A power transformer fault diagnosis method and device based on improved multi-dimensional expansion space

Abstract

The application discloses a power transformer fault diagnosis method and device based on an improved multi-dimensional expansion space, belongs to the technical field of electric power, and solves the problems that existing diagnosis methods are prone to noise interference, feature mining is insufficient, and clustering is prone to local optimization. The method comprises the following steps: acquiring dissolved gas data in transformer oil, oversampling a minority class sample by using an adaptive synthesis method and undersampling a majority class sample by using a DBSCAN algorithm to balance the data, establishing a degradation model containing common trends and random deviation components based on an improved high-rank tensor algorithm, calculating one-dimensional and two-dimensional features by using a multi-scale approximate entropy algorithm, fusing the features into multi-domain multi-scale features by using an adaptive channel attention model, inputting the features into a multi-dimensional expansion space learning network, optimizing network parameters by using an improved adaptive meerkat algorithm and introducing a differential evolution algorithm to assist global optimization, and outputting a fault type diagnosis result, which is used for power transformer fault diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of power technology, and in particular relates to a method and device for diagnosing power transformer faults based on an improved multidimensional extended space. Background Technology

[0002] In the electrical monitoring of power transformers, traditional methods suffer from difficulties in data acquisition and susceptibility to environmental noise interference. Furthermore, frequent fluctuations in equipment operating conditions lead to unstable monitoring signals, making it difficult to effectively acquire electrical quantity data that accurately reflects the transformer's internal state. More critically, the electrochemical state information within the transformer contained in the electrical monitoring signals exhibits nonlinear and multi-scale evolutionary characteristics. Existing monitoring methods struggle to explore the complex evolution of the transformer's internal electrochemical state under various fault conditions from limited electrical signals. This significantly limits the application of fault diagnosis methods based on traditional electrical monitoring.

[0003] In the field of power transformer fault diagnosis based on dissolved gases in oil, most existing diagnostic methods are designed only for main grid transformers. These methods have specific application scenarios and poor applicability to distribution network transformers or other different types of transformers. Furthermore, existing methods are mostly used for source tracing analysis after a transformer fault occurs, aiming to identify the cause of the fault rather than for real-time or near-real-time fault status identification and early warning during transformer operation. In addition, existing methods lack the ability to deeply mine complementary information from multi-dimensional, multi-scale electrochemical features in terms of data processing and feature extraction. This results in a large amount of valuable electrochemical monitoring data not being fully utilized, and the correlation between feature information being ignored, thus limiting the diagnostic methods' ability to identify different transformer fault modes.

[0004] In cluster analysis of power transformer fault states, existing technologies mostly employ direct feature concatenation to simply combine multiple features, or use linear transformations to reduce dimensionality or map features. However, electrochemical feature data under transformer fault states typically exhibit high dimensionality, nonlinearity, and strong correlation. Direct concatenation or linear transformations cannot effectively handle the complex relationships between features, making clustering operations prone to convergence to local optima, hindering the attainment of globally optimal clustering results, and directly impacting the stability and accuracy of diagnostic results. These issues mean that existing power transformer fault diagnosis methods still have significant room for improvement in terms of anti-interference capability, feature utilization efficiency, and diagnostic reliability. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a power transformer fault diagnosis method and apparatus based on an improved multidimensional extended space, thereby resolving the issues present in the prior art.

[0006] Firstly, to achieve the above objectives, the present invention provides a power transformer fault diagnosis method based on an improved multidimensional extended space, comprising the following steps: Acquire transformer operating status data, including dissolved gas content data in oil; The unbalanced sample set processing method is used to equalize the operating state data to obtain balanced state data. Based on the balanced state data, a degradation model of the power transformer is established. The degradation model is used to describe the evolution law of the internal electrochemical state of the transformer over time. Based on the state data output by the degradation model, the one-dimensional and two-dimensional features of the transformer operating state are calculated using a multi-scale approximate entropy algorithm. The one-dimensional and two-dimensional features are then fused using an adaptive channel attention model to obtain multi-domain multi-scale fused features. The multi-domain, multi-scale fused features are input into a multidimensional extended space learning network for cluster analysis. During the cluster analysis, an improved adaptive honey badger algorithm is used to optimize the parameters of the multidimensional extended space learning network, and a differential evolution algorithm is used to assist the improved adaptive honey badger algorithm in global optimization, outputting transformer fault type diagnosis results.

