Non-contact judging method and system for insulation aging degree of mutual inductor based on multi-spectral image fusion and deep residual network

Abstract

The application discloses a kind of mutual inductor insulation aging degree non-contact judging method and system based on multispectral image fusion and depth residual network.The method comprises the following steps: collecting spectrum preprocessing to obtain standardized one-dimensional spectral data;Two kinds of spectrum are spliced and then input into pre-training neural network, based on the correlation of spectral variables and the degree of polymerization and the collinearity between variables, variable weighting is carried out, intermediate data fusion is generated to realize weighted combination spectrum;Weighted combination spectrum is input into depth residual shrinkage network, each residual shrinkage module executes convolution extraction, adaptive channel threshold generation, soft threshold denoising and residual connection in turn, and outputs multi-channel feature map;After pooling, flattening and full connection layer regression calculation, the output is the predicted value of the degree of polymerization of insulation paper;Compared with threshold value, determine aging state and output evaluation result.Through multispectral fusion, the lack of single spectral information is made up, and the soft threshold mechanism of depth residual shrinkage network is used to filter out spectral noise to realize accurate prediction of aging degree.

Description

Technical Field

[0001] This invention belongs to the field of power equipment insulation condition monitoring technology, specifically relating to a non-contact method and system for judging the degree of insulation aging of instrument transformers based on multispectral image fusion and deep residual network. Background Technology

[0002] Instrument transformers are critical equipment in power systems, and their insulation aging directly affects grid safety. Currently, the degree of aging is mainly assessed by measuring the degree of polymerization (DP value) of the insulation paper through power outage sampling. However, this method is cumbersome, has poor timeliness, and damages the equipment's seal.

[0003] Existing technologies attempt to use single spectra such as infrared or Raman for non-contact detection, but they have significant drawbacks: single-spectral feature information is limited and cannot comprehensively reflect aging characteristics; moreover, spectral data has high dimensionality and noise, and severe multicollinearity among variables leads to low prediction accuracy and poor robustness of traditional modeling methods. In addition, existing methods lack the ability to accurately quantify the degree of aggregation and provide hierarchical early warning.

[0004] Therefore, there is an urgent need for a non-contact intelligent judgment method that integrates the advantages of multispectral analysis, has strong noise resistance, and can accurately quantify the degree of insulation aging. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a non-contact method for judging the insulation aging degree of current transformers based on multispectral image fusion and deep residual network, so as to solve the problems of relying on power outage sampling, incomplete single spectral features, strong collinearity of high-dimensional spectral data and weak noise resistance of prediction models in the prior art.

[0006] In a first aspect, embodiments of this application provide a non-contact method for determining the insulation aging degree of a current transformer based on multispectral image fusion and a deep residual network, the method comprising:

[0007] The Fourier transform infrared spectrum and Raman spectrum of the insulating oil sample of the transformer under test were acquired by non-contact acquisition, and preprocessed to obtain standardized one-dimensional spectral data.

[0008] Two spectra are spliced ​​along the variable dimension and then input into a pre-trained neural network with a single hidden layer. An objective function is constructed based on the correlation between each spectral variable and the degree of polymerization of the insulating paper and the collinearity between variables for pre-training. The variable weights are output and the spliced ​​spectra are weighted variable by variable to achieve intermediate data fusion and generate weighted combined spectra.

[0009] The weighted combined spectrum is input into a deep residual shrinking network composed of multiple residual shrinking modules stacked together. Each module sequentially performs convolution extraction, adaptive generation of shrinking thresholds for each channel, soft thresholding noise reduction and residual connection operations based on the shrinking thresholds, and outputs the final multi-channel feature map through the last module.

[0010] The final multi-channel feature map is subjected to regression calculation through a pooling layer, a flattening layer and a fully connected layer to output the predicted value of the degree of polymerization of insulating paper.

[0011] The predicted degree of polymerization of the insulating paper is compared with a preset aging threshold to determine the insulation aging status of the transformer and output a non-contact evaluation result.

[0012] Secondly, embodiments of this application provide a non-contact system for determining the insulation aging degree of a current transformer based on multispectral image fusion and a deep residual network, applied to the method described in the first aspect, the system comprising:

[0013] The spectral acquisition module is used to non-contactly acquire the Fourier transform infrared spectrum and Raman spectrum of the insulating oil sample of the transformer under test, and preprocess them to obtain standardized one-dimensional spectral data.

[0014] The data fusion module is used to splice two spectra along the variable dimension and input them into a pre-trained neural network with a single hidden layer. Based on the correlation between each spectral variable and the degree of polymerization of the insulating paper and the collinearity between variables, the objective function is constructed for pre-training. The module outputs the variable weights and performs variable-wise weighting on the spliced ​​spectrum to achieve intermediate data fusion and generate a weighted combined spectrum.

[0015] The deep feature extraction module is used to input the weighted combined spectrum into a deep residual shrinking network composed of multiple residual shrinking modules stacked together. Each module sequentially performs convolution extraction, adaptive generation of shrinking thresholds for each channel, soft thresholding noise reduction and residual connection operations based on the shrinking thresholds, and outputs the final multi-channel feature map through the final module.

[0016] The aging degree regression module is used to perform regression calculations on the final multi-channel feature map through the pooling layer, flattening layer and fully connected layer, and output the predicted value of the degree of polymerization of the insulating paper.

[0017] The aging status determination module is used to compare the predicted value of the degree of polymerization of the insulating paper with the preset aging threshold, determine the aging status of the current transformer insulation, and output non-contact evaluation results.

[0018] Thirdly, embodiments of this application provide an electronic device, including:

[0019] processor;

[0020] Memory used to store processor-executable instructions;

[0021] The processor is configured to implement, when executing the instructions, a non-contact method for determining the insulation aging degree of current transformers based on multispectral image fusion and deep residual networks as described in the first aspect.

[0022] Fourthly, embodiments of this application provide a computer-readable storage medium storing a program that instructs a device to execute the non-contact method for determining the insulation aging degree of a current transformer based on multispectral image fusion and deep residual networks as described in the first aspect.

