Improved oversampling-based power transmission line fault classification system and method

Abstract

Disclosed in the present invention are an improved oversampling-based power transmission line fault classification system and method. The method of the present invention comprises: first, by using historical fault information of State Grid as a database, obtaining a plurality of sub-datasets by means of division based on fault types, each sub-dataset consisting of transient current traveling waves at power transmission line fault moments; then, using a K-means method to perform clustering operation on minority classes of a fault dataset, and using BSMOTE to perform oversampling, so as to achieve balancing of various classes, thus forming a balanced dataset; then, calculating time domain and frequency domain features of transient current traveling waves in the balanced dataset, so as to form a feature dataset; finally, on the basis of CNN and SVM deep learning algorithms, establishing a classification model and training same by using the feature dataset, so as to obtain a trained fault classification model; and using the trained fault classification model to classify and identify faults. The method is highly accurate and effective, and can effectively extract fault feature quantities and quickly and accurately classify faults.

Description

Power transmission line fault classification system and method based on improved oversampling TECHNICAL FIELD

[0001] The present application relates to the field of power grid system power transmission line fault diagnosis, in particular to the aspects of fault feature extraction and fault type identification, and specifically to a power transmission line fault classification system and method based on improved oversampling. BACKGROUND

[0002] Power transmission lines bear the heavy responsibility of transmitting electric energy in the power grid and are a key link to ensure stable operation of the power grid. The overhead line-dominant power transmission lines are exposed to the wild for a long time, have the characteristics of many points, long lines, wide surfaces and complex working environment, and thus determine that the power transmission lines are the most prone to faults in the power system. The line hazards or line faults caused by lightning or non-lightning strikes of the power transmission lines bring economic losses to the national life and production. Identifying the fault types of the power transmission lines can help dispatchers to predict the power restoration time and make scheduling arrangements, provide technical assistance for line inspection and recording, and be conducive to quickly completing fault elimination so as to reduce the time required for power restoration. Therefore, correct fault type identification is crucial to improving stability and power transmission capacity.

[0003] Synthetic Minority Over-sampling Technique (SMOTE) is an improved algorithm based on traditional random oversampling, which avoids the large number of repeated occurrence of minority class samples in the data set and effectively reduces the probability of overfitting problem. However, the SMOTE algorithm does not select the minority class samples differently, so there are problems such as blindness and marginalization when synthesizing new samples.

[0004] The transmission line fault identification classification mainly includes two interlocking parts: fault feature research and classification method research. The transient waveform characteristics of the line are derived from the electrical quantity changes of the fault recording, which usually needs to be processed with digital signal processing technology for continuous value signal processing. Common digital signal transformation methods include Fourier transform, wavelet transform, modal decomposition, etc. Common time sequence waveform classification methods include support vector machine (SVM), logistic regression, adaptive enhancement (Adaboost), artificial neural network and deep learning algorithm. Convolutional neural network (CNN) develops from neural network. The CNN model increases the convolution layer and the pooling layer on the basis of the traditional neural network model. Through the convolution layer and the pooling layer to replace part of the full connection layer, and relying on the weight sharing and local perception field characteristics of the CNN model, the purpose of simplifying the model operation and improving the model training effect is achieved. The traditional CNN model uses the softmax layer as the output layer, and the essence of this layer is a layer of full connection layer, which is used to calculate the probability of different categories through linear combination of feature parameters, and finally realizes classification. The classification effect of the softmax layer is not as good as the classification algorithm based on traditional machine learning when it is used for non-linear and sparse data, so the SVM classifier is used to replace the softmax layer.

[0005] SUMMARY

[0006] The purpose of the present application is to provide a transmission line fault classification system and method based on improved oversampling. The present application proposes a BSMOTE improved oversampling method, and constructs a classification model based on CNN and SVM deep learning algorithm. The CNN and SVM model increases the attention layer after the pooling layer of the traditional CNN model. The main function of this layer is to further improve the selection of feature parameters. By updating the weight of the feature parameters, the feature parameters with better class distinction ability are selected, so as to improve the efficiency and accuracy of identification and classification.

