High-precision magnetizing inrush current waveform identification and classification method

Abstract

The invention provides a high-precision excitation inrush current waveform identification and classification method, and belongs to the technical field of electronic information. The technical problem that the complex excitation waveform is difficult to accurately identify by the existing method is solved. According to the technical scheme, the method comprises the following steps: S1, selecting nine feature variables based on preprocessed data; s2, performing a round of feature screening by using U test for the 9 feature variables; s3, performing two rounds of feature screening by using collinearity test and feature importance sorting; s4, on the basis of RandomForest, GradientBoosting and Light-GBM, a soft voting integrated classifier is constructed, and on the basis of the RandomForest, the GradientBoosting and the Light-GBM, the soft voting integrated classifier is constructed; and S5, carrying out excitation waveform classification based on a soft voting ensemble classifier. The method has the beneficial effects that the excitation waveforms can be accurately identified and classified.

Description

Technical Field

[0001] The present invention relates to the field of electronic information technology, and in particular to a high-precision identification and classification method of excitation surge current waveforms. Background Art

[0002] In the field of electronic information technology, the identification and classification of excitation inrush current waveforms are crucial for applications such as transformer protection. Different forms of excitation waveforms directly affect the dynamic behavior of magnetic flux inside the transformer core, so correctly identifying and classifying the waveform characteristics of excitation inrush current is of great significance for ensuring the safe and stable operation of the power system.

[0003] The excitation waveform is mainly reflected in the distribution law of the magnetic flux density changing with time. Common excitation waveforms include sine wave, triangular wave and trapezoidal wave, which correspond to different magnetic flux density change modes: the magnetic flux density of the sine wave changes sinusoidally with time, which is common in traditional AC power supplies; the magnetic flux density of the triangular wave increases and decays linearly with time, which is common in some DC to AC conversion circuits; the magnetic flux density of the trapezoidal wave remains constant for a period of time, and then changes rapidly to another constant value, which is common in switching power supplies.

[0004] In the transformer protection of power system, the current waveform is basically symmetrical when the transformer is in normal operation and internal fault. However, during the excitation inrush current, due to the large value of high-order harmonic components in the inrush current, the change curve of the excitation inrush current is a peak wave, resulting in an asymmetric current waveform. By analyzing the excitation waveform, the type and location of the fault can be determined, which helps to quickly isolate the fault and restore the normal operation of the system.

[0005] When designing electrical equipment such as transformers and motors, understanding the impact of different excitation waveforms on equipment performance can help engineers optimize the design and improve the efficiency and reliability of the equipment. By analyzing the excitation waveform, it is helpful to study the dynamic stability of the power system, especially when the system is disturbed, such as voltage fluctuations and frequency changes in the power system. Continuous monitoring and analysis of the excitation waveform can predict potential failures of the equipment, which facilitates early maintenance and improves the safety of the power system. In addition, in the simulation study of the power system, the classification of excitation waveforms can be used to simulate different operating conditions to study the behavior of the system under various conditions, which also plays an important role in promoting the modernization and intelligence of the power system.

[0006] However, there are relatively few methods for identifying excitation waveforms that rely on magnetic flux density to obtain characteristic variables, and efficient and high-precision identification has not yet been achieved. Therefore, it has become an urgent need to develop a high-precision identification and classification method for excitation inrush current waveforms to solve the above problems. Summary of the invention

[0007] The purpose of the present invention is to provide a high-precision identification and classification method for excitation inrush current waveforms, which can accurately identify and classify excitation waveforms. The present invention aims to improve the accuracy and response speed of transformer protection through advanced signal processing algorithms and machine learning techniques, and ensure the safe, stable and reliable operation of the power system. In addition, the system will provide optimization suggestions for power system design, enhance equipment performance, predict and prevent potential faults, thereby improving the efficiency and economy of the power system, and supporting the modernization and intelligent development of the power system.

