Dry-type transformer epoxy resin fatigue life estimation method based on machine learning

Abstract

The invention relates to the technical field of power equipment state evaluation and artificial intelligence, in particular to a dry-type transformer epoxy resin fatigue life estimation method based on machine learning, and the method comprises the steps: collecting transformer operation data and insulation characteristic data as basic data; carrying out anomaly removal, linear interpolation completion and Z-score standardization preprocessing on the acquired data through a 3 sigma criterion; time domain, frequency domain and derivative features are extracted to form initial features, and effective features are obtained through PCA dimension reduction; constructing a life label by combining the accelerated fatigue test and the decommissioned transformer data, and constructing a data set based on the life label and the effective features; an XGBoost-ELM hybrid model is trained, data drift is detected through KL divergence, and online learning and incremental updating of the model are carried out; the real-time data is input into the correction model after being subjected to preprocessing and feature processing, the remaining life is output, and third-level early warning is executed. The method improves the estimation precision and real-time performance, assists the maintenance decision, and guarantees the operation safety of the transformer.

Description

Technical Field

[0001] This invention relates to the fields of power equipment condition assessment and artificial intelligence technology, specifically to a method for predicting the epoxy resin fatigue life of dry-type transformers based on machine learning. Background Technology

[0002] Dry-type transformers, with their advantages of fire and explosion protection and convenient maintenance, are widely used in critical scenarios such as high-rise buildings and rail transportation. Epoxy resin, as its core insulation material, directly determines the reliability of transformer operation through its fatigue life. During long-term service, epoxy resin is subjected to the coupled effects of electrical, thermal, and mechanical stresses, making it prone to micro-damage accumulation and eventual failure. Therefore, accurately predicting its fatigue life is crucial for ensuring power grid safety.

[0003] Existing lifespan prediction technologies have significant limitations: First, traditional mechanistic modeling methods rely on complex physical assumptions, making it difficult to handle the large amount of nonlinear and strongly coupled monitoring data during transformer operation. The models have poor generalization ability, and parameters need to be readjusted when using transformers from different manufacturers. Second, single machine learning methods often use basic features such as temperature and load as inputs, ignoring the time-domain and frequency-domain correlation characteristics of the data. The prediction accuracy is significantly affected by the data quality, and the error often exceeds 20% under conditions of frequent load fluctuations. Third, there is a lack of data adaptation mechanisms. When monitoring data drifts, the model cannot be corrected in real time, leading to a gradual distortion of the prediction results.

[0004] Therefore, there is an urgent need for a machine learning prediction method that can deeply mine the value of multi-source data and has adaptive correction capabilities, in order to solve the problems of poor generalization, low accuracy and lack of adaptability of existing technologies. Summary of the Invention

[0005] The purpose of this invention is to address the problems of existing technologies that rely on single features, have weak generalization ability, and lack data drift correction mechanisms. This invention provides a machine learning-based method for predicting the epoxy resin fatigue life of dry-type transformers. Based on feature engineering, ensemble learning, and adaptive correction, this method achieves high-precision life prediction under different operating conditions and transformer models, providing a reliable basis for condition-based maintenance.

[0006] The method includes the following steps: S1. Construct a multi-source data acquisition system that includes physical monitoring and insulation characteristics, and deploy monitoring devices to collect transformer operation data and insulation characteristic data as basic data; S2. Outliers in the basic data are removed using the 3σ criterion, missing data are filled in using linear interpolation, and Z-score standardization is used to eliminate dimensional differences to obtain the preprocessed data. S3. Extract time-domain features and frequency-domain features from the preprocessed data and construct derived features to form an 18-dimensional initial feature set. After dimensionality reduction by principal component analysis (PCA), a 12-dimensional effective feature set is obtained. S4. Basic life data is obtained through accelerated fatigue testing. The true feature-life correspondence is corrected by combining the data of decommissioned transformers in the field and used as the life label. The dataset composed of the 12-dimensional effective feature set and the life label is divided into training set and test set in a 7:3 ratio. S5. Construct an XGBoost model based on a CART tree and determine the model parameters and splitting logic through iterative training; S6. Construct an XGBoost-ELM hybrid model, use XGBoost to extract high-order features from the training set, input the high-order features into the ELM model for lifetime prediction, output the remaining lifetime of epoxy resin, and use Bayesian optimization algorithm to optimize hyperparameters. S7. Calculate the KL divergence between the newly collected basic data and the training set data in real time, and determine whether the data has drifted by anomaly; S8. Incremental updates are performed using an online learning strategy, and the model accuracy is verified using a test set. Based on the accuracy verification results, it is determined whether to trigger full-scale retraining to obtain the corrected model. S9. After processing the real-time monitoring data through steps S2-S3, input it into the corrected XGBoost-ELM hybrid model. Based on the prediction results of the XGBoost-ELM hybrid model, output the remaining life value and provide maintenance suggestions according to the three-level early warning mechanism.