[0007] To achieve this purpose, the power transmission line fault classification system based on improved oversampling designed by the application comprises a fault data set generation module, a fault balanced data set generation module, a fault feature data set generation module and a fault classification model establishment module; the fault data set generation module is used for acquiring power transmission line fault transient current traveling wave signals, classifying the power transmission line fault transient current traveling wave signals according to fault types, and forming a fault data set; the fault balanced data set generation module is used for clustering and oversampling the fault data set formed by the fault data set generation module, and forming a fault balanced data set; the fault feature data set generation module is used for extracting time domain and frequency domain features of the transient current traveling wave signals in the fault balanced data set formed by the fault balanced data set generation module, and forming a fault feature data set; and the fault classification model establishment module is used for training a classification model based on CNN and SVM algorithms according to the fault feature data set obtained by the fault feature data set generation module, obtaining a trained fault classification model, and inputting a to-be-detected power transmission line fault transient current traveling wave signal into the fault classification model to obtain a fault type.

[0008] A power transmission line fault classification method based on improved oversampling, which comprises the following steps: step 1, acquiring power transmission line fault transient current traveling wave signals, classifying the power transmission line fault transient current traveling wave signals according to fault types, and forming a fault data set; step 2, clustering and oversampling the fault data set, and forming a fault balanced data set; step 3, extracting time domain and frequency domain features of the transient current traveling wave signals in the fault balanced data set, and forming a fault feature data set; and step 4, training a classification model based on CNN and SVM algorithms according to the fault feature data set, obtaining a trained fault classification model, and inputting a to-be-detected power transmission line fault transient current traveling wave signal into the fault classification model to obtain a fault type.

[0009] The application has the following beneficial effects: compared with the traditional SMOTE oversampling method, the new samples generated by the BSMOTE improved oversampling method proposed by the application will move towards the center of gravity, thereby moving away from the decision boundary, effectively overcoming the marginalization problem of the new samples generated by the SMOTE algorithm, and further improving the learning ability of the classification model for the minority class samples. The CNN and SVM model proposed by the application increases the attention layer after the pooling layer of the traditional CNN model, which can further improve the selection ability of the feature parameters and improve the efficiency and accuracy of the identification and classification. BRIEF DESCRIPTION OF DRAWINGS

[0010] Fig. 1 is a system structure diagram of the application;

[0011] Fig. 2 is a flowchart of a power transmission line fault unbalanced data oversampling and classification method;

[0012] Fig. 3 is a flow chart of a power transmission line fault unbalanced data K-means clustering specific implementation;

[0013] Fig. 4 is a schematic diagram of the BSMOTE algorithm generating new samples;

[0014] Fig. 5 is a schematic diagram of the CNN and SVM classification model structure;

[0015] Among them, 1-fault data set generation module, 2-fault balanced data set generation module, 3-fault feature data set generation module, 4-fault classification model establishment module. DETAILED DESCRIPTION

[0016] The application will be further described in detail below in combination with the drawings and specific embodiments:

[0017] In order to improve the learning effect of the classification model on the minority class samples and better realize the selection of feature parameters, improve the efficiency and accuracy of recognition classification, the application proposes an improved BSMOTE oversampling method, and constructs a classification model based on CNN and SVM deep learning algorithm. The CNN and SVM model increases an attention layer after the pooling layer of the traditional CNN model. The main function of the layer is to further improve the selection of feature parameters. By updating the weight of the feature parameters, the feature parameters with better class distinction ability are selected, so as to improve the efficiency and accuracy of recognition classification.

[0018] Embodiment 1

[0019] A power transmission line fault classification system based on improved oversampling, as shown in Fig. 1, includes a fault data set generation module 1, a fault balanced data set generation module 2, a fault feature data set generation module 3 and a fault classification model establishment module 4:

[0020] The fault data set generation module 1 is used to obtain the power transmission line fault transient current traveling wave signal, and the power transmission line fault transient current traveling wave signal is summarized according to the fault type to form a fault data set;

[0021] The fault balanced data set generation module 2 is used to cluster and oversample the fault data set formed by the fault data set generation module 1 to form a fault balanced data set;

[0022] The fault feature data set generation module 3 is used to extract the time domain and frequency domain features of the transient current traveling wave signal in the fault balanced data set formed by the fault balanced data set generation module 2 to form a fault feature data set;

[0023] The fault classification model establishment module 4 is used for training a classification model established based on a CNN and a SVM algorithm according to a fault feature data set obtained by the fault feature data set generation module 3, inputting a to-be-detected transmission line fault transient current traveling wave signal into the model to obtain a fault type predicted by the model, and then realizing classification identification of the transmission line fault, and the overall implementation process is as shown in FIG. 2.

