Oil paper insulation thermal fault identification method based on three-dimensional fluorescence spectrum image identification

Abstract

An oil paper insulation thermal fault identification method based on three-dimensional fluorescence spectrum image identification comprises the following steps: collecting oil paper insulation samples in different thermal fault states and normal states, carrying out fluorescence spectrum detection on insulation oil in the oil paper insulation samples, and obtaining three-dimensional fluorescence spectrum data of the insulation oil; converting the fluorescence data into a three-dimensional fluorescence spectrum image; the method comprises the following steps: preprocessing a three-dimensional fluorescence spectrum image, performing feature extraction on the three-dimensional fluorescence spectrum image by utilizing an FFI-Net algorithm module, and performing feature selection and dimension reduction processing on extracted features; constructing a fault recognition model based on a support vector machine classification algorithm, and training the fault recognition model by using the training set to obtain a trained fault recognition model; and inputting the three-dimensional fluorescence spectrum image of the to-be-detected insulating oil into the trained fault identification model, and outputting a fault identification result.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment insulation state identification, and particularly to an oil-paper insulation thermal fault identification method based on three-dimensional fluorescence spectrum image recognition. Background Art

[0002] As a key insulation structure of high-voltage electrical equipment such as power transformers, the performance of the oil-paper insulation system is directly related to the safe and stable operation of the equipment and even the entire power system. During long-term operation, the oil-paper insulation material may cause thermal faults due to various factors, such as local overheating, insulation aging, etc. If these faults cannot be detected and handled in time, it may lead to a serious decline in the insulation performance of the equipment, and even cause serious accidents such as short circuits and explosions, resulting in huge economic losses and social impacts. Traditional oil-paper insulation thermal fault diagnosis methods mainly include oil chromatography analysis, electrical tests, and infrared thermal imaging. Oil chromatography analysis judges the insulation state by detecting the content and component changes of dissolved gases in transformer oil, but this method has limitations such as a long detection period and low sensitivity to early faults, and it is difficult to meet the requirements of real-time monitoring and rapid diagnosis. Electrical tests such as measuring insulation resistance and dielectric loss factor can provide certain insulation state information, but usually need to be operated with power cut, and their ability to locate and quantitatively analyze local thermal faults is limited, and they cannot comprehensively and accurately reflect the overall state of the oil-paper insulation. Infrared thermal imaging technology can visually display the temperature distribution on the surface of the equipment, but its detection results are easily affected by external environmental factors, and there are certain limitations in detecting thermal faults inside the oil-paper insulation, and it is difficult to deeply reveal the essential characteristics of the faults.

[0003] In recent years, fluorescence spectroscopy technology has gradually received attention in the fields of material composition analysis and state detection. Fluorescence spectroscopy can provide rich information on the molecular structure and chemical environment of substances. For oil-paper insulation materials, their fluorescence characteristics will change significantly with the degree of thermal aging, so it has the potential for oil-paper insulation thermal fault diagnosis. However, applying fluorescence spectroscopy technology to oil-paper insulation thermal fault identification and effectively processing it in combination with image recognition algorithms still faces many challenges. The three-dimensional fluorescence spectrum data has a complex structure and contains a large amount of redundant information. How to efficiently extract the features related to thermal faults from it is the key issue. Traditional feature extraction methods are often based on manual experience, and it is difficult to fully mine the deep feature information in the data. Moreover, when dealing with high-dimensional data, problems such as improper feature selection and poor dimensionality reduction effect are likely to occur, affecting the accuracy and efficiency of fault identification. In addition, existing image recognition algorithms have limitations in feature selection and dimensionality reduction methods when processing three-dimensional fluorescence spectrum images, and they cannot effectively map the high-dimensional feature space to a low-dimensional space, resulting in limited performance of the classification model and difficulty in meeting the requirements of high-precision identification of oil-paper insulation thermal faults in practical engineering applications.

[0004] The above information disclosed in the background section is only used to enhance the understanding of the background of the present invention, and thus may include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention

[0005] The present invention provides an oil-paper insulation thermal fault recognition method based on three-dimensional fluorescence spectroscopy image recognition. Through feature extraction and classification models, combined with feature selection and dimensionality reduction techniques, accurate recognition and diagnosis of oil-paper insulation thermal faults are achieved, so as to overcome the deficiencies of the prior art, improve the accuracy, reliability and timeliness of power equipment fault diagnosis, and provide a strong guarantee for the safe and stable operation of the power system.

