A method, medium and system for predicting multi-point grounding defects of transformer core and clamps

Through multi-mode data preprocessing and feature fusion, combined with ant colony optimization and LSTM model fine-tuning, the problem of poor training stability of neural network models in the prediction of multi-point grounding defects of transformer cores and clamps is solved, and efficient multi-point grounding defects are achieved Detection and troubleshooting.

CN118428217BActive Publication Date: 2024-11-05ELECTRIC POWER RESEARCH INSTITUTE OF STATE GRID NINGXIA ELECTRIC POWER COMPANY +6
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202410526499.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-04-29
Publication Date
2024-11-05
Estimated Expiration
2044-04-29

AI Technical Summary

Technical Problem

When the prior art uses neural network models to predict multi-point grounding defects of transformer cores and clamps, the error data present in the actual sample leads to poor training stability and insufficient robustness.

Method used

Multimode data preprocessing, propensity error analysis and Bayesian analysis combined with ant colony optimization algorithm is used to build a fusion feature matrix, and the particle swarm algorithm is used to fine-tune the LSTM model to achieve robust detection of multi-point grounding defects.

Benefits of technology

The abnormal component feature representation and fusion capability of multi-source heterogeneous data is improved, the fitting ability of complex data distribution is enhanced, the discrimination ability and interpretability of detection is improved, and the fault diagnosis and prevention capabilities are good.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118428217B_ABST
    Figure CN118428217B_ABST
Patent Text Reader

Abstract

The invention provides a method, medium and system for predicting multi-point grounding defects of transformer core and clamps, belonging to the technical field of transformer core and clamps, and comprising: obtaining multi-mode data of transformer operation; preprocessing, deleting conventional components and extracting abnormal components; performing tendency error analysis and identifying a tendency error matrix of multi-point grounding defects; performing Bayesian analysis and identifying a Bayesian flying point matrix; fusing the tendency error matrix and the Bayesian flying point matrix to obtain an original feature matrix; optimizing the original feature matrix by using an ant colony algorithm to obtain an optimized feature matrix; calculating the correlation of abnormal components of each multi-mode data of the optimized feature matrix and constructing a fused feature matrix; training a pre-trained LSTM model to obtain a fine-tuned LSTM model; inputting the fused feature matrix into the fine-tuned LSTM model to obtain a multi-point grounding defect index of the core and clamps of the transformer to be tested.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of transformer cores and clamps, and in particular, relates to a method, medium and system for predicting multi-point grounding defects of transformer cores and clamps. Background Art

[0002] With the rapid development of industrial automation, Internet of Things, financial transactions and other fields, real-time multi-source heterogeneous data anomaly detection technology has become a hot topic in current research. Since data in the real world usually has complex noise, missing, skew and other problems, higher requirements are placed on the robustness and discrimination ability of real-time detection systems. At the same time, in order to ensure the efficiency of operation, the anomaly detection algorithm itself must also have high computational efficiency. Existing technologies still have some shortcomings in solving this problem, and new solutions are urgently needed.

[0003] Industrial process monitoring is one of the main application scenarios of anomaly detection of multi-source heterogeneous data. In modern intelligent manufacturing systems, multiple modal data such as electrical signals, temperature, vibration, and flow are often involved. These data reflect different aspects of the operating status of the equipment and have certain correlations with each other. Only by considering them comprehensively can the operating status of the system be fully evaluated. The traditional statistical process control (SPC) method only considers the statistical characteristics of single modal data and cannot improve the detection accuracy by using the correlation between multi-source data. At present, neural networks are developing rapidly, and methods for predicting multi-point grounding defects of transformer cores and clamps using neural network models have also emerged. Specifically, the low-dimensional potential representation of the input data is learned, and then the input is reconstructed back to the original space. However, this method requires a large number of labeled samples for supervised training. The erroneous data in the actual samples will cause the trained neural network model to have poor training stability and poor interpretability, resulting in poor robustness. Summary of the invention

[0004] In view of this, the present invention provides a method, medium and system for predicting multi-point grounding defects of transformer cores and clamps, which can solve the technical problem that when the prior art uses a neural network model to predict multi-point grounding defects of transformer cores and clamps, the erroneous data existing in the actual samples will cause the trained neural network model to have poor training stability and poor robustness.

[0005] The present invention is achieved in that:

[0006] A first aspect of the present invention provides a method for predicting multi-point grounding defects of a transformer core and a clamp, which comprises the following steps:

[0007] S10, obtaining multi-mode data during transformer operation, including electrical data, partial discharge ultrasonic data of the grounding points of the core and the clamps, grounding current data of the core and the clamps, and vibration data of the core and the clamps;

[0008] S20, preprocessing each multi-mode data, deleting regular components, and extracting abnormal components;

[0009] S30, performing a tendency error analysis on the abnormal components of each multi-mode data, and identifying a tendency error matrix of multi-point grounding defects;

[0010] S40, performing Bayesian analysis on the abnormal components of each multi-mode data to identify a Bayesian flying point matrix;

[0011] S50, fusing the tendency error matrix and the Bayesian flying point matrix to obtain an original feature matrix; and optimizing the original feature matrix using an ant colony algorithm to obtain an optimized feature matrix;

[0012] S60, calculating the correlation of abnormal components of each multi-mode data of the optimized feature matrix;

[0013] S70, constructing a fusion feature matrix for the abnormal components of each type of multimodal data according to the correlation;

[0014] S80, using a particle swarm algorithm based on the optimized feature matrix to train the pre-trained LSTM model to obtain a fine-tuned LSTM model, wherein the LSTM model is used to determine the multi-point grounding defect index of the core and the clamp of the transformer to be tested based on the fusion feature matrix of the transformer to be tested;

[0015] S90, inputting the fused feature matrix into the fine-tuned LSTM model to obtain the multi-point grounding defect index of the core and the clamp of the transformer to be tested.

[0016] On the basis of the above technical solution, the transformer core and clamp multi-point grounding defect prediction method of the present invention can also be improved as follows:

[0017] Among them, the specific implementation steps of step S10 are: for each modal data among the electrical data, partial discharge ultrasonic data of the grounding point of the core and the clamp, the grounding current data of the core and the clamp, and the vibration data of the core and the clamp, use wavelet transform to reduce noise and preprocess them, and use wavelet packet decomposition to extract their characteristic energy distribution; then use the simultaneous inverse problem optimization method to reconstruct the characteristic energy distribution into abnormal component characteristics; wherein the simultaneous inverse problem optimization method is solved by the alternating direction multiplier algorithm.

[0018] Among them, the specific implementation steps of step S20 are: for each type of abnormal component feature extracted from the modal data, a kernel principal component analysis algorithm is used to reduce its dimension and map it to a low-dimensional space, and the kernel principal component analysis algorithm uses a Gaussian kernel function to achieve implicit mapping; then the similarity distance between all sample points is calculated in the low-dimensional space to construct a similarity matrix.

[0019] Furthermore, for the constructed similarity matrix, the spectral clustering method is used to cluster the sample points with high similarity to obtain a cluster set; the principal component analysis is performed on the cluster set to determine the main direction of the tendency error, that is, the first column of the tendency error matrix; then the other columns of the tendency error matrix are iteratively determined until most of the variance is explained or the preset rank threshold is reached.

[0020] Among them, the specific implementation steps of step S40 are: for each abnormal component, use the expectation maximization algorithm to fit it as a mixture of several Gaussian distributions; calculate the probability density of each observation value under different Gaussian distributions, and update the probability according to the Bayesian theorem; mark the observation value with a probability less than a preset threshold as a flying point, and feature encode the flying point to obtain a Bayesian flying point matrix.

[0021] Among them, the specific implementation steps of step S50 are: splicing the tendency error matrix and the Bayesian flying point matrix by column to construct an original feature matrix; then using the ant colony optimization algorithm to optimize the original feature matrix, the ant colony optimization algorithm includes: constructing a fully connected feature construction graph, allowing multiple artificial ants to randomly walk on the graph to construct feature subset paths; evaluating the fitness of each subset path, leaving more concentrated pheromones on paths with higher fitness; after multiple iterations, determining the optimal feature subset as the optimized feature matrix.

