Predictive maintenance method for current transformer with iron core type structure

Abstract

The invention relates to the technical field of fault prediction of current transformers, in particular to a predictive maintenance method of a current transformer with an iron core type structure, which comprises the following steps: feature selection and fault classification; a wavelet transform method which has strong video analysis capability and is suitable for processing non-stationary signals of a current transformer is adopted, and a Spearman level correlation coefficient and mutual information combination method is adopted to ensure that extracted features have nonlinear correlation and can maximize the information amount, so that the generalization capability of a model is improved, and the accuracy of the model is improved. A classification prediction model is constructed by designing a random forest, so that high-dimensional data can be processed, the model is not sensitive to noise, and the performance of the model can be further optimized through a built-in feature importance evaluation mechanism; according to the method, the single-phase short-circuit fault and the double-phase short-circuit fault of the current transformer can be accurately predicted, and the robustness is high.

Description

Technical Field

[0001] The invention relates to the technical field of current transformer fault prediction, and in particular to a predictive maintenance method for a current transformer with an iron core structure. Background Art

[0002] The iron core structure current transformer is an important component commonly used in power systems. In order to reduce the occurrence of faults in the iron core structure current transformer, the traditional method is to have the staff conduct regular inspections of the iron core structure current transformer, which is time-consuming and labor-intensive, and the inspection effect is poor;

[0003] Therefore, there are now some current transformer fault data analysis and prediction systems based on big data, such as the invention patent with publication number CN114091726A, and a current sensor monitoring and alarm device and system for a low-voltage current transformer, such as the invention patent with publication number CN118473096B, which are used to predict faults of current transformers.

[0004] However, it was found during use that the existing fault prediction system has low accuracy in predicting single-phase short-circuit faults and double-phase short-circuit faults of current transformers, and has poor robustness. In addition, the processing effect of data features is poor. Therefore, a predictive maintenance method for core-type current transformers is urgently needed to improve the above problems. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a wavelet transform method with powerful video analysis capability and suitable for processing non-stationary signals of current transformers, and ensures that the extracted features have both nonlinear correlation and maximize the amount of information through the Spearman rank correlation coefficient and mutual information combination method, thereby improving the generalization ability of the model, and constructs a classification prediction model by designing a random forest, which can not only process high-dimensional data and is insensitive to noise, but also further optimize the performance of the model through its built-in feature importance evaluation mechanism, so that the method can accurately predict the single-phase short-circuit fault and the two-phase short-circuit fault of the current transformer and has a predictive maintenance method for a current transformer with an iron core structure with strong robustness.

[0006] A predictive maintenance method for a current transformer with an iron core structure of the present invention comprises the following steps:

[0007] S1. Data preprocessing: First, data cleaning is performed to remove some abnormal values ​​and illegal data. Since the original sensor data collected cannot be directly used for our subsequent processing, the data needs to be labeled, and then the fault state and normal state data are converted into digital forms that can be processed by the machine learning model. Wavelet transform technology is used to process the signals of current transformers with complex frequency components and transient characteristics;

[0008] S2, feature selection: Spearman coefficient and mutual information are used to jointly screen the data and retain features with strong monotonic relationships;

[0009] S3, fault classification: use random forest model for classification prediction;

[0010] The wavelet transform method with powerful video analysis capability and suitable for processing non-stationary signals of current transformers is adopted, and the Spearman rank correlation coefficient and mutual information are combined to ensure that the extracted features have both nonlinear correlation and maximize the amount of information, thereby improving the generalization ability of the model. A classification prediction model is constructed by designing a random forest, which can not only process high-dimensional data and is insensitive to noise, but also further optimize the performance of the model through its built-in feature importance evaluation mechanism, so that this method can accurately predict single-phase short-circuit faults and double-phase short-circuit faults of current transformers and has strong robustness.

[0011] Preferably, the wavelet transform technology comprises the following specific steps:

[0012] a1. Select a suitable wavelet basis function ψ(t). Here we choose Daubechies wavelet, which is convenient for video localization analysis of the signal and can effectively capture short-term mutations in the current transformer signal;

[0013] a2. Perform wavelet decomposition, perform discrete wavelet transform on the data, and decompose the signal x(t) into multiple levels through the selected wavelet basis function to form different frequency components to extract the characteristics of different frequency bands of the signal;

[0014] a3. Perform wavelet reconstruction. In the classification prediction task, the frequency information of the signal directly decomposed by wavelet is too complex or scattered. Therefore, the processed frequency information is restored to a signal suitable for further analysis. This not only retains the frequency information but also allows for subsequent processing. The inverse wavelet transform (IDWT) is used:

[0015]

[0016] Among them, c j,k are the wavelet coefficients obtained at the jth layer.

[0017] Preferably, the signal x(t) is expressed as:

[0018]

[0019] Among them, c j,k is the wavelet coefficient, and ψ j,k (t) is the translation and scaling of the wavelet basis function, and J is the number of decomposition levels.