[0006] An oil-paper insulation thermal fault recognition method based on three-dimensional fluorescence spectroscopy image recognition includes:

[0007] Step S1: Collect oil-paper insulation samples in different thermal fault states and normal states, perform fluorescence spectroscopy detection on the insulating oil in the oil-paper insulation samples, obtain its three-dimensional fluorescence spectroscopy data, and convert the fluorescence data into three-dimensional fluorescence spectroscopy images. The three-dimensional fluorescence spectroscopy images are divided into a training set and a test set;

[0008] Step S2: Preprocess the three-dimensional fluorescence spectroscopy images, and use the FFI-Net algorithm module to extract features from the three-dimensional fluorescence spectroscopy images. Among them, the preprocessing includes median filtering for removing noise and histogram equalization for improving image quality;

[0009] Step S3: Perform feature selection and dimensionality reduction processing on the extracted features. Among them, the random forest algorithm is used for feature selection, and the importance score of each feature is calculated , select the top k important features according to the score to form a feature subset , and then perform dimensionality reduction on the feature subset using linear discriminant analysis LDA to obtain the dimensionality-reduced feature vector ;

[0010] Step S4: Construct a fault recognition model based on the support vector machine classification algorithm and use the training set to train the fault recognition model to obtain a trained fault recognition model. The test set is used to test and verify the trained fault recognition model. If the accuracy rate of the test and verification is lower than 90%, adjust the feature extraction based on the support vector machine classification algorithm to optimize the fault recognition model until the accuracy rate of the test and verification is higher than 90%;

[0011] Step S5: Input the three-dimensional fluorescence spectroscopy image of the insulating oil to be detected into the trained fault recognition model, and output the fault recognition result.

[0012] In the described oil-paper insulation thermal fault recognition method based on three-dimensional fluorescence spectral image recognition, in step S1, different thermal faults include low-temperature overheating, medium-temperature overheating, and high-temperature overheating. Among them, the temperature range of low-temperature overheating is between 150°C and 300°C; the temperature range of medium-temperature overheating is between 300°C and 700°C; the temperature of high-temperature overheating is higher than 700°C.

[0013] In the described oil-paper insulation thermal fault recognition method based on three-dimensional fluorescence spectral image recognition, in step S1, the three-dimensional fluorescence spectral data includes a fluorescence excitation-emission matrix, and the fluorescence excitation-emission matrix is a matrix spectrum characterized by a three-dimensional coordinate of excitation wavelength - emission wavelength - fluorescence intensity.

[0014] In the described oil-paper insulation thermal fault recognition method based on three-dimensional fluorescence spectral image recognition, in step S2, the FFI-Net algorithm module includes a convolutional layer, a pooling layer, and a fully connected layer. The convolutional operation of the convolutional layer is expressed as:

[0015] ,

[0016] where, is the input feature map, is the convolutional kernel weight, is the bias term, is the activation function, is the output feature map. Through the combination of multiple convolutional layers and pooling layers, high-level semantic features in the three-dimensional fluorescence spectral image are extracted, and finally a feature vector including fluorescence intensity distribution features and wavelength-related features is obtained. .

[0017] In the described oil-paper insulation thermal fault recognition method based on three-dimensional fluorescence spectral image recognition, in step S3, the goal of linear discriminant analysis LDA is to find a projection matrix to maximize the between-class scatter and minimize the within-class scatter. Its optimization problem is expressed as:

[0018] ,

[0019] where, and are the between-class scatter matrix and the within-class scatter matrix respectively. By solving the generalized eigenvalue problem , the eigenvectors corresponding to the largest generalized eigenvalues are obtained, which constitute the projection matrix . Project the feature vector onto the low-dimensional space to obtain the feature vector after dimensionality reduction.

[0020] In the described oil-paper insulation thermal fault recognition method based on three-dimensional fluorescence spectrum image recognition, in step S4, the fault recognition model is represented by the following optimization problem:

[0021] ,

[0022] ,

[0023] where, is the weight vector, is the bias term, is the penalty parameter, is the sample label, is the feature vector after dimensionality reduction, is the slack variable. By solving the optimization problem, the support vector machine model parameters are obtained. m is the number of training samples,

[0024] The constraint conditions are:

[0025] ,

[0026] where, is the sample label, is the feature vector after dimensionality reduction,

[0027] The hyperplane equation is:

[0028] ,

[0029] For any input feature vector , if the calculated satisfies:

[0030] , then it is determined that the sample belongs to the positive class and there is an oil-paper insulation thermal fault,

[0031] , then it is determined that the sample belongs to the negative class and there is no thermal fault,

[0032] Model training: Initialize the weight vector w and the bias term b, set them to random small values or zero, and at the same time set the initial value of the penalty parameter C. Solve the optimization problem of SVM through the sequential minimal optimization algorithm, decompose the problem into multiple small optimization sub-problems, and iteratively update the optimization algorithm to reduce the objective function value until convergence or the maximum number of iterations is reached to obtain the trained SVM model parameters;

[0033] Evaluate the model performance using cross-validation, and count the model accuracy, recall rate, and F1 value. Adjust the penalty parameter C and the kernel function parameters according to the results.

[0034] In the described method for identifying thermal faults in oil-paper insulation based on three-dimensional fluorescence spectral image recognition, in step S4, for the support vector machine model, when the input is the feature vector after dimensionality reduction after that, the support vector machine model determines the class to which the sample belongs by calculating the distance to the hyperplane. The hyperplane equation is:

[0035]

[0036] where, is the weight vector of the support vector machine model, which determines the direction of the hyperplane, is the bias term, which determines the position of the hyperplane. The values of both are obtained by solving the above optimization problem.