[0024] In the technical solution, the fault types summarized by the fault data set generation module 1 include lightning strike faults and non-lightning strike faults; the non-lightning strike faults include windage yaw faults, ice damage faults and bird damage faults; and the transient current traveling wave signal sequence length in this embodiment is 1250.

[0025] In the technical solution, the fault balanced data set generation module 2 uses a K-means clustering algorithm for clustering, has good expansibility and efficiency for large data sets, is easy to implement and has relatively small calculation amount, and is therefore very effective in processing large-scale data, as shown in FIG. 3, and the specific implementation method is as follows:

[0026] Step 2.1, for the initial non-lightning strike fault sample set D1 in the fault data set, randomly extract k samples (k=2 for the first clustering) x1, x2,..., x k as the clustering center points;

[0027] Step 2.2, for the remaining samples in the initial non-lightning strike fault sample set D1 except the clustering centers x1, x2,..., x k , calculate the Euclidean distances of the remaining samples and the center points x j (j=1, 2,..., k) respectively by the following formula, and divide the remaining samples into the cluster C u (u=1, 2,..., k) where the Euclidean distance of the remaining sample is the smallest, and the calculation method of the Euclidean distance is as follows:

[0028] In the formula, d(x i ,x j ) is the Euclidean distance of the sample x i and the sample x j , x i =(x i1 ,x i2 ,...,x ih ) represents the i-th h-dimensional data sample; and x j =(x j1 ,x j2 ,...,x jh ) represents the j-th h-dimensional data sample.

[0029] Step 2.3. Calculate the average value of each cluster C obtained in step 2.2. u The average value m of the samples in (u = 1, 2,..., k) u Set the obtained average value as the center value of the cluster corresponding to the average value.

[0030] In the formula, m u The average value m of the data in cluster C u (u = 1, 2,..., k), x p The sample in cluster C u (u = 1, 2,..., k).

[0031] Step 2.4. Calculate the sum of squared errors, specifically, calculate the sum of squared errors between each sample in the initial non-lightning fault sample set D1 and the center value of the cluster to which the sample belongs, and the calculation formula is:

[0032] In the formula, SSE is the sum of squared errors, m u The average value m of the data in cluster C u (u = 1, 2,..., k), x p The sample in cluster C u (u = 1, 2,..., k).

[0033] Step 2.5. Iteratively execute steps 2.2 to 2.5 until the clustering division no longer changes or the SSE value tends to a certain limit value, at which point the clustering ends to obtain k clusters. If the sample quantity of each cluster ≥ m, then the value of k is increased by 1 and the clustering is restarted in step 2.1. If the sample quantity of a certain cluster < m, then the value of k is decreased by 1 and the clustering is restarted in step 2.1. The clustering ends when the value of k satisfies that the sample quantity of each cluster ≥ m and the value of k+1 does not satisfy that the sample quantity of each cluster ≥ m, and outputs the K-means clustering subset C u (u = 1, 2,..., k). In this embodiment, m is preferably 3.

[0034] In the above technical solution, the fault balance data set generation module 2 uses the improved SMOTE algorithm for oversampling to expand the clustered unbalanced data. The improved SMOTE algorithm only generates new samples between the triangular centers and vertices constructed in each cluster, so the generated new samples will move towards the center position, overcoming the edge problem of the SMOTE algorithm in generating new samples. As shown in FIG. 4, the specific implementation method is:

[0035] Step 3.1. Randomly select m samples x u ’(u = 1, 2,..., k) in any cluster C r ’(u = 1, 2,..., k) of the clustered fault data set. sx t Constructing m-gon vertex x m (m=r,s,...,t), calculate its barycentric sample x B , barycentric sample x B The calculation formula is as follows:

[0036] In the formula, x B represents the h-dimensional barycentric vector of the m-gon composed of vertices x m (m=r,s,...,t), x j =(x j1 ,x j2 ,...,x jh ) represents the jth h-dimensional data sample.

