Transformer winding deformation degree identification method and device, equipment and storage medium

Abstract

The application provides a transformer winding deformation degree identification method and device, equipment and a storage medium, and belongs to the technical field of transformer anomaly identification. The method comprises the following steps: dividing a frequency response signal of a transformer winding into multiple equal-length segments based on a logarithmic frequency segmentation method, calculating the correlation entropy of each two segments to construct a correlation entropy matrix; converting the correlation entropy matrix into a correlation entropy image with RGB channels, performing feature extraction on the correlation entropy image to obtain an initial feature map; performing at least one feature enhancement processing on the initial feature map to obtain a target enhanced feature map; and obtaining a deformation degree identification result of the transformer winding based on the target enhanced feature map. The application can improve the identification accuracy of the slight deformation of the transformer winding in the online monitoring scene of the motor train unit.

Description

Technical Field

[0001] This application belongs to the field of transformer anomaly identification technology, and more specifically, it relates to a method, device, equipment, and storage medium for identifying the degree of deformation of transformer windings. Background Technology

[0002] The detection of winding deformation in auxiliary transformers of high-speed trains is crucial for train operation safety. Winding deformation can alter frequency response characteristics, thereby affecting the stability of the train's power supply system. The current mainstream frequency response analysis (FRA) method identifies winding deformation by analyzing frequency response signals in the 1kHz-100MHz frequency band, but this technology has significant limitations when dealing with complex deformation characteristics.

[0003] Traditional methods for handling complex deformation features, such as segmenting through fixed windows or manually statistically analyzing features and analyzing peak offset and spectral amplitude distribution, are insufficient to capture subtle fluctuations across the entire frequency band simultaneously. This results in inadequate sensitivity to minor deformations such as the initial stage of inter-turn short circuits and slight winding displacements, failing to meet the accuracy requirements for identifying minor deformations of transformer windings in online monitoring scenarios for high-speed trains. Summary of the Invention

[0004] The purpose of this application is to provide a method, device, equipment, and storage medium for identifying the degree of deformation of transformer windings, so as to solve one or more technical problems mentioned in the background art and improve the identification accuracy of slight deformation of transformer windings in the online monitoring scenario of EMU trains.

[0005] A first aspect of this application provides a method for identifying the degree of deformation of transformer windings, including: The frequency response signal of the transformer winding is divided into multiple segments of equal length based on the logarithmic frequency segmentation method, and the correlation entropy of every two segments is calculated to construct a correlation entropy matrix. The correlation entropy matrix is ​​converted into a correlation entropy image with RGB channels, and features are extracted from the correlation entropy image to obtain an initial feature map; Perform at least one feature enhancement process on the initial feature map to obtain the target enhanced feature map; The deformation degree identification result of the transformer winding is obtained based on the target enhanced feature map; Each feature enhancement process includes: The feature map to be enhanced is subjected to horizontal pooling, vertical pooling, and feature extraction to obtain the target enhanced feature map. If the feature enhancement process is performed multiple times, the feature map to be enhanced is the initial feature map for the first feature enhancement process; for subsequent feature enhancement processes, the feature map to be enhanced is the feature map after the previous enhancement process.

[0006] A second aspect of this application provides a transformer winding deformation degree identification device, comprising: The first data processing unit is used to divide the frequency response signal of the transformer winding into multiple segments of equal length based on the logarithmic frequency segmentation method, and to calculate the correlation entropy of every two segments to construct a correlation entropy matrix. The second data processing unit is used to convert the correlation entropy matrix into a correlation entropy image with RGB channels, and to extract features from the correlation entropy image to obtain an initial feature map. The third data processing unit is used to perform at least one feature enhancement process on the initial feature map to obtain the target enhanced feature map; An anomaly identification unit is used to obtain the deformation degree identification result of the transformer winding based on the target enhanced feature map; The third data processing unit is specifically used to perform horizontal pooling, vertical pooling, and feature extraction on the feature map to be enhanced, so as to obtain the target enhanced feature map. If the feature enhancement process is performed multiple times, the feature map to be enhanced is the initial feature map for the first feature enhancement process; for subsequent feature enhancement processes, the feature map to be enhanced is the feature map after the previous enhancement process.

[0007] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the above-described method for identifying the degree of deformation of transformer windings.

[0008] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method for identifying the degree of deformation of transformer windings.

[0009] The beneficial effects of the transformer winding deformation identification method, device, equipment, and storage medium provided in this application embodiment are as follows: This implementation divides the frequency response signal of the transformer winding into multiple equal-length segments based on a logarithmic frequency segmentation method, and constructs a correlation entropy matrix by calculating the correlation entropy between every two segments. Logarithmic frequency segmentation can better adapt to the characteristics of the frequency response signal in different frequency bands, while the correlation entropy matrix can effectively capture the similarities and differences between signal segments. This segmentation method can solve the problem of not being able to obtain subtle changes in high-frequency bands above 10MHz in existing technologies, enabling effective identification of the degree of transformer winding deformation, and capturing the frequency characteristic correlation between segments through the correlation entropy matrix.

[0010] Furthermore, this embodiment converts the correlation entropy matrix into a correlation entropy image with RGB channels, transforming complex matrix data into an intuitive image format, which facilitates feature extraction. The initial feature map obtained in this way can more comprehensively reflect the state information of the transformer windings, helping to improve recognition accuracy.