[0022] Among them, the specific implementation steps of step S60 are: extracting the characteristic sub-matrices corresponding to each type of multimodal data from the optimized characteristic matrix; for each pair of abnormal component feature vectors in the sub-matrix, calculating the correlation coefficients between them, including the Pearson correlation coefficient, the Spearman correlation coefficient and the mutual information; and saving all the calculated correlation coefficient values ​​in the correlation matrix.

[0023] Among them, the specific implementation steps of step S70 are: based on the correlation matrix, the abnormal component feature vector of each multimodal data is regarded as a node in the abnormal component feature graph, and the correlation is used as the weight of the edge to construct a graph; the spectral clustering method is used on the abnormal component feature graph to find K fusion nodes that can summarize the entire topological structure; the feature vectors corresponding to the K fusion nodes are spliced ​​by column to obtain a fusion feature matrix.

[0024] Optionally, K is a positive integer greater than 5.

[0025] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the above-mentioned method for predicting multi-point grounding defects of transformer cores and clamps.

[0026] A third aspect of the present invention provides a transformer core and clamp multi-point grounding defect prediction system, which includes the above-mentioned computer-readable storage medium.

[0027] Compared with the prior art, the transformer core and clamp multi-point grounding defect prediction method, medium and system provided by the present invention have the following beneficial effects:

[0028] 1) It realizes the feature representation and fusion of abnormal components of multi-source heterogeneous data, which can make full use of the complementarity between different modal data, without the need for manual labeling of large amounts of training data, and improve the discriminative ability of anomaly detection.

[0029] 2) By encoding the flying point features of abnormal components, robust processing of complex situations such as noise, missing values, and outliers is achieved.

[0030] 3) Feature construction is realized based on the ant colony optimization algorithm, and high-order feature combinations are automatically generated, which further enhances the anomaly detection model's ability to fit complex data distributions.

[0031] 4) The modal source information is retained, and the anomaly detection results have good interpretability, which is easy for fault diagnosis and prevention.

[0032] In summary, the technical solution of the present invention solves the technical problem that when the prior art uses a neural network model to predict multi-point grounding defects of transformer cores and clamps, the erroneous data existing in the actual samples will cause the trained neural network model to have poor training stability and poor robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0034] Figure 1 A flow chart of the method provided by the present invention. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0036] like Figure 1 As shown, it is a flow chart of a method for predicting multi-point grounding defects of transformer core and clamps provided by the present invention. The method comprises the following steps:

[0037] S10, obtaining multi-mode data during transformer operation, including electrical data, partial discharge ultrasonic data of the grounding points of the core and the clamps, grounding current data of the core and the clamps, and vibration data of the core and the clamps;

[0038] S20, preprocessing each multi-mode data, deleting regular components, and extracting abnormal components;

[0039] S30, performing a tendency error analysis on the abnormal components of each multi-mode data, and identifying a tendency error matrix of multi-point grounding defects;

[0040] S40, performing Bayesian analysis on the abnormal components of each multi-mode data to identify a Bayesian flying point matrix;

[0041] S50, fusing the tendency error matrix and the Bayesian flying point matrix to obtain an original feature matrix; and optimizing the original feature matrix using an ant colony algorithm to obtain an optimized feature matrix;

[0042] S60, calculating the correlation of abnormal components of each multi-mode data of the optimized feature matrix;

[0043] S70, constructing a fusion feature matrix for the abnormal components of each multi-mode data according to the correlation;

[0044] S80, using a particle swarm algorithm based on an optimized feature matrix to train a pre-trained LSTM model to obtain a fine-tuned LSTM model, the LSTM model is used to determine the multi-point grounding defect index of the core and the clamp of the transformer to be tested based on the fusion feature matrix of the transformer to be tested;

[0045] S90. Input the fused feature matrix into the fine-tuned LSTM model to obtain the multi-point grounding defect index of the core and the clamp of the transformer to be tested.

[0046] The specific implementation methods of the above steps are described in detail below:

[0047] The specific implementation method of step S10 is: obtaining multi-mode data of the transformer during operation, including electrical data, partial discharge ultrasonic data of the grounding point of the core and the clamp, grounding current data of the core and the clamp, and vibration data of the core and the clamp. These four types of data cover multiple aspects such as electrical characteristics, partial discharge conditions, grounding current and mechanical vibration of the transformer during operation, and can fully reflect the operating status of the transformer. By simultaneously collecting these four types of data, the operating status of the transformer can be better described, providing a rich source of information for subsequent fault diagnosis and prediction.

[0048] Electrical data can be electrical parameters such as voltage, current, and power of the transformer body and other auxiliary equipment. These parameters can reflect the basic electrical characteristics of the transformer. The partial discharge ultrasonic data of the core and clamp grounding point refers to the partial discharge ultrasonic signals collected at different grounding points of the transformer core and clamps. These ultrasonic signals can reflect the partial discharge of the transformer insulation, thereby inferring the insulation state. The core and clamp grounding current data measures the grounding current flowing through different grounding points of the transformer core and clamps. Abnormalities in these current values ​​may indicate the presence of a grounding fault or other problems. Finally, the core and clamp vibration data refers to the vibration signals collected at different positions of the transformer core and clamps. It can reflect the mechanical state of the transformer core and clamps. For example, severe vibration may mean that the core and clamps have deformed or other mechanical damage.

[0049] The collection methods of these four types of data vary depending on the type of data. Electrical data can be obtained through voltage and current sensors installed on the transformer body or auxiliary equipment; partial discharge ultrasonic data requires ultrasonic sensors to be arranged at different grounding points of the transformer core and clamps for collection; ground current data can be directly measured with current sensors or ground ammeters; and vibration data requires acceleration sensors to be installed on the surface of the core and clamps to detect vibration. All of these sensors need to have sufficient accuracy and sampling frequency to ensure that high-quality data is obtained. In addition, in actual applications, different transformer types and operating environments may require adjustments to the layout and number of sensors to obtain the best data coverage.

[0050] The specific implementation of step S20 is: pre-processing each multi-mode data obtained, deleting the normal components, and extracting the abnormal components. The purpose of this step is to remove the normal components from the original data and only retain the abnormal components that may indicate a fault, thereby facilitating subsequent fault detection and diagnosis.

[0051] Specifically, for electrical data, time-frequency analysis methods such as wavelet analysis and empirical mode decomposition (EMD) can be used to decompose the data into components of different time scales and frequencies. Usually, electrical quantities such as current and voltage in normal operation will show a certain periodicity and stability, and the corresponding components are normal components; while in abnormal situations, due to the influence of faults, some mutations or random components will be introduced, which are abnormal components. By setting a suitable threshold, the normal components can be filtered out and only the abnormal components are retained.

[0052] For ultrasonic data, denoising (such as wavelet denoising) can be performed first, and then the signal can be feature extracted using statistical features such as energy and kurtosis to distinguish the features of abnormal discharge. Under normal circumstances, these statistical features of ultrasonic signals should fluctuate within a certain range; once an abnormality occurs, it will deviate from the normal range, and this part of the signal is the abnormal component.

[0053] The preprocessing of ground current data can first remove the power frequency component and harmonic component, then perform statistical analysis on the residual components, and identify abnormal components using indicators such as peak factor and pulse factor. Under normal circumstances, these statistical indicators of ground current should fluctuate within the normal range, and once an abnormality occurs, it will exceed the normal threshold range.

[0054] As for vibration data, the original signal needs to be detrended and de-baselined first, and then the abnormal components can be extracted based on the signal's time domain statistical characteristics (such as root mean square, kurtosis, etc.) or frequency domain characteristics (such as power spectrum density, etc.). Usually, these statistical characteristics of vibration signals should be relatively stable under normal working conditions, but will fluctuate violently once an abnormality occurs. This part of the fluctuation component is the abnormal component.