[0007] It should be further explained that in step S1: The transformer operating data includes the active and reactive power of the three-phase load, the high-voltage side voltage, the ambient temperature, and the winding temperature. The insulation characteristic data include dielectric loss factor, partial discharge quantity, and volume resistivity.

[0008] It should be further explained that the specific steps of step S2 are as follows: S21. Calculate the mean of each set of basic data based on the 3σ criterion. m and standard deviation s Calculate the anomaly detection interval: [ m -3 s , m +3 s Outliers falling outside the interval are removed, and missing data is filled using linear interpolation based on time series. S22. The data after completion is standardized using Z-score to eliminate dimensional differences, resulting in preprocessed data. The formula for the standardization process is: ; in, For data values, For data standard values, The mean, s The standard deviation is denoted as .

[0009] It should be further explained that the specific steps of step S3 are as follows: S31. Calculate the four types of time-domain characteristics of transformer operating data, including mean, variance, peak value and kurtosis; among them, the mean reflects the long-term operating status, the variance reflects the degree of fluctuation, the peak value reflects extreme operating conditions, and the kurtosis reflects abrupt change characteristics. S32. Perform Fourier transform on the active and reactive power, high-voltage side voltage, and partial discharge data of the three-phase load of the transformer to extract two types of frequency domain features, including spectral peak value and harmonic content; whereby the spectral peak value reflects the fluctuation frequency and the harmonic content reflects the degree of nonlinearity. S33. Based on the operating mechanism of power equipment, six types of derived characteristics are constructed, including the medium temperature-load coupling coefficient T / P and the temperature sensitivity of dielectric loss Δtan d / ΔT, partial discharge pulse density PD, voltage-to-load ratio U / P, temperature coefficient of resistance Δ r / ΔT and load fluctuation entropy; S34. Standardize the 18-dimensional initial feature variables of n samples to obtain the standardized matrix. ; S35. Solve the equation 18 eigenvalues ​​were obtained. and the corresponding 18 feature vectors The variance of the principal components is selected based on the magnitude of their eigenvalues, sorted from largest to smallest by cumulative variance contribution rate and individual principal component variance contribution rate. The top 12 eigenvalues ​​are then selected, and their corresponding 12 eigenvectors form the projection matrix. Wherein, cumulative variance contribution rate = Data projection dimensionality reduction uses a projection matrix W to perform a linear transformation on the standardized data Z, resulting in a 12-dimensional effective feature set. ,in, I is the identity matrix, and C is the covariance matrix, used to measure the correlation between features, according to the formula... Find the eigenvalues ​​and eigenvectors of the covariance matrix C.

[0010] It should be further explained that the specific steps of step S4 are as follows: S41. Select the same grade of epoxy resin as the in-service transformer, make standard specimens, simulate different stress conditions in an accelerated fatigue test chamber, and record the time from micro-damage to failure of the specimens as the basic life data. S42. Collect the operation monitoring data and actual service life of m retired transformers, combine them with the damage state of epoxy resin after dismantling, correct the accelerated fatigue test data, and obtain the true characteristic-life correspondence. S43. Divide the dataset consisting of 12-dimensional features and lifetime labels into a training set and a test set in a 7:3 ratio, where the training set contains 5000 samples and the test set contains 2143 samples.

[0011] It should be further explained that the specific steps of step S5 are as follows: S51. Build an XGBoost model, assuming K trees are trained, and the final prediction is the sum of the predictions from all trees. For the i-th... i For each sample, the final predicted value is: ; in, For the t-th tree pair The predicted value, Belongs to all CART tree sets, the iteration logic is as follows: , indicating that the prediction in round t = the prediction in the previous t-1 rounds + the prediction of the t-th tree; S52. For the newly added t-th tree, based on the aforementioned iterative logic, the objective function for the t-th round is rewritten as: ; Where n is the number of samples, For loss function, This is the complexity penalty term for the t-th decision tree; S53. Substituting the loss function into the objective function yields a simplified objective function containing only the first derivative. and second derivative Quadratic function, where: ; ; The simplified objective function is: ; S54. Assign samples to different leaf nodes, grouping them by leaf node. Assume the t-th tree has T leaves, and the sample set is: ; in, Let j be the set of samples contained in the j-th leaf node. Let be the weight of the j-th leaf node; S55. The simplified objective function is summed over the leaf nodes, and the optimal leaf weights are derived as follows: ; S56. Calculate the split gain: ; in, , These are the sums of the first and second derivatives of the samples in the left subtree, respectively. , These are the sum of the first and second derivatives of the samples in the right subtree, respectively. The threshold is used; when Gain > 0, the split occurs. The larger the size, the more difficult it is to split.

