Transformer fault diagnosis method and system based on IGWO-GBDT

Abstract

This invention relates to the technical field of transformer fault diagnosis, and in particular to a transformer fault diagnosis method and system based on the IGWO-GBDT. This method constructs an improved gray wolf optimization algorithm (IGWO) and optimizes the position update strategy of the gray wolf population through a nonlinear convergence factor, a dynamic weight allocation mechanism, and a Levy flight perturbation mechanism. The improved gray wolf optimization algorithm is used to globally optimize the hyperparameter combination of a gradient boosting tree to determine the optimal hyperparameter combination. The optimized gradient boosting tree is trained using a dissolved gas analysis dataset and then used to diagnose transformer fault types. This method uses the improved gray wolf optimization algorithm (IGWO) to optimize GBDT hyperparameters, improving diagnostic accuracy and efficiency.

Description

Technical Field

[0001] The present invention belongs to the technical field of power equipment fault diagnosis, and in particular relates to a transformer fault diagnosis method and system based on IGWO-GBDT. Background Art

[0002] As a core device in the power system, the operating status of the transformer is directly related to the safety and stability of the power grid. During the operation of the transformer, due to the long-term exposure to multiple stresses such as electrical, thermal, and mechanical stresses, various internal faults may occur, such as partial discharge, overheating, and insulation aging. If these faults are not diagnosed and handled in a timely manner, they may cause equipment damage or even cause serious power accidents. Therefore, accurate and efficient transformer fault diagnosis technology is an important means to ensure the reliable operation of the power system. Dissolved gas analysis (DGA), a technology widely used in transformer fault diagnosis, can effectively reflect the operating status and potential fault types of the transformer by monitoring changes in the dissolved gas content in the oil. However, in practical applications, DGA data often suffers from problems such as sample imbalance, complex features, and diverse fault categories, which poses severe challenges to traditional diagnostic methods.

[0003] Existing transformer fault diagnosis methods primarily include rule-based methods, statistical analysis methods, and machine learning methods. Rule-based methods, such as the IEC three-ratio method and the Duval triangle method, rely on expert-developed criteria. However, these methods are less adaptable to complex faults and are prone to misdiagnosis or omission. Statistical analysis methods analyze data through mathematical modeling, but their assumptions about data distribution may not align with reality, limiting diagnostic accuracy. In recent years, machine learning methods have garnered widespread attention in transformer fault diagnosis due to their powerful nonlinear modeling capabilities and adaptability to complex data. However, traditional machine learning models, such as support vector machines (SVMs) and random forests (RFs), have limited performance when dealing with high-dimensional data and sample imbalance. Furthermore, their hyperparameter optimization processes often rely on manual experience, making it difficult to achieve global optimization. Furthermore, the diagnostic performance of existing methods often degrades significantly when faced with insufficient fault sample diversity or high uncertainty samples.

[0004] To address these issues, a transformer fault diagnosis method is urgently needed that can effectively address sample imbalance, improve model generalization, and automatically optimize hyperparameters. Specifically, a method is needed to generate high-quality synthetic samples during the data preprocessing phase to balance the dataset. This method, combined with advanced optimization algorithms, can globally optimize the hyperparameters of the machine learning model, thereby improving the model's diagnostic accuracy and reliability. Furthermore, a more comprehensive evaluation mechanism is needed that comprehensively considers the diagnostic difficulty of different fault categories and the model's performance on high-uncertainty samples to further enhance the credibility and practicality of the diagnostic results. Summary of the Invention

[0005] This paper addresses the challenges of data imbalance, inefficient model hyperparameter optimization, and insufficient diagnostic accuracy in existing transformer fault diagnosis methods. By doing so, we propose a transformer fault diagnosis method and system based on IGWO-GBDT. This method utilizes an auxiliary classifier generative adversarial network (ACGAN) for data balancing, combined with an improved grey wolf optimization algorithm (IGWO) to optimize the hyperparameters of the gradient boosted tree (GBDT) model, and employs a weighted F1 score and uncertainty assessment to improve the fitness function. This approach overcomes the shortcomings of existing technologies and enables efficient and accurate diagnosis of transformer faults.