[0011] Furthermore, in the feature enhancement process, this embodiment performs pooling operations in both horizontal and vertical directions, enabling analysis and processing of the feature map from multiple dimensions to obtain more comprehensive feature information. This multi-dimensional processing approach helps enhance feature expression, allowing the model to better capture the characteristic manifestations of winding deformation in different directions and improve its adaptability to various deformation conditions. In addition, through multiple feature enhancement processes, the model can gradually filter out noise and interference factors in the signal, extracting more representative and stable features. Even if the input frequency response signal contains some noise or fluctuations, a reliable target enhanced feature map can still be obtained after multi-level feature enhancement processing, thereby ensuring the robustness of the transformer winding deformation degree identification results and reducing the possibility of misjudgment. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A flowchart illustrating a method for identifying the degree of deformation of transformer windings provided in an embodiment of this application; Figure 2 A schematic diagram illustrating the generation of a CEI image according to an embodiment of this application; Figure 3 An architecture diagram of an MSA-CNN model provided in an embodiment of this application; Figure 4 This is a structural block diagram of a transformer winding deformation degree identification device provided in an embodiment of this application; Figure 5 This is a schematic block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0014] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0015] Related technologies use deep learning models such as Support Vector Machines or Random Forests to process the features of the frequency response curve to obtain the identification results of transformer anomalies. The features in the above scheme are obtained by extracting the frequency response curve in segments using a fixed window. This scheme has the following drawbacks: Insufficient feature representation ability: The fixed window segmentation method cannot adapt to the multi-scale characteristics of frequency response curves and has a weak ability to represent cross-frequency band deformation.

[0016] Lack of attention mechanism: The above deep learning model cannot automatically focus on the frequency band sensitive to transformer winding deformation, and has limited ability to capture key features that affect deformation, which can easily lead to a decrease in the accuracy of moderate deformation identification.

[0017] The inability to quantify structural similarity differences between different frequency band segments: The above methods rely on the overall similarity of a single frequency response curve for comparison, failing to quantify the structural similarity differences between different frequency band segments. This results in low accuracy in transformer winding deformation location and weak noise immunity. When the frequency response curve exhibits non-deformation-related fluctuations due to electromagnetic interference, temperature fluctuations, and other factors during EMU operation, traditional methods are prone to misjudgment. Furthermore, they lack targeted processing mechanisms for minute feature changes in high-frequency bands above 10MHz, making it difficult to meet the multiple requirements for accuracy, real-time performance, and anti-interference capabilities in EMU online monitoring scenarios.

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following description will be provided in conjunction with the accompanying drawings and specific embodiments.

[0019] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating a method for identifying the degree of deformation of transformer windings according to an embodiment of this application. The method can be executed by an electronic device and may include steps S101 to S104.

[0020] S101: Divide the frequency response signal of the transformer winding into multiple segments of equal length based on the logarithmic frequency segmentation method, and calculate the correlation entropy of every two segments to construct a correlation entropy matrix.

[0021] In this embodiment, the frequency response signal of the transformer winding can be a preprocessed signal obtained from the original frequency response signal of the auxiliary transformer winding of the EMU.

[0022] Specifically, a frequency response analysis (FRA) instrument can be used to acquire the original frequency response signal of the auxiliary transformer winding of the EMU (Electric Multiple Unit) train. The testing process strictly follows the GB / T1094.3-2017 standard. The FRA outputs a sinusoidal excitation signal in the range of 1kHz to 100MHz, which is injected into the winding through a coupling capacitor, while the amplitude of the frequency response is acquired simultaneously. The voltage setting during the test does not exceed AC220V (±15%), and the scanner uses logarithmic intervals when performing frequency scanning, acquiring the same number of points every ten octaves. That is, the number of sampling points is the same in each of the five intervals: 1kHz~10kHz, 10kHz~100kHz, 100kHz~1MHz, 1MHz~10MHz, and 10MHz~100MHz. The original frequency response signal is a sequence of amplitude data at different frequency points, which is then analyzed by two-dimensional data pairs. Composition, in which This represents the frequency at frequency point n. This represents the amplitude at frequency point n.

[0023] In this embodiment, in order to eliminate data deviation and data fluctuation, the original frequency response signal needs to be preprocessed, such as by normalization (3Zscore normalization method), specifically as follows: ; Where μ is the mean amplitude of the original frequency response signal, and σ is the standard deviation. This represents the original amplitude at frequency point n. This processing standardizes the original frequency response signal to a frequency response signal with a mean of 0 and a standard deviation of 1, thereby eliminating amplitude deviations under different test conditions.

[0024] In this embodiment, after obtaining the frequency response signal of the transformer winding, it can be divided into multiple segments of equal length based on the logarithmic frequency segmentation method. Each segment covers the same frequency octave and contains the same data points. Therefore, the total number of segments in the normalized signal is [number missing]. m represents the number of data points in each segment, n is the frequency point, and the maximum value of n is the total number of data points.

[0025] In this embodiment, the steps for calculating the correlation entropy of every two segments and constructing the correlation entropy matrix are as follows: For every two segments, calculate the relevant entropy and construct the relevant entropy matrix C, specifically as follows: Calculate the joint probability of the a-th data point in the i-th segment and the b-th data point in the j-th segment. The calculation method is as follows: ; in, This represents the a-th data point in the i-th segment. This represents the b-th data point in the j-th segment.

[0026] Calculate the entropy between the i-th segment and the j-th segment: ; Calculate the marginal entropy value of the i-th segment: ; Calculate the marginal entropy value of segment j: ; in, Let be the marginal probability of the a-th data point in the i-th segment. Let be the marginal probability of the b-th data point in the j-th segment.

[0027] The correlation entropy matrix C is a k×k square matrix (k is the number of segments), and each element in this matrix... The cross-correlation entropy between the entropy values ​​of segment i and segment j is expressed by the formula: . It can also be used The symbol "-" indicates a joint association relationship.

[0028] (refer to Figure 2 ).

[0029] S102: Convert the correlation entropy matrix into a correlation entropy image with RGB channels, and extract features from the correlation entropy image to obtain an initial feature map.

[0030] In this embodiment, converting the correlation entropy matrix into a correlation entropy image with RGB channels includes: The correlation entropy matrix is ​​converted into a grayscale matrix, and the element values ​​in the grayscale matrix are mapped to the RGB channels respectively to obtain the correlation entropy image.

[0031] In this embodiment, the relevant entropy matrix can first be converted into a grayscale matrix using minimum-maximum normalization: ; in, This represents the minimum value in matrix C. 225 represents the maximum value in matrix C. 225 means that the correlation entropy matrix C is normalized and mapped to an integer matrix of 0 to 225. Multiplying by 255 corresponds to the 8-bit grayscale color space of the image.