[0012] It should be further explained that the specific steps of step S6 are as follows: S61. Input the training set into the XGBoost model, set the number of decision trees to 100, the learning rate to 0.05, and the maximum tree depth to 6. Extract the hidden high-order correlation features in the data through iterative training and output 6-dimensional high-order features. S62. Use the 6-dimensional features output by XGBoost as the input layer of ELM, set the number of nodes in the input layer to 6, the number of nodes in the hidden layer to 20, the activation function to be the Sigmoid function, and the output layer to be the remaining lifetime of epoxy resin. S63. The Bayesian optimization algorithm is used to jointly optimize the learning rate of XGBoost and the number of hidden layer nodes of ELM, with the objective function being to minimize the mean absolute error of the test set.

[0013] It should be further explained that step S7 specifically includes: Establish a data drift monitoring mechanism to calculate the KL divergence between newly collected data and training data in real time. Define the outlier as KL divergence / initial KL divergence. When the outlier is greater than 5%, it is determined that the data has drifted significantly.

[0014] It should be further explained that the specific steps of step S8 are as follows: S81. An online learning strategy is adopted, incorporating manually labeled lifetime data and new data after drift into the training set at a ratio of 10%, to incrementally train the XGBoost-ELM model and update the model parameters; S82. After each update, the model accuracy is verified using the test set. When the mean absolute error is greater than 8%, the entire dataset is retrained.

[0015] It should be further explained that the specific steps of step S9 are as follows: S91. After processing the real-time monitoring data through steps S2-S3, input it into the corrected XGBoost-ELM hybrid model and output the remaining life of the epoxy resin in days. S92. Set a three-level early warning system based on the remaining lifespan: Level 1 warning, remaining lifespan > 180 days: Normal operation, monthly monitoring; Level II warning, 90 days < remaining lifespan ≤ 180 days: Strengthen monitoring and formulate maintenance plans; Level 3 warning, remaining lifespan ≤ 90 days: Immediate shutdown for maintenance.

[0016] As can be seen from the above technical solutions, the present invention has the following advantages: This invention presents a machine learning-based method for predicting the epoxy resin fatigue life of dry-type transformers. It combines physical monitoring, a multi-source data system of insulation characteristics, and PCA dimensionality reduction to extract core features. This method does not rely on complex physical assumptions, is adaptable to dry-type transformers from different manufacturers, avoids repeated parameter adjustments, and improves cross-device generalization. It addresses the issue of insufficient accuracy in single machine learning by extracting time-domain, frequency-domain, and coupled derived features. A high-order feature correlation is mined using an XGBoost-ELM hybrid model, and hyperparameters are optimized using Bayesian methods, reducing prediction errors under fluctuating load conditions. Furthermore, it resolves the distortion caused by data drift by using KL divergence to detect drift in real time, combined with online learning incremental updates and a full retraining mechanism to achieve adaptive model correction. This invention effectively improves prediction accuracy and adaptability, assisting in power grid safety and maintenance decisions in critical scenarios. Attached Figure Description

[0017] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the machine learning-based method for predicting the fatigue life of epoxy resin in dry-type transformers according to the present invention.

[0019] Figure 2 This is a schematic diagram of a multi-source data monitoring device for the machine learning-based method for predicting the epoxy resin fatigue life of dry-type transformers according to the present invention.

[0020] Figure 3 This is a schematic diagram of the 12-dimensional effective feature set of the machine learning-based method for predicting the fatigue life of epoxy resin in dry-type transformers according to the present invention.

[0021] Figure 4 This is a schematic diagram of the XGBoost-ELM hybrid model structure of the machine learning-based method for predicting the epoxy resin fatigue life of dry transformers according to the present invention. Detailed Implementation

[0022] The various embodiments of the present invention will be described more fully in the detailed steps of the machine learning-based method for predicting the epoxy resin fatigue life of dry-type transformers, which will be described in detail below. The present invention may have various embodiments, and adjustments and changes may be made therein. However, it should be understood that there is no intention to limit the various embodiments of the present invention to the specific embodiments disclosed herein, but rather the present invention should be understood to cover all adjustments, equivalents, and / or alternatives falling within the spirit and scope of the various embodiments of the present invention.

[0023] It should be understood that, when used in this specification, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0024] The terms "one embodiment" or "some embodiments" used in this invention mean that one or more embodiments of the invention include the specific features, structures, or characteristics described in that embodiment. Therefore, the terms "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of the invention do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized.