[0006] The present invention is implemented by the following technical solution: a transformer fault diagnosis method based on IGWO-GBDT, the steps are as follows:

[0007] S1: Constructing dissolved gas analysis dataset;

[0008] S2: Construct an improved gray wolf optimization algorithm (IGWO) to optimize the position update strategy of the gray wolf population through nonlinear convergence factors, dynamic weight allocation mechanism and Levy flight disturbance mechanism;

[0009] S3: Use the improved gray wolf optimization algorithm to perform global optimization on the hyperparameter combination of the gradient boosting tree (GBDT) to determine the optimal hyperparameter combination;

[0010] S4: The optimized gradient boosting tree is trained using the dissolved gas analysis dataset, and the trained gradient boosting tree is used to diagnose transformer fault types.

[0011] Further preferably, step S1 includes: preprocessing the original dissolved gas analysis data of the transformer to obtain preprocessed dissolved gas analysis data, performing sample balancing processing on the preprocessed dissolved gas analysis data through an auxiliary classifier generative adversarial network (ACGAN), generating synthetic samples consistent with the distribution of real fault data, generating a dissolved gas analysis data set consisting of the synthetic samples and the dissolved gas analysis data, and dividing the dissolved gas analysis data set into a training set and a test set.

[0012] Further preferably, the pre-processed dissolved gas analysis data is subjected to sample balancing processing by using an auxiliary classifier generative adversarial network (ACGAN). The specific process is as follows:

[0013] The random noise vector z is concatenated with the fault category label c as the input of the generator to guide the generator to generate synthetic samples. The synthetic samples are expressed as , G represents the generator;

[0014] The discriminator not only contains a binary classifier for judging whether a sample is true or false, but also adds an auxiliary classifier for predicting the fault category of the sample;

[0015] The optimization objective of the auxiliary classifier generative adversarial network (ACGAN) consists of two parts. One is to measure the ability of the discriminator to distinguish between real samples and generated samples. , the other is to ensure the consistency of the discriminator and the generator in identifying the sample category ; At the same time, in the optimization process, feature matching items are introduced , the expression is:

[0016] ;

[0017] in, is the feature matching item, Indicates the first The feature extraction function of the layer, n is the number of discriminator layers, and Respectively represent real samples and synthetic samples, Represents the number of samples; the total loss function is:

[0018] ;

[0019] in, Generate the total adversarial network loss for the auxiliary classifier, and They are and The weight coefficient of .

[0020] Further preferably, the calculation formula of the nonlinear convergence factor is:

[0021] ;

[0022] in, is the nonlinear convergence factor, is the initial convergence factor, ranging from [0, 2]; is the current iteration number; is the maximum number of iterations; and are two nonlinear adjustment parameters, and , , e is a natural constant.

[0023] Further optimization is performed by introducing a dynamic weight allocation mechanism to dynamically adjust the weights of α wolf, β wolf, and δ wolf according to the current number of iterations. The weight calculation formula is as follows:

[0024] ;

[0025] ;

[0026] ;

[0027] in, 、 、 Represent the weights of α wolf, β wolf, and δ wolf respectively.

[0028] The optimization process using the Levy flight perturbation mechanism is:

[0029] exist The Lévy flight disturbance is introduced during the wolf's position update process, and its step length follows the following distribution:

[0030] ;

[0031] in, represents the step size of the Lévy distribution, is the scale parameter, is a positional parameter, is a random variable, represents the characteristic exponent of the Levy distribution, The mean is 0 and the variance is The normal distribution of represents the standard normal distribution, standard deviation Calculated by the gamma function;

[0032] After each location update, The wolf's position is subjected to Levy flight perturbation, and the perturbation intensity coefficient is set to , to control the jump amplitude; the position update formula is:

[0033] ;

[0034] ;

[0035] in, Express The Lévy flight perturbation imposed by the wolf's position, is the updated population position, is the population position before updating.