[0032] Secondly, the elements in the grayscale matrix are mapped to the R, G, and B channels respectively to obtain the Correntropy Image (CEI) (see reference). Figure 2In this embodiment, if the original size of the grayscale matrix is ​​inconsistent with the target size (the preset size of the CEI image, which can be set according to the experiment), bilinear interpolation is used for adjustment. The interpolation process preserves the relative distribution of frequency differences. The diagonal elements of the matrix form the main diagonal baseline of the CEI image because their own differences are 0. The grayscale distribution in the off-diagonal areas reflects the frequency domain similarity between segments. The more severe the winding deformation, the denser the off-diagonal highlight areas. This mapping method allows the frequency domain characteristics of the frequency response signal to be directly reflected in the image color space, so that the generated CEI image can be directly used for subsequent model processing, thereby improving the accuracy of identifying the degree of transformer winding deformation.

[0033] In this embodiment, the step of extracting features from the relevant entropy image to obtain an initial feature map, and then performing feature enhancement on the initial feature map to obtain the target enhanced feature map, can be achieved through a multi-scale self-attention convolutional neural network (Multi... This is achieved using Scale Attention Convolutional Neural Network (MSA-CNN).

[0034] The MSA-CNN model extracts features through a lightweight network and enhances feature discrimination by combining a multi-scale self-attention mechanism, ultimately achieving accurate classification of the degree of winding deformation.

[0035] Specifically, refer to Figure 3 The MSA-CNN model can include a first convolutional layer, at least one bottleneck module, a second convolutional layer, a third convolutional layer, a global average pooling layer, and a fourth convolutional layer connected in sequence. The number of convolutional layers and their positions in the model are not limited to the specific example mentioned above; other possible implementations are also within the scope of protection. If there are multiple bottleneck modules, they must be sequentially connected. The first convolutional layer is a 5×5 convolutional layer, which can extract features from the relevant entropy image, capturing coarse-grained spatial features to obtain an initial feature map. The bottleneck module can perform feature enhancement processing on the initial feature map to obtain a target enhanced feature map. The second, third, global average pooling layers, and fourth convolutional layers can perform anomaly classification and identification on the target enhanced feature map to obtain the deformation degree identification result of the transformer winding.

[0036] S103: Perform feature enhancement processing on the initial feature map at least once to obtain the target enhanced feature map.

[0037] Each feature enhancement process includes: performing horizontal pooling, vertical pooling, and feature extraction on the feature map to be enhanced to obtain the target enhanced feature map.

[0038] In this embodiment, the bottleneck module of the MSA-CNN model can perform feature enhancement processing on the initial feature map. If there are multiple bottleneck modules, feature enhancement processing can be performed sequentially multiple times to obtain the final target enhanced feature map. In this embodiment, if there are multiple bottleneck modules, that is, if the feature enhancement processing is performed multiple times, for the first feature enhancement processing, the feature map to be enhanced is the initial feature map; for subsequent feature enhancement processing, the feature map to be enhanced is the feature map after the previous enhancement processing.

[0039] In one embodiment, the feature map to be enhanced is subjected to horizontal pooling, vertical pooling, and feature extraction to obtain the target enhanced feature map, including: The spatial dimension of the feature map to be enhanced is reduced from the current dimension to a preset dimension to obtain the dimensionality-reduced feature map; The first enhanced feature map is obtained by performing depthwise convolution and pointwise convolution on the dimensionality-reduced feature map. The first enhanced feature map is subjected to horizontal pooling, vertical pooling and feature extraction to obtain a local enhanced feature map. The local enhanced feature map and the first enhanced feature map are then weighted and fused to obtain the second enhanced feature map. The spatial dimension of the second enhanced feature map is restored from the preset dimension to the current dimension to obtain the third enhanced feature map; the third enhanced feature map is fused with the feature map to be enhanced to obtain the target enhanced feature map.

[0040] In this embodiment, each bottleneck module of the MSA-CNN model may include a dimensionality reduction convolutional layer, a depthwise separable convolutional layer, an MSA attention layer, and an up-dimensional convolutional layer connected in sequence.

[0041] The dimension reduction convolutional layer uses 1×1 convolution to reduce the number of channels and computational cost of the feature map to be enhanced, and to reduce the dimension of the feature map (from 64 dimensions to 16 dimensions), thereby improving computational efficiency.

[0042] Depthwise separable convolutional layers use 3×3 convolutions to separate spatial features (the first enhanced feature map) from the dimensionality-reduced feature map through depthwise and pointwise convolutions, thereby enhancing feature representation capabilities while reducing the number of parameters. Here, "number of parameters" refers to the total number of trainable parameters for the depthwise separable convolutional layer. By separating spatial and channel convolutions, depthwise separable convolution significantly reduces the number of parameters in the convolution kernel while maintaining equivalent feature extraction capabilities. This reduces the model's computational complexity and memory footprint, improving inference efficiency.

[0043] The MSA attention layer introduces a multi-scale spatial-channel joint attention mechanism, which can simultaneously capture key features at different scales and orientations. It enhances local effective information and suppresses redundant noise through horizontal and vertical pooling. Then, it weights and fuses the local enhanced features with the original enhanced features, making similar features more aggregated and dissimilar features more significantly different, thereby improving the discriminative power between features and the robustness of the model.

[0044] The dimension-upgrading convolutional layer uses 1×1 convolution to restore the spatial dimension of the second enhanced feature map from the preset dimension to the current dimension (from 16 dimensions to 64 dimensions), and restores the number of channels.

