CN115565028A - Transformer oil aging detection method, detection system, electronic equipment and readable storage medium - Google Patents

Abstract

The patent of the present invention discloses a transformer oil aging detection method, detection system, electronic equipment and readable storage medium. The Stacking integrated learning method is to integrate the existing hyperspectral detection algorithms and fully combine the advantages of each single learner. Improve the generalization ability and detection accuracy of integrated learning by making full use of strengths and avoiding weaknesses, and realize the optimized detection of transformer aging state through this method. At the same time, compared with other insulating oil aging detection methods, the hyperspectral detection method has outstanding advantages such as convenience, high sensitivity, and strong applicability, and can accurately reflect the aging of transformer oil. This method can be widely used in the insulation of transformers. In the scene of oil aging detection.

Description

Transformer oil aging detection method, detection system, electronic equipment and readable storage medium

technical field

The invention belongs to the technical field of transformer state detection, and in particular relates to a transformer oil aging detection method, a detection system, electronic equipment and a readable storage medium.

Background technique

As a key equipment in the power system, the normal operation of the transformer is a necessary condition for ensuring power stability and improving power quality. The life of a transformer has a great correlation with the insulation performance of its insulating material. However, as the transformer is put into operation and accumulated, the quality of the transformer oil gradually declines, and the color of the oil gradually darkens from the initial transparent color. It means that the aging has been very serious and the insulation performance is poor. As the aging degree gradually deepens, it may cause unsafe operation of the transformer, and even cause more serious faults, threatening the stability of the power system.

Many scholars at home and abroad are engaged in the research on the detection method of transformer oil aging state. The existing detection methods of transformer oil aging state include direct method and indirect method. The indirect method is to determine the aging state of transformer oil through indirect detection of the content of other substances. The content of furfural or the content of characteristic gases in oil, and the direct method is to judge the aging state by directly detecting the physical and chemical properties of transformer oil, such as spectral detection technology. Compared with other aging detection methods, spectral detection technology is widely used in the detection process of transformer oil aging because it can realize non-contact detection. Among them, hyperspectral technology has the outstanding characteristics of high spectral resolution and multiple bands. And favored. Due to the phenomenon of insulation aging, the properties of the oil have changed, and the hyperspectral technology uses the difference in the hyperspectral lines of different material components to achieve the purpose of detection. Through the detection of transformer oil aging, the working status of the transformer can be grasped in time, which is convenient for subsequent maintenance and repair operations.

Contents of the invention

In order to solve the problems in the above solutions, the present invention provides a transformer oil aging detection method, detection system, electronic equipment and readable storage medium, which are used to monitor and detect the aging state of transformers, so as to facilitate timely grasp of the operating state and Arrange maintenance and other follow-up operations.

The purpose of the present invention can be achieved through the following technical solutions:

Provide a kind of transformer oil aging detection method, it comprises the following steps:

S1. Obtain hyperspectral image data of insulating oil samples with different aging degrees;

S2. Perform feature selection and processing operations on the obtained hyperspectral image data;

S3. Use the processed data to build a Stacking integrated learning model, detect the oil sample to be tested, input the obtained characteristic data into the built Stacking integrated learning model, and obtain the final prediction result, which can be used to detect the condition of the transformer oil.

Further, step S1 includes the following steps: making the insulating oil undergo accelerated aging in a constant temperature aging box, and preparing insulating oil samples of different aging degrees according to different aging times;

Further, step S2 includes the following steps:

S31: Using a hyperspectral imager to collect images of the sample to be tested to obtain original hyperspectral image data of the sample;

S32: Perform preprocessing operations on the obtained hyperspectral line data information, including operations such as black and white correction, multivariate scattering correction, and smoothing and denoising.

Further, it specifically includes the following steps:

S41: Randomly select a certain sample to establish a PLS model, and calculate the absolute value of the regression coefficient and the weight of each band point;

S42: Use EDP and ARS to select variables, and calculate the root mean square error value of cross-validation;

S43: Select the Monte Carlo times, compare 100 RMSECVs, and select the smallest subset among them as the optimal feature variable.

