Oil-immersed transformer abnormal state identification method based on electric-thermal-vibration multi-dimensional data fusion

Abstract

The invention discloses an electricity-heat-vibration multi-dimensional data fusion oil-immersed transformer abnormal state identification method, which comprises the following steps: analyzing multi-physical field coupling data on a transformer, extracting key characteristic parameters, and carrying out abnormal state identification by using an improved BES-LSSVM model. According to the method, a multi-source fusion data mining technology and a machine learning algorithm are combined, parameters are optimized by using a hybrid model of a BES optimization algorithm and a least square support vector machine, and classification and identification of abnormal working conditions are realized through a k-means clustering algorithm. According to the method, the accuracy and reliability of transformer abnormal state evaluation can be remarkably improved, and the fault occurrence rate and the maintenance cost are effectively reduced.

Description

Technical Field

[0001] The present invention relates to a transformer state identification method, in particular to an oil-immersed transformer abnormal state identification method based on electric-thermal-vibration multi-dimensional data fusion, and belongs to the technical field of power systems. Background Art

[0002] Transformers, as critical components of power systems, have experienced an increasing probability of failure over years of operation, directly impacting the safety and stability of the power grid. Research is needed to more efficiently and accurately identify the impact of multi-dimensional transformer information and complex operating conditions on abnormal transformer conditions. This could enable sensing of abnormal transformer conditions, assist with routine maintenance and inspection of transformer stations, reduce transformer failure rates, and ultimately improve the safety, reliability, and stability of power systems.

[0003] In the prior art, there is a distribution transformer abnormal state identification system that takes into account the missing detection data, such as the one disclosed in publication number CN117668729A. By adopting advanced data processing algorithms and artificial intelligence technology, it can accurately identify the abnormal state of the transformer even when a large amount of monitoring data is missing. This method not only improves the accuracy of transformer fault diagnosis, but also enhances the reliability and stability of the power system. At present, research on transformer state assessment mainly focuses on diagnostic technology based on artificial intelligence algorithms, state assessment technology based on expert systems, and assessment technology based on correlation. However, these methods often only focus on single-dimensional data, ignoring the impact of transformer multi-dimensional information and various complex working conditions on the state, such as electrical loss characteristics or thermal loss characteristics. It is difficult to fully reflect the health status of the transformer under complex working conditions, and lacks the reading and judgment of the overall state information of the transformer. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for identifying abnormal conditions of oil-immersed transformers by integrating electric, thermal and vibration multi-dimensional data in order to solve at least one of the above technical problems. This method comprehensively considers multiple physical field information such as electricity, heat and vibration, establishes a multi-source information database for different operating conditions of the transformer, introduces an algorithm to optimize the machine learning model parameters, and more comprehensively improves the accuracy of identifying abnormal conditions of the transformer.

[0005] The present invention achieves the above-mentioned object through the following technical solution: A method for identifying abnormal state of oil-immersed transformer by fusion of electric, thermal and vibration multi-dimensional data, comprising the following steps:

[0006] Step 1: Collect the real-time power signal, oil temperature signal and vibration signal of the transformer and establish a multivariate database;

[0007] Step 2: preprocess the data in the established multivariate database to characterize and normalize the sample data;

[0008] Step 3: Divide different working conditions based on the k-means clustering algorithm and generate working condition labels;

[0009] Step 4: Use the BES-LSSVM model to process key feature parameters, optimize the least squares support vector machine parameters using the vulture algorithm, and classify and identify abnormal working conditions using the k-means clustering algorithm;

[0010] Step 5: Use the historical data of the transformer to train the BES-LSSVM model, and use real-time data to test the model performance to optimize the model parameters; realize the identification of abnormal status of the transformer.

[0011] As a further solution of the present invention: in step one, the power signal is obtained from the nameplate parameters, factory information and background data of the operation and maintenance department; the oil temperature signal is measured by an oil temperature sensor, and the position of the oil temperature sensor measuring point is based on the heat dissipation characteristics of the transformer, taking into account the highest temperature points in the axial and radial directions; the vibration signal is measured by a vibration sensor, and the position of the vibration sensor measuring point is based on the vibration propagation path of the transformer. The measuring point is arranged in the radial direction in the area with the shortest radial vibration transmission path of the winding and the largest amplitude. In the axial direction, the measuring point is arranged at the starting point of the signal attenuation by analyzing the vibration signal attenuation path.

[0012] As a further solution of the present invention: in step 2, the collected raw data is normalized using the following formula:

[0013]

[0014] Where x ij is the jth eigenvalue of the i-th sample in the original data matrix; min(x j ) is the minimum value of the jth feature in all samples; max(x j ) is the maximum value of the jth feature in all samples; x ij ′ is the jth eigenvalue of the i-th sample after normalization, ranging from [0,1]; d is the total number of dimensions of the feature.

