A transformer intelligent state evaluation method based on SMOTETomek and graph convolution network

Abstract

The present application belongs to the technical field of power equipment big data operation and maintenance and artificial intelligence fault diagnosis, and particularly relates to a transformer intelligent state evaluation method based on SMOTE Tomek and graph convolution network. The method collects transformer temperature, strain, partial discharge, oil chromatogram and vibration multi-source monitoring data, adopts a SMOTE Tomek hybrid sampling algorithm to balance the training set samples, constructs a k-neighbor graph through multi-metric fusion of Gaussian kernel, cosine similarity and Pearson correlation coefficient, accurately depicts the nonlinear coupling relationship of multiple monitoring parameters, builds a residual graph convolution network containing residual connection to complete deep feature extraction, introduces game theory to calculate the transformer comprehensive health score by fusing subjective and objective weights, outputs the equipment health level and four types of typical fault recognition results according to the grading score, predicts the remaining service life based on the Weibull distribution, synchronously triggers a three-level grading early warning and generates an operation and maintenance decision report. The present application fully excavates the correlation characteristics of multi-source monitoring data, and the evaluation accuracy can reach more than 95%.

Description

Technical Field

[0001] This invention belongs to the field of power equipment big data operation and maintenance and artificial intelligence fault diagnosis technology, and in particular relates to a transformer intelligent state assessment method based on SMOTE Tomek and graph convolutional networks. Background Technology

[0002] Power transformers are core hub equipment in power transmission and distribution systems, and their operating status directly affects the reliability and safety of the power grid. With the increasing intelligence of the power grid, transformer monitoring dimensions are constantly being enriched, covering multi-source data such as oil chromatography, vibration, partial discharge, temperature, and strain. However, existing condition assessment technologies still have many limitations: First, the data category distribution is severely unbalanced. The proportion of normal transformer operation samples is extremely high, while fault samples are difficult to collect and few in number. Directly building a diagnostic model is prone to majority class bias, resulting in a high rate of missed fault detection and insufficient accuracy.

[0003] Secondly, the fusion of multi-source monitoring information is difficult. Traditional assessment methods often rely on single oil chromatographic indicators or fixed thresholds for judgment, making it difficult to uncover the nonlinear coupling correlation between multi-dimensional parameters such as temperature, vibration, partial discharge, and oil quality. They also fail to capture the coordinated change characteristics of early, subtle faults, resulting in a lag in early warning.

[0004] Third, the weighting of status assessments is highly subjective. Existing health scoring systems rely heavily on expert experience to set indicator weights, making them susceptible to human interference and unable to adapt to different operating conditions and individual equipment differences, thus limiting the stability and accuracy of assessment results.

[0005] Fourth, the assessment output dimensions are limited. Most solutions can only output the current health level, failing to simultaneously achieve accurate fault type identification and quantitative prediction of remaining service life, making it difficult to support predictive maintenance decisions throughout the entire lifecycle.

[0006] In recent years, graph convolutional networks have been gradually applied to the field of equipment diagnosis due to their ability to model non-Euclidean data. However, existing solutions mostly only build graph structures for a single data source, and deep networks are prone to feature oversmoothing and performance degradation. At the same time, they lack a scientific quantitative evaluation system, making it difficult to meet the high-precision operation and maintenance needs of engineering sites. Summary of the Invention

[0007] The purpose of this invention is to address the aforementioned technical problems by providing a smart transformer condition assessment method based on SMOTE Tomek and graph convolutional networks. This method solves the data imbalance problem through hybrid sampling, achieves deep feature fusion of multi-source data through multi-metric graph construction and residual graph convolutional networks, and enhances the objectivity of the assessment by combining game theory with subjective and objective weighting. Finally, it synchronously outputs the health level, fault type, and remaining life prediction results, thereby achieving high-precision transformer condition assessment and early fault warning.

