Multi-sensor overhead ground wire fault detection and classification method based on random forest

Abstract

The invention discloses a multi-sensor overhead ground wire fault detection and classification method based on a random forest, and belongs to the technical field of overhead ground wire detection, and the method comprises the steps: carrying out the uniform segmentation of collected overhead ground wire current signal data, and constructing a sample data set; selecting a variable VMD modulus reasonable search range, and determining an optimal VMD modulus for each current signal; iMF components generated by VMD decomposition are screened, a numerical matrix is constructed, and the numerical matrix is converted into a grayscale image to serve as a key data set for random forest classifier training and testing; and training a random forest classifier by using the grayscale image training set, and verifying the effectiveness of the classifier by means of the test set. According to the multi-sensor overhead ground wire fault detection and classification method based on the random forest, the problems of non-stationary nonlinearity and noise interference of overhead ground wire fault signals are solved, the fault recognition accuracy of a single sensor is remarkably improved, troubleshooting and repairing can be rapidly and effectively carried out, and the safety of personnel and equipment is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of overhead ground wire detection, and in particular to a multi-sensor overhead ground wire fault detection classification method based on random forest. Background Art

[0002] In the power transmission system, the overhead ground wire plays a vital role. It not only protects the transmission line from natural disasters such as lightning strikes, but also ensures the stable operation of the power system to a certain extent. However, various faults will inevitably occur in the long-term operation of the overhead ground wire, such as discharge faults caused by lightning strikes, grounding faults caused by insulation degradation, and disconnection faults caused by mechanical damage. The traditional overhead ground wire fault detection method mainly relies on manual inspection and some simple electrical quantity monitoring methods. Manual inspection is inefficient and difficult to detect faults in real time. There is a large time lag and it cannot meet the requirements of modern power systems for rapid fault response. The monitoring method based on a single electrical quantity, such as monitoring only current or voltage, can reflect the operating status of the line to a certain extent, but it has limited ability to distinguish complex fault conditions, is prone to misjudgment or missed judgment, and cannot accurately distinguish different types of faults, thus affecting the subsequent fault repair strategy formulation.

[0003] With the continuous development of sensor technology, various types of sensors can be applied to the monitoring of overhead ground wires, such as temperature sensors, vibration sensors, current sensors, etc. These sensors can obtain the operation information of overhead ground wires from different angles, but how to effectively integrate these multi-source sensor data and extract valuable fault features from them, and then realize accurate fault detection and classification has become an urgent problem to be solved. Summary of the invention

[0004] The purpose of the present invention is to provide a multi-sensor overhead ground wire fault detection and classification method based on random forest, which can solve the problems of non-stationary nonlinearity and noise interference of overhead ground wire fault signals, significantly improve the fault identification accuracy of a single sensor, and can quickly and effectively perform fault troubleshooting and repair to ensure the safety of personnel and equipment.

[0005] To achieve the above object, the present invention provides a multi-sensor overhead ground wire fault detection classification method based on random forest, comprising the following steps:

[0006] S1. Using a current sensor to collect current signals of overhead ground wires, dividing the collected current signal data into equal intervals to construct a sample data set;

[0007] S2, select the search domain range of VMD modulus, calculate the maximum value of the envelope kurtosis of each sample data under each modulus, and determine the VMD optimal modulus corresponding to the current signal;

[0008] S3, performing VMD decomposition operation on each segment of data that has been divided in S1, and after decomposing and obtaining IMF components, carrying out corresponding screening work;

[0009] S4, respectively extracting the components with the maximum correlation coefficient with the original signal from the eigenmode components of each order, then normalizing the selected components, arranging them in a specific order, and accumulating to construct a numerical matrix;

[0010] S5, converting the numerical matrix constructed in S4 into a grayscale image format, and generating a plurality of grayscale images according to the time series data of each current signal, and using the generated plurality of grayscale images as a training set and a test set of a random forest algorithm;

[0011] S6. Determine the optimal model parameters of the random forest and determine the parameters of the tree through cross-validation;

[0012] S7. Use random partitioning to separate the training set from the test set, use the training set to train the random forest model, enable the model to learn data features and rules, and build an effective classification decision mechanism;

[0013] S8. Use the test set to verify the performance and effectiveness of the random forest classifier and obtain the final fault classification result.

