A transformer insulation state evaluation method based on multi-scale spectral feature fusion

Abstract

The application discloses a transformer insulation state evaluation method based on multiscale spectral feature fusion, and belongs to the field of image processing. The method comprises the following steps: obtaining historical operation data and original spectral images of a transformer; performing multiscale decomposition on the original spectral images to obtain macroscopic overall spectral features and microscopic local spectral features; performing fusion processing on the spectral features of different scales to generate a fusion feature map; extracting a fusion feature vector reflecting an insulation state; and comparing the fusion feature vector with an evaluation standard feature library to obtain a transformer insulation state evaluation result. The application adopts hyperspectral imaging technology to obtain data, performs multiscale decomposition by using an adaptive Gaussian kernel and two-dimensional discrete wavelet transform, and realizes feature fusion by using deformable convolution and mutual information weighting, so that multi-dimensional information of an insulation medium can be effectively captured. The comparison with a support vector machine and historical operation data can significantly improve the accuracy and reliability of transformer insulation state evaluation.

Description

Technical Field

[0001] This invention relates to the field of image processing, and in particular to a method for evaluating the insulation status of transformers by multi-scale spectral feature fusion. Background Technology

[0002] As a core component of the power system, the health assessment of the insulation condition of power transformers is crucial. Currently, using spectral analysis technology to detect the insulation medium of transformers has become an effective method. Acquiring spectral images of the insulation medium through techniques such as hyperspectral imaging, and analyzing these images, can provide rich data support for transformer fault diagnosis. These technologies have been widely applied in image processing and pattern recognition, aiming to extract valuable information from complex image data.

[0003] However, existing methods for assessing insulation status based on spectral images still have some technical limitations. On the one hand, feature extraction at a single scale is insufficient to fully capture the complex information contained in the image, resulting in limited accuracy of the assessment results. On the other hand, traditional feature fusion methods typically employ simple overlay or stitching, failing to fully utilize the complementary information between features at different scales and lacking adaptive consideration of feature importance. This means that the robustness and accuracy of the assessment methods need improvement when facing varied and complex operating conditions. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a method for assessing transformer insulation status through multi-scale spectral feature fusion. It employs hyperspectral imaging technology to acquire multi-dimensional spectral information of the transformer insulation medium, and utilizes adaptive Gaussian kernel and two-dimensional discrete wavelet transform to perform multi-scale decomposition and feature fusion of the image. This method comprehensively captures the macroscopic and microscopic features of the insulation medium, significantly improving the accuracy and reliability of transformer insulation status assessment.

[0005] The above objectives can be achieved through the following approach: A method for assessing the insulation status of a transformer by multi-scale spectral feature fusion includes: acquiring original spectral images of the transformer's historical operating data and spectral information of the insulating medium in several bands; performing multi-scale decomposition on the original spectral images to obtain spectral features at different scales, wherein the spectral features at different scales include macroscopic overall spectral features and microscopic local spectral features; fusing the spectral features at different scales to generate a fused feature map; analyzing the fused feature map to extract a fused feature vector reflecting the insulation status of the transformer; and comparing the fused feature vector with a preset assessment standard feature library to obtain the insulation status assessment result of the transformer.

[0006] Optionally, the acquisition of the original spectral image of the transformer's historical operating data and the spectral information of several bands of the insulating medium includes: collecting the transformer's historical operating data, wherein the historical operating data includes operating years, load conditions, and environmental parameter data; scanning the transformer's insulating medium using hyperspectral imaging technology to acquire a hyperspectral data cube covering the visible to near-infrared bands; performing dimensionality reduction processing on the hyperspectral data cube, selecting sensitive bands of the insulation state according to the characteristic differences of the insulating medium, and obtaining several bands of spectral information; and reconstructing the image from the several bands of spectral information to generate the original spectral image.

[0007] Optionally, multi-scale decomposition of the original spectral image includes: performing spatial domain multi-scale decomposition using an adaptive Gaussian kernel based on the original spectral image, and obtaining macroscopic overall spectral features by dynamically adjusting the standard deviation of the Gaussian kernel; and performing frequency domain multi-scale decomposition using a two-dimensional discrete wavelet transform-based method based on the original spectral image, and obtaining microscopic local spectral features by automatically selecting the number of decomposition layers by combining the energy entropy criterion.

