A method and system for fault arc identification based on wavelet neural network

Abstract

The present application provides a method and system for fault arc identification based on wavelet neural network, which relates to the technical field of fault arc identification. The method comprises steps of signal acquisition, signal processing, first modeling, first training and fault identification; the system comprises a signal acquisition module, a signal processing module, a first modeling module, a first training module and a fault identification module. The present application trains the wavelet neural network model using historical current signals when a fault arc occurs and fault type labels, thereby improving the accuracy of identifying a fault arc.

Description

Technical Field

[0001] The present application relates to the technical field of fault arc identification, and in particular to a method and system for fault arc identification based on wavelet neural network. Background Art

[0002] In the safe operation of power systems and electrical equipment, arc faults are a serious safety hazard. Arc faults usually occur in situations such as poor electrical connections, insulation damage or equipment aging, which may cause fires, equipment damage or even personal injury. Therefore, accurate and rapid identification of arc faults is of great significance to ensure the safe and stable operation of power systems.

[0003] With the development of signal processing technology and machine learning technology, fault arc identification methods based on signal processing have gradually emerged. These methods identify fault arcs by automatically analyzing the time domain, frequency domain or time-frequency domain characteristics of electrical signals such as current and voltage. For example: the Chinese invention patent with an application publication date of August 6, 2024 and an application publication number of CN118445716A provides a detection method for automatically identifying series fault arcs, including the following steps: Step S1: Conduct a fault arc experiment and collect current data; Step S2: Build a model for detecting series fault arcs; Step S3: Set and adjust the parameters of the model established by step S2, train the model, and evaluate and compare its performance; Step S4: Realize the detection of series fault arcs, where the training set data mainly comes from the accumulation of historical data.

[0004] However, despite the great potential of these technologies, they still face many challenges in practical applications. In particular, since the number of arc fault samples in historical data is often very limited, it is difficult for the neural network model to obtain sufficient learning during the training process, which leads to low accuracy in identifying arc faults. Summary of the invention

[0005] In order to improve the accuracy of identifying arc faults, the present application provides an arc fault identification method and system based on wavelet neural network.

[0006] In the first aspect, the present application provides a method for fault arc identification based on wavelet neural network, which adopts the following technical solution:

[0007] A method for identifying arc faults based on wavelet neural network comprises the following steps:

[0008] Signal acquisition: including the first acquisition and the second acquisition;

[0009] First acquisition: collecting historical current signals when a fault arc occurs and corresponding fault type labels, and recording the historical current signals when a fault arc occurs as the first signal;

[0010] Second acquisition: collecting the current signal to be processed, recorded as the second signal;

[0011] Signal processing: including decomposition and signal reconstruction;

[0012] Decomposition: Decomposing the first signal using a wavelet packet transform algorithm to obtain wavelet packet coefficients of the target frequency band;

[0013] Signal reconstruction: reconstructing the signal using the wavelet packet coefficients to obtain the target signal, and using the target signal as the new first signal;

[0014] First modeling: establish a wavelet neural network model;

[0015] First training: using the first signal and the fault type label to train the wavelet neural network model to obtain a trained wavelet neural network model;

[0016] Fault identification: The second signal is input into the trained wavelet neural network model and the fault type is output.

[0017] By adopting the above technical solution, the present application can accumulate a large amount of fault sample data by collecting the historical current signal and the corresponding fault type label when the fault arc occurs, which improves the authenticity and pertinence of the fault sample data and helps to improve the accuracy of the wavelet neural network model in identifying the fault arc. The present application uses the wavelet packet transform algorithm to decompose the first signal, and can accurately extract information in different frequency bands. Compared with the traditional wavelet transform, the wavelet packet transform has higher frequency resolution and better time-frequency localization characteristics, so it can more effectively capture the characteristic signal generated by the fault arc. The present application uses the wavelet packet coefficient to reconstruct the signal, and can obtain the target signal, that is, the new historical current signal, which helps to remove noise and redundant information and improve the quality of the signal. The present application uses the first signal (the historical current signal after signal processing) and the fault type label to train the wavelet neural network model, so that the wavelet neural network model can learn the characteristics and classification rules of the fault arc, so as to have accurate fault identification ability.

[0018] Optionally, in the first acquisition step, a normal historical current signal is also acquired and recorded as a third signal, and after the signal processing step is performed and before the first modeling step is performed, the method further includes:

[0019] First extraction: extract features from the second signal to obtain n features;

[0020] Second modeling: building a CNN model;

[0021] Second training: the CNN model is trained using the first signal and the third signal. During the training process, the label of the third signal is set to 0, and the label of the first signal is set to 1.

[0022] First classification: input n features into the trained CNN model in sequence to obtain n output data;

[0023] Statistics: Count the number of 1s in n output data and record it as the first data;

[0024] First judgment: judge whether the first data is greater than a preset threshold, if so, execute the first modeling step; if not, execute the second acquisition step.

[0025] By adopting the above technical solution, the present application collects normal historical current signals, and uses normal historical current signals and historical current signals when a fault arc occurs to train the CNN model, so that the CNN model can learn the characteristic differences between the normal and fault states. After that, the second signal is feature extracted to obtain n features that can characterize the characteristics of the second signal. The extracted features are sequentially input into the trained CNN model to obtain n output data, and the number of labels 1 (i.e., fault features) is counted. If the number of 1s exceeds the preset threshold, the second signal belongs to the historical current signal when a fault arc occurs. Otherwise, it does not belong to the historical current signal when a fault arc occurs. The second acquisition step is performed to continue monitoring the current signal. The present application makes a preliminary judgment on the second signal through the trained CNN model to determine whether the second signal is normal current data.

[0026] Optionally, after performing the first training step and before performing the fault identification step, the method further includes:

[0027] Encoding: Encode n features separately to obtain n feature vectors;

[0028] First calculation: calculate the correlation between all feature pairs based on n feature vectors, select the maximum value among all correlations and record it as the second data;

[0029] Second judgment: judging whether the second data is less than a preset threshold value, if so, executing the deletion step; if not, executing the fault identification step;

[0030] Subtract: Delete one feature in the current feature pair and execute the first calculation step.

