Transformer operation state evaluation method and device, electronic equipment and storage medium

Abstract

The invention relates to a transformer operation state evaluation method and device, electronic equipment and a storage medium, and relates to the technical field of transformer fault diagnosis, and the method comprises the steps: carrying out the sample enhancement of a minority class fault sample set in a multi-dimensional feature system through employing an IMGOA algorithm, and constructing an enhanced sample set; performing visual evaluation on the enhanced sample set by adopting a principal component analysis method, and screening key feature subsets of which the cumulative contribution rate of the enhanced sample set which is qualified through evaluation reaches a preset threshold value based on a feature importance evaluation mechanism of a CatBoost classification model; and performing optimization on hyper-parameters of the CatBoost classification model by using a multi-strategy enhanced goose swarm optimization algorithm, and training the CatBoost classification model based on the key feature subset and the optimal hyper-parameter combination to construct a transformer operation state evaluation model. According to the invention, the diagnosis precision can be improved, the fault type can be judged more accurately, and the possibility of misjudgment or missed judgment is reduced.

Description

Technical Field

[0001] This application relates to the field of transformer fault diagnosis technology, specifically to a method, device, electronic equipment, and storage medium for evaluating the operating status of a transformer. Background Technology

[0002] Against the backdrop of the rapid development of smart grids, higher standards have been set for the operational stability, fault tolerance, and intelligent control levels of various power equipment. Transformers, as key devices responsible for voltage regulation and power distribution, play a pivotal role in the entire power transmission and distribution network. They not only affect the efficient transmission of electricity but also play a decisive role in the stability and security of the power grid. Due to their high equipment cost and complex operating environment, a failure often triggers widespread power outages, causing significant economic losses and social impact. Therefore, ensuring the long-term, highly reliable, and safe operation of transformers has become one of the core challenges in current power system operation and maintenance management.

[0003] In recent years, the rapid development of sensor technology, wireless communication, the Internet of Things, and big data platforms has made it possible to acquire real-time multi-dimensional operating status data of transformers. Modern power systems are widely equipped with online monitoring devices that continuously collect key operating information such as winding temperature, dissolved gases in oil, partial discharge, load current, and voltage. This massive, high-frequency, and high-dimensional data provides a solid data foundation for achieving precise and intelligent transformer fault diagnosis.

[0004] Among numerous online monitoring technologies, the method based on dissolved gas analysis (DGA) has become an important tool for judging potential thermal and discharge faults inside transformers due to its significant advantages such as rapid detection, high sensitivity, and non-invasiveness.

[0005] Traditional DGA diagnostic techniques, such as the three-ratio method, Key Gas method, Dornenburg method, and IEC method, mainly determine the fault type based on the proportion or content of gases. Although these methods are theoretically sound and easy to operate, their diagnostic accuracy is limited, especially in the early stages of fault category overlap or abnormality, where they are prone to judgment bias. Secondly, they are highly dependent on data, and are prone to misjudgment or missed diagnosis when the sample size is small or the characteristics are complex. Summary of the Invention

[0006] In view of this, this application provides a method, apparatus, electronic device, and storage medium for evaluating the operating status of transformers. The main purpose is to address the limitations of traditional DGA diagnostic techniques, which primarily rely on the proportion or content of gases to determine fault types. These techniques have limited diagnostic accuracy, especially during periods of overlapping fault categories or early anomalies, making them prone to misjudgment. Furthermore, they are highly dependent on data, and in situations with few samples or complex overlapping features, they are prone to misjudgment or missed detection.

[0007] According to a first aspect of this application, a method for evaluating the operating status of a transformer is provided, the method comprising: Historical oil chromatographic data of dissolved gases in transformer oil are obtained. Based on the historical oil chromatographic data, a multidimensional feature system is constructed that includes fused gas concentration, ratio features and normalization combination. An improved Markov chain Monte Carlo method is used to perform sample augmentation processing on the minority class fault sample set in the multidimensional feature system to construct an augmented sample dataset for model training. Principal component analysis was used to perform visual evaluation of the augmented sample dataset, and the feature importance evaluation mechanism built into the CatBoost classification model was used to select the key feature subsets in the augmented sample dataset that have reached the preset contribution rate threshold and passed the evaluation. A multi-strategy enhanced goose flock optimization algorithm is constructed, and the optimization objective is to minimize the model loss function. The key feature subset and the corresponding training labels are used as input data. The multi-strategy enhanced goose flock optimization algorithm is used to optimize the hyperparameters of the CatBoost classification model to obtain the optimal hyperparameter combination. The CatBoost classification model is retrained based on the key feature subset and the optimal hyperparameter combination to construct a transformer operation status assessment model, and the constructed transformer operation status assessment model is used to assess the operation status of the transformer.

[0008] In some embodiments of this disclosure, before constructing a multidimensional feature system comprising fused gas concentrations, ratio characteristics, and normalized combinations based on the historical oil chromatographic data, the method further includes: The historical oil chromatographic data is cleaned, including: imputing missing values ​​of non-critical variables using the median method; removing sample data corresponding to missing values ​​of critical variables; retaining the first record of duplicate samples based on their timestamps and deleting other duplicate records; and detecting outliers using a combined box plot and percentile method. If a detected outlier matches the corresponding fault label, it is retained; otherwise, it is considered a false anomaly and excluded. The historical oil chromatographic data is subjected to sample labeling and verification processing, which includes: initially labeling the sample data in the historical oil chromatographic data based on operation and maintenance records, fault repair logs and expert diagnostic conclusions, and cross-validating the initial labeling results using the three-ratio method. If the initial labeling results are consistent with the three-ratio method verification results, the initial labeling results are taken as the final labeling results. If the initial labeling results are inconsistent with the three-ratio method verification results, the initial labeling results are taken as the final labeling results, and the three-ratio method verification results are used as auxiliary reference information to correct the final labeling results. The construction of a multidimensional feature system based on the historical oil chromatographic data, comprising fused gas concentration, ratio characteristics, and normalized combinations, includes: Based on the historical oil chromatography data after data cleaning, sample labeling and validation, a multidimensional feature system is constructed, which includes fused gas concentration, ratio features and normalized combinations. The multidimensional feature system is a 32-dimensional multi-level diagnostic feature system.

[0009] In some embodiments of this disclosure, the method of employing an improved Markov chain Monte Carlo method to perform sample augmentation processing on the minority class fault sample set in the multidimensional feature system to construct an augmented sample dataset for model training includes: Based on the minority class of fault samples, a Markov chain Monte Carlo objective function is constructed to evaluate the sample quality, and the acceptance probability is determined based on the Markov chain Monte Carlo objective function, wherein the acceptance probability is a decay factor and an adaptive penalty factor that dynamically change with the iteration rounds. Starting from the current sample state, candidate samples are generated using the suggested distribution; Based on the acceptance probability, determine whether to accept the candidate sample; Repeat the above steps until a preset number of augmented sample datasets are generated, wherein the augmented sample datasets consist of a preset number of candidate samples. The formula for calculating the acceptance probability is:

[0010] In the formula, Indicates the probability of acceptance. Indicates the attenuation factor. Indicates the magnitude of the perturbation in the current iteration. Indicates candidate solutions The Monte Carlo objective function value of the Markov chain. Indicates the current solution The Monte Carlo objective function value of the Markov chain; The update formula for the current iteration perturbation amplitude is:

[0011] In the formula, Indicates the magnitude of the perturbation in the current iteration. Indicates the adaptive penalty factor. This indicates the magnitude of the disturbance in the previous iteration.

[0012] In some embodiments of this disclosure, the visualization evaluation processing of the enhanced sample dataset using principal component analysis includes: Obtain a real sample dataset, and merge the real sample dataset and the augmented sample dataset to construct an original data matrix; The original data matrix is ​​centered, and the covariance matrix of the centered original data matrix is ​​calculated. The covariance matrix is ​​subjected to eigenvalue decomposition to obtain multiple eigenvalues ​​and their corresponding eigenvectors; One or more principal components are selected based on the magnitude of the eigenvalues, and their corresponding eigenvectors are used to project the real sample dataset and the augmented sample dataset from the original high-dimensional space to a low-dimensional visualization space composed of the principal components. In the low-dimensional visualization space, the distribution of the real sample dataset and the augmented sample dataset are compared and analyzed to evaluate the quality of the augmented sample dataset.

