDO-SVM based transformer fault diagnosis method of claim 9, wherein: In step S3, the ICDO algorithm is selected to optimize the two parameters c and g of SVM, so as to improve the diagnosis accuracy.

Images