Further, the operation of feature selection and processing in step S3 is mainly to perform dimensionality reduction processing on the extracted feature quantities, and the steps of PCA dimensionality reduction processing include:

S51: Calculate the covariance matrix of n characteristic data sets and solve the eigenvalue and eigenvector of the matrix;

S52: Sort the eigenvalues from large to small, and select the eigenvectors corresponding to the largest k features to form a mapping matrix, thereby reducing the dimensionality of the original n features into k features.

Further, step S3 includes the following steps:

S61: Divide the preprocessed hyperspectral lines into a training set and a test set according to a ratio of 8:2, and train the model;

S62: Select decision tree, SVM, and ANN as the base learner. Due to the strong generalization ability of random forest, it is selected as the meta-classifier for fusion model training. The main features are extracted by the base learner, and then fused by the meta-learner. Generate aging prediction results;

The decision tree of the base learner in the step S62 specifically includes two stages, generation and pruning, and the pruning is realized by optimizing the loss function. Let the number of leaf nodes of the tree T be |T|, and t is the tree The leaf node of T, the leaf node has N _t sample points, and the number of sample points of class k is N _tk , k=1,2,...,K,H _t (T) is the information entropy of the leaf node , α is a parameter and (α≥0), the decision tree loss function can be defined as (the number of sample points of each leaf node is multiplied by the information entropy of the node, and finally a regular term is added):

可将其转化为求解带约束条件的最小值问题：, in the step S62, the specific implementation steps of seeking the best classification hyperplane about the base learner SVM are as follows: assume that the training samples Di=(xi, yi), i=1,2,...n, yi= {1,-1}, where xi represents the source data, yi represents the category value, the hyperplane is w*x+b=0, and the geometric interval between the hyperplane and the sample point is defined as

This can be transformed into solving a constrained minimum problem:

st.y _i [w(x _i )+b]-1≥0, i=1,2,...l

其中i代表第i个样本数据，j代表第j类，yi代表第i个样本的真实类别，通常的，s_j＝x_iw_j，将损失值推广至整个训练集可知，

The calculation method of the loss function is

Where i represents the i-th sample data, j represents the j-th class, and yi represents the true category of the i-th sample. Usually, s _j = x _i w _j , and the loss value can be extended to the entire training set.

作为激活函数，将其损失函数定义为输出向量的均方误差，即：

其中

表示输出向量第i个元素与标签第i个元素的差值。In the step S62, the ANN structure of the base learner includes an input layer, a hidden layer and an output layer, select

As an activation function, its loss function is defined as the mean square error of the output vector, namely:

in

Indicates the difference between the i-th element of the output vector and the i-th element of the label.

Further, step S62 includes the following steps:

S71: Input the training set into different base models for training, and obtain corresponding prediction results on the test set;

S72: Apply the prediction result of the base model as a feature to the meta-model, and input the average prediction value of the base model in the test set as the test set of the meta-model into the meta-model for training and fitting, and train a final aging detection model.

Further, the formula for black and white correction is:

In the formula, T _A is the hyperspectral image data after correction, T is the hyperspectral image data before correction, T _B is the standard black image data, and T _W is the standard white image data.

Further, the specific implementation method of multivariate scattering correction is:

S91: obtain the average spectral data value:

S92: Perform a linear regression of each spectrum with the mean spectrum:

S93: Corrected spectral data:

A transformer oil aging detection electronic device is provided, which includes:

a memory storing executable instructions; and

A processor configured to execute executable instructions in the memory to implement the method for aging transformer oil based on Stacking integrated learning according to any one of claims 1-9.

A readable storage medium is provided, on which executable instructions are stored, and when the executable instructions are executed by a processor, a method for aging transformer oil based on Stacking integrated learning is realized.

A transformer oil aging detection system is provided, which includes:

The hyperspectral image data acquisition module is used to acquire hyperspectral image data of insulating oil samples with different aging degrees;

The feature selection and processing operation module is used to perform feature selection and processing operations on the obtained hyperspectral image data;

The detection module uses the processed data to build a Stacking integrated learning model, detects the oil samples to be tested, and inputs the obtained characteristic data into the built Stacking integrated learning model to obtain the final prediction result, which can be used to detect the transformer oil. situation.

