An intelligent detection method for transformer winding deformation

Abstract

The present invention discloses an intelligent detection method for transformer winding deformation, which includes data acquisition, data augmentation, creating a transformer winding deformation detection model, winding deformation detection, and model calibration. The present invention relates to the technical field of transformer detection, specifically an intelligent detection method for transformer winding deformation. This solution innovatively introduces optimization strategies of probability correction, geometric deviation inspection, and information entropy evaluation to improve the stability and distribution rationality of augmented data, optimize data quality, and improve the adaptability of the model to complex deformation patterns and the intelligent level of transformer condition monitoring by constructing an innovative decision boundary, introducing an adaptive dynamic kernel function, designing a nonlinear boundary relaxation function, and optimizing the regularization term control of the Lagrangian function; by introducing a feedback data calibration mechanism, using contrastive learning to achieve feature alignment, enhancing the aggregation of data of the same category, distinguishing different categories of data, and effectively reducing the misjudgment rate.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer detection, and specifically refers to an intelligent transformer winding deformation detection method. Background Art

[0002] A transformer is a core device in the power system. During long-term operation, its windings may be deformed due to factors such as short-circuit current impact, mechanical vibration, and thermal expansion and contraction. Winding deformation can lead to a decline in electrical performance and even trigger serious power accidents. Therefore, accurately and reliably detecting winding deformation is of great significance.

[0003] Traditional data augmentation methods have problems such as distorted data distribution, single perturbation method, insufficient quality control, and difficulty in screening high-quality data. Traditional transformer winding deformation detection methods have problems such as rigid decision boundaries, fixed kernel functions, single relaxation strategies, and poor adaptability to local data structures. Traditional transformer winding deformation detection models have problems such as model drift, unstable features, and poor adaptability to new data during long-term operation. Summary of the Invention

[0004] In view of the above situation, to overcome the defects of the prior art, the present invention provides an intelligent transformer winding deformation detection method. Aiming at the problems existing in the traditional data augmentation method, such as data distribution distortion, single perturbation method, insufficient quality control, and difficulty in screening high-quality data, this solution innovatively introduces a multi-level optimization strategy of non-linear perturbation conversion, probability correction based on Markov chain, geometric deviation test, and information entropy evaluation, effectively improving the diversity, stability, and distribution rationality of the augmented data, ensuring that the new data can not only enrich the data set but also maintain consistency with the core features of the original data, thereby optimizing the data quality, enhancing the generalization ability of the model, and strengthening the robustness and accuracy of the transformer winding deformation detection; aiming at the problems existing in the traditional transformer winding deformation detection method, such as rigid decision boundary, fixed kernel function, single relaxation strategy, and poor adaptability to the local structure of data, this solution improves the adaptability of the model to complex deformation patterns, enhances the non-linear expression ability of data classification, and improves the detection accuracy and robustness by constructing a dynamic segmented hyperplane decision boundary, introducing an adaptive dynamic kernel function, designing a non-linear boundary relaxation function, and optimizing the regularization term control of the Lagrangian function, enabling the model to more accurately identify the tiny deformation of the transformer winding and improving the intelligent level of transformer condition monitoring; aiming at the problems existing in the traditional transformer winding deformation detection model during long-term operation, such as model drift, unstable features, and poor adaptability to new data, this solution improves the stability and generalization ability of the model and enhances its adaptability under actual working conditions by introducing a feedback data calibration mechanism, using contrastive learning to achieve feature alignment, enhancing the aggregation of data of the same category and distinguishing different categories of data, and at the same time calculating the data distribution deviation through a distribution shift detection mechanism and adjusting the model parameters, effectively reducing the misjudgment rate.

[0005] The technical solution adopted by the present invention is as follows: An intelligent transformer winding deformation detection method, which includes the following steps:

[0006] Step S1: Data acquisition;

[0007] Step S2: Data augmentation;

[0008] Step S3: Create a transformer winding deformation detection model;

[0009] Step S4: Winding deformation detection;

[0010] Step S5: Model calibration.

[0011] Further, in step S1, the data collection is to collect the electrical parameter data, mechanical vibration data, temperature data, and deformation data of historical transformers; the electrical parameter data refers to the voltage, current, frequency, three-phase voltage unbalance rate, transformer load rate, harmonic content, and impedance parameters of the transformer; the mechanical vibration data refers to the vibration signals of the windings; the temperature data includes the temperature of the windings and the ambient temperature; the deformation data refers to whether the transformer windings are deformed; meanwhile, an original data set is created.