[0036] Further preferably, the hyperparameters in step S3 include the number of trees, learning rate, maximum tree depth, subsample ratio, minimum sample split and minimum sample leaf node.

[0037] Further preferably, the gradient boosting tree (GBDT) introduces L1 regularization parameters and L2 regularization parameters during the training process, and adopts an adaptive regularization parameter adjustment mechanism to dynamically adjust the regularization parameters according to the performance of the gradient boosting tree during the training process.

[0038] Furthermore, when performing global optimization on the hyperparameter combination of the gradient boosting tree, the weighted F1 score is used as the fitness function, and the category weight adjustment mechanism and gradient boosting tree uncertainty evaluation are introduced. The calculation formula is as follows:

[0039] ;

[0040] Where C represents the number of fault categories, represents the number of type c fault samples, represents the F1 score of the c-th fault, represents the category weight of the c-th fault;

[0041] At the same time, the fitness function combines the uncertainty evaluation of the gradient boosting tree prediction to give an additional reward term R to the performance of the gradient boosting tree on high uncertainty samples. The calculation formula is:

[0042] ;

[0043] in, represents the prediction uncertainty of the gradient boosting tree for the c-th type fault sample, express Corresponding weight coefficient; final fitness function for:

[0044] ;

[0045] in, Represents the reward item weight.

[0046] The present invention also provides a transformer fault diagnosis system based on IGWO-GBDT, comprising a data acquisition and processing module, a model building module and a fault diagnosis module;

[0047] The data acquisition and processing module acquires raw dissolved gas analysis data and processes it to construct a dissolved gas analysis data set;

[0048] The model building module optimizes the position update strategy of the gray wolf population through a nonlinear convergence factor, a dynamic weight allocation mechanism, and a Levy flight perturbation mechanism, thereby constructing an improved gray wolf optimization algorithm (IGWO). The improved gray wolf optimization algorithm is used to perform a global optimization of the hyperparameter combination of the gradient boosting tree (GBDT) to determine the optimal hyperparameter combination. The optimized gradient boosting tree is then trained using a dissolved gas analysis dataset.

[0049] The fault diagnosis module deploys the trained gradient boosting tree for transformer fault type diagnosis.

[0050] This paper combines the Improved Grey Wolf Optimizer (IGWO) algorithm with a Gradient Boosted Tree (GBDT) algorithm to address data imbalance in transformer fault diagnosis, improving the model's generalization and diagnostic accuracy. This method enhances the efficiency and accuracy of hyperparameter optimization through a nonlinear convergence factor, a dynamic weight allocation mechanism, and a Levy flight perturbation mechanism. Furthermore, the method generates high-quality synthetic samples using an ACGAN, improving the class balance of the dataset and providing reliable support for transformer condition assessment and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 Schematic diagram of the overall process of the transformer fault diagnosis method based on IGWO-GBDT of the present invention. DETAILED DESCRIPTION

[0052] The present invention will be further described in detail below with reference to the accompanying drawings.

[0053] Reference Figure 1 ,A transformer fault diagnosis method based on IGWO-GBDT,the steps are as follows:

[0054] S1: Construct a dissolved gas analysis dataset; preprocess the original transformer dissolved gas analysis data (including outlier processing and feature normalization) to obtain preprocessed dissolved gas analysis data. Use the auxiliary classifier generative adversarial network (ACGAN) to perform sample balancing on the preprocessed dissolved gas analysis data to generate synthetic samples consistent with the distribution of real fault data. The dissolved gas analysis dataset consists of the synthetic samples and the dissolved gas analysis data, and the dissolved gas analysis dataset is divided into a training set and a test set.

[0055] S2: Construct an improved gray wolf optimization algorithm (IGWO) to optimize the position update strategy of the gray wolf population through nonlinear convergence factors, dynamic weight allocation mechanism and Levy flight disturbance mechanism;

[0056] S3: Use the improved gray wolf optimization algorithm to perform global optimization on the hyperparameter combination of the gradient boosting tree (GBDT) to determine the optimal hyperparameter combination;

[0057] S4: The optimized gradient boosting tree is trained using the dissolved gas analysis dataset, and the trained gradient boosting tree is used to diagnose transformer fault types.

