Transformer fault diagnosis method based on data generation-migration learning

Abstract

The present application relates to the technical field of detection of high-voltage electrical equipment in power systems, and particularly relates to a transformer fault diagnosis method based on data generation-migration learning. The method comprises the following steps: step S1, obtaining labeled data of oil dissolved gas analysis stored in a database, wherein the labeled data comprises gas content data and fault category labels; step S2, logarithmically transforming the gas content data into a form of approximate normal distribution; step S3, constructing a high-dimensional distribution for the data processed in step S2, using a Copula function to extract statistical information of the high-dimensional distribution, and then sampling oil dissolved gas analysis synthetic data samples according to the information; and step S4, after generating the oil dissolved gas analysis synthetic data, using an FT-Transformer model to pre-train and fine-tune, and performing fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of detection technology for high-voltage electrical equipment in power systems, and specifically to a transformer fault diagnosis method based on data generation-transfer learning. Background Technology

[0002] Power transformers are among the most critical pieces of equipment in power systems, and their safe and stable operation is of paramount importance. Currently, the vast majority of power transformers are oil-immersed, making the monitoring of their condition extremely crucial. Dissolved gas analysis (DGA) is one of the most important detection methods in oil-immersed transformer condition monitoring. Its advantage lies in its independence from electric and magnetic fields within the transformer, allowing for live testing while the transformer is in operation. The content of dissolved gases in the insulating oil is closely related to the operating condition of the power transformer, reflecting various internal fault states. Transformer fault diagnosis methods based on gas content data can achieve early fault diagnosis and real-time fault analysis during transformer operation.

[0003] Traditional fault diagnosis methods for DGA (Distributed Gauge Arrangement) fault diagnosis mainly include IEEE / IEC standards, Rogers ratio method, Duval three-ratio method, and Duval pentagon method. These traditional methods rely solely on discrimination criteria for diagnosis and have no data volume requirements, but they lack any learning ability, have overly absolute threshold boundaries, and thus have low fault diagnosis accuracy. Machine learning methods for DGA fault diagnosis mainly include support vector machines, Bayesian networks, multilayer perceptrons, and lightweight gradient boosting machines. These machine learning methods have a small number of model parameters and are suitable for diagnosis with limited data, but they struggle to uncover deep features from the data, and their fault diagnosis accuracy is also limited.

[0004] In recent years, deep learning technology has been widely applied in fields such as natural language processing and computer vision, resulting in significant improvements across these fields. This has created opportunities for achieving high-precision fault diagnosis of transformers. However, training deep learning requires a large amount of data. Since transformer anomalies are low-probability events, and the proportions of different anomalies vary greatly, the number of transformer fault samples is scarce, and the proportions of different fault types are disparate. Currently, the lack of DGA fault cases and data imbalance are significant factors limiting the development of transformer health monitoring. This leads to trained deep learning models lacking reliability, robustness, and generalization ability, making it difficult to achieve high-precision transformer fault diagnosis.

[0005] To address the issue of low model accuracy caused by insufficient sample size, it is essential to investigate methods for effectively expanding DGA data samples. However, DGA data is tabular data, and common data augmentation methods in the field of tabular data are still limited to minority class oversampling and other methods with limited generalization. Although a number of Generative Adversarial Networks (GANs) for tabular data synthesis have emerged in recent years, such as CTGAN, research on them is still in its early stages, and the quality of synthesized tabular data needs to be improved.

[0006] By analyzing the characteristics of DGA data for each fault type, it was found that after taking the logarithm log(·) of the DGA data, the distribution characteristics between each pair of gases have an obvious elliptical distribution pattern. This distribution pattern conforms to an approximate Gaussian distribution. The Copula function in multivariate distribution modeling can be used to study the joint distribution of multidimensional random variables. Therefore, it can be used to model DGA data. By sampling the modeled distribution model, any number of DGA composite data can be obtained.

[0007] Transfer learning is essential when task-specific real training data is scarce. It is widely used in fields such as computer vision and natural language processing, and pre-training-fine-tuning is a proven and effective transfer learning method. Therefore, it is necessary to study pre-training-fine-tuning methods based on DGA data. This involves first using pre-training on the DGA synthetic dataset to enable the deep learning model to learn the features of the synthetic data, and then fine-tuning the model using the DGA real dataset to achieve knowledge transfer.