[0012] Further, in step S2, the data augmentation specifically includes the following steps:

[0013] Step S21: Perturbation transformation, construct a non-linear mapping function to perturb the original data in the high-dimensional space to generate new data points, which is expressed as follows:

[0014] ;

[0015] where, represents the data after perturbation transformation, represents the original data, and represent the transformation weights, represents the rectified linear unit function, represents the hyperbolic tangent function, Dt represents the noise factor, represents the noise variance, represents the unit rectangle, represents that the noise factor follows a normal distribution with a mean of 0 and a variance of ;

[0016] Step S22: Probability correction, construct new data points based on the Markov chain to make the generated data conform to the original distribution while introducing perturbations, which is expressed as follows:

[0017] ;

[0018] where, represents the data after probability correction, represents the transition factor whose elements follow a beta distribution, represents the data sample randomly selected from the original data set, represents the cooperation coefficient;

[0019] Step S23: Create a seed data set, and add the new data generated after perturbation transformation and probability correction to the created seed data set;

[0020] Step S24: Geometric test, calculate the geometric deviation between the data in the seed data set and the original data in the high-dimensional space to ensure that the new data points do not deviate from the distribution core, which is expressed as follows:

[0021] ;

[0022] wherein, represents the geometric deviation between the data of the seed data set and the original data, i represents the index of the data point, N represents the total number of data points, represents the exponential function with the natural constant as the base, represents the i-th data point in the seed data set, represents the i-th original data point, represents taking the modulus;

[0023] Step S25: Information entropy test, calculate the change in the information entropy of the data in the seed data set, which is expressed as follows:

[0024] ;

[0025] wherein, represents the information entropy deviation between the data of the seed data set and the original data, and respectively represent the probability densities of the data of the seed data set and the original data, represents taking the absolute value, represents taking the logarithm;

[0026] Step S26: Data augmentation, calculate the geometric deviation and information entropy deviation between all the data in the seed data set and the original data, remove the data in the seed data set where both the geometric deviation and the information entropy deviation are in the maximum 50%, and add the remaining data in the seed data set to the original data.

[0027] Furthermore, in step S3, the creation of the transformer winding deformation detection model specifically includes the following steps:

[0028] Step S31: Create a transformation matrix group, the matrix group contains two matrices, one is an upper triangular matrix and the other is a diagonal matrix;

[0029] Step S32: Define the objective function, set whether the transformer winding is deformed as the label of the model, introduce a dynamic segmented hypersurface as the decision boundary, and the shape of the hypersurface is determined by the transformation matrix group, which is expressed as follows:

[0030] ;

[0031] wherein, and respectively represent the weight vector and bias vector of the model, represents the transformation matrix group, represents the parameter when obtaining the minimum value of the objective expression , and , 、 and represent the importance weights, represents the loss function, represents the class label of the i-th data point, represents the i-th data point, represents the output of the decision function, k represents the index of the matrix in the transformation matrix group, represents the k-th matrix in the transformation matrix group, represents taking the square of the Frobenius norm;

[0032] Step S33: Introduce a dynamic kernel function and adjust the shape and parameters of the kernel function according to the local structure of the input data, which is expressed as follows:

[0033] ;

[0034] where j represents the index of the data point, represents the j-th data point, represents the dynamic kernel function between the i-th and j-th data points, represents the bandwidth function, represents the bandwidth function value between the i-th and j-th data points, represents the cosine function, T represents the transpose symbol, represents a randomly selected matrix in the transformation matrix group, and represent the bandwidth weights, represents the median of the Euclidean distances between the i-th data point and all other data points;

[0035] Step S34: Generate the decision function, which is expressed as follows:

[0036] ;

[0037] where, represents the sign function, represents the Lagrange multiplier of the i-th data point, represents the decision bias;

[0038] Step S35: Define the constraint conditions, introduce a dynamic boundary relaxation function, and determine the relaxation degree of each sample point through a non-linear function, which is expressed as follows:

[0039] ;

[0040] where, represents the relaxation variable, represents the relaxation coefficient, represents the relaxation degree control factor;

[0041] Step S36: Generate the Lagrangian function, introduce a regularization term to control the variation range of the dual variables, introduce the constraint conditions into the objective function, and construct the Lagrangian function, which is expressed as follows:

[0042] ;

[0043] where, represents the set of Lagrange multipliers, represents the set of importance weights, represents the Lagrangian function, represents at the parameter of , , and the value of the Lagrangian function under the conditions of

[0044] Furthermore, in step S4, the winding deformation detection is to collect the electrical parameter data, mechanical vibration data and temperature data of the transformer in real time, input the data into the transformer winding deformation detection model, and the model judges whether the transformer winding is deformed.