[0014] Preferably, in S1, an equal-interval division operation is performed on the original current signal data according to an equidistant segmentation mode with a time interval of n.

[0015] Preferably, the search domain for the VMD modulus in S2 is set to K in the interval [2, 16].

[0016] Preferably, the detailed implementation process of S2 is as follows: a VMD decomposition operation is performed on the current signal collected by the current sensor, the modulus K is initially set to 2, the envelope kurtosis value of each modal signal under the set modulus K is calculated, and the maximum value of the envelope kurtosis under the modulus is obtained by comparison; then the process analysis is continued in a manner where K increases by 1 successively, that is, K=K+1, until the value of K reaches 16, so that the maximum value of the envelope kurtosis under each modulus can be obtained, and then the K value corresponding to the global maximum value of the envelope kurtosis can be accurately determined, thereby determining the most suitable VMD decomposition modulus.

[0017] Preferably, the specific implementation method of S3 is: use the correlation coefficient method to screen the decomposed IMF components, select the IMF component with the largest correlation coefficient as the data source for generating the grayscale image, and use the correlation coefficient method to effectively identify the IMF component that is most closely related to the original signal and the most representative.

[0018] Preferably, the correlation coefficient calculation formula of the IMF component is:

[0019]

[0020] Where l represents the length of the signal, and its value reflects the extension range of the signal in the time dimension; a(j) represents the jth segment of the original signal, which is the basic segmentation unit that constitutes the entire original signal sequence; I MF,i (j) is the j-th signal corresponding to the ith IMF component, which reflects the segmented signal situation under the specific IMF component; and r(i) is the quantitative value of the degree of association between the ith IMF component and the original signal a(t), that is, the correlation coefficient, which is obtained by comprehensive calculation of the above parameters and is used to measure the similarity and correlation between the IMF component and the original signal.

[0021] Preferably, S4 normalizes the screened IMF components according to the following formula:

[0022]

[0023] Among them, a and b are the values ​​before and after normalization respectively; F max , F min are the maximum and minimum values ​​of the original grayscale image respectively.

[0024] Preferably, in S4, the following specific method is used for converting the numerical matrix into a grayscale image: first, a grayscale image frame of size n×n is constructed, and then the width n of the image is accurately divided into two equal parts, thereby constructing two areas of n×n / 2; then, according to the size specifications of each of the two areas, the IMF components obtained by screening the sensor signals are arranged in a predetermined order and filled into the corresponding areas.

[0025] Preferably, the tree parameters in S6 include the number of trees, the maximum depth, the minimum number of sample splits, the minimum number of sample leaf nodes, and the minimum impurity reduction of leaf nodes.

[0026] Preferably, in the process of training the random forest in S7, the current signal classifier is trained with the help of the optimization characteristics of the Newton method.

[0027] Therefore, the present invention adopts the above-mentioned multi-sensor overhead ground wire fault detection and classification method based on random forest, which solves the problems of non-stationary nonlinearity and noise interference of overhead ground wire fault signals, significantly improves the fault identification accuracy of a single sensor, and can quickly and effectively perform fault troubleshooting and repair to ensure the safety of personnel and equipment.

[0028] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a flow chart of an overhead ground wire fault identification method according to an embodiment of a multi-sensor overhead ground wire fault detection classification method based on random forests of the present invention;

[0030] Figure 2 This is a confusion matrix display diagram of an embodiment of a multi-sensor overhead ground wire fault detection classification method based on random forest in the present invention. DETAILED DESCRIPTION

[0031] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0032] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0033] Embodiment 1

[0034] As an integrated learning algorithm, the random forest algorithm has excellent classification performance and robustness to data noise. It can process multivariate data, improve the accuracy and stability of classification by building multiple decision trees and integrating their results. When faced with complex overhead ground wire fault data, it is expected to fully tap the potential information in multi-sensor data, overcome the limitations of traditional methods, and achieve more efficient and accurate overhead ground wire fault detection and classification, thereby improving the safety and reliability of the power system and reducing operation and maintenance costs and fault losses.