[0037] Step 3.2, randomly generate a new non-lightning fault sample x B between the m-gon barycentric sample x m and the vertex x new (m=r,s,...,t), the calculation formula is as follows: x new =x m +rand(0,1)*(x B -x m )

[0038] Step 3.3, according to the sampling ratio, execute steps 3.1 and 3.2 in the k clusters C u '(u=1,2,...,k) of the clustered fault data set in a loop until the initial non-lightning fault sample quantity in the clustered fault data set is not less than the lightning fault sample quantity. The sampling ratio is majority class sample quantity / minority class sample quantity (integer+1, such as majority class sample quantity / minority class sample quantity=5.3, taking 6 as the sampling ratio), and the loop will be exited when the minority class sample quantity is equal to the majority class sample quantity.

[0039] In the above technical solution, the fault feature data set generation module 3 is used to extract the time domain features and frequency domain features of the transient current row wave signal in the fault balanced data set formed by the fault balanced data set generation module 2, obtain signal time domain features and frequency domain features, and generate a fault feature data set. When analyzing high-dimensional data, the model is prone to overfitting, that is, overfitting to training data and unable to generalize to new samples. Feature extraction of data can play a role in data dimension reduction, and these features can reflect various properties of signals and play a key role in fault recognition algorithms.

[0040] The time domain features include waveform mean a1, root mean square a2, square root amplitude a3, absolute value mean a4, nonlinear characteristic a5, higher order nonlinear characteristic a6, variance a7, peak value a8, minimum value a9, peak to peak value a 10 , ratio of root mean square and absolute value mean a 11 , ratio of root mean square and peak value a 12 , ratio of peak value and absolute value mean a 13 , ratio of peak value and overall amplitude a 14 , relationship between nonlinear characteristic and fluctuation a 15 , relationship between higher order nonlinear characteristic and fluctuation a 16 ; mainly focusing on the amplitude, amplitude, duration, etc. of the signal, the expression of the time domain feature is: a8 = max|x(i)| a9 = min|x(i)| a 10 = a8-a9

[0041] In the formula, i = 1, 2, …, N is the sample sampling point number, x(i) is the time domain signal sequence of the transient current traveling wave signal; a1 represents the mean of all elements of the time domain signal sequence of the transient current traveling wave signal, reflecting the average level of the signal in the entire time range; a2 represents the square root of the mean of all elements of the time domain signal sequence of the transient current traveling wave signal, representing the root mean square value of the signal, which is used to describe the effective value of the signal; a3 represents the square of the mean of the square root of the absolute value of all elements of the time domain signal sequence of the transient current traveling wave signal, which is used to describe the overall amplitude of the signal; a4 represents the mean of the absolute value of all elements of the time domain signal sequence of the transient current traveling wave signal, indicating the average level of the absolute value of the signal; a5 represents the mean of the cube of all elements of the time domain signal sequence of the transient current traveling wave signal minus a1, reflecting the nonlinear characteristic of the signal; a6 represents the mean of the fourth power of all elements of the time domain signal sequence of the transient current traveling wave signal minus a1, which is used to detect the higher order nonlinear characteristic of the signal; a7 represents the mean of the square of all elements of the time domain signal sequence of the transient current traveling wave signal minus a1, indicating the variance of the signal, which is used to measure the fluctuation of the signal; a8 represents the maximum value of the absolute value of all elements of the time domain signal sequence of the transient current traveling wave signal, which is used to detect the peak value of the signal; a9 represents the minimum value of the absolute value of all elements of the time domain signal sequence of the transient current traveling wave signal, which is used to detect the minimum value of the signal; a 10 represents the difference between a8 and a9, indicating the range between the peak value and the minimum value of the signal; a 11 represents the ratio of a2 and a4, which is used to describe the peak property of the signal; a 12 represents the ratio of a8 and a2, which is used to compare the root mean square and the peak value of the signal; a 13Ratio of a8 and a4, used to compare the peak value and the mean value of the absolute value of the signal;a 14 Ratio of a8 and a3, used to compare the peak value and the overall amplitude of the signal;a 15 Ratio of the cube of the square root of a5 and a7, used to describe the relationship between the non-linear characteristics and the fluctuation of the signal;a 16 Ratio of the fourth power of the square root of a6 and a7, used to describe the relationship between the higher-order non-linear characteristics and the fluctuation of the signal; wherein, a1-a4 and a8-a 10 Indicates the time-domain vibration amplitude and energy size, a5-a7 and a 11 ~a 16 Indicates the time series distribution of the time-domain signal.