[0031] By adopting the above technical solution, the present application encodes n features and converts them into feature vectors, which helps to better represent and distinguish different features in subsequent processing. When the second signal belongs to a certain fault type label, the features it contains should be strongly correlated. The present application calculates the correlation between all feature pairs based on n feature vectors, and selects the value of the maximum correlation as the second data. If the second data is less than the preset threshold, it means that the correlation between these features is very low, and there are some redundant features. It is necessary to delete the redundant features in the feature pairs corresponding to the second data, and recalculate the correlation between the remaining features.

[0032] Optionally, after executing the second judgment step and before executing the deletion step, the method further includes:

[0033] Third modeling: establishing a PNN model, wherein the PNN model includes m neurons;

[0034] Third training: using the first signal and the third signal to train the PNN model;

[0035] Calculate similarity: Calculate the similarity between the i-th feature vector and each neuron in the trained PNN model to obtain m similarity results;

[0036] Calculate output: Calculate the output probability of the trained PNN model based on m similarity results;

[0037] Probability judgment: judge whether the output probability is greater than a preset probability threshold, if so, execute the fault identification step; if not, execute the deletion step.

[0038] By adopting the above technical solution, the present application establishes a PNN model, which is a classifier based on a radial basis function network. It performs classification by learning the probability distribution of input data, and then uses the first signal and the third signal to train the PNN model. Through training, the PNN model can learn the characteristic differences between the normal and faulty states, and can classify according to these characteristics in subsequent processing. The present application obtains m similarity results by calculating the similarity between the i-th feature vector and each neuron in the trained PNN model, which is actually evaluating the similarity between the current feature vector and the type represented by each neuron in the PNN model. Based on these similarity results, the output probability of the PNN model can be calculated, that is, the probability that the current feature vector belongs to a certain type. The present application can determine whether to perform the fault identification step or the deletion step by judging whether the output probability is greater than the preset threshold. If the output probability is greater than the preset threshold, it means that the current feature vector has a high similarity with the fault type, so it can be considered that the current signal contains fault features, and the fault identification step should be performed at this time. If the output probability is not greater than the preset threshold, it may indicate that the current feature vector has a low similarity with the fault type, or the signal quality is poor. In this case, a deletion step can be considered to reduce redundant features.

[0039] Optionally, after executing the step of signal acquisition and before executing the step of signal processing, the method further includes:

[0040] Second extraction: using a global search algorithm to traverse the signal set consisting of historical current signals to obtain the fault arc characteristics;

[0041] Second classification: Use a classifier to classify each fault arc feature, obtain the type label of each fault arc feature, and integrate the fault arc features with the same type label into a feature set;

[0042] Assign weights: Use the random forest algorithm to assign weights to each fault arc feature in the feature set;

[0043] Feature fusion: performing feature weighted fusion based on the fault arc feature and the weight, obtaining the weighted fused fault arc feature, which is recorded as the first feature;

[0044] Second calculation: calculating the correlation between the first feature and the second signal, recorded as third data;

[0045] Data judgment: judge whether the third data is greater than a preset threshold. If so, use the first feature to replace the first signal to perform the signal processing step; if not, execute the second acquisition step.

[0046] By adopting the above technical solution, the present application uses a global search algorithm to extract fault arc features from all historical current signals, and can more accurately identify features related to fault arcs, namely, fault arc features. Afterwards, the present application uses a classifier to classify the fault arc features, and integrates features of the same type into a feature set. The present application visualizes the fault type label of the historical current signal onto the fault arc feature in the historical current signal, which can improve the accuracy of classification and enhance the generalization ability of the wavelet neural network model. The present application uses a random forest algorithm to assign weights to each fault arc feature in each feature set, taking into account the importance of different fault arc features in fault arc detection, which helps to improve the accuracy of fault identification. The fault arc features and weights are then weighted fused to obtain a first feature, which can adjust the proportion of the fault arc features in the first feature. The correlation between the first feature and the second signal is then calculated. If the correlation is greater than a preset threshold, it indicates that the first feature has a high correlation with the current signal to be processed, that is, the second signal is the current signal when a fault arc occurs, and the first feature is used to perform the signal processing step. Otherwise, it indicates that the current signal to be processed is a normal signal, and the second acquisition step is performed to continuously monitor the current signal.

[0047] Optionally, after performing the second classification step and before performing the weight assignment step, the method further includes:

[0048] Table building: building a management table, wherein the management table stores the fault arc features in the feature set;

[0049] Generate neighborhood features: Generate neighborhood features based on fault arc features using a generative adversarial network;

[0050] Evaluation: The evaluation result of the neighborhood feature is calculated using an evaluation function, and the calculation model of the evaluation function is as follows:

[0051] ;

[0052] in, is the evaluation function; N is the number of neighborhood features; is the jth eigenvalue of the i-th neighborhood feature; p is the number of eigenvalues ​​of the i-th neighborhood feature;

[0053] Third judgment: judging whether the evaluation result is greater than a preset threshold, if so, executing the step of adding neighborhood features; if not, executing the step of iteration;

[0054] Add neighborhood features: add neighborhood features to the management table;

[0055] Iteration: Integrate the neighborhood features whose evaluation results meet the preset threshold into a new feature set, and execute the steps of generating neighborhood features until the preset stop condition is met.

[0056] By adopting the above technical solution, neighborhood features are generated based on the fault arc features using a generative adversarial network. This step increases the diversity of features and helps to discover new feature combinations that are helpful for classification. This application uses an evaluation function to calculate the evaluation results of neighborhood features to evaluate the quality of each neighborhood feature, thereby reflecting the effectiveness of the neighborhood feature for the classification task. After that, it is determined whether to perform the step of adding neighborhood features or the step of performing iterations based on whether the evaluation result is greater than a preset threshold. If the evaluation result is good, a new neighborhood feature is added, and the neighborhood features that meet the conditions are added to the management table. This step increases the richness of the feature set; if the evaluation result is not good, the step of iteration is performed.