[0055] Different preprocessing methods correspond to different threshold settings, and the selection of thresholds requires a combination of experience and a large amount of data analysis. For example, for wavelet analysis, the threshold can be set based on the variance or other statistics of the wavelet coefficients at each scale; for EMD, anomalies can be identified based on the periodicity and energy distribution of each intrinsic mode function; for statistical features such as energy and kurtosis, the threshold range can be determined based on the distribution of these features under normal working conditions. Usually, 3σ is a more commonly used threshold reference value, that is, the abnormal component should exceed the range of 3 times the standard deviation of the normal value.

[0056] The specific implementation of step S30 is: performing a tendency error analysis on the abnormal components of each multi-mode data to identify the tendency error matrix of the multi-point grounding defect. The purpose of the tendency error analysis is to find the correlation or coupling relationship between the abnormal components caused by the multi-point grounding defect, so as to better characterize the characteristic mode of the defect.

[0057] Specifically, we first need to construct a high-dimensional data matrix containing all abnormal components, where each row represents a time point and each column represents an abnormal component. Then, we perform principal component analysis (PCA) or other dimensionality reduction techniques on this matrix to project the high-dimensional data into a low-dimensional feature space. In this low-dimensional space, due to the existence of multiple grounding defects, the abnormal components will show a certain correlation or coupling relationship, and these related components will gather in the same direction, forming a tendency error direction.

[0058] In order to quantify this bias error, we can calculate the distance from each sample point to all other sample points in the low-dimensional space, and then build a similarity matrix based on these distances. This matrix reflects the degree of similarity between different sample points. The larger the value, the more similar the two sample points are. Next, we perform spectral clustering analysis on the similarity matrix to cluster sample points with high similarity into the same cluster. Due to the bias error, these clusters will be distributed along a specific direction.

[0059] By performing principal component analysis on the distribution of these clusters, the main direction of the tendency error, that is, the direction of the first principal component, can be determined. This direction is a column in the tendency error matrix. By similar methods, other columns can be further solved until the rank of the matrix reaches the preset rank threshold or explains most of the variance; the explanation here means the meaning of representation and embodiment, and most of it can be set according to experience, such as more than 90%. The final matrix is ​​the desired tendency error matrix.

[0060] In practical applications, the above-mentioned tendency error analysis can be performed for different multi-mode data to obtain the corresponding tendency error matrix. For the setting of thresholds, the rank threshold can usually be 5 to 10, and the variance contribution rate threshold can be 80% to 95%.

[0061] It should be noted that the above method describes a method of analyzing the tendency error based on spectral clustering. In fact, other clustering algorithms (such as K-Means, density clustering, etc.) or unsupervised learning methods (such as autoencoders, etc.) can also be used to discover the pattern of tendency errors. Different methods correspond to different calculation steps, but in essence, they all discover the correlation and offset caused by defects from the distribution of abnormal components.

[0062] The specific implementation of step S40 is: perform Bayesian analysis on the abnormal components of each multi-mode data to identify the Bayesian flying point matrix. The purpose of this step is to find abnormal values ​​or outliers caused by defects from the abnormal components, and encode these abnormal values ​​to form a feature matrix.

[0063] Bayesian analysis is a reasoning method based on Bayesian theorem that uses prior knowledge and observation data to update the probability of an event. Here, each abnormal component is regarded as a random variable, and its distribution can be modeled using a Gaussian mixture model or other parameter models. Specifically, for each abnormal component, its distribution can be fitted as a mixture of several Gaussian distributions using the expectation maximization (EM) algorithm or other unsupervised clustering algorithms.

[0064] After obtaining these Gaussian distribution models, the probability density of each observation under different Gaussian distributions can be calculated, and these probabilities can be updated according to Bayes' theorem. Usually, most observations fall into the main Gaussian distribution, with a higher corresponding probability, while outliers fall into areas with lower probabilities. Therefore, a threshold can be set to mark observations with probabilities less than the threshold as outliers or flying spots.

[0065] For the observations marked as flying spots, they can be encoded into a feature vector, and the encoding method can be one-hot encoding or other appropriate encoding methods. In this way, for each abnormal component, a feature vector describing the distribution of its abnormal values ​​can be obtained. These feature vectors of all abnormal components are concatenated in columns to form the required Bayesian flying spot matrix.

[0066] In actual operation, the above Bayesian analysis can be performed on different multimodal data to obtain the corresponding Bayesian flying point matrix. For the probability threshold, it can usually be taken in the order of 0.01 to 0.05. In addition, in addition to the Gaussian mixture model, other density estimation methods (such as kernel density estimation, etc.) can also be used to model the distribution of abnormal components to discover flying points.

[0067] It should be pointed out that the Bayesian flying point matrix describes the position information of the outliers in the entire anomaly component distribution, which is complementary to the correlation or offset pattern described by the tendency error matrix. Therefore, in the subsequent feature fusion, both will be considered together.

[0068] The specific implementation method of step S50 is: fusing the tendency error matrix and the Bayesian flying point matrix to obtain the original feature matrix; and optimizing the original feature matrix using the ant colony algorithm to obtain the optimized feature matrix.

[0069] First, for the fusion part, the two matrices can be concatenated column by column to form a higher-dimensional feature matrix. This is equivalent to merging the tendency error features and flying point features into the same feature space. Since these two types of features describe different aspects of the distribution of abnormal components, fusing them is conducive to improving the integrity of feature representation.

[0070] After fusion, the original feature matrix can be further optimized to improve the separability and robustness of the features. Ant colony optimization algorithm is used here.

[0071] The ant colony algorithm is a random search algorithm that simulates the foraging behavior of ants. Here, feature selection is regarded as a combinatorial optimization problem: from all possible feature subsets, an optimal subset is selected so that the model trained on this subset has the best classification performance.

[0072] Specifically, we first need to build a fully connected feature construction graph, where each node represents a feature dimension. Then, we use multiple artificial ants to randomly walk on this graph, and each ant will construct a possible feature subset path. During the walking process, the ants will give priority to paths that have been selected more times in the past and contain better features, and there will also be a certain probability of exploration to choose new paths.

[0073] After each ant constructs a path, it can train a supervised model (such as support vector machine, random forest, etc.) based on the feature subset corresponding to the path, and evaluate its performance on the validation set. Based on this performance value, the fitness of the path can be calculated, and the corresponding concentration of pheromone can be left on the path. After multiple iterations, the pheromone concentration on the optimal path will be much higher than that on other paths, so that the best feature subset can be determined.

[0074] In the implementation process, the convergence speed and optimization of the algorithm can be balanced by controlling parameters such as the ant colony size, the number of iterations, and the pheromone concentration update strategy. Some other parameters need to be set, such as the pheromone volatility coefficient and the heuristic guidance parameter, which control the ants' preference for historical information and heuristic information when constructing solutions. It is usually necessary to tune these parameters in a large number of experiments to obtain satisfactory optimization results.

[0075] In addition to the ant colony algorithm, other optimization algorithms can also be considered to optimize the feature matrix, such as genetic algorithm, particle swarm optimization, simulated annealing, etc. Different algorithms have different advantages and disadvantages, and need to be selected according to the characteristics of the specific problem.

[0076] After the above optimization, an optimized feature matrix can be obtained, in which only the most discriminative feature dimensions for multi-point grounding defect detection are retained, and the expression of these features is also more compact and robust.

[0077] The specific implementation method of step S60 is: calculating the correlation of the abnormal components of each multimodal data in the optimized feature matrix.

[0078] The correlation reflects the degree of correlation between different abnormal components, which provides a basis for subsequent feature fusion.

[0079] Specifically, the correlation can be calculated using the following method:

[0080] First, it is necessary to extract the characteristic sub-matrices corresponding to each type of multi-mode data from the optimized feature matrix, such as electrical data features, ultrasonic data features, etc.

[0081] Next, for each pair of abnormal component feature vectors, the correlation coefficients between them can be calculated, including Pearson correlation coefficient, Spearman correlation coefficient, mutual information, etc. The larger the value of these correlation coefficients, the stronger the correlation between the two feature vectors.