[0035] Therefore, the present invention is based on a specific design concept. In actual work, the background noise of the fault signal is strong, and it is difficult to obtain comprehensive fault characteristics by analyzing the single sensor signal, which affects the recognition accuracy, and the sensor position selection depends on experience. Therefore, before random forest classification, the present invention uses VMD to pre-process the fault current signal, and arranges the screened IMF components in sequence at the feature level and converts them into grayscale images. Then the random forest structure is designed, the training set is input into the network training, and the network validity is verified with the test set, so as to realize the overhead ground wire fault detection and classification function. This design overcomes the shortcomings of traditional methods, integrates multi-sensor data and advanced algorithms, and improves the accuracy and reliability of fault detection and classification.

[0036] Based on the above design concept, the present invention proposes a multi-sensor overhead ground wire fault detection classification method based on random forest, such as Figure 1 As shown, the following steps are included:

[0037] S1. The current signal of the overhead ground wire is collected by using multiple current sensors, and then the collected current signal data is divided into equal intervals to construct a sample data set to provide a data basis for subsequent processing.

[0038] S2. Perform VMD decomposition processing on the collected current signal. Initially, set the modulus to K=2, and then calculate the envelope kurtosis value for each modal signal under the set modulus K. Through comparative analysis, determine the maximum value of the envelope kurtosis under this modulus, and then accurately determine the VMD optimal modulus corresponding to the current signal to ensure the effectiveness and accuracy of the decomposition. Then, continue the above calculation and comparison process in a manner that K increases by 1 (i.e., K=K+1) until the value of K reaches 16, thereby obtaining the maximum value of the envelope kurtosis under each modulus. On this basis, further determine the K value corresponding to the global maximum value of the envelope kurtosis. Among them, the envelope calculation formula for each mode is as follows:

[0039]

[0040] Where i is the i-th mode of K; is the absolute value (the result of HT of the i-th mode of K); is the i-th mode generated by VMD when the mode is K; t represents different moments. In addition, the envelope kurtosis of the i-th mode of K is calculated as follows:

[0041]

[0042] in, for The mean of for The standard deviation of; the numerator in formula (2) is The fourth-order central moment of .

[0043] The local maximum ek can be obtained K,max as follows:

[0044] ek K,max =max(ek 1 ,ek 2 ,…,ek K ) (3);

[0045] The search interval of K is set to [2, 16], and the search step is set to 1. In this way, 15 local peaks can be accurately obtained within this complete search range. Through careful comparison and in-depth analysis of these local peaks, the global maximum value ek can be successfully determined. g,max :

[0046] ek g,max =max(ek 2,max ,ek 3,max ,…ek 15,max ) (4);

[0047] According to formula (4), we can get ek g,max When , the corresponding K value is represented by K′, where K′ can be obtained from formula (5):

[0048] K'=argmax(ek g,max ) (5);

[0049] Among them, the argmax function represents the function parameter (set), so that ek g,max The value of the variable when it reaches its maximum value.

[0050] S3. In order to ensure that the constructed grayscale image can effectively retain the fault feature information in the original signal and avoid the adverse effects of interference components such as noise to the greatest extent, the correlation coefficient method is used to screen the various IMF components obtained after decomposition to obtain more representative and key information component data. By accurately calculating the correlation coefficient between each IMF component and the original signal, the IMF component with the largest correlation coefficient value is selected as the data basis for the subsequent generation of grayscale images. The calculation formula of the correlation coefficient is as follows:

[0051]

[0052] Where l represents the length of the signal, and its value reflects the extension range of the signal in the time dimension; a(j) represents the jth segment of the original signal, which is the basic segmentation unit that constitutes the entire original signal sequence; I MF,i (j) is the j-th signal corresponding to the ith IMF component, which reflects the segmented signal situation under the specific IMF component; and r(i) is the quantitative value of the degree of association between the ith IMF component and the original signal a(t), that is, the correlation coefficient, which is obtained by comprehensive calculation of the above parameters and is used to measure the similarity and correlation between the IMF component and the original signal.

[0053] S4. In order to compare the grayscale image more accurately and efficiently, the normalization process is carried out on the IMF components obtained by screening according to the following formula:

[0054]

[0055] Among them, a and b are the values ​​before and after normalization respectively; F max , F min are the maximum and minimum values ​​of the original grayscale image respectively.

[0056] S5. Construct a grayscale image of size n×n, and divide the image width n precisely into two equal parts, thereby constructing two regions of size n×n / 2. According to the specifications of the region, the IMF components screened out by each sensor signal are arranged in sequence and filled in, so as to successfully obtain the grayscale image of the current signal and realize the conversion and integration of data from signal to image.