[0013] In some embodiments of this disclosure, the construction of the multi-strategy enhanced goose flock optimization algorithm includes: An improved Tent chaotic mapping is used to initialize the goose population; Calculate the fitness value of each individual in the goose population, and determine the globally optimal guide individual for the current iteration based on the fitness-distance balance strategy; The positions of individuals in the goose population are subjected to sine and cosine perturbations to enhance local search capabilities; A lens imaging inverse learning strategy is introduced to generate inverse solutions to expand the search space; The goose population is updated based on the globally optimal guiding individual and the individual positions after sine and cosine perturbations, and the above steps are iteratively executed until the convergence condition is met, thus completing the construction of the multi-strategy enhanced goose flock optimization algorithm.

[0014] In some embodiments of this disclosure, determining the globally optimal guiding individual for the current iteration based on the fitness-distance balance strategy includes: Calculate the Euclidean distance between each individual and the current best individual; Normalize the fitness function values ​​and Euclidean distances of all individuals; The comprehensive score is calculated based on the normalized fitness function value and the Euclidean distance. The individual with the highest overall score is selected as the globally optimal guiding individual.

[0015] In some embodiments of this disclosure, the step of evaluating the operating status of the transformer using the constructed transformer operating status evaluation model includes: Obtain real-time oil chromatographic data of the transformer to be evaluated; The real-time oil chromatography data is preprocessed to construct a real-time feature vector that matches the input format of the transformer operating status assessment model; The real-time feature vector is input into the transformer operation status assessment model for multi-class probability prediction processing, and the probability distribution vector of the transformer to be evaluated is output. Each element in the probability distribution vector corresponds to the probability of different operating states of the transformer. The operating states include at least one of the following: normal state, partial discharge, low-energy discharge, arc discharge, low temperature overheating, medium temperature overheating, or high temperature overheating. Based on the probability distribution vector, the operating state with the highest probability value is selected as the final operating state evaluation result of the transformer to be evaluated.

[0016] According to a second aspect of this application, a transformer operating status assessment device is provided, the device comprising: The acquisition module is used to acquire historical oil chromatographic data of dissolved gases in transformer oil. Based on the historical oil chromatographic data, a multidimensional feature system is constructed, which includes fused gas concentration, ratio features and normalization combination. An improved Markov chain Monte Carlo method is used to perform sample augmentation processing on the minority class fault sample set in the multidimensional feature system to construct an augmented sample dataset for model training. The filtering module is used to perform visual evaluation processing on the augmented sample dataset using principal component analysis, and to use the feature importance evaluation mechanism built into the CatBoost classification model to filter the subset of key features in the augmented sample dataset that have reached a preset contribution rate threshold and have passed the evaluation. The first processing module is used to construct a multi-strategy enhanced goose flock optimization algorithm, and with minimizing the model loss function as the optimization objective, the key feature subset and the corresponding training labels are used as input data, and the multi-strategy enhanced goose flock optimization algorithm is used to optimize the hyperparameters of the CatBoost classification model to obtain the optimal hyperparameter combination. The evaluation module is used to retrain the CatBoost classification model based on the key feature subset and the optimal hyperparameter combination to construct a transformer operating status evaluation model, and to use the constructed transformer operating status evaluation model to evaluate the operating status of the transformer.

[0017] In some embodiments of this disclosure, the evaluation module is specifically used for: Obtain real-time oil chromatographic data of the transformer to be evaluated; The real-time oil chromatography data is preprocessed to construct a real-time feature vector that matches the input format of the transformer operating status assessment model; The real-time feature vector is input into the transformer operation status assessment model for multi-class probability prediction processing, and the probability distribution vector of the transformer to be evaluated is output. Each element in the probability distribution vector corresponds to the probability of different operating states of the transformer. The operating states include at least one of the following: normal state, partial discharge, low-energy discharge, arc discharge, low temperature overheating, medium temperature overheating, or high temperature overheating. Based on the probability distribution vector, the operating state with the highest probability value is selected as the final operating state evaluation result of the transformer to be evaluated.

[0018] According to a third aspect of this application, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause a computer to perform the method of the first aspect described above.

[0019] According to a fourth aspect of this application, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method of the first aspect described above.

[0020] By employing the above technical solutions, this application provides a transformer operating status assessment method, device, electronic equipment, and storage medium. Compared with existing technologies, this application can acquire historical oil chromatographic data of dissolved gases in transformer oil, construct a multi-dimensional feature system based on the historical oil chromatographic data that includes fused gas concentration, ratio characteristics, and normalized combinations, and use an improved Markov chain Monte Carlo method to perform sample augmentation processing on the minority class fault sample set in the multi-dimensional feature system to construct an augmented sample dataset for model training; use principal component analysis to perform visualization evaluation processing on the augmented sample dataset, and utilize the feature importance built into the CatBoost classification model. A performance evaluation mechanism is implemented to select a subset of key features whose cumulative contribution rate reaches a preset contribution rate threshold in the enhanced sample dataset. A multi-strategy enhancement goose flock optimization algorithm is constructed, with minimizing the model loss function as the optimization objective. The key feature subset and its corresponding training labels are used as input data, and the hyperparameters of the CatBoost classification model are optimized using the multi-strategy enhancement goose flock optimization algorithm to obtain the optimal hyperparameter combination. Based on the key feature subset and the optimal hyperparameter combination, the CatBoost classification model is retrained to construct a transformer operation status evaluation model, and the transformer operation status evaluation model is used to evaluate the operation status of the transformer.

[0021] The scheme described in this application constructs a multi-dimensional feature system by acquiring historical oil chromatographic data of dissolved gases in transformer oil, which includes integrated gas concentration, ratio characteristics, and normalized combinations. Compared to traditional methods that rely solely on gas proportions or contents to determine fault types, this multi-dimensional feature system more comprehensively reflects transformer operating status information, covers more features from various angles, and helps to more accurately determine fault types, reducing judgment biases caused by single features in fault category overlaps or early abnormal stages.

[0022] An improved Markov chain Monte Carlo method is employed to augment the minority class fault sample set in a multidimensional feature system, constructing an augmented sample dataset for model training. In cases of scarce samples, sample augmentation increases the amount of data available for training, enabling the model to learn richer fault features and reducing the risk of misclassification or omission due to insufficient samples.

[0023] Principal component analysis (PCA) was used to visually evaluate the augmented sample dataset and assess its quality. This ensures the reliability of the sample data used for subsequent training, avoids the impact of sample quality issues on model training performance, and further improves the accuracy of model judgments.

[0024] The CatBoost classification model was initially trained on a qualified augmentation sample dataset. Based on its built-in feature importance assessment mechanism, a subset of key features in the augmentation sample dataset whose cumulative contribution rate reached a preset contribution rate threshold was selected. From numerous features, key features that significantly contribute to fault diagnosis were selected, reducing interference from feature cross-complexity and allowing the model to focus on key features for analysis and judgment, thereby improving diagnostic accuracy.

[0025] A multi-strategy-enhanced goose flocking optimization algorithm is constructed, with minimizing the model loss function as the optimization objective. This algorithm is then used to optimize the hyperparameters of the CatBoost classification model to obtain the optimal hyperparameter combination. By optimizing the hyperparameters, the performance of the CatBoost classification model is further improved, making it more accurate and reliable in handling complex data and identifying fault types, and reducing the possibility of misclassification or missed classification.

[0026] The CatBoost classification model was retrained based on a subset of key features and the optimal hyperparameter combination to construct a transformer operating status assessment model. The model constructed after this series of optimizations integrates the advantages of multidimensional features, high-quality samples, key features, and optimized hyperparameters, effectively addressing the problems of traditional DGA diagnostic techniques and providing a more accurate assessment of transformer operating status.

[0027] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0028] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0029] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 A flowchart illustrating a transformer operating status assessment method provided in an embodiment of this application is shown. Figure 2 This application provides a feature importance cumulative contribution rate curve diagram according to an embodiment of the present application; Figure 3 This application provides a feature importance ranking histogram. Figure 4 A flowchart illustrating another transformer operating status assessment method provided in an embodiment of this application is shown; Figure 5 This paper illustrates a training structure framework diagram of a CatBoost classification model provided in an embodiment of this application. Figure 6 This paper illustrates a transformer fault diagnosis architecture provided in an embodiment of this application. Figure 7 This paper presents a comparative analysis chart of the classification performance of a basic and feature optimization model provided in an embodiment of this application. Figure 8 A schematic diagram of the structure of a transformer operating status assessment device provided in an embodiment of this application is shown. Detailed Implementation

[0031] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of this application, including various details to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0032] The following description, with reference to the accompanying drawings, describes a transformer operation status assessment method, apparatus, electronic device, and storage medium according to embodiments of this application.