[0037] For any arbitrarily input feature vector, if the calculated value satisfies:

[0038]

[0039] then it is determined that the sample belongs to the positive class, that is, there is a thermal fault in the oil-paper insulation; otherwise, it belongs to the negative class, that is, there is no thermal fault.

[0040] In the described method for identifying thermal faults in oil-paper insulation based on three-dimensional fluorescence spectral image recognition, when the fault recognition model is multi-class classification, the one-vs-rest strategy is adopted to construct multiple binary classification models, and the final fault type is determined by voting based on the classification results of all binary classification models.

[0041] In the described method for identifying thermal faults in oil-paper insulation based on three-dimensional fluorescence spectral image recognition, the fluorescence spectral detection device performs fluorescence spectral detection on the insulating oil in the oil-paper insulation sample. The excitation wavelength scanning band is 250 - 380 nm, and the emission wavelength scanning band is 260 - 700 nm; the slit of the fluorescence device is set to 2 mm; the scanning step size of the excitation band and the emission band is set to 2 nm; the scanning speed is set to 400 nm / min; and the integration time is set to 50 ms.

[0042] In the described method for identifying thermal faults in oil-paper insulation based on three-dimensional fluorescence spectral image recognition, the three-dimensional fluorescence spectral image includes a contour fluorescence spectrogram.

[0043] Compared with the prior art, the present invention has the following advantages:

[0044] (1) Improve the accuracy of fault identification: The present invention uses the FFI-Net algorithm to extract features from three-dimensional fluorescence spectroscopy images. This algorithm can learn the deep features in the images, including fluorescence intensity distribution features and wavelength-related features. Compared with traditional manual feature extraction methods, the FFI-Net algorithm can capture more comprehensively and accurately the complex feature information related to thermal faults in oil-paper insulation, effectively improving the representativeness and discrimination of features, and providing a more accurate basis for subsequent fault identification. In addition, by using the random forest algorithm for feature selection, the most valuable features for fault identification can be screened out according to the importance scores of features, redundant features can be removed, and the quality and discrimination of features can be improved. Further, linear discriminant analysis (LDA) is used for dimensionality reduction, which can reduce the feature dimension while retaining important feature information, reduce the computational complexity, improve the performance of the classification model, and thus improve the accuracy of fault identification. This optimized feature selection and dimensionality reduction method can better retain the effective information related to faults and remove irrelevant or redundant information, enabling the classification model to identify faults more accurately.

[0045] (2) Enhance the reliability of fault identification: The present invention uses the support vector machine classification algorithm to construct a fault identification model. The model has good classification performance and generalization ability, and can effectively handle high-dimensional data and complex classification problems. By training with a large amount of data with correct labels, the model can learn the feature patterns under different fault states, so that when facing new and unknown samples, it can accurately classify and identify, enhancing the reliability of fault identification. The support vector machine maximizes the classification margin by finding the optimal hyperplane, improving the generalization ability and robustness of the model; the random forest reduces the risk of overfitting by constructing multiple decision trees and making voting decisions, improving the stability and reliability of the model.

[0046] (3) Improve the ability of early fault detection: Three-dimensional fluorescence spectroscopy technology has high sensitivity to small changes in oil-paper insulation materials and can detect the fluorescence characteristic changes caused by early thermal faults. The present invention uses this technology combined with advanced image recognition algorithms to timely detect early thermal faults in oil-paper insulation, providing strong support for the preventive maintenance of power equipment, avoiding the further development and expansion of faults, and improving the ability of early fault detection. This high-sensitivity detection ability enables power equipment to be timely diagnosed and processed at the initial stage of faults, effectively reducing the impact of faults on the operation safety and reliability of equipment. Description of the Drawings

[0047] By reading the detailed description in the following preferred specific embodiments, various other advantages and benefits of the present invention will become clear to those of ordinary skill in the art. The accompanying drawings of the specification are only for the purpose of showing the preferred embodiments and are not considered to be a limitation of the present invention. Obviously, the drawings described below are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings based on these drawings without creative efforts. Moreover, throughout the drawings, the same reference numerals are used to represent the same components.

[0048] In the drawings:

[0049] Figure 1 is a schematic flow chart of a method for identifying thermal faults in oil-paper insulation based on three-dimensional fluorescence spectrum image recognition of the present invention;

[0050] Figure 2 are mineral insulating oils collected under normal conditions, low-temperature overheating, medium-temperature overheating, and high-temperature overheating in a specific embodiment 1 of a method for identifying thermal faults in oil-paper insulation based on three-dimensional fluorescence spectrum image recognition of the present invention;

[0051] Figure 3(a) is a three-dimensional fluorescence spectrum diagram drawn from the fluorescence data of the insulating oil under normal conditions in a specific embodiment 1 of a method for identifying thermal faults in oil-paper insulation based on three-dimensional fluorescence spectrum image recognition of the present invention;

[0052] Figure 3(b) is a three-dimensional fluorescence spectrum diagram drawn from the fluorescence data of the insulating oil after low-temperature overheating in a specific embodiment 1 of a method for identifying thermal faults in oil-paper insulation based on three-dimensional fluorescence spectrum image recognition of the present invention;

[0053] Figure 3(c) is a three-dimensional fluorescence spectrum diagram drawn from the fluorescence data of the insulating oil after medium-temperature overheating in a specific embodiment 1 of a method for identifying thermal faults in oil-paper insulation based on three-dimensional fluorescence spectrum image recognition of the present invention;

[0054] Figure 3(d) is a three-dimensional fluorescence spectrum diagram drawn from the fluorescence data of the insulating oil after high-temperature overheating in a specific embodiment 1 of a method for identifying thermal faults in oil-paper insulation based on three-dimensional fluorescence spectrum image recognition of the present invention.