[0008] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is specifically: a high-precision method for identifying and classifying excitation inrush current waveforms, comprising the following steps:

[0009] S1. Select 9 feature variables based on the preprocessed data;

[0010] S2. For the 9 characteristic variables, a round of feature screening was performed using the U test;

[0011] S3, using collinearity test and feature importance ranking to perform second round of feature screening;

[0012] S4. Build a soft voting ensemble classifier based on RandomForest, GradientBoosting and Light-GBM;

[0013] S5. Classify the excitation waveform based on the soft voting integrated classifier.

[0014] Step S1 specifically includes the following steps:

[0015] S1.1. Take 4 materials as material samples Y = {Y1, Y2, Y3, Y4}, and extract the magnetic flux density data of all sampling points of the 4 materials in a cycle under different working conditions;

[0016] S1.2. Based on the extracted data, 9 characteristic variables are set as characteristic samples X = {X1, X2, ..., X9}, where the characteristic variable X1 represents the average magnetic flux density B avg , the value of the average magnetic flux density is equal to the sum of the magnetic flux density at each moment in a period T divided by the number of sampling points, which is specifically expressed as follows:

[0017]

[0018] Among them, B i represents the magnetic flux density at the i-th moment, and n represents the total number of sampling points;

[0019] Characteristic variable X2, which represents the standard deviation of magnetic flux density B std , specifically expressed as follows:

[0020]

[0021] Characteristic variable X3, which represents the magnetic flux density skewness B skew , specifically expressed as follows:

[0022]

[0023] Characteristic variable X4, which represents the kurtosis of magnetic flux density B kurt , specifically expressed as follows:

[0024]

[0025] Characteristic variable X5, which represents the maximum value of magnetic flux density B max , specifically expressed as follows:

[0026] B max =max(B1, B2, ..., B n ) (5)

[0027] Characteristic variable X6, which represents the maximum value of magnetic flux density B min , specifically expressed as follows:

[0028] B min =min(B1, B2, ..., B n ) (6)

[0029] The characteristic variable X7 represents the magnetic flux density range Range, which is specifically expressed as follows:

[0030] Range = B max -B min (7)

[0031] The characteristic variable X8 represents the coefficient of variation CV, which is specifically expressed as follows:

[0032]

[0033] Characteristic variable X9, which represents the smoothed second-order derivative, is used because the magnetic flux density data of each sampling point is discrete. 2 B is used to approximate the second-order derivative B″, which is specifically expressed as follows:

[0034] Δ 2 B=B i+2 -2B i+1 +B i (9)

[0035]

[0036] Here, h represents the step size between sampling points.

[0037] Step S2 specifically includes the following steps:

[0038] S2.1. Determine the U test hypothesis, assuming that the population distributions from which the two independent samples come are the same;

[0039] S2.2, material Y j The data of any two of the nine feature variables under (j=1, 2, 3, 4) are merged into one data set. The group data is sorted in ascending order and the sorted data is assigned a rank. If the data values ​​are the same, the average rank is assigned;

[0040] S2.3. Based on the merged data set and the ranking obtained after sorting, calculate the U statistic, which is specifically expressed as follows:

[0041]

[0042] Among them, n1 and n2 are the sizes of the two groups of samples, r1 and r2 are the ranks of the first and second groups of samples, respectively.

[0043] S2.4. Compare the U values ​​of the two characteristic variables, take U = min (U1, U2), and use statistical software or U test table to find the p value corresponding to the U value;

[0044] S2.5. If the p-value is less than the significance level of 0.05, the characteristic variable is retained, and the final number of characteristic variables retained is N1.

[0045] Step S3 specifically includes the following steps:

[0046] S3.1. Based on the N1 characteristic variables, the VIF values ​​are calculated respectively, which are specifically expressed as follows:

[0047]

[0048] Among them, VIF h is the variance inflation factor of the hth feature variable, R h It is the coefficient of determination obtained by linear regression of the hth characteristic variable as the response variable with respect to other characteristic variables;

[0049] S3.2. Based on the VIF h , performed the variance inflation factor (VIF) test, and a VIF value higher than 10 indicated the existence of collinearity problems;

[0050] S3.3. If there is a feature variable with a VIF value higher than 10, the feature variable with the largest VIF value is eliminated, that is, N1-1 feature variables remain.