[0058] In step S1, during the data collection and preprocessing phase, the raw dissolved gas analysis (DGA) data generated during oil-immersed transformer operation is processed to generate a raw DGA dataset. Preprocessing (including outlier processing and feature normalization) is then performed to obtain a preprocessed DGA dataset. The preprocessed DGA dataset is then subjected to sample equalization using an auxiliary classifier generative adversarial network (ACGAN) model to generate a balanced DGA dataset. The balanced DGA dataset is then divided into training, validation, and test sets to prepare for subsequent model training and parameter optimization.

[0059] In practical applications, raw dissolved gas analysis (DGA) data usually contain hydrogen ( ), methane ( ), ethane ( ), ethylene ( ), acetylene ( ) and other gas content information. Raw dissolved gas analysis data collected by sensors may contain outliers, such as data deviations caused by equipment aging or external interference. To eliminate outliers, statistical anomaly detection methods, such as boxplots or Z-scores, are used to identify and delete data points that fall outside the normal range. After outlier processing is completed, the data is feature normalized to ensure that each feature has the same dimension and distribution range. This process uses min-max normalization or Z-score normalization to ensure good numerical stability of the data during subsequent modeling.

[0060] Since the number of samples of different fault categories in the original dissolved gas analysis data may vary significantly, directly using it will lead to insufficient learning of minority category samples in the model, thus affecting the diagnostic accuracy. Therefore, the sample balancing process of the preprocessed dissolved gas analysis data is performed using the auxiliary classifier generative adversarial network (ACGAN). The specific process is as follows:

[0061] Generator input: The random noise vector z is concatenated with the fault category label c as the input of the generator to guide the generator to generate synthetic samples of a specific category. The synthetic samples are represented as , G represents the generator. At the same time, the feature fusion network is used to perform deep fusion processing on the random noise vector z and the fault category label c, extract the key features in the random noise vector z and the fault category label c, and perform feature cross-interaction to form a more representative feature vector, which is then input into the generator main network to generate higher quality synthetic samples.

[0062] Discriminator: The discriminator not only includes a binary classifier for determining sample authenticity but also adds an auxiliary classifier for predicting the sample's fault category. In addition to retaining the original tasks of determining sample authenticity and predicting class labels, the discriminator now also adds a fault severity prediction branch to predict the severity of the fault corresponding to the input sample. A corresponding loss term related to fault severity prediction is added to the optimization objective. This allows the discriminator to simultaneously learn the authenticity, category, and fault severity characteristics of samples during training, improving its overall performance in distinguishing faulty samples.

[0063] Optimization goal: The optimization goal of the auxiliary classifier generative adversarial network (ACGAN) consists of two parts. One is to measure the ability of the discriminator to distinguish between real samples and generated samples. , the other is to ensure the consistency of the discriminator and the generator in identifying the sample category , which is expressed as follows:

[0064] ;

[0065] ;

[0066] in, Represents the real data distribution expectations, Represents the noise distribution and category distribution joint expectations, Indicates sampling real samples from real datasets and its corresponding fault category label c, and the paired data ( ) Calculate the expected value. represents the real sample, Represents the synthetic sample generated by the generator based on random noise z and fault category label c. The goal of model training is to maximize the generator G , maximize the discriminator D At the same time, in the optimization process, feature matching items are introduced , so that the sample features generated by the generator are closer to the real sample features in statistical distribution, thereby improving the quality and diversity of the generated samples. Its expression is:

[0067] ;

[0068] in, is the feature matching item, Indicates the first The feature extraction function of the layer, n is the number of discriminator layers, and Respectively represent real samples and synthetic samples, Represents the number of samples. The total loss function is:

[0069] ;

[0070] in, Generate the total adversarial network loss for the auxiliary classifier, and They are and The weight coefficient of .

