A transformer insulating oil fault detection method based on oil classification

Through the method based on oil classification, using k-meams clustering and random forest model, the prediction intervals are divided and the main prediction model is constructed, which solves the problem of a single prediction model adapting to the changes in insulation oil of different components, and improves the detection of insulation oil faults of transformer accuracy and efficiency.

CN118091280BActive Publication Date: 2024-11-22SOUTHWEST UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410150140.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-02-02
Publication Date
2024-11-22
Estimated Expiration
2044-02-02

AI Technical Summary

Technical Problem

In the prior art, in the detection of transformer insulation oil failure, a single prediction model cannot adapt to the changes in insulation oil of different components, resulting in a decrease in prediction accuracy, and designing models for each insulation oil will cause waste of computing power.

Method used

Using an oil-based classification method, the prediction interval is divided through the k-meams clustering algorithm, the main prediction model is constructed using a random forest model, the number of prediction models is reduced and the accuracy is improved, and the fault characteristics are obtained by combining multi-frequency ultrasonic detection and chromatographic analysis, and the fault characteristics are calculated. Predict trust and dynamically adjust interval relationships.

Benefits of technology

It significantly improves the accuracy of insulating oil fault monitoring of different components, reduces the number of prediction models, improves the efficiency and accuracy of monitoring, and provides comprehensive judgment support for predictive trust.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118091280B_ABST
    Figure CN118091280B_ABST
Patent Text Reader

Abstract

The present invention relates to the technical field of transformer fault detection, and in particular to a transformer insulating oil fault detection method based on oil classification. The present invention adopts the form of prediction interval division. For a main prediction interval and several slave prediction intervals under its jurisdiction, the same main prediction model can be used to complete the diagnosis and detection of insulating oil fault characteristics, which significantly reduces the number of prediction models and improves the accuracy of prediction. For the insulating oil in each slave prediction interval, in addition to providing the diagnosis result, it also includes the confidence of the structure, which is convenient for the staff to make a comprehensive judgment. During use, the present invention continuously divides the main and slave prediction intervals, which can gradually increase the accuracy of the prediction, and at the same time avoids a large number of prediction models. It ensures the accuracy of each prediction model and a wide range of use.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of transformer fault detection, and in particular to a transformer insulating oil fault detection method based on oil classification. Background Art

[0002] As the core equipment of the power system, transformers may experience faults such as partial discharge during long-term operation. How to monitor and ensure the safe operation of transformers has become one of the core tasks of the power system.

[0003] After long-term observation, it is found that when a transformer fails, the following fault characteristics may occur: increased oil density, increased viscosity, increased acid value, changes in water content, increased acetylene in the oil, increased methane, etc. This provides an idea for transformer fault monitoring, and early warning of transformer faults can be given by monitoring changes in fault characteristics.

[0004] The monitoring methods of fault characteristics have been developed and mature, including chromatography, multi-frequency ultrasonic detection, etc. With the popularity of computer deep learning algorithms, deep learning algorithms are also being used for the analysis and diagnosis of fault characteristics.

[0005] The shortcomings of the existing technology are: 1. Currently, most of them use a single prediction model for a certain type of insulating oil, or for multiple types of insulating oil; however, the components of insulating oils are not completely consistent. Even the same insulating oil will cause changes in components when operating in different environments. After the components of the insulating oil change, its fault characteristics will also change accordingly. A single prediction model cannot fit insulating oils of different components, which will lead to a decrease in prediction accuracy. 2. If a prediction model is designed for each type of insulating oil, there will be high computing power requirements, resulting in a waste of computing power. Summary of the invention

[0006] The invention discloses a transformer insulating oil fault detection method based on oil classification, which can significantly improve the accuracy of fault monitoring of insulating oils of different components under the condition of building a small number of prediction models.

[0007] It is achieved through such a technical solution, the specific method is as follows:

[0008] Obtain the component content and fault characteristics of the insulating oil to be tested;

[0009] Determine the prediction interval of the insulating oil to be tested based on the component content of the insulating oil to be tested;

[0010] According to the prediction interval of the insulating oil to be tested, determining the main prediction interval to which the insulating oil to be tested belongs;

[0011] The fault characteristics of the insulating oil to be tested are input into the main prediction model constructed by the main prediction interval to which it belongs, and the main prediction model outputs the diagnosis result and prediction confidence of the insulating oil to be tested.