[0008] In view of this, the present invention provides a transformer intelligent state assessment method based on SMOTE Tomek and graph convolutional networks, comprising the following steps: S1: Collect multi-source monitoring data of transformer temperature, strain, partial discharge, oil chromatography, and vibration, classify different health status categories according to health index, and split the dataset into training set and test set according to proportion; S2: The SMOTETomek hybrid sampling algorithm is used to perform class balancing preprocessing on the training set, and the balanced training set and test set data are normalized. S3: Map various monitoring parameters to graph nodes, use multi-metric fusion method to calculate the correlation strength between nodes, and construct a k-nearest neighbor graph structure to represent the coupling relationship of parameters; S4: Construct a residual graph convolutional network model containing residual connections, input k-nearest neighbor graph data into the model for deep feature extraction, and obtain a high-dimensional feature vector of multi-source data fusion; S5: Based on game theory, the deep fusion features and state evaluation indicators are weighted both subjectively and objectively to calculate the comprehensive health status score of the transformer. S6: Based on the comprehensive health status score, output the transformer health level, fault type, and remaining life prediction results.

[0009] Furthermore, in S1, the multi-source monitoring data includes dissolved gas data in the oil, chamber vibration acceleration signal, partial discharge signal, fiber optic temperature and strain monitoring data, and electrical parameter data; wherein the dissolved gas data in the oil includes time-series concentration data of H2, CH4, C2H6, C2H4, C2H2, CO, and CO2. The health status is divided into five levels—very good, good, moderate, poor, and very poor—based on the health index value, with a training set to test set ratio of 7:3.

[0010] Furthermore, in S2, the specific execution steps of the SMOTE Tomek hybrid sampling algorithm are as follows: Identify minority class fault samples in the training set, randomly select a benchmark sample in the minority class, calculate its Euclidean distance to the rest of the samples in the same class, and select k nearest neighbor samples. Randomly select the target sample from the k nearest neighbor samples, according to the formula. New minority class samples are generated and added to the training set to complete minority class oversampling, where x is the baseline sample and x is the base sample. n For the target nearest neighbor samples, rand(0,1) is a random number between 0 and 1; Traverse all samples in the training set, calculate the nearest neighbor of each sample using Euclidean distance, and label the sample pairs with different classes as Tomek Link pairs; Remove all boundary noise samples corresponding to Tomek Link pairs from the training set to complete majority class undersampling, so that the number of samples of each class in the training set is balanced.

[0011] Furthermore, in S3, the specific steps of the multi-metric fusion method are as follows: The similarity between any two graph nodes is calculated using three metrics: Gaussian kernel function, cosine similarity, and Pearson correlation coefficient, respectively, and three adjacency matrices are generated accordingly. The arithmetic mean of the corresponding elements of the three adjacency matrices is taken to obtain the fused composite adjacency matrix; For each node in the comprehensive adjacency matrix, only the k nearest neighbor nodes with the highest association strength are retained, and the connection weights of the remaining nodes are reset to 0, resulting in the final k-nearest neighbor graph adjacency matrix A.

[0012] Furthermore, the calculation formulas for the three measurement methods are as follows: Gaussian kernel function:

[0013] ,in x is the Gaussian kernel bandwidth parameter. i x j These are the feature vectors of node i and node j, respectively; Cosine similarity:

[0014] Pearson correlation coefficient:

[0015] Where T is the total length of the time-series sampling. , are the time-series mean values ​​of the features of node i and node j, respectively.

[0016] Furthermore, in S4, the residual graph convolutional network consists of 5 stacked feature extraction units. Each feature extraction unit includes a graph convolutional layer, a batch normalization layer, a ReLU activation function layer, and an EdgePool graph pooling layer in sequence. A residual connection is set every other feature extraction unit to jump the input features of the unit to the output of the unit. The end of the network is connected to two fully connected layers and a softmax classification layer in sequence to output a multi-source fusion feature vector and a preliminary classification result.

[0017] Furthermore, the output expression of a single feature extraction unit is: X' = ​​ReLU(BN(GConv(A·X·W))), where A is the k-nearest neighbor graph adjacency matrix, X is the input feature matrix, W is the trainable weight matrix of the graph convolutional layer, BN is the batch normalization operation, and ReLU is the linear rectified activation function. The output expression after the EdgePool graph pooling layer is: X output =EdgePool(X') The residual connection adds the input features of the nth feature extraction unit to the output features of the (n+1)th feature extraction unit element by element, and then inputs them into the subsequent network layers.