[0055] The present invention will be further explained below with reference to the accompanying drawings and embodiments. Specific Embodiments

[0056] The specific embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the specific embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0057] It should be noted that in the description and claims, certain terms are used to refer to specific components. Those skilled in the art should understand that technicians may use different terms to refer to the same component. The description and claims of this specification do not distinguish components based on the differences in terms, but rather on the functional differences of the components. As mentioned throughout the description and claims, "comprising" or "including" is an open-ended term and should be interpreted as "including but not limited to". The following description of the specification is for the purpose of a preferred implementation of the present invention, but the description is based on the general principles of the specification and is not intended to limit the scope of the present invention. The protection scope of the present invention shall be subject to what is defined by the appended claims.

[0058] To facilitate the understanding of the embodiments of the present invention, the following will further explain with specific embodiments in conjunction with the accompanying drawings, and each accompanying drawing does not constitute a limitation to the embodiments of the present invention.

[0059] As Figure 1 shown in FIGS. 3(d), the oil-paper insulation thermal fault recognition method based on three-dimensional fluorescence spectrum image recognition includes the following steps:

[0060] Step S1: Collect oil-paper insulation samples in different thermal fault states and normal states, perform fluorescence spectrum detection on the insulating oil in the oil-paper insulation samples to obtain their three-dimensional fluorescence spectrum data, and convert the fluorescence data into three-dimensional fluorescence spectrum images, and divide the three-dimensional fluorescence spectrum images into a training set and a test set;

[0061] Step S2: Preprocess the three-dimensional fluorescence spectrum images, and use the FFI-Net algorithm module to extract features from the three-dimensional fluorescence spectrum images. Among them, the preprocessing includes median filtering for noise removal and histogram equalization for improving image quality;

[0062] Step S3: Perform feature selection and dimensionality reduction processing on the extracted features. Among them, the random forest algorithm is used for feature selection, and the importance score of each feature is calculated , select the top k important features according to the score to form a feature subset , and then perform dimensionality reduction on the feature subset using linear discriminant analysis LDA to obtain the dimensionality-reduced feature vector ;

[0063] Step S4: Construct a fault recognition model based on the support vector machine classification algorithm and use the training set to train the fault recognition model to obtain a trained fault recognition model. The test set is used to test and verify the trained fault recognition model. If the accuracy rate of the test and verification is lower than 90%, adjust the feature extraction based on the support vector machine classification algorithm to optimize the fault recognition model until the accuracy rate of the test and verification is higher than 90%.

[0064] Step S5: Input the three-dimensional fluorescence spectral image of the insulating oil to be detected into the trained fault recognition model, and output the fault recognition result.

[0065] In the preferred embodiment of the method for identifying thermal faults in oil-paper insulation based on three-dimensional fluorescence spectral images, in step S1, different thermal faults include low-temperature overheating, medium-temperature overheating, and high-temperature overheating, where the low-temperature overheating temperature range is between 150°C and 300°C; the medium-temperature overheating temperature range is between 300°C and 700°C; and the high-temperature overheating temperature is higher than 700°C.

[0066] In the preferred embodiment of the method for identifying thermal faults in oil-paper insulation based on three-dimensional fluorescence spectral images, in step S1, the three-dimensional fluorescence spectral data includes a fluorescence excitation-emission matrix, and the fluorescence excitation-emission matrix is a matrix spectrum characterized by the three-dimensional coordinates of excitation wavelength - emission wavelength - fluorescence intensity.

[0067] In the preferred embodiment of the method for identifying thermal faults in oil-paper insulation based on three-dimensional fluorescence spectral images, in step S2, the FFI-Net algorithm module includes a convolutional layer, a pooling layer, and a fully connected layer. The convolution operation of the convolutional layer is expressed as:

[0068] ,

[0069] where, is the input feature map, is the convolutional kernel weight, is the bias term, is the activation function, is the output feature map. Through the combination of multiple convolutional layers and pooling layers, high-level semantic features in the three-dimensional fluorescence spectral image are extracted, and finally a feature vector including fluorescence intensity distribution features and wavelength-related features is obtained . The wavelength features include wavelength bands and wavelengths.

