Discharge ultrasonic signal classification method based on class specificity GAN and feature weighting
By using a method based on class-specific GAN and feature weighting, the problems of feature redundancy, insufficient generalization ability of small samples and inter-class feature confusion in partial discharge classification are solved, and high-precision partial discharge classification is achieved.
Patent Information
- Application Number
- CN202510728783.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-03
- Publication Date
- 2025-09-12
AI Technical Summary
Existing partial discharge classification methods have problems such as feature redundancy and noise sensitivity, insufficient generalization ability for small samples, and inter-class feature confusion.
A discharge ultrasonic signal classification method based on class-specific GAN and feature weighting was adopted. The dimensionality was reduced to 5 dimensions through principal component analysis. An independent GAN network was constructed and WGAN-GP was used to enhance the model stability. The Manhattan distance weight was assigned in combination with random forest feature importance to perform optimal matching classification.
The accuracy and robustness of partial discharge classification are improved, the feature differences between different categories are effectively highlighted, the model instability and inter-class interference in small sample training are alleviated, and the reliability of classification is improved.
Smart Images
Figure CN120632600A_ABST
Abstract
Description
Technical field:
[0001] The present invention relates to the technical field of insulation status monitoring of high-voltage electrical equipment, and in particular to a pattern recognition and classification method based on partial discharge signals. Background technology:
[0002] Partial discharge (PD) is an early sign of insulation degradation in high-voltage power equipment. Accurate detection is crucial for preventing equipment failures and ensuring safe grid operation. Ultrasonic detection technology, with its advantages of being non-invasive and resistant to electromagnetic interference, has become a mainstream method for PD monitoring. However, existing methods face the following challenges:
[0003] (1) Feature redundancy and noise sensitivity: Traditional methods rely on manual extraction of time domain or frequency domain features (such as pulse counts and spectrum peaks), which have problems of one-sided information and redundancy.
[0004] (2) Limited small sample classification performance: The PD samples available in actual projects are limited, which makes the learning model prone to overfitting.
[0005] (3) Inter-class feature confusion: Different discharge types (such as surface discharge and suspension discharge) behave similarly in the time-frequency domain, and the classification boundaries are blurred. Summary of the invention:
[0006] The technical problems to be solved by the present invention are: feature redundancy in partial discharge classification methods, insufficient generalization ability of small samples, and confusion of inter-class features.
[0007] To solve the above technical problems, the present invention provides a technical solution: a discharge ultrasonic signal classification method based on class-specific GAN and feature weighting, the steps of which are:
[0008] Step 1: Extract 21-dimensional features from the time domain, frequency domain, and time-frequency domain, and reduce the dimensionality to 5 dimensions through principal component analysis;
[0009] Step 2: Build independent GAN networks for each type of discharge, and use WGAN-GP (Wasserstein GAN with Gradient Penalty) gradient penalty to enhance the stability of the intra-class model;
[0010] Step 3: Assign Manhattan distance weights based on the importance of random forest features, highlight the contribution of key feature dimensions, and construct weighted Manhattan distance;
[0011] Step 4: Use the obtained weighted Manhattan distance to optimally match the samples generated by various GANs to obtain the sample classification results.
[0012] Furthermore, in step 1, the process of extracting the time domain features is: sequentially obtaining the peak amplitude, rise time, pulse width, zero crossing rate, kurtosis and skewness features of the time domain.
[0013] Furthermore, in step 1, the process of frequency domain feature extraction is: performing fast Fourier transform on the ultrasonic signal to obtain a spectrum, and extracting the spectrum centroid, sub-band energy ratio and main frequency component amplitude features.
[0014] Furthermore, in the step 1, the process of extracting the time-frequency domain features is as follows: using three-layer wavelet packet decomposition to obtain wavelet packet energy entropy and node energy ratio features.
[0015] Furthermore, in step 2, a separate GAN network is designed for each type of discharge, using a fully connected network to adapt low-dimensional features. Each adversarial generative network structure includes a generator (G) and a discriminator (D). The number of GANs matches the number of discharge categories.