[0018] Furthermore, in S5, the specific steps of the game theory-based comprehensive weighting are as follows: The subjective weight vectors of each traditional state evaluation index are determined by using the analytic hierarchy process and combining the experience of domain experts. The entropy weight method is used to calculate the objective weight vector of each feature component based on the information entropy of the deep fusion features; A game theory optimization model is constructed, with the objective function being to minimize the deviation between the comprehensive weight and the subjective and objective weights, and the optimal comprehensive weight coefficient is obtained by solving the problem. The indicators and feature components are weighted and summed according to the optimal comprehensive weight to obtain a comprehensive health status score with a value range of 0-100.

[0019] Furthermore, in S6: The health level classification rules are as follows: a comprehensive health score of ≥80 is very good, 60≤score<80 is good, 40≤score<60 is moderate, 20≤score<40 is poor, and a score<20 is very poor. Fault types are identified by the softmax classifier at the end of the residual graph convolutional network, which can identify four typical fault types: overheating fault, partial discharge fault, winding mechanical fault, and insulation deterioration fault. The remaining life prediction method uses a Weibull distribution model to model the degradation trend by fitting the degradation curve of historical health scores, and predicts the remaining time for the transformer to reach the failure threshold. Based on the numerical range of the comprehensive health score, short-term evolution trend, and fault type matching degree, a first-level trend warning, a second-level mode warning, and a third-level sudden change warning are triggered respectively, and a maintenance decision support report is generated simultaneously.

[0020] The beneficial effects of this invention are: 1. Effectively solves the problem of data imbalance. The SMOTE Tomek hybrid sampling algorithm is adopted, which combines minority class oversampling with majority class boundary denoising undersampling. This not only expands the scale of scarce fault samples, but also eliminates noisy overlapping samples at class boundaries, effectively alleviating the problem of missed fault detection caused by data bias in the model, and providing a balanced and high-quality dataset for model training.

[0021] 2. Achieve deep fusion of multi-source heterogeneous data. Multiple monitoring parameters, such as temperature, strain, partial discharge, oil chromatography, and vibration, are uniformly modeled as graph nodes. Adjacency relationships are constructed through the fusion of three metrics: Gaussian kernel function, cosine similarity, and Pearson correlation coefficient, accurately characterizing the nonlinear coupling relationships between multiple parameters. Combined with a graph convolutional network with residual connections, deep graph structural features are extracted. This avoids the problems of over-smoothing and performance degradation in deep networks, while fully mining collaborative fault information from multiple sources, improving the ability to detect early, subtle faults.

[0022] 3. Improve the objectivity and accuracy of status assessment. Introduce a game theory optimization framework to integrate subjective expert experience with objective data information, and adaptively calculate the optimal comprehensive weight. This avoids the subjective bias of a single weighting method, making the health status score more consistent with the actual operating status of the equipment, and the accuracy of status assessment is no less than 95%.

[0023] 4. Provides comprehensive operation and maintenance decision support. It simultaneously outputs health level, fault type identification results, and remaining service life prediction results, along with a three-level early warning mechanism and maintenance decision reports. This enables a full-chain assessment from condition awareness and fault location to service life prediction, providing comprehensive decision-making basis for predictive maintenance throughout the transformer's entire life cycle. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the 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.

[0025] Figure 1 This is a schematic diagram of the overall process of the transformer intelligent condition assessment method of the present invention.

[0026] Figure 2This is a schematic diagram illustrating the principle of constructing a k-nearest neighbor graph through multi-metric fusion in this invention.

[0027] Figure 3 This is a schematic diagram of the residual graph convolutional network of the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] In the description of this application, it should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. For ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following drawings denote similar items; therefore, once an item is defined in one drawing, it need not be further discussed in subsequent drawings.

