Transformer fault diagnosis method based on data reconstruction and IBKA-SVM

Abstract

The invention provides a transformer fault diagnosis method based on data reconstruction and IBKA-SVM, and is applied to the field of transformer fault diagnosis. Firstly, boundary synthesis minority class samples are combined with a Gaussian mixture model to carry out mixed sampling processing on a training set, and a kernel principal component analysis method is further used to carry out feature fusion and feature extraction on a constructed high-dimensional data set; secondly, chaotic mapping is introduced to improve the initialization position of a black-wing plinary algorithm, and a lens imaging reverse learning strategy is introduced into the black-wing plinary algorithm, so that the capability of jumping out of local optimum is improved; according to the method, the superiority of an improved black wing algorithm (IBKA) is verified through two test functions, and the advantages of the improved IBKA are verified through the two test functions; and then optimizing SVM model parameters by using an IBKA algorithm based on the characteristics of the data reconstruction, and constructing a transformer diagnosis model of the IBKA-SVM based on the data reconstruction. Compared with the prior art, the precision is higher, and the stability is higher.

Description

Technical Field

[0001] The present invention relates to the field of transformer fault diagnosis, and in particular to a transformer fault diagnosis method based on data reconstruction and IBKA-SVM. Background Art

[0002] In recent years, with the development of power systems and the continuous expansion of their operation scale, oil-immersed transformers, as key equipment in power systems, have a direct impact on the safety and stability of power systems due to their operating status. Therefore, early diagnosis of transformer faults is particularly important.

[0003] At present, dissolved gas analysis (DGA) in insulating oil is the most widely used method for transformer fault diagnosis. The gas content dissolved in oil has a potential connection with transformer faults. In order to effectively utilize this relationship, researchers have summarized some methods. At present, transformer fault diagnosis methods are mainly divided into traditional methods and artificial intelligence methods. Traditional methods mainly include IEC three-ratio method, Doernenburg method, ICUSU method, etc. These methods have defects such as missing coding and too absolute boundaries in actual use, resulting in poor diagnostic results. Artificial intelligence methods mainly include neural networks, support vector machines, random forests (RF), extreme learning machines (ELM), etc. They have been applied to the field of transformer fault diagnosis and have achieved good results. However, artificial intelligence methods also have their own problems. On the one hand, traditional artificial intelligence methods have poor feature extraction capabilities for original DGA data and high computational costs. On the other hand, when artificial intelligence methods face unbalanced data sets, the diagnostic results often tend to favor majority class samples, resulting in low model accuracy and other problems. Scholar Luo Wenxuan used capsule neural networks to effectively improve the accuracy of fault diagnosis, but the computational complexity is high and it is more suitable for large sample data. Scholar Bao Jinshan used algorithms to optimize SVM parameters to achieve good diagnostic results, with low computational cost and high diagnostic accuracy. Scholar Xie Guomin balanced the initial unbalanced samples, improving the diagnostic effect of unbalanced sample sets, but did not balance only the training set, which was slightly different from the actual test situation. Scholar LIU G used ADASYN combined with the quartering method to process unbalanced samples, improving the balance of samples, but its complex regions and classification boundaries have certain limitations, and it is easy to generate noise samples that reduce the generalization ability of the model. It is proposed to use the kernel principal component analysis method to optimize high-dimensional features as feature input, but the kernel principal component analysis method will lose information in small variance data, resulting in the loss of some important information. Summary of the invention

[0004] In view of the above problems, the present invention proposes a transformer fault diagnosis method based on IBKA-SVM with data reconstruction. First, mixed sampling and balanced samples are performed on the diagnostic model training set Borderline-Smote-GMM; secondly, feature expansion is performed on the data to strengthen the connection between the data and the fault type, and feature fusion is performed through the kernel principal component analysis method (KPCA) to reduce the feature dimension; then the Black Kite algorithm is improved by combining the Logistic chaos mapping strategy with the lens direction learning strategy, which improves the stability of the Black Kite algorithm and its ability to jump out of the local optimum; then the samples reconstructed by the data are input and the parameters of the SVM are optimized using the improved Black Kite algorithm to construct the IBKA-SVM transformer fault diagnosis model, which effectively improves the reliability and generalization ability of the model.

[0005] The present invention is implemented by adopting the following technical scheme: a transformer fault diagnosis method based on IBKA-SVM of data reconstruction is implemented by the following steps:

[0006] S1: Collect publicly available dissolved gas analysis datasets with labeled transformer fault types;

[0007] S2: The collected original data set is divided into a training set and a test set in a ratio of 7:3;

[0008] S3: Build a data reconstruction model and perform Borderline-Smote-GMM mixed sampling to balance the data samples of the divided training set. Perform feature reconstruction on the test set and the balanced training set. The upper layer includes a feature expansion module and a Z-score standardization module. The lower layer uses the kernel principal component analysis method to perform feature fusion and dimensionality reduction.