[0023] The beneficial effects of this invention are as follows: It employs non-contact spectral acquisition, eliminating the need for power outages during sampling and avoiding oil sample contamination, thus enabling rapid assessment of insulation status. By fusing infrared and Raman spectral information and combining correlation enhancement and collinearity suppression mechanisms, redundant variables are effectively eliminated, improving data quality. The introduction of a deep residual shrinkage network with soft thresholding adaptively filters out noise features, solving the problems of weak spectral signals and strong background interference, significantly improving the accuracy and robustness of aggregation degree prediction. The aggregation degree values ​​are converted into intuitive aging levels, and maintenance recommendations are output, providing a scientific basis for condition-based maintenance of power equipment. Attached Figure Description

[0024] Figure 1 This is a schematic flowchart of a non-contact method for determining the insulation aging degree of a current transformer based on multispectral image fusion and deep residual network, provided in an embodiment of this application.

[0025] Figure 2 This is a diagram illustrating the architecture of a non-contact system for determining the insulation aging degree of a current transformer based on multispectral image fusion and deep residual networks, provided in an embodiment of this application.

[0026] Figure 3 A schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0027] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0028] It should be noted that in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0029] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0030] Example 1

[0031] Figure 1 This is a schematic flowchart illustrating a non-contact method for determining the insulation aging degree of a current transformer based on multispectral image fusion and deep residual networks, provided as an embodiment of this application. Figure 1 As shown, a non-contact method for determining the insulation aging degree of current transformers based on multispectral image fusion and deep residual networks includes:

[0032] S1. The Fourier transform infrared spectrum and Raman spectrum of the insulating oil sample of the current transformer under test are acquired non-contactly, and preprocessed to obtain standardized one-dimensional spectral data. The acquisition and cleaning of the raw spectral signals are completed to eliminate instrument noise and optical path differences, unify the data scale, and provide a clean and regular input basis for subsequent analysis.

[0033] Specifically, in this embodiment, the preprocessing of the Fourier transform infrared spectrum and the Raman spectrum includes:

[0034] S1.1 The Savitzky-Golay filtering algorithm is used to smooth and denoise the two spectra respectively. The filtering window width set for the Fourier transform infrared spectrum is larger than that set for the Raman spectrum. The polynomial fitting order is set to the second to fourth order.

[0035] In practical implementation, for Fourier transform infrared (FTIR) spectroscopy, since its signal contains broad characteristic peaks and is easily affected by low-frequency baseline drift, the window width of the Savitzky-Golay filter is set to 25 data points to filter out broadband noise; for Raman spectroscopy, since its characteristic peaks are relatively sharp and its full width at half maximum (FWHM) is narrow, the window width is set to 9 data points to avoid losing molecular fingerprint information during the smoothing process; the polynomial fitting order for both spectra is selected as third order to achieve smoothing while preserving the higher-order characteristics of the spectral signal.

[0036] Savitzky-Golay filtering achieves smoothing through local polynomial least squares fitting, and its mathematical expression is as follows: Let the spectral data to be smoothed be... ( , (total number of spectral variables), window half-width is Window width is ,use polynomial of degree ( Fit the data points within the window:

[0037] ,

[0038] in, The coefficients are polynomials, determined by minimizing the weighted sum of squared residuals within the window:

[0039] ,

[0040] in, For the first The original values ​​of spectral intensity at each wavenumber position; For the first Smoothed spectral intensity values ​​at each wavenumber position; For indexing spectral variables, ; The FTIR value is half the window width. =12, Raman takes =4; Indicates the width of the filtering window; The order of the polynomial fitting is taken in this embodiment. =3; The first in the window The convolution weight coefficients at each position are given by the least squares analytical solution; The total number of variables for a single spectral data point. For the first The fitting coefficients of the order polynomial.

[0041] Set the window width for FTIR spectra. =25 (i.e.) =12), setting the window width for Raman spectroscopy. =9 (i.e.) =4), polynomial order =3. Convolution weight coefficients The analytical solution of the least squares fit is pre-calculated and fixed. In practice, the Savitzky-Golay filtering algorithm extracts spectral data within a window centered on each data point and uses the least squares method to determine the fitting coefficients. Then, the value of the polynomial at the center point is taken as the smoothed output. The window slides point by point across the spectral sequence, traversing the entire sequence. The system uses a set of data points to smooth the entire spectrum. For data points located at the edge of the spectrum that cannot fill the window, a mirror-image approach is used to expand the boundary values ​​to fill the gaps.

[0042] S1.2. Using a vector normalization method, the smoothed spectral data is divided by its own Euclidean length, mapping the spectral intensity to the interval between zero and one, thus obtaining the standardized one-dimensional spectral data. In the implementation process, the intensity vectors of the FTIR and Raman spectra processed in S1.1 are first extracted. Then, the Euclidean length of this vector (i.e., the square root of the sum of the squares of all intensity values) is calculated. Each intensity value in the vector is divided by this Euclidean length, making the processed spectral data a unit vector. This vector normalization process can eliminate the overall spectral intensity variation caused by fluctuations in light source intensity or differences in sample concentration, ensuring that spectral data collected from different batches only reflect differences in the chemical structure of the substances themselves during feature comparison, rather than differences in the absolute values ​​of physical quantities.

[0043] Vector normalization maps the smoothed spectral intensity vector to a unit vector, and its calculation is as follows: Let the smoothed spectral intensity vector be... , Given the total number of variables in the spectrum, the Euclidean length is... Defined as:

[0044] ,

[0045] Normalized spectral intensity values Calculate using the following formula:

[0046] , ,

[0047] in, This represents the smoothed spectral intensity vector. For the smoothed first Spectral intensity values ​​of each variable; The total number of variables for a single spectral data point; is the Euclidean length (L2 norm) of the spectral intensity vector. For the normalized first The spectral intensity values ​​of each variable, with a range of [0,1]; For indexing spectral variables.

[0048] Independent vector normalization is performed on the FTIR and Raman spectra separately. Specifically, the intensity vector of the smoothed FTIR spectrum is first extracted, and its Euclidean length is calculated. Then, for each intensity value in the vector Divide each vector by this length to make the processed FTIR spectral vector a unit vector. Similarly, the same operation is performed on Raman spectroscopy. This processing eliminates the overall intensity variation of the spectrum caused by fluctuations in light source intensity or differences in sample concentration, ensuring that the two spectral data are on a uniform scale before feature fusion. This allows subsequent analysis to reflect only the differences in the chemical structure of the substances themselves, rather than the differences in the absolute values ​​of physical quantities.