[0057] S6. Determine the optimal model parameters of random forest. Overhead ground wire fault detection has multiple fault modes and complex signal characteristics. Through cross-validation, the number of trees is determined to be 800, the maximum depth (max_depth) is 15, the minimum number of sample splits (min_samples_split) is 7, the minimum number of sample leaf nodes (min_samples_leaf) is 4, and the minimum impurity reduction of leaf nodes (min_impurity_decrease) is 0.008.

[0058] S7. Randomly divide the two current signal grayscale images generated by S5, determine the training set and the test set respectively, and use the training set to train the random forest.

[0059] The core principle of random forest is to find a maximum margin hyperplane in the sample space to achieve the optimal segmentation effect for the given two types of training sample data. The process of determining the hyperplane is equivalent to solving a convex quadratic programming optimization problem, which is described as follows:

[0060]

[0061] Among them, x i , x j is the training sample, α i , α j is the Lagrange multiplier, y i is the sample x i The category, y j is the sample x j The category of m is the size of the sample. Since the radial basis function (RBF) is used as the kernel function, it can be expressed as:

[0062]

[0063] Among them, k(x i ,x j ) is the kernel function, and σ is the variance of the RBF kernel function. Based on the Gaussian kernel function in formula (9), the support vector expansion f(x) can be obtained by solving formula (8):

[0064]

[0065] Among them, k(x,x i) is the kernel function and b is the displacement of the separating hyperplane.

[0066] S8. After the random forest model is trained, the model is tested using the test set to achieve accurate classification of overhead ground wire faults. In the test set, 300 sets of data are included in each electrical fault and mechanical fault category at each current level, and their corresponding labels are "Electrical" and "Mechanical" respectively. The test set is input into the trained overhead ground wire fault detection model for experimental verification. The verification results are presented in Figure 2 Among them, Figure 2 This is a display of the confusion matrix constructed using the test results at different current levels, where Figure 2 Figure (a) is the confusion matrix of the 300A test set. Figure 2 Figure (b) is the confusion matrix of the 1000A test set. Figure 2 Figure (c) in is the confusion matrix of the 1500A test set. Figure 2 Figure (d) is the confusion matrix of the 2000A test set. It can be found that the fault detection accuracy at different current levels is not completely consistent, and the distribution of misjudgments between electrical faults and mechanical faults is also uneven. The main reason is that the spectrum distribution of fault currents at different current levels is different. However, from an overall perspective, the detection accuracy of electrical faults and mechanical faults exceeds 99% at all current levels. In summary, the random forest multi-sensor overhead ground wire fault diagnosis method based on VMD preprocessing can effectively achieve the fault diagnosis goal and has high stability.

[0067] The present invention has the following advantages: by integrating multi-sensor data and applying the random forest algorithm, the problem of inaccurate complex fault identification by traditional electrical quantity monitoring methods is overcome, and different types of overhead ground wire faults can be accurately identified, effectively reducing misjudgments and missed judgments, improving the accuracy of fault diagnosis, and ensuring the safe and stable operation of the power system; compared with traditional manual inspections, this technology realizes the automation and efficiency of fault detection, can collect and process data in real time, quickly discover potential fault hazards, reduce the risk of power outages due to delayed detection, and improve the reliability and continuity of power supply; accurate fault detection and classification optimizes the operation and maintenance strategy of the power system, reduces unnecessary inspections, reduces manpower and time costs, avoids small faults from turning into large faults, improves the stability, efficiency and economic benefits of the power system, and promotes the sustainable development of the power industry.

[0068] Therefore, the present invention adopts the above-mentioned multi-sensor overhead ground wire fault detection and classification method based on random forest, which solves the problems of non-stationary nonlinearity and noise interference of overhead ground wire fault signals, significantly improves the fault identification accuracy of a single sensor, and can quickly and effectively perform fault troubleshooting and repair to ensure the safety of personnel and equipment.