[0051] S3.4, cyclically eliminate the feature variables with VIF values ​​greater than 10 in descending order of VIF values, until the VIF values ​​of all feature variables are less than 10, and finally retain N2 feature variables;

[0052] S3.5, evaluate the importance of N2 feature variables;

[0053] S3.6. Retain the top three feature variables with the highest importance score;

[0054] Step S4 specifically includes the following steps:

[0055] S4.1, based on the feature data set, dividing the samples in the feature data set into a test set and a training set according to a specific ratio;

[0056] S4.2, based on the training set, training the RandomForest classifier, the GradientBoosting classifier and the LightGBM classifier respectively;

[0057] S4.3. Set a weight coefficient v for each classifier i i , specifically expressed as follows:

[0058]

[0059] Among them, i=1 represents RandomForest classifier; i=2 represents GradientBoosting classifier; i=3 represents LightGBM classifier, AUC i is the probability of the three excitation waveforms being correctly predicted under classifier i;

[0060] S4.4. Based on the weight v of each classifier i and the excitation waveform prediction probability p ik , respectively calculate the average prediction probability of waveform k, which is specifically expressed as follows:

[0061]

[0062] Among them, P k is the weighted average predicted probability of waveform k, k = 1 represents a sine wave, k = 2 represents a triangle wave, k = 3 represents a trapezoidal wave, p ik represents the predicted probability of waveform k under classifier i;

[0063] S4.5. Integrate the trained RandomForest classifier, GradientBoosting classifier and LightGBM classifier into a soft voting classifier.

[0064] Step S5 specifically includes the following steps:

[0065] S5.1, based on the weighted average predicted probability P k , sorting the weighted average prediction probabilities of the sine wave, the triangle wave, and the trapezoidal wave, and obtaining the sequence number with the highest weighted average prediction probability, including:

[0066] The sequence number of the feature variable with the highest weighted average prediction probability is calculated according to formula (16), which is specifically expressed as follows:

[0067]

[0068] Where k is the final classified waveform label, k = 1 represents a sine wave, k = 2 represents a triangle wave, and k = 3 represents a trapezoidal wave;

[0069] S5.2, based on the test set, using the soft voting integrated classifier to classify the excitation waveform of the test set;

[0070] S5.3. Use ROC curve to analyze the prediction results.

[0071] Compared with the prior art, the present invention has the following beneficial effects:

[0072] (1) The present invention uses U test and collinearity analysis to perform two rounds of feature variable screening, which significantly improves the classification accuracy and enhances the model's adaptability to different materials and waveforms.

[0073] (2) The present invention combines three different classifiers and utilizes their complementary advantages to achieve optimal performance on specific tasks, demonstrating the powerful function of the soft voting integrated classifier in processing complex data.

[0074] (3) When integrating the predictions of different classifiers, the present invention provides a smoother and more reliable decision-making process by calculating the average probability rather than a simple majority vote. It demonstrates excellent robustness and consistency in the excitation waveform recognition task and ensures 100% classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0076] Figure 1 It is a schematic diagram of the overall process of a high-precision identification and classification method of excitation inrush current waveform proposed by the present invention;

[0077] Figure 2A schematic diagram of a process of "using U test to perform a round of feature screening" for a high-precision identification and classification method of excitation inrush current waveform proposed by the present invention;

[0078] Figure 3 A schematic diagram of the process of "using collinearity test and feature importance ranking to perform two rounds of feature screening" of a high-precision identification and classification method of excitation inrush current waveform proposed by the present invention;

[0079] Figure 4 A schematic diagram of the process of "building a soft voting integrated classifier" for a high-precision identification and classification method of excitation inrush current waveform proposed by the present invention;

[0080] Figure 5 A schematic diagram of the "excitation waveform classification" process of a high-precision identification and classification method of an excitation inrush current waveform proposed by the present invention;