[0008] Optionally, the fusion processing of the spectral features at different scales includes: performing multi-band joint convolution operations on the macroscopic overall spectral features and the microscopic local spectral features respectively, extracting spatial distribution features while maintaining the correlation between each band, and obtaining the processed features; using deformable convolution to enhance the processed features; calculating weight coefficients based on the mutual information of each feature, and using the weight coefficients to weight the macroscopic overall spectral features and the microscopic local spectral features to obtain a fused feature map.

[0009] Optionally, the extraction of the fused feature vector reflecting the insulation state of the transformer includes: reducing the dimensionality of the fused feature map to generate a preliminary feature vector; and based on the preliminary feature vector, performing a nonlinear transformation using a fully connected neural network to obtain the fused feature vector reflecting the insulation state of the transformer.

[0010] Optionally, comparing the fused feature vector with a preset evaluation standard feature library includes: calculating the cosine similarity between the fused feature vector and the preset evaluation standard feature library to obtain a preliminary classification result; based on the preliminary classification result, using a support vector machine classifier to perform classification to obtain a confidence score; and obtaining the transformer insulation status evaluation result based on the confidence score and historical operating data.

[0011] Optionally, the feature enhancement using deformable convolution includes: based on the processed features, using an offset learning network to predict the offset of the deformable convolution kernel; and performing a convolution operation between the offset of the deformable convolution kernel and the processed features to obtain the enhanced features.

[0012] Optionally, dimensionality reduction of the fused feature map includes: compressing the spatial dimension of the fused feature map using a global average pooling algorithm to obtain a feature matrix; and flattening the feature matrix to obtain a preliminary feature vector.

[0013] Optionally, the classification using a support vector machine classifier includes: applying a radial basis function to the preliminary classification result nonlinearly to obtain feature vectors in a high-dimensional feature space; and calculating a confidence score based on the distance between the feature vectors in the high-dimensional feature space and the feature vectors in the high-dimensional feature space.

[0014] Based on the same inventive concept, this invention also provides a transformer insulation condition assessment system based on multi-scale spectral feature fusion. The system includes: a data acquisition module for acquiring historical operating data of the transformer and original spectral images of several bands of spectral information of the insulating medium; a multi-scale decomposition module for performing multi-scale decomposition on the original spectral images to obtain spectral features at different scales; a feature fusion module for fusing the spectral features at different scales to generate a fused feature map; a feature analysis and extraction module for analyzing the fused feature map and extracting a fused feature vector reflecting the transformer insulation condition; and an assessment module for comparing the fused feature vector with a preset assessment standard feature library to obtain the transformer insulation condition assessment result.

[0015] Compared with the prior art, the present invention has the following advantages: 1. This invention employs a dual-channel heterogeneous multi-scale decomposition method, combining an adaptive Gaussian kernel and two-dimensional discrete wavelet transform, to extract macroscopic and microscopic features from the spatial and frequency domains of images, respectively. This method overcomes the limitations of single-scale analysis, achieving comprehensive capture of multi-dimensional information in spectral images, providing a solid data foundation for subsequent accurate diagnosis, and significantly improving the richness and effectiveness of feature representation.

[0016] 2. This invention proposes an innovative feature fusion strategy that adaptively fuses features at different scales using deformable convolution and mutual information weighting mechanisms. This method can dynamically adjust the fusion focus, exhibiting higher flexibility and accuracy, especially when processing anomalous regions in images. This intelligent fusion approach effectively solves the problems of information redundancy and loss of key information in traditional methods, greatly enhancing the robustness and reliability of the evaluation results.

[0017] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a method for evaluating the insulation status of a transformer by fusing multi-scale spectral features according to an embodiment of the present invention.

[0020] Figure 2 This is a spectral curve of the insulating medium according to an embodiment of the present invention.

[0021] Figure 3 This is a multi-scale feature mutual information heatmap according to an embodiment of the present invention.

[0022] Figure 4 This is a cosine similarity bar chart according to an embodiment of the present invention.

[0023] Figure 5 This is a schematic diagram of the structure of a transformer insulation status assessment system based on multi-scale spectral feature fusion according to an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Reference Figure 1 One embodiment of the present invention proposes a method for assessing the insulation condition of transformers through multi-scale spectral feature fusion. This method employs hyperspectral imaging technology to acquire multi-dimensional spectral information of the transformer insulation medium, and utilizes an adaptive Gaussian kernel and two-dimensional discrete wavelet transform to perform multi-scale decomposition and feature fusion of the image. This method can comprehensively capture the macroscopic and microscopic features of the insulation medium, significantly improving the accuracy and reliability of transformer insulation condition assessment.