[0070] In the preferred embodiment of the method for identifying thermal faults in oil-paper insulation based on three-dimensional fluorescence spectral images, in step S3, the goal of linear discriminant analysis LDA is to find a projection matrix to maximize the between-class scatter and minimize the within-class scatter, and its optimization problem is expressed as:

[0071] ,

[0072] Among them, and are the between-class scatter matrix and the within-class scatter matrix respectively. By solving the generalized eigenvalue problem , the eigenvectors corresponding to the largest generalized eigenvalues are obtained, which form the projection matrix . The eigenvector is projected onto the low-dimensional space to obtain the reduced-dimensional eigenvector .

[0073] In the preferred embodiment of the oil-paper insulation thermal fault recognition method based on three-dimensional fluorescence spectral image recognition, in step S4, the fault recognition model is represented by the following optimization problem:

[0074] ,

[0075] ,

[0076] Among them, is the weight vector, is the bias term, is the penalty parameter, is the sample label, is the reduced-dimensional eigenvector, is the slack variable. By solving the optimization problem, the support vector machine model parameters are obtained. m is the number of training samples,

[0077] The constraint conditions are:

[0078] ,

[0079] Among them, is the sample label, is the reduced-dimensional eigenvector,

[0080] The hyperplane equation is:

[0081] ,

[0082] For any input eigenvector , if the calculated satisfies:

[0083] , then it is determined that the sample belongs to the positive class and there is an oil-paper insulation thermal fault,

[0084] , then it is determined that the sample belongs to the negative class and there is no thermal fault,

[0085] Model training: Initialize the weight vector w and the bias term b, set them to random small values or zero, and at the same time set the initial value of the penalty parameter C. Solve the optimization problem of the SVM through the sequential minimal optimization algorithm, decompose the problem into multiple small optimization sub-problems, and iteratively update the optimization algorithm to reduce the value of the objective function until convergence or the maximum number of iterations is reached, obtaining the trained SVM model parameters;

[0086] Evaluate the model performance using cross-validation, count the model accuracy, recall rate, and F1 value, adjust the penalty parameter C and kernel function parameters according to the results, and obtain the support vector machine model parameters by solving the optimization problem, enabling the fault identification model to distinguish different fault states.

[0087] In the preferred embodiment of the oil-paper insulation thermal fault identification method based on three-dimensional fluorescence spectrum image recognition, in step S4, for the support vector machine model, when the input is the feature vector after dimensionality reduction After that, the support vector machine model determines the class to which the sample belongs by calculating the distance to the hyperplane. The hyperplane equation is:

[0088]

[0089] Among them, Is the weight vector of the support vector machine model, which determines the direction of the hyperplane, Is the bias term, which determines the position of the hyperplane. The values of both are obtained by solving the above optimization problem,

[0090] For any input feature vector, if the calculated value satisfies:

[0091]

[0092] Then it is determined that the sample belongs to the positive class, that is, there is an oil-paper insulation thermal fault, otherwise it belongs to the negative class, that is, there is no thermal fault.

[0093] In the preferred embodiment of the oil-paper insulation thermal fault identification method based on three-dimensional fluorescence spectrum image recognition, when the fault identification model is a multi-class classification, the one-vs-rest strategy is adopted to construct multiple binary classification models, and the final fault type is determined by voting based on the classification results of all binary classification models.

[0094] In the preferred embodiment of the oil-paper insulation thermal fault recognition method based on three-dimensional fluorescence spectrum image recognition, the fluorescence spectrum detection device performs fluorescence spectrum detection on the insulating oil in the oil-paper insulation sample. The excitation wavelength scanning band is 250-380 nm, the emission wavelength scanning band is 260-700 nm, the fluorescence reaction is strong, and the recognition effect is improved; the slit of the fluorescence device is set to 2 mm; the scanning step of the excitation band and the emission band is set to 2 nm; the scanning speed is set to 400 nm / min; the integration time is set to 50 ms, and the collected data is not only comprehensive but also conducive to the recognition and judgment of the support vector machine model.

[0095] In the preferred embodiment of the oil-paper insulation thermal fault recognition method based on three-dimensional fluorescence spectrum image recognition, the three-dimensional fluorescence spectrum image includes a contour fluorescence spectrum diagram.

[0096] In one embodiment, in constructing a fault recognition model based on the support vector machine classification algorithm and using the training set to train the fault recognition model to obtain a trained fault recognition model,

[0097] 1. Model construction

[0098] (1) Model selection

[0099] The present invention selects the support vector machine (SVM) as the fault recognition model. SVM is a supervised learning algorithm, which is widely used in classification and regression analysis, especially performing well in high-dimensional data classification and small sample data sets, which is in line with the characteristics of the data in the oil-paper insulation thermal fault diagnosis;

[0100] (2) Mathematical representation of the model

[0101] The constructed fault recognition model can be represented by the following optimization problem:

[0102]

[0103] Among them, is the weight vector, which determines the direction of the hyperplane;

[0104] is the bias term, which determines the position of the hyperplane;

[0105] is the penalty parameter, which is used to control the penalty degree for misclassified samples;

[0106] is the slack variable, which allows certain misclassification of some samples during classification to improve the generalization ability of the model;

[0107] m is the number of training samples.