[0030] This embodiment provides a smart state assessment method for transformers based on SMOTE Tomek and graph convolutional networks. The specific implementation steps are as follows: Step 1: Multi-source monitoring data acquisition and dataset partitioning Corresponding sensing devices are deployed in key components such as the transformer body, high-voltage bushings, windings, and cooling system to simultaneously collect multi-dimensional heterogeneous monitoring data, specifically including: Oil chromatography data: Concentration time-series data of seven characteristic gases, namely H2, CH4, C2H6, C2H4, C2H2, CO, and CO2, were obtained through an online oil chromatography monitoring device, with a sampling interval of 1 hour; Vibration data: Vibration acceleration signals on the surface of the oil tank were collected by a triaxial accelerometer. The original sampling frequency was 5kHz. After preprocessing, time-domain statistical features and frequency-domain energy features were extracted. Partial discharge data: Discharge signals are acquired using an ultra-high frequency partial discharge detection device, and discharge amplitude, discharge frequency, and phase distribution characteristics are extracted. Temperature and strain data: Top layer oil temperature, winding hot spot temperature, and winding strain monitoring data are collected through a fiber optic sensing system; Electrical parameter data: Three-phase voltage, three-phase current, active power, and reactive power operation data are collected through current and voltage transformers.

[0031] Based on historical transformer maintenance records, insulation test results, and industry health index standards, all samples were categorized into five operating states according to their health index values: 0.8–1.0 ("Very Good"), 0.6–0.8 ("Good"), 0.4–0.6 ("Moderate"), 0.2–0.4 ("Poor"), and 0–0.2 ("Very Poor"). The training and test sets were randomly divided in a 7:3 ratio. The training set was used for model training and weight calculation, while the test set was used to verify the model's evaluation accuracy.

[0032] Step 2: SMOTE Tomek Mixture Sampling and Data Normalization To address the class imbalance problem caused by a large number of normal samples and a small number of faulty samples in the training set, the SMOTE Tomek mixture sampling algorithm is used for balancing preprocessing. The specific process is as follows: (1) Minority class oversampling: Identify the fault class sample with the fewest samples in the training set, randomly select a reference sample x from the minority class samples, calculate the Euclidean distance from the reference sample to all other samples of the same class, and select the 5 closest samples as the nearest neighbor set; randomly select a target sample x from the nearest neighbor set. n According to the formula Generate new minority class samples and add them to the training set. Repeat the above operation until the number of minority class samples reaches the set equilibrium value.

[0033] (2) Majority undersampling: Traverse all samples in the training set and calculate the nearest neighbor of each sample by Euclidean distance. If a pair of samples that are the nearest neighbors belong to different categories, mark the sample pair as a Tomek Link pair. Remove the boundary noise samples in all Tomek Link pairs from the training set to eliminate the overlap and interference samples of the category boundaries.

[0034] After mixed sampling, the number of samples in the five health status categories of the training set is basically balanced, avoiding model bias towards the majority class and improving the recall rate of fault identification.

[0035] After sampling, the min-max normalization method is used to linearly scale all feature data in both the training and test sets. The calculation formula is as follows:

[0036] Where x is the original data value, min and max are the minimum and maximum values ​​of the corresponding feature columns, respectively. After normalization, the data value range is unified to [0,1], eliminating the influence of differences in the dimensions of different parameters on the model and scoring calculation.

[0037] Step 3: Constructing a k-nearest neighbor graph structure through multi-metric fusion Each monitoring parameter is mapped to a node in a graph structure, and the strength of the association between nodes is used as the weight of the edges to construct an undirected graph structure, as shown in Figure 2. To comprehensively characterize the association characteristics between parameters, three metrics are used to calculate node similarity: (1) Gaussian kernel function: It measures the similarity of vectors in a high-dimensional space through a nonlinear mapping. The calculation formula is as follows: ,in x is the Gaussian kernel bandwidth parameter. i x j These are the feature vectors of node i and node j, respectively; Cosine similarity: measures the degree of similarity between two feature vectors in terms of their directions. The formula is as follows:

[0038] Pearson correlation coefficient: measures the degree of linear correlation between the time-series trends of two variables. The formula is as follows: Where T is the total length of the time-series sampling. , are the time-series mean values ​​of the features of node i and node j, respectively.