[0016] Furthermore, in step 2, the independent GAN network training process for each type of discharge is as follows: when training each type of GAN, the input generator signal is a variety of real samples, and the sample features are reconstructed through the GAN. By training multiple GANs by separating sample categories, the feature differences between samples of different categories are highlighted, and the model difficulty is reduced and the stability is improved.
[0017] Furthermore, in step three, the weighted Manhattan distance used in the classification process is used to evaluate the distance between the feature vector output by the generator and the feature vector of the real sample, and by assigning different weights to different features, the contribution of effective features is highlighted.
[0018] Furthermore, in step 3, the weight distribution process is as follows: Random Forest calculates the importance of each dimension of the 5-dimensional features, and the feature importance score w i ', the normalized weight is:
[0019]
[0020] Where: w i Represents the importance weight of the i-th dimension feature. The larger the weight, the higher the contribution of the dimension to the classification.
[0021] Furthermore, in step 4, the classification rule is: for a sample to be identified, the sample is input into various trained GAN generators to generate samples, and then the weighted Manhattan distance is used to match the samples generated by various GANs with the original samples.
[0022] The beneficial effects of the present invention are:
[0023] 1. This paper constructs a separate GAN network for each type of discharge, highlighting the feature differences between samples of different categories while reducing model complexity. Furthermore, a gradient penalty (WGAN-GP) is used to improve the stability of small sample models. GAN focuses on improving the ability to generate small samples of specific types to ensure the fit of the generated model, while weighted Manhattan distance focuses on feature importance. The combination of the two improves the reliability of distance matching.
[0024] 2. During the initial processing of this application, the frequency band is divided into several sub-bands according to the frequency range of the ultrasonic signal, and the energy ratio is calculated. In addition, a three-layer wavelet packet is used to decompose the signal to generate eight feature nodes to capture the local characteristics of the signal in detail. Principal component analysis (PCA) is used for dimensionality reduction processing to eliminate redundant information and reduce feature dimensions and computational complexity.
[0025] 3. This application proposes to construct a GAN with exclusive categories for each type of sample, which is called "class-specific" GAN. Class-specific GAN avoids inter-class interference by independently modeling the characteristic distribution of each type of discharge signal, and combines the generative adversarial mechanism to implicitly expand the training data appropriately to buffer the quantitative defects of small samples.
[0026] In order to make the above and other objects, features and advantages of the present invention more clearly understood, preferred embodiments are given below with reference to the accompanying drawings for detailed description. Description of the drawings:
[0027] In order to more clearly illustrate the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only partial schematic diagrams of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0028] Figure 1 This is a flowchart of the application;
[0029] Figure 2 A block diagram of the class-specific adversarial generation network training process for this application;
[0030] Figure 3 This is the classification block diagram of this application. Specific implementation method:
[0031] Embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although certain embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.
[0032] In the present invention, unless otherwise clearly specified and limited, terms such as "installation", "setting", "connection", "fixation", and "screw-on" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integrated connection; it can be a mechanical connection, a direct connection, or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances.
[0033] The names of the messages or information exchanged between multiple devices in the embodiments of the present invention are only used for illustrative purposes and are not used to limit the scope of these messages or information.
[0034] Example
[0035] like Figure 1-3 As shown in FIG, a discharge ultrasonic signal classification method based on class-specific GAN and feature weighting is described, wherein the steps are:
[0036] Step S1, extracting 21-dimensional features from the time domain, frequency domain, and time-frequency domain, and reducing the dimensionality to 5 dimensions through principal component analysis;
[0037] Step S2: construct independent GAN networks for each type of discharge, and use WGAN-GP gradient penalty to enhance intra-class model stability;
[0038] Step S3: assign Manhattan distance weights based on random forest feature importance, highlight the contribution of key feature dimensions, and construct weighted Manhattan distance;
[0039] Step S4: Optimally match the samples generated by various GANs using the obtained weighted Manhattan distance to obtain the sample classification results.