[0082] For the Pearson correlation coefficient, the following formula can be used to calculate:

[0083] r=Σ((X-mean(X))(Y-mean(Y))) / (sqrt(Σ(X-mean(X))^2)*sqrt(Σ(Y-mean(Y))^2))

[0084] Where X and Y represent two eigenvectors, and mean(X) and mean(Y) represent their means.

[0085] The Spearman correlation coefficient is calculated based on the rank of the feature vector rather than the original value. The formula is as follows:

[0086] r=1-(6*Σ(d^2)) / (n*(n^2-1))

[0087] Where d is the difference in rank between corresponding elements of two eigenvectors, and n is the length of the vector.

[0088] Mutual information measures the correlation between two random variables from the perspective of information theory. Its calculation usually requires discretization or density estimation of the eigenvectors first.

[0089] In addition to the above three correlation coefficients, other correlation measures can also be considered, such as the maximum information coefficient, distance correlation coefficient, etc. They each have their own characteristics, and appropriate indicators need to be selected according to specific circumstances.

[0090] After calculating the correlation coefficients of each pair of abnormal component feature vectors, these values ​​can be saved in a correlation matrix, which reflects the correlation between different abnormal component features in the optimized feature matrix.

[0091] The specific implementation method of step S70 is: according to the calculated correlation, the abnormal components of each multimodal data are used to construct a fusion feature matrix.

[0092] The purpose of this step is to effectively fuse the abnormal components from different modal data to form a more complete and robust feature representation to provide input for subsequent LSTM model training.

[0093] Specifically, a graph-based feature fusion method can be used to treat the abnormal component feature vector of each modal data as a node in the graph, and the element value in the correlation matrix is ​​used as the weight of the edge between the nodes. In this way, an abnormal component feature graph is constructed.

[0094] In this graph, the abnormal component features of each modal data are concentrated together to form a local area, called a community. The abnormal component features with high correlation are closely connected and located in the center of the community. Different communities are connected through some edge nodes.

[0095] The goal of fusion is to find a group of nodes on this abnormal component feature graph that can well summarize the topological structure of the entire graph, and concatenate the feature vectors corresponding to these nodes as the fused feature representation.

[0096] Specifically, we can adopt the idea of ​​spectral clustering. First, we calculate the Laplace matrix of the graph:

[0097] L=DW

[0098] Where D is the degree matrix, W is the correlation matrix (as an adjacency matrix). Then solve the minimum K eigenvectors of the Laplacian matrix, and the K nodes corresponding to them are the required fusion node set.

[0099] By concatenating the eigenvectors corresponding to these K nodes by columns, we can obtain the fused feature matrix, which integrates the abnormal component information from each modal data and can better characterize the characteristic patterns of multi-point grounding defects.

[0100] In implementation, the selection of K can be automatically determined by methods such as deviation ratio test, or a fixed value can be given based on experience, such as 5 to 10 fusion nodes.

[0101] In addition to spectral clustering, other graph algorithms or multi-view learning techniques can also be tried to fuse the features of abnormal components. As long as they can take into account the correlation between features and retain sufficient discriminative power, they are all optional solutions.

[0102] The specific implementation method of step S80 is: using the particle swarm algorithm based on the optimized feature matrix to train the pre-trained LSTM model to obtain a fine-tuned LSTM model, which is used to judge the multi-point grounding defect index of the core and clamps of the transformer to be tested based on the fusion feature matrix of the transformer to be tested.

[0103] First, let's introduce the basic principles of the long short-term memory network (LSTM). LSTM is a special recursive neural network designed to better learn long-term dependencies in sequence data. Its core idea is to introduce a data structure called cell state, and control the update and output of the cell state through a special gating mechanism (including forget gate, input gate and output gate). This design enables LSTM to better alleviate the long-term dependency problem of traditional RNN and excels in extracting temporal pattern features from sequence data.

[0104] In this application scenario, the transformer's multimodal data can be regarded as a multivariate time series, and each time step corresponds to a vector containing all modal features. By inputting such sequence data, LSTM can automatically learn the time evolution pattern of multimodal data, thereby modeling and predicting the operating status of the transformer.

[0105] After obtaining the pre-trained model, the optimized feature matrix obtained previously can be used to further fine-tune it. Specifically, the particle swarm optimization (PSO) algorithm can be used to optimize the model parameters using the classification performance of LSTM on the validation set as the fitness function.

[0106] Particle swarm optimization is a random optimization algorithm based on swarm intelligence. It simulates the social behavior of each individual in the cluster making decisions based on their own experience and the experience of the group. In PSO, each parameter to be optimized can be regarded as a "particle", and all particles form a group. Each particle flies in the solution space according to a certain speed formula, and at the same time records its own historical optimal solution and the global optimal solution of the group. In the iterative process, each particle will adjust its position and speed according to these two optimal solutions, so as to gradually approach the optimal solution area.

[0107] Specifically for LSTM parameter optimization, all trainable parameters of LSTM (such as weight matrix, bias vector, etc.) can be straightened into a high-dimensional vector as a particle in PSO. For each candidate particle solution, it is used to initialize LSTM, and the classification performance of the model (such as accuracy, F1, etc.) is evaluated on the training set and validation set as the fitness value of the particle.

[0108] In the PSO iteration, the position and speed of the particles are continuously adjusted according to the historical optimal solution of the particle itself and the group, thereby updating the parameters of the LSTM model. After several generations of iterations, the entire particle group will converge to a near optimal solution, and the LSTM model with the best performance on the optimized feature set will be obtained.

[0109] When implementing, it is necessary to set some hyperparameters of PSO, such as the number of particles, maximum iteration number, inertia weight, acceleration constant, etc. These parameters will affect the convergence speed and optimization effect of the algorithm. Usually, cross-validation can be performed on the validation set first to find a relatively optimal set of parameter configurations.

[0110] In addition to PSO, you can also try other optimization algorithms to fine-tune the parameters of LSTM, such as genetic algorithms, simulated annealing, etc. They have some differences in specific operations, but in essence they all effectively search the parameter space of LSTM.

[0111] After the above fine-tuning, a high-performance LSTM model trained for the optimized feature set is obtained. When new transformer data to be tested is input, its fusion feature matrix can be calculated first, and then input into the LSTM for forward propagation reasoning to obtain the final multi-point grounding defect index evaluation result.

[0112] It should be noted that the above is only one possible solution for LSTM training and fine-tuning. In practical applications, other types of sequence models (such as GRU, Transformer, etc.), different optimization objectives (such as maximizing confidence, etc.), and a combination of multiple optimization algorithms can also be considered to obtain more robust and accurate model performance.

[0113] After the above steps, the invention also discloses a method for online monitoring and evaluation of multi-point grounding defects of transformers based on the obtained LSTM model.

[0114] Specifically, the method comprises the following steps:

[0115] S90) obtaining real-time streaming data of multi-mode data of the transformer to be tested, including electrical data, temperature data, vibration data, fluid data, etc.;

[0116] S100) for each modality data, extract the corresponding abnormal component features, and splice these abnormal component features into a fusion feature matrix to be tested;

[0117] S110) inputting the fused feature matrix to be tested into the trained LSTM model, and obtaining the evaluation result of the LSTM model on the multi-point grounding defect of the transformer, i.e., the defect index value, through forward propagation;

[0118] S120) comparing the defect index value with a preset threshold value, if it is greater than the threshold value, it is determined that there are multiple-point grounding defects, otherwise it is determined that the transformer is operating normally;

[0119] S130) determining the severity of the defect according to the defect index value, and locating the possible fault location in combination with the analysis result of the abnormal component;

[0120] S140) For the transformer determined to be defective, a fault warning is issued and maintenance suggestions are given.

[0121] It can be seen from the above steps that the invention constructs an end-to-end multi-mode data fusion and sequence modeling framework for online monitoring of multi-point grounding defects in transformers.

[0122] The first step is to obtain multi-source heterogeneous monitoring data of the transformer in real time, including multiple modes such as electrical, temperature, vibration, and insulating fluid. These raw data are used to extract abnormal component features in each mode through the anomaly detection method introduced earlier.