[0026] The method described in this embodiment specifically includes: Acquire raw spectral images of the transformer's historical operating data and the spectral information of the insulating medium in several bands; Specifically, raw hyperspectral image data of the transformer's insulation medium is acquired using optical imaging equipment; this is essentially a data cube containing both spatial and spectral dimensions. Simultaneously, historical operational data such as the transformer's years of operation, load conditions, and ambient temperature and humidity are collected from its operational records. These data collectively serve as inputs to this method, comprehensively describing the transformer's current state and historical context.

[0027] The original spectral image is decomposed into multiple scales to obtain spectral features at different scales, including macroscopic overall spectral features and microscopic local spectral features. Specifically, the acquired raw spectral image is decomposed. By employing different image processing techniques, the image is decomposed into macroscopic overall spectral features containing large-scale, low-frequency information and microscopic local spectral features containing small-scale, high-frequency information. Macroscopic features reflect the overall condition of the insulating medium, while microscopic features reveal subtle local anomalies or defects.

[0028] The spectral features at different scales are fused to generate a fused feature map; Specifically, the obtained macroscopic overall spectral features and microscopic local spectral features are fused. This fusion process aims to effectively integrate information at different scales, retaining key information while eliminating redundancy and enhancing the expressive power of the features, ultimately generating a fused feature map containing multi-scale information.

[0029] The fused feature map is analyzed to extract the fused feature vector reflecting the insulation state of the transformer; Specifically, the generated fused feature map is analyzed. By applying dimensionality reduction and mapping algorithms to process the fused feature map, a feature vector that can quantify and reflect the transformer's insulation state is extracted. This vector will serve as the unique numerical representation of the transformer's insulation state.

[0030] The fused feature vector is compared with a preset evaluation standard feature library to obtain the insulation status evaluation result of the transformer.

[0031] Specifically, the obtained fused feature vectors are input into a pre-established evaluation model. This model uses a pre-defined evaluation standard feature library as a reference, compares the input feature vectors with the standard vectors in the library, calculates similarity or performs classification, and finally outputs the insulation status evaluation result of the transformer.

[0032] By employing a series of image processing techniques, such as multi-scale decomposition of hyperspectral images, heterogeneous feature fusion, and feature vector extraction, the macroscopic and microscopic features of transformer insulation media can be captured comprehensively and accurately, significantly improving the accuracy and reliability of transformer insulation condition assessment.

[0033] Optionally, the original spectral image used to acquire historical operating data of the transformer and spectral information of the insulating medium in several bands includes: Collect historical operating data of the transformer, including operating years, load conditions, and environmental parameter data; Specifically, data such as equipment operating time, real-time load current, and ambient temperature and humidity are collected from transformer monitoring equipment or databases via data interfaces. This data is timestamped and stored as non-image data input reflecting the individual operating background of the transformer.

[0034] Based on the transformer insulation medium, a hyperspectral imaging technique is used to scan and acquire a hyperspectral data cube covering the visible to near-infrared bands. Specifically, a hyperspectral imager is used to perform non-contact scanning of transformer oil samples or solid insulation materials to capture continuous spectral information within a certain wavelength range. This information is organized in the form of a three-dimensional data cube, where two dimensions represent spatial location and the third dimension represents spectral wavelength.

[0035] For example, by using hyperspectral imaging, continuous spectral information in the 400 nm to 1000 nm band range is obtained, resulting in a spectral data cube with a size of 256×256 pixels containing 120 continuous bands, which comprehensively records the reflection or transmission characteristics of the insulating medium at different wavelengths.

[0036] The hyperspectral data cube is dimensionality reduced, and sensitive bands of the insulation state are selected based on the characteristic differences of the insulating medium to obtain several band spectral information. Specifically, to reduce data redundancy and highlight key information, this method performs dimensionality reduction on the hyperspectral data cube. This is achieved by applying algorithms such as principal component analysis or minimum noise separation to automatically identify and select the spectral bands that differ most significantly under different insulation conditions. These selected bands are the sensitive bands.

[0037] like Figure 2 As shown, the differences in spectral characteristics of the insulating medium under normal and aged conditions are intuitively demonstrated.

[0038] For example, after processing the hyperspectral data cube, it was found that the spectral characteristics of normal oil samples and aged oil samples differed most significantly in the three bands of 750 nm, 820 nm and 940 nm. Therefore, these three bands were selected as sensitive bands.