[0033] To address the limitations of traditional DGA diagnostic techniques, which primarily rely on gas ratios or concentrations to determine fault types, especially during fault category overlap or early anomaly stages, this approach suffers from limited accuracy and is prone to misjudgment. Furthermore, it is highly dependent on data, making it susceptible to misjudgments or omissions when samples are scarce or features are complex. This application provides a method for evaluating transformer operating conditions, such as... Figure 1 As shown, the method includes: Step 101: Obtain historical oil chromatographic data of dissolved gases in transformer oil. Based on the historical oil chromatographic data, construct a multidimensional feature system that includes fused gas concentration, ratio features and normalization combination. Then, use the improved Markov chain Monte Carlo method to perform sample augmentation processing on the minority class fault sample set in the multidimensional feature system to construct an augmented sample dataset for model training.

[0034] Historical oil chromatographic data may include a variety of characteristic gases (such as...) The concentration values ​​of ( ) and the corresponding fault repair records and operation and maintenance logs.

[0035] To improve model training quality and diagnostic accuracy, before constructing a multidimensional feature system based on historical oil chromatographic data that includes fused gas concentration, ratio features, and normalized combinations, data preprocessing of the historical oil chromatographic data is necessary. This preprocessing mainly includes three parts: data cleaning and normalization, sample labeling and validation strategies, and dataset construction and sample distribution analysis. Specifically, it may include: Historical oil chromatographic data underwent data cleaning, which included: imputation of missing values ​​for non-critical variables using the median method; removal of sample data corresponding to missing values ​​for critical variables; retention of the first record and deletion of duplicate records for duplicate samples; and detection of outliers using a combined box plot and percentile method. For detected outliers, if they matched the corresponding fault label, they were retained; otherwise, they were identified as false anomalies and excluded. Historical oil chromatographic data undergoes sample labeling and validation processing. This process includes: initially labeling sample data from historical oil chromatographic data based on maintenance records, fault repair logs, and expert diagnostic conclusions; and cross-validating the initial labeling results using the three-ratio method. If the initial labeling results are consistent with the three-ratio method validation results, the initial labeling results are used as the final labeling results. If the initial labeling results are inconsistent with the three-ratio method validation results, the initial labeling results are used as the final labeling results, and the three-ratio method validation results are used as auxiliary reference information to correct the final labeling results. A multidimensional feature system was constructed based on historical oil chromatographic data, including fused gas concentration, ratio characteristics, and normalized combinations, comprising: Based on historical oil chromatography data after data cleaning, sample labeling and validation, a multidimensional feature system was constructed, which includes fused gas concentration, ratio features and normalized combinations. The multidimensional feature system is a 32-dimensional multi-level diagnostic feature system.

[0036] In this embodiment, the acquired historical oil chromatographic data undergoes data cleaning and normalization to eliminate noise and anomalies. This may include: traversing all historical oil chromatographic data samples; for missing values ​​of non-critical features (such as minor gas concentrations or non-core ratios), imputation can be performed using the median of the fault category to which the feature belongs. For missing values ​​of critical features (such as the concentrations of the five characteristic gases), the sample is directly removed to ensure the integrity of the core data.

[0037] Check the historical oil chromatography data for duplicate samples with identical features and labels. For duplicate samples, retain only the earliest record (i.e., the first record) based on its timestamp and delete the rest to avoid data redundancy that could bias the model training.

[0038] Outlier detection can be performed using a combined strategy of box plots and percentile analysis. First, the quartiles (Q1, Q3) and interquartile range (IQR) of each feature can be calculated, and features in the range [Q1 - 1.5] can be identified. IQR, Q3 + 1.5 Data points outside the IQR interval are considered potential outliers. Then, a second confirmation can be made using percentile methods (such as the 99th percentile). For detected outliers, a second judgment can be made: if the outlier is mechanistically consistent with its corresponding fault label (e.g., in the case of an arc fault)... If the concentration is extremely high, it is considered a key abnormal sample reflecting the real fault and is retained; if the abnormal value is significantly inconsistent with the fault label (for example, under normal operating conditions), it is considered a key abnormal sample reflecting the real fault and is retained. If the concentration is abnormally high, it is considered false anomaly data and is excluded.

[0039] In this embodiment, the historical oil chromatographic data after cleaning and standardization is labeled and verified for fault categories. This may include: initially labeling the samples based on maintenance records, fault repair logs, and expert diagnostic conclusions, that is, assigning an initial fault type label (such as: normal, partial discharge, low-energy discharge, arc discharge, low-temperature overheating, medium-temperature overheating, high-temperature overheating, etc.) to each sample, and then using the three-ratio method for cross-validation. The three-ratio method can be used to calculate the fault type judgment based on rules for all samples.

[0040] The initial labeling results can be compared with the empirical rule results of the three-ratio method. If they are consistent, the initial fault type label can be directly used as the final labeling result.

[0041] If the two are inconsistent, the principle of "original label first, rule-based correction" is followed. That is, the initial fault type label is still used as the final labeling result, but the judgment result of the three-ratio method is used as auxiliary information to correct the final labeling result.

[0042] After completing the cleaning, outlier identification, and normalization of historical oil chromatography data, a 32-dimensional multi-level diagnostic feature system (i.e., a 32-dimensional multi-level diagnostic feature system) can be systematically designed and constructed based on the concentration characteristics, combination relationships, and generation mechanisms of dissolved gases under different fault modes, as input for subsequent model training.

[0043] The framework of the traditional Markov Chain Monte Carlo (MCMC) method is to assume the current state is... From the suggested distribution Candidate samples y are generated from the data, and the acceptance probability is used to determine the appropriate sample. Determine whether to accept the state transition. When using the Metropolis-Hastings criterion, the probability of acceptance is defined as:

[0044] In the formula, Indicates the probability of acceptance. Indicates the current state. Indicates the candidate state. Let represent the objective function for constructing candidate states. This represents the objective function for constructing the current state. Indicates from candidate state Move to the current state The probability density, Indicates the current state Next, generate candidate states. The probability density.

[0045] like At that time, Accept y as the next state Otherwise, leave the original state unchanged.

[0046] The limited number of samples for transformer fault data, especially for rare faults such as high-energy arcs, leads to class imbalance in the dataset. To address this issue, this embodiment employs an improved MCMC method to oversample and enhance minority class samples.

[0047] This improved MCMC method, based on the traditional Metropolis-Hastings algorithm, introduces a dynamic decay factor and an adaptive penalty factor to enhance the quality and distribution rationality of the generated samples. The specific process is as follows: Step 1: Based on the minority class of fault samples, construct a Markov chain Monte Carlo objective function to evaluate the sample quality, and determine the acceptance probability based on the Markov chain Monte Carlo objective function. The acceptance probability is determined by introducing a decay factor and an adaptive penalty factor that dynamically change with the iteration rounds to improve the quality and distribution rationality of the generated samples. Step 2: Starting from the current sample state, generate candidate samples using the suggested distribution; Step 3: Based on the acceptance probability, determine whether to accept the candidate sample; Repeat steps 2-3 until a preset number of augmented sample datasets are generated. The augmented sample dataset consists of a preset number of accepting candidate samples. To avoid candidate solutions deviating too far from the target distribution in the initial sampling stage, which could lead to the incorrect acceptance of anomalous samples, a decay factor is introduced based on the traditional Metropolis-Hastings acceptance probability, and the following iteratively adjusted acceptance criterion (i.e., the formula for calculating the acceptance probability) is constructed:

[0048] In the formula, Indicates the probability of acceptance. Indicates the attenuation factor. Indicates the magnitude of the perturbation in the current iteration. Indicates candidate solutions The Monte Carlo objective function value of the Markov chain. Indicates the current solution The Monte Carlo objective function value of the Markov chain; To further suppress the spread of bias caused by early anomalous sampling, an adaptive penalty factor (i.e., a fixed attenuation coefficient) is introduced. The sampling perturbation intensity is dynamically adjusted through an exponential decay mechanism, and its update formula is shown below:

[0049] In the formula, Indicates the magnitude of the perturbation in the current iteration. Indicates the adaptive penalty factor. This indicates the magnitude of the disturbance in the previous iteration.

[0050] Step 102: Use principal component analysis to perform visual evaluation of the augmented sample dataset, and use the feature importance evaluation mechanism built into the CatBoost classification model to select the key feature subsets in the augmented sample dataset whose cumulative contribution rate reaches the preset contribution rate threshold.