[0123] Then, the extracted abnormal component features are fused at the feature level to form a fused feature matrix as the input of the LSTM model. The fused feature matrix integrates the abnormal pattern information of multi-modal data, which is conducive to the model learning a more complete defect feature representation.

[0124] Next, the fused feature matrix is ​​input into the trained LSTM model according to the time step. The model will model the entire sequence data, capture the temporal evolution law, and output a defect index evaluation value.

[0125] The defect index value can reflect whether the current operating state of the transformer is abnormal. A reasonable threshold can be set based on experience. If the index value exceeds the threshold, it is judged that the transformer has multi-point grounding defects; otherwise, the operation is normal.

[0126] For transformers that are determined to have defects, this method can also estimate the severity of the defects based on the defect index value, and locate possible fault locations, such as which windings and grounding points have grounding currents, combined with the analysis results of abnormal components.

[0127] Finally, the method will issue a fault warning for the transformer with defects, reminding the staff that maintenance is needed. At the same time, it can also give corresponding maintenance suggestions, such as the need to replace the winding or rewind, so that the staff can switch equipment or take countermeasures in time to avoid further losses.

[0128] In general, this invention proposes an innovative online monitoring method for transformer multi-point grounding defects based on multi-mode data fusion and LSTM sequence modeling. Compared with traditional methods based on a single mode or artificial empirical rules, this method has the following advantages:

[0129] 1) It integrates multi-source heterogeneous data, making the information richer and more comprehensive;

[0130] 2) Automatically learn the temporal evolution pattern of multimodal data without the need for manual feature design;

[0131] 3) It can give continuous defect scores instead of simple binary classification results;

[0132] 4) Combined with abnormal component analysis, the fault location can be located to provide a basis for maintenance;

[0133] 5) The entire process is automated and online, enabling real-time monitoring of the transformer.

[0134] In order to better describe the specific implementation of the present invention, the formula is used below to specifically explain the solution of the present invention:

[0135] The specific implementation method of step S10:

[0136] For electrical dataX e , temperature data X t , vibration dataX v and Fluid Data X f Each modal data X m First, wavelet transform is used to reduce noise and preprocess the data to obtain the processed data. Wavelet transform has good time-frequency localization characteristics and can effectively remove high-frequency noise in the data while retaining the mutation characteristics of the signal.

[0137] Next, Perform wavelet packet decomposition to obtain wavelet packet coefficients at different scales

[0138]

[0139] Where j represents the scale, represents the kth wavelet packet coefficient at the jth scale. Wavelet packet decomposition can capture the energy distribution characteristics of the signal at different scales and different time positions.

[0140] Then, define the wavelet packet energy for:

[0141]

[0142] Where N is the data length. It reflects the energy of the kth wavelet packet coefficient at the jth scale.

[0143] All By splicing them together, we get the modal data X m The characteristic energy distribution vector This vector captures the X m The energy distribution characteristics at various scales and time positions have good discrimination capabilities for anomaly detection.

[0144] Next, the above energy distribution is transformed into Reconstruction into abnormal component features Specifically, we need to solve the following optimization problem:

[0145]

[0146] Among them, Φ is a prior dictionary matrix that can sparsely represent the characteristics of the signal; λ is the regularization coefficient used to control The optimization problem can be solved efficiently using the alternating direction multiplier algorithm (ADMM).

[0147] The ADMM algorithm decomposes the original optimization problem into two sub-problems that are easier to solve, and optimizes alternately until convergence, thereby obtaining the optimal The algorithm steps are as follows:

[0148] initialization

[0149] renew

[0150]

[0151] Among them, ρ is the algorithm parameter

[0152] renew

[0153]

[0154] is the soft threshold function, defined as:

[0155]

[0156] renew

[0157]

[0158] Repeat 2)-4) until convergence

[0159] Through the above steps, the modal data X is obtained. m The abnormal component characteristics This feature is able to capture sudden changes and unusual patterns in the data.

[0160] The specific implementation method of step S20:

[0161] For each modal data, the abnormal component features extracted First, the kernel principal component analysis algorithm (KPCA) is used to map it to a low-dimensional space to achieve dimensionality reduction. KPCA uses kernel techniques to implicitly map data to a high-dimensional or even infinite-dimensional space, and performs linear principal component analysis in this space to achieve a nonlinear dimensionality reduction effect.

[0162] Specifically, given M samples First calculate their kernel matrix K:

[0163]

[0164] Among them, φ is a nonlinear mapping function that maps samples to high-dimensional or even infinite-dimensional feature space; k is a kernel function, and the commonly used one is the Gaussian kernel:

[0165]

[0166] Where σ is the kernel bandwidth parameter. Using the kernel technique, the inner product between the mapped samples can be directly calculated without explicitly knowing the mapping φ.

[0167] After obtaining the kernel matrix K, we centralize it and get

[0168]

[0169] Among them, 1 is a column vector of all ones.

[0170] Then solve The eigenvalues ​​and eigenvectors satisfy:

[0171]

[0172] Take the first p largest eigenvalues ​​λ1≥λ2≥...≥λ p and the corresponding eigenvector Then any sample The projection in the low-dimensional space after dimensionality reduction is:

[0173]

[0174] Among them, K(:,k) represents the kth column of the kernel matrix K. In this way, the abnormal component features Mapped to p-dimensional low-dimensional space, we get low-dimensional feature representation

[0175] Next, the similarity distance d between all sample points is calculated in this low-dimensional space. ij :

[0176]

[0177] All d ij Constitute the similarity matrix D.

[0178] The specific implementation method of step S30:

[0179] For the similarity matrix D, the spectral clustering method is used to cluster the sample points with high similarity to obtain the cluster set

[0180] The core idea of ​​spectral clustering is to regard the sample as a node of a weighted undirected graph, use the similarity distance as the weight of the edge, and then cluster the samples by cutting the graph. Specifically, it is necessary to minimize the cutting cost function:

[0181]

[0182] The above formula represents the total cost of cutting the graph into K subgraphs. This optimization problem is equivalent to solving the eigenvalue problem of the Laplace matrix L:

[0183]

[0184] Where L = DW is the Laplace matrix, D is the degree matrix, and W is the similarity matrix. Take the eigenvectors corresponding to the first K smallest non-zero eigenvalues ​​of L For each sample i, press Map it to K-dimensional space, and then use K-means algorithm to cluster it in this space.

[0185] Getting the cluster set Then, for each cluster C i Conduct principal component analysis (PCA) to determine the main direction of the bias error Specifically, calculate C i The covariance matrix of all samples in is:

[0186]

[0187] in, C i The mean vector of the samples in . Solve for Σ i The eigenvector corresponding to the maximum eigenvalue of That is C i main direction.

[0188] Then, all The columns are concatenated to form the first column of the propensity error matrix U. Subsequent columns can be determined in a similar iterative manner until most of the variance is explained or the preset rank threshold is reached. The final U is the propensity error matrix.

[0189] Specific implementation of step S40:

[0190] For each abnormal component feature The Gaussian mixture model (GMM) and expectation maximization (EM) algorithm are used to fit it as a mixture of several Gaussian distributions.

[0191] First, assume It can be represented by a mixture of K Gaussian distributions:

[0192]

[0193] Among them, π k is the mixing coefficient of the kth mixed component, satisfying Indicates the mean The covariance matrix is ​​Σ k The probability density function of the Gaussian distribution.

[0194] For any observation Define its hidden variables Denotes the probability that it comes from the kth mixture component:

[0195]

[0196] We need to solve the function that maximizes the log-likelihood Parameter value in:

[0197]

[0198] because Direct optimization is difficult, so the expectation maximization (EM) algorithm is used to iteratively solve it. The EM algorithm consists of the E step and the M step:

[0199] Step E: Calculate the value of each observation Probability under different mixture compositions:

[0200]

[0201] M step: using the γ calculated in the E step jk , re-estimate the parameter Θ:

[0202]

[0203] Repeat steps E and M until convergence or the maximum number of iterations is reached. The final parameter value is This is the optimal estimate of GMM.