Further, the detection module includes:

The training set is divided into submodules, which are used to divide the preprocessed hyperspectral lines into a training set and a test set according to a ratio of 8:2;

The model training sub-module is used to input the training set into different base models for training, and obtain corresponding prediction results on the test set; apply the prediction results of the base model as features to the meta-model, and use the base model in the test set The average predicted value is input into the meta-model as the test set of the meta-model for training and fitting, and the final aging detection model is trained;

The aging prediction result generation sub-module is used to select decision tree, SVM, ANN as the base learner. Due to the strong generalization ability of random forest, it is selected as the meta-classifier for fusion model training, and the main features are extracted by the base learner, and then Fusion by meta-learners to finally generate aging prediction results;

The base learner decision tree specifically includes two stages, generation and pruning. The pruning is realized by optimizing the loss function. Let the number of leaf nodes of the tree T be |T|, t is the leaf node of the tree T, The leaf node has N _t sample points, wherein the number of sample points of class k is N _tk , k=1,2,...,K,H _t (T) is the information entropy of the leaf node, α is a parameter and (α≥0), the decision tree loss function can be defined as (the number of sample points of each leaf node is multiplied by the information entropy of the node, and finally the regular term is added):

可将其转化为求解带约束条件的最小值问题：The specific implementation steps for the base learner SVM to seek the best classification hyperplane are as follows: Assuming training samples Di=(xi,yi),i=1,2,...n,yi={1,-1}, Among them, xi represents the source data, yi represents the category value, the hyperplane is w*x+b=0, and the geometric interval between the hyperplane and the sample point is defined as

This can be transformed into solving a constrained minimum problem:

st.y _i [w(x _i )+b]-1≥0, i=1,2,...l

其中i代表第i个样本数据，j代表第j类，yi代表第i个样本的真实类别，通常的，s_j＝x_iw_j，将损失值推广至整个训练集可知，

The calculation method of the loss function is

Where i represents the i-th sample data, j represents the j-th class, and yi represents the true category of the i-th sample. Usually, s _j = x _i w _j , and the loss value can be extended to the entire training set.

作为激活函数，将其损失函数定义为输出向量的均方误差，即：

其中

表示输出向量第i个元素与标签第i个元素的差值。The basic learner ANN structure includes input layer, hidden layer and output layer, select

As an activation function, its loss function is defined as the mean square error of the output vector, namely:

in

Indicates the difference between the i-th element of the output vector and the i-th element of the label.

Compared with the detection technology of existing transformer oil aging, the effective benefit of the present invention is: the first, can realize the on-line detection to transformer oil aging; difference, improve the learning ability and generalization ability of the model; 3. The research on the aging of transformer oil aims to deeply study the relationship between the microcosmic and macroscopic aspects of the aging process, reveal the aging mechanism of transformer oil, and provide a scientific basis for the evaluation of the aging state of transformer oil. Fourth, the application of hyperspectral technology in transformer oil state detection can provide theoretical reference for other research and study.

Description of drawings

In order to illustrate the embodiments or technical solutions of the present invention more clearly, the following will briefly introduce the accompanying drawings used in the embodiments or technical solutions. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those skilled in the art, other drawings can also be obtained based on these drawings without creative effort.

Fig. 1 is a schematic diagram of a transformer oil aging detection method in the present invention.

detailed description

The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

In the specific implementation process, as shown in Figure 1, the transformer oil aging detection method includes the following steps:

S1. Obtain hyperspectral image data of insulating oil samples with different aging degrees;

S2. Perform feature selection and processing operations on the obtained hyperspectral image data;

S3. Use the processed data to build a Stacking integrated learning model, detect the oil sample to be tested, input the obtained characteristic data into the built Stacking integrated learning model, and obtain the final prediction result, which can be used to detect the condition of the transformer oil.

Step S1 includes the following steps: making the insulating oil undergo accelerated aging in a constant temperature aging box, and preparing insulating oil samples of different aging degrees according to different aging times;

Step S2 following steps:

S31: Using a hyperspectral imager to collect images of the sample to be tested to obtain original hyperspectral image data of the sample;

S32: Perform preprocessing operations on the obtained hyperspectral line data information, including operations such as black and white correction, multivariate scattering correction, and smoothing and denoising.