[0007] Optionally, the process of equalizing the operating status data using an imbalanced sample set processing method includes: Identify the minority class and majority class samples in the operational status data; For the minority class samples, an adaptive synthesis method is adopted to adaptively determine the amount of data to be synthesized for each minority class sample based on the proportion of majority class samples in the neighborhood of each minority class sample, and generate new minority class samples. For the majority class samples, the DBSCAN algorithm is used to identify the core points, boundary points, and noise points within the majority class samples, and the boundary points and noise points are deleted.

[0008] Optionally, the process of establishing a degradation model for the power transformer based on the equilibrium state data includes: The electrochemical index data under various fault conditions in the equilibrium state data are taken as deterministic components, and the independent and identically distributed measurement noise is taken as random noise components. The deterministic components and the random noise components are superimposed to form the deterioration state of the transformer at the monitoring point. The deterioration state of a transformer is decomposed into a common deterioration trend component and a random deviation component. The common deterioration trend component is used to describe the common aging pattern of the transformer as a whole, and the random deviation component is used to describe the deviation between the deterioration path of a specific transformer and the common deterioration trend.

[0009] Optionally, the process of calculating the one-dimensional and two-dimensional features of the transformer's operating state using the multi-scale approximate entropy algorithm includes: For the state data sequence output by the degradation model, coarse-grained time series of different scales are formed by calculating the average value of multiple consecutive time points. The approximate entropy of each coarse-grained time series is calculated in an increasing dimension manner to obtain a one-dimensional multi-scale approximate entropy. The continuous data points in the state data sequence are encoded into a two-dimensional matrix. The data in the two-dimensional matrix are normalized and scaled. The scaled data is mapped to angles and radii in polar coordinates. The correlation within different time intervals is measured by trigonometric function transformation of the angles and differences between points, and a two-dimensional multi-scale approximate entropy is obtained.

[0010] Optionally, the process of fusing the one-dimensional features and the two-dimensional features using an adaptive channel attention model includes: The one-dimensional multi-scale approximate entropy and the two-dimensional multi-scale approximate entropy are input into the cross-modal channel attention model, and the cross-modal channel attention model calculates channel attention weights for the one-dimensional features and the two-dimensional features respectively through a multilayer perceptron; The fusion factor is determined based on the channel attention weights, and the one-dimensional features and the two-dimensional features are weighted and fused using the fusion factor to output the multi-domain multi-scale fused features.

[0011] Optionally, the process of inputting the multi-domain, multi-scale fused features into a multi-dimensional extended spatial learning network for cluster analysis includes: In the input layer of the multidimensional extended space learning network, radial basis operations are used to transform the input features of each dimension based on the variance and mean sequences of the data in each dimension. In the summation layer of the multidimensional extended spatial learning network, the conditional probability density of each fault category is estimated using the Parzen window function; In the output layer of the multidimensional extended spatial learning network, each neuron represents a fault category, and the most likely fault type is determined based on the conditional probability density.

[0012] Optionally, the process of optimizing the parameters of the multidimensional extended space learning network using the improved adaptive honey badger algorithm includes: Initialize the honey badger population, where each individual honey badger represents a set of parameters of the multidimensional extended space learning network; During the exploration phase, the individual's location is updated based on the odor intensity and distance between the current individual and random individuals; During the development phase, the location update direction is jointly determined based on the odor intensity and distance between the current individual and the best individual in the population, as well as the odor intensity and distance between the current individual and the average of the three best individuals. The search step size parameter is dynamically updated based on the current distance of the individual to the prey and the distance of the entire population to the prey in order to control the search range of the algorithm. The mutation and crossover operations of the differential evolution algorithm are applied to the honey badger population to generate new individuals.