[0040] In step S1, the process of extracting time domain features is: sequentially obtaining the peak amplitude, rise time, pulse width, zero crossing rate, kurtosis and skewness features of the time domain.
[0041] Among them, it is proposed to combine the time domain, frequency domain and time-frequency domain features to fully reflect the physical characteristics of ultrasonic signals.
[0042] During processing, the ultrasound signal's frequency band is divided into several subbands based on its frequency range, and energy ratios are calculated. Furthermore, the signal is decomposed using a three-layer wavelet packet, generating eight feature nodes to capture the signal's local characteristics. To reduce feature dimensionality and computational complexity, PCA is used for dimensionality reduction.
[0043] For the ultrasonic signal x(t) sampled at 80 MHz, 21-dimensional features are extracted from the time domain, frequency domain, and time-frequency domain:
[0044] (1) Time domain features (6 dimensions)
[0045] ①Peak Amplitude:
[0046] V peak =max(|x(t)|) (1)
[0047] Where: x(t) is the time domain signal, V peak Indicates the maximum absolute value of the signal, reflecting the maximum intensity of the discharge pulse.
[0048] ②Rise Time:
[0049] T rise =t 90％ -t 10％ (2)
[0050] Where: t 10％ and t 90％ These are the time points when the signal reaches 10% and 90% of its peak value, respectively, describing the steepness of the pulse front.
[0051] ③ Pulse Width: The duration during which the signal amplitude exceeds 50% of the peak value, representing the energy concentration.
[0052] ④Zero-Crossing Rate (ZCR):
[0053]
[0054] Where: II is the indicator function (1 when crossing the zero point, otherwise 0), N is the number of signal sampling points, which measures the high-frequency oscillation characteristics of the signal.
[0055] ⑤Kurtosis:
[0056]
[0057] Where: μ is the signal mean, σ is the standard deviation, and K>0 means the pulse is sharper than the normal distribution.
[0058] ⑥Skewness:
[0059]
[0060] Where: S>0 means the right tail of the signal is longer, otherwise the left tail is longer.
[0061] (2) Frequency domain features (6 dimensions)
[0062] Perform a fast Fourier transform (FFT) on x(t) to obtain the spectrum X(f) and extract the following features:
[0063] ① Spectral Centroid:
[0064]
[0065] Where: f k is the kth frequency component, C represents the center position of the spectrum energy distribution, and reflects the main frequency range of the signal.
[0066] ②Subband energy ratio (4D):
[0067] According to the Nyquist sampling theorem, the highest frequency of the signal is half of the sampling frequency, which is 40 MHz. The 0-40 MHz spectrum is divided into four sub-bands (0-10 MHz, 10-20 MHz, 30-40 MHz, and 50-0 MHz). The energy proportion of each sub-band is calculated as follows:
[0068]
[0069] Where: a i and b i are the start and end frequencies of each sub-band, Indicates the energy proportion of the corresponding frequency band.
[0070] ③Dominant Frequency Magnitude:
[0071] A dom =max(|X(f)|) (8)
[0072] Where: Adom is the intensity of the frequency component corresponding to the maximum amplitude in the spectrum, which represents the dominant oscillation frequency of the signal.
[0073] (3) Time-frequency domain features (9 dimensions)
[0074] Using 3-layer wavelet packet decomposition (the mother wavelet is db4), the following features are extracted:
[0075] ① Wavelet Packet Energy Entropy:
[0076]
[0077] Where: E j is the energy of the jth node, and H represents the complexity of measuring the energy distribution (the higher the entropy value, the more dispersed the distribution).
[0078] ②Node energy ratio (8 dimensions):
[0079]
[0080] Where: It represents the ratio of the energy of the jth node to the total energy.
[0081] The process of reducing dimensionality to 5 dimensions through principal component analysis is as follows:
[0082] The principal component analysis (PCA) is used to reduce the 21-dimensional features to 5 dimensions. Let the original feature matrix be Its covariance matrix is: Through eigenvalue decomposition, the first five principal components are selected.