[0025] To make the objectives, features, and advantages of this invention more apparent and understandable, specific embodiments and accompanying drawings will be used to clearly and completely describe the technical solutions protected by this invention. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0026] Please see Figure 1 The diagram shows a flowchart of a machine learning-based method for predicting the fatigue life of epoxy resin in dry-type transformers. The method includes the following steps: S1. Construct a multi-source data acquisition system that includes physical monitoring and insulation characteristics, and deploy monitoring devices to collect transformer operation data and insulation characteristic data as basic data; S2. Outliers in the basic data are removed using the 3σ criterion, missing data are filled in using linear interpolation, and Z-score standardization is used to eliminate dimensional differences to obtain the preprocessed data. S3. Extract time-domain features and frequency-domain features from the preprocessed data and construct derived features to form an 18-dimensional initial feature set. After dimensionality reduction by principal component analysis (PCA), a 12-dimensional effective feature set is obtained. S4. Basic life data is obtained through accelerated fatigue testing. The true feature-life correspondence is corrected by combining the data of decommissioned transformers in the field and used as the life label. The dataset composed of the 12-dimensional effective feature set and the life label is divided into training set and test set in a 7:3 ratio. S5. Construct an XGBoost model based on a CART tree and determine the model parameters and splitting logic through iterative training; S6. Construct an XGBoost-ELM hybrid model, use XGBoost to extract high-order features from the training set, input the high-order features into the ELM model for lifetime prediction, output the remaining lifetime of epoxy resin, and use Bayesian optimization algorithm to optimize hyperparameters. S7. Calculate the KL divergence between the newly collected basic data and the training set data in real time, and determine whether the data has drifted by anomaly; S8. Incremental updates are performed using an online learning strategy, and the model accuracy is verified using a test set. Based on the accuracy verification results, it is determined whether to trigger full-scale retraining to obtain the corrected model. S9. After processing the real-time monitoring data through steps S2-S3, input it into the corrected XGBoost-ELM hybrid model. Based on the prediction results of the XGBoost-ELM hybrid model, output the remaining life value and provide maintenance suggestions according to the three-level early warning mechanism.

[0027] It should be noted that, in this embodiment, the above method is applicable to the full life cycle management of dry-type transformers of 10kV~35kV level.

[0028] As a refinement and extension of the specific implementation of the above embodiments, in order to fully illustrate the specific implementation process in this embodiment, another method for predicting the epoxy resin fatigue life of dry-type transformers based on machine learning is provided. The implementation object is a 10kV dry-type transformer (model SCB13-1000 / 10) with epoxy resin model E-44; the monitoring equipment includes a temperature sensor (accuracy ±0.1℃), a power sensor (accuracy class 0.5), a partial discharge detector (detection range 10pC~1000pC), and a dielectric loss tester (accuracy ±0.0001); the computing platform is an Intel Xeon E5-2678 v3 CPU, an NVIDIA Tesla V100 GPU, and a Python 3.9 software environment (Scikit-learn, XGBoost library).

[0029] The methods include: S1. Construct a multi-source data acquisition system that includes physical monitoring and insulation characteristics, and deploy monitoring devices to collect transformer operation data and insulation characteristic data as basic data; in step S1: The transformer operating data includes the active and reactive power of the three-phase load, the high-voltage side voltage, the ambient temperature, and the winding temperature. The insulation characteristic data include dielectric loss factor, partial discharge quantity, and volume resistivity.

[0030] like Figure 2 As shown, the monitoring device includes: Core equipment: Dry-type transformer; Sensing modules: PD (partial discharge detection device), temperature sensor, voltage / current measuring device, power factor meter, power sensor, dielectric loss tester, and Tanδ detection device; Transmission and Platform Modules: Data Acquisition Terminal, RS-485 Module, PQ Device (Communication Relay), 5G / WiFi Communication Module, Cloud Platform.

[0031] The PD and temperature sensors are directly deployed on the transformer body; voltage / current measuring devices are connected to the transformer's electrical circuit; power factor analyzers, dielectric loss testers, Tanδ detectors, and power sensors are connected to the corresponding test points on the transformer via cables to collect equipment status parameters. Signals from each sensing module are aggregated to a data acquisition terminal, which transmits data via an RS-485 module, and then uploads the data to the cloud platform via a 5G / WiFi communication module (in conjunction with PQ equipment for relay), enabling the acquisition and management of multi-source data.

[0032] S2. Outliers in the basic data are removed using the 3σ criterion, missing data are filled in using linear interpolation, and Z-score standardization is used to eliminate dimensional differences, resulting in preprocessed data; the specific steps of step S2 are as follows: S21. Calculate the mean of each set of basic data based on the 3σ criterion. m and standard deviation s Calculate the anomaly detection interval: [ m -3 s , m +3 s Outliers falling outside the interval are removed, and missing data is filled using linear interpolation based on time series. S22. The data after completion is standardized using Z-score to eliminate dimensional differences, resulting in preprocessed data. The formula for the standardization process is: ; in, For data values, For data standard values, The mean, s The standard deviation is denoted as .