[0051] In the embodiments of this application, during the training process, in order to monitor the distribution changes of generated samples and real samples in the low-dimensional space in real time, principal component analysis (PCA) can be used after each iteration to compare the distribution of generated data and real data, so as to comprehensively evaluate the performance of the generative model. Specifically, this may include: Obtain the real sample dataset, merge the real sample dataset and the augmented sample dataset to construct the original data matrix; The original data matrix is ​​centered, and the covariance matrix of the centered original data matrix is ​​calculated. The covariance matrix is ​​subjected to eigenvalue decomposition to obtain multiple eigenvalues ​​and their corresponding eigenvectors; One or more principal components are selected based on the magnitude of the eigenvalues, and their corresponding eigenvectors are used to project the real sample dataset and the augmented sample dataset from the original high-dimensional space to a low-dimensional visualization space composed of principal components. In this low-dimensional visualization space, the distribution of the real sample dataset and the augmented sample dataset can be compared and analyzed to evaluate the quality of the augmented sample dataset.

[0052] In this embodiment, the original data can first be organized into an n×p original data matrix, where n is the number of samples and p is the number of features. The covariance matrix of the centered data is then calculated. :

[0053] In the formula, Represents the covariance matrix. This represents the transpose of the original data matrix. This represents the original data matrix, and n represents the number of samples.

[0054] The covariance matrix is ​​subjected to eigenvalue decomposition, and the specific formula is shown below:

[0055] In the formula, The eigenvectors (i.e., principal component directions) of the covariance matrix are represented. Represents the corresponding eigenvalue. This represents the eigenvector.

[0056] In this embodiment of the application, the CatBoost classification model embeds a feature importance evaluation module, which can automatically evaluate the impact of each feature on the objective function during training. Each time a decision tree is constructed, CatBoost calculates the decrease in loss function for each feature across all nodes used for splitting in the decision tree model. In terms of its characteristic importance indicators It can be represented as:

[0057] In the formula, This represents the importance index of the Zth feature. This represents the total number of decision trees included in the CatBoost model. Indicates the feature used in the t-th decision tree The set of all nodes to which node splitting occurs. This indicates that the nth node of the t-th tree uses a feature. The gain obtained after splitting is the decrease in the objective function at that node.

[0058] After calculating the feature importance index, the cumulative contribution rate is typically used to determine the selection threshold for key features. Specifically, all features are sorted from highest to lowest importance index value, and their cumulative contribution rates are calculated. The first Z features are considered when their cumulative contribution rate first reaches the set threshold. The minimum feature set of is expressed by the formula:

[0059] In the formula, Indicates the cumulative contribution rate. Let Z represent the importance index of the Z-th feature, and M represent the total number of initial features. This indicates the set cumulative contribution rate threshold. Commonly used thresholds are 0.90 or 0.95, which means selecting the top Z features with a cumulative contribution rate of 90% or 95%.

[0060] Figure 2 The cumulative contribution rate curve of feature importance is shown, indicating that the top 15 features can explain most of the data variability, and the effect of further increasing the number of features on improving the contribution rate gradually decreases. Figure 3 The bar chart shows the ranking of feature importance, illustrating the significance of each feature in the model.

[0061] Step 103: Construct a multi-strategy enhanced goose flock optimization algorithm, and take minimizing the model loss function as the optimization objective. Use the key feature subset and the corresponding training labels as input data, and use the multi-strategy enhanced goose flock optimization algorithm to optimize the hyperparameters of the CatBoost classification model to obtain the optimal hyperparameter combination.

[0062] For the embodiments of this application, the operation process of the Goose Optimization Algorithm (GOOSE) can be summarized into five key stages: (1) Population initialization: Initialization phase: Setting up a flock of geese ,in, (i < m) is a parameter vector representing the parameter combination for the i-th goose, where m represents the total number of individuals in the population. This parameter combination is denoted as . Where n is the number of parameters, the goal is to maximize the accuracy of the decision tree classifier.

[0063] In the formula, Let represent the objective function of the parameter combination for the i-th goose. This indicates the number of correctly classified fault samples. This indicates the number of faulty samples that were misclassified. This represents the number of correctly classified normal samples. This indicates the number of normal samples that were misclassified.

[0064] (2) Information generation and transmission stage: Use a random number between [0,1] as the information and transmission time that the i-th goose transmits to the group. Total time for:

[0065] In the formula, This represents the total time taken, where N represents the total number of individuals in the flock. This represents the information and time that the i-th goose transmits to the group.

[0066] (3) Search direction determination stage: The search direction of a flock of geese is determined by information transmission mechanisms and the distance between individuals. This is expressed as the Euclidean distance between the current individual and the globally optimal individual:

[0067] In the formula, This represents the Euclidean distance between the current individual and the globally optimal individual. This represents the position vector of the i-th individual. This represents the globally optimal position vector.

[0068] (4) Parameter update stage: The parameter vector is updated based on the average information transmission time and individual distance. A dynamic adjustment factor is introduced during the update process to enhance the adaptability of the algorithm. The update formula is as follows:

[0069] In the formula, This represents the position vector of the i-th goose at the (t+1)-th iteration. Let represent the position vector of the i-th goose at the t-th iteration. Represents the random disturbance term. Represents the globally optimal position vector. Indicates control factor. This represents the information and time that the i-th goose transmits to the group.

[0070] (5) Local perturbation and convergence mechanism stage: To avoid getting trapped in local optima and to increase the diversity of the parameter space, a decay factor is introduced. The amplitude of the disturbance is controlled and dynamically changed with the number of iterations:

[0071] In the formula, Indicates the dynamic decay factor. This represents the initial attenuation coefficient. Indicates the current iteration number. This indicates the maximum number of iterations.

[0072] Although the GOOSE algorithm demonstrates strong global search capabilities and robustness in continuous optimization problems, it still has certain performance limitations when dealing with complex optimization tasks involving high dimensions, multiple peaks, and dense distributions of local maxima. Therefore, an improved Goose Optimization Algorithm (IMGOA) framework is proposed, integrating the approaches of four strategies. The improved flowchart is shown below. Figure 4 As shown.

[0073] Constructing a multi-strategy-enhanced goose flock optimization algorithm may specifically include: Step 1: Initialize the goose population using an improved Tent chaotic mapping; In this step, in order to improve the uneven distribution of the initial population in the goose flock algorithm and thus enhance the global search capability, Tent chaotic mapping is used to generate the initial goose flock population.

[0074] Assuming initial value [0,1], the values ​​of the preceding and following iterations are related, then when When the value is 0 or close to 0, the mapping value will either get stuck in a loop or become reversible, where k is the number of iterations.

[0075]

[0076] In the formula, This represents the mapping value in the k-th iteration. This represents the next state mapping value.

[0077] exist When the value is 0, a random perturbation term is introduced to preserve its chaotic properties:

[0078] In the formula, This represents the mapping value in the k-th iteration. This represents the adjustment parameter, which is a random number between 0 and 1. This represents the mapping value in the (k+1)th iteration.

[0079] From the above formula, it can be seen that if the flock of geese is Then the application of chaotic variables in the solution space is represented as:

[0080] in, Indicates the location of an individual goose in the flock. Represents chaotic sequence values. This represents the upper bound of the solution space. This represents the lower bound of the solution space.

[0081] Step 2: Calculate the fitness value of each individual in the goose population, and determine the globally optimal guide individual for the current iteration based on the fitness-distance balance strategy; Specifically, determining the globally optimal guiding individual for the current iteration based on a fitness-distance balance strategy may include: Calculate any individual in the current population With the current best individual The Euclidean distance between them is given by the following formula:

[0082] In the formula, Represents any individual in the current population. With the current best individual The Euclidean distance between them This represents the parameter value of the i-th individual in the j-th dimension. This represents the parameter value of the current optimal individual in the j-th dimension, and n represents the number of dimensions in the problem.

[0083] The fitness function values ​​of all individuals and European distance Normalization is performed to balance the influence of the two indicators. The formula for fitness normalization is:

[0084] In the formula, This represents the normalized fitness function value. The fitness function value represents the individual's fitness. This represents the minimum fitness value of the current population. This represents the maximum fitness value of the current population. This represents a small constant, used to prevent the denominator from being zero.