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A multi-sensor overhead ground wire fault detection classification method based on random forest, characterized in that: The following steps are involved: S1. Using a current sensor to collect current signals of overhead ground wires, dividing the collected current signal data into equal intervals to construct a sample data set; S2, select the search domain range of VMD modulus, calculate the maximum value of the envelope kurtosis of each sample data under each modulus, and determine the VMD optimal modulus corresponding to the current signal; S3, performing VMD decomposition operation on each segment of data that has been divided in S1, and after decomposing and obtaining IMF components, carrying out corresponding screening work; S4, respectively extracting the components with the maximum correlation coefficient with the original signal from the eigenmode components of each order, then normalizing the selected components, arranging them in a specific order, and accumulating to construct a numerical matrix; S5, converting the numerical matrix constructed in S4 into a grayscale image format, and generating a plurality of grayscale images according to the time series data of each current signal, and using the generated plurality of grayscale images as a training set and a test set of a random forest algorithm; S6. Determine the optimal model parameters of the random forest and determine the parameters of the tree through cross-validation; S7. Use random partitioning to distinguish the training set from the test set, use the training set to train the random forest model, enable the model to learn data features and rules, and build an effective classification decision mechanism; S8. Use the test set to verify the performance and effectiveness of the random forest classifier and obtain the final fault classification result.

2. The multi-sensor overhead ground wire fault detection classification method based on random forest according to claim 1 is characterized by: In S1, an equal-interval division operation is performed on the original current signal data according to an equal-interval segmentation mode with a time interval of n.

3. The multi-sensor overhead ground wire fault detection classification method based on random forest according to claim 1 is characterized in that: The search domain for the VMD modulus in S2 is set to K in the interval [2, 16].

4. The multi-sensor overhead ground wire fault detection classification method based on random forest according to claim 3 is characterized by: The detailed implementation process of S2 is as follows: VMD decomposition operation is performed on the current signal collected by the current sensor, the modulus K is initially set to 2, the envelope kurtosis value of each modal signal under the set modulus K is calculated, and the maximum value of the envelope kurtosis under the modulus is obtained by comparison; then, the process analysis is continued in a manner of K increasing by 1, that is, K=K+1, until K reaches 16, so that the maximum value of the envelope kurtosis under each modulus can be obtained, and then the K value corresponding to the global maximum value of the envelope kurtosis can be accurately determined, thereby determining the most suitable VMD decomposition modulus.

5. The multi-sensor overhead ground wire fault detection classification method based on random forest according to claim 1 is characterized by: The specific implementation method of S3 is: use the correlation coefficient method to screen the decomposed IMF components, select the IMF component with the largest correlation coefficient as the data source for generating the grayscale image, and use the correlation coefficient method to effectively identify the IMF component that is most closely related to the original signal and the most representative.

6. The multi-sensor overhead ground wire fault detection classification method based on random forest according to claim 5 is characterized by: The correlation coefficient calculation formula of the IMF component is: Where l represents the length of the signal, and its value reflects the extension range of the signal in the time dimension; a(j) represents the jth segment of the original signal, which is the basic segmentation unit that constitutes the entire original signal sequence; I MF,i (j) is the j-th signal corresponding to the ith IMF component, which reflects the segmented signal situation under the specific IMF component; and r(i) is the quantitative value of the degree of association between the ith IMF component and the original signal a(t), that is, the correlation coefficient, which is obtained by comprehensive calculation of the above parameters and is used to measure the similarity and correlation between the IMF component and the original signal.

7. The multi-sensor overhead ground wire fault detection classification method based on random forest according to claim 1 is characterized by: The S4 performs normalization processing on the screened IMF components according to the following formula: Among them, a and b are the values ​​before and after normalization respectively; F max , F min are the maximum and minimum values ​​of the original grayscale image respectively.

8. The multi-sensor overhead ground wire fault detection classification method based on random forest according to claim 1 is characterized by: In said S4, the conversion of the numerical matrix into a grayscale image is carried out in the following specific manner: firstly, a grayscale image frame of size n×n is constructed, and then the width n of the image is precisely divided into two equal parts, thereby constructing two regions of n×n / 2; Then, according to the size specifications of the two areas, the IMF components obtained by filtering the sensor signals are arranged in a predetermined order and filled into the corresponding areas.

9. The multi-sensor overhead ground wire fault detection classification method based on random forest according to claim 1 is characterized by: The parameters of the tree in S6 include the number of trees, the maximum depth, the minimum number of sample splits, the minimum number of sample leaf nodes, and the minimum impurity reduction of leaf nodes.

10. The multi-sensor overhead ground wire fault detection classification method based on random forest according to claim 1, characterized in that: In the process of training the random forest in S7, the current signal classifier is trained with the help of the optimization characteristics of the Newton method.

Images