[0015] As a further solution of the present invention: in step 3, the working condition division of the k-means clustering algorithm specifically includes:

[0016] The feature samples are divided into k categories, and the objective function is:

[0017]

[0018] In the formula, J is the objective function value, which represents the sum of the squares of the distances from all samples to the cluster centers to which they belong; k is the number of clusters, which means that the samples are divided into k categories; x iis the feature vector of the i-th sample, with dimension d; S j is the jth cluster set, which contains all samples belonging to this class; μ j is the center vector of the jth cluster, with a dimension of d, and is updated iteratively as follows:

[0019]

[0020] Where, |S j | is the number of samples in the jth cluster; the cluster center μ is updated iteratively j , until convergence, generate the working condition label l j ∈{1,…,k}.

[0021] As a further solution of the present invention: in step 4, the extraction of key feature parameters includes:

[0022] A power transmission efficiency calculation unit, used to calculate the power transmission efficiency of the transformer and its changing speed;

[0023] Oil temperature change rate calculation unit, used to measure and calculate the oil temperature change rate and its acceleration relative to the ambient temperature rise;

[0024] The vibration fundamental frequency amplitude prediction unit is used to predict the vibration fundamental frequency amplitude of the transformer box surface and calculate the fundamental frequency amplitude error.

[0025] As a further solution of the present invention: in step 4, the process of establishing a transformer evaluation model based on BES-LSSVM is:

[0026] Construct the LSSVM classification model. The optimization problem of LSSVM is defined as:

[0027]

[0028] The constraints are:

[0029]

[0030] In the formula, w is the weight vector, b is the bias term, γ is the penalty coefficient, is the kernel function mapping, e i is the error term;

[0031] Solved by the Lagrange multiplier method, the final classification function is:

[0032]

[0033] Where K(x,x i ) is the kernel function, and the radial basis kernel function is used:

[0034]

[0035] The BES algorithm optimizes parameters, and the optimization goal is to minimize the classification error; the fitness function is defined as:

[0036] f(θ)=1-Accuracy,θ==C,σ]

[0037] Where θ is the parameter combination to be optimized; c is the regularization parameter of the LSSVM model, which controls the model complexity and overfitting degree; σ is the width parameter of the radial basis kernel function (RBF), which determines the mapping range of the feature space; Accuracy is the classification accuracy (percentage) of the model under the current parameter θ; f(θ) is the fitness function value, which represents the classification error (1-accuracy).

[0038] As a further solution of the present invention: the BES algorithm steps are as follows:

[0039] (1) Initialize the population:

[0040] Randomly generate N individual vulture positions, each of which represents a set of parameters (γ, σ);

[0041] (2) Fitness calculation:

[0042] Take classification accuracy as fitness function:

[0043]

[0044] (3) Selection stage:

[0045] The vulture selects the best area to search and updates its position:

[0046] x i,new =x best +α·r1·(x mean -x i )

[0047] Where x best is the current optimal solution, x mean is the average position of the population, r1 is a random number, and α is the search intensity coefficient;

[0048] (4) Search phase:

[0049] The vulture flies in a spiral to accurately catch its prey. The position update formula is:

[0050] x i,new =x i +r2·(x best -x i )+r E ·(x j -x k )

[0051] Where r2,r E is a random number, x j ,x k are randomly selected individuals;

[0052] (5) Subduction stage:

[0053] The vulture quickly dives to the optimal solution, and the update formula is:

[0054]

[0055] Where β is the subduction factor, r G is a random number;

[0056] (6) Iteration termination:

[0057] Repeat the above steps until the maximum number of iterations is reached and output the optimal parameters (γ*,σ*).

[0058] As a further solution of the present invention: in step 5, the model is trained using historical data, and the model performance is tested using real-time data to optimize the model parameters;

[0059] The LSSVM model is trained using the optimized parameters, and the classification decision function is:

[0060]

[0061] Where f(x) is the classification decision function, and the output value is +1 (normal) or -1 (abnormal); α is the Lagrange multiplier, which reflects the support degree of the i-th sample to the classification hyperplane; y i is the label of the i-th sample, y i ∈{-1,+1};K(x i ,x) is the kernel function, which calculates the similarity between sample x and the sample to be classified x, and uses the radial basis kernel function (RBF):

[0062]

[0063] Where v is the kernel function width parameter, which controls the mapping range of the feature space; σ is the bias term, which adjusts the position of the classification hyperplane; if f(x) = 1, it is judged as an abnormal condition; otherwise, it is normal.

[0064] The beneficial effects of the present invention are:

[0065] 1) The present invention arranges power, oil temperature, and vibration sensors to collect multi-physics field coupling data, extracts key characteristic parameters (such as power transmission efficiency, oil temperature change rate relative to ambient temperature, and vibration fundamental frequency amplitude), and uses an improved BES-LSSVM model to identify abnormal conditions;

[0066] 2) This method combines multi-source fusion data mining technology and machine learning algorithms, and uses the Bald Eagle Search (BES) optimization algorithm to optimize the kernel function parameters and regularization parameters of the LSSVM model, thereby improving the classification accuracy and generalization ability of the model. The k-means clustering algorithm is used to classify and identify abnormal operating conditions, thereby enhancing the reliability of the evaluation results.