[0083] In step S2, the independent GAN network of each type of discharge design adopts a fully connected network to adapt the low-dimensional features, and each type of adversarial generation network structure includes a generator (G) and a discriminator (D).
[0084] Among them, the training process of the network is:
[0085] The generator receives a 5-dimensional discharge sample vector as input, transforms it through a 3-layer fully connected network (5→128→64→5), uses the LeakyReLU (α=0.2) activation function in the hidden layer, and uses a linear layer to generate a 5-dimensional feature vector in the output layer. During the training process, whether to add a small amount of random noise as the input of the generator can be decided based on the network performance.
[0086] The discriminator receives a 5-dimensional feature vector (real or generated) as input, and performs feature extraction and discrimination through a 3-layer fully connected network (5→128→64→1). The hidden layer uses the LeakyReLU (α=0.2) activation function, and the output layer uses a linear layer. For a certain type of sample, Figure 2 The training process shown obtains the generator of this type of samples.
[0087] Loss function: Use gradient penalty to optimize the generator and discriminator to improve training stability and prevent small sample training crashes.
[0088] In step S3, a weighted Manhattan distance is constructed to highlight the contributions of important features by assigning different weights to different features.
[0089] In step S4, the weighted Manhattan distance is used to evaluate the distance between the feature vector output by the generator and the feature vector of the original real sample. The classification method is as follows: Figure 3 shown.
[0090] Assume that there are g types of discharge signals, and k represents the kth type of sample. Perform the following steps:
[0091] (1) Use the test set data to train g types of GANs.
[0092] (2) Generate reconstructed feature vector: Take a sample to be identified X text Input all GAN generators to obtain K reconstructed feature vectors, each feature vector is recorded as X k gen (k=1,2...,g).
[0093] (3) Calculate X separately text Weighted Manhattan distance to each reconstructed feature vector (K in total):
[0094]
[0095] Where: Dk represents X text With X k gen The weighted Manhattan distance of x. i represents all dimensions of the feature, ranging from 1 to 5. test，i and The i-th dimension features of a test sample and the k-th class sample generated for the sample respectively.
[0096] (4) Compare the sizes of D1, D2…Dk…Dg, and the category k corresponding to the smallest Dk is the prediction result.
[0097] The weight distribution process is as follows: Random Forest calculates the importance of each dimension of the 5-dimensional features, and the feature importance score w i ', the normalized weight is:
[0098]
[0099] Where: w i Represents the importance weight of the i-th dimension feature. The larger the weight, the higher the contribution of the dimension to the classification.
[0100] The following is an experimental analysis of the technical solution proposed in this application
[0101] Using the multidimensional feature extraction method proposed in this paper, 50 data sets each for plate-to-plate discharge, suspended discharge, corona discharge, and multi-tip-to-plate discharge were generated, and the feature vectors were reduced to 5 dimensions. Class-specific GAN models corresponding to each of the four discharge types were trained. Ten data sets for each of the four types of discharge signals were then selected and fed into the trained models. The weighted Manhattan distance was used to determine the sample type. Table 1 shows a comparison of these models with traditional random forest (maximum depth 3) and SVM (RBF kernel, regularization parameter) classification methods.
[0102] Table 1 Recognition effect
[0103]
[0104] In addition, to verify the effectiveness of the weighted Manhattan coefficient and increase the complexity of the samples to be tested, we selected five groups of white noise signals and five groups of stained noise signals and compared them using the weighted Manhattan distance in this paper and the ordinary unweighted Manhattan distance, as shown in Table 2. It can be seen that the characteristic of the weighted Manhattan distance, which emphasizes primary features and weakens secondary features, is significantly effective for the accurate classification of complex samples.