[0049] S2. The two spectra are concatenated along the variable dimension and input into a pre-trained neural network with a single hidden layer. An objective function is constructed based on the correlation between each spectral variable and the degree of polymerization of the insulating paper, as well as collinearity among variables. This function is then pre-trained, outputting variable weights and weighting the concatenated spectra variable-by-variable to achieve intermediate-level data fusion and generate a weighted combined spectrum. Using statistical constraints, highly correlated and independent spectral features are selected. Weighted fusion addresses the issue of one-sided information from a single spectral spectrum, extracting the most representative comprehensive feature vector.

[0050] Specifically, in this embodiment, the pre-training of the objective function based on the correlation between each spectral variable and the degree of polymerization of the insulating paper, as well as the collinearity among the variables, includes:

[0051] S2.1 Define the objective function as consisting of a correlation enhancement term and a multicollinearity suppression term; the correlation enhancement term is used to maximize the absolute value of the Pearson correlation coefficient between the weighted combined spectral data and the degree of polymerization of the insulating paper; the multicollinearity suppression term is used to minimize the variance inflation factor among the weighted spectral variables.

[0052] In practice, the objective function consists of two mutually balancing sub-terms: the correlation enhancement term guides the model to assign greater weights to spectral bands that are highly correlated with the degree of aging by calculating the absolute value of the Pearson correlation coefficient between the weighted fused spectrum and the degree of polymerization of the insulating paper; the multicollinearity suppression term introduces the variance inflation factor (VIF) as an evaluation index. The higher the VIF value, the more severe the multicollinearity among variables. This term aims to reduce redundant information among different spectral variables, prevent model overfitting, and thus screen out characteristic bands with strong independence.

[0053] The Pearson correlation coefficient is used to measure the degree of linear correlation between spectral variables and the degree of aggregation. It is defined as follows:

[0054] Let the first The values ​​of the spectral variables in all training samples are The corresponding measured degree of polymerization of the insulating paper is , Let be the total number of training samples. and The Pearson correlation coefficient between them is:

[0055] ,

[0056] in, For the first In the training samples, the th Intensity values ​​of each spectral variable For the first Measured values ​​of the degree of polymerization of insulating paper for each training sample. For the first The mean of each spectral variable across all training samples; The mean of the aggregation degree measurements for all training samples; For the first Pearson correlation coefficients between spectral variables and degree of aggregation.

[0057] For each variable in the spliced ​​spectrum Calculate its correlation with the aggregation degree label respectively. Pearson correlation coefficient between In the specific calculation, first calculate the mean of each variable and the degree of aggregation. and Then calculate the product of biases for each sample. The summation is used as the numerator, and the square root of the product of the sum of squares of deviations of each variable and the sum of squares of deviations of aggregation degree is used as the denominator. The closer a value is to 1, the stronger the linear correlation between the variable and the degree of aging, and the greater the weight it should receive in subsequent weighting.

[0058] The pre-training objective function constructed in this invention consists of a correlation enhancement term and a multicollinearity suppression term, and its complete mathematical form is as follows: Let there be a total of There are variables, and the variable weight vector is... ,in For the first The weighting coefficients of each variable. The weighted spectral variable value is... .

[0059] The pre-training objective function is defined as:

[0060] ,

[0061] in, For the weighted first The variance inflation factor for each variable is defined as follows:

[0062] ,

[0063] In the above formula, For the weighted average One variable is the dependent variable, and the rest... The coefficient of determination obtained by performing linear regression with weighted variables as independent variables. This represents the total number of variables in the spliced ​​spectrum (FTIR variables + Raman variables); For the first The weighting coefficients of each spectral variable are output by a single hidden layer neural network. For the first In the nth sample The spectral intensity value after weighting of each variable; For the first Pearson correlation coefficients between the variables and the degree of aggregation; For the weighted first The variance inflation factor (VIF) of each variable is used to measure multicollinearity. For the first The variables affect the rest The linear regression determination coefficients of the variables, with values ​​ranging from [0,1]; The weighting coefficient for the relevance enhancement term is taken in this embodiment. =1.0; The weighting coefficients for the collinearity suppression term are taken in this embodiment. =0.5; The loss function is used for pre-training the importance factor block; For training sample index; For indexing spectral variables.

[0064] A single-hidden-layer neural network is used as the carrier for weight optimization, with the stitched spectral data as input and the objective function... The optimization objective is to minimize the value of the objective function. The Adaptive Moment Estimation (Adam) algorithm is used to perform gradient descent iterations, dynamically adjusting the connection weights between the input layer and the hidden layer until the gradient of the objective function approaches zero or the maximum number of iterations is reached. In specific implementations, The value is 1.0. A value of 0.5 is used to ensure that the correlation enhancement term is dominant, while effectively suppressing multicollinearity among variables. After convergence, the connection weight matrix from the input layer to the hidden layer reflects the importance of each spectral variable in the prediction of the degree of aggregation. This matrix is ​​extracted and reshaped into a one-dimensional vector. , which serves as the basis for subsequent spectral weighting.

[0065] S2.2. Iteratively optimize the objective function using gradient descent, adjust the connection weights of the single hidden layer neural network until convergence, and output the converged connection weights from the input layer to the hidden layer as the variable weights.

[0066] In this embodiment, a single-hidden-layer neural network is used as the carrier for weight optimization. The stitched spectral data is taken as input, and the optimization objective is to minimize the objective function. The Adaptive Moments Estimation (Adam) algorithm is used to perform gradient descent iterations, dynamically adjusting the connection weights between the input layer and the hidden layer until the gradient of the objective function approaches zero or the maximum number of iterations is reached. At this point, the converged connection weight matrix reflects the importance of each spectral variable to the degree of aggregation prediction. This matrix is ​​extracted and reshaped into a one-dimensional vector, which serves as the basic weight coefficients for subsequent spectral weighting.

[0067] Specifically, in this embodiment, the step of weighting the spliced ​​spectra variable by variable to achieve intermediate data fusion and generate a weighted combined spectrum includes:

[0068] S2.3 Map the variable weights to the corresponding spectral variable positions to form a weight vector consistent with the dimensions of the spliced ​​spectrum. During implementation, the variable weights obtained in step S2.2 are mapped one-to-one according to their positions in the original spliced ​​spectrum to construct a weight vector with dimensions completely consistent with the spliced ​​spectrum. Specifically, if the spliced ​​spectrum consists of the first m FTIR variables and the last n Raman variables, the weight vector will also contain m+n elements, and the i-th weight value strictly corresponds to the intensity value of the i-th spectral wavelength, thus ensuring that each spectral channel can obtain an independent weighting coefficient according to its importance.