[0042] The frequency domain features include a mean value b1 of the frequency domain signal, a variance b2 of the frequency domain signal, a skewness b3 of the frequency domain signal, a kurtosis b4 of the frequency domain signal, a power spectral density b5 of the frequency domain signal, a root mean square bandwidth b6 of the frequency domain signal, a variance bandwidth b7 of the frequency domain signal, a kurtosis bandwidth b8 of the frequency domain signal, a pulse factor b9 of the frequency domain signal, a relationship b 10 between the bandwidth and the root mean square of the frequency domain signal, a sharp factor b 11 of the frequency domain signal, a peak-to-peak value factor b 12 of the frequency domain signal, and a waveform factor b 13 of the frequency domain signal; the specific expression is:

[0043] In the formula, n=1, 2, …, K, n is the spectral line number, K is the total number of spectral lines, s(n) is the frequency amplitude value of the n-th spectral line of the transient current traveling wave signal, f n is the frequency value of the n-th spectral line of the transient current traveling wave signal; b1 represents the mean value of the frequency domain signal, indicating the average level of the frequency domain signal; b2 represents the variance of the frequency domain signal, used to measure the dispersion degree of the frequency domain signal; b3 represents the skewness of the frequency domain signal, used to describe the asymmetry of the frequency domain signal; b4 represents the kurtosis of the frequency domain signal, used to describe the peak value property of the frequency domain signal; b5 represents the power spectral density of the frequency domain signal, indicating the power distribution of the signal; b6 represents the root mean square bandwidth of the frequency domain signal, used to represent the bandwidth characteristics of the frequency domain signal; b7 represents the variance bandwidth of the frequency domain signal, used to represent the bandwidth characteristics of the frequency domain signal; b8 represents the kurtosis bandwidth of the frequency domain signal, used to represent the bandwidth characteristics of the frequency domain signal; b9 represents the pulse factor of the frequency domain signal, used to represent the pulse characteristics of the signal; b 10 represents the ratio of b6 and b5, used to describe the relationship between the bandwidth and the root mean square of the frequency domain signal; b 11 represents the sharp factor of the frequency domain signal, used to describe the sharp peak property of the signal; b 12 represents the peak-to-peak value factor of the frequency domain signal, used to describe the peak-to-peak value property of the signal; b13 Waveform factor representing the frequency domain signal, used to describe the fluctuation property of the signal; wherein b1 represents the size of the frequency domain vibration energy, b2~b4, b6, b 10 ~b 13 Indicates the concentration or dispersion degree of the spectrum, b5, b7~b9 indicates the change of the main frequency band position.

[0044] In the above technical solution, the fault feature data set generation module 3 is used to divide the fault feature data set into a training set and a test set according to a set proportion, and the set proportion is preferably 8:2.

[0045] In the above technical solution, the fault classification model establishment module 4 establishes a classification model based on CNN and SVM algorithms, as shown in FIG. 5, the classification model includes an input layer, a convolution layer, a pooling layer, an attention layer, a full connection layer and an output layer; the advantages of CNN and SVM algorithms are that CNN can automatically learn the feature representation of data, and SVM can effectively process high-dimensional data and nonlinear classification problems. In addition to the advantages in data classification accuracy, the CNN and SVM method also has strong robustness and generalization ability.

[0046] The input layer uses the training set divided by the fault feature data set generation module 3 as input to train the classification model;

[0047] The convolution layer completes the selection of feature parameters through two characteristics of local perception field and weight sharing, wherein the convolution kernel size is set to 3 and the number is 32;

[0048] The pooling layer is used to reduce the dimension of the data after convolution of the convolution layer, so as to realize further extraction of features. Common pooling methods include maximum pooling and average pooling, that is, extracting the average value and maximum value of the local feature vector. The global average pooling is adopted in the present application;

[0049] The attention layer continuously trains the input data through the attention mechanism, increases the weight of the feature parameter with strong class distinction ability, and finally achieves the purpose of improving the overall classification effect. By reducing the weight of unimportant feature parameters, the influence of non-important features in the feature set can be reduced;

[0050] The full connection layer accesses the output of the attention layer, and adopts full connection to connect each neuron with all the neurons of the previous layer. In the CNN and SVM model, the full connection layer is after the attention layer, and the output of the attention layer is used as the input. In order to improve the performance of the convolutional neural network, the full connection layer selects ReLu function as the activation function;

[0051] The output layer inputs the output of the full connection layer into the SVM classifier for classification.