Specifically include the following steps:

S41: Randomly select a certain sample to establish a PLS model, and calculate the absolute value of the regression coefficient and the weight of each band point;

S42: Use EDP and ARS to select variables, and calculate the root mean square error value of cross-validation;

S43: Select the Monte Carlo times, compare 100 RMSECVs, and select the smallest subset among them as the optimal feature variable.

The operation of feature selection and processing in step S3 is mainly to perform dimensionality reduction processing on the extracted feature quantities, and the steps of PCA dimensionality reduction processing include:

S51: Calculate the covariance matrix of n characteristic data sets and solve the eigenvalue and eigenvector of the matrix;

S52: Sort the eigenvalues from large to small, and select the eigenvectors corresponding to the largest k features to form a mapping matrix, thereby reducing the dimensionality of the original n features into k features.

Step S3 comprises the following steps:

S61: Divide the preprocessed hyperspectral lines into a training set and a test set according to a ratio of 8:2, and train the model;

S62: Select decision tree, SVM, and ANN as the base learner. Due to the strong generalization ability of random forest, it is selected as the meta-classifier for fusion model training. The main features are extracted by the base learner, and then fused by the meta-learner. Generate aging prediction results;

The decision tree of the base learner in the step S62 specifically includes two stages, generation and pruning, and the pruning is realized by optimizing the loss function. Let the number of leaf nodes of the tree T be |T|, and t is the tree The leaf node of T, the leaf node has N _t sample points, and the number of sample points of class k is N _tk , k=1,2,...,K,H _t (T) is the information entropy of the leaf node , α is a parameter and (α≥0), the decision tree loss function can be defined as (the number of sample points of each leaf node is multiplied by the information entropy of the node, and finally a regular term is added):

可将其转化为求解带约束条件的最小值问题：, in the step S62, the specific implementation steps of seeking the best classification hyperplane about the base learner SVM are as follows: assume that the training samples Di=(xi, yi), i=1,2,...n, yi= {1,-1}, where xi represents the source data, yi represents the category value, the hyperplane is w*x+b=0, and the geometric interval between the hyperplane and the sample point is defined as

This can be transformed into solving a constrained minimum problem:

st.y _i [w(x _i )+b]-1≥0, i=1,2,...l

其中i代表第i个样本数据，j代表第j类，yi代表第i个样本的真实类别，通常的，s_j＝x_iw_j，将损失值推广至整个训练集可知，

The calculation method of the loss function is

Where i represents the i-th sample data, j represents the j-th class, and yi represents the true category of the i-th sample. Usually, s _j = x _i w _j , and the loss value can be extended to the entire training set.

作为激活函数，将其损失函数定义为输出向量的均方误差，即：

其中

表示输出向量第i个元素与标签第i个元素的差值。In the step S62, the ANN structure of the base learner includes an input layer, a hidden layer and an output layer, select

As an activation function, its loss function is defined as the mean square error of the output vector, namely:

in

Indicates the difference between the i-th element of the output vector and the i-th element of the label.

Step S62 comprises the following steps:

S71: Input the training set into different base models for training, and obtain corresponding prediction results on the test set;

S72: Apply the prediction result of the base model as a feature to the meta-model, and input the average prediction value of the base model in the test set as the test set of the meta-model into the meta-model for training and fitting, and train a final aging detection model.

The formula for black and white correction is:

In the formula, T _A is the hyperspectral image data after correction, T is the hyperspectral image data before correction, T _B is the standard black image data, and T _W is the standard white image data.

The specific implementation method of multivariate scattering correction is as follows:

S91: obtain the average spectral data value:

S92: Perform a linear regression of each spectrum with the mean spectrum:

S93: Corrected spectral data:

The transformer oil aging detection electronic equipment includes:

a memory storing executable instructions; and

A processor configured to execute executable instructions in the memory to implement the method for aging transformer oil based on Stacking integrated learning according to any one of claims 1-9.