[0045] Furthermore, in step S5, the model calibration specifically includes the following steps:

[0046] Step S51: Obtain the feedback data, obtain the electrical parameter data, mechanical vibration data and temperature data of the transformer input into the model, and record the judgment result of the model on whether the transformer winding is deformed, and combine the data into a feedback data set;

[0047] Step S52: Feature alignment, for the transformer data in the feedback data set, use the contrast learning method to make the distance between data of the same category closer in the high-dimensional space and data of different categories farther away, which is expressed as follows:

[0048] ;

[0049] where, represents the feature alignment loss, represents the initial data of the model in the data set for model training, represents the feedback data in the feedback data set, represents whether the initial data of the model and the feedback data belong to the same category. When they belong to the same category is equal to 1, and when they belong to different categories is equal to 0, m represents the margin factor, represents taking the maximum value;

[0050] Step S53: Feedback calibration. Introduce a distribution shift detection mechanism, calculate the deviation between the initial data distribution of the model and the feedback data distribution, and perform model calibration, which is expressed as follows:

[0051] ;

[0052] where, represents the distribution deviation, represents taking the overall deviation distance, represents the overall deviation distance between the initial data distribution of the model and the feedback data distribution, represents the model parameters after feedback calibration, represents the initial model parameters, represents the calibration coefficient, represents the gradient of the distribution deviation with respect to the model parameters.

[0053] The beneficial effects achieved by the present invention using the above solution are as follows:

[0054] (1) Aiming at the problems of data distribution distortion, single perturbation method, insufficient quality control, and difficulty in screening high-quality data existing in traditional data augmentation methods, this solution innovatively introduces a multi-level optimization strategy of non-linear perturbation transformation, probability correction based on Markov chain, geometric deviation test, and information entropy evaluation, effectively improving the diversity, stability, and distribution rationality of the augmented data, ensuring that the new data can not only enrich the data set but also maintain consistency with the core features of the original data, thereby optimizing data quality, enhancing the generalization ability of the model, and enhancing the robustness and accuracy of transformer winding deformation detection.

[0055] (2) Aiming at the problems of rigid decision boundaries, fixed kernel functions, single relaxation strategies, and poor adaptability to local data structures existing in traditional transformer winding deformation detection methods, this solution improves the adaptability of the model to complex deformation patterns, enhances the non-linear expression ability of data classification, improves detection accuracy and robustness, enables the model to more accurately identify minor deformations of transformer windings, and improves the intelligent level of transformer condition monitoring by constructing a dynamic segmented hyper-surface decision boundary, introducing an adaptive dynamic kernel function, designing a non-linear boundary relaxation function, and optimizing the regularization term control of the Lagrangian function.

[0056] (3) Aiming at the problems of model drift, unstable features, and poor adaptability to new data existing in traditional transformer winding deformation detection models during long-term operation, this solution improves the stability and generalization ability of the model and enhances its adaptability under actual working conditions by introducing a feedback data calibration mechanism, using contrastive learning to achieve feature alignment, enhancing the aggregation of same-class data and distinguishing different-class data, and at the same time calculating the data distribution deviation through a distribution shift detection mechanism and adjusting the model parameters, effectively reducing the misjudgment rate. Description of the Drawings

[0057] Figure 1 Schematic diagram of an intelligent transformer winding deformation detection method provided by the present invention;

[0058] Figure 2 Schematic diagram of data augmentation provided by the present invention;

[0059] Figure 3 Schematic diagram for creating a transformer winding deformation detection model;

[0060] Figure 4 Schematic diagram for model calibration.

[0061] The accompanying drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, but do not constitute a limitation to the present invention. Specific embodiments

[0062] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of the present invention.