[0052] In another embodiment, the power line fault classification system based on improved oversampling described above comprises a processor, wherein the processor is configured to execute the above-mentioned program modules existing in the memory, including a fault data set generation module 1, a fault balanced data set generation module 2, a fault feature data set generation module 3, and a fault classification model establishment module 4.

[0053] Embodiment 2

[0054] A power line fault classification method based on improved oversampling using the above-mentioned system comprises the following steps:

[0055] Step 1, obtaining power line fault transient current traveling wave signals, classifying the power line fault transient current traveling wave signals according to fault types to form a fault data set;

[0056] Step 2, clustering and oversampling the fault data set formed in the step 1 to form a fault balanced data set;

[0057] Step 3, extracting time domain and frequency domain features of the transient current traveling wave signals in the fault balanced data set formed in the step 2 to form a fault feature data set;

[0058] Step 4, training the classification model based on CNN and SVM algorithm according to the fault feature data set obtained in the step 3 to obtain a trained fault classification model, and inputting the to-be-detected power line fault transient current traveling wave signals into the fault classification model to obtain the fault type.

[0059] Embodiment 3

[0060] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to realize the steps of the method of embodiment 2.

[0061] Embodiment 4

[0062] A computer program product comprises a computer program, and the computer program is executed by a processor to realize the steps of the method of embodiment 2.

[0063] The contents not described in detail in the specification belong to the prior art known to those skilled in the art. Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can be in the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can be in the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0064] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks or in the flowchart one or more blocks and / or in the block or blocks of the block diagram.

[0065] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart block or blocks or in the flowchart one or more blocks and / or in the block or blocks of the block diagram.

[0066] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks or in the flowchart one or more blocks and / or in the block or blocks of the block diagram.

[0067] Finally, it should be noted that the above-described embodiments are merely used to illustrate the technical solutions of the present application, but are not intended to limit the scope of protection of the present application. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that: after reading the present application, those skilled in the art can make various changes, modifications or equivalent replacements to the specific embodiments of the present application. However, these changes, modifications or equivalent replacements are all within the scope of protection of the claims of the present application.

Claims

1. A transmission line fault classification system based on improved oversampling, characterized by: It includes a fault data set generation module (1), a fault balance data set generation module (2), a fault feature data set generation module (3) and a fault classification model establishment module (4): The fault data set generation module (1) is used to obtain the transmission line fault transient current traveling wave signal, summarize the transmission line fault transient current traveling wave signal according to the fault type, and form a fault data set; The fault balance data set generation module (2) is used to cluster and oversample the fault data set generated by the fault data set generation module (1) to form a fault balance data set; The fault feature data set generation module (3) is used to extract the time domain and frequency domain features of the transient current traveling wave signal in the fault balance data set generated by the fault balance data set generation module (2) to form a fault feature data set; The fault classification model establishment module (4) is used to train a classification model established based on the CNN and SVM algorithms according to the fault feature data set obtained by the fault feature data set generation module (3) to obtain a trained fault classification model, and input the transient current traveling wave signal of the transmission line fault to be detected into the fault classification model to obtain the fault type.

2. The improved oversampling-based transmission line fault classification system according to claim 1, characterized in that: The fault types summarized by the fault data set generation module (1) include lightning faults and non-lightning faults; the non-lightning faults include windage faults, ice damage faults and bird damage faults.