[0071] Through the above improvements, the optimized auxiliary classifier generative adversarial network can generate higher quality synthetic samples, while improving the comprehensive discrimination performance of the discriminator for fault samples, providing a more balanced and accurate data basis for subsequent transformer fault diagnosis based on IGWO-GBDT, and improving the performance and reliability of the entire fault diagnosis method.

[0072] After the data set is partitioned, the model parameter optimization and verification phase begins. The core of this phase is to globally optimize the hyperparameter combinations of the GBDT model using the Improved Grey Wolf Optimization (IGWO) algorithm. The Improved Grey Wolf Optimization (IGWO) algorithm improves upon the standard Grey Wolf Optimization algorithm in three aspects: a nonlinear convergence factor, a dynamic weight allocation mechanism, and a Levy flight perturbation mechanism.

[0073] A nonlinear convergence factor is used to replace the linear convergence factor in the standard Grey Wolf Optimization (GWO) algorithm. The calculation formula of the nonlinear convergence factor is:

[0074] ;

[0075] in, is the nonlinear convergence factor, is the initial convergence factor, ranging from [0, 2]; is the current iteration number; is the maximum number of iterations; and are two nonlinear adjustment parameters, and , , e is a natural constant.

[0076] This nonlinear convergence factor can provide a larger exploration space in the early stage of algorithm iteration and enhance the global search capability; as the number of iterations increases, the convergence factor gradually decreases, allowing the algorithm to focus on local search in the later stage and improve convergence accuracy.

[0077] A dynamic weight allocation mechanism is used to replace the fixed weight allocation in the standard Grey Wolf Optimization (GWO) algorithm, as follows:

[0078] A dynamic weight allocation mechanism is introduced to dynamically adjust the weights of α wolf, β wolf, and δ wolf according to the current number of iterations to balance global exploration and local development capabilities. The weight calculation formula is as follows:

[0079] ;

[0080] ;

[0081] ;

[0082] in, 、 、 Represent the weights of α wolf, β wolf, and δ wolf respectively.

[0083] In the initial iterations, the α-wolf is given a larger weight to highlight its global guiding role. As the number of iterations increases, the α-wolf's weight is gradually reduced, while the β-wolf and δ-wolf weights are increased. This enhances local search capabilities, accelerates convergence, and improves optimization accuracy. Through these improvements, the dynamic weight allocation mechanism can better adapt to the needs of different stages in the optimization process, balance global exploration and local development capabilities, and enhance the optimization performance of the IGWO algorithm, making it more suitable for optimizing GBDT hyperparameters in transformer fault diagnosis.

[0084] The Levy flight perturbation mechanism is used to enhance population diversity and help the algorithm escape from the local optimal solution, as follows:

[0085] exist The Lévy flight disturbance is introduced during the wolf's position update process, and its step length follows the following distribution:

[0086] ;

[0087] in, represents the step size of the Lévy distribution, is the scale parameter, is a positional parameter, is a random variable, represents the characteristic exponent of the Levy distribution, =1.5, The mean is 0 and the variance is The normal distribution of represents the standard normal distribution, standard deviation It is calculated by the gamma function, specifically:

[0088] ;

[0089] in, is the gamma function.

[0090] After each location update, The wolf's position is subjected to Levy flight perturbation, and the perturbation intensity coefficient is set to , to control the jump amplitude. The specific position update formula is:

[0091] ;

[0092] ;

[0093] in, Express The Lévy flight perturbation imposed by the wolf's position, is the updated population position, is the population position before updating.

[0094] Dynamically adjust the disturbance intensity coefficient based on the current population diversity indicators (such as the average distance between individuals in the population, the standard deviation of the fitness value, etc.) When population diversity is low, increase To enhance the global search capability; when the population diversity is high, reduce Through fine local search, we can more effectively escape from the local optimal solution and enhance the global optimization ability and convergence speed of the algorithm.