[0013] Secondly, the present invention also provides a power transformer fault diagnosis device based on an improved multidimensional extended space, used to implement a power transformer fault diagnosis method based on an improved multidimensional extended space, the device comprising: The data acquisition module is used to acquire the transformer's operating status data, which includes data on the content of dissolved gases in the oil. The data balancing and degradation modeling module is used to balance the operating state data using an unbalanced sample set processing method to obtain balanced state data, and to establish a degradation model of the power transformer based on the balanced state data. The degradation model is used to describe the evolution law of the internal electrochemical state of the transformer over time. The multi-domain, multi-scale feature fusion module is used to calculate the one-dimensional and two-dimensional features of the transformer operating state based on the state data output by the degradation model using a multi-scale approximate entropy algorithm, and then fuse the one-dimensional and two-dimensional features using an adaptive channel attention model to obtain multi-domain, multi-scale fused features. The fault diagnosis output module is used to input the multi-domain, multi-scale fused features into a multi-dimensional extended space learning network for cluster analysis. During the cluster analysis, an improved adaptive honey badger algorithm is used to optimize the parameters of the multi-dimensional extended space learning network, and a differential evolution algorithm is used to assist the improved adaptive honey badger algorithm in global optimization, outputting the transformer fault type diagnosis result.

[0014] Thirdly, the present invention also provides a computer terminal device, comprising: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the power transformer fault diagnosis method based on the improved multidimensional extended space in the first aspect described above.

[0015] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the steps of the power transformer fault diagnosis method based on the improved multidimensional extended space described in the first aspect above.

[0016] Compared with the prior art, the present invention has the following advantages and technical effects: This invention provides a power transformer fault diagnosis method and device based on an improved multidimensional extended space. An improved high-rank tensor algorithm is used to establish a power transformer state degradation model, which can effectively uncover the nonlinear and multi-scale evolution of the transformer's internal electrochemical state under various fault conditions. An improved adaptive channel attention model effectively extracts the nonlinear and non-stationary characteristics of the transformer's internal state, significantly improving sensitivity to early, minor faults. Cluster analysis of multi-domain, multi-scale features is performed using a multidimensional extended space learning method, an improved adaptive honey badger algorithm, and a differential evolution algorithm. This fully utilizes the complementary information between different features, avoids the limitations of single features, and effectively avoids the problem of clustering operations converging to local optima. This invention improves the model's resistance to noise and operating condition fluctuations, significantly enhancing the stability and accuracy of the diagnostic results, and can be effectively applied to the precise diagnosis of different types of power transformer faults. Attached Figure Description

[0017] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart illustrating a power transformer fault diagnosis method based on an improved multidimensional extended space, according to an embodiment of the present invention. Detailed Implementation

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0020] Example 1 like Figure 1 As shown, this embodiment provides a power transformer fault diagnosis method based on an improved multidimensional extended space, including: Acquire transformer operating status data, including dissolved gas content data in oil; The unbalanced sample set processing method is used to equalize the operating state data to obtain balanced state data. Based on the balanced state data, a degradation model of the power transformer is established. The degradation model is used to describe the evolution law of the internal electrochemical state of the transformer over time. Based on the state data output by the degradation model, the one-dimensional and two-dimensional features of the transformer operating state are calculated using a multi-scale approximate entropy algorithm. The one-dimensional and two-dimensional features are then fused using an adaptive channel attention model to obtain multi-domain multi-scale fused features. The multi-domain, multi-scale fused features are input into a multidimensional extended space learning network for cluster analysis. During the cluster analysis, an improved adaptive honey badger algorithm is used to optimize the parameters of the multidimensional extended space learning network, and a differential evolution algorithm is used to assist the improved adaptive honey badger algorithm in global optimization, outputting transformer fault type diagnosis results.

[0021] Furthermore, the process of equalizing the operating status data using an imbalanced sample set processing method includes: Identify the minority class and majority class samples in the operational status data; For the minority class samples, an adaptive synthesis method is adopted to adaptively determine the amount of data to be synthesized for each minority class sample based on the proportion of majority class samples in the neighborhood of each minority class sample, and generate new minority class samples. For the majority class samples, the DBSCAN algorithm is used to identify the core points, boundary points, and noise points within the majority class samples, and the boundary points and noise points are deleted.