[0057] Optionally, after performing the step of allocating weights and before performing the step of fusing features, the method further includes:

[0058] Feature screening: including sorting, feature selection and integration;

[0059] Sorting: sort the fault arc features in each feature set in descending order of weight to obtain the fault arc feature sequence;

[0060] Feature selection: In the fault arc feature sequence, select the first k fault arc features;

[0061] Integration: The selected fault arc features are reintegrated into a new feature set and the weight assignment step is performed.

[0062] By adopting the above technical solution, the fault arc features in each feature set are sorted in descending order according to the weight, which helps to clearly understand the importance of each fault arc feature. Through sorting, it is possible to intuitively see which fault arc features contribute the most to the classification task and which fault arc features may be relatively redundant or irrelevant. In the fault arc feature sequence, the first k fault arc features are selected, and relatively important fault arc features can be selected from the original feature set based on the sorting results. By selecting the first k features, the dimension of the feature set can be reduced, the computational complexity can be reduced, and the removal of redundant or irrelevant fault arc features can reduce the risk of overfitting. After that, the selected fault arc features are reintegrated into a new feature set, and the step of assigning weights is performed, so that the screened fault arc features can be recombined into a new and more refined feature set.

[0063] Optionally, after performing the decomposition step and before performing the signal reconstruction step, the method further includes:

[0064] Constructing a matrix: integrating the wavelet packet coefficients into a two-dimensional matrix, normalizing the two-dimensional matrix, and obtaining a normalized two-dimensional matrix, which is recorded as the first matrix;

[0065] Generating an image: converting each element in the first matrix into a grayscale value, and generating a grayscale image based on the grayscale value;

[0066] Integrating feature matrix: extracting SIFI feature descriptors of the grayscale image, and integrating the SIFI feature descriptors into a feature matrix;

[0067] Visualization processing: using the t-SNE algorithm to visualize the feature matrix to obtain feature distribution results;

[0068] Fourth judgment: judge whether the characteristic distribution result meets expectations. If so, execute the signal reconstruction step; if not, execute the decomposition step.

[0069] By adopting the above technical solution, the wavelet packet coefficients are integrated into a two-dimensional matrix and normalized, which helps to eliminate the dimensional differences between the data, convert the elements in the first matrix into gray values, and generate a grayscale image. This process realizes the conversion from numerical data to image data. The grayscale image can intuitively display the distribution and characteristics of the wavelet packet coefficients, providing a visualization basis for subsequent feature extraction. The SIFI feature descriptors of the grayscale image are extracted, and these feature descriptors are integrated into a feature matrix. The feature matrix is ​​visualized using the t-SNE (t-distributed Stochastic Neighbor Embedding) algorithm to obtain the feature distribution results. This step helps to intuitively understand the information distribution in the feature matrix and provides a basis for judging whether the feature distribution meets expectations. The present application also makes a judgment based on the feature distribution results. If it meets expectations, the signal reconstruction step is executed; if it does not meet expectations, the decomposition step is re-executed, that is, the signal reconstruction is performed only when the feature distribution meets expectations, thereby improving the reliability and accuracy of the entire method.

[0070] Optionally, after executing the fourth judgment step and before executing the decomposition step, the method further includes:

[0071] Check: Check whether the first signal has classification errors. If so, reassign the fault type label to the first signal; if not, increase the number of decomposition layers.

[0072] By adopting the above technical solution, the classification accuracy of the signal can be improved by checking whether the first signal has classification errors. If the first signal has no classification errors, but the classification result still does not meet expectations, it means that the current number of signal decomposition layers is insufficient to reveal subtle features or patterns in the signal. In this case, increasing the number of decomposition layers can provide more information and help analyze the signal more accurately.

[0073] In the second aspect, the present application provides a fault arc identification system based on wavelet neural network, which adopts the following technical solution:

[0074] A fault arc identification system based on wavelet neural network, comprising:

[0075] A signal acquisition module, comprising a first acquisition unit and a second acquisition unit;

[0076] A first acquisition unit, used for acquiring a historical current signal when a fault arc occurs and a corresponding fault type label, and recording the historical current signal when a fault arc occurs as a first signal;

[0077] A second acquisition unit, used for acquiring a current signal to be processed, recorded as a second signal;

[0078] A signal processing module, which is in communication connection with the first acquisition unit, comprises a decomposition unit and a signal reconstruction unit;

[0079] A decomposition unit, used to decompose the first signal using a wavelet packet transform algorithm to obtain wavelet packet coefficients of a target frequency band;

[0080] A signal reconstruction unit, used to reconstruct the signal using the wavelet packet coefficients to obtain a target signal, and use the target signal as a new first signal;

[0081] The first modeling module is used to establish a wavelet neural network model;

[0082] A first training module is connected to the signal processing module and the first modeling module for training the wavelet neural network model using the first signal and the fault type label to obtain a trained wavelet neural network model;

[0083] The fault identification module is connected to the second acquisition unit and the first training module for inputting the second signal into the trained wavelet neural network model and outputting the fault type.

[0084] By adopting the above technical scheme, the first acquisition unit of the present application is responsible for collecting the historical current signal and its corresponding fault type label when a fault arc occurs, the second acquisition unit is responsible for collecting the current signal to be processed, the decomposition unit uses the wavelet packet transform algorithm to decompose the historical current signal when a fault arc occurs, and extracts information of different frequency bands, the signal reconstruction unit uses the wavelet packet coefficients obtained by decomposition to reconstruct the signal to obtain the target signal, the first training module: uses the first signal (the historical current signal after signal processing) and its corresponding fault type label to train the wavelet neural network model, the fault identification module inputs the second signal (the current signal to be processed) into the trained wavelet neural network model, and outputs the predicted fault type according to the learned knowledge, the entire system can realize accurate and real-time identification of fault arcs through effective signal processing and intelligent identification technology, and provides strong support for the safe and stable operation of the power system.