[0095] Through the above improvements, the Levy flight perturbation mechanism can more flexibly adapt to different stages of the optimization process, improve population diversity, enhance the algorithm's ability to escape local optimal solutions, and improve the efficiency and effectiveness of the improved grey wolf optimization algorithm (IGWO) in optimizing the hyperparameter combination of gradient boosting tree (GBDT) in transformer fault diagnosis.

[0096] In step S3, the improved grey wolf optimization algorithm (IGWO) is used to perform global optimization on the hyperparameter combination of the gradient boosting tree (GBDT). The hyperparameters include the number of trees, learning rate, maximum tree depth, subsample ratio, minimum sample split and minimum sample leaf node. The search range of each hyperparameter is as follows:

[0097] Number of trees ;

[0098] Learning rate ;

[0099] Maximum depth ;

[0100] Subsample proportion ;

[0101] Minimum sample split ;

[0102] Minimum sample leaf node ;

[0103] L1 regularization parameter ;

[0104] L2 regularization parameter ;

[0105] The termination condition of the hyperparameter optimization is that the performance of the gradient boosting tree is not significantly improved after 10 consecutive iterations, and the hyperparameters that have a greater impact on the performance of the gradient boosting tree are adjusted first during the hyperparameter optimization process.

[0106] The gradient boosting tree (GBDT) introduces L1 regularization parameters and L2 regularization parameters during the training process, and adopts an adaptive regularization parameter adjustment mechanism to dynamically adjust the regularization parameters according to the performance of the gradient boosting tree during the training process. The calculation formula is:

[0107] ;

[0108] Among them, M is the number of leaf nodes of the decision tree, is the predicted value of the leaf node. The adjustment formula of the regularization parameter is:

[0109] ;

[0110]

[0111] in, is the L1 regularization parameter for the t-th iteration, is the L1 regularization parameter for the t+1th iteration, is the L2 regularization parameter of the t-th iteration, is the L2 regularization parameter for the t+1th iteration, To adjust the rate, is the rate of change of performance of the gradient boosting tree in consecutive iterations.

[0112] In step S3, the weighted F1 score of the validation set is used as the fitness function, and the category weight adjustment mechanism and gradient boosting tree uncertainty assessment are introduced. The calculation formula is as follows:

[0113] ;

[0114] Where C represents the number of fault categories, represents the number of type c fault samples, represents the F1 score of the c-th fault, The weight of the category c fault is used to reflect the importance of this category in fault diagnosis. The weight coefficient can be manually set or automatically adjusted according to the risk level or diagnostic difficulty of the fault type.

[0115] At the same time, the fitness function combines the uncertainty evaluation of the gradient boosting tree prediction to give an additional reward term R to the performance of the gradient boosting tree on high uncertainty samples. The calculation formula is:

[0116] ;

[0117] in, represents the prediction uncertainty of the gradient boosting tree for the c-th type fault sample, express The corresponding weight coefficient. The final fitness function for:

[0118] ;

[0119] in, represents the reward item weight. By combining the class weight adjustment mechanism with the gradient boosting tree prediction uncertainty assessment, the improved fitness function can more comprehensively reflect the comprehensive performance of the gradient boosting tree on different fault categories. It also encourages the gradient boosting tree to improve diagnostic accuracy on high-uncertainty samples, providing more reliable results for transformer fault diagnosis.

[0120] After completing hyperparameter optimization, the model testing and diagnosis result output stage begins. In this stage, the gradient boosting tree with optimized parameters is tested using a test set to evaluate its performance in actual fault diagnosis tasks. Test metrics include key performance indicators such as accuracy, recall, and F1 score. Accuracy reflects the proportion of samples correctly classified by the gradient boosting tree for all samples, recall measures the ability of the gradient boosting tree to identify a certain type of fault sample, and the F1 score comprehensively considers the performance of accuracy and recall. During the test, by comparing the prediction results of the gradient boosting tree with the true labels, the values ​​of various indicators are calculated, and the performance of the gradient boosting tree on different fault categories is analyzed. Finally, the transformer fault diagnosis results are output based on the test results to provide decision support for transformer condition assessment and maintenance.