[0008] FT-Transformer is a Transformer network designed specifically for tabular data. It can capture the correlation between multiple variables in tabular data through the attention mechanism. It achieves excellent results in tabular data classification tasks that other types of deep models have failed to achieve. Therefore, it is suitable for transformer fault diagnosis of DGA data. Summary of the Invention

[0009] Given the scarcity of real-world DGA fault data for transformers and the imbalance in sample categories, this disclosure utilizes a Copula function to model the multidimensional distribution of logarithmic gas content data for each category. A large amount of synthetic DGA data is obtained through model sampling. A pre-trained and fine-tuned FT-Transformer model is then implemented: pre-trained using synthetic DGA data and fine-tuned using a training set of real DGA data. This improves the fault diagnosis accuracy on the real DGA data test set. Ultimately, this model can be used for real-time DGA data monitoring and fault diagnosis in transformers, achieving power transformer condition monitoring and high-precision fault diagnosis.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] (1) Read the original labeled dataset of transformer DGA from the database. This real DGA dataset contains data on the content of each gas and labels of the fault category.

[0012] (2) Perform data cleaning on the real DGA dataset, replacing 0 with a very small value, such as 10. -2 To avoid the occurrence of Nan values ​​after subsequent logarithmic transformation;

[0013] (3) Take a logarithmic transformation log(·) with a certain base for the gas content column in the cleaned DGA real dataset, such as the natural logarithmic transformation ln(·), so that the distribution of the gas content column is approximately normal.

[0014] (4) According to the category to which the real DGA dataset belongs (partial discharge, low energy discharge, high energy discharge, low temperature overheating, medium temperature overheating, high temperature overheating and normal), it is divided into training set and test set according to the ratio. For example, training set train: test set test = 8:2, where the ratio of each category in the training set and the test set is the same.

[0015] (5) For each label category (partial discharge, low energy discharge, high energy discharge, low temperature overheating, medium temperature overheating, high temperature overheating and normal), the DGA training set data is modeled using the Gaussian Copula function for multivariate distribution modeling.

[0016] (6) For each category of Gaussian Copula model, sample and generate an equal number of DGA synthetic data. For example, sample 10,000 sample points for each category, add a label column, and combine the synthetic data of each label category into a DGA synthetic dataset. This synthetic dataset is a dataset with an equal number of samples for each category.

[0017] (7) Standardize the training set of the real DGA dataset and standardize the synthetic DGA dataset using the standardization parameters of the training set;

[0018] (8) The FT-Transformer model was pre-trained using the standardized DGA synthetic dataset;

[0019] (9) Fine-tune the pre-trained FT-Transformer model using the training set of the standardized DGA real dataset;

[0020] (10) The test set of the real DGA dataset is standardized using the training set standardization parameters;

[0021] (11) Test the FT-Transformer model using the test set of the standardized DGA real dataset;

[0022] (12) After passing the test, it can be applied to the monitoring and fault diagnosis of real transformer real-time DGA data.

[0023] A transformer fault diagnosis method based on data generation and transfer learning includes the following steps:

[0024] Step S1: Obtain the labeled data of dissolved gas analysis in oil stored in the database. The labeled data includes the content data of each gas and the fault category label.

[0025] Step S2: The gas content data is logarithmically transformed into an approximate normal distribution.

[0026] Step S3: Construct a high-dimensional distribution for the data processed in step S2. Use the Copula function to extract the statistical information of the high-dimensional distribution, and then use this information to sample the synthetic data samples for dissolved gas analysis in oil.

[0027] Step S4: After generating the synthetic data for dissolved gas analysis in oil, the FT-Transformer model can be used for pre-training and fine-tuning, and fault diagnosis can be performed.

[0028] This technical solution is further optimized so that the types of gases include hydrogen, methane, ethane, ethylene, and acetylene.

[0029] This technical solution is further optimized so that the fault categories include partial discharge (PD), low-energy discharge (D1), high-energy discharge (D2), low-temperature overheating (T1), medium-temperature overheating (T2), high-temperature overheating (T3), and normal state.

[0030] In a further optimization of this technical solution, step S1 also includes data cleaning of the real dataset of dissolved gas analysis in the oil, replacing 0 with a minimum value.