[0039] The spectral information of the aforementioned bands is used to reconstruct the image, generating the original spectral image.

[0040] Specifically, the obtained spectral information in the sensitive bands is reconstructed. By mapping this discrete spectral information onto a two-dimensional image space, one or more original spectral images containing key spectral information are generated.

[0041] For example, the spectral information of the three selected sensitive bands is superimposed as the three RGB color channels to generate a color raw spectral image for image processing, which intuitively reflects the characteristics of the insulating medium.

[0042] Optionally, performing multi-scale decomposition on the original spectral image includes: Based on the original spectral image, an adaptive Gaussian kernel is used for spatial domain multi-scale decomposition. By dynamically adjusting the standard deviation of the Gaussian kernel, the macroscopic overall spectral features are obtained. Specifically, performing Gaussian filtering on the original spectral image is essentially a kernel-based image smoothing process. The standard deviation of the Gaussian kernel controls the filtering scale; a larger standard deviation results in a wider smoothing range and more macroscopic features. To achieve adaptive decomposition, the standard deviation of the Gaussian kernel is dynamically adjusted by analyzing the local texture or gradient information of the image. (Gaussian kernel function) The definition is as follows: , in, and Represents pixel coordinates. It is the base of the natural logarithm. It is the standard deviation of the Gaussian kernel, and its value is based on the local variance of the image. Dynamic adjustment, such as: , It is a preset scaling factor. This method of dynamically adjusting the Gaussian kernel parameter can automatically adapt the filtering scale according to the image content, thereby extracting the macroscopic features that best represent the overall condition of the insulating medium in the spatial domain.

[0043] For example, suppose that in the spectral image being processed, the local variance of a region... The calculated value is 10, and the preset scaling factor is... When the value is 0.5, the system uses... The standard deviation of the Gaussian kernel is dynamically adjusted to smooth the texture of the region and extract macroscopic features.

[0044] Based on the original spectral image, a frequency domain multi-scale decomposition method based on two-dimensional discrete wavelet transform is used. By combining the energy entropy criterion, the number of decomposition layers is automatically selected to obtain microscopic local spectral features.

[0045] Specifically, a two-dimensional discrete wavelet transform is performed on the original spectral image. Wavelet transform decomposes the image into sub-images of different frequency bands, including a low-frequency sub-image and multiple high-frequency sub-images. The high-frequency sub-images contain microscopic details such as image edges and textures, and can serve as microscopic local spectral features. To determine the optimal number of decomposition levels, the energy entropy criterion is used. Energy entropy is a measure of the complexity of image information; when decomposing to a certain level, if the rate of change of energy entropy falls below a certain threshold, further decomposition is considered to be of little benefit, and decomposition automatically stops. (Image energy entropy) The definition is as follows: , in, This is the normalized energy distribution of the wavelet coefficients. When the rate of change of energy entropy... Stop decomposition at that time. It is the first The energy entropy of the layer, The preset threshold is typically set to 1%. This criterion ensures that sufficient microscopic details are extracted while avoiding redundancy caused by over-decomposition.

[0046] For example, when performing wavelet decomposition on an original spectral image, the energy entropy of the second layer is calculated to be 1200, and the energy entropy of the third layer is 1205. If a preset threshold is set... It is 1%, because Therefore, the decomposition is automatically stopped, and the high-frequency detail components of the third layer are extracted as microscopic local spectral features.

[0047] Optionally, the fusion processing of the spectral features at different scales includes: Multi-band joint convolution operations are performed on both macroscopic overall spectral features and microscopic local spectral features to extract spatial distribution features while maintaining the correlation between bands, resulting in processed features. Specifically, multi-band joint convolution is applied to both macroscopic overall spectral features and microscopic local spectral features. This convolution operation is particularly effective when processing hyperspectral images, as it performs sliding window operations simultaneously in two spatial dimensions and one spectral dimension using a three-dimensional convolution kernel. This process can capture the correlation between different bands at the same spatial location and extract the spatial distribution patterns of different spectral features. For example, for an input feature map F (with dimensions of... ,in For spatial dimensions, (where F is the channel dimension) is a pixel value in the output feature map F of the multi-band joint convolution. Calculated using the following formula: , in, It is a three-dimensional convolution kernel. This represents the offset of the convolution kernel. In this way, processed features representing macroscopic and microscopic features can be obtained respectively.