[0105] Table 2 Comparison of unweighted Manhattan recognition results
[0106]
[0107] Experimental results show that the accuracy of the proposed method is more than 5% higher than that of support vector machine (SVM), random forest and unweighted Manhattan distance; in addition, it has good feature characterization ability for samples, which is reflected in its good screening ability for complex samples (white noise and stained noise signals) that do not belong to known discharge categories; the proposed method provides a high-precision and robust solution for partial discharge classification in small sample scenarios.
[0108] Note that the above are only preferred embodiments of the present invention and the technical principles employed. Those skilled in the art will appreciate that the present invention is not limited to the specific embodiments herein, and that various obvious changes, readjustments, and substitutions are possible for those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments and may include many other equivalent embodiments without departing from the scope of the present invention. The scope of the present invention is determined by the scope of the appended claims.
Claims
1. A method for classifying discharge ultrasonic signals based on class-specific GAN and feature weighting, the steps of which are: Step 1: Extract 21-dimensional features from the time domain, frequency domain, and time-frequency domain, and reduce the dimensionality to 5 dimensions through principal component analysis; Step 2: Build independent GAN networks for each type of discharge and use WGAN-GP gradient penalty to enhance intra-class model stability; Step 3: Assign Manhattan distance weights based on the importance of random forest features, highlight the contribution of key feature dimensions, and construct weighted Manhattan distance; Step 4: Use the obtained weighted Manhattan distance to optimally match the samples generated by various GANs to obtain the sample classification results.
2. The method for classifying discharge ultrasonic signals based on class-specific GAN and feature weighting according to claim 1 is characterized by: In step 1, the process of time domain feature extraction is: sequentially obtaining the peak amplitude, rise time, pulse width, zero crossing rate, kurtosis and skewness features of the time domain.
3. The method for classifying discharge ultrasonic signals based on class-specific GAN and feature weighting according to claim 1 is characterized by: In the step 1, the process of frequency domain feature extraction is: performing fast Fourier transform on the ultrasonic signal to obtain a spectrum, and extracting the spectrum centroid, sub-band energy ratio and main frequency component amplitude features.
4. The method for classifying discharge ultrasonic signals based on class-specific GAN and feature weighting according to claim 1 is characterized by: In the step 1, the process of time-frequency domain feature extraction is: using three-layer wavelet packet decomposition to obtain wavelet packet energy entropy and node energy ratio features.
5. The method for classifying discharge ultrasonic signals based on class-specific GAN and feature weighting according to claim 1 is characterized by: In step 2, an independent GAN network is designed for each type of discharge, and a fully connected network is used to adapt the low-dimensional features. Each adversarial generative network structure includes a generator (G) and a discriminator (D). The number of GANs is consistent with the number of discharge categories.
6. The method for classifying discharge ultrasonic signals based on class-specific GAN and feature weighting according to claim 1 is characterized by: In step 2, the independent GAN network training process for each type of discharge is as follows: when training each type of GAN, the input generator signal is a variety of real samples, and the sample features are reconstructed through the GAN. By training multiple GANs by separating sample categories, the feature differences between samples of different categories are highlighted, which reduces model complexity and improves stability.
7. The method for classifying discharge ultrasonic signals based on class-specific GAN and feature weighting according to claim 1 is characterized by: In step 3, the weighted Manhattan distance used in the classification process is used to evaluate the distance between the feature vector output by the generator and the feature vector of the real sample. By assigning different weights to different features, the contribution of effective features is highlighted.
8. The method for classifying discharge ultrasonic signals based on class-specific GAN and feature weighting according to claim 1 is characterized by: In step 3, the weight distribution process is as follows: Random Forest calculates the importance of each dimension of the 5-dimensional features, and the feature importance score w i ', the normalized weight is: Where: w i Represents the importance weight of the i-th dimension feature. The larger the weight, the higher the contribution of the dimension to the classification.
9. The method for classifying discharge ultrasonic signals based on class-specific GAN and feature weighting according to claim 1, characterized in that: In step 4, the classification rule is: for a sample to be identified, the sample is input into various types of trained GAN generators to generate samples, and then the weighted Manhattan distance is used to match the samples generated by various types of GANs with the original samples.