[0085] In summary, the present application includes at least one of the following beneficial technical effects:

[0086] 1. This application can accumulate a large amount of fault sample data by collecting historical current signals and their corresponding fault type labels when a fault arc occurs, thereby improving the authenticity and pertinence of the fault sample data and helping to improve the accuracy of the wavelet neural network model in identifying the fault arc. This application uses a wavelet packet transform algorithm to decompose the first signal, and can accurately extract information from different frequency bands. Compared with traditional wavelet transforms, wavelet packet transforms have higher frequency resolution and better time-frequency localization characteristics, so they can more effectively capture the characteristic signals generated by the fault arc. This application uses wavelet packet coefficients to reconstruct the signal, and can obtain the target signal, that is, the new historical current signal, which helps to remove noise and redundant information and improve the quality of the signal. This application uses the first signal (historical current signal after signal processing) and the fault type label to train the wavelet neural network model, so that the wavelet neural network model can learn the characteristics and classification rules of the fault arc, thereby having accurate fault identification capabilities.

[0087] 2. The present application collects normal historical current signals, and uses normal historical current signals and historical current signals when a fault arc occurs to train the CNN model, so that the CNN model can learn the characteristic differences between normal and faulty states. After that, feature extraction is performed on the second signal to obtain n features that can characterize the characteristics of the second signal. The extracted features are sequentially input into the trained CNN model to obtain n output data, and the number of labels with 1 (i.e., fault features) is counted. If the number of 1s exceeds the preset threshold, the second signal belongs to the historical current signal when a fault arc occurs. Otherwise, it does not belong to the historical current signal when a fault arc occurs. The second acquisition step is performed to continue monitoring the current signal. The present application makes a preliminary judgment on the second signal through the trained CNN model to determine whether the second signal is normal current data.

[0088] 3. The present application uses a global search algorithm to extract fault arc features from all historical current signals, which can more accurately identify features related to fault arcs, namely, fault arc features. Afterwards, the present application uses a classifier to classify the fault arc features, and integrates features of the same type into a feature set. The present application concretizes the fault type label of the historical current signal to the fault arc feature in the historical current signal, which can improve the accuracy of classification and enhance the generalization ability of the wavelet neural network model. The present application uses a random forest algorithm to assign weights to each fault arc feature in each feature set, taking into account the importance of different fault arc features in fault arc detection, which helps to improve the accuracy of fault identification. Afterwards, the fault arc features and weights are weightedly fused to obtain the first feature, which can adjust the proportion of the fault arc feature in the first feature, and then calculate the correlation between the first feature and the second signal. If the correlation is greater than the preset threshold, it indicates that the first feature has a high correlation with the current signal to be processed, that is, the second signal is the current signal when the fault arc occurs, and the first feature is used to perform the signal processing step. Otherwise, it indicates that the current signal to be processed is a normal signal, and the second acquisition step is performed to continuously monitor the current signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 is a flow chart of Example 1 of the present application;

[0090] Figure 2 This is a flowchart from S23 first extraction to S28 first judgment in Example 2 of the present application;

[0091] Figure 3 It is a flowchart with S61 coding to S69 deleted in Example 2 of the present application;

[0092] Figure 4 is a flow chart of Example 3 of the present application;

[0093] Figure 5 This is a flow chart of Example 4 of the present application. DETAILED DESCRIPTION

[0094] The following combination Figures 1 to 5 This application is described in further detail.

[0095] Embodiment 1: This embodiment discloses a method for identifying arc faults based on wavelet neural network. Figure 1 The method includes: S1 signal acquisition, S2 signal processing, S3 first modeling, S4 first training and S5 fault identification. First, historical current signals and corresponding fault type labels are collected, and then the current signal to be processed is collected. Then, the historical current signal is decomposed to obtain the wavelet packet coefficient of the target frequency band, and then the signal is reconstructed based on the wavelet packet coefficient to obtain the target signal. Then, the wavelet neural network model is trained using the target signal and the fault type label, and then the current signal to be processed is input into the trained wavelet neural network model to output the fault type. The process of this embodiment is as follows:

[0096] S1 signal acquisition, including S11 first acquisition and S12 second acquisition.

[0097] S11 is the first acquisition, which collects historical current signals and corresponding fault type labels when a fault arc occurs, and records the historical current signals when a fault arc occurs as the first signal. The first signal is a current signal detected in a real environment in the past, and the fault type label is a fault type marked based on the characteristics of the first signal.

[0098] S12 is a second acquisition, in which the current signal to be processed is acquired and recorded as a second signal. The second signal is a current signal for which the fault type currently needs to be identified.

[0099] S2 signal processing, including S21 decomposition and S22 signal reconstruction.

[0100] S21 decomposition: the wavelet packet transform algorithm selects an appropriate number of decomposition layers and frequency band division method according to the characteristics and requirements of the first signal, and then decomposes the first signal layer by layer until a predetermined number of decomposition layers is reached or a specific frequency band division requirement is met.

[0101] In each layer of decomposition, the wavelet packet transform algorithm calculates the wavelet packet coefficients on each frequency band of the layer. These coefficients reflect the characteristics of the signal on the frequency band, and then extracts the wavelet packet coefficients of the target frequency band.

[0102] S22 signal reconstruction uses the wavelet packet coefficients of the target frequency band in the S21 decomposition to perform inverse wavelet packet transform to reconstruct the signal, obtain the target signal, and use the target signal as the new first signal.

[0103] The fault type label of the original first signal is used as the fault type label of the new first signal. Although the reconstructed target signal has been processed, its fault characteristics are consistent with the original first signal, so the fault type label of the original first signal can be retained.

[0104] S3 first builds a wavelet neural network model.

[0105] The wavelet neural network model includes an input layer, a hidden layer, and an output layer. The neurons in the hidden layer use nonlinear wavelet basis functions (such as Morlet wavelet, etc.) instead of traditional Sigmoid functions, and the weights from the input layer to the hidden layer and the threshold of the hidden layer are replaced by the scale expansion factor and time translation factor of the wavelet function respectively.

[0106] S4 is a first training, in which the wavelet neural network model is trained using the first signal and the fault type label to obtain a trained wavelet neural network model.

[0107] S5 fault identification, inputs the second signal into the trained wavelet neural network model, and the trained wavelet neural network model predicts and analyzes the second signal based on the learned knowledge and experience, and outputs the fault type.