[0031] In a further optimization of this technical solution, the steps for synthesizing data in step S3 are as follows:

[0032] (1) The logarithmically converted oil dissolved gas analysis training set data is divided into 7 sub-data sets according to the labels: partial discharge, low energy discharge, high energy discharge, low temperature overheating, medium temperature overheating, high temperature overheating and normal state. Each sub-data set is synthesized using the following (2) to (5);

[0033] (2) Assuming that the distribution of the logarithmic gas content in each column of the dissolved gas analysis in the oil is, for example, Gaussian distribution, Beta distribution, truncated Gaussian distribution, uniform distribution, gamma distribution, or non-parametric Gaussian kernel density estimation, the distribution with the largest p value obtained by the Kolmogorov-Smirnov test is taken as the distribution of that column, and the distribution functions F1(x1), F2(x2), F3(x3), F4(x4), and F5(x5) of the 5 columns of data are obtained. In this embodiment, they are all Gaussian distributions.

[0034] (3) Treat a sample row as a vector X = (x1, x2, ..., x5), and use the inverse Gaussian distribution transform to convert X into Y = [Φ (-1) (F1(x1)),...,Φ (-1) (F5(x5))];

[0035] (4) After all sample rows have been transformed, calculate the transformed covariance matrix Σ;

[0036] (5) Substitute into formula (1.2) to obtain the Gaussian Copula function;

[0037] C(u1,…,u n )=Φ Σ (Φ (-1) (u1),...,Φ (-1) (u n (1.3)

[0038] (6) After the above steps, 7 Copula functions are obtained. The following uses each Gaussian Copula function to generate synthetic data for dissolved gas analysis in 7 types of oil.

[0039] (7) Sample vector v from a random n-dimensional Gaussian distribution;

[0040] (8) Find the Cholesky decomposition, LL T =Σ, thus obtaining vector u = Lv;

[0041] (9) Inverse transformation to obtain synthetic data

[0042] In a further optimization of this technical solution, after generating the dissolved gas analysis synthesis data in oil in step S4, the data needs to be standardized.

[0043] Further optimization of this technical solution, the training process of the FT-Transformer model is as follows:

[0044] In the pre-training phase, the original input is a synthetic dataset of dissolved gas analysis in oil, including 5-dimensional logarithmic data of gas content and 1-dimensional fault category data. A feature tokenizer is used to increase the dimensionality of each feature column and generate corresponding weight vectors. Numerical feature column: First, each dimension of the logarithmic gas content data of dissolved gas analysis in oil is multiplied element-wise with the corresponding weight vector, and a bias term is added. Categorical feature column: First, the DGA categorical features are one-hot encoded, then the encoded vector is multiplied with the weight matrix, and a bias term is added. Since there are no categorical features in the DGA gas content, this operation is unnecessary. Label column: The labels need to be converted to integer labels of [0, 1, ..., n-1], where n is the number of categories.

[0045] Its formula is expressed as:

[0046]

[0047]

[0048]

[0049] In the formula, num represents numerical features, cat represents categorical features, knum represents the number of numerical features (5 in DGA data), and kcat represents the number of categorical features (0 in DGA data). Let represent the one-hot vector of the j-th categorical feature. stack[·] denotes stacking vectors to form a feature representation matrix.

[0050] After feature processing, all the transformed vectors are stacked together to form the final feature representation. Then, these feature vectors are input into multiple Transformer blocks for transformation. Each Transformer block consists of a self-attention mechanism and a feedforward neural network, which are used to learn the dependencies between the features of the DGA and extract high-level feature representations.

[0051] Finally, the final output is generated by using the embedding vector corresponding to the [CLS] token at the last position in the Transformer block. This output is used to obtain the DGA fault classification result.

[0052] Unlike existing technologies, the above technical solution has the following beneficial effects:

[0053] The data generation technology used in this disclosure is based on the Gaussian Copula model. The generated synthetic DGA data retains the statistical and distributional characteristics of the original data well, and has higher reliability compared to data generated by generative adversarial networks. The FT-Transformer model used for fault diagnosis can better identify the fault type in the DGA data, and its accuracy is higher than that of traditional machine learning models and threshold discrimination models. The pre-training-fine-tuning technique in transfer learning is used to further improve the capabilities of the FT-Transformer model. Pre-training on a label-balanced synthetic dataset can avoid the impact of imbalanced labels in the real DGA data on the model, thereby improving the accuracy of fault diagnosis. Attached Figure Description

[0054] Figure 1 This is a flowchart of the method.