[0069] S2.4 Calculate the Hadamard product of the weight vector and the spliced ​​spectrum to obtain the preliminary weighted spectrum. In this step, the constructed weight vector is multiplied element-wise with the corresponding spliced ​​spectral data, i.e., the Hadamard product. Through this operation, elements with larger values ​​in the weight vector amplify the signal intensity of the corresponding spectral band, while elements with values ​​close to zero suppress or filter irrelevant and noise bands, thereby enhancing key aging characteristics and weakening interference information at the data level, generating the preliminary weighted spectrum.

[0070] The Hadamard product is used to perform element-wise weighted summation of the weight vector and spectral data. Its mathematical expression is as follows: Let the spliced ​​spectral vector be... The weight vector is Then the weighted spectral vector The The elements are: , The vector form is expressed as: , where ⊙ represents the Hadamard product (element-by-element multiplication) operator. To splice spectral vectors, For splicing the spectrum, the first The intensity values ​​of each variable (after normalization). This is the weight vector output by the importance factor block; For the first Weighting coefficients of each spectral variable, The weighted spectral vector; For the weighted first Intensity values ​​of each spectral variable; For indexing spectral variables, .

[0071] Obtaining the weight vector Then, it is compared with the normalized spliced ​​spectral vector. Perform element-wise multiplication. In practice, iterate through each variable of the spliced ​​spectrum. , will the The original intensity values ​​of each variable Multiply by the corresponding weighting coefficient The weighted intensity value is obtained. Variables with weights close to 1 are amplified, while variables with weights close to 0 are suppressed, thereby strengthening key aging characteristics and weakening irrelevant interference information at the data level.

[0072] S2.5. Perform principal component analysis (PCA) along the variable dimension on the preliminary weighted spectrum, and extract the principal components with a cumulative contribution rate that meets the preset requirement as feature vectors. To further compress the data dimensionality and eliminate residual microcollinearity, principal component analysis (PCA) is performed on the preliminary weighted spectrum. PCA transforms potentially correlated variables into a set of linearly uncorrelated principal components through orthogonal transformation. In this embodiment, the preset threshold for the cumulative contribution rate is set to 95%, that is, the top k principal components that can explain 95% of the information in the original weighted spectrum are extracted. These principal components are used as feature vectors, which not only retain the most core spectral features, but also significantly reduce the input dimensionality of subsequent deep networks.

[0073] S2.6. The feature vectors are reconstructed into a one-dimensional data sequence, which is then input into the deep residual shrinkage network as the weighted combined spectrum. Finally, the extracted multi-dimensional principal component feature vectors are serialized and reconstructed. Specifically, the principal component data are flattened and spliced ​​sequentially to form a one-dimensional data sequence. This sequence is the final weighted combined spectrum after feature selection, weighted fusion, and dimensionality reduction, containing the most significant aging characteristics of the insulating paper. This one-dimensional sequence is input into the deep residual shrinkage network, which utilizes its unique soft thresholding module to further remove noise from deep features, thereby achieving high-precision inversion prediction of the degree of polymerization of the insulating paper.

[0074] S3. The weighted combined spectrum is input into a deep residual shrinking network composed of multiple stacked residual shrinking modules. Each module sequentially performs convolution extraction on the input feature map, adaptively generates shrinking thresholds for each channel, performs soft thresholding noise reduction and residual connection operations based on the shrinking thresholds, and outputs the final multi-channel feature map through the final module. By using a deep network with an attention mechanism, noise interference hidden in the feature map is adaptively identified and removed, and the deep aging feature patterns behind the spectral data are mined, achieving noise reduction and enhancement of the features.

[0075] Specifically, in this embodiment, each module sequentially performs convolution extraction on the input feature map, adaptively generates shrinkage thresholds for each channel, performs soft thresholding noise reduction and residual connection operations based on the shrinkage thresholds, specifically including:

[0076] S3.1. Perform the first convolution operation on the input feature map to extract local spectral features. In specific implementation, the preprocessed and feature-fused one-dimensional weighted combined spectral data is reshaped into a feature map format suitable for input to the convolutional network and input into the current module of the deep residual shrinking network. The module first scans the input feature map through a one-dimensional convolutional layer (first convolution operation), and uses the sliding calculation of the convolution kernel in the spectral sequence dimension to extract local spectral features containing the correlation information of adjacent bands, generating a primary feature map containing multi-channel information, providing basic data for subsequent adaptive threshold calculation.

[0077] S3.2. The extracted local spectral features are input into a global average pooling layer. After compressing the spatial dimension, they are mapped through a fully connected layer and a non-linear activation function to adaptively generate the shrinkage threshold for each feature channel. This step aims to achieve adaptive threshold generation. First, the multi-channel feature map output from S3.1 is processed by a global average pooling layer (GAP) to compress the entire spectral dimension information of each channel into a global feature value to reduce the data dimensionality. Subsequently, these global feature values ​​are sequentially input into a fully connected layer and a sub-network with a non-linear activation function (such as Sigmoid or ReLU) for mapping learning. This sub-network outputs a set of positive numbers equal to the number of feature channels based on the characteristics of the input data, serving as the shrinkage threshold for each feature channel, reflecting the model's adaptive judgment ability regarding the noise level of different channels.

[0078] The shrinkage threshold for each feature channel is adaptively generated by the attention mechanism, and its calculation formula is as follows: Let the multi-channel feature map output after the first convolution operation be... ,in For the number of channels, The feature length for each channel is determined. First, each channel is compressed into a global feature value using global average pooling:

[0079] , ,

[0080] in, For the first The first channel The feature values ​​at each position.

[0081] Then, the global feature vector Nonlinear transformation is performed on the input fully connected layer:

[0082] ,

[0083] .

[0084] Final The shrinkage threshold for each channel is:

[0085] ,

[0086] in, For the first Adaptive shrinkage threshold for each channel, This is the normalized threshold coefficient output by the Sigmoid function. For the first Global average absolute value characteristics of each channel; This represents the global feature vector for all channels; Here is the weight matrix of the first fully connected layer, with dimension 1. ; This is the bias vector for the first fully connected layer; This is the weight matrix of the second fully connected layer, with dimension 1. , This is the bias vector for the second fully connected layer; This refers to the dimensionality reduction ratio. The activation function for the rectified linear unit; This is the Sigmoid activation function. For the position index of the feature sequence, .