[0045] In this embodiment, residual connections are used between bottleneck modules to alleviate gradient vanishing and promote feature reuse and information flow. The residual connection between adjacent bottleneck modules involves adding the input feature map (the feature map to be enhanced) of the previous bottleneck module to the third enhanced feature map (the output feature map of that module, processed through dimensionality reduction, depthwise separable convolution, MSA attention, and dimensionality increase convolution). This forms a jump connection structure: input → module processing → output + input. This operation directly transfers shallow features to deeper layers, effectively alleviating the gradient vanishing problem in deep network training, while promoting the reuse and information flow of features at different levels, thus improving the model's feature representation ability and training stability.

[0046] This embodiment reduces the computational complexity and feature extraction difficulty of the feature map to be enhanced by first reducing dimensionality and then increasing it. Furthermore, feature enhancement through multiple bottleneck modules significantly improves the discriminative power between features and the robustness of the model.

[0047] S104: The deformation degree identification result of the transformer winding is obtained based on the target enhanced feature map.

[0048] In this embodiment, the deformation degree identification result of the transformer winding is obtained based on the target enhanced feature map, which can be achieved by the second convolutional layer, the third convolutional layer, the global average pooling layer, and the fourth convolutional layer of the MSA-CNN model.

[0049] The second convolutional layer is a 3×3 convolutional layer, which further integrates the features output by the bottleneck module to enhance feature representation. Specifically, it performs sliding window convolution calculations on the feature map output by the bottleneck module using a 3×3 convolutional kernel, weighting and aggregating multi-channel features in the local space of the feature map, integrating scattered local features into more globally relevant abstract features, and further enhancing feature representation while retaining key spatial structure information. The third convolutional layer is a 3×3 convolutional layer that extracts abstract features from the target enhancement feature map and compresses the size of the feature map (from 64×64 to 28×28). The global average pooling layer performs global average pooling on the spatial dimensions of the target enhancement feature map, compressing the 28×28 feature map to 1×1, reducing the number of parameters and aggregating global information from the feature map. The fourth convolutional layer is a fully connected convolutional layer that outputs the final feature representation and performs classification to obtain the classification result. That is, it can obtain the deformation degree identification result of the transformer winding.

[0050] As can be seen from the above, firstly, this implementation divides the frequency response signal of the transformer winding into multiple equal-length segments based on the logarithmic frequency segmentation method, and calculates the correlation entropy between every two segments to construct a correlation entropy matrix. Logarithmic frequency segmentation can better adapt to the characteristics of the frequency response signal in different frequency bands, while the correlation entropy matrix can effectively capture the similarity and differences between signal segments. This segmentation method can solve the problem that existing technologies cannot obtain subtle changes in high-frequency bands above 10MHz, enabling effective identification of the degree of transformer winding deformation, and capturing the frequency characteristic correlation between segments through the correlation entropy matrix.

[0051] Secondly, this embodiment also converts the correlation entropy matrix into a correlation entropy image with RGB channels, transforming complex matrix data into an intuitive image form, which facilitates feature extraction using deep learning models. The initial feature map obtained in this way can more comprehensively reflect the state information of the transformer windings, helping to improve recognition accuracy.

[0052] Finally, in the feature enhancement process, this embodiment performs pooling operations in both horizontal and vertical directions, enabling analysis and processing of the feature map from multiple dimensions to obtain more comprehensive feature information. This multi-dimensional processing method helps enhance feature expression, allowing the model to better capture the characteristic manifestations of winding deformation in different directions and improve its adaptability to various deformation conditions. Furthermore, through multiple feature enhancement processes, the model can gradually filter out noise and interference factors in the signal, extracting more representative and stable features. Even if the input frequency response signal contains some noise or fluctuations, a reliable target enhanced feature map can still be obtained after multi-level feature enhancement processing, thereby ensuring the robustness of the transformer winding deformation degree identification results and reducing the possibility of misjudgment.

[0053] In one embodiment of this application, horizontal pooling, vertical pooling, and feature extraction are performed on the first enhanced feature map to obtain a local enhanced feature map, including: The first enhanced feature map is divided into multiple feature groups according to the total number of channels, and the multiple feature groups correspond to different frequency bands; For each feature group, horizontal pooling, vertical pooling, and local feature extraction are performed to obtain the local enhanced feature map of each feature group; The local enhanced feature maps of each feature group are fused to obtain the local enhanced feature map.

[0054] In this embodiment, the MSA attention layer divides the input first enhanced feature map into g feature groups according to the total number of channels c, with each group containing c / g channels. In this embodiment, g is set to 4. These four feature groups correspond to different frequency bands and can be used to explore the semantic features of different frequency bands related to these feature groups. According to the power industry standard DL / T911-2016 "Frequency Response Analysis Method for Winding Deformation of Power Transformers", different feature groups correspond to different frequency bands. The signals of different frequency bands correspond to the deformation features of different positions and degrees of windings. The low-frequency band is mainly related to the overall macroscopic displacement of the winding, the mid-frequency band focuses on the local deformation between turns and between coils, and the high-frequency band reflects the microstructural defects between turns and strands. Moreover, the channel dimension of the correlation entropy image itself carries the frequency information of the frequency response signal. Processing by feature group and frequency band can avoid interference between features of different frequency bands and achieve differentiated feature enhancement.

[0055] The first enhanced feature map x Dividing the total number of channels c into g feature groups can be represented as follows: x =[ x 0 , x 1 ,…, x g 1], denoted as x ∈R (c / g)×h×w c / g represents the number of channels in each feature group after the total number of channels c in the first enhanced feature map is divided into g feature groups on average, which is also the number of channels in each feature group. For example, in the embodiment, c is set to 4, then c / g is the channel dimension corresponding to each feature group, ensuring the balanced splitting of channel dimensions during frequency band processing.

[0056] Where h represents the height of the first enhanced feature map and w represents the width of the first enhanced feature map.

[0057] The MSA attention layer employs three parallel paths to extract features from multiple feature groups. Each feature group requires three parallel paths to obtain its corresponding local enhanced feature map. The first and second paths use 1×1 convolutions, while the third path uses 3×3 convolutions. In the two 1×1 convolution paths, horizontal and vertical pooling are performed respectively, and then the outputs of the two paths are summed using pointwise multiplication. The third path uses 3×3 convolutions to directly extract local features from the feature groups.