[0020] Preferably, c j,k and ψ j,k (t) are respectively expressed as:

[0021] cj,k=∫x(t)ψj,k(t)dt,

[0022] Preferably, the specific steps of S2 are as follows:

[0023] b1. Perform a preliminary screening and use the Spearman coefficient to filter out features that have almost no monotonic relationship with the label, reduce some data dimensions, and retain features with a strong monotonic relationship;

[0024] b2. After the initial screening, use mutual information to evaluate whether the remaining features have nonlinear or more complex effects on the labels to reduce the omission of important features;

[0025] b3. Finally, during the screening process, features that meet the threshold 1 set for the Spearman coefficient or the threshold 2 set for the mutual information are selected.

[0026] Preferably, the specific steps of S3 are as follows:

[0027] c1. After the previous preprocessing and feature selection are completed, the data is divided into training set and test set, and the model is first created using Pipeline, including SMOTE and random forest classifier. Since the fault data in the current transformer data is a minority, we use SMOTE to generate minority class samples to balance the training data;

[0028] c2. Use grid search to find the best hyperparameter combination of the model through cross-validation, and then use the best hyperparameter combination to predict the test set.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] 1. The wavelet transform method, which has powerful video analysis capabilities and is suitable for processing non-stationary signals of current transformers, is used. The Spearman rank correlation coefficient and mutual information are combined to ensure that the extracted features have both nonlinear correlation and maximize the amount of information, thereby improving the generalization ability of the model;

[0031] 2. Design a random forest to build a classification prediction model, which can not only process high-dimensional data and is insensitive to noise, but also further optimize the performance of the model through its built-in feature importance evaluation mechanism;

[0032] 3. This method can accurately predict single-phase short-circuit faults and double-phase short-circuit faults of current transformers and has strong robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flow chart of a predictive maintenance method of a current transformer with an iron core structure according to the present invention;

[0034] Figure 2 It is a schematic flow chart of the data preprocessing steps of the present invention;

[0035] Figure 3 It is a schematic diagram of the feature selection process of the present invention;

[0036] Figure 4 It is the ROC curve diagram of the present invention;

[0037] Figure 5 is a confusion matrix diagram of the present invention;

[0038] Figure 6 It is a result evaluation report diagram of the present invention. DETAILED DESCRIPTION

[0039] In order to facilitate understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. The present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive.

[0040] like Figures 1 to 3 As shown, a predictive maintenance method for a current transformer with an iron core structure includes the following steps:

[0041] S1. Data preprocessing: First, data cleaning is performed to remove some abnormal values ​​and illegal data. Since the original sensor data collected cannot be directly used for our subsequent processing, the data needs to be labeled, and then the fault state and normal state data are converted into digital forms that can be processed by the machine learning model. Wavelet transform technology is used to process the signals of current transformers with complex frequency components and transient characteristics;

[0042] S2, feature selection: Spearman coefficient and mutual information are used to jointly screen the data and retain features with strong monotonic relationships;

[0043] S3, fault classification: use random forest model for classification prediction;

[0044] The specific steps of the wavelet transform technology are as follows:

[0045] a1. Select a suitable wavelet basis function ψ(t). Here we choose Daubechies wavelet, which is convenient for video localization analysis of the signal and can effectively capture short-term mutations in the current transformer signal;

[0046] a2. Perform wavelet decomposition, perform discrete wavelet transform on the data, and decompose the signal x(t) into multiple levels through the selected wavelet basis function to form different frequency components to extract the characteristics of different frequency bands of the signal;

[0047] a3. Perform wavelet reconstruction. In the classification prediction task, the frequency information of the signal directly decomposed by wavelet is too complex or scattered. Therefore, the processed frequency information is restored to a signal suitable for further analysis. This not only retains the frequency information but also allows for subsequent processing. The inverse wavelet transform is used:

[0048]

[0049] Among them, c j,k is the wavelet coefficient obtained at the jth layer;

[0050] The signal x(t) is expressed as:

[0051]

[0052] Among them, c j,k is the wavelet coefficient, and ψ j,k (t) is the translation and scaling of the wavelet basis function, and J is the number of decomposition levels;

[0053] c j,k and ψ j,k (t) are respectively expressed as:

[0054] cj,k=∫x(t)ψj,k(t)dt,

[0055] The specific steps of S2 are as follows:

[0056] b1. Perform a preliminary screening and use the Spearman coefficient to filter out features that have almost no monotonic relationship with the label, reduce some data dimensions, and retain features with a strong monotonic relationship;

[0057] b2. After the initial screening, use mutual information to evaluate whether the remaining features have nonlinear or more complex effects on the labels to reduce the omission of important features;

[0058] b3. Finally, during the screening process, features that meet the threshold 1 set for the Spearman coefficient or the threshold 2 set for the mutual information are selected;

[0059] The specific steps of S3 are as follows:

[0060] c1. After the previous preprocessing and feature selection are completed, the data is divided into training set and test set, and the model is first created using Pipeline, including SMOTE and random forest classifier. Since the fault data in the current transformer data is a minority, we use SMOTE to generate minority class samples to balance the training data;

[0061] c2. Use grid search to find the best hyperparameter combination of the model through cross-validation, and then use the best hyperparameter combination to predict the test set.