[0087] The threshold is learned automatically through an attention mechanism, without relying on manual setting. In practice, the multi-channel feature map output from the first convolutional layer is first processed... Perform global average pooling to calculate the mean of the absolute values ​​of all positional features for each channel. Then the global feature vector The fully connected layers of the input compression-excitation network (SENet) architecture: The first fully connected layer reduces the dimension from Reduced to ( =16), after ReLU activation, the second fully connected layer will reduce the dimension from rise The output normalized threshold coefficient is activated by Sigmoid. Finally, Global eigenvalues ​​of the corresponding channels Multiplying them together yields the adaptive shrinkage threshold for each channel. Each channel Independent calculation adapts to different channel noise levels. This mechanism ensures the threshold is always positive. >0), to avoid soft thresholding degenerating into an identity mapping.

[0088] S3.3. Based on the shrinkage threshold, perform soft thresholding on the feature maps of each channel. Features with absolute values ​​less than the corresponding channel's shrinkage threshold are set to zero, while features with absolute values ​​greater than or equal to the shrinkage threshold are retained, thus filtering out noise features. After obtaining the shrinkage threshold for each channel, the module performs soft thresholding on the local spectral feature maps extracted in S3.1. Soft thresholding, as a nonlinear transformation function, sets any feature value in the feature map to zero if its absolute value is less than the shrinkage threshold of the channel to which the feature belongs; otherwise, it shrinks the feature value towards zero by the threshold value. This operation can effectively identify and filter out noise-related features with low amplitudes in the feature map, while retaining effective spectral features with high amplitudes, thus achieving a dual function of feature selection and data denoising.

[0089] Soft thresholding is the core denoising mechanism of deep residual shrinking networks, and its function is defined as follows: Let the feature value in a certain feature channel be... The shrinkage threshold corresponding to this channel is ( If >0), then the output of the soft thresholding function is... for:

[0090] ,

[0091] This function can also be written equivalently as:

[0092] ,

[0093] in, This represents the original activation value of a feature point in the feature map. For symbolic functions, When it is 1, The time is -1. The time is 0.

[0094] After obtaining the shrinkage threshold of each channel Then, soft thresholding is performed channel-by-channel on the local spectral feature map extracted in S3.1. For each feature value within a certain feature channel... Determine its absolute value and threshold. Relationship: If If the feature is zero, it is considered noise and set to zero; if Then the feature is preserved and shrinks towards zero. The derivative of the soft thresholding function is 0 or 1, which is consistent with the derivative characteristics of the ReLU activation function. This is beneficial for gradient backpropagation in deep networks and can effectively suppress gradient vanishing and gradient exploding.

[0095] S3.4. Perform an element-wise addition residual connection operation on the soft-thresholding denoised feature map and the original input feature map of the module to output the feature map of the current residual shrinking module. To alleviate the gradient vanishing problem during deep network training and prevent the loss of useful information, this step introduces a residual learning mechanism. The soft-thresholding denoised feature map from S3.3 is added element-wise to the input feature map of the current residual shrinking module. This identity mapping structure ensures that the network performance does not degrade as the number of layers increases, allowing the high-quality features filtered by noise to be fused with the original feature information, ultimately outputting a feature map rich in aging information and with a significantly improved signal-to-noise ratio, which can be used by subsequent network layers or fully connected regression layers to accurately predict the degree of polymerization of insulating paper.

[0096] S3.4 The soft thresholding operation performed on the feature map of each channel based on the shrinkage threshold specifically includes: sorting the feature values ​​in each feature channel according to their absolute values, retaining features whose absolute values ​​are greater than or equal to the shrinkage threshold of the corresponding channel, and setting features whose absolute values ​​are less than the shrinkage threshold to zero, thereby filtering out noise features; wherein, in the step of adaptively generating the shrinkage threshold of each channel, it also includes applying a non-negative constraint to the generated threshold, forcibly restricting the shrinkage threshold to a positive range, so as to prevent feature misretention caused by non-positive thresholds, and ensure that the soft thresholding operation has stable feature selection capability.

[0097] In the threshold generation process of S3.2, an absolute value function or ReLU activation function is introduced to impose a non-negativity constraint on the output, ensuring that the shrinkage threshold always lies within the range greater than zero. This constraint avoids the risk of the soft thresholding function degenerating into an identity mapping when the threshold is not positive, ensuring that the network maintains stable feature selection and noise reduction performance at different training stages.

[0098] S4. The final multi-channel feature map is subjected to regression calculation through a pooling layer, a flattening layer, and a fully connected layer to output the predicted value of the degree of polymerization of the insulating paper. The high-dimensional feature map is transformed into a low-dimensional numerical output, and a nonlinear mapping relationship is established between spectral features and the physical aging degree (degree of polymerization) of the insulating paper to achieve accurate quantification of aging.

[0099] Specifically, in this embodiment, the step of performing regression calculations on the final multi-channel feature map through a pooling layer, a flattening layer, and a fully connected layer to output the predicted degree of polymerization of the insulating paper includes:

[0100] S4.1 The final multi-channel feature map is input into a global average pooling layer, and each feature channel is compressed along the spatial dimension to generate a channel-level feature vector. In specific implementation, the final multi-channel feature map with rich semantic information obtained after processing by a multi-level residual shrinkage module is input into the global average pooling layer. This layer operates independently on each channel of the feature map, calculating the arithmetic mean of all spatial location feature values ​​within that channel, thereby compressing the two-dimensional feature map of size C×L into a one-dimensional vector of length C. This process reduces the number of subsequent parameters, suppresses overfitting, and retains the global response intensity information of each feature channel, generating a comprehensive channel-level feature vector.

[0101] S4.2. Input the channel-level feature vectors into the flattening layer to convert them into a one-dimensional feature sequence. This step follows from the output of S4.1, further inputting the channel-level feature vectors generated by the global average pooling layer into the flattening layer. The flattening layer reshapes the dimensions of the input data, strictly converting tensors that may carry multi-dimensional structural information into a one-dimensional array format (i.e., a one-dimensional feature sequence) without changing the total amount of data. This one-dimensional feature sequence eliminates redundant dimensions in the data structure, making it conform to the specific requirements of the fully connected layer for the input data format, thus enabling seamless connection to the subsequent regression calculation module and ensuring smooth data flow in the neural network.