[0204] Based on the estimated Each observation can be calculated Probability density under different Gaussian distributions:

[0205]

[0206] According to Bayes' theorem, the posterior probability is further calculated:

[0207]

[0208] For observations with smaller probability values, they are considered as outliers or flying points. Set a threshold ∈ if For all k, then Marked as flying spot.

[0209] Next, the flying point marked Perform feature encoding to obtain the flying point feature vector The encoding method can be simple indicator encoding or more complex kernel function-based encoding.

[0210] All By concatenating columns, we get the Bayesian flying point matrix V.

[0211] The specific implementation method of step S50:

[0212] Concatenate the previously obtained tendency error matrix U and Bayesian flying point matrix V by columns to construct the original feature matrix X0:

[0213] X0=[U,V]

[0214] Since there may be some redundant and irrelevant features in X0, it is necessary to optimize them to obtain a more concise and discriminative feature subset. Ant Colony Optimization (ACO) algorithm is used to optimize X0.

[0215] The ACO algorithm draws on the behavioral characteristics of ants releasing pheromones during foraging and selecting paths based on pheromone accumulation to effectively solve combinatorial optimization problems. Specifically, first, a fully connected feature construction graph G = (V, E) is constructed, where V is the set of feature nodes, that is, the column set of X0; E is the edge set between feature nodes.

[0216] Then, m artificial ants are placed to walk randomly on the graph G. Each ant selects the next node with a certain probability, and after a period of time, a feature subset path is formed. For the kth ant, the probability of moving from node i to node j is:

[0217]

[0218] Among them, J k(i) is the set of nodes that ant k can choose to move to when it is currently at node i; τ(i,j) represents the pheromone concentration on edge (i,j), which is used to accumulate empirical knowledge; η(j) is the expected value information of node j, which can be set as the importance score of the feature; α and β are the pheromone factor and expected value factor, respectively.

[0219] After a path is formed, its fitness f needs to be evaluated. Commonly used fitness functions include classification or regression accuracy, F-score, etc. The higher the fitness, the better the quality of the feature subset. According to the fitness information, a certain amount of pheromone is left on each path:

[0220]

[0221] Among them, Q is a constant. After N iterations, more pheromones will be accumulated on the high fitness paths, making it easier for subsequent ants to choose these high-quality paths.

[0222] Finally, when the algorithm converges, the feature subset corresponding to the path with the highest fitness is selected as the optimized feature matrix X.

[0223] Specific implementation of step S60:

[0224] From the optimized feature matrix X, the feature submatrix X corresponding to each modal data can be extracted respectively. e , X t , X v , X f .

[0225] Next, for the submatrix X m Each pair of abnormal component feature vectors in and Calculate the correlation between them. Commonly used correlation measures are:

[0226] Pearson correlation coefficient:

[0227]

[0228] in, and They are and The Pearson correlation coefficient measures the degree of linear correlation between two vectors.

[0229] Spearman correlation coefficient:

[0230]

[0231] Among them, rank(·) represents the sorting operation. The Spearman correlation coefficient measures the degree of correlation between the equal sequences of two vectors.

[0232] Mutual Information:

[0233]

[0234] in, and They are and The marginal probability distribution of is their joint probability distribution. Mutual information measures the mutual dependence between two random variables.

[0235] According to the above metrics, construct the abnormal component feature similarity matrix S of each modality data m ,in:

[0236] or or

[0237] Next, all S m Concatenate by columns to construct the abnormal component feature fusion matrix S:

[0238] S=[S e ,S t ,S v ,S f ]

[0239] Specific implementation of step S70:

[0240] The abnormal component feature fusion matrix S captures the correlation between the abnormal component features of each modality data. In order to further explore the high-order statistical characteristics between the abnormal component features, S is input into a deep belief network (DBN).

[0241] DBN is a probabilistic generative model consisting of multiple restricted Boltzmann machines (RBMs) that can learn the probability distribution of input data layer by layer. A standard DBN includes a visual layer (input layer) Several hidden layers And the parameter θ representing the joint distribution. The feedforward structure is:

[0242]

[0243] Among them, each It can be represented by an RBM model:

[0244]

[0245]

[0246] Among them, Z(θ) is the partition function; and is the bias vector; W (l) is the weight matrix. The parameter θ includes all biases and weights.

[0247] Through layer-by-layer unsupervised pre-training and layer-by-layer fine-tuning, DBN can learn high-order statistical features from input data. Specifically, the contrastive divergence algorithm is first used to perform unsupervised pre-training on each pair of adjacent layers of RBM, so that it can extract the probability distribution features of the input data. Then, DBN is regarded as a whole and supervised fine-tuned using labeled data, so that it can retain the learned probability distribution features while having the ability to discriminate abnormal data.

[0248] After DBN training is completed, the abnormal component feature fusion matrix S is input into it, and after layer-by-layer transformation of DBN, a high-order feature representation is obtained. This high-order feature representation not only retains the correlation information between the abnormal component features, but also captures their deeper statistical characteristics, thereby greatly enhancing the discriminative ability of anomaly detection.

[0249] The specific implementation method of step S80:

[0250] Represent the high-level features obtained previously Input into the supervised classifier to classify and distinguish abnormal data from normal data. Commonly used supervised classifiers include support vector machine (SVM), logistic regression, decision tree, etc. The specific classifier to be used can be flexibly selected according to the actual situation.

[0251] Taking SVM as an example, its model can be expressed as:

[0252]

[0253] in, Enter Mapping to high-dimensional feature space; and b are model parameters. For a given training data set where y i ∈{-1,1} is the label value, and the goal of SVM is to solve the following optimization problem:

[0254]

[0255] Where C is the regularization coefficient; ξ i It is a slack variable that allows some samples to cross the boundary to meet the separability condition.

[0256] By solving the dual problem, we can get the analytical solution of SVM:

[0257]

[0258] Among them, α i is the dual variable; N s is the number of support vectors; S is the set of support vectors.

[0259] For a given test sample Map it to a high-dimensional space and use the trained SVM model for classification prediction:

[0260]

[0261] During the classifier training process, cross-validation and other methods can be used to optimize model hyperparameters, such as C and kernel function parameters, to obtain the best classification performance.

[0262] In addition to SVM, you can also try other classifiers, such as logistic regression:

[0263]

[0264] Among them, σ(·) is the Sigmoid function. Through maximum likelihood estimation, the parameters of logistic regression can be solved and b.

[0265] Or use nonlinear classifiers such as decision trees, which can fit the complex distribution of sample data well.

[0266] Regardless of the classifier used, after obtaining the discrimination results of abnormal data and normal data, a normal or abnormal warning signal can be issued to the entire system based on this, thereby realizing real-time online monitoring. At the same time, since the original abnormal features of multimodal data are retained in the extraction, when an abnormality is detected, the cause of the abnormality can be easily traced and explained.

[0267] It should be noted that the variables and formulas involved in the above description are explained in the following table:

[0268]

[0269]

[0270]

[0271] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the above-mentioned method for predicting multi-point grounding defects of transformer cores and clamps.

[0272] The third aspect of the present invention provides a transformer core and clamp multi-point grounding defect prediction system, which includes the above-mentioned computer-readable storage medium. In the above-mentioned embodiment, the core and the clamp can be regarded as a whole or as two independent individuals, which does not affect the implementation of the steps of the present invention. That is, in the steps of the embodiment, the process of each step of the present invention is implemented on a separate core or a separate clamp.

[0273] Specifically, the principle of the present invention is:

[0274] 1) Extraction and fusion of abnormal component features

[0275] Traditional anomaly detection methods usually directly build models based on the original features of raw data, and cannot fully utilize the inherent correlation between different modal data. This paper proposes a novel anomaly component feature representation method, which aims to mine the cross-information between multi-source data and enhance the detection discrimination ability.

[0276] Specifically, we first perform a dimensionality reduction transformation on each modal data and map its original high-dimensional features into a low-dimensional space. In this low-dimensional space, the cluster properties are enhanced, and the distribution difference between abnormal samples and normal samples is more obvious. Then, based on this projection representation, we construct the abnormal component feature representation of each modal data to capture the essential characteristics of the abnormal data in each modality.