[0055] Figure 2 A diagram illustrating the data processing for the FT-Transformer model;

[0056] Figure 3 It consists of a feature tokenizer module and a Transformer layer. Detailed Implementation

[0057] To explain in detail the technical content, structural features, objectives, and effects of the technical solution, the following description is provided in conjunction with specific embodiments and accompanying drawings.

[0058] This invention discloses a transformer fault diagnosis method based on data augmentation-transfer learning, achieving high-precision fault diagnosis of transformers, such as... Figure 1 As shown, Figure 1 The method flowchart shows the specific steps as follows:

[0059] Step S1: Obtain the real transformer DGA tagged data stored in the database. The DGA tagged data includes gas content data and fault category labels. Gas types include hydrogen (H2), methane (CH4), ethane (C2H6), ethylene (C2H4), and acetylene (C2H2), and may also include other gas types. Fault categories include partial discharge (PD), low-energy discharge (D1), high-energy discharge (D2), low-temperature overheating (T1), medium-temperature overheating (T2), high-temperature overheating (T3), and normal state (Normal), and may also include other fault categories. This invention uses five gas types (hydrogen, methane, ethane, ethylene, and acetylene) and seven fault categories (partial discharge, low-energy discharge, high-energy discharge, low-temperature overheating, medium-temperature overheating, high-temperature overheating, and normal state) as an example. To avoid the gas content data of the subsequent DGA showing a Nan value after logarithmic transformation, the original real DGA dataset needs to be cleaned by replacing the 0s in the gas content column of the DGA data with a very small number. In this example, 10 is used.-2 ;

[0060] Step S2: Since the gas content in the real DGA dataset follows a long-tailed distribution, direct standardization would make the data difficult for the model to learn. To facilitate subsequent modeling, standardization, and fault diagnosis using the Gaussian Copula function, it is necessary to perform a logarithmic transformation (log(·)) on the gas content column of the DGA data, converting the distribution of each gas content into an approximately normal distribution for easier subsequent processing. This invention uses the natural logarithmic transformation (ln(·)) as an example. The DGA data is divided into a training set (train) and a test set (test) according to category labels (partial discharge, low-energy discharge, high-energy discharge, low-temperature overheating, medium-temperature overheating, high-temperature overheating, and normal state). The proportion of each category in the two datasets needs to be consistent, i.e., stratified segmentation according to label categories is required. In this example, the ratio of the training set (train) to the test set (test) is 8:2.

[0061] Step S3: After the data segmentation is completed, the training set train in the real DGA dataset can be used to synthesize the data. The five gases, hydrogen, methane, ethane, ethylene, and acetylene, are treated as a 5-dimensional distribution. A Gaussian Copula function is used to extract the statistical information of the 5-dimensional distribution for each of the seven categories of DGA data. Then, based on this information, a 5-dimensional Gaussian distribution is sampled and inversely transformed to obtain the DGA synthetic data sample.

[0062] Copula, meaning "connection" in Latin, was first introduced by Sklar in 1959 when answering a question about the relationship between multidimensional distribution functions and low-dimensional marginal distributions. It is widely used in statistics and finance. The principle of the Copula function is as follows:

[0063] The Copula function can separate the correlation between random variables from the marginal distribution of the variables for study. Therefore, it can be used to model and analyze the joint distribution of arbitrary dimensions, and thus can be used to model the 5-dimensional logarithmic gas content data in DGA data.

[0064] For an n-dimensional joint distribution H(x1,…x) n )=P(X1≤x1,…X n ≤x n The following processing can be performed on it. First, through each single variable X... i The cumulative distribution function (cdf) F i The standard uniformly distributed random variable U is obtained. i =F i (X i Then, using the Copula function C(u1,u2,...,u...) n This describes the relationship structure between n-dimensional random variables.

[0065]

[0066] Where X1, X2, ..., X n It is a random variable, x1, x2, ..., x n It is a sample of random variables, U1, U2, ..., U n Let u1, u2, ..., u be random variables that follow a standard uniform distribution. n It is a sample of a random variable that follows a standard uniform distribution, F i (-1) (u i ) is called F i The pseudo-inverse function is defined as F i (-1) (u)=inf{x:F(x)≥u}.