[0063] In the description of the present invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc. indicating the orientation or positional relationship are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation to the present invention.

[0064] Embodiment 1, refer to Figure 1 , an intelligent transformer winding deformation detection method provided by the present invention, the method includes the following steps:

[0065] Step S1: Data acquisition, acquiring electrical parameter data, mechanical vibration data, temperature data, and deformation data of historical transformers;

[0066] Step S2: Data augmentation, performing data augmentation through perturbation transformation, probability correction, creating a seed data set, geometric test, and information entropy test;

[0067] Step S3: Creating a transformer winding deformation detection model, creating a transformer winding deformation detection model by creating a transformation matrix group, defining an objective function, introducing a dynamic kernel function, generating a decision function, defining constraint conditions, and generating a Lagrangian function;

[0068] Step S4: Winding deformation detection, collect the electrical parameter data, mechanical vibration data and temperature data of the transformer in real time, input the data into the transformer winding deformation detection model, and the model determines whether the transformer winding is deformed;

[0069] Step S5: Model calibration, which is achieved by obtaining feedback data, feature alignment and feedback calibration.

[0070] Embodiment 2, refer to Figure 1 , this embodiment is based on the above embodiment. In step S1, the data collection is specifically to collect the electrical parameter data, mechanical vibration data, temperature data and deformation data of historical transformers; the electrical parameter data refers to the voltage, current, frequency, three-phase voltage unbalance rate, transformer load rate, harmonic content and impedance parameters of the transformer; the mechanical vibration data refers to the vibration signal of the winding; the temperature data includes the temperature of the winding and the temperature of the environment; the deformation data refers to whether the transformer winding is deformed; at the same time, create an original data set, set each piece of collected data as a feature, each group of data features forms a data vector, and add all the data vectors formed by the collected data to the original data set to form a complete original data set.

[0071] Embodiment 3, refer to Figure 1 and Figure 2 , this embodiment is based on the above embodiment. In step S2, the data augmentation specifically includes the following steps:

[0072] Step S21: Perturbation transformation, construct a non-linear mapping function to perturb the original data in the high-dimensional space to generate new data points, which is expressed as follows:

[0073] ;

[0074] Among them, represents the data after perturbation transformation, represents the original data, and represent the transformation weights, represents the rectified linear unit function, represents the hyperbolic tangent function, Dt represents the noise factor, represents the noise variance, represents the unit rectangle, represents that the noise factor follows a normal distribution with a mean of 0 and a variance of ;

[0075] Step S22: Probability correction, construct new data points based on the Markov chain, so that the generated data conforms to the original distribution while introducing perturbations, which is expressed as follows:

[0076] ;

[0077] Among them, represents the data after probability correction, represents the transition factor for the elements to follow the beta distribution, represents the data sample randomly selected from the original dataset, represents the synergy coefficient;

[0078] Step S23: Create a seed dataset, and add the new data generated after perturbation transformation and probability correction to the created seed dataset;

[0079] Step S24: Geometric test, calculate the geometric deviation between the data in the seed dataset and the original data in the high-dimensional space to ensure that the new data points do not deviate from the distribution core, which is expressed as follows:

[0080] ;

[0081] Among them, represents the geometric deviation between the data in the seed dataset and the original data, i represents the index of the data point, and N represents the total number of data points, represents the exponential function with the natural constant as the base, represents the i-th data point in the seed dataset, represents the i-th original data point, represents taking the modulus;

[0082] Step S25: Information entropy test, calculate the change in the information entropy of the data in the seed dataset, which is expressed as follows:

[0083] ;

[0084] Among them, represents the information entropy deviation between the data in the seed dataset and the original data, and respectively represent the probability densities of the data in the seed dataset and the original data, represents taking the absolute value, represents taking the logarithm;

[0085] Step S26: Data augmentation, calculate the geometric deviation and information entropy deviation between all the data in the seed dataset and the original data, remove the data in the seed dataset where both the geometric deviation and the information entropy deviation are in the maximum 50%, and add the remaining data in the seed dataset to the original data.