[0048] For example, for a spectral feature map containing 120 bands, a three-dimensional convolution kernel is used to perform multi-band joint convolution, where one pixel value By the pixel and its surrounding neighboring pixel values With convolution kernel We obtain the result by weighted summation.

[0049] Deformable convolution is used to enhance the processed features; Specifically, deformable convolution is applied to the processed features. Unlike traditional convolution kernels with fixed sampling positions, the sampling positions of deformable convolution are learnable; their offsets are predicted by an independent network branch based on the gradient distribution of the current feature map. This allows the convolution kernel to adaptively focus on key regions in the image related to insulation states, especially irregularly shaped defects or anomalous regions, thereby effectively enhancing the expressive power of the features. The output position of deformable convolution... eigenvalues Calculated using the following formula: , in, The sampling grid is for traditional convolution. The position in the grid. It's weight. It is a learnable offset.

[0050] For example, when an irregularly shaped aging region of insulating medium appears in the spectral image, the deformable convolutional sampling grid will adaptively shift, the shift amount being... The network predicts the data, thus accurately covering the irregular area and extracting more precise local defect features.

[0051] The weighting coefficients are calculated based on the mutual information of each feature, and the macroscopic overall spectral features and microscopic local spectral features are weighted using the weighting coefficients to obtain a fused feature map.

[0052] Specifically, mutual information is used to measure the degree of information dependence between macroscopic overall spectral features and microscopic local spectral features. Higher mutual information indicates more shared information between the two features; conversely, lower mutual information indicates more complementary information. Based on this relationship, weighting coefficients are calculated to weight the two features. Higher weighting coefficients indicate that the feature contains more unique and valuable information, and vice versa. Mutual Information Calculated using the following formula: , in, It is a feature and joint probability distribution and These are their marginal probability distributions. Based on the magnitude of mutual information, a normalized weighting coefficient can be determined. Then the feature maps are fused. Calculated using the following formula: , in, and It consists of macroscopic and microscopic characteristics. and These are the corresponding weighting coefficients.

[0053] like Figure 3 As shown, this demonstrates the information correlation and complementarity between features at different scales.

[0054] For example, the mutual information between macroscopic global spectral features and microscopic local spectral features is calculated. It was determined that microscopic features contain more unique information, and therefore were assigned to... Weighting coefficients It will be higher than Then, the final fused feature map is obtained through a weighted formula. .

[0055] Optionally, the extraction of the fused feature vector reflecting the insulation state of the transformer includes: The fused feature map is dimensionality reduced to generate a preliminary feature vector; Specifically, dimensionality reduction is performed on the fused feature map to reduce the amount of data and extract core information. This step uses techniques such as global average pooling to convert the two-dimensional fused feature map into a one-dimensional vector. Global average pooling calculates the average value of each channel in the feature map to obtain a single value, ultimately compressing the entire feature map into a vector that reflects its overall features.

[0056] For example, a fused feature map of size 7×7×128 can be converted into a one-dimensional preliminary feature vector of length 128 after global average pooling, where each element of the vector represents the average activation intensity of the corresponding channel.

[0057] Based on the preliminary feature vectors, a fully connected neural network is used for nonlinear transformation to obtain a fused feature vector reflecting the insulation state of the transformer.

[0058] Specifically, the initial feature vector obtained from dimensionality reduction is used as input and fed into a fully connected neural network for nonlinear transformation. This neural network consists of multiple layers of neurons, with each layer's neurons connected to all neurons in the previous layer. Nonlinear transformation can capture the complex relationships between features, enhancing the model's expressive power. Neurons achieve nonlinear transformation through activation functions. [Image of a neuron in the neural network] Output Calculated using the following formula: , in, It is the first of the input vectors element It is a connection input and neurons The weight, It is a bias term. It is a non-linear activation function, such as the modified linear unit function: .

[0059] For example, a preliminary feature vector of length 128 is input into a fully connected neural network containing two hidden layers. The features are nonlinearly transformed by a modified linear unit function, and finally a fused feature vector of length 64 is output. This vector is the final representation reflecting the insulation state of the transformer.

[0060] Optionally, comparing the fused feature vector with a preset evaluation standard feature library includes: Calculate the cosine similarity between the fused feature vector and the preset evaluation standard feature library to obtain preliminary classification results; Specifically, the obtained fused feature vector Compared with the standard feature vectors stored in the preset evaluation standard feature library A comparison is performed. The comparison uses a cosine similarity algorithm, which measures the similarity between two vectors by calculating the cosine of the angle between them; the closer the value is to 1, the more similar they are. Similarity scores are calculated using vectors from all standard vectors in the database, and the standard vector with the highest score is selected as the initial classification result. The cosine similarity between two vectors is then calculated. Calculated using the following formula: , in, It is the dot product of two vectors. and These are their respective L2 norms.