[0067] If F1, F2, ..., F n If the sequence is continuous, then C is unique; otherwise, C exists only in Ran(F1) × … × Ran(F1). n Ran(F) is the only one on the list. i ) represents F i All possible values. Conversely, if C is a Copula function, it can be seen from formula (1.1) that the correlation structure C is the joint distribution function of n-dimensional random variables whose marginal distributions all follow a uniform distribution on [0,1].

[0068] Copula functions come in various types. Since the DGA data exhibits a distinct elliptical distribution after logarithmic transformation, and the Gaussian Copula function is a commonly used Copula function that is well-suited for fitting this type of distribution, the Gaussian Copula function is used to fit the training set of the real DGA dataset.

[0069] Suppose X = (X1,...,X) n )~N n (0,Σ), where Σ is the covariance matrix and Φ represents the standard normal distribution, then the n-variable Gaussian Copula function is:

[0070] C(u1,…,u n )=Φ Σ (Φ (-1) (u1),...,Φ (-1) (u n (1.6)

[0071] Using continuous distributions F1(x1), F2(x2), ..., F n (x n Replace u1, u2, ..., u in formula (1.2)n This yields a multivariate distribution H(x1,…x) with a Gaussian correlation structure. n Assuming that the 5-dimensional marginal distributions of the DGA all follow a Gaussian distribution with certain parameters, then u1, u2, ..., u in the above formula can be expressed as u1, u2, ..., u2. n Replace it with the corresponding Gaussian distribution.

[0072] Similarly, the Gaussian Copula function can be used to sample from the multivariate distribution and perform an inverse transformation to obtain the desired DGA synthetic data.

[0073] The steps for synthesizing DGA data are as follows:

[0074] (1) The logarithmically converted DGA training set data is divided into 7 sub-data sets according to the labels (partial discharge, low energy discharge, high energy discharge, low temperature overheating, medium temperature overheating, high temperature overheating and normal state). Each sub-data set is synthesized using the following (2) to (5).

[0075] (2) Assume the distribution of logarithmic gas content in each column of the DGA, such as Gaussian distribution, Beta distribution, truncated Gaussian distribution, uniform distribution, gamma distribution, or a non-parametric Gaussian kernel density estimate. Use the Kolmogorov-Smirnov test, and the distribution with the largest p-value is taken as the distribution of that column. The distribution functions F1(x1), F2(x2), F3(x3), F4(x4), and F5(x5) of the five columns of data are obtained. In this embodiment, they are all Gaussian distributions.

[0076] (3) Treat a sample row as a vector X = (x1, x2, ..., x5), and use the inverse Gaussian distribution transform to convert X into Y = [Φ (-1) (F1(x1)),...,Φ (-1) (F5(x5))).

[0077] (4) After all sample rows have been transformed, calculate the transformed covariance matrix Σ.

[0078] (5) Substitute into formula (1.2) to obtain the Gaussian Copula function.

[0079] (6) After the above steps, seven Copula functions are obtained. The following uses each Gaussian Copula function to generate synthetic data for the seven types of DGA.

[0080] (7) Sample vector v from a random n-dimensional Gaussian distribution.

[0081] (8) Find the Cholesky decomposition, LL T =Σ, thus obtaining vector u = Lv.

[0082] (9) Inverse transformation to obtain synthetic data

[0083] Through the steps described above, any number of synthetic data points can be obtained. Equal numbers of samples are sampled from the seven Gaussian Copula models, and these seven types of synthetic data samples are added to the label column to form a balanced DGA synthetic dataset. It is unnecessary to further perform an exponential transformation on the DGA synthetic data to obtain a distribution similar to the original data. This is because the distribution of the original DGA data is very unfavorable for subsequent fault diagnosis; directly using the logarithmically transformed data is more beneficial for subsequent work.

[0084] Step S4: After generating the DGA synthetic data, the data needs to be standardized to facilitate model learning. Standardization requires using the mean and variance of the training set (train) from the real DGA dataset. First, the training set (train) of the real DGA dataset is standardized, and then the standardized parameters (mean and variance) of the training set are used to standardize the DGA synthetic dataset. The DGA synthetic dataset and the training set (train) from the real DGA dataset can be used for pre-training and fine-tuning of the FT-Transformer model, as well as for fault diagnosis.