The readable storage medium has executable instructions stored thereon, and when the executable instructions are executed by a processor, the transformer oil aging method based on Stacking integrated learning is realized.

The transformer oil aging detection system includes:

The hyperspectral image data acquisition module is used to acquire hyperspectral image data of insulating oil samples with different aging degrees;

The feature selection and processing operation module is used to perform feature selection and processing operations on the obtained hyperspectral image data;

The detection module uses the processed data to build a Stacking integrated learning model, detects the oil samples to be tested, and inputs the obtained characteristic data into the built Stacking integrated learning model to obtain the final prediction result, which can be used to detect the transformer oil. situation.

Detection modules include:

The training set is divided into submodules, which are used to divide the preprocessed hyperspectral lines into a training set and a test set according to a ratio of 8:2;

The model training sub-module is used to input the training set into different base models for training, and obtain corresponding prediction results on the test set; apply the prediction results of the base model as features to the meta-model, and use the base model in the test set The average predicted value is input into the meta-model as the test set of the meta-model for training and fitting, and the final aging detection model is trained;

The aging prediction result generation sub-module is used to select decision tree, SVM, ANN as the base learner. Due to the strong generalization ability of random forest, it is selected as the meta-classifier for fusion model training, and the main features are extracted by the base learner, and then Fusion by meta-learners to finally generate aging prediction results;

The base learner decision tree specifically includes two stages, generation and pruning. The pruning is realized by optimizing the loss function. Let the number of leaf nodes of the tree T be |T|, t is the leaf node of the tree T, The leaf node has N _t sample points, wherein the number of sample points of class k is N _tk , k=1,2,...,K,H _t (T) is the information entropy of the leaf node, α is a parameter and (α≥0), the decision tree loss function can be defined as (the number of sample points of each leaf node is multiplied by the information entropy of the node, and finally the regular term is added):

可将其转化为求解带约束条件的最小值问题：The specific implementation steps for the base learner SVM to seek the best classification hyperplane are as follows: Assuming training samples Di=(xi,yi),i=1,2,...n,yi={1,-1}, Among them, xi represents the source data, yi represents the category value, the hyperplane is w*x+b=0, and the geometric interval between the hyperplane and the sample point is defined as

This can be transformed into solving a constrained minimum problem:

st.y _i [w(x _i )+b]-1≥0, i=1,2,...l

其中i代表第i个样本数据，j代表第j类，yi代表第i个样本的真实类别，通常的，s_j＝x_iw_j，将损失值推广至整个训练集可知，

The calculation method of the loss function is

Where i represents the i-th sample data, j represents the j-th class, and yi represents the true category of the i-th sample. Usually, s _j = x _i w _j , and the loss value can be extended to the entire training set.

作为激活函数，将其损失函数定义为输出向量的均方误差，即：

其中

表示输出向量第i个元素与标签第i个元素的差值。The basic learner ANN structure includes input layer, hidden layer and output layer, select

As an activation function, its loss function is defined as the mean square error of the output vector, namely:

in

Indicates the difference between the i-th element of the output vector and the i-th element of the label.

In the specific implementation process, the Stacking integration model process includes the following steps:

S1: Divide the preprocessed hyperspectral lines into a training set and a test set in a ratio of 8:2;

S2: The decision tree, SVM, and ANN are used as the base learner respectively. Due to the strong generalization ability of the random forest, it is selected as the meta-classifier for fusion model training. The main features are extracted by the base learner, and then fused by the meta-learner. Finally, the aging prediction result is generated;

Specifically, the model training process includes the following steps:

S1: Input the training set into different base models for training, and obtain corresponding prediction results on the test set;

S2: Input the features obtained from the base model as the training set into the meta-model, and use the average predicted value of the base model in the test set as the test set of the meta-model into the meta-model for training and fitting to obtain the final aging detection result.

In particular, the coefficient of determination, root mean square error, and mean absolute error are selected as evaluation indicators to analyze and evaluate the detection situation of each model, and to verify whether the predictive ability of the integrated model is better than that of a single classifier.