[0085] The formula for distance normalization is:

[0086] In the formula, This represents the normalized Euclidean distance. Represents any individual in the current population. With the current best individual The Euclidean distance between them This represents the maximum value of the current population distance. This represents the minimum distance value in the current population. This represents a small constant, used to prevent the denominator from being zero.

[0087] Based on the normalized fitness function value and Euclidean distance, the FDB comprehensive score is calculated as follows:

[0088] In the formula, This represents the overall score of the i-th individual. Indicates the weighting factor. This represents the normalized fitness function value. This represents the normalized Euclidean distance.

[0089] After calculating the score for all individuals in the population, the individual with the highest overall score can be selected. As the globally optimal guiding entity, update the globally optimal position:

[0090] In the formula, Indicates the globally optimal position. This represents the optimal position of an individual.

[0091] To further improve the algorithm's adaptability, this application designs a dynamic adaptive weight factor. The algorithm adjusts its weights in real time based on the search status during iterations to achieve a reasonable balance between broad-area exploration in the early stages and localized development in the later stages. The dynamic weight update formula is as follows:

[0092] In the formula, This represents a dynamically adaptive weighting factor. Indicates the current iteration number. Modulo operation is represented. This indicates the maximum number of iterations.

[0093] Step 3: Perform sine and cosine perturbations on the positions of individuals within the goose flock to enhance local search capabilities; In this step, the sine and cosine perturbation update formulas are as follows:

[0094] In the formula, This represents the position of the i-th individual in the (t+1)-th generation. This represents the position of the i-th individual in generation t. The magnitude factor, which controls the search range, typically decreases with the number of iterations. Let represent the perturbation angle factor, which follows a uniform distribution in the interval [0, 2π]. This represents the pulling strength factor between an individual and the optimal solution. This represents the random number [0,1] used to select the sine or cosine function perturbation. This represents the current optimal solution.

[0095] Step 4: Introduce a lens imaging reverse learning strategy to generate a reverse solution to expand the search space; The goose population is updated based on the globally optimal guiding individual and the individual positions after sine and cosine perturbations, and steps 2-4 are iteratively executed until the convergence condition is met, thus completing the construction of the multi-strategy enhanced goose flock optimization algorithm.

[0096] In this step, the lens imaging back-learning strategy can expand the search space and improve global search capabilities by generating virtual symmetric solutions based on the relationship between individuals and the optimal solution. Back-learning position calculation formula:

[0097] In the formula, This represents the lens imaging position of an individual in the d-th dimension. This indicates the position of the lens focal point, which is typically taken as the midpoint of the search space or the position of the current optimal solution. Represents the reflection coefficient. This represents the individual's current position in the d-th dimension.

[0098] If the search space is To prevent candidate solutions generated by the lens imaging reverse learning strategy from exceeding the legal range of the search space, the newly generated solutions undergo the following boundary processing:

[0099] In the formula, This represents the lens imaging position of an individual in the d-th dimension. This represents the lower bound of the search space. It represents the upper bound of the search space.

[0100] The CatBoost classification model uses the difference between predicted and actual probabilities to fit a loss function, and progressively optimizes the model's output probabilities by building multiple weak learners to minimize the loss function. The optimization objective of the loss function is as follows:

[0101] In the formula, Let N represent the optimization objective of the loss function, and C represent the total number of samples and the total number of fault categories. The true label representing sample i belonging to class C (using one-hot encoding) is used. This represents the probability that the model predicts sample i belongs to class C.

[0102] In the process of constructing the decision tree, the CatBoost classification model aims to minimize the aforementioned loss function by continuously selecting appropriate features and split points to effectively reduce the loss function. Specifically, for specific features... For the corresponding split point s, the gain of the objective function is defined as the decrease in the loss function before and after the split, which can be expressed as:

[0103] In the formula, Representation of features and the gain objective function of its corresponding split point s, This represents the set of all samples in the node before the split. This represents the set of samples corresponding to the leftmost child node after partitioning according to the splitting features and split points. This represents the set of samples corresponding to the right child nodes after partitioning according to the splitting features and split points. This represents the first-order gradient of the loss function with respect to the current model's predictions. This represents the second gradient of the loss function with respect to the current model's predictions. This represents the regularization parameters used to prevent overfitting in the tree model. The training structure framework of the model is as follows: Figure 5 As shown.

[0104] For the CatBoost classification model, the hyperparameter settings directly affect the transformer fault diagnosis results. CatBoost has multiple hyperparameters, and their settings involve various aspects, including controlling the model's fitting ability, convergence speed, robustness, and expressive power. Appropriate hyperparameter configuration can significantly improve the model's performance in high-dimensional, multi-class, and imbalanced transformer fault identification tasks. To better optimize the model, an improved goose herding optimization algorithm can be used to tune the hyperparameters of the CatBoost classification model, aiming to find the optimal hyperparameter combination.

[0105] The CatBoost classification model, as an efficient gradient boosting tree model, has multiple hyperparameters. The settings of these hyperparameters directly affect the model's fitting ability, robustness, and convergence speed. To improve the overall performance of the CatBoost classification model in the transformer fault identification task, four key hyperparameters (i.e., the optimal hyperparameter combination) were selected for optimization: maximum number of iterations, learning rate, maximum tree depth, and randomness.

[0106] Step 104: Retrain the CatBoost classification model based on the key feature subset and the optimal hyperparameter combination to construct a transformer operation status assessment model, and use the transformer operation status assessment model to assess the operation status of the transformer.

[0107] For the embodiments of this application, in order to achieve efficient modeling and accurate classification of transformer fault data with multiple categories, imbalanced samples, and complex feature coupling, an intelligent diagnostic model based on the fusion of the CatBoost classification model and the goose flocking optimization algorithm with multi-strategy enhancement is proposed. The transformer fault diagnosis architecture is as follows: Figure 6 As shown, it specifically includes the following four stages: (1) Data preprocessing and feature construction: An initial feature set is constructed based on historical operating samples collected from dissolved gas analysis of transformer oil. The dataset includes typical fault gases and comprises 32 original features. During preprocessing, a min-max normalization method is used to linearly scale the features, uniformly mapping them to the [0,1] interval to avoid the impact of inconsistent dimensions on the model learning process. To address the problem of imbalanced class distribution, an improved MCMC method is used to oversample minority class samples, enhancing the model's sensitivity and robustness to rare faults.

[0108] (2) Preliminary training and feature selection of the CatBoost model: Based on the normalized and augmented data, the CatBoost classifier can be used for benchmark modeling. CatBoost has a built-in feature importance evaluation mechanism that can quantify the contribution of each feature to the classification task. Features are ranked by importance according to their cumulative contribution to the decrease in the objective function during training, and a subset of key features is selected using a cumulative contribution rate threshold method. Features with the top 95% cumulative contribution rate are selected to build a simplified model. The final feature dimension is compressed to approximately 15 features.

[0109] (3) Hyperparameter optimization driven by the IMGOA algorithm: An improved multi-strategy goose optimization algorithm (IMGOA) is introduced to jointly optimize the core hyperparameters of CatBoost. IMGOA utilizes mechanisms such as Tent mapping initialization, fitness-distance balance search, sine and cosine perturbation, and lens imaging back learning to select the minimum mean squared error loss function as the optimization objective, ensuring that the final hyperparameter combination achieves the best balance between accuracy and generalization ability.

[0110] (4) Model retraining and diagnostic deployment: After obtaining the optimal combination of hyperparameters output by the IMGOA algorithm, it is applied to the CatBoost classifier to remodel the training set and construct the final IMGOA-CatBoost integrated diagnostic model (i.e., transformer operating status assessment model).

[0111] To effectively evaluate the model, three test functions were selected for analysis.

[0112] (1)Shifted Sphere Function The Shifted Sphere Function (F1) is a widely used benchmark function in the field of optimization algorithms. It is characterized by its simplicity, continuity, unimodality, and convexity, making it a fundamental tool for evaluating the performance of optimization algorithms. It is based on the standard sphere function, shifting the global optimum from the origin to a new position. Its mathematical expression is:

[0113] In the formula, This represents the objective function value, where n represents the number of dimensions in the problem. Represents the components of each dimension.

[0114] (2) Schwefel's Problem 1.2 Schwefel's Problem 1.2 is often used to evaluate the performance of optimization algorithms when dealing with dependencies between variables. It exhibits unimodality and indivisibility, and its mathematical expression is:

[0115] In the formula, The objective function value is represented by d, and the problem dimension is represented by d. Represents the components of each dimension.