[0067] 3) This method can significantly improve the accuracy and stability of transformer operating status assessment, effectively reduce the failure rate and maintenance costs. The BES-LSSVM intelligent assessment model established by the present invention can effectively solve the problem that traditional methods rely on single physical field data and are difficult to fully reflect the true operating status of the transformer, improve the accuracy and stability of transformer status assessment, and have certain guiding significance for real-time and effective detection of transformer status and ensuring the safety, stability and reliability of power system operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is a schematic diagram of the process of the present invention;

[0069] Figure 2 This is a schematic diagram of a specific process of using the K-means clustering algorithm to classify different operating states of a transformer;

[0070] Figure 3 This is a schematic diagram of the specific process of processing key feature parameters using the BES-LSSVM model in the present invention. DETAILED DESCRIPTION

[0071] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0072] Example 1, as Figures 1 to 3 As shown, this embodiment provides a method for identifying abnormal conditions of an oil-immersed transformer by fusion of electric, thermal, and vibration multi-dimensional data, comprising the following steps:

[0073] Step 1: Collect the real-time power signal, oil temperature signal and vibration signal of the transformer and establish a multivariate database;

[0074] Step 2: preprocess the data in the established multivariate database to characterize and normalize the sample data;

[0075] Step 3: Divide different working conditions based on the k-means clustering algorithm and generate working condition labels;

[0076] Step 4: Use the BES-LSSVM model to process the key characteristic parameters, optimize the least squares support vector machine (LSSVM) parameters through the condor algorithm (BES), and classify and identify abnormal working conditions through the k-means clustering algorithm;

[0077] Step 5: Use the historical data of the transformer to train the BES-LSSVM model, and use real-time data to test the model performance to optimize the model parameters; realize the identification of abnormal status of the transformer.

[0078] In summary, this paper combines the BES optimization algorithm with the LSSVM model to provide a novel method for identifying abnormal conditions in power transformers. This method comprehensively utilizes multi-physics data, including electrical, thermal, and vibration data, to accurately identify abnormal transformer conditions, thereby improving the safety and stability of power systems. Experimental verification demonstrates that this method demonstrates higher accuracy and reliability in transformer condition assessment and has broad application prospects.

[0079] In addition to all the technical features of the first embodiment, the second embodiment also includes:

[0080] In step 1, the power signal is obtained from the nameplate parameters, factory information, and backend data from the operation and maintenance department; the oil temperature signal is measured by an oil temperature sensor. The location of the oil temperature sensor measurement point is based on the heat dissipation characteristics of the transformer, taking into account the highest temperature points in the axial and radial directions; the vibration signal is measured by a vibration sensor. The location of the vibration sensor measurement point is based on the vibration propagation path of the transformer. The measurement point is arranged in the radial area with the shortest radial vibration transmission path and the largest amplitude of the winding. In the axial direction, the measurement point is arranged at the starting point of the signal attenuation by analyzing the vibration signal attenuation path.

[0081] The collected raw data are normalized using the following formula:

[0082]

[0083] Where x ij is the jth eigenvalue of the i-th sample in the original data matrix; min(x j ) is the minimum value of the jth feature in all samples; max(x j ) is the maximum value of the jth feature in all samples; x ij ′ is the jth eigenvalue of the i-th sample after normalization, ranging from [0,1]; d is the total number of dimensions of the feature.

[0084] In addition to all the technical features of the first embodiment, this embodiment also includes:

[0085] In step 3, the working condition division of the k-means clustering algorithm specifically includes:

[0086] The feature samples are divided into k categories, and the objective function is:

[0087]

[0088] Where J is the objective function value, which represents the sum of the squares of the distances from all samples to the cluster centers to which they belong; k is the number of clusters, which means that the samples are divided into k categories (such as normal, slightly abnormal, severely abnormal, etc.); x i is the feature vector of the i-th sample, with dimension d (such as electrical, thermal, and vibration features); S j is the jth cluster set, which contains all samples belonging to this class; μ j is the center vector of the jth cluster, with a dimension of d, and is updated iteratively as follows:

[0089]

[0090] Where, |S j | is the number of samples in the jth cluster; the cluster center μ is updated iteratively j , until convergence, generate the working condition label l j ∈{1,…,k}.

[0091] It should be noted that in the present invention, the number of clusters k is determined based on the actual category of the transformer operating state and the elbow rule. The elbow rule determines the optimal number of clusters by calculating the sum of squares (SSE) under different k values, where SSE is defined as follows:

[0092]

[0093] Among them, S j represents the sample set in the jth cluster, c j represents the center of the jth cluster, d(x i ,c j ) represents the sample x i and cluster center c j The distance between them.

[0094] Sample allocation, after initializing the cluster centers, each sample is assigned to the nearest cluster based on the distance between its features and each cluster center. The distance is calculated using the Euclidean distance, which is defined as follows:

[0095]

[0096] Among them, d represents the dimension of the feature, x i,O and c j,O Represents sample x respectivelyi and cluster center c j The eigenvalue in the mth dimension.

[0097] Update the cluster center. After all samples are assigned to the corresponding clusters, update the center of each cluster. The cluster center is updated by calculating the mean of all samples in each cluster, as shown below:

[0098]

[0099] Among them, |S j | represents the number of samples in the jth cluster.