[0012] The advantage of this embodiment is that, for a main prediction interval and several slave prediction intervals under its jurisdiction, the same main prediction model can be used to complete the diagnosis and detection of insulating oil fault characteristics, which significantly reduces the number of prediction models and improves the accuracy of the prediction; for the insulating oil in each slave prediction interval, in addition to providing the diagnosis results, it also includes the confidence of the structure, which is convenient for the staff to make a comprehensive judgment.

[0013] Furthermore, the fault characteristics include: water content, dielectric loss, surface tension, breakdown voltage, acid value, volume resistivity, viscosity and density.

[0014] Furthermore, the fault characteristics of the insulating oil to be tested are obtained, including multi-frequency ultrasonic detection and chromatographic analysis.

[0015] Furthermore, according to the content of insulating oil components, the method of dividing the prediction interval is as follows:

[0016] Obtaining a number of insulating oils with different component contents as sample insulating oils;

[0017] For each sample insulating oil, a test voltage is applied for a preset time to obtain several fault characteristics of each sample insulating oil;

[0018] A characteristic sample is constructed with all the fault characteristics of each sample insulating oil;

[0019] The k-means clustering algorithm is used to cluster the characteristic samples and divide them into several cluster intervals, each of which includes several insulating oils with different component contents.

[0020] The advantage of this embodiment is that by applying the same test voltage, the characteristic changes of different insulating oils can be highlighted, and the same change characteristics can be clustered into one category. Fault monitoring can be achieved through a single prediction model, thereby reducing the number of prediction models while ensuring monitoring accuracy.

[0021] Furthermore, the method for selecting the main prediction interval is: selecting the interval containing the largest amount of insulating oil with different component contents as the main prediction interval, and the remaining intervals as the slave prediction intervals.

[0022] The advantage of this embodiment is that there are more possibilities for using the prediction intervals containing a large number of prediction intervals, and using them as the main prediction intervals can increase the range of high-accuracy predictions of the main prediction model.

[0023] Furthermore, the main prediction model is a random forest model, and the training method is:

[0024] Obtain historical samples of insulating oil within the main prediction interval and use the historical samples to train the random forest model.

[0025] Furthermore, the calculation method of the predicted trust is as follows:

[0026] Get the center position of the main prediction interval, calculate the Euclidean distance from the center position of the main prediction interval to each sample insulating oil, and take the farthest Euclidean distance as the maximum boundary length of the main prediction interval;

[0027] Get the center positions of the remaining slave prediction intervals, calculate the Euclidean distance from the center position of the slave prediction interval to the center position of the main prediction interval, and divide the maximum boundary length of the main prediction interval by the Euclidean distance from the center position of the prediction interval to the center position of the main prediction interval as the prediction confidence of the sample insulating oil in the slave prediction interval.

[0028] The advantage of this embodiment is that, in the absence of actual use verification, the ratio of the Euclidean distance is used as the basis for calculating the prediction confidence, and the prediction confidence is associated with the distance from the slave interval to the main interval, which can increase the accuracy of the prediction confidence.

[0029] Furthermore, during the fault detection process,

[0030] If the prediction confidence is lower than the second threshold, the affiliation between the slave prediction interval and the master prediction interval is deleted;

[0031] If the prediction confidence is higher than the first threshold, the slave prediction interval will be merged into the master prediction interval.

[0032] The advantage of this embodiment is that the Euclidean distance can reflect some characteristics of the prediction confidence, but cannot completely replace the prediction confidence. Therefore, the accuracy of fault monitoring can be further increased through subsequent use and correction.

[0033] Furthermore, the prediction confidence from the prediction interval is corrected as follows:

[0034] Inputting the fault characteristics of the insulating oil to be tested belonging to the secondary prediction interval into the primary prediction model to obtain the prediction diagnosis result;

[0035] The number of correct prediction results is divided by the total number of predictions to obtain the prediction accuracy rate;

[0036] The middle value between the prediction accuracy and the prediction confidence of the prediction interval before correction is taken as the prediction confidence after correction and update.