[0116] (3) Schwefel's Problem 2.21 Schwefel's Problem 2.21 is a standard benchmark function used to evaluate the performance of global optimization algorithms. This function is defined as the maximum absolute value of all elements in the decision variable vector. Due to its simple structure and well-defined properties, Schwefel's Problem 2.21 is frequently used as a fundamental benchmark function to assess the basic convergence ability and efficiency of optimization algorithms in finding the optimal solution. It is particularly suitable for verifying whether an algorithm can quickly and accurately locate the global optimum without interference from local optima. Its mathematical expression is:

[0117] In the formula, Represents the objective function value. Represents the components of each dimension.

[0118] Figure 7The confusion matrices of the base model and this model are compared. The base model has lower classification accuracy in some fault categories, particularly in the "medium-temperature overheating" and "low-energy discharge" categories, where misclassification is significant. In contrast, this model significantly improves the accuracy for all categories, especially in the "partial discharge" and "arc discharge combined with overheating" categories, where the accuracy reaches 100%.

[0119] In this embodiment of the disclosure, the transformer's operating status is evaluated using the constructed transformer operating status evaluation model, which may specifically include: Obtain real-time oil chromatographic data of the transformer to be evaluated; Real-time oil chromatography data are preprocessed to construct real-time feature vectors that match the input format of the transformer operating status assessment model; The real-time feature vector is input into the transformer operation status assessment model for multi-class probability prediction processing, and the output is the probability distribution vector of the transformer to be evaluated. Each element in the probability distribution vector corresponds to the probability of different operating states of the transformer. The operating states include at least one of the following: normal state, partial discharge, low-energy discharge, arc discharge, low temperature overheating, medium temperature overheating or high temperature overheating. Based on the probability distribution vector, the operating state with the highest probability value is selected as the final operating state evaluation result of the transformer to be evaluated.

[0120] In this embodiment, an online oil chromatography monitoring device can first be installed on the transformer in the substation. This device can collect dissolved gas data in the transformer oil in real time. Through a data acquisition system connected to the monitoring device, real-time oil chromatography data is automatically acquired at regular intervals (e.g., hourly). The collected real-time oil chromatography data may include hydrogen (… ), methane ( ), ethane ( ), ethylene ( ), acetylene ( ), carbon monoxide ( ),carbon dioxide( Concentration information for various gases, including )

[0121] The system can check for missing values ​​in the collected real-time oil chromatogram data. If missing values ​​are found, for non-critical gas variables, the median of the same type is used for imputation; if critical variables (such as the concentration of acetylene, a gas that has a significant impact on fault diagnosis) are missing, the data set is discarded. For duplicate samples, the data from the first record is retained based on the timestamp. A combined strategy of box plots and percentile method is used for outlier detection. Critical outlier samples that are consistent with the fault label (e.g., an abnormally high acetylene concentration is detected when a partial discharge fault is known to exist) are retained; false outlier data that is inconsistent with the operating status (e.g., random abnormally high values ​​caused by monitoring device malfunctions) are excluded.

[0122] Based on the concentration characteristics, combination relationships, and generation mechanisms of dissolved gases under different fault modes, a 32-dimensional multi-level diagnostic feature system is constructed, which may include gas concentration characteristics (such as the absolute concentration values ​​of various gases), ratio characteristics, local normalization characteristics (values ​​after local normalization of the concentrations of different gases), and combination ratio characteristics (ratios calculated by combining the concentrations of multiple gases).

[0123] The constructed 32-dimensional multi-level diagnostic feature system can be linearly scaled using a min-max normalization method, uniformly mapping each feature value to the [0,1] interval. This normalization process avoids the impact of inconsistent dimensions on the model learning process.

[0124] The transformer operating status assessment model constructed in this application is an IMGOA-CatBoost integrated diagnostic model. During the training phase, the model has been fully trained using a large amount of historical transformer oil chromatographic data, and the key hyperparameters of the CatBoost classification model (such as maximum number of iterations, learning rate, maximum tree depth, and stochastic intensity) have been optimized through the multi-strategy enhanced IMGOA algorithm to achieve the best balance between accuracy and generalization ability.

[0125] The preprocessed real-time feature vector is input into the trained IMGOA-CatBoost model, which performs multi-class probability prediction based on the input feature vector. The model outputs a probability distribution vector, where each element corresponds to the probability of different operating states of the transformer. In this embodiment, the operating states include normal state, partial discharge, low-energy discharge, arc discharge, low-temperature overheating, medium-temperature overheating, and high-temperature overheating. For example, the output probability distribution vector may be [0.85, 0.05, 0.03, 0.02, 0.03, 0.01, 0.01], corresponding to the probabilities of normal state, partial discharge, low-energy discharge, arc discharge, low-temperature overheating, medium-temperature overheating, and high-temperature overheating, respectively.

[0126] For the probability distribution vector [0.85, 0.05, 0.03, 0.02, 0.03, 0.01, 0.01] of the above output, the probability value of 0.85 corresponding to the normal state is the largest. Therefore, the final operating state evaluation result of the transformer to be evaluated is determined to be the normal state.

[0127] In summary, based on the transformer operating status assessment method provided in this application, compared with existing technologies, this application can obtain historical oil chromatographic data of dissolved gases in transformer oil, construct a multi-dimensional feature system based on the historical oil chromatographic data including fused gas concentration, ratio characteristics, and normalized combinations, and use an improved Markov chain Monte Carlo method to perform sample augmentation processing on the minority class fault sample set in the multi-dimensional feature system to construct an augmented sample dataset for model training; principal component analysis is used to perform visualization evaluation processing on the augmented sample dataset, and the feature importance evaluation mechanism built into the CatBoost classification model is used to screen the evaluation... The algorithm identifies a subset of key features whose cumulative contribution rate in the qualified augmented sample dataset reaches a preset contribution rate threshold. It then constructs a multi-strategy augmentation goose-flock optimization algorithm, using the key feature subset and corresponding training labels as input data. This algorithm optimizes the hyperparameters of the CatBoost classification model to obtain the optimal hyperparameter combination. Based on the key feature subset and the optimal hyperparameter combination, the CatBoost classification model is retrained to construct a transformer operating status assessment model. Finally, this model is used to assess the operating status of the transformer.

[0128] The scheme described in this application constructs a multi-dimensional feature system by acquiring historical oil chromatographic data of dissolved gases in transformer oil, which includes integrated gas concentration, ratio characteristics, and normalized combinations. Compared to traditional methods that rely solely on gas proportions or contents to determine fault types, this multi-dimensional feature system more comprehensively reflects transformer operating status information, covers more features from various angles, and helps to more accurately determine fault types, reducing judgment biases caused by single features in fault category overlaps or early abnormal stages.

[0129] An improved Markov chain Monte Carlo method is employed to augment the minority class fault sample set in a multidimensional feature system, constructing an augmented sample dataset for model training. In cases of scarce samples, sample augmentation increases the amount of data available for training, enabling the model to learn richer fault features and reducing the risk of misclassification or omission due to insufficient samples.

[0130] Principal component analysis (PCA) is used to visually evaluate the augmented sample dataset and assess its quality. This ensures the reliability of the sample data used for subsequent training, avoids the impact of sample quality issues on model training performance, and further improves the accuracy of model judgments.

[0131] The CatBoost classification model was initially trained on a qualified augmentation sample dataset. Based on its built-in feature importance assessment mechanism, a subset of key features in the augmentation sample dataset whose cumulative contribution rate reached a preset contribution rate threshold was selected. From numerous features, key features that significantly contribute to fault diagnosis were selected, reducing interference from feature cross-complexity and allowing the model to focus on key features for analysis and judgment, thereby improving diagnostic accuracy.

[0132] A multi-strategy-enhanced goose flocking optimization algorithm is constructed, with minimizing the model loss function as the optimization objective. This algorithm is then used to optimize the hyperparameters of the CatBoost classification model to obtain the optimal hyperparameter combination. By optimizing the hyperparameters, the performance of the CatBoost classification model is further improved, making it more accurate and reliable in handling complex data and identifying fault types, and reducing the possibility of misclassification or missed classification.

[0133] The CatBoost classification model was retrained based on a subset of key features and the optimal hyperparameter combination to construct a transformer operating status assessment model. The model constructed after this series of optimizations integrates the advantages of multidimensional features, high-quality samples, key features, and optimized hyperparameters, effectively addressing the problems of traditional DGA diagnostic techniques and providing a more accurate assessment of transformer operating status.