[0277] Furthermore, multiple similarity metrics such as Pearson correlation coefficient, Spearman correlation coefficient and mutual information are defined, and the abnormal component feature fusion matrix is ​​constructed based on these metrics. This matrix not only contains the similarity information of the abnormal component features within each modality, but also integrates the interdependence of the abnormal component features between different modalities. In this way, the complementarity of multi-source heterogeneous data can be fully utilized to improve the discriminative ability of anomaly detection.

[0278] 2) Encoding of abnormal component flying point characteristics

[0279] Since real data often contains noise, missing values, outliers and other complex situations, these "abnormal" data should not be simply filtered out in the anomaly detection task, but need to be encoded and utilized in a certain way. The present invention proposes an abnormal component flying point feature encoding method, which can effectively mine the valuable information contained in the abnormal data.

[0280] Specifically, the abnormal component feature fusion matrix is ​​projected into a low-dimensional space, and the cluster distribution is parameterized using the Gaussian mixture model (GMM). Samples with small posterior probabilities with all Gaussian distributions are considered to be truly "abnormal" data, i.e., flying points. These flying points are encoded as flying point feature vectors and input into the downstream feature optimization module together with the abnormal component features.

[0281] 4) Online feature optimization based on ant colony intelligence

[0282] Since real data usually have complex noise, missing values ​​and data skew problems, it is difficult to achieve ideal detection performance by relying only on the original abnormal component features and flying point features. Therefore, an online feature optimization strategy is designed to realize feature construction based on the ant colony optimization algorithm and automatically generate new high-order combined features.

[0283] Specifically, the abnormal component features and flying point features are listed as the initial feature set to construct a feature construction graph. The ant colony optimization algorithm performs heuristic search on the feature graph, generates new high-order features by combining feature nodes, and evaluates the feature fitness according to the classification or regression accuracy. The optimized feature matrix outputted in the end can better fit the complex distribution of the data and improve the accuracy and robustness of anomaly detection.

[0284] The following is a specific embodiment, which is intended to illustrate the specific implementation process and technical effects of the present invention.

[0285] This embodiment takes a 500kV main transformer in a 500kV large substation as the research object. The main transformer is a three-phase two-cycle distributed structure with a rated capacity of 1000MVA. The iron core adopts a five-legged beam structure, and grounding devices are provided at five iron core grounding points. After long-term operation, the main transformer has partial discharge and vibration abnormalities caused by multi-point grounding defects. In order to timely discover and diagnose this defect, the multi-mode data fusion LSTM model of the present invention is used for online monitoring.

[0286] The first step (S10) is to obtain the multi-mode data of the main transformer during operation. Specifically, it includes:

[0287] (1) Electrical data: Collect the three-phase line voltage, line current and neutral point current on the high and low voltage sides of the main transformer, with a sampling frequency of 5 kHz;

[0288] (2) Ultrasonic data: Ultrasonic sensors were installed at five core grounding points to collect partial discharge ultrasonic signals with a sampling frequency of 20 kHz;

[0289] (3) Ground current data: directly measure the ground current values ​​at five grounding points with a sampling frequency of 1 kHz;

[0290] (4) Vibration data: Vibration sensors were installed on the top of each foot of the core to collect the core vibration signal with a sampling frequency of 1 kHz.

[0291] In the second step (S20), each modal data is preprocessed to separate abnormal components.

[0292] Wavelet analysis was used to decompose the electrical data, with the wavelet basis as db8, the decomposition scale as 8, and the threshold as 3 times the variance of the wavelet coefficients.

[0293] The ultrasonic data were preprocessed by bandpass filtering, with a passband of 20 to 100 kHz;

[0294] Harmonic analysis is used to remove harmonics from ground current data and retain the residual component;

[0295] Short-time Fourier analysis is used to preprocess the vibration signal, and the spectral components that exceed 3 times the standard deviation of the normal range are retained.

[0296] Through the above method, the abnormal components corresponding to each modal data are obtained.

[0297] In the third step (S30), a tendency error analysis is performed on each abnormal component.

[0298] All abnormal components are concatenated by column to form a large matrix X. PCA is used to project X onto the first 10 principal components to obtain the reduced-dimensional matrix Xpca. The distance from each point in Xpca to all other points is calculated to construct the similarity matrix S.

[0299] Spectral clustering analysis was performed on S to obtain five main clusters. PCA was performed on each cluster to determine the direction of its principal component, which was used as a column of the tendency error matrix T. The final T was a 10*5 matrix.

[0300] In the fourth step (S40), Bayesian analysis is performed on each abnormal component.

[0301] For each abnormal component, a Gaussian mixture model is used to model its distribution, and the parameters of the three Gaussian distributions are estimated by the expectation maximization algorithm. The probability density value of each data point under these Gaussian distributions is calculated.

[0302] The points with probability values ​​lower than 0.01 are marked as flying points, and these flying points are encoded using one-hot encoding. This operation is performed for all abnormal components to obtain a 5*100 Bayesian flying point matrix B.

[0303] In the fifth step (S50), the tendency error matrix T and the Bayesian flying point matrix B are concatenated column by column to obtain the original 10*105 feature matrix F.

[0304] Ant colony algorithm is used to optimize F, and the optimized 7*30 feature matrix F' is obtained. The algorithm parameters are set as follows: ant colony size 30, maximum iteration number 1000, pheromone volatility coefficient 0.5, α=1, β=2.

[0305] In the sixth step (S60), the Pearson correlation coefficient of each pair of abnormal component feature vectors in F' is calculated to obtain a 30*30 correlation matrix R.

[0306] In the seventh step (S70), based on R, an abnormal component feature map G is constructed, and 9 fusion nodes are determined using the spectral clustering method. The corresponding feature vectors are concatenated to obtain an optimized 9*1 fusion feature vector f.

[0307] In the eighth step (S80), the pre-trained LSTM model is fine-tuned using the particle swarm algorithm.

[0308] The LSTM parameters to be optimized are expanded into a 12800-dimensional vector as the particle encoding in PSO. The population size is set to 50, the maximum number of iterations is 1000, w = 0.9, c1 = c2 = 2.0.

[0309] Taking classification accuracy as the fitness function, cross-validation optimization was performed on the training set (10,800 samples) and the validation set (2,700 samples).

[0310] At the final convergence, the accuracy on the validation set reached 95.6%, and the optimized LSTM model M was obtained.

[0311] In the ninth step (S90-S140), M is used to perform online monitoring on the main transformer.

[0312] The electrical, ultrasonic, ground current and vibration data streams are obtained from the online monitoring system of the main transformer. The data of the latest 120 time steps are extracted by sliding, the abnormal components are extracted, and the feature fusion is performed to obtain f.

[0313] Input f into M to calculate the defect index. If it is greater than 0.8, it is determined that there are multiple grounding defects. At the same time, the distribution of abnormal components is analyzed and the fault location is located based on empirical knowledge.

[0314] During the one-year online real-time monitoring of the main transformer, the method of the present invention detected the following fault conditions by statistically analyzing the defect index output by M:

[0315] (1) The model detected a high defect index (0.82-0.95) for several consecutive days, and the fault location was located at the grounding point of the E-phase iron core. On-site inspection found that there was indeed looseness causing poor grounding, and timely tightening solved the problem.

[0316] (2) The defect index remained at around 0.88. Analysis showed that the abnormality was caused by partial discharge. Combined with other external factors, it was judged that it might be partial discharge of phase A winding. The problem was finally solved by replacing the bushing.

[0317] (3) The defect index gradually rose to above 0.9. Analysis showed that it might be multi-point grounding, and the vibration abnormality was relatively serious. The detection found that the main transformer was displaced, resulting in a small multi-point grounding gap.

[0318] In contrast, the conventional fault diagnosis method based on empirical rules only detects the above-mentioned fault (2), with a missing rate of 50%, while the method of the present invention can achieve efficient and comprehensive fault detection.