[0022] Specifically, the implementation process of this embodiment includes: Step 1: Preprocessing of internal state data of power transformers and establishment of deterioration model Step 1.1 Preprocessing of internal state data of power transformer The internal status data of the transformer includes five characteristics, corresponding to eight transformer fault types, including normal type (NT), high-energy discharge (HD), low-energy discharge (LD), high-temperature overheating (HO), medium-temperature overheating (ITO), medium-low-temperature overheating (ILO), low-temperature overheating (LO), and partial discharge (PD). The specific data are the content of dissolved hydrogen (H2), methane (CH4), ethane (C2H6), ethylene (C2H4), and acetylene (C2H2) gases in the transformer oil under each fault type, as well as the content ratio of different gas types.

[0023] Methods for handling imbalanced sample sets: First, an adaptive synthesis method is adopted to generate new samples based on the learning difficulty of minority class samples. The proportion of majority class in the neighborhood is calculated using k-nearest neighbors to determine the amount of data to be synthesized for each minority class.

[0024] (1) Calculate the total quantity that needs to be synthesized: ; in: IThis represents the total number of samples that need to be synthesized. m 1 represents the number of samples in the majority class. m s These represent the number of samples in the minority class; β ∈ [0, 1], where ∈ is the parameter of the equilibrium sample state. β When the value is 1, the sample is perfectly balanced.

[0025] (2) Let the minority class sample set be x i Identification using Euclidean distance x i Each sample x in of K Find the nearest neighbor samples and calculate the majority class samples in K The proportion of the nearest neighbor samples is calculated as follows: ; in, Q Δi represents the nearest neighbor sample in Euclidean distance; Δi is the number of samples in the majority class. r i For majority class samples K The proportion of the nearest neighbor samples.

[0026] (3) From x in of K Select one minority class sample from the nearest neighbor samples. x is The synthesized data is as follows: ; in, s in For newly synthesized samples; l A random number between 0 and 1.

[0027] (4) Repeat the above steps until the new sample synthesis for each minority class sample is satisfied.

[0028] Then, for the majority class samples, the data density within each majority class is identified using the DBSCAN algorithm. Based on the sample points... p Neighborhood radius and minimum number of points are used to determine sample points p The types of sample points are as follows: core points, sample points, and noise points. For any given point... p Its neighborhood is p With center radius ò The circle, that is: ; in, N ò ( p) is the sample point p The neighborhood; this ( p , q )for p , q The Euclidean distance between two points.

[0029] By judging sample points N ò ( p The relationship between the number of midpoints and the minimum point MinPts determines the type of sample points. The relationship is as follows: ; If point p It is not a core point in itself, but it is located at a certain core point. ò In the neighborhood, then p It is a boundary point, if the point p It is not a core point, nor does it belong to any core point. ò In the neighborhood, then p It is a noise point.

[0030] Based on sample points p Undersampling is performed on the type of noise points and boundary points in each cluster. Noise points are deleted in the sparse boundary point region.

[0031] Furthermore, the process of establishing a degradation model for the power transformer based on the equilibrium state data includes: The electrochemical index data under various fault conditions in the equilibrium state data are taken as deterministic components, and the independent and identically distributed measurement noise is taken as random noise components. The deterministic components and the random noise components are superimposed to form the deterioration state of the transformer at the monitoring point. The deterioration state of a transformer is decomposed into a common deterioration trend component and a random deviation component. The common deterioration trend component is used to describe the common aging pattern of the transformer as a whole, and the random deviation component is used to describe the deviation between the deterioration path of a specific transformer and the common deterioration trend.

[0032] Specifically, the implementation process of this embodiment includes: Step 1.2 Establishment of a degradation model for power transformers based on an improved high-rank tensor algorithm By establishing a degradation model for power transformers, this study explores the nonlinear and multi-scale evolution of the internal electrochemical state of transformers under various fault conditions. The power transformer is monitored at various points. A The degradation state is as follows: ; ; In the formula: This refers to the internal electrochemical index data of power transformers under various fault conditions after processing with an unbalanced sample set. It represents independent and identically distributed measurement noise, conforming to a normal distribution with a mean of 0 and a variance of σ².

[0033] ; In the formula: Describe the potential common degradation trends of power transformers; It has a mean of 0 and a covariance of cov( A , p The random deviation component of the power transformer reflects the degradation path and common degradation trend.