[0100] Iterative optimization, sample allocation, and cluster center update are performed iteratively until the convergence condition is met. The convergence condition can be that the change in cluster center is less than a preset threshold, or the reduction in SSE in consecutive iterations is less than a certain value.

[0101] Determine the clustering results. When the algorithm converges, each sample is finally assigned to a cluster, forming k clusters. Each cluster represents a specific transformer operating condition.

[0102] After clustering is complete, abnormal operating conditions can be identified by analyzing the distribution of cluster characteristics. For example, if the number of samples in a cluster is significantly smaller than that in other clusters, or if the characteristics of the cluster are significantly different from normal operating conditions, then the cluster may represent an abnormal operating condition.

[0103] The resulting clustering results can be used for real-time monitoring and anomaly warnings within the transformer condition monitoring system. By assigning newly acquired data to corresponding clusters in real time, the system can quickly identify abnormal operating conditions and issue warnings, thereby improving the safety and reliability of transformer operation.

[0104] In addition to all the technical features of the first embodiment, this embodiment also includes:

[0105] In step 4, the extraction of key feature parameters includes:

[0106] A power transmission efficiency calculation unit, used to calculate the power transmission efficiency of the transformer and its changing speed;

[0107] Oil temperature change rate calculation unit, used to measure and calculate the oil temperature change rate and its acceleration relative to the ambient temperature rise;

[0108] The vibration fundamental frequency amplitude prediction unit is used to predict the vibration fundamental frequency amplitude of the transformer box surface and calculate the fundamental frequency amplitude error.

[0109] In step 4, the process of establishing a transformer evaluation model based on BES-LSSVM is as follows:

[0110] Construct the LSSVM classification model. The optimization problem of LSSVM is defined as:

[0111]

[0112] The constraints are:

[0113]

[0114] In the formula, w is the weight vector, b is the bias term, γ is the penalty coefficient, is the kernel function mapping, e i is the error term;

[0115] Solved by the Lagrange multiplier method, the final classification function is:

[0116]

[0117] Where K(x,x i ) is the kernel function, and the present invention adopts the radial basis kernel function:

[0118]

[0119] The BES algorithm optimizes parameters, and the optimization goal is to minimize the classification error; the fitness function is defined as:

[0120] f(θ)=1-Accuracy,θ==C,σ]

[0121] Where θ is the parameter combination to be optimized; c is the regularization parameter of the LSSVM model, which controls the model complexity and overfitting degree; σ is the width parameter of the radial basis kernel function (RBF), which determines the mapping range of the feature space; Accuracy is the classification accuracy (percentage) of the model under the current parameter θ; f(θ) is the fitness function value, which represents the classification error (1-accuracy).

[0122] The BES algorithm steps are as follows:

[0123] (1) Initialize the population:

[0124] Randomly generate N individual vulture positions, each of which represents a set of parameters (γ, σ);

[0125] (2) Fitness calculation:

[0126] Take classification accuracy as fitness function:

[0127]

[0128] (3) Selection stage:

[0129] The vulture selects the best area to search and updates its position:

[0130] x i,new =x best +α·r1·(x mean -x i )

[0131] Where x best is the current optimal solution, x mean is the average position of the population, r1 is a random number, and α is the search intensity coefficient;

[0132] (4) Search phase:

[0133] The vulture flies in a spiral to accurately catch its prey. The position update formula is:

[0134] x i,new =x i +r2·(x best -x i )+r E ·(x j -x k )

[0135] Where r2,r E is a random number, x j ,x k are randomly selected individuals;

[0136] (5) Subduction stage:

[0137] The vulture quickly dives to the optimal solution, and the update formula is:

[0138]

[0139] Where β is the subduction factor, r G is a random number;

[0140] (6) Iteration termination:

[0141] Repeat the above steps until the maximum number of iterations is reached and output the optimal parameters (γ*,σ*).

[0142] Example 5: In addition to all the technical features of Example 1, this example also includes:

[0143] In step 5, the model is trained using historical data and the model performance is tested using real-time data to optimize the model parameters;

[0144] Using the optimized parameter θ * Train the LSSVM model, and the classification decision function is:

[0145]

[0146] Where f(x) is the classification decision function, and the output value is +1 (normal) or -1 (abnormal); α is the Lagrange multiplier, which reflects the support degree of the i-th sample to the classification hyperplane; y i is the label of the i-th sample, y i ∈{-1,+1};K(x i ,x) is the kernel function, which calculates the similarity between sample x and the sample to be classified x, and uses the radial basis kernel function (RBF):

[0147]

[0148] Where σ is the kernel function width parameter, which controls the mapping range of the feature space; σ is the bias term, which adjusts the position of the classification hyperplane; if f(x) = 1, it is judged as an abnormal condition; otherwise, it is normal.

[0149] Example 6, a method for identifying abnormal conditions of an oil-immersed transformer by fusion of electrical, thermal, and vibration multi-dimensional data, comprising the following steps:

[0150] Step 1: Multi-physics data acquisition and preprocessing. Multi-source sensors are deployed on the power transformer to collect electrical, thermal, and vibration data, including power loss, oil temperature, and vibration signals. The collected data is normalized to ensure accuracy and consistency.