[0009] S4: The traditional black kite algorithm is hybridly improved by combining the Logistic chaotic mapping strategy with the lens imaging reverse learning strategy.

[0010] S5: Use the improved black-winged kite algorithm IBKA to optimize the best parameter combination of the SVM model penalty factor c and kernel function parameter g;

[0011] S6: Input the low-dimensional sensitive training set after data reconstruction into the optimal parameter combination SVM model for training, and use the trained model for fault diagnosis.

[0012] Furthermore, the step S3 comprises:

[0013] The Borderline-Smote oversampling model divides the minority class samples P into three categories: Safe, Danger, and Noise according to the K nearest neighbor distance, and only generates minority samples for the Danger class. The steps are as follows.

[0014] Step 1: Calculate each sample P in the minority class sample P according to the K nearest neighbor formula i The Euclidean distance of the nearest K samples, and the minority class samples are divided into three categories according to the K nearest neighbor distance

[0015]

[0016] Where: x i and i are two sample points in n-dimensional space respectively; dist(X,Y) is the Euclidean distance between the two sample points.

[0017] Step 2: Sample point x i There are m nearest neighbors, and among the m nearest neighbors, n sample points belong to the majority class samples.

[0018]

[0019] Step 3: Perform SMOTE interpolation on Danger samples.

[0020] X new =X i +λ×(X j -X i )

[0021] Where: X new is the synthesized new sample; X j Yes X i The minority class sample set among the m nearest neighbors of ; λ takes random values ​​in the range of 0 to 1.

[0022] Preferably, the GMM-based cluster undersampling method can identify the most representative data characteristics of the majority class sample data and perform more reasonable sampling processing based on these characteristics. GMM is a parameterized probability distribution model that divides the majority class samples into multiple Gaussian components, each Gaussian component represents a cluster of clusters, and the following posterior probability is used to calculate the probability that each data point belongs to each component.

[0023]

[0024] Where x is the measured data point; Φ(x|θ k ) represents the Gaussian distribution density of the kth sub-model; α k represents the probability of the kth sub-model being selected; θ k is μ k and Parameter set of μ k and represents the mean and variance of the kth sub-model.

[0025] Preferably, the feature expansion method recommended by IEC and IEEE is used to expand the feature quantity, and the concentration data of five gases, H2, CH4, C2H2, C2H4, and C2H6, are used as initial features to further expand the feature quantity, thereby obtaining 21 features to be selected.

[0026] Preferably, the following formula is used to perform Z-score normalization to put the feature quantities on the same scale.

[0027] Y i =(X i -μ i ) / δ i

[0028] Where: X i is the original value of the i-th reselected feature, Y i Yes X i The value after normalization, μ i is the mean of the ith feature, δ i The standard deviation of the ith feature.

[0029] Preferably, a kernel principal component analysis method is used to perform feature fusion and feature extraction on the high-dimensional features at this time.

[0030] Furthermore, the step S4 comprises:

[0031] The black kite algorithm is introduced, which is inspired by the survival strategy of the black kite. It has the characteristics of strong evolutionary ability, fast search speed and strong optimization ability. Its basic steps are:

[0032] 1) Population initialization

[0033] X i =BK lb +rand(BK ub -BK lb )

[0034] Where: i is an integer between 1 and the maximum population size, BK lb With BK ub represents the lower and upper bounds of the ith black kite in dimension j, and rand is a randomly selected value between [0,1].

[0035] During the initialization process, the black kite selects the individual with the best fitness value as the initial leader in the initial group. The formula is the initial leader X L The formula of .

[0036] f best =min(f(X i ))

[0037] X L=X(find(f best = = f(X i )))

[0038] Where: f best represents the individual with the best fitness value, X L Is the leader of the group.

[0039] 2) Aggressive behavior

[0040] The attack behavior stage belongs to the global search stage, and its mathematical model is:

[0041]

[0042] Where: and They are represented as the position of the ith black kite in the jth dimension at the t and t+1 iterations respectively, r is a random number between 0 and 1, and p is a constant of 0.9.

[0043] 3) Migration behavior

[0044] Below is a mathematical model of the migratory behavior of black-winged kites:

[0045]

[0046] In the formula, represents the score leader of the black-winged kite in the jth dimension at the tth iteration, C(0,1) represents the Cauchy mutation, and F i represents the j-th dimension current position of any black-winged kite in the t-th iteration, F ri The fitness value of any black kite at a random position in the jth dimension in the tth iteration.