[0039] The arithmetic mean of the corresponding elements of the three adjacency matrices obtained from the above three metrics is used to obtain the fused comprehensive adjacency matrix. To reduce redundant connections and improve computational efficiency, for each node in the comprehensive adjacency matrix, only the top 4 adjacent nodes with the highest association strength are retained, and the connection weights of the remaining nodes are reset to 0, finally obtaining the adjacency matrix A of the k-nearest neighbor graph, thus completing the graph structure construction.

[0040] Step 4: Deep Feature Extraction from Residual Map Convolutional Network A residual graph convolutional network with residual connections is constructed as the backbone network for feature extraction. The specific structure and operational logic are as follows: Feature Extraction Units: The network consists of five stacked feature extraction units. Each unit comprises a graph convolutional layer (GConv), a batch normalization layer (BN), a ReLU activation function layer, and an EdgePool graph pooling layer. The output of a single feature extraction unit is calculated as follows: X' = ​​ReLU(BN(GConv(A·X·W))), where A is the k-nearest neighbor graph adjacency matrix, X is the input feature matrix, W is the trainable weight matrix of the graph convolutional layer, BN is the batch normalization operation, used to accelerate network convergence and suppress overfitting, and ReLU is the linear rectified activation function, which introduces non-linear fitting capability.

[0041] Each feature extraction unit is followed by an EdgePool graph pooling layer. By learning node importance scores, key nodes are selected and retained, and neighborhood features are aggregated, reducing the graph size and computational cost. The output after pooling is: X output =EdgePool(X').

[0042] (2) Residual Connections: A residual connection is set for every feature extraction unit, that is, the input of the first feature extraction unit is connected to the output of the second feature extraction unit, and the input of the third feature extraction unit is connected to the output of the fourth feature extraction unit. The gradient vanishing and feature oversmoothing problems of deep graph convolutional networks are alleviated by identity mapping, thereby improving the feature extraction capability of deep networks. The residual connection adds the input features and the output features of the corresponding unit element by element before inputting them into the subsequent network layers.

[0043] (3) Classification output: After the 5 feature extraction units, two fully connected layers are connected in sequence. The first fully connected layer has a dimension of 64 and the second layer has a dimension of 32. Finally, the softmax classification layer outputs the classification probability of 5 health states and outputs a 128-dimensional multi-source fusion feature vector for subsequent health score calculation.

[0044] During the model training phase, the k-nearest neighbor graph data from the training set is input into the residual graph convolutional network. Iterative training is performed using the cross-entropy loss function and the Adam optimizer. The model classification accuracy is verified through the test set, thus completing the offline model training.

[0045] Step 5: Calculate the health score using a combination of subjective and objective weighting methods based on game theory. To balance expert experience with objective data and improve the scientific rigor and accuracy of health scores, a game theory approach is used to comprehensively weight subjective and objective factors. The specific steps are as follows: (1) Subjective weight calculation: Experts in the field of power equipment operation and maintenance were invited to construct a judgment matrix for eight traditional key state evaluation indicators, including total hydrocarbon content, acetylene concentration, vibration intensity, partial discharge, and top oil temperature, using the analytic hierarchy process (AHP). After passing the consistency test, the subjective weight vector w1 was calculated.

[0046] (2) Objective weight calculation: For the 128-dimensional multi-source fusion features output by the residual graph convolutional network, the entropy weight method is used to calculate the information entropy of each feature component. The difference coefficient is calculated based on the size of the information entropy, and then the objective weight vector w2 of each feature component is obtained.

[0047] (3) Game theory optimization solution: Construct the comprehensive weight expression w = α1w1 + α2w2, take the minimization of the total deviation between the comprehensive weight and the subjective weight and the objective weight as the objective function, establish the game optimization model, and obtain the optimal weight coefficients α1 and α2 by solving, and then obtain the optimal comprehensive weight vector.

[0048] (4) Calculation of comprehensive health score: According to the optimal comprehensive weight, the normalized indicators and feature components are weighted and summed, and the results are linearly mapped to the 0-100 range to obtain the comprehensive health status score of the transformer. The higher the score, the better the health status of the equipment.