[0108] Among them, the weight vector w and the bias term b will be learned and adjusted through an optimization algorithm, and the initial values can be small random values or zero. The penalty parameter C needs to be adjusted according to the specific dataset, and the optimal value is determined through the cross-validation method.

[0109] The constraint conditions are:

[0110]

[0111] Among them, is the sample label (positive class or negative class), is the feature vector after dimensionality reduction.

[0112] The hyperplane equation is:

[0113]

[0114] For any input feature vector , if the calculated satisfies:

[0115] , then it is determined that the sample belongs to the positive class (there is a thermal fault in oil-paper insulation)

[0116] , then it is determined that the sample belongs to the negative class (there is no thermal fault)

[0117] (3) Multi-class classification strategy

[0118] In the present invention, the model is a multi-class classification, so the one-vs-all strategy is adopted to construct multiple binary classification models. Specifically, for each fault type (such as low-temperature overheating, medium-temperature overheating, high-temperature overheating), a binary classification model is trained respectively. In the prediction stage, the sample to be detected is input into all binary classification models, and the final fault type is determined through comprehensive judgment according to the classification results of all models.

[0119] 2. Model training

[0120] (1) Training data preparation

[0121] The data processed by using the patent steps S1-S3 is used as the training set, which includes the feature vector after dimensionality reduction and the corresponding fault type labels.

[0122] (2) Parameter initialization

[0123] Initialize the weight vector w and the bias term b, set them to small random values or zero, and at the same time set the initial value of the penalty parameter C.

[0124] (3) Optimization algorithm selection

[0125] The SVM optimization problem is solved by the sequential minimal optimization (SMO) algorithm, which decomposes the problem into multiple small optimization sub-problems and solves them efficiently.

[0126] (4) Iterative Optimization

[0127] The optimization algorithm is iteratively updated to reduce the objective function value until convergence or the maximum number of iterations is reached, and the trained SVM model parameters are obtained.

[0128] (5) Model verification and adjustment

[0129] Use cross-validation to evaluate model performance, and calculate model accuracy, recall rate, F1 value, etc. According to the results, adjust hyperparameters such as C and kernel function parameters to improve model performance.

[0130] 3. Fault identification

[0131] (1) Input preprocessing

[0132] The insulating oil sample to be tested is subjected to fluorescence spectrum detection, and three-dimensional fluorescence spectrum data is obtained and converted into an image, and pre-processed by median filtering, histogram equalization and other pre-processing operations, which are consistent with steps S1 to S2 of the present invention.

[0133] (2) Feature extraction and processing

[0134] The FFI-Net algorithm is used to extract the features of the preprocessed image, including the fluorescence intensity distribution features and the wavelength-related features, and then the random forest algorithm is used for feature selection to screen the important features to form a feature subset, and finally the LDA is used for further dimensionality reduction, which is consistent with step S3 of the present invention.

[0135] (3) Model prediction

[0136] The reduced feature vector is input into the trained SVM model, and the model calculates the distance from the feature vector to the hyperplane according to the hyperplane equation to determine the category it belongs to. For multi-category classification, multiple binary classification models are used to comprehensively determine the fault type.

[0137] (4) Result output and interpretation

[0138] The model outputs fault identification results, such as "no thermal fault, low-temperature thermal fault", etc., and can also provide information such as confidence level to assist decision-making.

[0139] Example 1

[0140] A method for identifying the electrical aging stage of oil-paper insulation based on the characteristic parameters of parallel factors of three-dimensional fluorescence spectrum has the following specific steps:

[0141] S1 Collect a total of 200 groups of mineral insulating oil samples under low-temperature overheating, medium-temperature overheating, high-temperature overheating, and normal conditions. Among them, 150 groups are used as the training set, and 50 groups are used as the test set. Perform fluorescence spectroscopy detection on the insulating oil in the samples. The fluorescence spectroscopy detection device used in this embodiment is the Hitachi fluorescence spectrophotometer F-7000. The excitation wavelength scanning band is 250 - 380 nm, and the emission wavelength scanning band is 260 - 700 nm; the slit of the fluorescence device is set to 2 mm; the scanning step size of the excitation band and the emission band is set to 2 nm; the scanning speed is set to 400 nm / min; the integration time is set to 50 ms. Convert the measured three-dimensional fluorescence spectroscopy data into an image form, as shown in Figure 3(a), Figure 3(b), Figure 3(c), and Figure 3(d);

[0142] S2 Use the median filtering method and the histogram equalization method to preprocess the obtained three-dimensional fluorescence spectroscopy image, and then use the FFI-Net algorithm to extract features from the preprocessed three-dimensional fluorescence spectroscopy image. The FFI-Net algorithm includes a convolutional layer, a pooling layer, and a fully connected layer. By training to learn the feature patterns related to the thermal fault of oil-paper insulation in the image, the convolution operation of the convolutional layer can be expressed as:

[0143]

[0144] Among them, is the input feature map, is the convolutional kernel weight, is the bias term, is the activation function, is the output feature map. Through the combination of multiple convolutional layers and pooling layers, high-level semantic features in the image are extracted, and finally the feature vector is obtained. These features include fluorescence intensity distribution features and wavelength-related features;