[0108] This embodiment uses the historical current signal when a fault arc occurs and the corresponding fault type label to train the wavelet neural network model, so that the wavelet neural network model can focus on learning the relationship between the historical current signal when a fault arc occurs and the corresponding fault type label. The trained wavelet neural network model can accurately and quickly give the fault type of the current signal to be processed.

[0109] Example 2: Reference Figure 2 The difference between this embodiment and the first embodiment is that in the first acquisition in S11, a normal historical current signal is also acquired and recorded as a third signal, and after executing the signal processing in S2 and before executing the first modeling in S3, it also includes:

[0110] S23 first extraction, extracting features of the second signal in the time domain and the frequency domain to obtain n features. Extracting features such as statistics, peak values, and mean values ​​of the second signal in the time domain, and extracting spectrum features such as amplitude and phase of the second signal in the frequency domain.

[0111] S24 is the second modeling, establishing a CNN model. The activation function of the output layer of the CNN model adopts the Relu function. The Relu function solves the problem of gradient vanishing in the positive interval and its calculation and convergence speed are much faster than those of the Sigmoid function. When the input value is non-negative, the gradient of the Relu function is 1, which can effectively solve the problem of gradient vanishing.

[0112] S25 is the second training, using the first signal and the third signal to train the CNN model. During the training process, the label of the normal current signal (ie, the third signal) is set to 0, and the label of the first signal is set to 1.

[0113] S26 is the first classification, and each feature extracted in S23 is regarded as an independent input sample. The n features are input into the trained CNN model in turn to obtain n output data, which represent the classification results of each feature by the CNN model.

[0114] S27 counts, traverses n output data, counts the number of labels with 1, and records them as the first data.

[0115] S28 is the first judgment, judging whether the first data is greater than a preset threshold. If so, it is considered that the possibility of a fault signal is high, and then S3 is executed for the first modeling; if not, it is considered that the possibility of a fault signal is low, and S12 is executed for the second acquisition to continue monitoring the current signal.

[0116] Then, the same as in Example 1, continue to perform S3 first modeling and S4 first training, refer to Figure 3 After performing the first training in S4, this embodiment further includes:

[0117] S61 encoding encodes n features respectively to obtain n feature vectors.

[0118] S62 is the first calculation, which calculates the correlation between each pair of eigenvectors by using methods such as the Pearson correlation coefficient and the Spearman rank correlation coefficient, and selects the maximum value of all correlations, which is recorded as the second data.

[0119] S63 is a second judgment, determining whether the second data is less than a preset threshold value. If so, executing S64 is a third modeling; if not, executing S5 is a fault identification.

[0120] S64: third modeling, establishing a PNN model, wherein the PNN model includes m neurons.

[0121] S65: third training, using the first signal and the third signal to train the PNN model.

[0122] S66 calculates the similarity. For each feature vector, the similarity between it and each neuron in the PNN model (such as cosine similarity, Euclidean distance, etc.) is calculated to obtain m similarity results.

[0123] S67 calculates the output, and calculates the output probability of the trained PNN model based on the m similarity results, that is, these similarity values ​​are used as inputs to the radial basis function to generate a probability estimate for each fault type.

[0124] The output probability of the PNN model refers to the probability that the input data calculated by the PNN model belongs to each fault type based on the above similarity results. These probability values ​​are calculated by passing the similarity values ​​to an output layer (the softmax layer is used in this embodiment), which converts the similarity values ​​into probability values, the sum of which is 1, and each probability value represents the possibility that the input data belongs to the corresponding fault type.

[0125] S68 probability judgment, judging whether the output probability is greater than a preset probability threshold, if so, executing S5 fault identification; if not, executing S69 deletion.

[0126] S69 is deletion, deleting one feature in the current feature pair, and performing S62 the first calculation.

[0127] Assume that there are five features A1, A2, A3, A4, and A5. Encode the five features respectively and obtain the feature vectors A1=[0.1, 0.3, 0.4], A2=[0.2, 0.3, 0.5], A3=[0.2, 0.6, 0.4], A4=[0.3, 0.7, 0.1], and A5=[0.8, 0.2, 0.1].

[0128] Taking the Pearson correlation coefficient as an example, the correlation between feature vectors A1 and A2 is calculated:

[0129] ;

[0130] ;

[0131] ;

[0132] ;

[0133] ;

[0134] ;

[0135] ;

[0136] ;

[0137] .

[0138] That is, the correlation between eigenvectors A1 and A2 is about 0.6.

[0139] Similarly, the correlation between eigenvectors A1 and A3 is about 0.2. The correlation between eigenvectors A1 and A4 is about 0.048. The correlation between eigenvectors A1 and A5 is about -0.985. The correlation between eigenvectors A2 and A3 is about 0.98. The correlation between eigenvectors A2 and A4 is about -0.243. The correlation between eigenvectors A2 and A5 is about -0.985. The correlation between eigenvectors A3 and A4 is about 0.949. The correlation between eigenvectors A3 and A5 is about 0.143. The correlation between eigenvectors A4 and A5 is about 0.857.

[0140] The second data is: 0.985. The negative sign indicates that the eigenvectors A2 and A5 are negatively correlated, but the correlation between the two is the largest.

[0141] Assuming that the preset threshold is 0.99 and the second data is less than the preset threshold, the third modeling is executed in S64.

[0142] In other embodiments, if the preset threshold is 0.98, the second data is greater than the preset threshold, and S5 fault identification is performed.

[0143] Then, the third modeling in S64 and the third training in S65 are performed.

[0144] Then, for each feature vector, the similarity between it and each neuron in the PNN model (such as cosine similarity, Euclidean distance, etc.) is calculated to obtain m similarity results.

[0145] It is assumed that the PNN model includes 3 neurons, and the representative vectors of the three neurons are as follows:

[0146] Neuron 1: B1 = [0.5, 0.5, 0.5];

[0147] Neuron 2: B2 = [0.1, 0.9, 0.2];

[0148] Neuron 3: B3 = [0.8, 0.1, 0.6].