[0058] In this embodiment, the steps of obtaining the local enhanced feature map of each feature group through horizontal pooling, vertical pooling, and local feature extraction employ a multi-scale spatial-channel joint attention mechanism, specifically: The multi-scale spatial-channel joint attention mechanism of the MSA attention layer includes a spatial attention component, a channel attention component, and a feature-weighted fusion component.

[0059] The spatial attention component uses multi-scale convolutional kernels to extract spatial features in different directions for each feature group; the channel attention component uses global average pooling and max pooling in parallel to generate channel attention weights after fusion, thereby enhancing the expression of channel features.

[0060] After multiplying the spatial features and channel attention weights, the result is a locally enhanced feature map obtained by passing the Sigmoid activation function.

[0061] In one embodiment, horizontal pooling, vertical pooling, and local feature extraction are performed on each feature group to obtain a local enhanced feature map for each feature group, including: Horizontal pooling is performed on each feature group to obtain the distribution characteristics of the relevant entropy values ​​in the frequency dimension; Vertical pooling is performed on each feature group to obtain the distribution characteristics of the relevant entropy values ​​in the magnitude dimension; The first tensor is obtained by multiplying the distribution characteristics of the correlation entropy values ​​in the frequency dimension and the distribution characteristics of the correlation entropy values ​​in the amplitude dimension. The second tensor is obtained by performing local feature extraction on the feature group; The local enhanced feature map for each feature group is obtained based on the first and second tensors.

[0062] In this embodiment, for each feature group: In the first path, dimensionality reduction is performed first using a 1×1 convolution, followed by one-dimensional horizontal pooling of the feature group to obtain the distribution characteristics of the relevant entropy values ​​in the frequency dimension. In the second path, similarly, dimensionality reduction is performed first using a 1×1 convolution, followed by one-dimensional vertical pooling of the feature group to obtain the distribution characteristics of the relevant entropy values ​​in the amplitude dimension. In the third path, local features (also known as joint local entropy features) are directly extracted using a 3×3 convolution.

[0063] For the first and second paths, the features obtained are multiplied together to obtain the first tensor; for the third path, the output is directly used as the second tensor. Based on the first and second tensors, a local enhanced feature map for each feature group is obtained. The feature multiplication in this embodiment is not a simple numerical multiplication, but a standard operation in feature fusion in deep learning, typically referring to element-wise multiplication or matrix multiplication (matrix dot product / outer product). The purpose is to allow features from two directions to "modulate" each other, generating more discriminative joint features.

[0064] In one embodiment, obtaining a local enhanced feature map for each feature group based on a first tensor and a second tensor includes: The global spatial information of the first tensor is encoded to obtain the first encoded result. The dimension of the second tensor is adjusted to be the same as that of the first tensor. The first encoded result and the second tensor after adjusting the dimension are multiplied by a matrix to obtain the first spatial attention map. The global spatial information of the second tensor is encoded to obtain the encoded second result. The dimension of the first tensor is adjusted to be the same as that of the second tensor. The encoded second result is then multiplied by the first tensor after the dimension adjustment to obtain the second spatial attention map. The local enhanced feature map for each feature group is determined based on the first spatial attention map and the second spatial attention map.

[0065] In this embodiment, the first spatial attention layer of the MSA attention layer is used to encode the global spatial information of the first tensor using global average pooling to obtain the encoded first result. The dimension of the second tensor is adjusted to be the same as the dimension of the first tensor output by the 1×1 convolution path. Then, matrix multiplication is performed on the encoded first result and the dimension-adjusted second tensor, followed by sigmoid activation to obtain the first spatial attention map. This first spatial attention map can highlight regions with high entropy changes.

[0066] The second spatial attention layer of the MSA attention layer encodes the global spatial information of the second tensor using global average pooling to obtain the encoded second result. It then adjusts the dimension of the first tensor to match the dimension of the second tensor output by the 3×3 convolution path. A matrix dot product is then performed on the encoded second result and the dimension-adjusted first tensor, followed by sigmoid activation to obtain the second spatial attention map. This second spatial attention map preserves pixel-level positional relationships.

[0067] Finally, the MSA attention layer enhances the global background correlation of pixels and highlights local features by fusing the high entropy value change region and the corresponding pixel-level position correlation.

[0068] In one embodiment, before obtaining the local enhanced feature map of each feature group based on the first spatial attention map and the second spatial attention map, the recognition method further includes: Global average pooling is performed on the distribution characteristics of the relevant entropy values ​​in the frequency dimension, the distribution characteristics of the relevant entropy values ​​in the amplitude dimension, and the second tensor to obtain the first path weight, the second path weight, and the third path weight; The first path weight, the second path weight, and the third path weight are concatenated into channels, and the concatenation result is input into the multilayer perceptron model to obtain the channel attention weights. Specifically, the local enhanced feature map for each feature group is obtained based on the first spatial attention map and the second spatial attention map, including: The second spatial attention map is multiplied element-wise with the channel attention weights, and then added element-wise with the first spatial attention map to obtain the sum. The spatial attention mask is obtained based on the sum. The distribution characteristics of the relevant entropy values ​​in the frequency dimension, the distribution characteristics of the relevant entropy values ​​in the amplitude dimension, and the second tensor are weighted and summed with their respective path weights. The weighted summation result is then multiplied element-wise with the spatial attention mask to obtain the local enhanced feature map of the feature group.

[0069] In this embodiment, the channel attention component enables the MSA attention layer to perform global average pooling on the distribution characteristics of the relevant entropy values ​​in the frequency dimension, the distribution characteristics of the relevant entropy values ​​in the amplitude dimension, and the second tensor to obtain the first path weights. v 1. Second path weight v 2 and third path weights v 3; Then, concatenate the weights of the first path, the second path, and the third path, and input them into the multi-incremental perceptron model. After sigmoid activation, the channel attention weights are obtained. : ; in, This represents the Sigmoid activation function. This represents a multilayer perceptron.

