Partial discharge mode identification method based on combination of wavelet scattering network and sample entropy

Abstract

The invention relates to the technical field of partial discharge mode recognition, in particular to a partial discharge mode recognition method based on a wavelet scattering network in combination with sample entropy. According to the method, partial discharge signal data is acquired and preprocessed, a three-layer wavelet scattering network is constructed by using wavelet scattering transformation to extract signal features, multi-scale coefficient features with translation invariance and local deformation stability are obtained, feature optimization is performed in combination with multi-scale sample entropy analysis, and the partial discharge signal data is obtained. And finally, a support vector machine (SVM) is adopted to carry out classification and identification. Through the method, different types of partial discharge signals can be efficiently and accurately identified, the identification precision and the generalization ability of the model are remarkably improved, and especially under the condition of small samples, the problems of low identification precision and poor generalization ability caused by insufficient data in a traditional method can be effectively avoided. The method has high application value and can be used for fault detection and early warning of power equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of partial discharge pattern recognition, and in particular to a partial discharge pattern recognition method based on wavelet scattering network combined with sample entropy. Background Technique

[0002] Partial discharge is a common insulation fault phenomenon in power equipment, especially in cables. Its occurrence is usually accompanied by cable aging, damage to insulation materials, and degradation of electrical performance. Partial discharge will cause partial breakdown of the cable insulation medium, and may lead to cable faults in severe cases, threatening the safety and stable operation of the power system. Therefore, early detection and accurate recognition of partial discharge signals in cables are of great significance for preventive maintenance and fault diagnosis of power equipment. Existing partial discharge recognition methods usually rely on a large amount of data to construct PRPD / PRPS maps for recognition. The PRPD / PRPS maps are used for pattern recognition by showing the time-domain and frequency-domain characteristics of partial discharge. The advantage of this method is that it can intuitively reflect the characteristics of different partial discharges, but its disadvantage is that a large amount of experimental data is often required to construct an accurate map, which poses high requirements for data acquisition and processing in the real-time monitoring and large-scale application of power equipment. In addition, with the introduction of machine learning methods, many studies have adopted data augmentation techniques to enhance sample diversity and improve the robustness of the model. However, this method depends on the accuracy of the data, and noise may be introduced during the augmentation process, affecting the recognition effect. Existing partial discharge signal recognition methods still face certain challenges in small-sample learning and computational efficiency. Therefore, there is an urgent need for an efficient and accurate signal recognition method to achieve accurate recognition of different types of partial discharges. Summary of the Invention

[0003] Aiming at the above problems, the purpose of the present invention is to provide a partial discharge pattern recognition method based on wavelet scattering network combined with sample entropy, so as to solve the problems of low recognition rate and poor generalization ability in the prior art, and provide a small-sample, efficient and accurate partial discharge pattern recognition method.

[0004] To achieve the above purpose, the present invention adopts the following technical solutions: A partial discharge pattern recognition method based on a diffusion model and a graph attention network, which is characterized by including the following steps:

[0005] S1: Obtain actual partial discharge signal data and preprocess the signal;

[0006] S2: Construct a three-layer wavelet scattering network through wavelet scattering transform, set the optimal wavelet scattering parameters according to the time-frequency diagram, input the processed partial discharge signal data for feature extraction, and obtain the third-order wavelet scattering coefficients of the partial discharge signal;

[0007] S3: Calculate the multi-scale coefficient sample entropy for each of the 1st and 2nd order scattering coefficients, analyze the complexity of the coefficients, and combine the coefficient entropy values of each sample into new features;

[0008] S4: Use the obtained new features as input, use SVM for partial discharge type recognition, adopt the grid search method to optimize the kernel function parameters, select the optimal parameter combination to improve the classification accuracy and the generalization ability of the model, and use accuracy, F1 score, and AUC as evaluation functions to measure the overall classification performance of the model;

[0009] Furthermore, the specific step S1 is as follows: Collect the actual cable partial discharge data. To ensure the accuracy of subsequent analysis and the representativeness of the signal, only the discharge signal is analyzed here, and the signal is intercepted. For 3 discharge modes, the partial discharge signal within a single power frequency cycle is truncated as a sample, and normalization processing is performed simultaneously.

[0010] Furthermore, the specific step S2 is as follows: The wavelet scattering network applies the wavelet scattering transform multiple times in a cascaded manner to gradually extract the high-order features of the signal. Adopt a three-layer wavelet scattering network architecture to convert the partial discharge signal into a high-level feature representation and be able to capture the change characteristics of the signal at multiple scales.