Claims (13)

1. The transformer oil aging detection method is characterized by comprising the following steps:

s1, acquiring hyperspectral image data of insulating oil samples with different aging degrees;

s2, performing feature selection and processing operation on the acquired hyperspectral image data;

and S3, constructing a Stacking integrated learning model by using the processed data, detecting the oil sample to be detected, inputting the obtained characteristic data into the constructed Stacking integrated learning model, and obtaining a final prediction result, namely the final prediction result can be used for detecting the condition of the transformer oil.

2. The transformer oil aging detection method according to claim 1, wherein the step S1 comprises the following steps: and (3) carrying out accelerated aging on the insulating oil in a constant-temperature aging box, and preparing insulating oil samples with different aging degrees according to different aging times.

3. The transformer oil aging detection method according to claim 1, wherein the step S2 comprises the following steps:

s31: acquiring an image of a sample to be detected by using a hyperspectral imager to obtain original hyperspectral image data of the sample;

s32: and carrying out preprocessing operation on the obtained hyperspectral spectral line data information, wherein the preprocessing operation comprises operations of black and white correction, multivariate scattering correction, smoothing and drying removal and the like.

4. The transformer oil aging detection method according to claim 1, characterized by comprising the following steps:

s41: randomly extracting a certain sample to establish a PLS model, and calculating the absolute value of a regression coefficient and the weight of each waveband point;

s42: carrying out variable selection by using the EDP and the ARS, and calculating a root mean square error value of cross validation;

s43: and selecting the Monte Carlo times, comparing 100 RMSECVs, and selecting the minimum subset as the optimal characteristic variable.

5. The transformer oil aging detection method according to claim 1, wherein the operation of feature selection and processing in step S3 is mainly to perform dimensionality reduction on the extracted feature quantity, and the step of PCA dimensionality reduction includes:

s51: calculating covariance matrixes of the n feature data sets and solving eigenvalues and eigenvectors of the matrixes;

s52: sorting the eigenvalues from big to small, selecting the eigenvector corresponding to the largest k characteristics to form a mapping matrix, and thus reducing the dimension of the original n characteristics to convert the original n characteristics into k characteristics.

6. The transformer oil aging detection method according to claim 1, wherein the step S3 comprises the following steps:

s61: dividing the preprocessed hyperspectral spectral lines into a training set and a testing set according to the proportion of 8;

s62: selecting a decision tree, an SVM and an ANN as a base learner, selecting the decision tree, the SVM and the ANN as a meta classifier for carrying out fusion model training due to strong generalization capability of random forests, extracting main characteristics by the base learner, fusing by the meta learner, and finally generating an aging prediction result;

the decision tree of the base learner in step S62 specifically includes two stages, generation and pruning, where the pruning is implemented by optimizing a loss function, where the number of leaf nodes of the tree T is set to | T |, T is a leaf node of the tree T, and the leaf node has N _t A number of sample points of class k being N _tk K =1,2, \8230;, K, H _t (T) is the information entropy of the leaf node, α is a parameter and (α ≧ 0), the decision tree loss function can be defined as (the number of sample points per leaf node is multiplied by the information entropy of the node, and finally a regularization term is added):

the specific implementation step of seeking the optimal classification hyperplane for the base learner SVM in step S62 is as follows, assuming that training samples Di = (xi, yi), i =1,2,.. N, yi = {1, -1}, where xi represents source data, yi represents a class value, and the hyperplane is w × + b =0, the method is implemented by using training samples Di = (xi, yi), i =1,2, and/or n, yi = {1, -1}, where x represents a source data, y represents a class value, and the hyperplane is defined as w × + b =0The geometric interval between the pseudo-hyperplane and the sample point is

This can be translated into solving the minimum problem with constraints:

st.y _i [w(x _i )+b]-1≥0,i＝1,2,......l

the loss function is calculated by

Where i represents the ith sample data, j represents the jth class, y _i Representing the real class of the ith sample, usually, s _j ＝x _i w _j The loss value is popularized to the whole training set to be known,

in step S62, the ANN structure of the base learner is selected from an input layer, a hidden layer and an output layer

As an activation function, its loss function is defined as the mean square error of the output vector, i.e.:

wherein

Representing the difference between the ith element of the output vector and the ith element of the tag.