[0034] Furthermore, the process of calculating the one-dimensional and two-dimensional features of the transformer's operating state using the multi-scale approximate entropy algorithm includes: For the state data sequence output by the degradation model, coarse-grained time series of different scales are formed by calculating the average value of multiple consecutive time points. The approximate entropy of each coarse-grained time series is calculated in an increasing dimension manner to obtain a one-dimensional multi-scale approximate entropy. The continuous data points in the state data sequence are encoded into a two-dimensional matrix. The data in the two-dimensional matrix are normalized and scaled. The scaled data is mapped to angles and radii in polar coordinates. The correlation within different time intervals is measured by trigonometric function transformation of the angles and differences between points, and a two-dimensional multi-scale approximate entropy is obtained.

[0035] Furthermore, the process of fusing the one-dimensional features and the two-dimensional features using an adaptive channel attention model includes: The one-dimensional multi-scale approximate entropy and the two-dimensional multi-scale approximate entropy are input into the cross-modal channel attention model, and the cross-modal channel attention model calculates channel attention weights for the one-dimensional features and the two-dimensional features respectively through a multilayer perceptron; The fusion factor is determined based on the channel attention weights, and the one-dimensional features and the two-dimensional features are weighted and fused using the fusion factor to output the multi-domain multi-scale fused features.

[0036] Specifically, the implementation process of this embodiment includes: Step 2: Based on the improved adaptive channel attention model, calculate the multi-domain, multi-scale approximate entropy of the transformer's internal state under different fault modes. An improved multi-scale approximate entropy algorithm for multi-source feature fusion is introduced to extract feature parameters from gas ratio data.

[0037] (1) Suppose that the result obtained after step 1 includes The time series of data points is u(1): ; (2) Generate a set of dimensions as m : X (1), X (2), X (3), ..., X ( N - m A vector of +1, where represents the window length.

[0038] ; (3) Increase the dimension by 1 to m +1, the approximate entropy of this sequence is defined as: ; in, ApEn ( m , r ) represents the approximate entropy of the sequence; f m ( r ) is the dimension m The average value of all data in the data chain; f m+1 ( r ) is the dimension m +1 is the average of all data in the data chain.

[0039] Length is N The sequence was obtained by following the steps described above. ApEn The estimated value is denoted as: ; When the scale is n At that time, the coarse-grained time series was calculated n The average value at consecutive time points is used to obtain the one-dimensional feature multi-scale approximate entropy: ; in, t It is a time series.

[0040] (4) Translate a section m The transformer data from each data point is encoded into one. m * m A two-dimensional matrix, containing the data X Normalized to [0,1]: ; in, This is the scaled-down transformer data sequence; For the first i In dimensional data chain N The maximum value of each data point;X for N A one-dimensional data chain of data points.

[0041] The scaled transformer data sequence Mapped to angle a Map the time radius R In polar coordinates, it is represented as: ; in, This is the scaled-down transformer data sequence; The angle of the mapping; R The mapping radius; Using time nodes, divide the interval [0,1] into M Equal portions.

[0042] The correlation between different time intervals is measured and encoded by trigonometric function transformations of the angles and differences between each point, and then approximate entropy of the transformer is obtained by calculation in Cartesian coordinate system: ; ; An adaptive channel attention model is used to transform the one-dimensional multi-scale approximate entropy into a two-dimensional multi-scale approximate entropy.

[0043] ; ; in, α For fusion factor; s (MLPatt()) is a cross-modal channel attention model; t Indicates a time scale; The one-dimensional feature of the transformer is approximated by a multi-scale entropy. The multi-scale approximate entropy of the two-dimensional features of the transformer; To improve the multi-scale approximate entropy of multi-source feature fusion.

[0044] Furthermore, the process of inputting the multi-domain, multi-scale fused features into a multi-dimensional extended space learning network for cluster analysis includes: In the input layer of the multidimensional extended space learning network, radial basis operations are used to transform the input features of each dimension based on the variance and mean sequences of the data in each dimension. In the summation layer of the multidimensional extended spatial learning network, the conditional probability density of each fault category is estimated using the Parzen window function; In the output layer of the multidimensional extended spatial learning network, each neuron represents a fault category, and the most likely fault type is determined based on the conditional probability density.