[0033] S3. Extract time-domain and frequency-domain features from the preprocessed data and construct derived features to form an 18-dimensional initial feature set. After principal component analysis (PCA) feature dimensionality reduction, a 12-dimensional effective feature set is obtained. The 12-dimensional effective feature set is as follows: Figure 3 As shown, the features include: 4-dimensional time-domain features: mean winding temperature (18.2%), load variance (15.7%), winding temperature kurtosis (1.5%), and load peak value (1.5%). 2-dimensional frequency-domain features: load spectrum peak value (8.5%), partial discharge harmonic content (4.2%). 6-dimensional derived features: temperature-load coupling coefficient (12.3%), dielectric loss temperature sensitivity (10.8%), partial discharge pulse density (7.6%), load fluctuation entropy (5.4%), voltage-load ratio (3.1%), and resistance temperature coefficient (2.9%). The cumulative contribution of the 12-dimensional effective feature set is 96.3%.

[0034] The specific steps of step S3 are as follows: S31. Calculate the four types of time-domain characteristics of transformer operating data, including mean, variance, peak value and kurtosis; among them, the mean reflects the long-term operating status, the variance reflects the degree of fluctuation, the peak value reflects extreme operating conditions, and the kurtosis reflects abrupt change characteristics. S32. Perform Fourier transform on the active and reactive power, high-voltage side voltage, and partial discharge data of the three-phase load of the transformer to extract two types of frequency domain features, including spectral peak value and harmonic content; whereby the spectral peak value reflects the fluctuation frequency and the harmonic content reflects the degree of nonlinearity. S33. Based on the operating mechanism of power equipment, six types of derived characteristics are constructed, including the medium temperature-load coupling coefficient T / P and the temperature sensitivity of dielectric loss Δtan d / ΔT, partial discharge pulse density PD, voltage-to-load ratio U / P, temperature coefficient of resistance Δ r / ΔT and load fluctuation entropy; S34. Standardize the 18-dimensional initial feature variables of n samples to obtain the standardized matrix. ; S35. Solve the equation 18 eigenvalues ​​were obtained. and the corresponding 18 feature vectors The variance of the principal components is selected based on the magnitude of their eigenvalues, sorted from largest to smallest by cumulative variance contribution rate and individual principal component variance contribution rate. The top 12 eigenvalues ​​are then selected, and their corresponding 12 eigenvectors form the projection matrix. Wherein, cumulative variance contribution rate = Data projection dimensionality reduction uses a projection matrix W to perform a linear transformation on the standardized data Z, resulting in a 12-dimensional effective feature set. ,in, I is the identity matrix, and C is the covariance matrix, used to measure the correlation between features, according to the formula... Find the eigenvalues ​​and eigenvectors of the covariance matrix C.

[0035] S4. Obtain basic life data through accelerated fatigue testing, and combine it with data from decommissioned transformers in the field to correct and obtain the true feature-life correspondence as the life label. Divide the dataset consisting of the 12-dimensional effective feature set and the life label into a training set and a test set in a 7:3 ratio. The specific steps of step S4 are as follows: S41. Select the same grade of epoxy resin as the in-service transformer, make standard specimens, simulate different stress conditions in an accelerated fatigue test chamber, and record the time from micro-damage to failure of the specimens as the basic life data. S42. Collect the operation monitoring data and actual service life of m retired transformers, combine them with the damage state of epoxy resin after dismantling, correct the accelerated fatigue test data, and obtain the true characteristic-life correspondence. S43. Divide the dataset consisting of 12-dimensional features and lifetime labels into a training set and a test set in a 7:3 ratio, where the training set contains 5000 samples and the test set contains 2143 samples.

[0036] S5. Construct an XGBoost model based on a CART tree, and determine the model parameters and splitting logic through iterative training; the specific steps of step S5 are as follows: S51. Build an XGBoost model, assuming K trees are trained, and the final prediction is the sum of the predictions from all trees. For the i-th... i For each sample, the final predicted value is: ; in, For the t-th tree pair The predicted value, Belongs to all CART tree sets, the iteration logic is as follows: , indicating that the prediction in round t = the prediction in the previous t-1 rounds + the prediction of the t-th tree; S52. For the newly added t-th tree, based on the aforementioned iterative logic, the objective function for the t-th round is rewritten as: ; Where n is the number of samples, For loss function, This is the complexity penalty term for the t-th decision tree; S53. Substituting the loss function into the objective function yields a simplified objective function containing only the first derivative. and second derivative Quadratic function, where: ; ; The simplified objective function is: ; S54. Assign samples to different leaf nodes, grouping them by leaf node. Assume the t-th tree has T leaves, and the sample set is: ; in, Let j be the set of samples contained in the j-th leaf node. Let be the weight of the j-th leaf node; S55. The simplified objective function is summed over the leaf nodes, and the optimal leaf weights are derived as follows: ; S56. Calculate the split gain: ; in, , These are the sums of the first and second derivatives of the samples in the left subtree, respectively. , These are the sum of the first and second derivatives of the samples in the right subtree, respectively. The threshold is used; when Gain > 0, the split occurs. The larger the size, the more difficult it is to split.