[0011] The wavelet scattering transform is divided into 3 steps: First, the wavelet filter convolves the signal to extract the local structure of each layer of the signal; after the wavelet convolution operation, a non-linear process is added to extract the non-linear features of the signal so that the network can process complex patterns. After the non-linear process, a smoothing operation is then performed. Thus, coefficient features with both translational invariance and local deformation stability are obtained.

[0012]

[0013] Among them, α(t) is the signal, is the low-pass filter, and are different wavelet bases; λ and μ are scale functions that control the scale of the filter, and S is the scattering coefficient, i.e., the wavelet scattering feature.

[0014] According to the characteristics of the partial discharge signal, optimize the scattering parameters such as the scale function, time-invariant characteristics, and quality factor to generate the best wavelet scattering coefficient matrix.

[0015] When generating the scattering coefficient time-frequency diagram for the obtained wavelet scattering coefficients, evaluate the best scattering parameters as the optimized network model, and adopt the PSD consistency algorithm as the evaluation index.

[0016] According to the definition of the Fourier transform, the energy spectral density calculation formula is:

[0017]

[0018] Among them, P(f) is the power spectral density of the signal at frequency f, x(t) is the signal, j is the imaginary unit, and t is time.

[0019] The PSD consistency can be evaluated by the difference between the actual PSD of the calculated scattering coefficient time-frequency diagram and the theoretical PSD of the original signal. The following two functions are used to measure the quality of the scattering parameters. Good parameters can obtain time-frequency coefficients that can characterize most of the signal energy and are almost uniformly distributed:

[0020] Characterized by the mean square error (MSE) function. The smaller the MSE, the closer the PSD calculated from the time-frequency diagram is to the PSD of the original signal, indicating a higher accuracy of the time-frequency method.

[0021]

[0022] Among them: P actua l(f i ) is the actual PSD at the frequency point f calculated from the time-frequency diagram. P i (f theoretical (f i ) is the theoretical PSD of the original signal.

[0023] Characterized by the spectral correlation coefficient. The closer the correlation coefficient value is to 1, the higher the consistency between the PSD calculated by the time-frequency method and the PSD of the original signal.

[0024]

[0025] and are the means of the actual and theoretical power spectral densities.

[0026] Find the best parameters from the above steps, and generate wavelet scattering coefficient features that can represent the signal through a three-layer wavelet scattering network

[0027] Furthermore, the specific content of step S3 is as follows: The scattering coefficient is also a set of time series, and the sample entropy algorithm is a method for evaluating the complexity of time series data. The sample entropy compresses the redundant information of the scattering coefficient to a certain extent, retains the essential complexity of the signal. At the same time, the scattering coefficients of different samples may vary in time series, but the complexity of the coefficient samples is generally the same and has good classification ability, which can improve the classification accuracy.

[0028] Adopt the multi-scale sample entropy function to optimize the features of each generated first- and second-order scattering coefficient. Through the calculated sample entropy values, combine them into a new feature vector. The formulas for calculating the sample entropy for the first-order and second-order scattering coefficients are as follows:

[0029] For the first-order scattering coefficient S (1) α(t, λ):

[0030]

[0031] For the first-order scattering coefficient S (2) α(t, λ):

[0032]

[0033] where A (1)m (r) and B (1)m (r) are the matching numbers of the first-order coefficients, A (2)m (r) and B (2)m (r) are the matching numbers of the second-order coefficients, m is the embedding dimension, and r is the tolerance.

[0034] Furthermore, the specific step S4 is as follows: Input the multi-scale scattering coefficient sample entropy into the SVM model. To optimize the parameters of the SVM model, use the grid search method for tuning, and select the optimal parameter combination of the kernel function type, sum function parameter, and regularization parameter C. Train the SVM model using the training set and evaluate it using the test set, and calculate the accuracy, F1 score, and AUC value.

[0035] Accuracy:

[0036]

[0037] F1 score:

[0038]

[0039] AUC value:

[0040]

[0041] Among them, TP is the true positive example, representing the number of samples correctly predicted as the positive class, TN is the true negative example, representing the number of samples correctly predicted as the negative class. FP is the false positive example, representing the number of negative class samples wrongly predicted as the positive class. FN is the false negative example, representing the number of positive class samples wrongly predicted as the negative class. The calculation methods of each index comprehensively evaluate the performance of the classification model.

[0042] The beneficial effects of the present invention are as follows:

[0043] (1) In the field of partial discharge, a wavelet scattering network is introduced. By extracting wavelet features with translational invariance and local deformation stability at multiple scales, it can effectively capture the local detail changes in the signal, improving the recognition ability of complex signal patterns. The wavelet scattering network can enhance the generalization ability of the model through its multi-scale feature extraction ability, thereby improving the recognition accuracy under small samples.