[0151] Step 2: Multi-physics data fusion and feature extraction. Based on the transformer's multi-physics coupling characteristics, an electric-thermal-vibration multi-source data fusion model is constructed. Characteristic parameters representing the transformer's operating status are extracted from the fused data, including power transmission efficiency, the rate of change of oil temperature relative to ambient temperature, and the amplitude of the transformer casing's surface vibration fundamental frequency. Based on the transformer's multi-physics coupling characteristics, an electric-thermal-vibration multi-source data fusion model is constructed. The fusion model uses a weighted fusion approach, assigning different weights to different physical field data based on their importance, ensuring that the fused data fully reflects the transformer's operating status.

[0152] Based on the multi-physics coupling characteristics of transformers, a multi-source data fusion model for electricity, heat, and vibration was constructed. By extracting key characteristic parameters such as power transmission efficiency, the rate of change of oil temperature relative to ambient temperature, and the amplitude of the fundamental vibration frequency, a comprehensive characterization of the transformer's operating status was achieved. Compared to traditional single-physics field data, the multi-physics fusion model more accurately reflects the transformer's true operating status, avoiding the limitations of a single data source.

[0153] Assume that the power transformer operation data set is Where x i ∈R P is the multi-physics field eigenvector y i∈{0,1} is the state label (0: normal, 1: abnormal). The feature vector x includes:

[0154] Power transfer efficiency: Where, P in is the input power, P out is the output power;

[0155] Oil temperature change rate: Where, T oil is the oil temperature, T env For ambient temperature:

[0156] Vibration fundamental frequency amplitude: Extract the amplitude of the fundamental frequency component A of the vibration signal through FFT fund , and calculate the error e=|A fund -A ref |, where A is the reference amplitude.

[0157] Normalize the original data:

[0158]

[0159] Step 3: BES-LSSVM model construction and optimization. The extracted feature parameters are classified using the LSSVM model to construct a preliminary transformer state identification model. The BES optimization algorithm is used to optimize the kernel function parameters and regularization parameters of the LSSVM model to improve the model's classification accuracy and generalization ability. The BES algorithm simulates the hunting behavior of vultures and performs global and local searches to ensure the optimality of the model parameters. The BES (Bald Eagle Search) optimization algorithm is used to optimize the kernel function parameters and regularization parameters of the LSSVM (Least Squares Support Vector Machine) model. The BES algorithm simulates the hunting behavior of vultures and performs global and local searches to ensure the optimality of the model parameters. Experimental results show that the BES-LSSVM model improves classification accuracy by approximately 12% and the F1 score by approximately 15% compared to the traditional LSSVM model, significantly improving the model's classification accuracy and generalization ability.

[0160] 3.1 LSSVM classification model

[0161] The optimization problem of LSSVM is defined as:

[0162]

[0163] The constraints are:

[0164]

[0165] In the formula, w is the weight vector, b is the bias term, γ is the penalty coefficient, is the kernel function mapping, e i is the error term.

[0166] Solved by the Lagrange multiplier method, the final classification function is:

[0167]

[0168] Where K(x,x i ) is the kernel function, and the present invention adopts the radial basis kernel function:

[0169]

[0170] 3.2 BES algorithm optimization parameters

[0171] The optimization goal is to minimize the classification error, and the fitness function is defined as:

[0172] f(θ)=1-Accuracy,θ==C,σ]

[0173] Where θ is the parameter combination to be optimized; c is the regularization parameter of the LSSVM model, which controls the model complexity and overfitting degree; σ is the width parameter of the radial basis kernel function (RBF), which determines the mapping range of the feature space; Accuracy is the classification accuracy (percentage) of the model under the current parameter θ; f(θ) is the fitness function value, which represents the classification error (1-accuracy).

[0174] The BES algorithm steps are as follows:

[0175] 1. Initialize the population:

[0176] Randomly generate N individual vulture positions, each of which represents a set of parameters (γ,σ).

[0177] 2. Fitness calculation:

[0178] Take classification accuracy as fitness function:

[0179]

[0180] 3. Selection stage:

[0181] The vulture selects the best area to search and updates its position:

[0182] x i,new =x best +α·r1·(x mean -x i )

[0183] Where x best is the current optimal solution, xmean is the average position of the population, r1 is a random number, and α is the search intensity coefficient.

[0184] 4. Search phase:

[0185] The vulture flies in a spiral to accurately catch its prey. The position update formula is:

[0186] x i,new =x i +r2·(x best -x i )+r E ·(x j -x k )

[0187] Where r2,r E is a random number, x j ,x k are randomly selected individuals.

[0188] 5. Dive phase:

[0189] The vulture quickly dives to the optimal solution, and the update formula is:

[0190]

[0191] Where β is the subduction factor, r G is a random number.

[0192] 6. Iteration termination:

[0193] Repeat the above steps until the maximum number of iterations is reached and output the optimal parameters (γ*,σ*).