[0070] In this embodiment, the first spatial attention map focuses on key regions with high entropy changes in the global scope, while the second spatial attention map focuses on preserving pixel-level local positional interaction details. By adding elements one by one, the global focusing capability and the local fine correlation capability can be combined. After normalization by the Sigmoid function, a spatial attention mask that can be applied to the feature map is formed, which can realize dynamic weighting of spatial information, ultimately enhancing the global background correlation of all pixels and improving the robustness and discriminative power of feature representation.

[0071] In one embodiment of this application, obtaining the deformation degree identification result of the transformer winding based on the target enhanced feature map includes: The target augmentation feature map is obtained by convolution and reducing the number of channels; Perform global averaging and fully connected convolution operations on the spatial dimension of the target enhanced feature map after channel reduction to obtain the deformation degree identification result of the transformer winding. In this embodiment, reference Figure 3 The second and third convolutional layers of the MSA-CNN model further extract features, avoiding feature omissions, while compressing the target augmentation feature map to 28×28. The global average pooling layer performs global average pooling on the spatial dimensions of the compressed target augmentation feature map, compressing the 28×28 feature map to 1×1, reducing the number of parameters and facilitating the aggregation of global information. The fourth convolutional layer is a fully connected layer, enabling analysis of global information and anomaly classification to obtain the deformation degree identification result of the transformer winding.

[0072] Corresponding to the transformer winding deformation identification method in the above embodiment, Figure 4 This is a structural block diagram of a transformer winding deformation degree identification device provided in one embodiment of this application. For ease of explanation, only the parts relevant to the embodiment of this application are shown. References Figure 4 The transformer winding deformation identification device 20 includes: a first data processing unit 21, a second data processing unit 22, a third data processing unit 23, and an anomaly identification unit 24.

[0073] The first data processing unit 21 is used to divide the frequency response signal of the transformer winding into multiple segments of equal length based on the logarithmic frequency segmentation method, and to calculate the correlation entropy of each pair of segments to construct a correlation entropy matrix. The second data processing unit 22 is used to convert the correlation entropy matrix into a correlation entropy image with RGB channels, and to extract features from the correlation entropy image to obtain an initial feature map. The third data processing unit 23 is used to perform at least one feature enhancement process on the initial feature map to obtain the target enhanced feature map; Anomaly identification unit 24 is used to obtain the deformation degree identification result of the transformer winding based on the target enhanced feature map; Specifically, the third data processing unit 23 performs horizontal pooling, vertical pooling, and feature extraction on the feature map to be enhanced during each feature enhancement process to obtain the target enhanced feature map. If the feature enhancement process is performed multiple times, the feature map to be enhanced is the initial feature map for the first feature enhancement process; for subsequent feature enhancement processes, the feature map to be enhanced is the feature map after the previous enhancement process.

[0074] In one embodiment of this application, the third data processing unit 23 is specifically used to: reduce the spatial dimension of the feature map to be enhanced from the current dimension to a preset dimension to obtain a dimension-reduced feature map; The first enhanced feature map is obtained by performing depthwise convolution and pointwise convolution on the dimensionality-reduced feature map. The first enhanced feature map is subjected to horizontal pooling, vertical pooling and feature extraction to obtain a local enhanced feature map. The local enhanced feature map and the first enhanced feature map are then weighted and fused to obtain the second enhanced feature map. The spatial dimension of the second enhanced feature map is restored from the preset dimension to the current dimension to obtain the third enhanced feature map; the third enhanced feature map is fused with the feature map to be enhanced to obtain the target enhanced feature map.

[0075] In one embodiment of this application, the third data processing unit 23 is specifically used to: divide the first enhanced feature map into multiple feature groups according to the total number of channels, with the multiple feature groups corresponding to different frequency bands; For each feature group, horizontal pooling, vertical pooling, and local feature extraction are performed to obtain the local enhanced feature map of each feature group; The local enhanced feature maps of each feature group are fused to obtain the local enhanced feature map.

[0076] In one embodiment of this application, the third data processing unit 23 is specifically used to: perform horizontal pooling on each feature group to obtain the distribution characteristics of the relevant entropy value in the frequency dimension; Vertical pooling is performed on each feature group to obtain the distribution characteristics of the relevant entropy values ​​in the magnitude dimension; The first tensor is obtained by multiplying the distribution characteristics of the correlation entropy values ​​in the frequency dimension and the distribution characteristics of the correlation entropy values ​​in the amplitude dimension. The second tensor is obtained by performing local feature extraction on the feature group; The local enhanced feature map for each feature group is obtained based on the first and second tensors.

[0077] In one embodiment of this application, the third data processing unit 23 is specifically used for: The global spatial information of the first tensor is encoded to obtain the first encoded result. The dimension of the second tensor is adjusted to be the same as that of the first tensor. The first encoded result and the second tensor after adjusting the dimension are multiplied by a matrix to obtain the first spatial attention map. The global spatial information of the second tensor is encoded to obtain the encoded second result. The dimension of the first tensor is adjusted to be the same as that of the second tensor. The encoded second result is then multiplied by the first tensor after the dimension adjustment to obtain the second spatial attention map. The local enhanced feature map for each feature group is determined based on the first spatial attention map and the second spatial attention map.

[0078] In one embodiment of this application, the transformer winding deformation degree identification device 20 further includes a weight calculation unit, which is used to perform global average pooling on the distribution characteristics of the relevant entropy value in the frequency dimension, the distribution characteristics of the relevant entropy value in the amplitude dimension, and the second tensor to obtain the first path weight, the second path weight, and the third path weight. The first path weight, the second path weight, and the third path weight are concatenated into channels, and the concatenation result is input into the multilayer perceptron model to obtain the channel attention weights. The third data processing unit 23 is specifically used for: The second spatial attention map is multiplied element-wise with the channel attention weights, and then added element-wise with the first spatial attention map to obtain the sum. The spatial attention mask is obtained based on the sum. The distribution characteristics of the relevant entropy values ​​in the frequency dimension, the distribution characteristics of the relevant entropy values ​​in the amplitude dimension, and the second tensor are weighted and summed with their respective path weights. The weighted summation result is then multiplied element-wise with the spatial attention mask to obtain the local enhanced feature map of the feature group.