[0085] FT-Transformer is a method that applies the Transformer architecture to tabular data. This method leverages the advantages of the Transformer when processing tabular data. See the diagram below for an illustration of FT-Transformer processing tabular data. Figure 2 First, each sample is increased in dimensionality using the feature tokenizer module, then trained using the Transformer module, and finally the prediction result is obtained using the Predict module.

[0086] In the pre-training phase, the original input is the DGA synthetic dataset, which includes 5-dimensional logarithmic data of gas content (numerical feature column) and 1-dimensional fault category data (label column). A feature tokenizer is used to increase the dimensionality of each feature column and generate corresponding weight vectors. These weight vectors capture the relationships and importance between the gas features. Numerical feature column: First, each dimension of the DGA logarithmic gas content data is multiplied element-wise with the corresponding weight vector, and a bias term is added. Categorical feature column: First, the DGA categorical features are one-hot encoded, then the encoded vector is multiplied with the weight matrix, and a bias term is added. Since there are no categorical features in the DGA gas content, this operation is unnecessary. Label column: The labels need to be converted to integer labels of [0, 1, ..., n-1], where n is the number of categories.

[0087] Its formula is expressed as:

[0088]

[0089] In the formula, num represents numerical features, cat represents categorical features, knum represents the number of numerical features (5 in DGA data), and kcat represents the number of categorical features (0 in DGA data). Let represent the one-hot vector of the j-th categorical feature. stack[·] denotes stacking vectors to form a feature representation matrix.

[0090] After feature processing, all transformed vectors are stacked together to form the final feature representation. These feature vectors are then fed into multiple Transformer blocks for transformation. Each Transformer block consists of a self-attention mechanism and a feedforward neural network, see... Figure 3 (b) is used to learn the dependencies between features of DGA and extract high-level feature representations.

[0091] Finally, the final output is generated using the embedding vector corresponding to the [CLS] token at the last position in the Transformer block. This output is used to obtain the DGA fault classification result.

[0092] The core idea of ​​FT-Transformer is to convert tabular data into feature vectors and then use the Transformer module to transform these vectors. This design enables FT-Transformer to better process tabular data and extract useful feature representations, thereby further improving the modeling and prediction capabilities of tabular data.

[0093] The FT-Transformer model is pre-trained using the DGA synthetic dataset to learn the basic features of the DGA data. This allows the model to learn the basic structure of the DGA data and the relationships between its features in advance, enabling the model to reach an optimal state and achieve the goal of knowledge transfer.

[0094] The fine-tuning phase involves copying a target model from the pre-trained source model and further training it using the target dataset. At this point, the parameters in the FT-Transformer model have learned from the DGA synthetic dataset, and this knowledge remains applicable to the DGA real dataset. Since the format and categories of the DGA synthetic dataset and the DGA real dataset are completely identical, the structure of the source model can be fully preserved, allowing for the copying of a target model and fine-tuning.

[0095] The FT-Transformer target model was optimized using the training set of the real DGA dataset. After simple iterative training, the fault diagnosis performance of the target model was further improved. For fine-tuning the target model, either the first few layers were fixed while the last few were fine-tuned, or the weights and biases of all layers were fine-tuned. Since the DGA synthetic dataset is extremely similar to the real dataset, it can be assumed that the first few layers of the model do not need to be trained; only the last Transformer layer and the last classification layer need to be trained. In this embodiment, a model with 3 layers was used, but other numbers of layers can also be used.

[0096] Step S5: Use the standardized parameters of the training set (train) of the real DGA dataset to standardize the test set (test). Use the standardized test set of the real DGA dataset to test the trained model and verify the training effect. After passing the verification, the model parameters can be saved for use in real-time DGA monitoring and fault diagnosis of transformers.

[0097] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Unless otherwise specified, an element defined by the phrase "comprising..." or "including..." does not exclude the presence of additional elements in the process, method, article, or terminal device that includes said element. Additionally, in this document, "greater than," "less than," "exceeding," etc., are understood to exclude the stated number; "above," "below," "within," etc., are understood to include the stated number.