[0102] S4.3 Input the one-dimensional feature sequence into a regression network consisting of at least two fully connected layers, wherein the first fully connected layer performs nonlinear mapping and dimensionality reduction on the one-dimensional feature sequence, and the last fully connected layer outputs a continuous value of a single node, which is the predicted value of the degree of polymerization of the insulating paper.

[0103] In the implementation process, a one-dimensional feature sequence is input into a regression network consisting of two or more fully connected layers stacked together. The first fully connected layer receives the sequence, performs a linear transformation through the weight matrix, and maps and reorganizes the feature space using a nonlinear activation function (such as ReLU), achieving further compression and abstraction of the feature dimension. Subsequently, the data flows into the last fully connected layer, which contains only one output node. This layer is responsible for linearly combining and mapping the high-dimensional abstract features into a continuous real value. Since the degree of polymerization (DP value) of insulating paper is a continuous physical quantity, the value of this output node is directly used as a quantitative prediction result of the aging degree of insulating paper, completing the end-to-end regression inversion from spectral features to physicochemical indicators.

[0104] Before the final fully connected layer outputs the predicted degree of polymerization of the insulating paper, a loss function constraint step is included: constructing a composite loss function containing an absolute error term and a smoothing constraint term. The absolute error term is used to measure the consistency between the predicted value and the actual degree of polymerization measurement value, and the smoothing constraint term is used to penalize the output change rate of the regression network to suppress the overfitting of the prediction result to spectral noise. During the model training phase, the parameters of the regression network and the front-end deep residual shrinkage network are iteratively updated by minimizing the composite loss function.

[0105] In practical implementation, to enhance the model's robustness to spectral noise and improve its generalization ability, the following composite loss function is constructed during the training phase: The absolute error term uses Mean Absolute Error (MAE), which measures the accuracy of the prediction by calculating the average absolute deviation between the predicted and measured polymerization degrees of the model output. MAE has a stronger tolerance for outliers than Mean Squared Error. The smoothing constraint term calculates the first-order difference or gradient norm along the input spectral dimension of the continuous values ​​output by the regression network and applies an L1 or L2 norm penalty to this rate of change, thereby constraining the prediction curve to remain smooth and suppressing overfitting caused by drastic fluctuations in random noise in the spectrum. During backpropagation, the weighted sum of these two terms is used as the final loss, simultaneously updating the learnable parameters of the back-end fully connected regression network and the front-end deep residual shrinkage network. This ensures the model has noise suppression capabilities while maintaining prediction accuracy, achieving reliable inversion of the polymerization degree of insulating paper.

[0106] This invention employs a composite loss function during the model training phase, which includes an absolute error term and a smoothing constraint term. Its complete definition is as follows:

[0107] Let the total number of training samples be , No. The predicted aggregation degree of each sample is The measured degree of polymerization is Composite loss function Defined as:

[0108] ,

[0109] The first term is the Mean Absolute Error (MAE) term. It is used to measure the consistency between predicted and measured values:

[0110] ,

[0111] The second term is the smoothing constraint term. This is used to penalize the rate of change of predicted values ​​of adjacent samples:

[0112] .

[0113] in, For the model to the first Predicted degree of polymerization of insulating paper for each training sample; For the first Measured degree of polymerization of insulating paper for each training sample (obtained by viscosity method); The total number of training samples; To determine the weighting coefficients for the smoothing constraint terms, this embodiment takes... =0.01.

[0114] During the model training phase, the composite loss function will be used. The objective of backpropagation is to calculate the predicted value for each sample using the absolute error (MAE) term. Compared with measured values The absolute deviations between the predicted values ​​are averaged, which provides good robustness to outliers. A smoothing constraint term calculates and averages the absolute values ​​of the first-order differences between the predicted values ​​of adjacent samples (sorted by aging time), applying an L1 norm penalty to the rate of change of the prediction curve. This forces the predicted values ​​to remain smooth and continuous along the aging process, suppressing overfitting of the model to random noise in the spectrum. During backpropagation, all learnable parameters of the deep residual shrinking network and the regression network are updated simultaneously using gradient descent. In this embodiment... The value is set to 0.01, which makes the absolute error term the dominant term and effectively suppresses the sharp fluctuations in the prediction curve, so that the model achieves the optimal balance between prediction accuracy and noise robustness.

[0115] S5. Compare the predicted degree of polymerization of the insulating paper with the preset aging threshold to determine the insulation aging status of the transformer and output the non-contact evaluation result. This establishes a bridge between quantitative results and engineering applications, transforming abstract numerical values ​​into intuitive operation and maintenance levels, and providing a reference basis that can be directly used for on-site maintenance decisions.

[0116] Specifically, in this embodiment, comparing the predicted degree of polymerization of the insulating paper with a preset aging threshold to determine the insulation aging state of the current transformer and outputting a non-contact evaluation result includes:

[0117] S5.1 At least two aging level thresholds are pre-set, including an initial threshold representing insulation in a newly installed or good condition, and a termination threshold representing insulation in a severely aged or failed state. In specific implementation, based on international electrical standards and the aging characteristics of transformer insulation paper, aging level thresholds for the degree of polymerization (DP value) of the insulation paper are pre-set. The initial threshold is set to DP=500, corresponding to the state of newly installed or slightly aged insulation paper, at which point the cellulose molecular chains are intact and the mechanical strength is high. The termination threshold is set to DP=250, corresponding to the state of severely aged insulation paper or nearing the end of its life, at which point the mechanical strength has significantly decreased, and there is a risk of insulation breakdown. These two thresholds constitute a three-level scale for insulation condition assessment, providing a benchmark for subsequent automated grading determination.

[0118] S5.2. The predicted value of the degree of polymerization of the insulating paper is compared with the initial threshold and the termination threshold in sequence: if the predicted value is greater than or equal to the initial threshold, the insulation state of the current transformer is determined to be normal; if the predicted value is less than or equal to the termination threshold, the insulation state of the current transformer is determined to be severely aged; if the predicted value is between the initial threshold and the termination threshold, the insulation state of the current transformer is determined to be slightly aged or in a state of concern.

[0119] This step implements a specific logical judgment process. The system sequentially substitutes the predicted value of the insulation paper polymerization degree output by the deep residual shrinkage network (e.g., DP=380) into the aforementioned threshold range for judgment: if the predicted value is ≥500, the system returns a judgment that the insulation status is normal; if the predicted value is ≤250, the system returns a warning of severe insulation aging; if the predicted value falls within the range of 250 to 500, the system returns a prompt of mild aging / concern status. This three-level judgment mechanism can transform continuous regression prediction values ​​into intuitive engineering decision-making basis, making it easier for operation and maintenance personnel to quickly grasp the health level of the equipment.