[0134] Based on the above Figure 1 The specific implementation of the method shown in this embodiment provides a transformer operating status assessment device, such as... Figure 8 As shown, the device includes: an acquisition module 31, a screening module 32, a first processing module 33, and an evaluation module 34; The acquisition module 31 is used to acquire historical oil chromatographic data of dissolved gases in transformer oil, construct a multi-dimensional feature system based on the historical oil chromatographic data that includes fused gas concentration, ratio features and normalization combination, and use an improved Markov chain Monte Carlo method to perform sample augmentation processing on the minority class fault sample set in the multi-dimensional feature system to construct an augmented sample dataset for model training. The screening module 32 is used to perform visual evaluation processing on the augmented sample dataset using principal component analysis, and to use the feature importance evaluation mechanism built into the CatBoost classification model to screen the subset of key features in the augmented sample dataset that have reached a preset contribution rate threshold and are qualified for evaluation. The first processing module 33 is used to construct a multi-strategy enhanced goose flock optimization algorithm, and with minimizing the model loss function as the optimization objective, the key feature subset and the corresponding training labels are used as input data, and the multi-strategy enhanced goose flock optimization algorithm is used to optimize the hyperparameters of the CatBoost classification model to obtain the optimal hyperparameter combination. Evaluation module 34 is used to retrain the CatBoost classification model based on the key feature subset and the optimal hyperparameter combination to construct a transformer operating status evaluation model, and to use the constructed transformer operating status evaluation model to evaluate the operating status of the transformer.

[0135] In specific application scenarios, the device includes: a second processing module 35; The second processing module 35 is used to perform data cleaning on the historical oil chromatographic data. The data cleaning includes: imputing missing values ​​of non-critical variables in the historical oil chromatographic data using a median method; removing sample data corresponding to missing values ​​of critical variables in the historical oil chromatographic data; retaining the first record of duplicate sample data based on its timestamp and deleting other duplicate records; and detecting outliers in the historical oil chromatographic data using a combined box plot and percentile method strategy. For detected outliers, if they match the corresponding fault label, they are retained; if they do not match the corresponding fault label, they are determined to be pseudo-outlier data and excluded. The historical oil chromatographic data is subjected to sample labeling and verification processing. The sample labeling and verification processing includes: initially labeling the sample data in the historical oil chromatographic data based on operation and maintenance records, fault repair logs, and expert diagnostic conclusions; and cross-validating the initial labeling results using the three-ratio method. If the initial labeling results are consistent with the three-ratio method verification results, the initial labeling results are taken as the final labeling results. If the initial labeling results are inconsistent with the three-ratio method verification results, the initial labeling results are taken as the final labeling results, and the three-ratio method verification results are used as auxiliary reference information to correct the final labeling results.

[0136] In specific application scenarios, the acquisition module 31 can be used to construct a multi-dimensional feature system based on the historical oil chromatography data after data cleaning, sample labeling and verification processing. This system includes fused gas concentration, ratio features and normalized combinations. The multi-dimensional feature system is a 32-dimensional multi-level diagnostic feature system.

[0137] In a specific application scenario, the acquisition module 31 can be used to construct a Markov chain Monte Carlo objective function for evaluating sample quality based on the minority class fault sample set, and to determine the acceptance probability based on the Markov chain Monte Carlo objective function. The acceptance probability is achieved by introducing a decay factor and an adaptive penalty factor that dynamically change with the iteration rounds to improve the quality and distribution rationality of the generated samples. Starting from the current sample state, candidate samples are generated through a suggested distribution. Based on the acceptance probability, it is determined whether to accept the candidate sample. Repeat the above steps until a preset number of augmented sample datasets are generated, wherein the augmented sample datasets consist of a preset number of candidate samples. The formula for calculating the acceptance probability is:

[0138] In the formula, Indicates the probability of acceptance. Indicates the attenuation factor. Indicates the magnitude of the perturbation in the current iteration. Indicates candidate solutions The Monte Carlo objective function value of the Markov chain. Indicates the current solution The Monte Carlo objective function value of the Markov chain; The update formula for the current iteration perturbation amplitude is:

[0139] In the formula, Indicates the magnitude of the perturbation in the current iteration. Indicates the adaptive penalty factor. This indicates the magnitude of the disturbance in the previous iteration.

[0140] In specific application scenarios, the filtering module 32 can be used to obtain a real sample dataset, merge the real sample dataset and the enhanced sample dataset to construct an original data matrix; The original data matrix is ​​centered, and the covariance matrix of the centered original data matrix is ​​calculated. The covariance matrix is ​​subjected to eigenvalue decomposition to obtain multiple eigenvalues ​​and their corresponding eigenvectors; One or more principal components are selected based on the magnitude of the eigenvalues, and their corresponding eigenvectors are used to project the real sample dataset and the augmented sample dataset from the original high-dimensional space to a low-dimensional visualization space composed of the principal components. In the low-dimensional visualization space, the distribution of the real sample dataset and the augmented sample dataset are compared and analyzed to evaluate the quality of the augmented sample dataset.

[0141] In specific application scenarios, the first processing module 33 can be used to initialize a goose population using an improved Tent chaotic mapping; calculate the fitness value of each individual in the goose population, and determine the globally optimal guiding individual for the current iteration based on a fitness-distance balance strategy; apply sine and cosine perturbations to the positions of individuals in the goose population to enhance local search capabilities; and introduce a lens imaging inverse learning strategy to generate inverse solutions to expand the search space. The goose population is updated based on the globally optimal guiding individual and the individual positions after sine and cosine perturbations, and the above steps are iteratively executed until the convergence condition is met, thus completing the construction of the multi-strategy enhanced goose flock optimization algorithm.

[0142] In a specific application scenario, the first processing module 33 can be used to calculate the Euclidean distance between each individual and the current best individual; Normalize the fitness function values ​​and Euclidean distances of all individuals; The comprehensive score is calculated based on the normalized fitness function value and the Euclidean distance. The individual with the highest overall score is selected as the globally optimal guiding individual.

[0143] In a specific application scenario, the evaluation module 34 can be used to acquire real-time oil chromatographic data of the transformer to be evaluated; preprocess the real-time oil chromatographic data to construct a real-time feature vector that matches the input format of the transformer operating status evaluation model; input the real-time feature vector into the transformer operating status evaluation model for multi-class probability prediction processing, and output the probability distribution vector of the transformer to be evaluated. Each element in the probability distribution vector corresponds to the probability of different operating states of the transformer. The operating states include at least one of the following: normal state, partial discharge, low-energy discharge, arc discharge, low-temperature overheating, medium-temperature overheating, or high-temperature overheating; and select the operating state with the highest probability value as the final operating status evaluation result of the transformer to be evaluated based on the probability distribution vector.

[0144] It should be noted that other corresponding descriptions of the functional units involved in the transformer operating status assessment device provided in this embodiment can be found in [reference needed]. Figure 1 The corresponding description in [the document] will not be repeated here.

[0145] Based on the above, Figure 1 Accordingly, this embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method. Figure 1 The method shown.

[0146] Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause a computer device (such as personal computer, server, or network device, etc.) to execute the methods of various implementation scenarios of this application.

[0147] Based on the above, Figure 1 The method shown, and Figure 8 To achieve the above objectives, the present application also provides an electronic device, comprising a storage medium and a processor; the storage medium for storing a computer program; and the processor for executing the computer program to implement the above-described virtual device embodiments. Figure 1 The method shown.

[0148] Optionally, the aforementioned physical devices may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB interfaces, card reader interfaces, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Wi-Fi interfaces), etc.

[0149] Those skilled in the art will understand that the physical device structure provided in this embodiment does not constitute a limitation on the physical device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0150] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the aforementioned physical device, supporting the operation of information processing programs and other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software in the information processing physical device.

[0151] It should be noted that, in this document, relational terms such as "first" and "second" are used merely 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 term "comprising" or any other variations thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus 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 apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0152] The above are merely specific embodiments of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to these embodiments, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for evaluating the operating status of a transformer, characterized in that, The method includes: Historical oil chromatographic data of dissolved gases in transformer oil are obtained. Based on the historical oil chromatographic data, a multidimensional feature system is constructed that includes fused gas concentration, ratio features and normalization combination. An improved Markov chain Monte Carlo method is used to perform sample augmentation processing on the minority class fault sample set in the multidimensional feature system to construct an augmented sample dataset for model training. Principal component analysis was used to perform visual evaluation of the augmented sample dataset, and the feature importance evaluation mechanism built into the CatBoost classification model was used to select the key feature subsets in the augmented sample dataset that have reached the preset contribution rate threshold and passed the evaluation. A multi-strategy enhanced goose flock optimization algorithm is constructed, and the optimization objective is to minimize the model loss function. The key feature subset and the corresponding training labels are used as input data. The multi-strategy enhanced goose flock optimization algorithm is used to optimize the hyperparameters of the CatBoost classification model to obtain the optimal hyperparameter combination. The CatBoost classification model is retrained based on the key feature subset and the optimal hyperparameter combination to construct a transformer operating status assessment model, and the constructed transformer operating status assessment model is used to assess the operating status of the transformer.