[0145] S3 Perform feature selection and dimensionality reduction processing on the extracted features. Use the random forest algorithm for feature selection, calculate the importance score of each feature, and select the top k important features according to the score to form a feature subset , and then perform dimensionality reduction on the feature subset using linear discriminant analysis (LDA). The goal of the linear discriminant analysis is to find a projection matrix that maximizes the between-class scatter and minimizes the within-class scatter. Its optimization problem can be expressed as:

[0146]

[0147] Among them, and are the between-class scatter matrix and the within-class scatter matrix respectively. By solving the generalized eigenvalue problem , obtain The eigenvectors corresponding to the largest generalized eigenvalues form a projection matrix , project the eigenvector into a low-dimensional space to obtain the eigenvector after dimensionality reduction , so as to reduce the computational complexity and improve the classification effect;

[0148] S4 constructs a fault identification model based on the support vector machine classification algorithm and uses the training set to train the model. The fault identification model can be represented by the following optimization problem:

[0149]

[0150]

[0151] Among them, is the weight vector, is the bias term, is the penalty parameter, is the sample label, is the eigenvector after dimensionality reduction, is the slack variable. By solving this optimization problem, the support vector machine model parameters are obtained so that the model can accurately distinguish different fault states;

[0152] For the constructed fault identification model, when the input is the eigenvector after dimensionality reduction , the model determines the class to which the sample belongs by calculating the distance to the hyperplane. The hyperplane equation is:

[0153]

[0154] Among them, is the weight vector of the support vector machine model, which determines the direction of the hyperplane, is the bias term, which determines the position of the hyperplane. The values of both are obtained by solving the above optimization problem.

[0155] For any input eigenvector, if the calculated value satisfies:

[0156]

[0157] Then it is determined that the sample belongs to the positive class, that is, there is a thermal fault in the oil-paper insulation, otherwise it belongs to the negative class, that is, there is no thermal fault. In this embodiment, a one-versus-all strategy is used to construct multiple binary classification models, and voting or comprehensive judgment is performed according to the classification results of all models, and the fault types of the test set are output as: no thermal fault, low-temperature thermal fault, medium-temperature thermal fault, and high-temperature thermal fault.

[0158] The experimental test results are shown in Table 1 below:

[0159]

[0160] As shown in Table 1, in this embodiment, by drawing three-dimensional fluorescence spectra of mineral insulation after different thermal faults and using the FFI-Net feature extraction algorithm and the support vector machine classification algorithm to establish a fault recognition model, accurate judgment of different thermal fault states of oil-paper insulation is achieved, with an average accuracy rate of 94.5%. It can effectively solve the deficiencies of the existing oil-paper insulation thermal fault recognition methods, improve the accuracy, reliability and early fault detection ability of fault recognition, provide a reliable technical means for power equipment fault diagnosis, and has broad application prospects and practical engineering value.

[0161] Although the embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the above specific embodiments and application fields. The above specific embodiments are merely illustrative and guiding, rather than restrictive. Those of ordinary skill in the art can also make many forms under the inspiration of this specification and without departing from the scope protected by the claims of the present invention, and these all belong to the scope of protection of the present invention.

Claims

1. An oil-paper insulation thermal fault recognition method based on three-dimensional fluorescence spectral image recognition, characterized in that It includes the following steps: Step S1: Collect oil-paper insulation samples in different thermal fault states and normal states, perform fluorescence spectroscopy detection on the insulating oil in the oil-paper insulation samples, obtain their three-dimensional fluorescence spectrum data, and convert the fluorescence data into three-dimensional fluorescence spectrum images. The three-dimensional fluorescence spectrum images are divided into a training set and a test set; Step S2: Preprocess the three-dimensional fluorescence spectrum images, and use the FFI-Net algorithm module to extract features from the three-dimensional fluorescence spectrum images. Among them, the preprocessing includes median filtering for noise removal and histogram equalization for improving image quality; Step S3: Perform feature selection and dimensionality reduction on the extracted features. Among them, the random forest algorithm is used for feature selection, and the importance score of each feature is calculated , and the top k important features are selected according to the score to form a feature subset , and then for the feature subset , linear discriminant analysis (LDA) is used for dimensionality reduction to obtain the dimensionality-reduced feature vector ; Step S4: Construct a fault recognition model based on the support vector machine classification algorithm and use the training set to train the fault recognition model to obtain a trained fault recognition model. The test set is used to test and verify the trained fault recognition model. If the accuracy of the test and verification is lower than 90%, adjust the feature extraction based on the support vector machine classification algorithm to optimize the fault recognition model until the accuracy of the test and verification is higher than 90%; Step S5: Input the three-dimensional fluorescence spectrum image of the insulating oil to be detected into the trained fault recognition model, and output the fault recognition result.