[0049] Step 6: Outputting Status Results and Implementing Hierarchical Early Warnings Based on the comprehensive health status score and network classification results, three types of assessment results are output simultaneously, along with a corresponding tiered early warning mechanism: (1) Health level classification: Based on the comprehensive health score range, it is divided into five levels: score ≥80 is "very good", 60≤score<80 is "good", 40≤score<60 is "moderate", 20≤score<40 is "poor", and score<20 is "very poor".

[0050] (2) Fault type identification: The probability of each type of fault is output by the softmax classifier at the end of the residual graph convolutional network. It can identify four typical faults: overheating fault, partial discharge fault, winding mechanical fault, and insulation deterioration fault, and output the fault type corresponding to the highest probability.

[0051] (3) Remaining life prediction: The historical health score time series data of the transformer is accumulated, and the equipment performance degradation trajectory is fitted by Weibull distribution. The remaining life of the transformer when the comprehensive health score drops to 30 points is used as the failure threshold.

[0052] Simultaneously, a three-tiered early warning mechanism is established: when the health score shows a continuous downward trend but remains within the normal range, a level-one trend warning is triggered, prompting attention to changes in equipment status; when the feature vector shows a high degree of matching with typical fault modes, a level-two mode warning is triggered, indicating an early fault risk; when the health score drops sharply or reaches the fault threshold, a level-three sudden change warning is triggered, indicating the need for emergency shutdown and maintenance. After an early warning is triggered, a maintenance decision support report is automatically generated, including anomaly feature analysis, fault risk assessment, and recommended maintenance solutions.

[0053] When running online, the real-time collected transformer monitoring data is processed through the same normalization and k-nearest neighbor graph construction process, and then input into the trained residual graph convolutional network and scoring system to output transformer status assessment results and early warning information in real time.

[0054] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A smart state assessment method for transformers based on SMOTE Tomek and graph convolutional networks. The feature is that it includes the following steps: S1 collects multi-source monitoring data on transformer temperature, strain, partial discharge, oil chromatography, and vibration. Based on the health index, it classifies different health status categories and proportionally splits the dataset into training and testing sets. S2 uses the SMOTE Tomek hybrid sampling algorithm to perform class balancing preprocessing on the training set, and then normalizes the balanced training and test set data. S3 maps various monitoring parameters to graph nodes, uses a multi-metric fusion method to calculate the correlation strength between nodes, and constructs a k-nearest neighbor graph structure that represents the coupling relationship of parameters; S4 constructs a residual graph convolutional network model containing residual connections, inputs k-nearest neighbor graph data into the model for deep feature extraction, and obtains a high-dimensional feature vector of multi-source data fusion. S5 uses game theory to comprehensively assign subjective and objective weights to the deep fusion features and state evaluation indicators, and calculates the comprehensive health status score of the transformer. S6 outputs the transformer health level, fault type, and remaining life prediction results based on the comprehensive health status score.

2. The transformer intelligent state assessment method based on SMOTE Tomek and graph convolutional networks according to claim 1, characterized in that: The multi-source monitoring data includes dissolved gas data in oil, chamber vibration acceleration signal, partial discharge signal, fiber optic temperature and strain monitoring data, and electrical parameter data; among which, the dissolved gas data in oil includes time-series concentration data of H2, CH4, C2H6, C2H4, C2H2, CO, and CO2; the health status is divided into five levels according to the health index value: very good, good, moderate, poor, and very poor, and the ratio of training set to test set is 7:

3.

3. The transformer intelligent state assessment method based on SMOTE Tomek and graph convolutional networks according to claim 2, characterized in that, The specific execution steps of the SMOTE Tomek hybrid sampling algorithm are as follows: S21 Identify minority class fault samples in the training set, randomly select a benchmark sample in the minority class, calculate its Euclidean distance to the rest of the samples in the same class, and select k nearest neighbor samples. S22 randomly selects the target sample from the k nearest neighbor samples, according to the formula. New minority class samples are generated and added to the training set to complete minority class oversampling, where x is the baseline sample and x is the base sample. n For the target nearest neighbor samples, rand(0,1) is a random number between 0 and 1; S23 Iterate through all samples in the training set, calculate the nearest neighbor of each sample using Euclidean distance, and label the sample pairs with different classes as Tomek Link pairs; S24 removes the boundary noise samples corresponding to all Tomek Link pairs from the training set, completing the majority class undersampling and balancing the number of samples in each class of the training set.