[0194] Step 4: Abnormal state identification and classification. The k-means clustering algorithm is used to classify the characteristic samples into different operating conditions, and labels are assigned to the sample operating condition types. The optimized BES-LSSVM model is used to classify the transformer's operating conditions and identify abnormal operating conditions. The k-means clustering algorithm is used to classify the characteristic samples into different operating conditions, and labels are assigned to the sample operating condition types. This algorithm can effectively distinguish between normal and abnormal operating conditions, avoiding the misjudgment problem caused by uneven data distribution in traditional methods. Experimental verification shows that the k-means clustering algorithm can significantly improve the accuracy of identifying abnormal operating conditions and enhance the reliability of the evaluation results.

[0195] 4.1. k-means working condition division

[0196] The feature samples are divided into k categories, and the objective function is:

[0197]

[0198] Where J is the objective function value, which represents the sum of the squares of the distances from all samples to the cluster center. k is the number of clusters, which means that the samples are divided into k categories (such as normal, slightly abnormal, severely abnormal, etc.). i is the feature vector of the i-th sample, with dimension d (such as electrical, thermal, and vibration features). j is the jth cluster set, which contains all samples belonging to this class. j is the center vector of the jth cluster, with a dimension of d, and is updated iteratively to

[0199]

[0200] Where, |S j | is the number of samples in the jth cluster. The cluster center μ is updated iteratively j , until convergence, generate the working condition label l j ∈{1,…,k}.

[0201] It should be noted that the analysis of the working condition division results:

[0202] After normalizing the data, clustering was performed using the k-means algorithm. Based on the principle that smaller DBI values yield better clustering results, a k value of 4 was determined to be optimal for clustering. Therefore, this embodiment divides the transformer operating conditions into four categories, as shown in Table 1:

[0203] Table 1: Standardized sample values of cluster center points for each operating condition type

[0204]

[0205] Considering the data characteristics of the selected dataset, the four operating conditions selected above represent the primary characteristics of the field monitoring data and can effectively reflect the operating status of the transformer. Although other operating condition combinations exist, these combinations are relatively rare in the dataset and lack significant representativeness. Therefore, this example focuses on analyzing the practical significance of the four operating conditions obtained from the classification.

[0206] Operating Condition 1: I < 0.5, T < 0.5, U > 0.5. The transformer current and oil temperature are both low, but the voltage is high. Under this condition, the transformer load is light, but the high voltage may place significant stress on the transformer's insulation system. Despite the low oil temperature and good heat dissipation, the high voltage may increase the risk of partial discharge or insulation degradation. Therefore, while the overall operating status is relatively stable, close monitoring of the insulation system's health is still necessary to prevent potential failures.

[0207] Operating Condition 2: I > 0.5, T > 0.5, U < 0.5. The transformer's current and oil temperature are both high, but the voltage is low. Under this condition, the transformer is heavily loaded, potentially leading to increased electrical losses and elevated oil temperatures. While low-voltage operation reduces mechanical stress and vibration, high oil temperatures can affect insulation performance and increase the burden on the cooling system. Furthermore, high-current operation can increase thermal stress on the transformer's internal components, impacting the equipment's stability and service life. Therefore, this operating condition requires enhanced cooling system maintenance and close monitoring of oil temperature fluctuations.

[0208] Operating Condition 3: I < 0.5, T > 0.5, U < 0.5. The transformer current is low, but the oil temperature is high and the voltage is also low. Under this condition, the transformer load is light, but the high oil temperature may indicate poor heat dissipation or a cooling system failure. While low-voltage operation reduces mechanical stress, high oil temperature may adversely affect insulation performance. Therefore, it is important to pay close attention to the cooling system's operating status and take timely measures to reduce the oil temperature to ensure safe transformer operation.

[0209] Operating Condition 4: I > 0.5, T < 0.5, U > 0.5. The transformer current is high, the oil temperature is low, but the voltage is also high. Under this condition, the transformer is heavily loaded, but due to the low oil temperature, heat dissipation is excellent, helping to maintain stable equipment operation. However, high-voltage operation may increase stress on the insulation system, leading to an increased risk of partial discharge. Therefore, although the overall operating status is relatively stable, the health of the insulation system must be closely monitored to prevent potential failures. Furthermore, high-current operation may increase thermal stress on the transformer's internal components, requiring enhanced monitoring and maintenance.

[0210] Based on the above working condition division, the sample set is divided according to the ratio of 8:2. The specific division is shown in Table 2:

[0211] Table 2: Sample dataset distribution

[0212] Working condition type Number of sample sets Number of training sets Number of test sets Working condition 1 3190 Working condition 2 3447 Working condition three 2148 Working condition four 1217 total 10002

[0213] 4.2 BES-LSSVM classifier training

[0214] Using the optimized parameter θ * Train the LSSVM model, and the classification decision function is:

[0215]

[0216] Where, f(x): classification decision function, the output value is +1 (normal) or -1 (abnormal). α: Lagrange multiplier, reflecting the support degree of the i-th sample for the classification hyperplane. y i : The label of the i-th sample, yi ∈{-1,+1}. K(x i ,x) kernel function, calculates the similarity between sample x and sample x to be classified, and uses radial basis kernel function (RBF):

[0217]

[0218] Where σ is the kernel width parameter, which controls the mapping range of the feature space. b is the bias term, which adjusts the position of the classification hyperplane. If f(x) = 1, the condition is considered abnormal; otherwise, it is considered normal.