[0081] Figure 6 This is a comparison diagram of the first-order and second-order derivatives of the smoothed sine wave based on the differential approximation method of the present invention;

[0082] Figure 7 This is a comparison diagram of the first-order and second-order derivatives of the smoothed triangular wave based on the differential approximation method of the present invention;

[0083] Figure 8 This is a comparison diagram of the first-order and second-order derivatives of the smoothed trapezoidal wave based on the differential approximation method of the present invention;

[0084] Fig. 9 This is a feature importance scoring diagram of the material of the present invention;

[0085] Fig.10 This is a feature importance scoring diagram of material 2 of the present invention;

[0086] Fig.11 This is a graph showing the importance of three features of the material of the present invention;

[0087] Fig.12 This is a graph showing the importance of four features of the material of the present invention;

[0088] Fig.13 A ROC curve diagram of materials for a high-precision method for identifying and classifying excitation inrush current waveforms of the present invention.

[0089] Fig.14 This is a ROC curve diagram of Material 1 of Example 2.

[0090] Fig.15 This is the ROC curve diagram of Material 2 of Example 2.

[0091] Fig.16 This is the ROC curve diagram of Material 3 of Example 2.

[0092] Fig.17 This is the ROC curve diagram of Material 4 in Example 2. DETAILED DESCRIPTION

[0093] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. Of course, the specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.

[0094] Example 1

[0095] See also Figure 1 As shown, this embodiment provides a high-precision method for identifying and classifying excitation inrush current waveforms, comprising the following steps:

[0096] S1. Select 9 feature variables based on the preprocessed data;

[0097] S1.1. Take 4 materials as material samples Y = {Y1, Y2, Y3, Y4}, and extract the magnetic flux density data of all sampling points of the 4 materials in a cycle under different working conditions;

[0098] S1.2. Based on the extracted data, 9 characteristic variables are set as characteristic samples X = {X1, X2, ..., X9}, where the characteristic variable X1 represents the average magnetic flux density B avg , the value of the average magnetic flux density is equal to the sum of the magnetic flux density at each moment in a period T divided by the number of sampling points, which is specifically expressed as follows:

[0099]

[0100] Among them, B i represents the magnetic flux density at the i-th moment, and n represents the total number of sampling points;

[0101] Characteristic variable X2, which represents the standard deviation of magnetic flux density B std , specifically expressed as follows:

[0102]

[0103] Characteristic variable X3, which represents the magnetic flux density skewness B skew , specifically expressed as follows:

[0104]

[0105] Characteristic variable X4, which represents the kurtosis of magnetic flux density B kurt , specifically expressed as follows:

[0106]

[0107] Characteristic variable X5, which represents the maximum value of magnetic flux density B max , specifically expressed as follows:

[0108] B max =max(B1, B2, ..., B n ) (5)

[0109] Characteristic variable X6, which represents the maximum value of magnetic flux density B min , specifically expressed as follows:

[0110] B min =min(B1, B2, ..., B n ) (6)

[0111] The characteristic variable X7 represents the magnetic flux density range Range, which is specifically expressed as follows:

[0112] Range = B max -B min (7)

[0113] The characteristic variable X8 represents the coefficient of variation CV, which is specifically expressed as follows:

[0114]

[0115] Characteristic variable X9, which represents the smoothed second-order derivative, is used because the magnetic flux density data of each sampling point is discrete. 2 B is used to approximate the second-order derivative B″, which is specifically expressed as follows:

[0116] Δ 2 B=B i+2 -2B i+1 +B i (9)

[0117]

[0118] Here, h represents the step size between sampling points.