[0047] Preferably, the Logistic chaotic mapping strategy described in step S4 is introduced into the population initialization stage, which improves the diversity and global search capability of the initial population. The Logistic chaotic mapping expression is:

[0048] X i+1 =rX i (1-X i )

[0049] Where: X i is the population of generation i, X i+1 is the population of the i+1 band, r is the control parameter, when r∈(3.57,4], the system enters a chaotic state, showing the characteristics of sensitive dependence on initial conditions, and when r is 4, the system is completely chaotic. Take r equal to 4.

[0050] Preferably, the lens imaging reverse learning strategy described in S4 guides the algorithm attack behavior, and generates a lens reverse position with the current optimal position black kite to improve population diversity and expand the search range. Its expression is:

[0051]

[0052] In the formula, let h / h′=K, and transform the above formula to get

[0053]

[0054] Then we get the expression in D-dimensional space

[0055]

[0056] In the above formula, X j The original position of the black kite in D-dimensional space during the migration process, X′ j is the position of the black-winged kite after the D-dimensional space transformation, ub j lb j are the upper and lower bounds of the hyperparameters in the space, respectively.

[0057] Furthermore, the step S5 comprises:

[0058] S501: Introduce logic chaos map to initialize the population;

[0059] S502: Introducing the lens imaging reverse learning strategy;

[0060] S503: The leader of the population attacks;

[0061] S504: Migrate to the black-winged kite with the optimal fitness value;

[0062] S505: Determine whether the constraint conditions are met, and if so, output the optimal hyperparameter combination; if not, return to S503;

[0063] Beneficial Effects

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] 1) The Borderline-Smote-GMM method is used to mix and process unbalanced samples. The key information in the fault data is effectively retained for minority samples and redundant samples are avoided due to oversampling. The majority samples are deleted, which can reduce sample redundancy while retaining the main characteristic structure of the data. At the same time, through this mixed sampling strategy, the representativeness of the samples can be optimized while ensuring data integrity, thereby improving the generalization ability of the model.

[0066] 2) Establish a feature reconstruction model, expand the feature quantity to make the feature more closely related to the result, and use the KPCA method to fuse the features to establish low-dimensional sensitive features, so as to avoid the high-dimensional data that will lead to an increased risk of overfitting, as well as increased data sparsity, and the distance measurement becomes unreliable, resulting in decreased model performance and calculation difficulties.

[0067] 3) The BKA algorithm is improved by mixing the logic chaos strategy with the lens reverse learning strategy. The improved IBKA algorithm overcomes the shortcomings of the BKA algorithm, such as poor initial position and easy falling into local optimality. It can find the optimal hyperparameter combination faster and more stably, and effectively improve the diagnosis accuracy and efficiency of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 is a flow chart of the method of the present invention;

[0069] Figure 2 Schematic diagram of feature expansion method;

[0070] Figure 3 Pareto chart for kernel principal component analysis;

[0071] Figure 4 Schematic diagram of lens reverse learning;

[0072] Figure 5 This is a performance comparison chart of the hybrid improved black kite algorithm IBKA and other algorithms; DETAILED DESCRIPTION

[0073] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0074] The present invention is achieved by adopting the following technical solutions:

[0075] Figure 1 A flow chart of a transformer fault diagnosis method based on IBKA-SVM of data reconstruction is shown in the present application, referring to Figure 1 , which is achieved by the following steps:

[0076] S1: The collected publicly available dissolved gas analysis dataset with labeled transformer fault types is used as the original dataset;

[0077] S2: The collected original data set is divided into a training set and a test set in a ratio of 7:3;

[0078] S3: Build a data reconstruction model and perform Borderline-Smote-GMM mixed sampling to balance the data samples of the divided training set. Perform feature reconstruction on the test set and the balanced training set. The upper layer includes a feature expansion module and a Z-score standardization module. The lower layer uses the kernel principal component analysis method to perform feature fusion and dimensionality reduction.

[0079] In the S3:

[0080] The Borderline-Smote oversampling model divides the minority class samples P into three categories: Safe, Danger, and Noise according to the K nearest neighbor distance, and only generates minority samples for the Danger class. The steps are as follows.

[0081] Step 1: Calculate each sample P in the minority class sample P according to the K nearest neighbor formula i The Euclidean distance of the nearest K samples, and the minority class samples are divided into three categories according to the K nearest neighbor distance

[0082]

[0083] Where: x i and i are two sample points in n-dimensional space respectively; dist(X,Y) is the Euclidean distance between the two sample points.