[0045] Specifically, the implementation process of this embodiment includes: Step 3: Based on the improved multidimensional extended space learning method, cluster analysis is performed on the features from Step 2 to obtain the diagnostic results.

[0046] Step 3.1 Improve the multidimensional extended space learning method In the input layer, the radial basis function is: ; in, This is a variance series for each dimension; It is a series of mean values ​​for each dimension.

[0047] In the summation layer, each neural network unit is connected only to the corresponding type of model neuron and is estimated according to the Parzen window function method and various types of conditional probability densities. m The probability density function estimate for the dimensional Parzen window is: ; in, m For a certain dimension, b Point values ​​in each data column.

[0048] The final layer is the output layer, where each neuron represents a category. The most likely transformer fault category is selected as the output result through an improved adaptive honey badger algorithm.

[0049] Furthermore, the process of optimizing the parameters of the multidimensional extended space learning network using the improved adaptive honey badger algorithm includes: Initialize the honey badger population, where each individual honey badger represents a set of parameters of the multidimensional extended space learning network; During the exploration phase, the individual's location is updated based on the odor intensity and distance between the current individual and random individuals; During the development phase, the location update direction is jointly determined based on the odor intensity and distance between the current individual and the best individual in the population, as well as the odor intensity and distance between the current individual and the average of the three best individuals. The search step size parameter is dynamically updated based on the current distance of the individual to the prey and the distance of the entire population to the prey in order to control the search range of the algorithm. The mutation and crossover operations of the differential evolution algorithm are applied to the honey badger population to generate new individuals.

[0050] Specifically, the implementation process of this embodiment includes: Step 3.2 Improved Adaptive Honey Badger Algorithm Improving the behavior of honey badger populations is divided into two phases: an "exploration phase" and an "exploitation phase," with equal execution probabilities. Their movement behaviors are controlled by the following equations: Old Honey Badger Individual Location Update: ; Determining the oscillation coefficient: ; New honey badger locations updated: ; in, For the honey badger population (each transformer data link) the first i Dimensional data, i ∈[1,2,…, D ]; D For data dimensions; I r The odor intensity between the random individual and the current individual; d r This represents the distance between the current individual and a random individual. The mean individual generated from the three best individuals in the population. i dimension; I μ The odor intensity is the difference between the mean individual and the current individual; d μ This represents the distance between the current individual and the mean individual; The best individual in the population is the first i dimension; r 1- r 7 is a distinct random number between 0 and 1; β It is the honey badger's ability to obtain food and β ≥1; F ∈{-1,1} is used to change the search direction; d r It is the distance between the prey and the corresponding individual.

[0051] Calculate the radiation intensity coefficient of the prey: ; Calculate the area term S of radiation intensity: ; Calculate the prey distance term: ; in, r 5 is a random number between 0 and 1. S It refers to concentration intensity; ; in, C It is a constant and C ≥1, t and t max These represent the current iteration number and the maximum iteration number, respectively.

[0052] To prevent the algorithm from getting stuck in local optima later on, the distance from the individual to the prey and the total population distance to the prey are used to update the algorithm. c , making c Determined by the relative positions of the populations, this enhances the algorithm's optimization capabilities in later stages. .

[0053] Step 3.3 Differential Evolution Algorithm Differential evolution is a method for generating a new vector based on mutation and crossover operations. The mutation process generates a new vector by weighting two random vectors: ; in, , and This represents three distinct individuals in a population. F This indicates that the variation factor is a constant, and F ≥0.

[0054] Crossover operations can enrich population diversity ; in, CR It is the cross factor, which is usually between 0 and 1. The fault type is identified as follows: Further analysis reveals the following transformer fault type: ; in, U fault This represents the most likely type of transformer fault.