4. The transformer intelligent state assessment method based on SMOTE Tomek and graph convolutional networks according to claim 3, characterized in that, The specific steps of the multi-metric fusion method are as follows: S31 uses three metrics—Gaussian kernel function, cosine similarity, and Pearson correlation coefficient—to calculate the similarity between any two graph nodes, generating three adjacency matrices accordingly. S32 takes the arithmetic mean of the corresponding elements of the three adjacency matrices to obtain the merged composite adjacency matrix; S33 For each node in the comprehensive adjacency matrix, only the k neighboring nodes with the highest association strength are retained, and the connection weights of the remaining nodes are reset to 0, resulting in the final k-nearest neighbor graph adjacency matrix A.

5. The transformer intelligent state assessment method based on SMOTE Tomek and graph convolutional networks according to claim 4, characterized in that, The calculation formulas for the three measurement methods are as follows: Gaussian kernel function: ， in x is the Gaussian kernel bandwidth parameter. i x j These are the feature vectors of node i and node j, respectively; Cosine similarity: Pearson correlation coefficient: Where T is the total length of the time-series sampling. , are the time-series mean values ​​of the features of node i and node j, respectively.

6. The transformer intelligent state assessment method based on SMOTE Tomek and graph convolutional networks according to claim 5, characterized in that: The residual graph convolutional network consists of five stacked feature extraction units. Each feature extraction unit includes a graph convolutional layer, a batch normalization layer, a ReLU activation function layer, and an EdgePool graph pooling layer. A residual connection is set every other feature extraction unit to jump the input features of the unit to the output of the unit. The network ends with two fully connected layers and a softmax classification layer, outputting a multi-source fusion feature vector and a preliminary classification result.

7. The transformer intelligent state assessment method based on SMOTE Tomek and graph convolutional networks according to claim 6, characterized in that, The output expression of a single feature extraction unit is: X' = ​​ReLU(BN(GConv(A·X·W))), where A is the k-nearest neighbor graph adjacency matrix, X is the input feature matrix, W is the trainable weight matrix of the graph convolutional layer, BN is the batch normalization operation, and ReLU is the linear rectified activation function. The output expression after the EdgePool graph pooling layer is: X output =EdgePool(X') The residual connection adds the input features of the nth feature extraction unit to the output features of the (n+1)th feature extraction unit element by element, and then inputs them into the subsequent network layers.

8. The transformer intelligent state assessment method based on SMOTE Tomek and graph convolutional networks according to claim 7, characterized in that, The specific steps of the game theory-based comprehensive weighting system are as follows: S51 uses the analytic hierarchy process (AHP) and combines the experience of domain experts to determine the subjective weight vectors of each traditional state evaluation index. S52 uses the entropy weight method to calculate the objective weight vector of each feature component based on the information entropy of the deep fusion features; S53 Construct a game theory optimization model with the objective function of minimizing the deviation between the comprehensive weight and the subjective and objective weights, and solve for the optimal comprehensive weight coefficient. S54 performs a weighted summation of each indicator and feature component according to the optimal comprehensive weight, and maps it to obtain a comprehensive health status score with a value range of 0-100.

9. The transformer intelligent state assessment method based on SMOTE Tomek and graph convolutional networks according to claim 8, characterized in that: The health level classification rules are as follows: a comprehensive health score of ≥80 is very good, 60≤score<80 is good, 40≤score<60 is moderate, 20≤score<40 is poor, and a score<20 is very poor. Fault types are identified by the softmax classifier at the end of the residual graph convolutional network, which can identify four typical fault types: overheating fault, partial discharge fault, winding mechanical fault, and insulation deterioration fault. The remaining life prediction method uses a Weibull distribution model to model the degradation trend by fitting the degradation curve of historical health scores, and predicts the remaining time for the transformer to reach the failure threshold. Based on the numerical range of the comprehensive health score, short-term evolution trend, and fault type matching degree, a first-level trend warning, a second-level mode warning, and a third-level sudden change warning are triggered respectively, and a maintenance decision support report is generated simultaneously.

Images