[0319] At the same time, the experimenters counted the performance indicators of the method of the present invention, the traditional empirical rule method, and the deep learning method based on single modal analysis on 100 sets of test data, as follows:

[0320] method Accuracy Recall F1 Method of the present invention 97.3% 96.2% 96.7% Rule of thumb method 87.5% 81.4% 84.3% Single-mode LSTM (Electrical Data) 92.7% 90.3% 91.5% Single-modal LSTM (ultrasound data) 89.6% 87.8% 88.7%

[0321] It can be seen that the method of the present invention comprehensively utilizes multi-source heterogeneous data and learns richer fault mode representations. Therefore, it is significantly better than the empirical rule method and the single-modality-based deep learning method in terms of accuracy, recall rate, F1 score and other indicators.

[0322] It is particularly important to point out that the recall rate of the method of the present invention is as high as 96.2%, which means that most cases with multi-point grounding defects can be successfully detected, which ensures the practicality and reliability of the method. However, the empirical rule method is subject to the limitations of the rules, and many more hidden defects are often difficult to capture, with a recall rate of only 81.4%.

[0323] In addition, the comprehensive F1 score of the method of the present invention reaches 96.7%, while the single-modal LSTM is only 91.5% and 88.7% respectively. It can be seen that the fusion of multimodal data can indeed significantly improve the classification performance of the model. This is because different modal data contain complementary fault mode information, which can better describe the overall picture of the defect after fusion. Similarly, it also proves that the robustness of the solution of the present invention is high.

[0324] In terms of time and space complexity, the method of the present invention is slightly higher than the single-modal method, but still remains within a controllable range. The main complexity of the algorithm lies in the two links of feature optimization and LSTM fine-tuning, with time complexity of O(n^3) and O(n^2), respectively, where n is the number of samples. The space complexity mainly depends on the number of parameters of the LSTM model, which is proportional to the length of the input sequence.

[0325] By optimizing the algorithm's parameter adjustment and model pruning, the algorithm's computational overhead can be controlled without sacrificing too much performance. In addition, existing parallel computing and GPU acceleration technologies can also greatly improve the real-time and efficiency of the method of the present invention.

[0326] In summary, this embodiment fully demonstrates the effectiveness and superiority of the multi-mode data fusion LSTM model of the present invention for transformer multi-point grounding defect detection. This method can not only detect defects with high precision, but also locate the fault location, providing an important decision-making basis for equipment maintenance. At the same time, its efficient online detection capability also gives it a good prospect for engineering applications. It is believed that with the continuous improvement of computing power and data, this technology will play an increasingly important role in fields such as power system status monitoring and fault diagnosis.

[0327] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for predicting multi-point grounding defects of transformer core and clamps, characterized in that: The following steps are involved: S10, obtaining multi-mode data of the transformer during operation, including electrical data, partial discharge ultrasonic data of the grounding point of the core and the clamp, grounding current data of the core and the clamp, and vibration data of the core and the clamp; the electrical data includes data of three-phase line voltage, line current, and neutral point current; S20, preprocessing each multi-mode data, deleting regular components, and extracting abnormal components; S30, performing a tendency error analysis on the abnormal components of each multi-mode data, and identifying a tendency error matrix of multi-point grounding defects; S40, performing Bayesian analysis on the abnormal components of each multi-mode data to identify a Bayesian flying point matrix; S50, fusing the tendency error matrix and the Bayesian flying point matrix to obtain an original feature matrix; and optimizing the original feature matrix using an ant colony algorithm to obtain an optimized feature matrix; S60, calculating the correlation of abnormal components of each multi-mode data of the optimized feature matrix; S70, constructing a fusion feature matrix for the abnormal components of each type of multimodal data according to the correlation; S80, using a particle swarm algorithm based on the optimized feature matrix to train the pre-trained LSTM model to obtain a fine-tuned LSTM model, wherein the LSTM model is used to determine the multi-point grounding defect index of the core and the clamp of the transformer to be tested based on the fusion feature matrix of the transformer to be tested; S90, inputting the fused feature matrix into the fine-tuned LSTM model to obtain the multi-point grounding defect index of the core and the clamp of the transformer to be tested; The specific steps of step S30 are as follows: for the constructed similarity matrix, the sample points with high similarity are clustered by the spectral clustering method to obtain a cluster set; the cluster set is subjected to principal component analysis to determine the main direction of the tendency error, that is, the first column of the tendency error matrix; then, the other columns of the tendency error matrix except the first column are iteratively determined until most of the variance is explained or a preset rank threshold is reached, thereby obtaining the tendency error matrix; The specific implementation steps of step S40 are: for each abnormal component, fitting it as a mixture of several Gaussian distributions using the expectation maximization algorithm; calculating the probability density of each observation value under different Gaussian distributions, and updating the probability according to the Bayesian theorem; marking the observation value with a probability less than a preset threshold as a flying point, and feature encoding the flying point to obtain a Bayesian flying point matrix; The specific implementation steps of step S50 are: splicing the tendency error matrix and the Bayesian flying point matrix by column to construct an original feature matrix; then optimizing the original feature matrix using an ant colony optimization algorithm, wherein the ant colony optimization algorithm includes: constructing a fully connected feature construction graph, allowing multiple artificial ants to randomly walk on the graph to construct feature subset paths; Evaluate the fitness of each subset path and leave more concentrated pheromones on the paths with higher fitness; after multiple iterations, determine the optimal feature subset as the optimized feature matrix.

2. A transformer core and clamp multi-point grounding defect prediction method according to claim 1, characterized in that: The specific implementation steps of step S10 are: for each modal data among the electrical data, the partial discharge ultrasonic data of the grounding point of the core and the clamp, the grounding current data of the core and the clamp, and the vibration data of the core and the clamp, use wavelet transform to reduce noise and preprocess them, and use wavelet packet decomposition to extract their characteristic energy distribution; Then, the characteristic energy distribution is reconstructed into abnormal component characteristics by using a simultaneous inverse problem optimization method; Among them, the simultaneous inverse problem optimization method is solved using the alternating direction multiplier algorithm.

3. A transformer core and clamp multi-point grounding defect prediction method according to claim 1, characterized in that: The specific implementation steps of step S20 are: for each abnormal component feature extracted from the modal data, a kernel principal component analysis algorithm is used to reduce its dimension and map it to a low-dimensional space, and the kernel principal component analysis algorithm uses a Gaussian kernel function to achieve implicit mapping; then the similarity distance between all sample points is calculated in the low-dimensional space to construct a similarity matrix.

4. A transformer core and clamp multi-point grounding defect prediction method according to claim 1, characterized in that: The specific implementation steps of step S60 are: extracting the characteristic sub-matrices corresponding to each type of multimodal data from the optimized characteristic matrix; for each pair of abnormal component characteristic vectors in the sub-matrix, calculating the correlation coefficients between them, including the Pearson correlation coefficient, the Spearman correlation coefficient and the mutual information; and saving all the calculated correlation coefficient values ​​in the correlation matrix.

5. A transformer core and clamp multi-point grounding defect prediction method according to claim 1, characterized in that: The specific implementation steps of step S70 are: based on the correlation matrix, the abnormal component feature vector of each multimodal data is regarded as a node in the abnormal component feature graph, and the correlation is used as the weight of the edge to construct a graph; the spectral clustering method is used on the abnormal component feature graph to find K fusion nodes that can summarize the entire topological structure; the feature vectors corresponding to the K fusion nodes are spliced ​​by column to obtain a fusion feature matrix.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the method for predicting multi-point grounding defects of transformer cores and clamps according to any one of claims 1 to 5.

7. A transformer core and clamp multi-point grounding defect prediction system, characterized in that: Contains the computer-readable storage medium of claim 5.

Citation Information

Patent Citations

  • Transformer bushing defect identification method and system based on sequence mode decomposition

    CN113344475A

  • Shield tunneling load advanced prediction method and system based on Bayesian optimization LSTM

    CN115081749A

  • Holographic microwave breast lump detection tendency error analysis method, medium and system

    CN116883320A

  • Sea water temperature-salinity-depth instrument tendency error analysis method, medium and system

    CN117928625A

  • Data-driven high-precision integrated navigation data fusion method

    WO2021063136A1