[0119] S2. For the 9 characteristic variables, a round of feature screening was performed using the U test;

[0120] S2.1. Determine the U test hypothesis, assuming that the population distributions from which the two independent samples come are the same;

[0121] S2.2, material Y j The data of any two of the nine feature variables under (j=1,2,3,4) are merged into one data set. The group data is sorted in ascending order and the sorted data is assigned a rank. If the data values ​​are the same, the average rank is assigned;

[0122] S2.3. Based on the merged data set and the ranking obtained after sorting, calculate the U statistic, which is specifically expressed as follows:

[0123]

[0124]

[0125] Among them, n1 and n2 are the sizes of the two groups of samples, r1 and r2 are the rank sums of the first and second groups of samples, respectively;

[0126] S2.4. Compare the U values ​​of the two characteristic variables, take U = min (U1, U2), and use statistical software or U test table to find the p value corresponding to the U value;

[0127] S2.5. If the p-value is less than the significance level of 0.05, the characteristic variable is retained, and the final number of characteristic variables retained is N1.

[0128] S3, using collinearity test and feature importance ranking to perform second round of feature screening;

[0129] S3.1. Based on the N1 characteristic variables, the VIF values ​​are calculated respectively, which are specifically expressed as follows:

[0130]

[0131] Among them, VIF h is the variance inflation factor of the hth feature variable, R h It is the coefficient of determination obtained by linear regression of the hth characteristic variable as the response variable with respect to other characteristic variables;

[0132] S3.2. Based on the VIF h , performed the variance inflation factor (VIF) test, and a VIF value higher than 10 indicated the existence of collinearity problems;

[0133] S3.3. If there is a feature variable with a VIF value higher than 10, the feature variable with the largest VIF value is eliminated, that is, N1-1 feature variables remain.

[0134] S3.4, cyclically eliminate the feature variables with VIF values ​​greater than 10 in descending order of VIF values, until the VIF values ​​of all feature variables are less than 10, and finally retain N2 feature variables;

[0135] S3.5, evaluate the importance of N2 feature variables;

[0136] S3.6. Retain the top three feature variables with the highest importance score;

[0137] S4. Build a soft voting ensemble classifier based on RandomForest, GradientBoosting and Light-GBM;

[0138] S4.1. Based on the feature data set, samples in the feature data set are divided into a test set and a training set according to a specific ratio; S4.2. Based on the training set, a RandomForest classifier, a GradientBoosting classifier and a LightGBM classifier are trained respectively;

[0139] S4.3. Set a weight coefficient v for each classifier i i , specifically expressed as follows:

[0140]

[0141] Among them, i=1 represents RandomForest classifier; i=2 represents GradientBoosting classifier; i=3 represents LightGBM classifier, AUC i is the probability of the three excitation waveforms being correctly predicted under classifier i;

[0142] S4.4. Based on the weight u of each classifier i and the excitation waveform prediction probability p ik , respectively calculate the average prediction probability of waveform k, which is specifically expressed as follows:

[0143]

[0144] Among them, P k is the weighted average predicted probability of waveform k, k = 1 represents a sine wave, k = 2 represents a triangle wave, k = 3 represents a trapezoidal wave, p ik represents the predicted probability of waveform k under classifier i;

[0145] S4.5. Integrate the trained RandomForest classifier, GradientBoosting classifier and LightGBM classifier into a soft voting classifier.

[0146] S5. Classify the excitation waveform based on the soft voting integrated classifier.

[0147] S5.1, based on the weighted average predicted probability P k , sorting the weighted average prediction probabilities of the sine wave, the triangle wave, and the trapezoidal wave, and obtaining the sequence number with the highest weighted average prediction probability, including:

[0148] The sequence number of the feature variable with the highest weighted average prediction probability is calculated according to formula (16), which is specifically expressed as follows:

[0149]

[0150] Where k is the final classified waveform label, k = 1 represents a sine wave, k = 2 represents a triangle wave, and k = 3 represents a trapezoidal wave;

[0151] S5.2, based on the test set, using the soft voting integrated classifier to classify the excitation waveform of the test set;

[0152] S5.3. Use ROC curve to analyze the prediction results.

[0153] Table 1 Symbols

[0154]

[0155]

[0156] In order to verify the effectiveness of the implementation of the present invention, Python is used to perform numerical simulation verification.