[0079] In one embodiment of this application, the anomaly identification unit 24 is specifically used for: The target augmentation feature map is obtained by convolution and reducing the number of channels; Global averaging and fully connected convolution operations are performed on the spatial dimension of the target enhanced feature map after channel reduction to obtain the deformation degree identification result of the transformer winding.

[0080] In summary, the transformer winding deformation degree identification method and device provided in this application have at least the following beneficial effects: (1) Robust feature quantization capability of the correlation entropy matrix This embodiment transforms the segmented differences of the frequency response signal into an absolute difference matrix of mutual information entropy through a correlation entropy matrix construction method, overcoming the limitations of traditional single-amplitude comparison of frequency curves. Correlation entropy, as a second-order similarity measure in kernel space, is inherently robust to non-Gaussian noise. Its segmented difference analysis mechanism effectively suppresses interference from single noise points. Compared to traditional difference quantification methods, the correlation entropy matrix achieves a nonlinear characterization of signal distribution differences through joint probability density calculation, enabling it to capture subtle deformation features ignored by existing technologies. It is particularly suitable for detecting early minor deformations in the windings of auxiliary transformers in high-speed trains.

[0081] (2) Enhanced visualization features of related entropy image transformation The process of mapping the relevance entropy matrix to a grayscale image is essentially transforming an abstract frequency domain signal into visual features that can be efficiently processed by deep learning models. This transformation strategy has a dual advantage: Multi-dimensional feature decoupling: Through a logarithmic segmentation strategy, the gray values ​​of the relevant entropy matrix simultaneously encode the intensity of frequency band differences and segment correlations, forming a two-dimensional feature distribution similar to a spectrogram, which facilitates the MSA-CNN model to simultaneously capture macroscopic morphology and microscopic differences.

[0082] Improved anti-interference capability: The generation process of grayscale images naturally suppresses high-frequency noise during signal acquisition. At the same time, preprocessing techniques such as histogram equalization enhance feature contrast, reducing the model's sensitivity to external factors such as changes in illumination and differences in equipment.

[0083] (3) Multi-scale features of the MSA-CNN model Multi-path feature extraction: The MSA attention layer captures the global correlation and local details of frequency bands through parallel paths of 1×1 horizontal pooling, vertical pooling, and 3×3 convolution.

[0084] Space-channel joint weighting: The dual spatial attention mechanism can dynamically focus on deformation-sensitive areas, thereby suppressing interference signals in non-critical frequency bands. The measured positioning accuracy of local deformation is improved to the pixel level.

[0085] Lightweight network architecture: By using depthwise separable convolution and residual connection design in the bottleneck module, end-to-end processing from signal acquisition to deformation level output is achieved while maintaining the parameter magnitude, which can meet the real-time requirements of EMU online monitoring.

[0086] See Figure 5 , Figure 5 This is a schematic block diagram of an electronic device provided according to an embodiment of this application. Figure 5The electronic device 300 in this embodiment may include one or more processors 301, one or more input devices 302, one or more output devices 303, and one or more memories 304. The processors 301, input devices 302, output devices 303, and memories 304 communicate with each other via a communication bus 305. The memories 304 store computer programs, including program instructions. The processors 301 execute the program instructions stored in the memories 304. Specifically, the processors 301 are configured to invoke the program instructions to perform the functions of the units in the above-described device embodiments, for example... Figure 4 The functions of the first data processing unit 21, the second data processing unit 22, the third data processing unit 23, and the anomaly recognition unit 24 are shown.

[0087] It should be understood that, in the embodiments of this application, the processor 301 may be a central processing unit (CPU), but it may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0088] Input device 302 may include a touchpad, a fingerprint sensor (for collecting the user's fingerprint information and fingerprint orientation information), a microphone, etc., and output device 303 may include a display (LCD, etc.), a speaker, etc.

[0089] The memory 304 may include read-only memory and random access memory, and provides instructions and data to the processor 301. A portion of the memory 304 may also include non-volatile random access memory. For example, the memory 304 may also store device type information.

[0090] In specific implementations, the processor 301, input device 302, and output device 303 described in the embodiments of this application can execute the implementation method described in the transformer winding deformation degree identification method provided in the embodiments of this application, or they can execute the implementation method of the electronic device described in the embodiments of this application, which will not be repeated here.

[0091] In another embodiment of this application, a computer-readable storage medium is provided. This computer-readable storage medium stores a computer program, which includes program instructions. When executed by a processor, the program instructions implement all or part of the processes in the methods described above. Alternatively, the computer program can instruct related hardware to complete the process. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0092] The computer-readable storage medium can be an internal storage unit of the electronic device in any of the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., provided on the electronic device. Furthermore, the computer-readable storage medium can include both internal and external storage units of the electronic device. The computer-readable storage medium is used to store computer programs and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0093] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0094] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electronic devices and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0095] In the several embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces or units, or it may be an electrical, mechanical, or other form of connection.

[0096] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0097] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0098] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for identifying the degree of deformation of transformer windings, characterized in that, include: The frequency response signal of the transformer winding is divided into multiple segments of equal length based on the logarithmic frequency segmentation method, and the correlation entropy of every two segments is calculated to construct a correlation entropy matrix. The correlation entropy matrix is ​​converted into a correlation entropy image with RGB channels, and features are extracted from the correlation entropy image to obtain an initial feature map; The initial feature map is subjected to at least one feature enhancement process to obtain the target enhanced feature map; The deformation degree identification result of the transformer winding is obtained based on the target enhanced feature map; Each feature enhancement process includes: The feature map to be enhanced is subjected to horizontal pooling, vertical pooling, and feature extraction to obtain the target enhanced feature map. Wherein, if the feature enhancement process is performed multiple times, for the first feature enhancement process, the feature map to be enhanced is the initial feature map; for subsequent feature enhancement processes, the feature map to be enhanced is the feature map after the previous enhancement process.