[0037] S6. Construct an XGBoost-ELM hybrid model. XGBoost is used to extract high-order features from the training set, which are then input into the ELM model for lifetime prediction. The output is the remaining lifetime of the epoxy resin. Bayesian optimization is used to optimize the hyperparameters. The structure of the XGBoost-ELM hybrid model is as follows: Figure 4 As shown, it includes: a data input layer, an XGBoost feature enhancement module, an XGBoost output layer, an ELM input layer, an ELM hidden layer, an ELM output layer, a result output layer, and an error feedback module; the data input layer, XGBoost feature enhancement module, XGBoost output layer, ELM input layer, ELM hidden layer, ELM output layer, and result output layer are connected in sequence, the ELM output layer is connected to the error feedback module, and the error feedback module is connected to the XGBoost feature enhancement module.

[0038] The specific steps of step S6 are as follows: S61. Input the training set into the XGBoost model, set the number of decision trees to 100, the learning rate to 0.05, and the maximum tree depth to 6. Extract the hidden high-order correlation features in the data through iterative training and output 6-dimensional high-order features. S62. Use the 6-dimensional features output by XGBoost as the input layer of ELM, set the number of nodes in the input layer to 6, the number of nodes in the hidden layer to 20, the activation function to be the Sigmoid function, and the output layer to be the remaining lifetime of epoxy resin. S63. The Bayesian optimization algorithm is used to jointly optimize the learning rate of XGBoost and the number of hidden layer nodes of ELM, with the objective function being to minimize the mean absolute error of the test set.

[0039] S7. Calculate the KL divergence between the newly collected baseline data and the training set data in real time, and determine whether data drift has occurred through outlier; Step S7 specifically includes: Establish a data drift monitoring mechanism to calculate the KL divergence between newly collected data and training data in real time. Define the outlier as KL divergence / initial KL divergence. When the outlier is greater than 5%, it is determined that the data has drifted significantly.

[0040] S8. Incremental updates are performed using an online learning strategy, and the model accuracy is verified using a test set. Based on the accuracy verification results, it is determined whether to trigger full-data retraining to obtain the corrected model. The specific steps of step S8 are as follows: S81. An online learning strategy is adopted, incorporating manually labeled lifetime data and new data after drift into the training set at a ratio of 10%, to incrementally train the XGBoost-ELM model and update the model parameters; S82. After each update, the model accuracy is verified using the test set. When the mean absolute error is greater than 8%, the entire dataset is retrained.

[0041] S9. After processing the real-time monitoring data in steps S2-S3, input it into the corrected XGBoost-ELM hybrid model. Based on the prediction results of the XGBoost-ELM hybrid model, output the remaining lifespan value and provide maintenance recommendations according to the three-level early warning mechanism. The specific steps of step S9 are as follows: S91. After processing the real-time monitoring data through steps S2-S3, input it into the corrected XGBoost-ELM hybrid model and output the remaining life of the epoxy resin in days. S92. Set a three-level early warning system based on the remaining lifespan: Level 1 warning, remaining lifespan > 180 days: Normal operation, monthly monitoring; Level II warning, 90 days < remaining lifespan ≤ 180 days: Strengthen monitoring and formulate maintenance plans; Level 3 warning, remaining lifespan ≤ 90 days: Immediate shutdown for maintenance.

[0042] In some embodiments, the following steps are included: Data collection: Eight types of data were collected for one year, resulting in 105,120 sets of raw data. After cleaning, 104,068 sets of valid data were retained. Feature construction: 18 initial features were extracted, and 12 features were retained after PCA dimensionality reduction (cumulative contribution rate 96.3%). Model training: After Bayesian optimization, the XGBoost learning rate was set to 0.05, the number of hidden layer nodes in ELM was set to 20, the MAE on the training set was 4.8%, and the MAE on the test set was 6.2%. Drift correction: After 6 months of operation, the data anomaly rate reached 6.5%, triggering an incremental update. After the update, the MAE of the test set recovered to 5.9%. Lifespan output: Input monitoring data in real time, the model outputs a remaining lifespan of 826 days, corresponding to a level 1 warning, and monthly monitoring is recommended.

[0043] Three existing methods were selected and compared with the method in this embodiment to verify the advantages of the present invention. The comparison data is shown in Table 1. Table 1 Comparison Data

[0044] It should be noted that, in terms of basic prediction accuracy, the MAE of the traditional damage mechanics model on the test set reaches 18.5%, while the MAE of the XGBoost-ELM hybrid model of this invention on the test set is 6.2%, which is better than the single ELM model (10.3%) and the XGBoost model (7.9%). Regarding multi-device adaptability, the accuracy fluctuation of the traditional damage mechanics model is 12.3%, while the accuracy fluctuation of the XGBoost-ELM hybrid model of this invention is only 1.8%, demonstrating stronger cross-device stability. Faced with data drift (changes in data distribution), the error of the traditional damage mechanics model soars to 25.1%, while the error of the XGBoost-ELM hybrid model of this invention is only 5.9%, demonstrating superior robustness against data changes. In summary, this invention comprehensively outperforms traditional models and existing machine learning models in terms of prediction accuracy, device adaptability stability, and resistance to data drift.