[0037] The advantage of this embodiment is that a correction method is proposed, in which the intermediate value between the predicted value and the actual value is used as the correction value, which takes into account the correction speed and avoids the situation of over-correction.

[0038] If the prediction confidence is lower than the second threshold, delete the affiliation between the slave prediction interval and the master prediction interval, and then perform the following operations:

[0039] Re-cluster all unaffiliated slave prediction intervals to divide the master and slave prediction intervals;

[0040] Construct a forecast model for the new primary forecast interval;

[0041] Calculate the forecast confidence of the new slave forecast interval to the master forecast interval.

[0042] The advantage of this embodiment is that by continuously dividing the main and slave prediction intervals in this way, the accuracy of the prediction can be gradually increased during use, while also avoiding a large increase in prediction models, thereby ensuring the accuracy of each prediction model and a wide range of use. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings of the present invention are as follows.

[0044] Figure 1 This is the principle diagram of the k-means clustering algorithm.

[0045] Figure 2 This is the schematic diagram of the random forest algorithm.

[0046] Figure 3 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0047] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0048] A transformer insulating oil fault detection method based on oil classification, such as Figure 1 As shown, the specific steps are as follows:

[0049] S1. Divide the prediction interval into master and slave intervals.

[0050] In step S1, the specific steps are as follows:

[0051] S11. Obtain several insulating oils with different component contents as sample insulating oils.

[0052] S12. Apply a test voltage to each sample insulating oil for a preset time to obtain several fault characteristics of each sample insulating oil.

[0053] S13. Construct a feature sample based on all the fault features of each sample of insulating oil.

[0054] S14. The k-means clustering algorithm is used to cluster the characteristic samples and divide them into several clustering intervals, each of which includes several insulating oils with different component contents.

[0055] S15. Select the interval containing the largest amount of insulating oil with different component contents as the main prediction interval, and the remaining intervals as the slave prediction intervals.

[0056] k-means clustering algorithm, such as Figure 1 As shown, the details are as follows:

[0057] According to the sample data, given the sample set D = {x 1 , x 2 , ..., x n}, where x i is the i-th fault feature. There are 8 fault features, i=8, which are water content, dielectric loss, surface tension, breakdown voltage, acid value, volume resistivity, viscosity and density;

[0058] Initialize the selected cluster number k, and select k samples from the sample set D randomly or by calculation as the initial cluster centers {y 1 ,y 2 , ..., y k};

[0059] Calculate the fault feature x i and cluster center y j The distance d ij As the basis for clustering, the calculation formula is as follows:

[0060]

[0061] Among them, p is the characteristic sample number;

[0062] Determine x i Cluster labeling: x i Classified as its distance d ij The cluster C corresponding to the cluster center with the smallest value λ (λ=1,2,...,k), complete a clustering;

[0063] The objective function is designed as follows:

[0064]

[0065] Iterative update: Based on the last clustering, calculate the new cluster center y', repeat clustering until the center y' no longer changes, and the cluster corresponding to the final cluster center is the set of insulating oils with the same fault characteristics. The calculation formula for updating the cluster center is as follows:

[0066]

[0067] Where C is the cluster center cluster obtained by the last clustering, and x is the fault feature in the cluster.

[0068] S2. Construct the main prediction model of the main prediction interval.

[0069] In step S2, the main prediction model is a random forest model, and the training method is: obtaining historical samples of insulating oil in the main prediction interval, and training the random forest model with the historical samples.

[0070] Random forest algorithm, such as Figure 2 As shown, several decision trees are first constructed, and then the final diagnosis result is obtained by integrating the prediction results of the decision trees.

[0071] In the process of forming a decision tree, each new node needs to select a new attribute as the basis for splitting. This embodiment uses the importance of the feature variable as the basis for splitting.