[0084] Step 2: Sample point x i There are m nearest neighbors, and among the m nearest neighbors, n sample points belong to the majority class samples.

[0085]

[0086] Step 3: Perform SMOTE interpolation on Danger samples.

[0087] X new =X i +λ×(X j -X i )

[0088] Where: X new is the synthesized new sample; X j Yes X i The minority class sample set among the m nearest neighbors of ; λ takes random values ​​in the range of 0 to 1.

[0089] Preferably, the GMM-based cluster undersampling method can identify the most representative data characteristics of the majority class sample data and perform more reasonable sampling processing based on these characteristics. GMM is a parameterized probability distribution model that divides the majority class samples into multiple Gaussian components, each Gaussian component represents a cluster of clusters, and the following posterior probability is used to calculate the probability that each data point belongs to each component.

[0090]

[0091] Where x is the measured data point; Φ(x|θ k ) represents the Gaussian distribution density of the kth sub-model; α k represents the probability of the kth sub-model being selected; θ k is μ k and The parameter set of μ k and represents the mean and variance of the kth sub-model.

[0092] Preferably, Figure 2 As shown in the figure, the feature expansion method is used to expand the feature quantity, and the five gas concentration data of H2, CH4, C2H2, C2H4, and C2H6 are used as initial features to expand the feature quantity, and 21 candidate features are obtained. In the figure, from number 6 to number 21 are all expanded fault features, among which total hydrocarbon = CH4 + C2H6 + C2H4 + C2H2, total gas = H2 + CH4 + C2H6 + C2H4 + C2H2. Since the subsequent support vector machine usually assumes that the features are on the same scale, the above 21 features are subjected to Z-score standardization using the following formula:

[0093] Y i =(X i -μ i ) / δ i

[0094] Where: X i is the original value of the i-th reselected feature, Y i Yes X i The value after normalization, μ i is the mean of the ith feature, δ i The standard deviation of the ith feature.

[0095] Further, using Figure 3 The Pareto chart is used to find the dimensions with a cumulative contribution of 95%, and the kernel principal component analysis method is used to perform feature fusion and feature extraction.

[0096] S4: The traditional black kite algorithm is hybridly improved by combining the Logistic chaotic mapping strategy with the lens imaging reverse learning strategy.

[0097] In S4: the following formula is used to introduce the Logistic chaos mapping strategy into the population initialization stage, thereby improving the diversity and global search capability of the initial population. The Logistic chaos mapping expression is:

[0098] X i+1 =rX i (1-X i )

[0099] Where: X i is the population of generation i, X i+1 is the population of the i+1 band, r is the control parameter, when r∈(3.57,4], the system enters a chaotic state, showing the characteristics of sensitive dependence on initial conditions, and when r is 4, the system is completely chaotic. Take r equal to 4.

[0100] The lens imaging reverse learning strategy guides the algorithm's attack behavior, using the current optimal position of the black kite to generate a lens reverse position to improve population diversity and expand the search range. Figure 4 As shown, its principle expression is:

[0101]

[0102] In the formula, let h / h′=K, and transform the above formula to get

[0103]

[0104] Then we get the expression in D-dimensional space

[0105]

[0106] In the above formula, X j The original position of the black kite in D-dimensional space during the migration process, X′ j is the position of the black-winged kite after the D-dimensional space transformation, ub j lb j are the upper and lower bounds of the hyperparameters in the space, respectively.

[0107] like Figure 5 As shown in the figure, the Eggholder function has a complex search space and multiple local minima when tested on two typical test functions f1 and f2. The Griewank function has many local minima and a unique global minimum. It can be seen that the IBKA algorithm is easier to jump out of the local optimum than other algorithms, which further proves the reliability and effectiveness of IBKA.

[0108]

[0109] S5: Use the improved black kite algorithm I BKA to optimize the best parameter combination of the SVM model penalty factor c and kernel function parameter g;

[0110] In S5: there are the following five steps:

[0111] S501: Introduce logic chaos map to initialize the population;

[0112] S502: Introducing the lens imaging reverse learning strategy;

[0113] S503: The leader of the population attacks;

[0114] S504: Migrate to the black-winged kite with the optimal fitness value;

[0115] S505: Determine whether the constraint conditions are met, and if so, output the optimal hyperparameter combination; if not, return to S503;

[0116] S6: Input the low-dimensional sensitive training set after data reconstruction into the optimal parameter combination SVM model for training, and use the trained model for fault diagnosis.