[0219] Step 5. Implement and apply the system. Integrate the above method into the power transformer status monitoring system to realize real-time monitoring of the transformer operating status and identification of abnormal conditions. The system can automatically identify the abnormal state of the transformer according to the changes in multi-physical field data, and provide early warning information to assist operation and maintenance personnel in troubleshooting and processing. Integrate the above method into the power transformer status monitoring system to realize real-time monitoring of the transformer operating status and identification of abnormal conditions. The system can automatically identify the abnormal state of the transformer according to the changes in multi-physical field data, and provide early warning information to assist operation and maintenance personnel in troubleshooting and processing. Experimental results show that the system can effectively reduce the failure rate of the transformer and significantly improve the safety and stability of the power system.

[0220] Cross-validation is used to evaluate performance, and the classification accuracy and F1 score are defined as:

[0221]

[0222] Where TP and TN are the number of correctly classified abnormal samples and normal samples, respectively. FP and FN are the number of misclassified abnormal samples and normal samples, respectively.

[0223] Experimental results show that BES-LSSVM improves the accuracy by about 12% and the F1 score by about 15% compared with traditional LSSVM.

[0224] Step 6: Model Validation and Optimization. The BES-LSSVM model was validated using real-world data to evaluate its classification accuracy and generalization capabilities. Based on the validation results, model parameters were further optimized to improve model accuracy and stability. The BES-LSSVM model was validated using real-world data to evaluate its classification accuracy and generalization capabilities. The validation process used a cross-validation method, dividing the dataset into a training set and a test set for model training and performance evaluation, respectively. Based on the validation results, model parameters were further optimized to improve model accuracy and stability. The optimization process involved adjusting the search parameters of the BES algorithm and the kernel function parameters of the LSSVM model.

[0225] The BES-LSSVM model is embedded in the transformer monitoring system, and the real-time data stream processing flow is as follows:

[0226] 1. Data acquisition module obtains |x ra |;

[0227] 2. Feature extraction module calculates η, ΔT, A fund ;

[0228] 3. Input BES-LSSVM classifier and output state label y pred ;

[0229] 4. If Trigger early warning signals and record abnormal operating condition types.

[0230] Step 7: Output and apply the results. Output the transformer operating status classification results, including the identification results of normal and abnormal states. Apply this method to transformer status monitoring in power systems to improve the safety and stability of power systems.

[0231] This paper combines the BES optimization algorithm with the LSSVM model to provide a new method for identifying abnormal conditions in power transformers. This method can comprehensively utilize multi-physics field data such as electrical, thermal, and vibration to accurately identify abnormal transformer conditions, thereby improving the safety and stability of the power system.

[0232] By analyzing multi-physics coupling data on the transformer, key characteristic parameters (such as power transmission efficiency, rate of change of oil temperature relative to ambient temperature, and vibration fundamental frequency amplitude) are extracted. An improved BES-LSSVM model is then used to identify abnormal conditions. This method combines multi-source fusion data mining technology with machine learning algorithms, optimizing parameters using a hybrid model of the Bald Eagle Search (BES) optimization algorithm and the least squares support vector machine (LSSVM). The k-means clustering algorithm is then used to classify and identify abnormal operating conditions.

[0233] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

[0234] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A method for identifying abnormal state of oil-immersed transformers by fusion of electric, thermal and vibration multi-dimensional data, characterized in that: The following steps are involved: Step 1: Collect the real-time power signal, oil temperature signal and vibration signal of the transformer and establish a multivariate database; Step 2: preprocessing the data in the multivariate database to characterize and normalize the sample data; Step 3: Divide different working conditions based on the k-means clustering algorithm and generate working condition labels; Step 4: Use the BES-LSSVM model to process key feature parameters, optimize the least squares support vector machine parameters using the vulture algorithm, and classify and identify abnormal working conditions using the k-means clustering algorithm; Step 5: Use the historical data of the transformer to train the BES-LSSVM model, and use real-time data to test the model performance to optimize the model parameters; realize the identification of abnormal status of the transformer.

2. The method for identifying abnormal state of oil-immersed transformers by fusion of electric, thermal and vibration multi-dimensional data according to claim 1 is characterized in that: In the step 1: The power signal is obtained from the nameplate parameters, factory information and backend data of the operation and maintenance department; The oil temperature signal is measured by the oil temperature sensor. The oil temperature sensor measurement point is based on the heat dissipation characteristics of the transformer, taking into account the highest temperature points in the axial and radial directions. The vibration signal is measured by a vibration sensor. The position of the vibration sensor measuring point is based on the vibration propagation path of the transformer. In the radial direction, the measuring point is arranged in the area with the shortest radial vibration transmission path and the largest amplitude of the winding. In the axial direction, the vibration signal attenuation path is analyzed and the measuring point is arranged at the starting point of the signal attenuation.