[0072] When constructing a random forest, the importance index of the feature variables is calculated, the feature variables are sorted, and the important feature variables are extracted; the specific method is as follows:

[0073] For 9 feature variables x 1 、x 2 , ..., x 9 , there are C categories and I decision trees, then the estimated probability of node m being in the Cth category is

[0074] The Gini index of node m is:

[0075]

[0076] Feature variable x j The importance score at node m is

[0077]

[0078] Among them, G l and G r They are the Gini indexes of the new nodes l and r after the node m branches out.

[0079] Feature variable x j The set that appears in the i-th tree is M, then the feature variable x j The importance of the i-th tree is expressed as

[0080]

[0081] Feature variable x j The importance in the model is expressed as

[0082]

[0083] S3. Obtain the component content and fault characteristics of the insulating oil to be tested.

[0084] In step S3, the fault characteristics include: water content, dielectric loss, surface tension, breakdown voltage, acid value, volume resistivity, viscosity and density.

[0085] Methods for obtaining fault characteristics of the insulating oil to be tested include multi-frequency ultrasonic detection.

[0086] S4, determining a prediction interval of the insulating oil to be tested based on the component content of the insulating oil to be tested;

[0087] After clustering, a query table of prediction intervals corresponding to insulating oil components is generated, and the prediction intervals of the insulating oil to be tested are obtained through the query table.

[0088] S5. According to the prediction interval of the insulating oil to be tested, determine the main prediction interval to which the insulating oil to be tested belongs.

[0089] In step S5, if the subordinate prediction interval where the insulating oil to be detected is located has not had its affiliation deleted, the original main prediction interval is selected; if the affiliation has been deleted, a new prediction interval is selected.

[0090] S6. Input the fault characteristics of the insulating oil to be tested into the main prediction model constructed by the main prediction interval to which it belongs, and the main prediction model outputs the diagnosis result and prediction confidence of the insulating oil to be tested.

[0091] Specifically, the calculation method of the predicted trust is as follows:

[0092] S61. Obtain the center position of the main prediction interval, calculate the Euclidean distance from the center position of the main prediction interval to each sample insulating oil, and use the farthest Euclidean distance as the maximum boundary length of the main prediction interval.

[0093] The maximum boundary length is calculated as:

[0094]

[0095]

[0096] In the formula, is the Euclidean distance of the jth sample insulating oil in the main prediction interval, is the value of the i-th characteristic variable of the j-th sample insulating oil in the main prediction interval, is the central value of the ith characteristic variable in the main prediction interval.

[0097] S62. Obtain the center positions of the remaining slave prediction intervals, and calculate the Euclidean distance from the center position of the slave prediction interval to the center position of the master prediction interval.

[0098] The calculation formula is as follows:

[0099]

[0100] In the formula, is the Euclidean distance from the i-th characteristic variable in the k-th slave prediction interval to the i-th characteristic variable in the main prediction interval.

[0101] S63. The prediction confidence of the sample insulating oil in the secondary prediction interval is obtained by dividing the maximum boundary length of the primary prediction interval by the Euclidean distance from the center position of the prediction interval to the center position of the primary prediction interval.

[0102]

[0103] Where trustlevel k is the prediction confidence of the kth prediction interval.

[0104] S7. Iterative updates during use.

[0105] In step S7, the prediction confidence of the prediction interval is corrected, and the specific steps are as follows:

[0106] S711, input the fault characteristics of the insulating oil to be tested belonging to the secondary prediction interval into the primary prediction model to obtain the prediction diagnosis result f rf (x i ).

[0107] S712: Divide the number of correct prediction results by the total number of predictions to obtain the prediction accuracy rate.

[0108] In step S712, after the result is detected, is the actual test result, then the prediction is judged to be correct;

[0109] The calculation formula for prediction accuracy is as follows:

[0110]

[0111] S713. Take the middle value of the prediction accuracy and the prediction confidence of the prediction interval before correction as the prediction confidence after correction and update.

[0112] In step S713, the calculation formula of the corrected and updated prediction confidence is as follows:

[0113]

[0114] In step S7, if the prediction confidence is lower than the second threshold, the affiliation between the slave prediction interval and the master prediction interval is deleted; if the prediction confidence is higher than the first threshold, the slave prediction interval is merged into the master prediction interval.