[0061] like Figure 4As shown, the cosine similarity between the fused feature vector of the transformer to be evaluated and different standard feature libraries is presented.

[0062] For example, a fused feature vector is The feature vector of the "normal" state in the preset evaluation standard feature library is: ,but and The cosine similarity is: , If the calculated feature vector has a higher similarity to the "slight aging" state, the preliminary classification result is "slight aging".

[0063] Based on the preliminary classification results, a support vector machine classifier is used for classification to obtain a confidence score. Specifically, the preliminary classification results are used as input to a Support Vector Machine (SVM) classifier. The SVM classifier separates samples of different classes by finding an optimal separating hyperplane in a multidimensional feature space. This hyperplane maximizes the distance to the nearest sample point, thereby improving the classification's generalization ability. The confidence score is obtained by calculating the distance from the sample point to this optimal separating hyperplane; the greater the distance, the higher the confidence score. The optimal separating hyperplane can be represented by the following formula: , in, It is the normal vector of the hyperplane. It is the input feature vector. It is a bias term. Confidence score can be Calculated.

[0064] For example, a support vector machine classifier processes a feature vector whose initial classification result is "normal". Assume the normal vector of the optimal classifying hyperplane is... Bias term The confidence score of the vector to the hyperplane is then calculated as follows: , The confidence score indicates the degree of certainty regarding the classification result.

[0065] Based on the confidence score and historical operating data, the insulation status assessment results of the transformer are obtained.

[0066] Specifically, the obtained confidence score is combined with historical operating data such as the transformer's service life, load conditions, and environmental parameters. For example, if the confidence score is high, but historical operating data shows that the transformer has exceeded its service life or has been operating under high load and harsh conditions for a long time, the assessment result will be revised. Finally, through a decision logic or rule set, the final insulation condition assessment result is obtained. For example, a transformer has a confidence score of 80%, initially classified as being in a "normal" state. However, its historical operating data shows that the transformer has been in operation for 25 years, far exceeding its design life. In this case, the initial result will be revised, resulting in a final assessment of "normal state requiring enhanced monitoring," and a corresponding early warning will be generated.

[0067] Optionally, the feature enhancement of the processed features using deformable convolution includes: Based on the processed features, an offset learning network is used to predict the offset of deformable convolution kernels. Specifically, the processed features are fed into a lightweight offset learning network. This network learns and generates an offset field with the same spatial size as the input feature map through a series of convolutional operations. Each element in this offset field corresponds to a non-integer offset value of a convolutional sampling point in two-dimensional space, which can adaptively adjust according to the content of the feature map, thereby accurately capturing subtle changes and irregular defects in the transformer insulation medium.

[0068] For example, a feature map containing an aged region of insulating medium is input into an offset learning network. The network analyzes the shape and texture of the aged region and generates an offset field, where the offset value corresponding to the aged region causes the sampling points of the deformable convolution kernel to move towards the edges and key points of the region to better focus on the defect.

[0069] The offset of the deformable convolution kernel is convolved with the processed features to obtain the enhanced features.

[0070] Specifically, the predicted offset field is combined with the processed feature map. During the convolution operation, the sampling grid of each convolution kernel is no longer fixed, but dynamically deformed according to the value of the offset field. This dynamically deformed convolution operation allows the convolution kernel to adapt more flexibly and accurately to various morphological characteristics of the insulating medium, especially irregularly shaped defects, thereby achieving effective feature enhancement.

[0071] For example, when processing a defect region with irregular edges, the deformable convolution kernel dynamically adjusts the sampling points from a traditional rectangular grid to an irregular shape that can closely fit the defect edges, guided by the offset field. This avoids collecting irrelevant background information and extracts more accurate local features.

[0072] Optionally, dimensionality reduction of the fused feature map includes: The fused feature map is compressed in spatial dimension using a global average pooling algorithm to obtain a feature matrix; Specifically, a global average pooling algorithm is performed on the fused feature map. This algorithm calculates the average value of all pixels in each feature channel of the fused feature map, compressing the two-dimensional spatial information of that channel into a single value. This process generates a value representing the overall intensity of each feature channel, ultimately transforming the entire fused feature map into a 1×1×C feature matrix, where C is the number of channels. For example, a fused feature map of size 7×7×128 is compressed into a feature matrix of size 1×1×128 after global average pooling.