3. The improved oversampling-based transmission line fault classification system according to claim 2, characterized in that: The fault balance data set generation module (2) uses the K-means clustering algorithm to cluster the fault data set. The specific implementation method is as follows: Step 2.1: For the initial non-lightning fault sample set D1 in the fault data set, randomly select k samples. When clustering for the first time, k = 2, x1, x2, ..., x k , as the center point of the cluster; Step 2.2: For the remaining samples x1, x2, ..., x in the initial non-lightning fault sample set D1 except the cluster center, k , respectively calculate the remaining samples and the center point x j (j=1,2,...,k) and divide the remaining samples into cluster C where the cluster center with the smallest Euclidean distance to the remaining samples is located u (u=1,2,...,k), the Euclidean distance is calculated as follows: Where, d(x i ,x j ) is the sample x i With sample x j The Euclidean distance, x i =(x i1 ,x i2 ,…,x ih ) represents the i-th h-dimensional data sample; x j =(x j1 ,x j2 ,…,x jh ) represents the jth h-dimensional data sample; Step 2.3, calculate each cluster C obtained in step 2.2 i The mean value m of the samples in (i=1,2,…,k) i , setting the obtained average value as the center value of the cluster corresponding to the average value; Where m i Cluster C i The average value of the data in (i=1,2,…,k), x p Cluster C i samples in (i=1,2,…,k); Step 2.4, calculate the sum of squared errors between each sample in the initial non-lightning fault sample set D1 and the center value of the cluster to which it belongs, using the following formula: Where SSE is the sum of squared errors, m i Cluster C i The average value of the data in (i=1,2,…,k), x p Cluster C i samples in (i=1,2,…,k); Step 2.5, iteratively execute steps 2.2 to 2.5 until the cluster division no longer changes or the error sum of squares tends to a certain limit value. At this point, clustering is completed and k clusters are obtained. If the sample size of each cluster is ≥m, the k value is increased by 1 and the process returns to step 2.1 for re-clustering. If the sample size of each cluster is <m, the k value is reduced by 1 and the process returns to step 2.1 for re-clustering. Clustering is completed when the k value satisfies the sample size of each cluster ≥m and k+1 does not satisfy the sample size of each cluster ≥m. Output the K-means cluster subset C i '(i=1,2,...,k).

4. The improved oversampling-based transmission line fault classification system according to claim 2, characterized in that: The fault balance data set generation module (2) uses the improved SMOTE algorithm to perform oversampling to expand the unbalanced data in the clustered fault data set, specifically: Step 3.1, in any cluster C of the fault data set after clustering u '(u=1,2,...,k) randomly select m samples x r ,x s ,..,x t Construct the vertex x of the m-gon m (m=r,s,...,t), calculate its centroid sample x B , centroid sample x B The calculation formula is as follows: Where x B Represented by vertex x m The h-dimensional centroid vector of the m-gon formed by (m=r,s,...,t), x j =(x j1 ,x j2 ,…,x jh ) represents the jth h-dimensional data sample; Step 3.2, sample x at the center of the m-gon B With vertex x m A new non-lightning fault sample x is randomly generated between (m=r,s,...,t) new , the calculation formula is as follows: x new =x m +rand(0,1)*(x B -x m ) Step 3.3, according to the sampling rate, k clusters C of the fault data set after clustering i '(i=1,2,...,k) loop execution Step 3.1 and step 3.2 are performed until the number of initial non-lightning fault samples in the clustered fault data set is not less than the number of lightning fault samples.