[0157] This embodiment screened the magnetic flux density data of 1024 sampling points of material 1, material 2, material 3 and material 4 in one cycle under different working conditions, and extracted 9 characteristic variables, including average magnetic flux density, magnetic flux density standard deviation, magnetic flux density skewness, magnetic flux density kurtosis, magnetic flux density maximum value, magnetic flux density minimum value, magnetic flux density range, coefficient of variation and smoothed second-order derivative, based on the above data. Figure 6 , 7 8 are comparison diagrams of the first-order and second-order derivatives of the sine wave after smoothing based on the differential approximation method of the present invention. It can be seen that the image of the second-order derivative has a higher degree of distinction between the three different waveforms than the first-order derivative. There is no convex point in the second-order derivative image of the sine wave after smoothing, and there are 2 and 4 convex points in the second-order derivative images of the triangular wave and trapezoidal wave after smoothing, respectively. Therefore, the smoothed second-order derivative is selected as the feature variable.

[0158] The U test was used to screen the characteristic variables, as shown in Table 2.

[0159] Table 2 Hypothesis test results of four material characteristic variables

[0160]

[0161]

[0162] From the data in Table 2, it can be seen that in Material 1, there are 6 characteristic variables with p values ​​less than 0.05, namely, average magnetic flux density, standard deviation, skewness, kurtosis, range and smoothed second-order derivative; in Material 2, there are 6 characteristic variables with p values ​​less than 0.05, namely, average magnetic flux density, standard deviation, skewness, kurtosis, range and smoothed second-order derivative; in Material 3, there are 7 characteristic variables with p values ​​less than 0.05, namely, average magnetic flux density, standard deviation, skewness, kurtosis, range, coefficient of variation and smoothed second-order derivative; in Material 4, there are 5 characteristic variables with p values ​​less than 0.05, namely, average magnetic flux density, standard deviation, skewness, kurtosis and smoothed second-order derivative. The p values ​​of the remaining characteristic variables are all greater than 0.05, indicating that these characteristic variables have no statistical significance in the discriminant waveform and are directly eliminated. At this point, one round of characteristic variables is completed, including the average magnetic flux density, standard deviation, skewness, kurtosis, range, coefficient of variation and smoothed second-order derivative. These seven characteristic variables enter the second round of feature screening.

[0163] The second round of feature screening adopted the collinearity test method, and performed multicollinearity tests on 6 feature variables in Material 1, 6 feature variables in Material 2, 7 feature variables in Material 3, and 5 feature variables in Material 4. The original VIF values ​​of the feature variables of the four materials were calculated, and the specific data are shown in Table 3. It can be seen from Table 3 that the collinearity of the standard deviation and range are strong, and the multicollinearity of skewness and kurtosis is low. The feature variables with VIF greater than 10 are eliminated in order from large to small until the VIF values ​​of all feature variables are less than 10. After VIF feature screening, the VIF values ​​of the retained feature variables of the four materials are shown in Table 4.

[0164] Table 3 Original VIF values ​​of characteristic variables in four materials

[0165]

[0166] Table 4 VIF values ​​of characteristic variables retained after screening of four materials

[0167]

[0168] The characteristic variables retained by the four materials were ranked according to their importance scores. Fig. 9 , 10 , 11, and 12 are the importance score diagrams of the four materials respectively. Finally, the characteristic variables with the top three scores are retained. The specific contents are shown in Table 5.

[0169] Table 5 Characteristic variable results after screening of four materials

[0170]

[0171] The subsequent classification model training will use the characteristic variables in Table 5. The training set contains 3400 data of material one, 3000 data of material two, 3200 data of material three, and 2800 data of material four.

[0172] The soft voting ensemble classifier was used to predict the excitation waveforms of 80 data in the test set (20 for each of the four materials), including 20 sine waves, 44 triangle waves, and 16 trapezoidal waves. A 100% classification accuracy was achieved in the classification task of the three waveforms, and the four classifiers showed the best performance in the classification task of the three waveforms of the four materials. The results show that the extracted feature variables can effectively characterize the characteristics of the waveforms in each classifier, ensuring the correct recognition of different waveforms by the classifier, and fully demonstrating the robustness and consistency of the classifier in classifying the three waveforms under different material conditions.