[0073] The feature matrix is ​​flattened to obtain preliminary feature vectors.

[0074] Specifically, the feature matrix is ​​flattened. Flattening is an operation that converts a multidimensional array into a one-dimensional vector, arranging all elements of the feature matrix sequentially into a single vector. This one-dimensional vector is the initial feature vector, which retains the key information of the fused feature map in each channel and provides a standardized input format for subsequent fully connected neural network processing.

[0075] For example, a feature matrix of size 1×1×128 is flattened and then converted into a preliminary feature vector containing 128 elements.

[0076] Optionally, the classification using a support vector machine classifier includes: The preliminary classification results are nonlinearly mapped using radial basis functions to obtain feature vectors in a high-dimensional feature space. Specifically, for the preliminary classification results, a radial basis function (RBF) is used as the kernel function to nonlinearly map them to a high-dimensional feature space. Insulation state samples that are difficult to separate linearly in the low-dimensional space can be more easily separated by a hyperplane in the high-dimensional space after this nonlinear mapping, thus improving classification accuracy. The kernel function calculates the inner product in the high-dimensional space by calculating the distance between sample points in the low-dimensional space, thereby avoiding the complexity of explicitly calculating the high-dimensional mapping. Two input vectors... and RBF kernel function between Calculated using the following formula: , in, It is a parameter that controls the width of the kernel function. It is the square of the Euclidean distance between the two vectors.

[0077] For example, the feature vector of a preliminary classification result is: The support vectors are Assuming parameters It is 0.5. (Through calculation) This yields the similarity between two vectors in a high-dimensional space.

[0078] The confidence score is calculated based on the distance between the feature vectors in the high-dimensional feature space and the feature vectors in the high-dimensional feature space.

[0079] Specifically, in the high-dimensional feature space mapped by the RBF kernel function, the optimal classification hyperplane that can separate sample points of different insulation state categories is sought. The confidence score is calculated by mapping the sample point to the high-dimensional space and then calculating its distance to the optimal classification hyperplane. The farther the distance from the hyperplane, the higher the certainty of the sample point's category, and the higher the confidence score. The decision function of the support vector machine classifier... This can be considered a confidence score, calculated using the following formula: , in, It is the feature vector to be classified. These are the support vectors obtained through training. It is a Lagrange multiplier. is the class label of the support vector, and N is the number of support vectors. It is the RBF kernel function.

[0080] For example, after RBF kernel function mapping, when calculating the decision function value of a feature vector to be classified, the kernel function operation is performed with multiple support vectors to finally obtain the decision function value. Since this value is much greater than 0, the classification result can be determined to be positive. At the same time, the value of 1.5 can also be used as its confidence score, indicating that the classification result has a high degree of certainty.

[0081] Based on the same inventive concept, such as Figure 5 As shown, the present invention also provides a transformer insulation condition assessment system based on multi-scale spectral feature fusion, the system comprising: The data acquisition module is used to acquire the original spectral images of the transformer's historical operating data and the spectral information of the insulating medium in several bands; The multi-scale decomposition module is used to perform multi-scale decomposition on the original spectral image to obtain spectral features at different scales. The feature fusion module is used to fuse the spectral features at different scales to generate a fused feature map; The feature analysis and extraction module is used to analyze the fused feature map and extract the fused feature vector reflecting the insulation state of the transformer; The evaluation module is used to compare the fused feature vector with a preset evaluation standard feature library to obtain the insulation status evaluation result of the transformer.

[0082] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0083] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A method for evaluating the insulation status of transformers by fusing multi-scale spectral features, characterized in that, The method includes: Acquire raw spectral images of the transformer's historical operating data and the spectral information of the insulating medium in several bands; The original spectral image is decomposed into multiple scales to obtain spectral features at different scales, including macroscopic overall spectral features and microscopic local spectral features. The spectral features at different scales are fused to generate a fused feature map; The fused feature map is analyzed to extract the fused feature vector reflecting the insulation state of the transformer; The fused feature vector is compared with a preset evaluation standard feature library to obtain the insulation status evaluation result of the transformer.