[0121] Another embodiment of the present invention provides a transformer fault diagnosis system based on IGWO-GBDT, comprising a data acquisition and processing module, a model building module, and a fault diagnosis module;

[0122] The data acquisition and processing module acquires raw dissolved gas analysis data and processes it to construct a dissolved gas analysis data set;

[0123] The model building module optimizes the position update strategy of the gray wolf population through a nonlinear convergence factor, a dynamic weight allocation mechanism, and a Levy flight perturbation mechanism, thereby constructing an improved gray wolf optimization algorithm (IGWO). The improved gray wolf optimization algorithm is used to perform a global optimization of the hyperparameter combination of the gradient boosting tree (GBDT) to determine the optimal hyperparameter combination. The optimized gradient boosting tree is then trained using a dissolved gas analysis dataset.

[0124] The fault diagnosis module deploys the trained gradient boosting tree for transformer fault type diagnosis.

[0125] Throughout the implementation process, the technical solution of the present invention solves key problems in the prior art through multiple innovations. First, the ACGAN model is used to perform sample balancing on the data set to generate high-quality synthetic samples, solving the problem of class imbalance in the data set. Secondly, the IGWO algorithm is used to perform global optimization on the hyperparameter combination of the gradient boosting tree, which significantly improves the generalization ability and diagnostic accuracy of the gradient boosting tree. Finally, the L1 regularization parameter and the L2 regularization parameter are introduced in the training process of the gradient boosting tree, and an adaptive regularization parameter adjustment mechanism is adopted to dynamically adjust the regularization parameters according to the performance of the gradient boosting tree during the training process, further enhancing the stability and robustness of the gradient boosting tree. The combination of the above technical solutions enables the present invention to effectively respond to the complex and changeable transformer fault diagnosis needs in actual application scenarios, and provides reliable technical support for the operation and maintenance of power equipment.

[0126] The above description merely represents preferred embodiments of the present invention and is not intended to limit the present invention in any other manner. Any person skilled in the art may utilize the above disclosure to modify or modify the present invention into equivalent embodiments. However, any simple modifications, equivalent variations, and modifications to the above embodiments that do not depart from the technical content of the present invention and are based on the technical essence of the present invention remain within the scope of protection of the present invention.

Claims

1. A transformer fault diagnosis method based on IGWO-GBDT, characterized in that the steps as follows: S1: Constructing dissolved gas analysis dataset; S2: Construct an improved gray wolf optimization algorithm to optimize the gray wolf population position update strategy through nonlinear convergence factor, dynamic weight allocation mechanism and Levy flight disturbance mechanism; S3: Use the improved gray wolf optimization algorithm to perform global optimization on the hyperparameter combination of the gradient boosting tree to determine the optimal hyperparameter combination; when performing global optimization on the hyperparameter combination of the gradient boosting tree, use the weighted F1 score as the fitness function, and introduce a category weight adjustment mechanism and gradient boosting tree uncertainty assessment. The calculation formula is as follows: ； Where C represents the number of fault categories, represents the number of type c fault samples, represents the F1 score of the c-th fault, represents the category weight of the c-th fault; At the same time, the fitness function combines the uncertainty evaluation of the gradient boosting tree prediction to give an additional reward term R to the performance of the gradient boosting tree on high uncertainty samples. The calculation formula is: ； in, represents the prediction uncertainty of the gradient boosting tree for the c-th type fault sample, express Corresponding weight coefficient; final fitness function for: ； in, Represents the reward item weight; S4: The optimized gradient boosting tree is trained using the dissolved gas analysis dataset, and the trained gradient boosting tree is used to diagnose transformer fault types.

2. The transformer fault diagnosis method according to claim 1, wherein the steps S1 includes: preprocessing the original dissolved gas analysis data of the transformer to obtain preprocessed dissolved gas analysis data, performing sample balancing processing on the preprocessed dissolved gas analysis data through the auxiliary classifier generative adversarial network, generating synthetic samples consistent with the distribution of real fault data, and a dissolved gas analysis data set composed of the synthetic samples and the dissolved gas analysis data, and dividing the dissolved gas analysis data set into a training set and a test set.