[0098] Although the above embodiments have been described, those skilled in the art, once they understand the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the above descriptions are merely embodiments of the present invention and do not limit the scope of patent protection of the present invention. Any equivalent structural or procedural transformations made using the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A transformer fault diagnosis method based on data generation-transfer learning, characterized in that, Includes the following steps: Step S1: Obtain the labeled data of dissolved gas analysis in oil stored in the database. The labeled data includes the content data of each gas and the fault category label. Step S2: The gas content data is logarithmically transformed into an approximate normal distribution. Step S3 involves constructing a high-dimensional distribution from the data processed in Step S2, extracting statistical information from the high-dimensional distribution using the Copula function, and then sampling synthetic data samples for dissolved gas analysis in oil based on this information. The steps for synthesizing data in Step S3 are as follows: (1) The logarithmically converted oil dissolved gas analysis training set data is divided into 7 sub-data sets according to the labels: partial discharge, low energy discharge, high energy discharge, low temperature overheating, medium temperature overheating, high temperature overheating and normal state. Each sub-data set is synthesized using the following (2)~(5); (2) Assuming that the distribution of logarithmic gas content in each column of the dissolved gas analysis in oil is Gaussian, the distribution with the largest p-value obtained by the Kolmogorov-Smirnov test is taken as the distribution of that column, and the distribution functions of the five columns of data are obtained. , , , , ; (3) Treat a sample row as a vector Using the inverse transform of the Gaussian distribution to Convert to ; (4) After all sample rows have been transformed, calculate the transformed covariance matrix. ; (5) Substitute into formula (1.2) to obtain the Gaussian Copula function; （1.1） (6) After the above steps, we obtain 7 Copula functions. The following section uses each Gaussian Copula function... The function generates synthetic data from dissolved gas analysis in seven types of oil. (7) Sample vectors from a random n-dimensional Gaussian distribution. ; (8) Find the Cholesky decomposition. , to obtain vector ; (9) Inverse transformation to obtain synthetic data ; Step S4: After generating the synthetic data for dissolved gas analysis in oil, the FT-Transformer model is pre-trained and fine-tuned, and fault diagnosis is performed. The training process of the FT-Transformer model is as follows: In the pre-training phase, the original input is a synthetic dataset of dissolved gas analysis in oil, including 5-dimensional logarithmic data of gas content and 1-dimensional fault category data. A feature tokenizer is used to increase the dimensionality of each feature column and generate corresponding weight vectors. For the numerical feature column: first, element-wise multiplication is performed between each dimension of the logarithmic gas content data of dissolved gas analysis in oil and the corresponding weight vector, plus a bias term. For the categorical feature column: first, one-hot encoding is performed on the DGA categorical features, then the encoded vector is multiplied by the weight matrix, plus a bias term. For the label column: the labels need to be converted to... Integer labels, where n is the number of categories; Its formula is expressed as: （1.2） In the formula, num represents numerical features, cat represents categorical features, knum represents the number of numerical features (5 in DGA data), and kcat represents the number of categorical features (0 in DGA data). Let represent the one-hot vector of the j-th classification feature; stack[·] represents stacking the vectors to form a feature representation matrix; After feature processing, all the transformed vectors are stacked together to form the final feature representation. Then, these feature vectors are input into multiple Transformer blocks for transformation. Each Transformer block consists of a self-attention mechanism and a feedforward neural network, which are used to learn the dependencies between the features of the DGA and extract high-level feature representations. Finally, the final output is generated by using the embedding vector corresponding to the [CLS] token at the last position in the Transformer block. This output is used to obtain the DGA fault classification result.

2. The transformer fault diagnosis method based on data generation-transfer learning as described in claim 1, characterized in that, The types of gases include hydrogen, methane, ethane, ethylene, and acetylene.

3. The transformer fault diagnosis method based on data generation-transfer learning as described in claim 1, characterized in that, The fault categories include partial discharge (PD), low-energy discharge (D1), high-energy discharge (D2), low-temperature overheating (T1), medium-temperature overheating (T2), high-temperature overheating (T3), and normal state.

4. The transformer fault diagnosis method based on data generation-transfer learning as described in claim 1, characterized in that, Step S1 further includes cleaning the real dataset of dissolved gas analysis in the oil by replacing 0 with a minimum value.

5. The transformer fault diagnosis method based on data generation-transfer learning as described in claim 1, characterized in that, After generating the dissolved gas analysis synthesis data in oil in step S4, the data needs to be standardized.

Images