[0055] Example 2 Based on the same general inventive concept, this invention also provides a power transformer fault diagnosis device based on an improved multidimensional extended space. The power transformer fault diagnosis device based on an improved multidimensional extended space provided by this invention is described below. The power transformer fault diagnosis device based on an improved multidimensional extended space described below can be referred to in correspondence with the power transformer fault diagnosis method based on an improved multidimensional extended space described above. The device includes: The data acquisition module is used to acquire the transformer's operating status data, which includes data on the content of dissolved gases in the oil. The data balancing and degradation modeling module is used to balance the operating state data using an unbalanced sample set processing method to obtain balanced state data, and to establish a degradation model of the power transformer based on the balanced state data. The degradation model is used to describe the evolution law of the internal electrochemical state of the transformer over time. The multi-domain, multi-scale feature fusion module is used to calculate the one-dimensional and two-dimensional features of the transformer operating state based on the state data output by the degradation model using a multi-scale approximate entropy algorithm, and then fuse the one-dimensional and two-dimensional features using an adaptive channel attention model to obtain multi-domain, multi-scale fused features. The fault diagnosis output module is used to input the multi-domain, multi-scale fused features into a multi-dimensional extended space learning network for cluster analysis. During the cluster analysis, an improved adaptive honey badger algorithm is used to optimize the parameters of the multi-dimensional extended space learning network, and a differential evolution algorithm is used to assist the improved adaptive honey badger algorithm in global optimization, outputting the transformer fault type diagnosis result.

[0056] Example 3 (equivalent to Example 2 above) In this embodiment, a computer terminal device is provided, including: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the above-described power transformer fault diagnosis method based on the improved multidimensional extended space.

[0057] In this embodiment, a computer-readable storage medium is also provided, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the above-described power transformer fault diagnosis method based on the improved multidimensional extended space.

[0058] This invention provides a power transformer fault diagnosis method and device based on an improved multidimensional extended space. An improved high-rank tensor algorithm is used to establish a power transformer state degradation model, which can effectively uncover the nonlinear and multi-scale evolution of the transformer's internal electrochemical state under various fault conditions. An improved adaptive channel attention model effectively extracts the nonlinear and non-stationary characteristics of the transformer's internal state, significantly improving sensitivity to early, minor faults. Cluster analysis of multi-domain, multi-scale features is performed using a multidimensional extended space learning method, an improved adaptive honey badger algorithm, and a differential evolution algorithm. This fully utilizes the complementary information between different features, avoids the limitations of single features, and effectively avoids the problem of clustering operations converging to local optima. This invention improves the model's resistance to noise and operating condition fluctuations, significantly enhancing the stability and accuracy of the diagnostic results, and can be effectively applied to the precise diagnosis of different types of power transformer faults.

[0059] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A power transformer fault diagnosis method based on improved multi-dimensional expansion space, characterized in that, Includes the following steps: Acquire transformer operating status data, including dissolved gas content data in oil; The unbalanced sample set processing method is used to equalize the operating state data to obtain balanced state data. Based on the balanced state data, a degradation model of the power transformer is established. The degradation model is used to describe the evolution law of the internal electrochemical state of the transformer over time. Based on the state data output by the degradation model, the one-dimensional and two-dimensional features of the transformer operating state are calculated using a multi-scale approximate entropy algorithm. The one-dimensional and two-dimensional features are then fused using an adaptive channel attention model to obtain multi-domain multi-scale fused features. The multi-domain, multi-scale fused features are input into a multidimensional extended space learning network for cluster analysis. During the cluster analysis, an improved adaptive honey badger algorithm is used to optimize the parameters of the multidimensional extended space learning network, and a differential evolution algorithm is used to assist the improved adaptive honey badger algorithm in global optimization, outputting transformer fault type diagnosis results.

2. The method according to claim 1, characterized in that, The process of equalizing the operating status data using an imbalanced sample set processing method includes: Identify the minority class and majority class samples in the operational status data; For the minority class samples, an adaptive synthesis method is adopted to adaptively determine the amount of data to be synthesized for each minority class sample based on the proportion of majority class samples in the neighborhood of each minority class sample, and generate new minority class samples. For the majority class samples, the DBSCAN algorithm is used to identify the core points, boundary points, and noise points within the majority class samples, and the boundary points and noise points are deleted.

3. The method according to claim 1, characterized in that, The process of establishing a degradation model for a power transformer based on the equilibrium state data includes: The electrochemical index data under various fault conditions in the equilibrium state data are taken as deterministic components, and the independent and identically distributed measurement noise is taken as random noise components. The deterministic components and the random noise components are superimposed to form the deterioration state of the transformer at the monitoring point. The deterioration state of a transformer is decomposed into a common deterioration trend component and a random deviation component. The common deterioration trend component is used to describe the common aging pattern of the transformer as a whole, and the random deviation component is used to describe the deviation between the deterioration path of a specific transformer and the common deterioration trend.