3. The transformer fault diagnosis method according to claim 2, wherein: The pre-processed dissolved gas analysis data is subjected to sample balancing through the auxiliary classifier generative adversarial network. The specific process is as follows: The random noise vector z is concatenated with the fault category label c as the input of the generator to guide the generator to generate synthetic samples. The synthetic samples are expressed as , G represents the generator; The discriminator not only contains a binary classifier for judging whether a sample is true or false, but also adds an auxiliary classifier for predicting the fault category of the sample; The optimization objective of the auxiliary classifier generative adversarial network (ACGAN) consists of two parts. One is to measure the ability of the discriminator to distinguish between real samples and generated samples. , the other is to ensure the consistency of the discriminator and the generator in identifying the sample category ; At the same time, in the optimization process, feature matching items are introduced , the expression is: ； in, is the feature matching item, Indicates the first The feature extraction function of the layer, n is the number of discriminator layers, and Respectively represent real samples and synthetic samples, Represents the number of samples; the total loss function is: ； in, Generate the total adversarial network loss for the auxiliary classifier, and They are and The weight coefficient of .

4. The transformer fault diagnosis method according to claim 1, wherein: The calculation formula of the nonlinear convergence factor is: ； in, is the nonlinear convergence factor, is the initial convergence factor, ranging from [0, 2]; is the current iteration number; is the maximum number of iterations; and are two nonlinear adjustment parameters, and , , e is a natural constant.

5. The transformer fault diagnosis method according to claim 4 is characterized in that: Dynamic weight allocation mechanism dynamically adjusts the weights of α wolf, β wolf, and δ wolf according to the current number of iterations. The weight calculation formula is as follows: ； ； ； in, 、 、 Represent the weights of α wolf, β wolf, and δ wolf respectively.

6. The transformer fault diagnosis method according to claim 5, characterized in that: The optimization process using the Levy flight perturbation mechanism is: The Levy flight disturbance is introduced in the position update process of the ω wolf, and its step length obeys the following distribution: ； in, represents the step size of the Levy distribution, is the scale parameter, is a positional parameter, is a random variable, represents the characteristic exponent of the Levy distribution, The mean is 0 and the variance is The normal distribution of represents the standard normal distribution, standard deviation Calculated by the gamma function; After each position update, a Levy flight perturbation is applied to the position of the ω wolf, and the perturbation intensity coefficient is set to ∈[0.01,0.1], to control the jump amplitude; the position update formula is: ； ； in, represents the Lévy flight perturbation imposed on the position of ω wolf, is the updated population position, is the population position before updating.

7. The transformer fault diagnosis method according to claim 1, wherein: The hyperparameters in step S3 include the number of trees, learning rate, maximum tree depth, subsample ratio, minimum sample split and minimum sample leaf node.

8. The transformer fault diagnosis method according to claim 1, wherein: The gradient boosting tree introduces L1 regularization parameters and L2 regularization parameters during the training process, and adopts an adaptive regularization parameter adjustment mechanism to dynamically adjust the regularization parameters according to the performance of the gradient boosting tree during the training process.

9. A transformer fault diagnosis system based on IGWO-GBDT, comprising a data acquisition and processing module, a model building module, and a fault diagnosis module, for implementing the transformer fault diagnosis method according to claim 1, characterized in that: The data acquisition and processing module collects raw dissolved gas analysis data and processes it to construct a dissolved gas analysis data set: The model building module optimizes the position update strategy of the gray wolf population through a nonlinear convergence factor, a dynamic weight allocation mechanism, and a Levy flight disturbance mechanism, thereby constructing an improved gray wolf optimization algorithm; the improved gray wolf optimization algorithm is used to perform a global optimization of the hyperparameter combination of the gradient boosting tree to determine the optimal hyperparameter combination; The optimized gradient boosting tree was trained using the dissolved gas analysis dataset; The fault diagnosis module deploys the trained gradient boosting tree for transformer fault type diagnosis.

Images