[0120] S5.3 Generate a visual assessment report containing the insulation status level, the predicted value of the degree of polymerization of the insulation paper, and maintenance recommendations based on the judgment result, and complete the output of the non-contact assessment result.

[0121] After the assessment is completed, the system automatically generates a structured, visual evaluation report. The report not only displays the quantified insulation condition level (e.g., mild aging) and specific predicted DP values ​​(e.g., predicted DP value: 380), but also outputs targeted maintenance recommendations based on the assessment results (e.g., for mild aging, it recommends shortening the inspection cycle and strengthening oil chromatography monitoring). The report is output in GUI or PDF format, achieving a fully contactless, intelligent closed-loop assessment process from spectral acquisition to diagnostic conclusions, guiding on-site maintenance work without requiring power outages for sampling.

[0122] Example 2

[0123] like Figure 2 As shown, this application provides a non-contact system architecture diagram for judging the insulation aging degree of current transformers based on multispectral image fusion and deep residual networks, which is applied to the method described in Embodiment 1, including:

[0124] The spectral acquisition module 210 is used to non-contactly acquire the Fourier transform infrared spectrum and Raman spectrum of the insulating oil sample of the transformer under test, and preprocess them to obtain standardized one-dimensional spectral data.

[0125] The data fusion module 220 is used to splice two spectra along the variable dimension and input them into a pre-trained neural network with a single hidden layer. Based on the correlation between each spectral variable and the degree of polymerization of the insulating paper and the collinearity between variables, the objective function is constructed for pre-training. The variable weights are output and the spliced ​​spectra are weighted variable by variable to realize intermediate data fusion and generate weighted combined spectra.

[0126] The deep feature extraction module 230 is used to input the weighted combined spectrum into a deep residual shrinking network composed of multiple residual shrinking modules stacked together. Each module sequentially performs convolution extraction, adaptive generation of shrinking thresholds for each channel, soft thresholding noise reduction and residual connection operations based on the shrinking thresholds, and outputs the final multi-channel feature map through the final module.

[0127] The aging degree regression module 240 is used to perform regression calculations on the final multi-channel feature map through the pooling layer, flattening layer and fully connected layer, and output the predicted value of the degree of polymerization of the insulating paper.

[0128] The aging state determination module 250 is used to compare the predicted value of the degree of polymerization of the insulating paper with the preset aging threshold, determine the aging state of the current transformer insulation, and output non-contact evaluation results.

[0129] Figure 3 This is an electronic device provided in one embodiment of this application. For example... Figure 3 As shown, the electronic device includes at least the following components: processor 301 and memory 300, communication interface 303, and bus 302.

[0130] In this embodiment of the application, memory 300 is used to store executable instructions of processor 301, which, when configured to execute instructions, implements the method as described in the first aspect.

[0131] In embodiments of this application, a computer-readable storage medium includes instructions that instruct a device to perform the method as described in the first aspect. For example, the instructions instruct the device to perform... Figure 1 The method is shown in the process steps.

[0132] In one embodiment of this application, the program operating in the electronic device may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). Information processed by these systems is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (FlashROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.

[0133] It should be noted that a portion of the electronic device described above can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0134] It should be noted that the computer mentioned here refers to a computer built into an electronic device, employing hardware including an operating system and peripheral devices. Furthermore, computer-readable recording media refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage systems such as hard drives built into the computer.

[0135] Furthermore, computer-readable recording media can include: media that dynamically stores programs for short periods of time, such as communication lines used when transmitting programs via networks like the Internet or communication lines like telephone lines; and media that store programs for fixed periods of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining them with programs already recorded in the computer.

[0136] Furthermore, the electronic device in the above embodiments can also be implemented as an assembly (system group) composed of multiple systems. Each system constituting the system group can possess some or all of the functions or functional blocks of the electronic device in the above embodiments. As a system group, it is sufficient to have all the functions or functional blocks of the electronic device.

[0137] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A non-contact method for determining the insulation aging degree of instrument transformers based on multispectral image fusion and deep residual networks, characterized in that, Includes the following steps: The Fourier transform infrared spectrum and Raman spectrum of the insulating oil sample of the transformer under test were acquired by non-contact acquisition, and preprocessed to obtain standardized one-dimensional spectral data. Two spectra are spliced ​​along the variable dimension and then input into a pre-trained neural network with a single hidden layer. An objective function is constructed based on the correlation between each spectral variable and the degree of polymerization of the insulating paper and the collinearity between variables for pre-training. The variable weights are output and the spliced ​​spectra are weighted variable by variable to achieve intermediate data fusion and generate weighted combined spectra. The weighted combined spectrum is input into a deep residual shrinking network composed of multiple residual shrinking modules stacked together. Each module sequentially performs convolution extraction, adaptive generation of shrinking thresholds for each channel, soft thresholding noise reduction and residual connection operations based on the shrinking thresholds, and outputs the final multi-channel feature map through the last module. The final multi-channel feature map is subjected to regression calculation through a pooling layer, a flattening layer and a fully connected layer to output the predicted value of the degree of polymerization of insulating paper. The predicted degree of polymerization of the insulating paper is compared with a preset aging threshold to determine the insulation aging status of the transformer and output a non-contact evaluation result.

2. The method according to claim 1, characterized in that, The preprocessing of the Fourier transform infrared spectrum and Raman spectrum specifically includes: The Savitzky-Golay filtering algorithm was used to smooth and denoise the two spectra respectively. The filtering window width set for the Fourier transform infrared spectrum was larger than that set for the Raman spectrum. The order of the polynomial fitting was set to the second to the fourth order. By employing a vector normalization method, the smoothed spectral data is divided by its own Euclidean length, and the spectral intensity is mapped to the interval between zero and one to obtain the standardized one-dimensional spectral data.

3. The method according to claim 1, characterized in that, The pre-training of the objective function, based on the correlation between each spectral variable and the degree of polymerization of the insulating paper, and the collinearity among the variables, specifically includes: The objective function is defined as consisting of a correlation enhancement term and a multicollinearity suppression term; The correlation enhancement term is used to maximize the absolute value of the Pearson correlation coefficient between the weighted combined spectral data and the degree of polymerization of the insulating paper; The multicollinearity suppression term is used to minimize the variance inflation factor among the weighted spectral variables. The objective function is iteratively optimized using gradient descent, and the connection weights of the single hidden layer neural network are adjusted until convergence. The converged connection weights from the input layer to the hidden layer are then output as the variable weights.