[0115] After deleting the relationship between the slave prediction interval and the master prediction interval, do the following:

[0116] S721. Re-cluster all the unaffiliated slave prediction intervals to divide the master and slave prediction intervals.

[0117] S722. Construct a prediction model for a new main prediction interval.

[0118] S723. Calculate the prediction confidence from the new slave prediction interval to the master prediction interval.

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A transformer insulating oil fault detection method based on oil classification, characterized in that: The specific method is as follows: Obtain the component content and fault characteristics of the insulating oil to be tested; Determine the prediction interval of the insulating oil to be tested based on the component content of the insulating oil to be tested; According to the prediction interval of the insulating oil to be tested, determining the main prediction interval to which the insulating oil to be tested belongs; The fault characteristics of the insulating oil to be tested are input into the main prediction model constructed by the main prediction interval to which it belongs, and the main prediction model outputs the diagnosis result and prediction confidence of the insulating oil to be tested; According to the content of insulating oil components, the method of dividing the prediction interval is as follows: Obtaining a number of insulating oils with different component contents as sample insulating oils; For each sample insulating oil, a test voltage is applied for a preset time to obtain several fault characteristics of each sample insulating oil; A characteristic sample is constructed with all the fault characteristics of each sample insulating oil; The k-means clustering algorithm is used to cluster the characteristic samples and divide them into several cluster intervals, each of which includes several insulating oils with different component contents. The method for selecting the main prediction interval is: select the interval containing the largest number of insulating oils with different component contents as the main prediction interval, and the remaining intervals as the secondary prediction intervals.

2. The transformer insulating oil fault detection method based on oil classification according to claim 1 is characterized in that: The fault characteristics include: water content, dielectric loss, surface tension, breakdown voltage, acid value, volume resistivity, viscosity and density.

3. The transformer insulating oil fault detection method based on oil classification according to claim 1 is characterized in that: The method for obtaining fault characteristics of insulating oil to be detected includes performing multi-frequency ultrasonic detection on the insulating oil to be detected.

4. The transformer insulating oil fault detection method based on oil classification according to claim 1 is characterized in that: The main prediction model is a random forest model, and the training method is: Obtain historical samples of insulating oil within the main prediction interval and use the historical samples to train the random forest model.

5. The transformer insulating oil fault detection method based on oil classification according to claim 1 is characterized in that: The calculation method of predicted confidence is as follows: Get the center position of the main prediction interval, calculate the Euclidean distance from the center position of the main prediction interval to each sample insulating oil, and take the farthest Euclidean distance as the maximum boundary length of the main prediction interval; Get the center positions of the remaining slave prediction intervals, calculate the Euclidean distance from the center position of the slave prediction interval to the center position of the main prediction interval, and divide the maximum boundary length of the main prediction interval by the Euclidean distance from the center position of the prediction interval to the center position of the main prediction interval as the prediction confidence of the sample insulating oil in the slave prediction interval.

6. The transformer insulating oil fault detection method based on oil classification according to claim 5 is characterized in that: During the fault detection process, If the prediction confidence is lower than the second threshold, the affiliation between the slave prediction interval and the master prediction interval is deleted; If the prediction confidence is higher than the first threshold, the slave prediction interval will be merged into the master prediction interval.

7. The transformer insulating oil fault detection method based on oil classification according to claim 6 is characterized in that: Correct the confidence of the prediction from the prediction interval as follows: Inputting the fault characteristics of the insulating oil to be tested belonging to the secondary prediction interval into the primary prediction model to obtain the prediction diagnosis result; The number of correct predictions is divided by the total number of predictions to obtain the prediction accuracy rate; The middle value between the prediction accuracy and the prediction confidence of the prediction interval before correction is taken as the prediction confidence after correction and update.

8. The transformer insulating oil fault detection method based on oil classification according to claim 6 is characterized in that: If the prediction confidence is lower than the second threshold, delete the affiliation between the slave prediction interval and the master prediction interval, and then perform the following operations: Re-cluster all unaffiliated slave prediction intervals to divide the master and slave prediction intervals; Construct a forecast model for the new primary forecast interval; Calculate the forecast confidence of the new slave forecast interval to the master forecast interval.