Claims

1. A method for predicting the fatigue life of epoxy resin in dry-type transformers based on machine learning, characterized in that, The method includes the following steps: S1. Construct a multi-source data acquisition system that includes physical monitoring and insulation characteristics, and deploy monitoring devices to collect transformer operation data and insulation characteristic data as basic data; S2. Outliers in the basic data are removed using the 3σ criterion, missing data are filled in using linear interpolation, and Z-score standardization is used to eliminate dimensional differences to obtain the preprocessed data. S3. Extract time-domain features and frequency-domain features from the preprocessed data and construct derived features to form an 18-dimensional initial feature set. After dimensionality reduction by principal component analysis (PCA), a 12-dimensional effective feature set is obtained. S4. Basic life data is obtained through accelerated fatigue testing. The true feature-life correspondence is corrected by combining the data of decommissioned transformers in the field and used as the life label. The dataset composed of the 12-dimensional effective feature set and the life label is divided into training set and test set in a 7:3 ratio. S5. Construct an XGBoost model based on a CART tree and determine the model parameters and splitting logic through iterative training; S6. Construct an XGBoost-ELM hybrid model, use XGBoost to extract high-order features from the training set, input the high-order features into the ELM model for lifetime prediction, output the remaining lifetime of epoxy resin, and use Bayesian optimization algorithm to optimize hyperparameters. S7. Calculate the KL divergence between the newly collected basic data and the training set data in real time, and determine whether the data has drifted by anomaly; S8. Incremental updates are performed using an online learning strategy, and the model accuracy is verified using a test set. Based on the accuracy verification results, it is determined whether to trigger full data retraining to obtain the corrected model. S9. After processing the real-time monitoring data through steps S2-S3, input it into the corrected XGBoost-ELM hybrid model. Based on the prediction results of the XGBoost-ELM hybrid model, output the remaining life value and provide maintenance suggestions according to the three-level early warning mechanism.

2. The method for predicting the fatigue life of epoxy resin in dry-type transformers based on machine learning according to claim 1, characterized in that, In step S1: The transformer operating data includes the active and reactive power of the three-phase load, the high-voltage side voltage, the ambient temperature, and the winding temperature. The insulation characteristic data include dielectric loss factor, partial discharge quantity, and volume resistivity.

3. The method for predicting the fatigue life of epoxy resin in dry-type transformers based on machine learning according to claim 1, characterized in that, The specific steps of step S2 are as follows: S21. Calculate the mean of each set of basic data based on the 3σ criterion. μ and standard deviation σ Calculate the anomaly detection interval: [ μ -3 σ , μ +3 σ Outliers falling outside the interval are removed, and missing data is filled using linear interpolation based on time series. S22. The data after completion is standardized using Z-score to eliminate dimensional differences, resulting in preprocessed data. The formula for the standardization process is: ； in, For data values, For data standard values, The mean, σ The standard deviation is denoted as .

4. The method for predicting the fatigue life of epoxy resin in dry-type transformers based on machine learning according to claim 2, characterized in that, The specific steps of step S3 are as follows: S31. Calculate the four types of time-domain characteristics of transformer operating data, including mean, variance, peak value and kurtosis; among them, the mean reflects the long-term operating status, the variance reflects the degree of fluctuation, the peak value reflects extreme operating conditions, and the kurtosis reflects abrupt change characteristics. S32. Perform Fourier transform on the active and reactive power, high-voltage side voltage, and partial discharge data of the three-phase load of the transformer to extract two types of frequency domain features, including spectral peak value and harmonic content; whereby the spectral peak value reflects the fluctuation frequency and the harmonic content reflects the degree of nonlinearity. S33. Based on the operating mechanism of power equipment, six types of derived characteristics are constructed, including the medium temperature-load coupling coefficient T / P and the temperature sensitivity of dielectric loss Δtan δ / ΔT, partial discharge pulse density PD, voltage-to-load ratio U / P, temperature coefficient of resistance Δ ρ / ΔT and load fluctuation entropy; S34. Standardize the 18-dimensional initial feature variables of n samples to obtain the standardized matrix. ; S35. Solve the equation 18 eigenvalues ​​were obtained. and the corresponding 18 feature vectors The variance of the principal components is selected based on the magnitude of their eigenvalues, sorted from largest to smallest by cumulative variance contribution rate and individual principal component variance contribution rate. The top 12 eigenvalues ​​are then selected, and their corresponding 12 eigenvectors form the projection matrix. Wherein, cumulative variance contribution rate = Data projection dimensionality reduction uses a projection matrix W to perform a linear transformation on the standardized data Z, resulting in a 12-dimensional effective feature set. ,in, I is the identity matrix, and C is the covariance matrix, used to measure the correlation between features, according to the formula... Find the eigenvalues ​​and eigenvectors of the covariance matrix C.