4. The method according to claim 3, characterized in that, The step of weighting the spliced ​​spectra variable by variable to achieve intermediate data fusion and generate weighted combined spectra specifically includes: The variable weights are mapped to the corresponding spectral variable positions to form a weight vector consistent with the spliced ​​spectral dimension. Calculate the Hadamard product of the weight vector and the spliced ​​spectrum to obtain the preliminary weighted spectrum; Principal component analysis is performed on the preliminary weighted spectrum along the variable dimension, and principal components with cumulative contribution rates reaching preset requirements are extracted as feature vectors; The feature vector is reconstructed into a one-dimensional data sequence, which is then used as the weighted combined spectrum input to the deep residual shrinkage network.

5. The method according to claim 1, characterized in that, Each module sequentially performs convolution extraction on the input feature map, adaptively generates shrinkage thresholds for each channel, performs soft thresholding noise reduction based on the shrinkage thresholds, and performs residual connection operations, specifically including: Perform the first convolution operation on the input feature map to extract local spectral features; The extracted local spectral features are input into a global average pooling layer. After compressing the spatial dimension, they are mapped to a nonlinear activation function through a fully connected layer to adaptively generate the shrinkage threshold corresponding to each feature channel. Based on the shrinkage threshold, a soft thresholding operation is performed on the feature map of each channel, setting features whose absolute value is less than the corresponding channel shrinkage threshold to zero, and retaining features whose absolute value is greater than or equal to the shrinkage threshold, thereby filtering out noise features; The feature map after soft thresholding and noise reduction is joined with the original input feature map of the module by an element-wise residual concatenation operation, and the feature map of the current residual shrinking module is output.

6. The method according to claim 5, characterized in that, The soft thresholding operation performed on the feature maps of each channel based on the shrinkage threshold specifically includes: The feature values ​​in each feature channel are sorted according to their absolute values. Features with absolute values ​​greater than or equal to the corresponding channel shrinkage threshold are retained, while features with absolute values ​​less than the shrinkage threshold are set to zero, thereby filtering out noise features. The step of adaptively generating the shrinkage threshold for each channel also includes applying a non-negative constraint to the generated threshold, forcing the shrinkage threshold to be limited to the positive range, so as to prevent feature misretention caused by non-positive thresholds and ensure that the soft thresholding operation has stable feature selection capability.

7. The method according to claim 1, characterized in that, The step of performing regression calculations on the final multi-channel feature map through a pooling layer, a flattening layer, and a fully connected layer to output a predicted value of the degree of polymerization of the insulating paper specifically includes: The final multi-channel feature map is input into a global average pooling layer, and each feature channel is compressed along the spatial dimension to generate a channel-level feature vector. The channel-level feature vectors are input into the flattening layer and converted into a one-dimensional feature sequence. The one-dimensional feature sequence is input into a regression network consisting of at least two fully connected layers. The first fully connected layer performs nonlinear mapping and dimensionality reduction on the one-dimensional feature sequence, and the last fully connected layer outputs a continuous value of a single node. This continuous value is the predicted value of the degree of polymerization of the insulating paper.

8. The method according to claim 7, characterized in that, Before the predicted degree of polymerization of the insulating paper is output from the final fully connected layer, a loss function constraint step is also included: A composite loss function is constructed, which includes an absolute error term and a smoothing constraint term. The absolute error term is used to measure the consistency between the predicted value and the actual aggregation degree measurement value. The smoothing constraint term is used to penalize the output change rate of the regression network to suppress the overfitting of the prediction result to spectral noise. During the model training phase, the parameters of the regression network and the front-end deep residual shrinkage network are iteratively updated by minimizing the composite loss function.

9. The method according to claim 1, characterized in that, The step of comparing the predicted degree of polymerization of the insulating paper with a preset aging threshold to determine the insulation aging state of the current transformer and outputting a non-contact evaluation result specifically includes: At least two aging level thresholds are preset, including an initial threshold that characterizes the insulation as newly installed or in good condition, and a termination threshold that characterizes the insulation as severely aged or in failure condition. The predicted value of the degree of polymerization of the insulating paper is compared with the initial threshold and the termination threshold in sequence: if the predicted value is greater than or equal to the initial threshold, the insulation state of the current transformer is determined to be normal; if the predicted value is less than or equal to the termination threshold, the insulation state of the current transformer is determined to be severely aged; if the predicted value is between the initial threshold and the termination threshold, the insulation state of the current transformer is determined to be slightly aged or in a state of concern. Based on the determination results, a visual assessment report is generated that includes the insulation condition level, the predicted value of the degree of polymerization of the insulation paper, and maintenance recommendations, thus completing the output of the non-contact assessment results.

10. A non-contact system for determining the insulation aging degree of a current transformer based on multispectral image fusion and deep residual networks, applied to the method described in any one of claims 1 to 9, characterized in that, The system includes: The spectral acquisition module is used to non-contactly acquire the Fourier transform infrared spectrum and Raman spectrum of the insulating oil sample of the transformer under test, and preprocess them to obtain standardized one-dimensional spectral data. The data fusion module is used to splice two spectra along the variable dimension and input them into a pre-trained neural network with a single hidden layer. Based on the correlation between each spectral variable and the degree of polymerization of the insulating paper and the collinearity between variables, the objective function is constructed for pre-training. The module outputs the variable weights and performs variable-wise weighting on the spliced ​​spectrum to achieve intermediate data fusion and generate a weighted combined spectrum. The deep feature extraction module is used to input the weighted combined spectrum into a deep residual shrinking network composed of multiple residual shrinking modules stacked together. Each module sequentially performs convolution extraction, adaptive generation of shrinking thresholds for each channel, soft thresholding noise reduction and residual connection operations based on the shrinking thresholds, and outputs the final multi-channel feature map through the final module. The aging degree regression module is used to perform regression calculations on the final multi-channel feature map through the pooling layer, flattening layer and fully connected layer, and output the predicted value of the degree of polymerization of the insulating paper. The aging status determination module is used to compare the predicted value of the degree of polymerization of the insulating paper with the preset aging threshold, determine the aging status of the current transformer insulation, and output non-contact evaluation results.

Images