5. The improved oversampling-based transmission line fault classification system according to claim 1, characterized in that: The fault feature data set generation module (3) is used to extract the time domain features and frequency domain features of the transient current traveling wave signal in the fault balance data set generated by the fault balance data set generation module (2), obtain the signal time domain features and frequency domain features, and generate a fault feature data set; The time domain features include waveform mean a1, root mean square a2, root square amplitude a3, absolute value mean a4, nonlinear characteristics a5, high-order nonlinear characteristics a6, variance a7, peak value a8, minimum value a9, peak-to-peak value a1, and the peak-to-peak value a2. 10 , the ratio of the root mean square and the absolute value a 11 , the ratio of the root mean square to the peak value a 12 , the ratio of the peak value to the absolute mean value a 13 , the ratio of peak to overall amplitude a 14 , the relationship between nonlinear characteristics and volatility 15 , the relationship between higher-order nonlinear characteristics and volatility 16 ; The expression of the time domain feature is: a8=max|x(i)| a9=min|x(i)| a 10 =a8-a9 Wherein: i = 1, 2, ..., N is the number of sample points, x(i) is the time domain signal sequence of the transient current traveling wave signal; a1 represents the mean of all elements of the time domain signal sequence of the transient current traveling wave signal; a2 represents the square root of the mean of all elements of the time domain signal sequence of the transient current traveling wave signal; a3 represents the square of the mean of the square roots of the absolute values ​​of all elements of the time domain signal sequence of the transient current traveling wave signal; a4 represents the mean of the absolute values ​​of all elements of the time domain signal sequence of the transient current traveling wave signal; a5 represents the sum of all elements of the time domain signal sequence of the transient current traveling wave signal minus a1. The cube of the transient current traveling wave signal is used to reflect the nonlinear characteristics of the signal; a6 represents the mean of all elements of the time domain signal sequence of the transient current traveling wave signal minus the fourth power of a1, which is used to detect the higher-order nonlinear characteristics of the signal; a7 represents the mean of all elements of the time domain signal sequence of the transient current traveling wave signal minus the square of a1, which represents the variance of the signal; a8 represents the maximum absolute value of all elements of the time domain signal sequence of the transient current traveling wave signal, which is used to detect the peak value of the signal; a9 represents the minimum absolute value of all elements of the time domain signal sequence of the transient current traveling wave signal, which is used to detect the minimum value of the signal; 10 Indicates the range between the peak value and the minimum value of the signal; a 11 Used to describe the peak nature of a signal; a 12 Used to compare the RMS and peak values ​​of a signal; a 13 Used to compare the peak value and the average of the absolute value of the signal; a 14 Used to compare the peak and overall amplitude of a signal; a 15 Used to describe the relationship between the nonlinear characteristics and volatility of the signal; a 16 Used to describe the relationship between higher-order nonlinear characteristics and volatility of signals; The frequency domain features include the mean b1 of the frequency domain signal, the variance b2 of the frequency domain signal, the skewness b3 of the frequency domain signal, the kurtosis b4 of the frequency domain signal, the power spectrum density b5 of the frequency domain signal, the root mean square bandwidth b6 of the frequency domain signal, the variance bandwidth b7 of the frequency domain signal, the kurtosis bandwidth b8 of the frequency domain signal, the pulse factor b9 of the frequency domain signal, and the relationship between the bandwidth and the root mean square b 10 , the peak factor b of the frequency domain signal 11 , the peak-to-peak factor b of the frequency domain signal 12 , the waveform factor b of the frequency domain signal 13 ; The specific expression is: Where: n=1,2,...,K, n is the spectral line number, K is the total number of spectral lines, s(n) is the frequency amplitude of the nth spectral line of the transient current traveling wave signal, f n is the frequency value of the nth spectrum line of the transient current traveling wave signal; b1 represents the mean value of the frequency domain signal; b2 represents the variance of the frequency domain signal; b3 represents the skewness of the frequency domain signal; b4 represents the kurtosis of the frequency domain signal; b5 represents the power spectrum density of the frequency domain signal; b6 represents the root mean square bandwidth of the frequency domain signal; b7 represents the variance bandwidth of the frequency domain signal; b8 represents the kurtosis bandwidth of the frequency domain signal; b9 represents the pulse factor of the frequency domain signal; b 10 Used to describe the relationship between the bandwidth and root mean square of a frequency domain signal; b 11 Indicates the peak factor of the frequency domain signal; b 12 Indicates the peak-to-peak factor of the frequency domain signal; b 13 Represents the shape factor of the frequency domain signal.

6. The improved oversampling-based transmission line fault classification system according to claim 1, characterized in that: The fault feature data set generation module (3) is used to divide the fault feature data set into a training set and a test set according to a set ratio.

7. The improved oversampling-based transmission line fault classification system according to claim 6, characterized in that: The fault classification model building module (4) is based on the CNN and SVM algorithms and is composed of an input layer, a convolution layer, a pooling layer, an attention layer, a fully connected layer, and an output layer. The input layer uses the training set divided by the fault feature data set generation module (3) as input to train the classification model; the convolution layer selects feature parameters; the pooling layer is used to reduce the dimension of the data after convolution of the convolution layer; the attention layer continuously trains the input data through the attention mechanism; the fully connected layer accesses the output of the attention layer and uses a fully connected method to connect each neuron with all neurons in the previous layer; the output layer inputs the output of the fully connected layer into the SVM classifier for classification.

8. A method for classifying power transmission line faults based on improved oversampling using the system of claim 1, characterized in that: The following steps are involved: Step 1: Acquire a transmission line fault transient current traveling wave signal, summarize the transmission line fault transient current traveling wave signal according to the fault type, and form a fault data set; Step 2: cluster and oversample the fault data set to form a fault-balanced data set; Step 3: extract the time domain and frequency domain features of the transient current traveling wave signal in the fault balance data set to form a fault feature data set; Step 4: According to the fault feature data set, the classification model established based on the CNN and SVM algorithms is trained to obtain a trained fault classification model. The transient current traveling wave signal of the transmission line fault to be detected is input into the fault classification model to obtain the fault type.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to claim 8 are implemented.

10. A computer program product comprising a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to claim 8 are implemented.

Images