[0086] By performing the above operations, in view of the problems of data distribution distortion, single perturbation method, insufficient quality control, and difficulty in screening high-quality data existing in traditional data augmentation methods, this solution innovatively introduces a multi-level optimization strategy of non-linear perturbation transformation, probability correction based on Markov chain, geometric deviation test, and information entropy evaluation, effectively improving the diversity, stability, and distribution rationality of augmented data, ensuring that the new data can not only enrich the data set but also maintain consistency with the core features of the original data, thereby optimizing data quality, enhancing the generalization ability of the model, and enhancing the robustness and accuracy of transformer winding deformation detection.

[0087] Example 4, refer to Figure 1 and Figure 3 , based on the above example, in step S3, the creation of the transformer winding deformation detection model specifically includes the following steps:

[0088] Step S31: Create a transformation matrix group, which contains two matrices, one is an upper triangular matrix and the other is a diagonal matrix;

[0089] Step S32: Define the objective function. Set whether the transformer winding is deformed as the label of the model, and introduce a dynamic segmented hypersurface as the decision boundary. The shape of the hypersurface is determined by the transformation matrix group, which is expressed as follows:

[0090] ;

[0091] Among them, and respectively represent the weight vector and bias vector of the model, represents the transformation matrix group, represents the parameter when obtaining the minimum value of the target expression , and , , and represent the importance weights, represents the loss function, represents the class label of the i-th data point, represents the i-th data point, represents the output of the decision function, k represents the index of the matrix in the transformation matrix group, represents the k-th matrix in the transformation matrix group, represents taking the square of the Frobenius norm;

[0092] Step S33: Introduce a dynamic kernel function, and adjust the shape and parameters of the kernel function according to the local structure of the input data, which is expressed as follows:

[0093] ;

[0094] Among them, j represents the index of the data point, represents the j-th data point, represents the dynamic kernel function between the i-th and j-th data points, represents the bandwidth function, represents the bandwidth function value between the i-th and j-th data points, represents the cosine function, and T represents the transpose symbol, represents a matrix randomly selected from the transformation matrix group, and represents the bandwidth weight, represents the median of the Euclidean distances between the i-th data point and all other data points;

[0095] Step S34: Generate a decision function, which is expressed as follows:

[0096] ;

[0097] Among them, represents the sign function, represents the Lagrange multiplier of the i-th data point, represents the decision bias;

[0098] Step S35: Define the constraint conditions, introduce a dynamic boundary relaxation function, and determine the relaxation degree of each sample point through a non-linear function, which is expressed as follows:

[0099] ;

[0100] Among them, represents the relaxation variable, represents the relaxation coefficient, represents the relaxation degree control factor;

[0101] Step S36: Generate the Lagrangian function, introduce a regularization term to control the variation range of the dual variables, introduce the constraint conditions into the objective function, and construct the Lagrangian function, which is expressed as follows:

[0102] ;

[0103] Among them, represents the set of Lagrange multipliers, represents the set of importance weights, represents the Lagrangian function, represents at the parameter 、 、 and the value of the Lagrangian function under the conditions of.

[0104] By performing the above operations, in view of the problems existing in the traditional transformer winding deformation detection method, such as rigid decision boundary, fixed kernel function, single relaxation strategy, and poor adaptability to the local structure of data, this solution improves the adaptability of the model to complex deformation modes, enhances the non-linear expression ability of data classification, improves the detection accuracy and robustness by constructing a dynamic segmented hypersurface decision boundary, introducing an adaptive dynamic kernel function, designing a non-linear boundary relaxation function, and optimizing the regularization term control of the Lagrangian function, enabling the model to more accurately identify the minor deformations of the transformer winding and improving the intelligent level of transformer condition monitoring.

[0105] Example 5, refer to Figure 1 and Figure 4 , based on the above example, in step S5, the model calibration specifically includes the following steps:

[0106] Step S51: Obtain feedback data, obtain the electrical parameter data, mechanical vibration data, and temperature data of the transformer input into the model, and record the judgment result of whether the transformer winding is deformed by the model, and combine the data into a feedback data set;

[0107] Step S52: Feature alignment, for the transformer data in the feedback data set, use the contrast learning method to make the distance between data of the same category closer in the high-dimensional space and data of different categories farther away, expressed as follows:

[0108] ;

[0109] Among them, represents the feature alignment loss, represents the initial data of the model in the data set for model training, represents the feedback data in the feedback data set, represents whether the initial data of the model and the feedback data belong to the same category. When they belong to the same category is equal to 1, and when they belong to different categories is equal to 0, m represents the interval factor, represents taking the maximum value;

[0110] Step S53: Feedback calibration, introduce a distribution shift detection mechanism, calculate the deviation between the distribution of the initial data of the model and the distribution of the feedback data, and perform model calibration, expressed as follows:

[0111] ;

[0112] Among them, represents the distribution deviation, represents taking the overall deviation distance, Represents the overall deviation distance between the initial data distribution of the model and the feedback data distribution. Represents the model parameters after feedback calibration. Represents the initial parameters of the model. Represents the calibration coefficient. Represents the gradient of the distribution deviation with respect to the model parameters.