[0149] Taking the calculation model of cosine similarity as an example, the similarity between feature vector A1 and neuron 1 in the PNN model is calculated as follows:

[0150] ;

[0151] Similarly, the similarity between eigenvector A1 and neuron 2 is about 0.35; the similarity between eigenvector A1 and neuron 3 is about 0.37; the similarity between eigenvector A2 and neuron 1 is about 0.69; the similarity between eigenvector A2 and neuron 2 is about 0.45; the similarity between eigenvector A2 and neuron 3 is about 0.47; the similarity between eigenvector A3 and neuron 1 is 0.943; the similarity between eigenvector A3 and neuron 2 is about 0.70; the similarity between eigenvector A3 and neuron 3 is about 0.61; the similarity between eigenvector A4 and neuron 1 is 0.41; the similarity between eigenvector A4 and neuron 2 is about 0.95; the similarity between eigenvector A4 and neuron 3 is about 0.59; the similarity between eigenvector A5 and neuron 1 is 0.77; the similarity between eigenvector A5 and neuron 2 is about 0.52; the similarity between eigenvector A5 and neuron 3 is about 0.97.

[0152] The output probability of the trained PNN model is calculated based on m similarity results:

[0153] ;

[0154] ;

[0155] in, is the output probability of the i-th fault type; is an intermediate variable; is the similarity result between the ith feature vector and the jth neuron; m is the number of neurons in the PNN model (m also represents the number of similarity results).

[0156] Then the probability that the fault type label of the second signal is the same as that of the feature vector A1, that is, the output probability of the PNN model when the fault type label of the second signal is the same as that of the feature vector A1 is:

[0157] ;

[0158] ;

[0159] About 0.54; About 0.751; is 0.65; About 0.75.

[0160] Similarly, the probability that the second signal has the same fault type label as feature vector A2 is approximately 0.1671; the probability that the second signal has the same fault type label as feature vector A3 is approximately 0.2324; the probability that the second signal has the same fault type label as feature vector A4 is 0.2011; and the probability that the second signal has the same fault type label as feature vector A5 is approximately 0.2321.

[0161] Assume that the similarity between the first eigenvector, that is, eigenvector A1 and each neuron in the PNN model is calculated, and the output probability is about 0.1671. In this embodiment, the preset probability threshold is 0.2, then 0.1671<0.2, and S69 deletion is executed.

[0162] In this embodiment, when the correlation between the feature vector pairs is low, the PNN model is used to further analyze the similarity between the feature vector and the fault type, and the similarity is further judged. This embodiment further refines the judgment criteria for the similarity, effectively reducing the risk of erroneously deleting the feature vector of the current signal to be processed due to misjudgment, thereby significantly improving the accuracy of fault identification.

[0163] Example 3: Reference Figure 4 The difference between this embodiment and embodiment 1 is that after performing S1 signal acquisition and before performing S2 signal processing, it also includes:

[0164] S7 signal preprocessing, including S71 second extraction, S72 second classification, S721 table building, S722 generating neighborhood features, S723 evaluation, S724 third judgment, S725 adding neighborhood features, S726 iteration, S73 weight allocation, S74 feature screening, S75 feature fusion, S76 second calculation and S77 data judgment.

[0165] S71 is the second extraction, which uses a global search algorithm to traverse the signal set composed of historical current signals to obtain fault arc characteristics. These fault arc characteristics include key parameters such as the amplitude, frequency, and phase of the current signal.

[0166] S72 second classification, using a classifier to classify each fault arc feature, obtain the category label of each fault arc feature, and integrate the fault arc features with the same category label into a feature set.

[0167] S721 creates a table, constructing a management table, in which the fault arc features in the feature set are stored.

[0168] S722 generates neighborhood features, and generates neighborhood features based on the fault arc features using a generative adversarial network.

[0169] S723 evaluation, using an evaluation function to calculate the evaluation result of the neighborhood feature, the calculation model of the evaluation function is as follows:

[0170] ;

[0171] in, is the evaluation function; N is the number of neighborhood features; is the jth eigenvalue of the i-th neighborhood feature; p is the number of eigenvalues ​​of the i-th neighborhood feature;

[0172] S724 is the third judgment, judging whether the evaluation result is greater than a preset threshold. If so, it means that the performance of the current neighborhood feature is good, and it can be considered to be added to the management table, and then S725 is executed to add the neighborhood feature; if not, it means that the performance of the current neighborhood feature is not good, and S726 is executed for iteration.

[0173] S725 adds a neighborhood feature, and adds the neighborhood feature to the management table.

[0174] S726 is an iteration, in which the neighborhood features whose evaluation results meet the preset threshold are integrated into a new feature set, and S722 is executed to generate neighborhood features until a preset stop condition is met (for example, a preset number of iterations is reached).

[0175] S73 assigns weights. After a new feature set is formed, in order to distinguish the importance of different features in fault identification, a random forest algorithm is used to assign weights to each fault arc feature in each feature set.

[0176] S74 Feature screening, including S741 sorting, S742 feature selection and S743 integration.

[0177] S741 sorting: sorting the fault arc features in each feature set in descending order of weight to obtain a fault arc feature sequence.

[0178] S742 feature selection, select the first k fault arc features in the fault arc feature sequence.

[0179] S743 integration, reintegrates the selected fault arc features into a new feature set, and executes S73 to allocate weights until a preset stop condition is met (such as reaching a preset number of cycles or the number of fault arc features in the fault arc feature sequence reaches a minimum requirement), and executes S75 feature fusion.

[0180] S75 feature fusion, performing feature weighted fusion based on the fault arc feature and the weight, multiplying each fault arc feature by its corresponding weight, and adding the results to obtain the weighted fused fault arc feature, which is recorded as the first feature.

[0181] S76 second calculation, calculating the correlation between the first feature and the second signal, recorded as third data. In this embodiment, the correlation between the first feature and the second signal is the same as that in S69 deletion, and the process is: respectively encoding the first feature and the second signal, and calculating the correlation coefficient based on the Pearson correlation coefficient.