5. The method for predicting the fatigue life of epoxy resin in dry-type transformers based on machine learning according to claim 1, characterized in that, The specific steps of step S4 are as follows: S41. Select the same grade of epoxy resin as the in-service transformer, prepare standard specimens, simulate different stress conditions in an accelerated fatigue test chamber, and record the time from micro-damage to failure of the specimens as the basic life data. S42. Collect the operation monitoring data and actual service life of m retired transformers, combine them with the damage state of epoxy resin after dismantling, correct the accelerated fatigue test data, and obtain the true characteristic-life correspondence. S43. Divide the dataset consisting of 12-dimensional features and lifetime labels into a training set and a test set in a 7:3 ratio, where the training set contains 5000 samples and the test set contains 2143 samples.

6. The method for predicting the fatigue life of epoxy resin in dry-type transformers based on machine learning according to claim 1, characterized in that, The specific steps of step S5 are as follows: S51. Build an XGBoost model, assuming K trees are trained, and the final prediction is the sum of the predictions from all trees. For the i-th... i For each sample, the final predicted value is: ； in, For the t-th tree pair The predicted value, Belongs to all CART tree sets, the iteration logic is as follows: , indicating that the prediction in round t = the prediction in the previous t-1 rounds + the prediction of the t-th tree; S52. For the newly added t-th tree, based on the aforementioned iterative logic, the objective function for the t-th round is rewritten as: ； Where n is the number of samples, For loss function, This is the complexity penalty term for the t-th decision tree; S53. Substituting the loss function into the objective function yields a simplified objective function containing only the first derivative. and second derivative Quadratic function, where: ； ； The simplified objective function is: ； S54. Assign samples to different leaf nodes, grouping them by leaf node. Assume the t-th tree has T leaves, and the sample set is: ； in, Let j be the set of samples contained in the j-th leaf node. Let be the weight of the j-th leaf node; S55. The simplified objective function is summed over the leaf nodes, and the optimal leaf weights are derived as follows: ； S56. Calculate the split gain: ； in, , These are the sums of the first and second derivatives of the samples in the left subtree, respectively. , These are the sum of the first and second derivatives of the samples in the right subtree, respectively. The threshold is used; when Gain > 0, the split occurs. The larger the size, the more difficult it is to split.

7. The method for predicting the fatigue life of epoxy resin in dry-type transformers based on machine learning according to claim 1, characterized in that, The specific steps of step S6 are as follows: S61. Input the training set into the XGBoost model, set the number of decision trees to 100, the learning rate to 0.05, and the maximum tree depth to 6. Extract the hidden high-order correlation features in the data through iterative training and output 6-dimensional high-order features. S62. Use the 6-dimensional features output by XGBoost as the input layer of ELM, set the number of nodes in the input layer to 6, the number of nodes in the hidden layer to 20, the activation function to be the Sigmoid function, and the output layer to be the remaining lifetime of epoxy resin. S63. The Bayesian optimization algorithm is used to jointly optimize the learning rate of XGBoost and the number of hidden layer nodes of ELM, with the objective function being to minimize the mean absolute error of the test set.

8. The method for predicting the fatigue life of epoxy resin in dry-type transformers based on machine learning according to claim 1, characterized in that, Step S7 specifically includes: Establish a data drift monitoring mechanism to calculate the KL divergence between newly collected data and training data in real time. Define the outlier as KL divergence / initial KL divergence. When the outlier is greater than 5%, it is determined that the data has drifted significantly.

9. The method for predicting the fatigue life of epoxy resin in dry-type transformers based on machine learning according to claim 1, characterized in that, The specific steps of step S8 are as follows: S81. An online learning strategy is adopted, incorporating manually labeled lifetime data and new data after drift into the training set at a ratio of 10%, to incrementally train the XGBoost-ELM model and update the model parameters; S82. After each update, the model accuracy is verified using the test set. When the mean absolute error is greater than 8%, the entire dataset is retrained.

10. The method for predicting the fatigue life of epoxy resin in dry-type transformers based on machine learning according to claim 1, characterized in that, The specific steps of step S9 are as follows: S91. After processing the real-time monitoring data through steps S2-S3, input it into the corrected XGBoost-ELM hybrid model and output the remaining life of the epoxy resin in days. S92. Set a three-level early warning system based on the remaining lifespan: Level 1 warning, remaining lifespan > 180 days: Normal operation, monthly monitoring; Level II warning, 90 days < remaining lifespan ≤ 180 days: Strengthen monitoring and formulate maintenance plans; Level 3 warning, remaining lifespan ≤ 90 days: Immediate shutdown for maintenance.

Images