2. The method for identifying the degree of deformation of transformer windings as described in claim 1, characterized in that, The feature map to be enhanced is subjected to horizontal pooling, vertical pooling, and feature extraction to obtain the target enhanced feature map, including: The spatial dimension of the feature map to be enhanced is reduced from the current dimension to a preset dimension to obtain the dimensionality-reduced feature map; The reduced feature map is then subjected to depthwise convolution and pointwise convolution to obtain the first enhanced feature map. The first enhanced feature map is subjected to horizontal pooling, vertical pooling and feature extraction to obtain a local enhanced feature map. The local enhanced feature map and the first enhanced feature map are then weighted and fused to obtain a second enhanced feature map. The spatial dimension of the second enhanced feature map is restored from the preset dimension to the current dimension to obtain the third enhanced feature map; the third enhanced feature map is fused with the feature map to be enhanced to obtain the target enhanced feature map.

3. The method for identifying the degree of deformation of transformer windings as described in claim 2, characterized in that, The step of performing horizontal pooling, vertical pooling, and feature extraction on the first enhanced feature map to obtain a local enhanced feature map includes: The first enhanced feature map is divided into multiple feature groups according to the total number of channels, and the multiple feature groups correspond to different frequency bands; For each feature group, horizontal pooling, vertical pooling, and local feature extraction are performed to obtain the local enhanced feature map of each feature group; The local enhanced feature map is obtained by fusing the local enhanced feature maps of each feature group.

4. The method for identifying the degree of deformation of transformer windings as described in claim 3, characterized in that, The process of performing horizontal pooling, vertical pooling, and local feature extraction on each feature group to obtain a local enhanced feature map for each feature group includes: Horizontal pooling is performed on each feature group to obtain the distribution characteristics of the relevant entropy values ​​in the frequency dimension; Vertical pooling is performed on each feature group to obtain the distribution characteristics of the relevant entropy value in the amplitude dimension; The first tensor is obtained by multiplying the distribution characteristics of the relevant entropy values ​​in the frequency dimension and the distribution characteristics of the relevant entropy values ​​in the amplitude dimension. The second tensor is obtained by performing local feature extraction on the feature group; The local enhanced feature map of each feature group is obtained based on the first tensor and the second tensor.

5. The method for identifying the degree of deformation of transformer windings as described in claim 4, characterized in that, The process of obtaining the local enhanced feature map for each feature group based on the first tensor and the second tensor includes: The global spatial information of the first tensor is encoded to obtain the first encoded result. The dimension of the second tensor is adjusted to be the same as that of the first tensor. The first encoded result and the second tensor after dimension adjustment are multiplied by matrix dot to obtain the first spatial attention map. The global spatial information of the second tensor is encoded to obtain the encoded second result. The dimension of the first tensor is adjusted to be the same as the dimension of the second tensor. The encoded second result is then multiplied by the first tensor after the dimension adjustment to obtain the second spatial attention map. The local enhanced feature map for each feature group is determined based on the first spatial attention map and the second spatial attention map.

6. The method for identifying the degree of deformation of transformer windings as described in claim 5, characterized in that, Before obtaining the local enhanced feature map for each feature group based on the first spatial attention map and the second spatial attention map, the method further includes: Global average pooling is performed on the distribution characteristics of the relevant entropy value in the frequency dimension, the distribution characteristics of the relevant entropy value in the amplitude dimension, and the second tensor to obtain the first path weight, the second path weight, and the third path weight; The first path weight, the second path weight, and the third path weight are concatenated into channels, and the concatenation result is input into the multilayer perceptron model to obtain the channel attention weights. The step of obtaining the local enhanced feature map for each feature group based on the first spatial attention map and the second spatial attention map includes: The second spatial attention map is multiplied element-wise with the channel attention weights, and then added element-wise with the first spatial attention map to obtain the sum result. The spatial attention mask is obtained based on the sum result. The distribution features of the relevant entropy values ​​in the frequency dimension, the distribution features of the relevant entropy values ​​in the amplitude dimension, and the second tensor are weighted and summed with their respective path weights. The weighted summation result is then multiplied element-wise with the spatial attention mask to obtain the local enhanced feature map of the feature group.

7. The method for identifying the degree of deformation of transformer windings as described in claim 1, characterized in that, The determination of the deformation degree of the transformer winding based on the target enhanced feature map includes: The target enhancement feature map is convolved and the number of channels is reduced to obtain a target enhancement feature map with reduced channels; A global averaging operation and a fully connected convolution operation are performed on the spatial dimension of the target enhanced feature map after channel reduction to obtain the deformation degree identification result of the transformer winding.

8. A device for identifying the degree of deformation of transformer windings, characterized in that, include: The first data processing unit is used to divide the frequency response signal of the transformer winding into multiple segments of equal length based on the logarithmic frequency segmentation method, and to calculate the correlation entropy of every two segments to construct a correlation entropy matrix. The second data processing unit is used to convert the correlation entropy matrix into a correlation entropy image with RGB channels, and to perform feature extraction on the correlation entropy image to obtain an initial feature map. The third data processing unit is used to perform at least one feature enhancement process on the initial feature map to obtain the target enhanced feature map; An anomaly identification unit is used to obtain the deformation degree identification result of the transformer winding based on the target enhanced feature map; The third data processing unit is specifically used to perform horizontal pooling, vertical pooling, and feature extraction on the feature map to be enhanced, so as to obtain the target enhanced feature map. Wherein, if the feature enhancement process is performed multiple times, for the first feature enhancement process, the feature map to be enhanced is the initial feature map; for subsequent feature enhancement processes, the feature map to be enhanced is the feature map after the previous enhancement process.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, 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 to 7.

Images