[0113] By performing the above operations, for the problems of model offset, unstable features, and poor adaptability to new data in the long-term operation of the traditional transformer winding deformation detection model, this solution introduces a feedback data calibration mechanism, uses contrastive learning to achieve feature alignment, enhances the aggregation of the same category of data, distinguishes different categories of data, and at the same time calculates the data distribution deviation through the distribution offset detection mechanism and adjusts the model parameters, thereby improving the stability and generalization ability of the model, enhancing its adaptability under actual working conditions, and effectively reducing the misjudgment rate.

[0114] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device.

[0115] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention.

[0116] The above describes the present invention and its implementation manners, and this description is not restrictive. What is shown in the drawings is only one of the implementation manners of the present invention, and the actual structure is not limited thereto. In short, if those of ordinary skill in the art are inspired by it and design similar structural manners and embodiments without creative efforts without departing from the purpose of the present invention, they shall fall within the protection scope of the present invention.

Claims

1. An intelligent transformer winding deformation detection method, characterized in that: The method comprises the following steps: Step S1: data collection, collecting historical transformer electrical parameter data, mechanical vibration data, temperature data and deformation data; Step S2: Data augmentation, data augmentation is performed through perturbation conversion, probability correction, creation of seed data set, geometric test and information entropy test; Step S21: perturbation conversion, constructing a nonlinear mapping function to perturb the original data in a high-dimensional space and generate new data points, which are expressed as follows: ； in, represents the data after disturbance conversion, Represents the original data, and represents the conversion weight, represents the linear rectification function, represents the hyperbolic tangent function, Dt represents the noise factor, represents the noise variance, represents the unit rectangle, It means that the noise factor has a mean of 0 and a variance of Normal distribution of Step S22: Probability correction, constructing new data points based on the Markov chain so that the generated data conforms to the original distribution while introducing perturbations, as shown below: ； in, Represents the data after probability correction, The transition factor indicating that the element follows the Beta distribution, represents a data sample randomly selected from the original data set, represents the synergy coefficient; Step S23: creating a seed data set, and adding the new data generated after the disturbance conversion and probability correction to the created seed data set; Step S24: Geometric check, calculate the geometric deviation between the data of the seed data set and the original data in the high-dimensional space to ensure that the new data point does not deviate from the distribution core, as shown below: ； in, represents the geometric deviation between the data of the seed data set and the original data, i represents the index of the data point, and N represents the total number of data points. represents an exponential function with a natural constant as base, represents the i-th data point in the seed dataset, represents the i-th original data point, Indicates modulus; Step S25: Information entropy test, calculating the information entropy change of the data in the seed data set, expressed as follows: ； in, Represents the information entropy deviation between the data of the seed data set and the original data, and Represent the probability density of the data of the seed data set and the original data respectively, Indicates taking the absolute value, It means taking logarithm; Step S26: data augmentation, calculating the geometric deviation and information entropy deviation between all data in the seed data set and the original data, removing the data in the seed data set with the largest geometric deviation and information entropy deviation at 50%, and adding the remaining data in the seed data set to the original data; Step S3: creating a transformer winding deformation detection model by creating a transformation matrix group, defining an objective function, introducing a dynamic kernel function, generating a decision function, defining constraint conditions, and generating a Lagrangian function; Step S31: Create a transformation matrix group, the matrix group includes two matrices, one is an upper triangular matrix, and the other is a diagonal matrix; Step S32: define the objective function, set whether the transformer winding is deformed as the label of the model, introduce a dynamic segmented hypersurface as the decision boundary, and the shape of the hypersurface is determined by the transformation matrix group, which is expressed as follows: ； in, and Represent the weight vector and bias vector of the model respectively, represents the transformation matrix group, Indicates the parameter used to find the minimum value of the target expression , and , , and represents the importance weight, represents the loss function, represents the category label of the i-th data point, represents the i-th data point, represents the output of the decision function, k represents the index of the matrix in the transformation matrix group, represents the kth matrix in the transformation matrix group, represents the square of the Frobenius norm; Step S33: Introduce a dynamic kernel function, and adjust the shape and parameters of the kernel function according to the local structure of the input data, as shown below: ； Where j represents the index of the data point, represents the jth data point, represents the dynamic kernel function between the i-th and j-th data points, represents the bandwidth function, represents the bandwidth function value between the i-th and j-th data points, represents the cosine function, T represents the transposed symbol, represents a randomly selected matrix from the transformation matrix group, and represents the bandwidth weight, Represents the median of the Euclidean distance between the i-th data point and all other data points; Step S4: winding deformation detection, using a transformer winding deformation detection model to detect the deformation of the transformer winding; Step S5: Model calibration, which is achieved by obtaining feedback data, feature alignment and feedback calibration.