[0044] (2) Combining the wavelet scattering coefficients with sample entropy optimizes the feature expression method and enhances the model's ability to depict the signal complexity. Through the calculation of sample entropy, the complexity of the signal can be effectively evaluated, further enhancing the distinctiveness of the features, improving the accuracy of partial discharge type recognition and the generalization ability of the model, and effectively improving the classification performance in different types of partial discharge signals. Description of the Drawings

[0045] To more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present application and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0046] Figure 1 It is a flowchart of a partial discharge pattern recognition method based on a wavelet scattering network combined with sample entropy provided by an embodiment of the present application;

[0047] Figure 2 It is a schematic diagram of a three-layer wavelet scattering network structure provided by an embodiment of the present application.

[0048] Figure 3 It is a schematic diagram of a structure combining wavelet scattering network coefficients with sample entropy provided by an embodiment of the present application. Detailed Embodiments

[0049] To make the objectives, technical solutions, and advantages of the present application clearer, the following will elaborate on the related technical solutions in conjunction with the drawings. It should be noted that the embodiments described here are only part of the content of the present application, not all embodiments. All other embodiments that can be obtained by those of ordinary skill in the art based on the embodiments in the present application without creative efforts should be included within the protection scope of the present application.

[0050] Figure 1 It is a flowchart of a partial discharge pattern recognition method based on a wavelet scattering network combined with sample entropy provided by an embodiment of the present application. The partial discharge pattern recognition method based on a wavelet scattering network combined with sample entropy specifically includes the following steps:

[0051] Step S101, detect different partial discharge signals through sensors and collect the corresponding partial discharge signal waveforms for input.

[0052] As an embodiment, step S102 includes: the types of signals actually detected and collected are not limited to UHF signals, ultrasonic signals, electromagnetic wave signals, current and voltage pulse signals, etc.

[0053] Step S102, preprocess the partial discharge signals.

[0054] As an embodiment, step S102 includes: truncate the partial discharge signals; normalize the amplitude of the truncated signals to the range of [0, 1].

[0055] Among them, the truncation process is to unify the specifications of the discharge signals and make fault samples.

[0056] Step S103, input the preprocessed partial discharge signals into the wavelet scattering network to generate features. Among them, the schematic diagram of the wavelet scattering network architecture is as Figure 2 shown.

[0057] As an embodiment, step S103 includes: construct a three-layer wavelet scattering network, find the optimal scattering parameters, and generate wavelet coefficient features.

[0058] The process of the three-layer wavelet scattering network is to extract the multi-scale features of the signal through a continuous wavelet scattering process for the data preprocessed in S102: the preprocessed partial discharge signals are convolved with the wavelet basis to extract the low-frequency features of the signal. In this process, the low-frequency coefficients obtained by convolution are non-linearly processed to enhance the translation invariance and local deformation stability of the signal. After smoothing processing, the corresponding wavelet scattering coefficient matrix is obtained.

[0059] Among them, by calculating the mean square error (MSE) and spectral correlation coefficient between the actual PSD of the scattering coefficient time-frequency diagram and the theoretical PSD of the original signal, the parameters of the wavelet scattering network are optimized, and finally the wavelet scattering coefficient features that can effectively represent the signal are generated.

[0060] Step S104, extract sample entropy features by combining the wavelet scattering matrix obtained in S103 with the multi-scale sample entropy algorithm. Among them, the schematic diagram of the recognition process of combining the wavelet scattering network coefficients with sample entropy is as Figure 3 shown.

[0061] As an embodiment, step S104 includes: calculate the coefficient sample entropy and combine new features.

[0062] Among them, sample entropy calculation is performed on the multi-scale wavelet scattering coefficients of the 2nd and 3rd layers containing the main energy in S104. The optimal embedding dimension and tolerance are selected, the sample entropy of the coefficients in the time series is calculated, and the obtained sample entropy structure is arranged into a one-dimensional vector according to the coefficient serial number for classification processing.

[0063] Step S105: Classify the feature vectors described in the above step S104 using SVM, and output the local discharge pattern recognition result.

[0064] As an embodiment, step S106 includes: parameter tuning; classification result output; pattern matching.

[0065] The parameter tuning is to tune the hyperparameters of the SVM model through grid search and select the optimal parameters for model training.

[0066] The classification result output is to output the classification result of the SVM model, including the classification label and actual label of each sample, and evaluate the classification accuracy using the accuracy, F1 score, and AUC value.

[0067] The pattern matching is to match the node feature vectors to the preset local discharge patterns to identify different types of local discharge pattern types.

[0068] Step S106: Output the final local discharge pattern recognition result for fault detection and warning of power equipment.