[0182] S77 data judgment, judge whether the third data is greater than the preset threshold value. If so, it means that there is a high correlation between the first feature and the second signal. Therefore, the first feature can be used to replace the original historical current signal for subsequent signal processing, that is, the first feature is used to replace the original historical current signal to perform S2 signal processing; if not, it means that the correlation between the first feature and the second signal is low or there is no obvious correlation, that is, the second signal is not the current signal when a fault arc occurs, and then continue to execute S12 second acquisition and monitoring of the current signal.

[0183] This embodiment extracts the fault arc features in the historical current signal, and performs weighted fusion processing on the fault arc features to obtain the first feature, and then calculates the correlation between the first feature and the second signal to determine whether the correlation between the two is greater than a preset threshold. If it is greater, it indicates that the second signal is a current signal generated when the fault arc occurs, and the signal processing can be performed. If it is not greater, it indicates that the second signal is a normal current signal, and then the current signal can be continued to be monitored. This embodiment also expands the feature set after the second classification step, and after assigning weights, the feature set is reduced based on the importance of the fault arc features, thereby optimizing the fault arc features.

[0184] Example 4: Reference Figure 5 , after performing S21 decomposition and before performing S22 signal reconstruction, it also includes:

[0185] S81 constructs a matrix and integrates the wavelet packet coefficients into a two-dimensional matrix in time sequence or frequency sequence. Each row or column of this matrix represents the characteristics of the signal at different frequencies or time periods.

[0186] Then, the two-dimensional matrix is ​​normalized (such as minimum-maximum normalization, Z-score normalization, etc.) to obtain a normalized two-dimensional matrix, which is recorded as the first matrix.

[0187] Assume that the decomposed signal contains 4 frequency bands (to simplify the explanation, it may actually contain more), and data is collected at 8 time points in each frequency band. According to the time sequence of the collection, a 4x8 two-dimensional matrix is ​​constructed as follows:

[0188] ;

[0189] The two-dimensional matrix C is normalized by the minimum-maximum normalization algorithm to obtain the first matrix C'. The first matrix C' is as follows:

[0190] .

[0191] S82 generates an image, converts each element in the first matrix into a grayscale value, and generates a grayscale image based on the grayscale value. By converting the elements in the first matrix into grayscale values, the characteristic information in the matrix can be intuitively displayed in the form of an image. This graphical representation method helps to better understand the distribution and characteristics of the data.

[0192] Converting each element in the first matrix C' to a grayscale value results in the following:

[0193] .

[0194] S83 integrates feature matrices, extracts SIFI feature descriptors of the grayscale image, and integrates the SIFI feature descriptors into a feature matrix.

[0195] S84 visualization processing, using the t-SNE algorithm to visualize the feature matrix. Through the t-SNE algorithm, the data points in the feature matrix can be mapped to a two-dimensional plane to obtain feature distribution results.

[0196] S85 is the fourth judgment, judging whether the characteristic distribution result meets expectations, if so, executing S22 signal reconstruction; if not, executing S86 inspection.

[0197] The standard for the feature distribution results to meet expectations is: the clustering effect is obvious or the data points are concentrated.

[0198] S86 checks whether there is a classification error in the first signal. If so, reassigns a fault category label to the first signal; if not, increases the number of decomposition layers.

[0199] This embodiment can perform feature extraction, dimensionality reduction processing and visualization analysis on the signal, providing strong support for subsequent fault diagnosis and signal reconstruction.

[0200] Embodiment 5: This embodiment discloses a fault arc identification system based on wavelet neural network, the system comprising:

[0201] The signal acquisition module includes a first acquisition unit and a second acquisition unit.

[0202] The first acquisition unit is responsible for acquiring the historical current signal when the fault arc occurs, and simultaneously recording the corresponding fault category label, and recording the historical current signal when the fault arc occurs as the first signal.

[0203] The second acquisition unit is responsible for acquiring the current signal to be processed and recording the current signal to be processed as the second signal.

[0204] The signal processing module is connected to the first acquisition unit in communication, and includes a decomposition unit and a signal reconstruction unit.

[0205] The decomposition unit is responsible for decomposing the first signal by using a wavelet packet transform algorithm to obtain wavelet packet coefficients of a target frequency band.

[0206] The signal reconstruction unit is responsible for reconstructing the signal using the wavelet packet coefficients to obtain the target signal and use the target signal as a new historical current signal.

[0207] The first modeling module is responsible for establishing the wavelet neural network model.

[0208] The first training module is connected to the signal processing module and the first modeling module for communication, and is responsible for training the wavelet neural network model using the first signal and the fault category label to obtain the trained wavelet neural network model.

[0209] The fault identification module is connected to the second acquisition unit and the first training module for inputting the second signal into the wavelet neural network model and outputting the fault category.

[0210] The signal acquisition module, signal processing module, first modeling module, first training module and fault identification module in this embodiment together constitute a complete arc fault identification system, which can automatically collect current signals, extract feature information, train classification models and perform fault identification, providing strong guarantee for the safe operation of the power system.

[0211] The above are all preferred embodiments of the present application, and the protection scope of the present application is not limited thereto. Therefore, any equivalent changes made according to the structure, shape, and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for identifying arc faults based on wavelet neural network, characterized in that: include: Signal acquisition: including the first acquisition and the second acquisition; First acquisition: collecting historical current signals when a fault arc occurs and corresponding fault type labels, and recording the historical current signals when a fault arc occurs as the first signal; Second acquisition: collecting the current signal to be processed, recorded as the second signal; Signal processing: including decomposition and signal reconstruction; Decomposition: Decomposing the first signal using a wavelet packet transform algorithm to obtain wavelet packet coefficients of the target frequency band; Signal reconstruction: reconstructing the signal using the wavelet packet coefficients to obtain the target signal, and using the target signal as the new first signal; First modeling: establish a wavelet neural network model; First training: using the first signal and the fault type label to train the wavelet neural network model to obtain a trained wavelet neural network model; Encoding: Encode n features separately to obtain n feature vectors; First calculation: calculate the correlation between all feature pairs based on n feature vectors, select the maximum value among all correlations and record it as the second data; Second judgment: judging whether the second data is less than a preset threshold value, if so, executing the third modeling step; if not, executing the fault identification step; Third modeling: establishing a PNN model, wherein the PNN model includes m neurons; Third training: using the first signal and the third signal to train the PNN model; Calculate similarity: Calculate the similarity between the i-th feature vector and each neuron in the trained PNN model to obtain m similarity results; Calculate output: Calculate the output probability of the trained PNN model based on m similarity results; Probability judgment: judge whether the output probability is greater than a preset threshold, if so, execute the fault identification step; if not, execute the deletion step; Deletion: Delete one feature in the current feature pair and execute the first calculation step; Fault identification: The second signal is input into the trained wavelet neural network model and the fault type is output.