2. The intelligent transformer winding deformation detection method according to claim 1 is characterized in that: In step S3, the transformer winding deformation detection model is created, specifically comprising the following steps: Step S31: Create a transformation matrix group; Step S32: define the objective function; Step S33: introducing a dynamic kernel function; Step S34: Generate a decision function, which is expressed as follows: ； in, represents the symbolic function, represents the Lagrange multiplier of the ith data point, represents decision bias; Step S35: Define constraint conditions, introduce dynamic boundary relaxation function, and determine the relaxation degree of each sample point through nonlinear function, which is expressed as follows: ； in, represents the slack variable, represents the relaxation coefficient, represents the relaxation degree control factor; Step S36: Generate a Lagrangian function, introduce a regularization term to control the range of change of the dual variable, introduce constraints into the objective function, and construct a Lagrangian function, which is expressed as follows: ； in, represents the set of Lagrange multipliers, represents the importance weight set, represents the Lagrangian function, Indicates that the parameter is , , and The value of the Lagrangian function under the condition of .

3. The intelligent transformer winding deformation detection method according to claim 1 is characterized in that: In step S5, the model calibration specifically includes the following steps: Step S51: obtaining feedback data, obtaining electrical parameter data, mechanical vibration data and temperature data of the transformer input into the model, and recording the model's judgment result on whether the transformer winding is deformed, and combining the data into a feedback data set; Step S52: feature alignment. For the transformer data of the feedback data set, a contrastive learning method is used to make the distance between the data of the same category closer in the high-dimensional space and the data of different categories farther apart, as shown below: ； in, represents the feature alignment loss, Represents the model initial data in the dataset used for model training, represents the feedback data in the feedback dataset, Indicates whether the model initial data and feedback data belong to the same category. If they belong to the same category, Equal to 1, belonging to different categories is equal to 0, m represents the interval factor, Indicates taking the maximum value; Step S53: Feedback calibration, introducing a distribution offset detection mechanism, calculating the deviation between the initial data distribution of the model and the feedback data distribution, and performing model calibration, as shown below: ； in, represents the distribution deviation, Indicates the overall deviation distance. Represents the overall deviation distance between the model's initial data distribution and the feedback data distribution, represents the model parameters after feedback calibration, represents the initial parameters of the model, represents the calibration coefficient, represents the gradient of the distribution deviation with respect to the model parameters.

4. The intelligent transformer winding deformation detection method according to claim 1 is characterized in that: In step S1, the data collection is to collect electrical parameter data, mechanical vibration data, temperature data and deformation data of the historical transformer; the electrical parameter data refers to the voltage, current, frequency, three-phase voltage imbalance rate, transformer load rate, harmonic content and impedance parameters of the transformer; the mechanical vibration data refers to the vibration signal of the winding; the temperature data includes the temperature of the winding and the temperature of the environment; the deformation data refers to whether the transformer winding is deformed; and the original data set is created at the same time.

5. The intelligent transformer winding deformation detection method according to claim 1 is characterized in that: In step S4, the winding deformation detection collects the electrical parameter data, mechanical vibration data and temperature data of the transformer in real time, inputs the data into the transformer winding deformation detection model, and the model determines whether the transformer winding is deformed.

Images