3. The method for identifying abnormal state of oil-immersed transformers by fusion of electric, thermal and vibration multi-dimensional data according to claim 1 is characterized by: In the second step, the collected raw data is normalized using the following formula: Where x ij is the jth eigenvalue of the i-th sample in the original data matrix; min(x j ) is the minimum value of the jth feature in all samples; max(x j ) is the maximum value of the jth feature in all samples; x ij ′ is the jth eigenvalue of the i-th sample after normalization, ranging from [0,1]; d is the total number of dimensions of the feature.

4. The method for identifying abnormal state of an oil-immersed transformer by fusion of electric, thermal and vibration multi-dimensional data according to claim 1 is characterized in that: In step 3, the working condition division of the k-means clustering algorithm specifically includes: The feature samples are divided into k categories, and the objective function is: In the formula, J is the objective function value, which represents the sum of the squares of the distances from all samples to the cluster centers to which they belong; k is the number of clusters, which means that the samples are divided into k categories; x i is the feature vector of the i-th sample, with dimension d; S j is the jth cluster set, which contains all samples belonging to this class; μ j is the center vector of the jth cluster, with a dimension of d, and is updated iteratively as follows: Where, |S j | is the number of samples in the jth cluster; the cluster center μ is updated iteratively j , until convergence, generate the working condition label l j ∈{1,…,k}.

5. The method for identifying abnormal state of oil-immersed transformers by fusion of electric, thermal and vibration multi-dimensional data according to claim 1, characterized in that: In step 4, the extraction of key feature parameters includes: A power transmission efficiency calculation unit, used to calculate the power transmission efficiency of the transformer and its changing speed; Oil temperature change rate calculation unit, used to measure and calculate the oil temperature change rate to ambient temperature rise and its acceleration; The vibration fundamental frequency amplitude prediction unit is used to predict the vibration fundamental frequency amplitude of the transformer box surface and calculate the fundamental frequency amplitude error.

6. The method for identifying abnormal state of oil-immersed transformers by fusion of electric, thermal and vibration multi-dimensional data according to claim 1, characterized in that: In step 4, the process of establishing a transformer evaluation model based on BES-LSSVM is as follows: Construct the LSSVM classification model. The optimization problem of LSSVM is defined as: The constraints are: In the formula, w is the weight vector, b is the bias term, γ is the penalty coefficient, is the kernel function mapping, e i is the error term; Solved by the Lagrange multiplier method, the final classification function is: Where K(x,x i ) is the kernel function, and the radial basis kernel function is used: The BES algorithm optimizes parameters, and the optimization goal is to minimize the classification error. The fitness function is defined as: f(θ)=1-Accuracy,θ==C,σ] Where θ is the parameter combination to be optimized; c is the regularization parameter of the LSSVM model, which controls the model complexity and overfitting degree; σ is the width parameter of the radial basis kernel function, which determines the mapping range of the feature space; Accuracy is the classification accuracy of the model under the current parameter θ; f(θ) is the fitness function value, which represents the classification error.

7. The method for identifying abnormal state of oil-immersed transformers by fusion of electric, thermal and vibration multi-dimensional data according to claim 6, characterized in that: The BES algorithm steps are as follows: (1) Initialize the population: Randomly generate N individual vulture positions, each of which represents a set of parameters (γ, σ); (2) Fitness calculation: Take classification accuracy as fitness function: (3) Selection stage: The vulture selects the best area to search and updates its position: x i,new =x best +α·r1·(x mean -x i ) Where x best is the current optimal solution, x mean is the average position of the population, r1 is a random number, and α is the search intensity coefficient; (4) Search phase: The vulture flies in a spiral to accurately catch its prey. The position update formula is: x i,new =x i +r2·(x best -x i )+r E ·(x j -x k ) Where r2,r E is a random number, x j ,x k are randomly selected individuals; (5) Subduction stage: The vulture quickly dives to the optimal solution, and the update formula is: Where β is the subduction factor, r G is a random number; (6) Iteration termination: Repeat the above steps until the maximum number of iterations is reached and output the optimal parameters (γ*,σ*).

8. The method for identifying abnormal state of oil-immersed transformers by fusion of electric, thermal and vibration multi-dimensional data according to claim 7, characterized in that: In step 5, the model is trained using historical data, and the model performance is tested using real-time data to optimize the model parameters; The LSSVM model is trained using the optimized parameters, and the classification decision function is: Where f(x) is the classification decision function, and the output value is +1 or -1; α is the Lagrange multiplier, which reflects the support degree of the i-th sample to the classification hyperplane; y i : The label of the first sample, y i ∈{-1,+1}; K(x i ,x) is the kernel function, which calculates the similarity between sample x and the sample to be classified x, and uses the radial basis kernel function (RBF): Where σ is the kernel function width parameter, which controls the mapping range of the feature space; σ is the bias term, which adjusts the position of the classification hyperplane; if f(x) = 1, it is judged as an abnormal condition; otherwise, it is normal.

Images