4. The method according to claim 1, characterized in that, The process of calculating the one-dimensional and two-dimensional features of the transformer operating state using the multi-scale approximate entropy algorithm includes: For the state data sequence output by the degradation model, coarse-grained time series of different scales are formed by calculating the average value of multiple consecutive time points. The approximate entropy of each coarse-grained time series is calculated in an increasing dimension manner to obtain a one-dimensional multi-scale approximate entropy. The continuous data points in the state data sequence are encoded into a two-dimensional matrix. The data in the two-dimensional matrix are normalized and scaled. The scaled data is mapped to angles and radii in polar coordinates. The correlation within different time intervals is measured by trigonometric function transformation of the angles and differences between points, and a two-dimensional multi-scale approximate entropy is obtained.

5. The method according to claim 4, characterized in that, The process of fusing the one-dimensional features and the two-dimensional features using an adaptive channel attention model includes: The one-dimensional multi-scale approximate entropy and the two-dimensional multi-scale approximate entropy are input into the cross-modal channel attention model, and the cross-modal channel attention model calculates channel attention weights for the one-dimensional features and the two-dimensional features respectively through a multilayer perceptron; The fusion factor is determined based on the channel attention weights, and the one-dimensional features and the two-dimensional features are weighted and fused using the fusion factor to output the multi-domain multi-scale fused features.

6. The method according to claim 1, characterized in that, The process of inputting the multi-domain, multi-scale fused features into a multi-dimensional extended spatial learning network for cluster analysis includes: In the input layer of the multidimensional extended space learning network, radial basis operations are used to transform the input features of each dimension based on the variance and mean sequences of the data in each dimension. In the summation layer of the multidimensional extended spatial learning network, the conditional probability density of each fault category is estimated using the Parzen window function; In the output layer of the multidimensional extended spatial learning network, each neuron represents a fault category, and the most likely fault type is determined based on the conditional probability density.

7. The method according to claim 1, characterized in that, The process of optimizing the parameters of the multidimensional extended space learning network using the improved adaptive honey badger algorithm includes: Initialize the honey badger population, where each individual honey badger represents a set of parameters of the multidimensional extended space learning network; During the exploration phase, the individual's location is updated based on the odor intensity and distance between the current individual and random individuals; During the development phase, the location update direction is jointly determined based on the odor intensity and distance between the current individual and the best individual in the population, as well as the odor intensity and distance between the current individual and the average of the three best individuals. The search step size parameter is dynamically updated based on the current distance of the individual to the prey and the distance of the entire population to the prey in order to control the search range of the algorithm. The mutation and crossover operations of the differential evolution algorithm are applied to the honey badger population to generate new individuals.

8. A power transformer fault diagnosis device based on an improved multidimensional extended space, characterized in that, For implementing the method according to any one of claims 1-7, the apparatus comprises: The data acquisition module is used to acquire the transformer's operating status data, which includes data on the content of dissolved gases in the oil. The data balancing and degradation modeling module is used to balance the operating state data using an unbalanced sample set processing method to obtain balanced state data, and to establish a degradation model of the power transformer based on the balanced state data. The degradation model is used to describe the evolution law of the internal electrochemical state of the transformer over time. The multi-domain, multi-scale feature fusion module is used to calculate the one-dimensional and two-dimensional features of the transformer operating state based on the state data output by the degradation model using a multi-scale approximate entropy algorithm, and then fuse the one-dimensional and two-dimensional features using an adaptive channel attention model to obtain multi-domain, multi-scale fused features. The fault diagnosis output module is used to input the multi-domain, multi-scale fused features into a multi-dimensional extended space learning network for cluster analysis. During the cluster analysis process, an improved adaptive honey badger algorithm is used to optimize the parameters of the multi-dimensional extended space learning network, and a differential evolution algorithm is used to assist the improved adaptive honey badger algorithm in global optimization, outputting the transformer fault type diagnosis result.

9. A computer terminal device, characterized in that, include: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors perform the steps of the method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-7.

Images