2. The method for evaluating transformer insulation status by multi-scale spectral feature fusion according to claim 1, characterized in that, The original spectral images used to acquire historical operating data of the transformer and spectral information of the insulating medium in several bands include: Collect historical operating data of the transformer, including operating years, load conditions, and environmental parameter data; Based on the transformer insulation medium, a hyperspectral imaging technique is used to scan and acquire a hyperspectral data cube covering the visible to near-infrared bands. The hyperspectral data cube is dimensionality reduced, and sensitive bands of the insulation state are selected based on the characteristic differences of the insulating medium to obtain several band spectral information. The spectral information of the aforementioned bands is used to reconstruct the image, generating the original spectral image.

3. The method for evaluating transformer insulation status by multi-scale spectral feature fusion according to claim 1, characterized in that, Multi-scale decomposition of the original spectral image includes: Based on the original spectral image, an adaptive Gaussian kernel is used for spatial domain multi-scale decomposition. By dynamically adjusting the standard deviation of the Gaussian kernel, the macroscopic overall spectral features are obtained. Based on the original spectral image, a frequency domain multi-scale decomposition method based on two-dimensional discrete wavelet transform is used. By combining the energy entropy criterion, the number of decomposition layers is automatically selected to obtain microscopic local spectral features.

4. The method for evaluating the insulation status of a transformer by multi-scale spectral feature fusion according to claim 1, characterized in that, The fusion processing of the spectral features at different scales includes: Multi-band joint convolution operations are performed on both macroscopic overall spectral features and microscopic local spectral features to extract spatial distribution features while maintaining the correlation between bands, resulting in processed features. Deformable convolution is used to enhance the processed features; The weighting coefficients are calculated based on the mutual information of each feature, and the macroscopic overall spectral features and microscopic local spectral features are weighted using the weighting coefficients to obtain a fused feature map.

5. The method for evaluating the insulation status of a transformer by multi-scale spectral feature fusion according to claim 1, characterized in that, The extraction of the fused feature vector reflecting the insulation state of the transformer includes: The fused feature map is dimensionality reduced to generate a preliminary feature vector; Based on the preliminary feature vectors, a fully connected neural network is used for nonlinear transformation to obtain a fused feature vector reflecting the insulation state of the transformer.

6. The method for evaluating transformer insulation status by multi-scale spectral feature fusion according to claim 1, characterized in that, The comparison of the fused feature vector with a preset evaluation standard feature library includes: Calculate the cosine similarity between the fused feature vector and the preset evaluation standard feature library to obtain preliminary classification results; Based on the preliminary classification results, a support vector machine classifier is used for classification to obtain a confidence score. Based on the confidence score and historical operating data, the insulation status assessment results of the transformer are obtained.

7. The method for evaluating the insulation status of a transformer by multi-scale spectral feature fusion according to claim 4, characterized in that, The feature enhancement of the processed features using deformable convolution includes: Based on the processed features, an offset learning network is used to predict the offset of deformable convolution kernels. The offset of the deformable convolution kernel is convolved with the processed features to obtain the enhanced features.

8. The method for evaluating the insulation status of a transformer by multi-scale spectral feature fusion according to claim 5, characterized in that, Dimensionality reduction of the fused feature map includes: The fused feature map is compressed in spatial dimension using a global average pooling algorithm to obtain a feature matrix; The feature matrix is ​​flattened to obtain preliminary feature vectors.

9. The method for evaluating the insulation status of a transformer by multi-scale spectral feature fusion according to claim 6, characterized in that, The classification using a support vector machine classifier includes: The preliminary classification results are nonlinearly mapped using radial basis functions to obtain feature vectors in a high-dimensional feature space. The confidence score is calculated based on the distance between the feature vectors in the high-dimensional feature space and the feature vectors in the high-dimensional feature space.

10. A transformer insulation condition assessment system based on multi-scale spectral feature fusion is applied to the transformer insulation condition assessment method based on multi-scale spectral feature fusion as described in any one of claims 1-9, characterized in that, The system includes: The data acquisition module is used to acquire the original spectral images of the transformer's historical operating data and the spectral information of the insulating medium in several bands; The multi-scale decomposition module is used to perform multi-scale decomposition on the original spectral image to obtain spectral features at different scales. The feature fusion module is used to fuse the spectral features at different scales to generate a fused feature map; The feature analysis and extraction module is used to analyze the fused feature map and extract the fused feature vector reflecting the insulation state of the transformer; The evaluation module is used to compare the fused feature vector with a preset evaluation standard feature library to obtain the insulation status evaluation result of the transformer.

Images