[0173] Fig.13 This is a ROC curve diagram of Material 1 of a high-precision method for identifying and classifying excitation inrush current waveforms of the present invention, where Class 1, 2, and 3 represent sine waves, triangle waves, and trapezoidal waves, respectively. The ROC curve shows that for Material 1, the four classifiers trained with the selected characteristic variables (kurtosis, skewness, and smoothed second-order derivative) all have a point (0,1) in the upper left corner, indicating that all examples of the three waveforms in Material 1 are correctly classified, and the classification accuracy reaches 100%, which is a perfect classifier.

[0174] The ROC curves of the other three materials are the same. Therefore, the high-precision excitation inrush current waveform identification and classification method proposed in the present invention is real and effective.

[0175] Example 2

[0176] In this embodiment, for material 1, skewness, smoothed second-order derivative, and average magnetic flux density are selected as characteristic variables for training classifiers; for material 2, skewness, average magnetic flux density, and smoothed second-order derivative are selected as characteristic variables for training classifiers; for material 3, skewness, coefficient of variation, and average magnetic flux density are selected as characteristic variables for training classifiers; for material 4, skewness, smoothed second-order derivative, and average magnetic flux density are selected as characteristic variables for training classifiers. The ROC curves of the four materials are shown as follows: Fig.14 , 15 , 16, and 17, where Class 1, 2, and 3 represent sine wave, triangle wave, and trapezoidal wave, respectively.

[0177] from Fig.14 , 15, 16, and 17, it can be seen that due to the differences in the selection of feature variables, the classification accuracy of the three waveforms of the four materials failed to reach 100%, but the accuracy of the soft voting integrated classifier was significantly better than the other three single classifiers as a whole, indicating that the selection of feature variables is crucial, and the performance of the soft voting integrated classifier is better than that of the single classifier.

[0178] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A high-precision method for identifying and classifying excitation inrush current waveforms, characterized in that: The following steps are involved: S1. Select 9 feature variables based on the preprocessed data; S2. For the 9 characteristic variables, a round of feature screening was performed using the U test; S3, using collinearity test and feature importance ranking to perform second round of feature screening; S4. Build a soft voting ensemble classifier based on RandomForest, GradientBoosting and Light-GBM; S5. Classify the excitation waveform based on the soft voting integrated classifier.

2. The high-precision identification and classification method of excitation inrush current waveform according to claim 1 is characterized in that: The step S1 comprises the following steps: S1.

1. Take 4 materials as material samples Y = {Y1, Y2, Y3, Y4}, and extract the magnetic flux density data of all sampling points of the 4 materials in a cycle under different working conditions; S1.2, based on the extracted data, set 9 feature variables as feature samples X = {X1, X2, ..., X9}, Among them, the characteristic variable X1, which represents the average magnetic flux density B avg , the value of the average magnetic flux density is equal to the sum of the magnetic flux density at each moment in a period T divided by the number of sampling points, expressed as follows: Among them, B i represents the magnetic flux density at the i-th moment, and n represents the total number of sampling points; Characteristic variable X2, which represents the standard deviation of magnetic flux density B std , which is expressed as follows: Characteristic variable X3, which represents the magnetic flux density skewness B skew , which is expressed as follows: Characteristic variable X4, which represents the kurtosis of magnetic flux density B kurt , which is expressed as follows: Characteristic variable X5, which represents the maximum value of magnetic flux density B max , which is expressed as follows: B max =max(B1,B2,...,B n ) (5) Characteristic variable X6, which represents the maximum value of magnetic flux density B min , which is expressed as follows: B min =min(B1,B2,,...,B n ) (6) The characteristic variable X7, which represents the magnetic flux density range Range, is expressed as follows: Range=B max -B min (7) The characteristic variable X8, which represents the coefficient of variation CV, is expressed as follows: Characteristic variable X9, which represents the smoothed second-order derivative, is used because the magnetic flux density data of each sampling point is discrete. 2 B is used to approximate the second-order derivative B″, which is expressed as follows: Δ 2 B=B i+2 -2B i+1 +B i (9) Here, h represents the step size between sampling points.