2. The arc fault identification method based on wavelet neural network according to claim 1 is characterized in that: In the first acquisition step, a normal historical current signal is also acquired and recorded as a third signal, and after executing the signal processing step and before executing the first modeling step, the method further includes: First extraction: extract features from the second signal to obtain n features; Second modeling: building a CNN model; Second training: the CNN model is trained using the first signal and the third signal. During the training process, the label of the third signal is set to 0, and the label of the first signal is set to 1. First classification: input n features into the trained CNN model in sequence to obtain n output data; Statistics: Count the number of 1s in n output data and record it as the first data; First judgment: judge whether the first data is greater than a preset threshold, if so, execute the first modeling step; if not, execute the second acquisition step.

3. The arc fault identification method based on wavelet neural network according to claim 1 or 2, characterized in that: After executing the step of signal acquisition and before executing the step of signal processing, the method further includes: Second extraction: using a global search algorithm to traverse the signal set consisting of historical current signals to obtain the fault arc characteristics; Second classification: Use a classifier to classify each fault arc feature, obtain the type label of each fault arc feature, and integrate the fault arc features with the same type label into a feature set; Assign weights: Use the random forest algorithm to assign weights to each fault arc feature in the feature set; Feature fusion: performing feature weighted fusion based on the fault arc feature and the weight, obtaining the weighted fused fault arc feature, which is recorded as the first feature; Second calculation: calculating the correlation between the first feature and the second signal, recorded as third data; Data judgment: judge whether the third data is greater than a preset threshold. If so, use the first feature to replace the first signal to perform the signal processing step; if not, execute the second acquisition step.

4. The arc fault identification method based on wavelet neural network according to claim 3 is characterized in that: After performing the step of second classification and before performing the step of assigning weights, the method further includes: Table building: building a management table, wherein the management table stores the fault arc features in the feature set; Generate neighborhood features: Generate neighborhood features based on fault arc features using a generative adversarial network; Evaluation: The evaluation result of the neighborhood feature is calculated using an evaluation function, and the calculation model of the evaluation function is as follows: ； in, is the evaluation function; N is the number of neighborhood features; is the jth eigenvalue of the i-th neighborhood feature; p is the number of eigenvalues ​​of the i-th neighborhood feature; Third judgment: judging whether the evaluation result is greater than a preset threshold, if so, executing the step of adding neighborhood features; if not, executing the step of iteration; Add neighborhood features: add neighborhood features to the management table; Iteration: Integrate the neighborhood features whose evaluation results meet the preset threshold into a new feature set, and execute the steps of generating neighborhood features until the preset stop condition is met.

5. The method for fault arc identification based on wavelet neural network according to claim 4 is characterized in that: After executing the step of allocating weights and before executing the step of feature fusion, the following steps are also included: Feature screening: including sorting, feature selection and integration; Sorting: sort the fault arc features in each feature set in descending order of weight to obtain the fault arc feature sequence; Feature selection: In the fault arc feature sequence, select the first k fault arc features; Integration: The selected fault arc features are reintegrated into a new feature set and the weight assignment step is performed.

6. The arc fault identification method based on wavelet neural network according to claim 1 or 2, characterized in that: After performing the decomposition step and before performing the signal reconstruction step, the method further includes: Constructing a matrix: integrating the wavelet packet coefficients into a two-dimensional matrix, normalizing the two-dimensional matrix, and obtaining a normalized two-dimensional matrix, which is recorded as the first matrix; Generating an image: converting each element in the first matrix into a grayscale value, and generating a grayscale image based on the grayscale value; Integrating feature matrix: extracting SIFI feature descriptors of the grayscale image, and integrating the SIFI feature descriptors into a feature matrix; Visualization processing: using the t-SNE algorithm to visualize the feature matrix to obtain feature distribution results; Fourth judgment: judge whether the characteristic distribution result meets expectations. If so, execute the signal reconstruction step; if not, execute the decomposition step.

7. The method for fault arc identification based on wavelet neural network according to claim 6 is characterized in that: After executing the fourth determination step and before executing the decomposition step, the method further includes: Check: Check whether the first signal has classification errors. If so, reassign the fault type label to the first signal; if not, increase the number of decomposition layers.

8. A fault arc identification system based on wavelet neural network, the system is used to execute the method according to any one of claims 1 to 7, characterized in that: include: A signal acquisition module, comprising a first acquisition unit and a second acquisition unit; A first acquisition unit, used for acquiring a historical current signal when a fault arc occurs and a corresponding fault type label, and recording the historical current signal when a fault arc occurs as a first signal; A second acquisition unit, used for acquiring a current signal to be processed, recorded as a second signal; A signal processing module, which is in communication connection with the first acquisition unit, comprises a decomposition unit and a signal reconstruction unit; A decomposition unit, used to decompose the first signal using a wavelet packet transform algorithm to obtain wavelet packet coefficients of a target frequency band; A signal reconstruction unit, used to reconstruct the signal using the wavelet packet coefficients to obtain a target signal, and use the target signal as a new first signal; The first modeling module is used to establish a wavelet neural network model; A first training module is connected to the signal processing module and the first modeling module for training the wavelet neural network model using the first signal and the fault type label to obtain a trained wavelet neural network model; The fault identification module is connected to the second acquisition unit and the first training module for inputting the second signal into the trained wavelet neural network model and outputting the fault type.

Images