[0069] Finally, it should be noted that the above are only the preferred embodiments of the present application and are not used to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for partial discharge pattern recognition based on a diffusion model and a graph attention network, characterized in that, The method includes: Obtain partial discharge signal data and perform data preprocessing; Use a three-layer wavelet scattering network to extract scattering coefficient features from the partial discharge signal data; Process the first- and second-order scattering coefficients at multiple scales using the sample entropy algorithm and combine them into new features; Input the features and preset labels into an SVM for classification and output the recognition result.

2. The method according to claim 1, wherein During the data preprocessing, it also includes: Truncate the signal data, take the partial discharge signal within a useful single power frequency cycle, and reduce the useless signal segments; Normalize the processed signal samples.

3. The method according to claim 1, wherein The wavelet scattering coefficient features also include: Construct a three-layer wavelet scattering network through wavelet scattering transform, Find the optimal wavelet scattering parameters for the partial discharge signal samples.

4. The method according to claim 3, wherein The construction of the wavelet scattering network is achieved through the following steps: The wavelet scattering network applies wavelet scattering transform multiple times in a cascaded manner The wavelet scattering transform is divided into 3 steps: First, the signal is convolved with a wavelet filter to extract the local structure of each layer of the signal; after the wavelet convolution operation, a nonlinear process is added to extract the nonlinear features of the signal, enabling the network to handle complex patterns. After the nonlinear process, a smoothing operation is then performed. Thus, coefficient features with both translational invariance and local deformation stability are obtained. where α(t) is a signal, is a low-pass filter, and are different wavelet bases; λ and μ are scaling functions that control the scale of the filter, and S is the scattering coefficient, i.e., the wavelet scattering feature.

5. The method according to claim 3, wherein The optimization of the scattering parameters includes the following steps: According to the characteristics of the partial discharge signal, optimize scattering parameters such as the scaling function, time-invariant characteristics, and quality factor to generate the best wavelet scattering coefficient matrix. When generating the scattering coefficient time-frequency diagram for the obtained wavelet scattering coefficients, evaluate the best scattering parameters as the optimized network model, and use the PSD consistency algorithm as the evaluation index. According to the definition of Fourier transform, the energy spectral density calculation formula is: where P(f) is the power spectral density of the signal at frequency f, x(t) is the signal, j is the imaginary unit, and t is time. PSD consistency can be evaluated by calculating the difference between the actual PSD of the calculated scattering coefficient time-frequency diagram and the theoretical PSD of the original signal. Use the following two functions to measure the quality of the scattering parameters. Good parameters result in time-frequency coefficients that can represent most of the signal energy and have an almost consistent distribution: The mean square error (MSE) function is characterized. The smaller the MSE, the closer the PSD calculated from the time-frequency diagram is to the PSD of the original signal, indicating a higher accuracy of the time-frequency method. Where: P actua l(f i ) is the frequency point f calculated from the time-frequency diagram i The actual PSD on P theoretical (f i ) is the theoretical PSD of the original signal. The spectral correlation coefficient is characterized. The closer the correlation coefficient value is to 1, the higher the consistency between the PSD calculated by the time-frequency method and the PSD of the original signal. and is the mean of the actual and theoretical power spectral densities. Find the best parameters from the above steps, and generate wavelet scattering coefficient features that can represent the signal through a three-layer wavelet scattering network.

6. The method according to claim 1, characterized in that The process of generating the sample entropy features is achieved through the following steps: Use the multi-scale sample entropy function to optimize the features of each generated first- and second-order scattering coefficient. By calculating the obtained sample entropy values, combine them into a new feature vector. The formulas for calculating the sample entropy for the first- and second-order scattering coefficients are as follows: For the first-order scattering coefficient S (1) α(t,λ): For the first-order scattering coefficient S (2) α(t,λ): Among them, it is A (1)m (r) and B (1)m (r) is the matching number of the first-order coefficient, A (2)m (r) and B (2)m (r) is the matching number of the second-order coefficient, m is the embedding dimension, and r is the tolerance.

7. The method according to claim 1, characterized in that, The SVM classification includes the following steps: Input the multi-scale scattering coefficient sample entropy into the SVM model to optimize the parameters of the SVM model. Use the grid search method for tuning to select the optimal parameter combination of the kernel function type, kernel function parameters, and regularization parameter C. Train the SVM model using the training set, evaluate it using the test set, and calculate the accuracy, F1 score, and AUC value. Accuracy: F1 score: AUC value: Among them, TP is the true positive, representing the number of samples correctly predicted as the positive class, TN is the true negative, representing the number of samples correctly predicted as the negative class. FP is the false positive, representing the number of negative class samples incorrectly predicted as the positive class. FN is the false negative, representing the number of positive class samples incorrectly predicted as the negative class. The calculation methods of each index comprehensively evaluate the performance of the classification model.

Images