[0062] Example

[0063] A case analysis is conducted using sample data from an iron core structured current transformer device in a 110kv third generation smart substation;

[0064] The original data set was obtained from the historical detection data of the sensors of the substation. We extracted some data and conducted 8 repeated tests. Each test consisted of 1,000 sample data, including data in healthy state, data with single-phase short circuit fault, and data with double-phase short circuit fault. The distribution of each sample data is shown in Table 1 below:

[0065] Table 1 Data status statistics in the dataset

[0066] Data No. Current transformer status Sample size D1 Health status (HLTY) 700 D2 Single-phase short circuit fault (1HCF) 200 D3 Two-phase short circuit fault (2HCF) 100

[0067] Some monitoring data are shown in Table 2:

[0068] Table 2 Partial current transformer monitoring data table

[0069]

[0070]

[0071] like Figure 4 As shown in Figure 1, the ROC curve results are obtained after training the model on the training set through our predictive maintenance technology and testing the classification prediction performance of the trained model on the test set.

[0072] like Figure 5 As shown, in order to comprehensively evaluate our model and further understand the classification effect of the model, we also plotted the confusion matrix.

[0073] like Figure 6 As shown in the final result evaluation report, 0 represents the 1HCF state, 1 represents the 2HCF state, and 2 represents the HLTY state.

[0074] Analysis of the above results shows that the overall average prediction rate of this technology has reached 94%, especially the prediction effect on the data in the HLTY state is very accurate, reaching an average of 100%, and the prediction effect on the data in the 1HCF state is also 84% on average. It can be seen that a predictive maintenance method for a current transformer with an iron core structure has the ability to accurately predict single-phase short-circuit faults and two-phase short-circuit faults of the current transformer and has strong robustness.

[0075] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A predictive maintenance method for a current transformer with an iron core structure, characterized in that: The following steps are involved: S1. Data preprocessing: First, data cleaning is performed to remove some abnormal values ​​and illegal data. Since the original sensor data collected cannot be directly used for our subsequent processing, the data needs to be labeled, and then the fault state and normal state data are converted into digital forms that can be processed by the machine learning model. Wavelet transform technology is used to process the signals of current transformers with complex frequency components and transient characteristics; S2, feature selection: Spearman coefficient and mutual information are used to jointly screen the data and retain features with strong monotonic relationships; S3. Fault classification: Use random forest model for classification prediction.

2. The predictive maintenance method of a current transformer with an iron core structure as claimed in claim 1, characterized in that: The specific steps of the wavelet transform technology are as follows: a1. Select a suitable wavelet basis function ψ(t). Here we choose Daubechies wavelet, which is convenient for video localization analysis of the signal and can effectively capture short-term mutations in the current transformer signal; a2. Perform wavelet decomposition, perform discrete wavelet transform on the data, and decompose the signal x(t) into multiple levels through the selected wavelet basis function to form different frequency components to extract the characteristics of different frequency bands of the signal; a3. Perform wavelet reconstruction. In classification prediction tasks, if the frequency information of the signal directly decomposed by wavelet is too complex or scattered, the processed frequency information is restored to a signal suitable for further analysis. This not only retains the frequency information but also allows for subsequent processing. The inverse wavelet transform is used.

3. The predictive maintenance method of a current transformer with an iron core structure as claimed in claim 2, characterized in that: The signal x(t) is expressed as: Among them, c j，k is the wavelet coefficient, and ψ j，k (t) is the translation and scaling of the wavelet basis function, and J is the number of decomposition levels.

4. The predictive maintenance method of a current transformer with an iron core structure as claimed in claim 3, characterized in that: c j，k and ψ j，k (t) are respectively expressed as: c j，k =τx(t)ψ j，k (t)dt, 5. The predictive maintenance method of a current transformer with an iron core structure as claimed in claim 1, characterized in that: The specific steps of S2 are as follows: b1. Perform a preliminary screening and use the Spearman coefficient to filter out features that have almost no monotonic relationship with the label, reduce some data dimensions, and retain features with a strong monotonic relationship; b2. After the initial screening, use mutual information to evaluate whether the remaining features have nonlinear or more complex effects on the labels to reduce the omission of important features; b3. Finally, during the screening process, features that meet the threshold 1 set for the Spearman coefficient or the threshold 2 set for the mutual information are selected.

6. The predictive maintenance method of a current transformer with an iron core structure as claimed in claim 1, characterized in that: The specific steps of S3 are as follows: c1. After the previous preprocessing and feature selection are completed, the data is divided into training set and test set, and the model is first created using Pipeline, including SMOTE and random forest classifier. Since the fault data in the current transformer data is a minority, we use SMOTE to generate minority class samples to balance the training data; c2. Use grid search to find the best hyperparameter combination of the model through cross-validation, and then use the best hyperparameter combination to predict the test set.

Images