7. The transformer oil aging detection method according to claim 6, wherein the step S62 comprises the following steps:

s71: inputting the training set into different base models for training, and obtaining corresponding prediction results on the test set;

s72: and applying the prediction result of the base model as a characteristic to the meta-model, inputting the average prediction value of the base model in the test set as the test set of the meta-model into the meta-model for training and fitting, and training a final aging detection model.

8. The transformer oil aging detection method according to claim 3, wherein the formula of black and white correction is as follows:

in the formula T _A For the corrected hyperspectral image data, T is the hyperspectral image data before correction, T _B Is standard black image data, T _W Is standard white image data.

9. The transformer oil aging detection method according to claim 3, wherein the multivariate scattering correction is specifically realized by the following steps:

s91: obtaining an average spectral data value:

s92: a unary linear regression of each spectrum with the average spectrum was performed:

s93: corrected spectral data:

10. the utility model provides a transformer oil aging testing electronic equipment which characterized in that includes:

a memory storing executable instructions; and

a processor configured to execute the executable instructions in the memory to implement the Stacking ensemble learning based transformer oil aging method of any of claims 1-9.

11. A readable storage medium having stored thereon executable instructions, wherein the executable instructions when executed by a processor implement the Stacking ensemble learning based transformer oil aging method according to any one of claims 1 to 9.

12. A transformer oil aging detection system, comprising:

the hyperspectral image data acquisition module is used for acquiring hyperspectral image data of insulating oil samples with different aging degrees;

the characteristic selection and processing operation module is used for carrying out characteristic selection and processing operation on the acquired hyperspectral image data;

and the detection module is used for constructing a Stacking integrated learning model by using the processed data, detecting the oil sample to be detected, inputting the obtained characteristic data into the constructed Stacking integrated learning model, and obtaining a final prediction result, namely the condition of the transformer oil can be detected.

13. The transformer oil degradation detection system of claim 12, wherein the detection module comprises:

the training set division submodule is used for dividing the preprocessed hyperspectral spectral lines into a training set and a test set according to the proportion of 8;

the model training submodule is used for inputting the training set into different base models for training and obtaining corresponding prediction results on the test set; applying the prediction result of the base model as a characteristic to the meta-model, inputting the average prediction value of the base model in the test set as the test set of the meta-model into the meta-model for training and fitting, and training a final aging detection model;

the aging prediction result generation submodule is used for selecting a decision tree, an SVM and an ANN as a base learning device, the random forest is selected as a meta classifier to perform fusion model training due to the strong generalization capability of the random forest, main characteristics are extracted by the base learning device, the meta learning device performs fusion, and an aging prediction result is finally generated;

the decision tree of the base learner specifically comprises two stages, namely generation and pruning, wherein the pruning is realized by optimizing a loss function, the number of leaf nodes of the tree T is set to be | T |, T is the leaf node of the tree T, and the leaf node has N _t A number of sample points of class k being N _tk K =1,2, \ 8230;, K, H _t (T) is the information entropy of the leaf node, α is a parameter and (α ≧ 0), the decision tree loss function can be defined as (the number of sample points per leaf node is multiplied by the information entropy of the node, and finally a regularization term is added):

the implementation of the base learner SVM to find the best classification hyperplane is performed by assuming that training samples Di = (xi, yi), i =1, 2.... N, yi = {1, -1}, where xi represents source data, yi represents class value, hyperplane is w x + b =0, and the geometric separation of the hyperplane from the sample points is defined as w = (x, y) } where x represents source data, and where y represents class value, and where w + x + b =0

This can be translated into solving the minimum problem with constraints:

st.y _i [w(x _i )+b]-1≥0,i＝1,2,......l

the loss function is calculated by

Where i represents the ith sample data, j represents the jth class, yi represents the true class of the ith sample, and usually, s _j ＝x _i w _j The loss value is popularized to the whole training set to be known,

the ANN structure of the base learner includes an input layer, a hidden layer and an output layer, and selects

As an activation function, its loss function is defined as the mean square error of the output vector, i.e.:

wherein

Representing the difference of the ith element of the output vector and the ith element of the label.

Images