2. The transformer operating status assessment method according to claim 1, characterized in that, Before constructing a multidimensional feature system based on the historical oil chromatographic data, including fused gas concentrations, ratio characteristics, and normalized combinations, the method further includes: The historical oil chromatographic data is cleaned, including: imputing missing values ​​of non-critical variables using the median method; removing sample data corresponding to missing values ​​of critical variables; retaining the first record of duplicate samples based on their timestamps and deleting other duplicate records; and detecting outliers using a combined box plot and percentile method. If a detected outlier matches the corresponding fault label, it is retained; otherwise, it is considered a false anomaly and excluded. The historical oil chromatographic data is subjected to sample labeling and verification processing, which includes: initially labeling the sample data in the historical oil chromatographic data based on operation and maintenance records, fault repair logs and expert diagnostic conclusions, and cross-validating the initial labeling results using the three-ratio method. If the initial labeling results are consistent with the three-ratio method verification results, the initial labeling results are taken as the final labeling results. If the initial labeling results are inconsistent with the three-ratio method verification results, the initial labeling results are taken as the final labeling results, and the three-ratio method verification results are used as auxiliary reference information to correct the final labeling results. The construction of a multidimensional feature system based on the historical oil chromatographic data, comprising fused gas concentration, ratio characteristics, and normalized combinations, includes: Based on the historical oil chromatography data after data cleaning, sample labeling and validation, a multidimensional feature system is constructed, which includes fused gas concentration, ratio features and normalized combinations. The multidimensional feature system is a 32-dimensional multi-level diagnostic feature system.

3. The transformer operating status assessment method according to claim 1, characterized in that, The improved Markov chain Monte Carlo method is used to perform sample augmentation processing on the minority class fault sample set in the multidimensional feature system to construct an augmented sample dataset for model training, including: Based on the minority class of fault samples, a Markov chain Monte Carlo objective function is constructed to evaluate the sample quality, and the acceptance probability is determined based on the Markov chain Monte Carlo objective function, wherein the acceptance probability is a decay factor and an adaptive penalty factor that are introduced to dynamically change with the iteration rounds. Starting from the current sample state, candidate samples are generated using the suggested distribution; Based on the acceptance probability, determine whether to accept the candidate sample; Repeat the above steps until a preset number of augmented sample datasets are generated, wherein the augmented sample datasets consist of a preset number of candidate samples. The formula for calculating the acceptance probability is: In the formula, Indicates the probability of acceptance. Indicates the attenuation factor. Indicates the magnitude of the perturbation in the current iteration. Indicates candidate solutions The Monte Carlo objective function value of the Markov chain. Indicates the current solution The Monte Carlo objective function value of the Markov chain; The update formula for the current iteration perturbation amplitude is: In the formula, Indicates the magnitude of the perturbation in the current iteration. Indicates the adaptive penalty factor. This indicates the magnitude of the disturbance in the previous iteration.

4. The transformer operating status assessment method according to claim 1, characterized in that, The visualization evaluation process of the enhanced sample dataset using principal component analysis includes: Obtain a real sample dataset, and merge the real sample dataset and the augmented sample dataset to construct an original data matrix; The original data matrix is ​​centered, and the covariance matrix of the centered original data matrix is ​​calculated. The covariance matrix is ​​subjected to eigenvalue decomposition to obtain multiple eigenvalues ​​and their corresponding eigenvectors; One or more principal components are selected based on the magnitude of the eigenvalues, and their corresponding eigenvectors are used to project the real sample dataset and the augmented sample dataset from the original high-dimensional space to a low-dimensional visualization space composed of the principal components. In the low-dimensional visualization space, the distribution of the real sample dataset and the augmented sample dataset are compared and analyzed to evaluate the quality of the augmented sample dataset.

5. The transformer operating status assessment method according to claim 1, characterized in that, The construction of the multi-strategy enhanced goose flock optimization algorithm includes: An improved Tent chaotic mapping is used to initialize the goose population; Calculate the fitness value of each individual in the goose population, and determine the globally optimal guide individual for the current iteration based on the fitness-distance balance strategy; The positions of individuals in the goose population are subjected to sine and cosine perturbations to enhance local search capabilities; A lens imaging inverse learning strategy is introduced to generate inverse solutions to expand the search space; The goose population is updated based on the globally optimal guiding individual and the individual positions after sine and cosine perturbations, and the above steps are iteratively executed until the convergence condition is met, thus completing the construction of the multi-strategy enhanced goose flock optimization algorithm.

6. The method according to claim 5, characterized in that, The method for determining the globally optimal guiding individual for the current iteration based on the fitness-distance balance strategy includes: Calculate the Euclidean distance between each individual and the current best individual; Normalize the fitness function values ​​and Euclidean distances of all individuals; The comprehensive score is calculated based on the normalized fitness function value and the Euclidean distance. The individual with the highest overall score is selected as the globally optimal guiding individual.

7. The transformer operating status assessment method according to claim 1, characterized in that, The process of evaluating the operating status of the transformer using the constructed transformer operating status evaluation model includes: Obtain real-time oil chromatographic data of the transformer to be evaluated; The real-time oil chromatography data is preprocessed to construct a real-time feature vector that matches the input format of the transformer operating status assessment model; The real-time feature vector is input into the transformer operation status assessment model for multi-class probability prediction processing, and the probability distribution vector of the transformer to be evaluated is output. Each element in the probability distribution vector corresponds to the probability of different operating states of the transformer. The operating states include at least one of the following: normal state, partial discharge, low-energy discharge, arc discharge, low temperature overheating, medium temperature overheating, or high temperature overheating. Based on the probability distribution vector, the operating state with the highest probability value is selected as the final operating state evaluation result of the transformer to be evaluated.

8. A transformer operating status assessment device, characterized in that, include: The acquisition module is used to acquire historical oil chromatographic data of dissolved gases in transformer oil. Based on the historical oil chromatographic data, a multidimensional feature system is constructed, which includes fused gas concentration, ratio features and normalization combination. An improved Markov chain Monte Carlo method is used to perform sample augmentation processing on the minority class fault sample set in the multidimensional feature system to construct an augmented sample dataset for model training. The filtering module is used to perform visual evaluation processing on the augmented sample dataset using principal component analysis, and to use the feature importance evaluation mechanism built into the CatBoost classification model to filter the subset of key features in the augmented sample dataset that have reached a preset contribution rate threshold and have passed the evaluation. The first processing module is used to construct a multi-strategy enhanced goose flock optimization algorithm, and with minimizing the model loss function as the optimization objective, the key feature subset and the corresponding training labels are used as input data, and the multi-strategy enhanced goose flock optimization algorithm is used to optimize the hyperparameters of the CatBoost classification model to obtain the optimal hyperparameter combination. The evaluation module is used to retrain the CatBoost classification model based on the key feature subset and the optimal hyperparameter combination to construct a transformer operating status evaluation model, and to use the constructed transformer operating status evaluation model to evaluate the operating status of the transformer.

9. The transformer operating status assessment device according to claim 8, characterized in that, The evaluation module is specifically used for: Obtain real-time oil chromatographic data of the transformer to be evaluated; The real-time oil chromatography data is preprocessed to construct a real-time feature vector that matches the input format of the transformer operating status assessment model; The real-time feature vector is input into the transformer operation status assessment model for multi-class probability prediction processing, and the probability distribution vector of the transformer to be evaluated is output. Each element in the probability distribution vector corresponds to the probability of different operating states of the transformer. The operating states include at least one of the following: normal state, partial discharge, low-energy discharge, arc discharge, low temperature overheating, medium temperature overheating, or high temperature overheating. Based on the probability distribution vector, the operating state with the highest probability value is selected as the final operating state evaluation result of the transformer to be evaluated.

10. An electronic device comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the transformer operating status evaluation method according to any one of claims 1 to 7.

Images