Claims

1. A transformer fault diagnosis method based on data reconstruction and IBKA-SVM, characterized by: This is accomplished by the following steps: S1: Collect publicly available dissolved gas analysis datasets with labeled transformer fault types; S2: The collected original data set is divided into a training set and a test set in a ratio of 7:3; S3: Build a data reconstruction model, perform Borderline-Smote-GMM mixed sampling to balance the data samples of the divided training set, and reconstruct the features of the test set and the balanced training set. The upper layer includes a feature expansion module and a Z-score standardization module, and the lower layer uses the kernel principal component analysis method to perform feature fusion and dimensionality reduction. S4: Combining the Logistic chaotic mapping strategy with the lens imaging reverse learning strategy to improve the traditional black kite algorithm; S5: Use the improved black-winged kite algorithm IBKA to optimize the best parameter combination of the SVM model penalty factor c and kernel function parameter g; S6: Input the low-dimensional sensitive training set after data reconstruction into the optimal parameter combination SVM model for training, and use the trained model for fault diagnosis.

2. The transformer fault diagnosis method based on IBKA-SVM of data reconstruction according to claim 1, characterized in that: The specific steps of establishing the proposed Borderline-Smote-GMM hybrid sampling model and feature reconstruction model in step S3 are as follows: Step 1: Calculate each sample P in the minority class sample P according to the K nearest neighbor formula i The Euclidean distance of the nearest K samples, and the minority class samples are divided into three categories according to the K nearest neighbor distance Where: x i and i are two sample points in n-dimensional space respectively; dist(X,Y) is the Euclidean distance between the two sample points; Step 2: Sample point x i There are m nearest neighbors, and among the m nearest neighbors, n sample points belong to the majority class samples; Step 3: Perform SMOTE interpolation on Danger samples X new =X i +λ×(X j -X i ) Where: X new is the synthesized new sample; X j Yes X i The minority class sample set among the m nearest neighbors of ; λ takes random values ​​in the range of 0 to 1; Step 4. GMM is a parameterized probability distribution model that divides the majority class samples into multiple Gaussian components. Each Gaussian component represents a cluster. The following posterior probability is used to calculate the probability that each data point belongs to each component. Where x is the measured data point; Φ(x|θ k ) represents the Gaussian distribution density of the kth sub-model; α k represents the probability of the kth sub-model being selected; θ k is μ k and Parameter set of μ k and represents the mean and variance of the kth sub-model; Step 5: Use the feature expansion method recommended by IEC and IEEE to expand the feature quantity, take the concentration data of five gases, H2, CH4, C2H2, C2H4, and C2H6, as the initial features and then expand the feature quantity to obtain 21 features to be selected; Use the following formula to perform Z-score standardization and put the feature quantities on the same scale: Y i =(X i -m i ) / d i Where: X i is the original value of the i-th reselected feature, Y i Yes X i The value after normalization, μ i is the mean of the ith feature, δ i The standard deviation of the i-th feature; Step 6: Use the kernel principal component analysis method to perform feature fusion and feature extraction on high-dimensional features.

3. The transformer fault diagnosis method based on IBKA-SVM of data reconstruction according to claim 1, characterized in that: In step S4, the traditional black kite algorithm is improved by hybridization using the Logistic chaotic mapping strategy and the lens imaging reverse learning strategy. The specific method is as follows: The Logistic chaotic mapping strategy is introduced into the population initialization stage using the following formula, which improves the diversity and global search capability of the initial population. The Logistic chaotic mapping expression is: X i+1 =rX i (1-X i ) Where: X i is the population of generation i, X i+1 is the population of band i+1, r is the control parameter, when r∈(3.57,4], the system enters a chaotic state, showing the characteristics of sensitive dependence on initial conditions, when r is 4, the system is completely chaotic, and r is taken to be 4; The lens imaging reverse learning strategy guides the algorithm's attack behavior, and generates a lens reverse position based on the current optimal position of the black kite to improve population diversity and expand the search range, as shown in Figure 4. Its principle expression is: In the formula, let h / h′=K, and transform the above formula to get Then we get the expression in D-dimensional space In the above formula, X j The original position of the black kite in D-dimensional space during the migration process, X′ j is the position of the black-winged kite after the D-dimensional space transformation, ub j lb j are the upper and lower bounds of the hyperparameters in the space, respectively.

4. The transformer fault diagnosis method based on IBKA-SVM of data reconstruction according to claim 1 is characterized in that: The specific steps in step S5 are: S501: Introduce logic chaos map to initialize the population; S502: Introducing the lens imaging reverse learning strategy; S503: The leader of the population attacks; S504: Migrate to the black-winged kite with the optimal fitness value; S505: Determine whether the constraint conditions are met. If so, output the optimal hyperparameter combination. If not, return to S503.

Images