3. The high-precision identification and classification method of excitation inrush current waveform according to claim 2 is characterized in that: The step S2 comprises the following steps: S2.

1. Determine the U test hypothesis, assuming that the population distributions from which the two independent samples come are the same; S2.2, material Y j The data of any two of the nine feature variables under (j=1,2,3,4) are merged into one data set. The group data is sorted in ascending order and the sorted data is assigned a rank. If the data values ​​are the same, the average rank is assigned; S2.

3. Based on the merged data set and the ranking obtained after sorting, calculate the U statistic, which is expressed as follows: Among them, n1 and n2 are the sizes of the two groups of samples, r1 and r2 are the rank sums of the first and second groups of samples, respectively; S2.

4. Compare the U values ​​of the two characteristic variables, take U = min (U1, U2), and use statistical software or U test table to find the p value corresponding to the U value; S2.

5. If the p-value is less than the significance level of 0.05, the characteristic variable is retained, and the final number of characteristic variables retained is N1.

4. The high-precision identification and classification method of excitation inrush current waveform according to claim 3 is characterized in that: The step S3 comprises the following steps: S3.

1. Based on the N1 characteristic variables, the VIF values ​​are calculated respectively, as shown below: Among them, VIF h is the variance inflation factor of the hth feature variable, R h It is the coefficient of determination obtained by linear regression of the hth characteristic variable as the response variable with respect to other characteristic variables; S3.

2. Based on the VIF h , perform variance inflation factor VIF test, a VIF value higher than 10 indicates collinearity problem; S3.

3. If there is a feature variable with a VIF value higher than 10, then remove the feature variable with the largest VIF value, that is, N1-1= feature variables remain; S3.4, cyclically eliminate the feature variables with VIF values ​​greater than 10 in descending order of VIF values, until the VIF values ​​of all feature variables are less than 10, and finally retain N2 feature variables; S3.5, evaluate the importance of N2 feature variables; S3.

6. Retain the top three feature variables in importance score.

5. The high-precision identification and classification method of excitation inrush current waveform according to claim 4 is characterized in that: The step S4 comprises the following steps: S4.1, based on the feature data set, dividing the samples in the feature data set into a test set and a training set according to a specific ratio; S4.2, based on the training set, training the RandomForest classifier, the GradientBoosting classifier and the LightGBM classifier respectively; S4.

3. Set a weight coefficient υ for each classifier i i , which is expressed as follows: Among them, i=1 represents RandomForest classifier; i=2 represents GradientBoosting classifier; i=3 represents LightGBM classifier, AUC i is the probability of the three excitation waveforms being correctly predicted under classifier i; S4.

4. Weights υ based on each classifier i and the excitation waveform prediction probability p ik , respectively calculate the average predicted probability of waveform k, expressed as follows: Among them, P k is the weighted average predicted probability of waveform k, k = 1 represents a sine wave, k = 2 represents a triangle wave, k = 3 represents a trapezoidal wave, p ik represents the predicted probability of waveform k under classifier i; S4.

5. Integrate the trained RandomForest classifier, GradientBoosting classifier and LightGBM classifier into a soft voting classifier.

6. The high-precision identification and classification method of excitation inrush current waveform according to claim 5 is characterized in that: The step S5 comprises the following steps: S5.1, based on the weighted average predicted probability P k , sorting the weighted average prediction probabilities of the sine wave, the triangle wave, and the trapezoidal wave, and obtaining the sequence number with the highest weighted average prediction probability, including: According to formula (16), the sequence number of the feature variable with the highest weighted average prediction probability is calculated as follows: Where k is the final classified waveform label, k = 1 represents a sine wave, k = 2 represents a triangle wave, and k = 3 represents a trapezoidal wave; S5.2, based on the test set, using the soft voting integrated classifier to classify the excitation waveform of the test set; S5.

3. Use ROC curve to analyze the prediction results.

Images