2. The oil-paper insulation thermal fault recognition method based on three-dimensional fluorescence spectrum image recognition according to claim 1, characterized in that Preferably, in step S1, different thermal faults include low-temperature overheating, medium-temperature overheating, and high-temperature overheating. Among them, the low-temperature overheating temperature range is between 150°C and 300°C; the medium-temperature overheating temperature range is between 300°C and 700°C; the high-temperature overheating temperature is higher than 700°C.

3. A method for identifying thermal faults in oil-paper insulation based on three-dimensional fluorescence spectral image recognition according to claim 1, characterized in that, In step S1, the three-dimensional fluorescence spectrum data includes a fluorescence excitation-emission matrix, and the fluorescence excitation-emission matrix is a matrix spectrum characterized by a three-dimensional coordinate of excitation wavelength-emission wavelength-fluorescence intensity.

4. A method for identifying thermal faults of oil-paper insulation based on three-dimensional fluorescence spectral image recognition according to claim 1, characterized in that, In step S2, the FFI-Net algorithm module includes a convolutional layer, a pooling layer, and a fully connected layer. The convolutional operation of the convolutional layer is expressed as: ， Among them, is the input feature map, is the convolution kernel weight, is the bias term, is the activation function, is the output feature map. Through the combination of multiple convolutional layers and pooling layers, a three-dimensional fluorescence spectral image is extracted to obtain a feature vector including fluorescence intensity distribution characteristics and wavelength characteristics .

5. The oil-paper insulation thermal fault recognition method based on three-dimensional fluorescence spectrum image recognition according to claim 1, characterized in that, In step S3, the goal of linear discriminant analysis LDA is to find a projection matrix to maximize the between-class scatter and minimize the within-class scatter, and its optimization problem is expressed as: ， Among them, and are the between-class scatter matrix and the within-class scatter matrix respectively. By solving the generalized eigenvalue problem , we obtain eigenvectors corresponding to the largest generalized eigenvalues, which form the projection matrix . Project the eigenvector onto the low-dimensional space to obtain the eigenvector after dimensionality reduction.

6. The oil-paper insulation thermal fault recognition method based on three-dimensional fluorescence spectrum image recognition according to claim 5, characterized in that, In step S4, the fault recognition model is expressed by the following optimization problem: ， ， Among them, is the weight vector, is the bias term, is the penalty parameter, is the sample label, is the feature vector after dimensionality reduction, is the slack variable. By solving the optimization problem, the support vector machine model parameters are obtained. m is the number of training samples. The constraint condition is: ， Among them, is the sample label, is the feature vector after dimensionality reduction, The hyperplane equation is: ， For any input feature vector , if the calculated satisfies: , it is determined that the sample belongs to the positive category and there is a thermal fault in oil-paper insulation, , it is determined that the sample belongs to the negative category and there is no thermal fault, Model training: Initialize the weight vector w and the bias term b, set them to random small values or zero, and at the same time set the initial value of the penalty parameter C. Solve the optimization problem of SVM through the sequential minimal optimization algorithm, decompose the problem into multiple small optimization sub-problems, and the optimization algorithm iteratively updates to reduce the objective function value until convergence or the maximum number of iterations is reached to obtain the trained SVM model parameters; Use cross-validation to evaluate the model performance, count the model accuracy, recall rate, and F1 value, and adjust the penalty parameter C and kernel function parameters according to the results.

7. A method for identifying thermal faults of oil-paper insulation based on three-dimensional fluorescence spectrum image recognition according to claim 6, characterized in that, In the step S4, for the support vector machine model, when the input is the feature vector after dimensionality reduction the support vector machine model determines the category to which the sample belongs by calculating the distance to the hyperplane, and the hyperplane equation is: ， Among them, is the weight vector of the support vector machine model, which determines the direction of the hyperplane, is the bias term, which determines the position of the hyperplane. The values of both are obtained by solving the above optimization problem. For any input feature vector, if the calculated value satisfies: ， Then it is determined that the sample belongs to the positive class, that is, there is an oil-paper insulation thermal fault, otherwise it belongs to the negative class, that is, there is no thermal fault.

8. A method for identifying thermal faults of oil-paper insulation based on three-dimensional fluorescence spectral image recognition according to claim 7, characterized in that When the fault recognition model is a multi-class classification, adopt the one-vs-rest strategy to construct multiple binary classification models, and make a voting judgment based on the classification results of all binary classification models to determine the final fault type.

9. The oil-paper insulation thermal fault recognition method based on three-dimensional fluorescence spectrum image recognition according to claim 1, characterized in that, The fluorescence spectroscopy detection device performs fluorescence spectroscopy detection on the insulating oil in the oil-paper insulation sample. The excitation wavelength scanning band is 250 - 380 nm, and the emission wavelength scanning band is 260 - 700 nm; the slit of the fluorescence device is set to 2 mm; the scanning step of the excitation band and the emission band is set to 2 nm; the scanning speed is set to 400 nm / min; the integration time is set to 50 ms.

10. A method for identifying thermal faults of oil-paper insulation based on three-dimensional fluorescence spectral image recognition according to claim 1, characterized in that, The three-dimensional fluorescence spectroscopy image includes a contour fluorescence spectroscopy graph.

Images