Transformer on-load tap-changer vibration signal multivariate feature extraction method and system

Abstract

The invention discloses a transformer on-load tap-changer vibration signal multivariate feature extraction method and system, and relates to the technical field of on-load tap-changer fault diagnosis, and the method comprises the steps: optimizing a first parameter of feature mode decomposition; based on the optimized first parameter, decomposing a vibration signal switched by the on-load tap-changer to obtain a first feature; performing signal image conversion and image feature extraction based on the decomposed vibration signal switched by the on-load tap-changer to obtain a second feature; fusing the first feature and the second feature to form a high-dimensional feature space matrix; performing dimension reduction processing on the high-dimensional matrix to obtain multivariate features for analysis and identification; according to the method, signal features can be efficiently extracted, the accuracy and robustness of OLTC fault diagnosis are improved, and the method has important engineering application value and can be widely applied to the field of OLTC state monitoring and fault detection.

Description

Technical Field

[0001] The present invention relates to the technical field of on-load tap-changer fault diagnosis, and particularly to a method and system for extracting multiple features of vibration signals of a transformer on-load tap-changer. Background Art

[0002] The on-load tap-changer (OLTC) is a core component of a transformer, and its safe and stable operation is crucial for the reliability of the power system. Therefore, it is of great significance to monitor the state and diagnose the faults of the OLTC. Mechanical faults are the main fault types of the OLTC, and mechanical faults cause serious damage to the OLTC and the transformer, thereby affecting the normal operation of power equipment and systems. Therefore, monitoring the mechanical state of the operating OLTC, predicting the possibility of fault occurrence and discriminating the fault type have important value for the reliable operation of the power system. The mechanical vibration signals generated during the switching process of the OLTC contain rich information. By using an acceleration sensor to perform non-destructive detection on its vibration signals, the state information and operation mode of the transmission mechanism are extracted for mechanical fault identification, and automatic diagnosis of faults such as mechanical jamming, poor contact, deformation, wear, and insufficient lubrication is carried out.

[0003] In the existing OLTC vibration signal analysis methods, envelope analysis methods based on Hilbert transform and wavelet decomposition, Fourier analysis, EMD, EEMD, VMD, and FMD all have their respective limitations. For example, the envelope analysis method based on Hilbert transform and wavelet decomposition has advantages in multi-scale signal decomposition, but it is prone to losing information in the low-frequency part, and this shortcoming will affect the accuracy of fault identification. Fourier analysis can effectively analyze the frequency-domain characteristics, but it has poor effects when analyzing the time-varying characteristics of non-stationary signals. This defect limits its application in the fault diagnosis of complex mechanical equipment, especially in OLTC equipment, where frequency-domain analysis is difficult to reveal the instantaneous characteristics of signals. The EMD algorithm decomposes the vibration signals of the entire OLTC switching process and extracts its characteristic energy spectrum. However, since the OLTC mechanical vibration signals contain multiple frequency components, these components are intertwined, resulting in the situation where different frequency components are mixed in the same IMF during EMD decomposition. Since EMD is adaptive, the decomposition result highly depends on the specific characteristics of the signal, resulting in the inability to determine the number of IMFs in advance during each decomposition. Too many IMFs will lead to overfitting, while too few IMFs cannot fully capture all the important characteristics of the signal. The EEMD algorithm alleviates the mode mixing phenomenon in the EMD algorithm, but the EEMD algorithm adds white noise to suppress mode mixing and performs multiple EMD operations, resulting in a large increase in the amount of calculation. The VMD algorithm mainly focuses on the spectral characteristics of the signal and does not explicitly consider fault characteristics such as impact and periodicity. The input parameters of FMD have a relatively obvious impact on its performance and efficiency. Summary of the Invention

[0004] In view of the above problems, the present invention is proposed.

[0005] Therefore, the technical problems solved by the present invention are as follows: the existing OLTC vibration signal analysis methods have defects such as the loss of low-frequency information affecting the accuracy of fault identification, poor effect in analyzing the time-varying characteristics of non-stationary signals, the decomposition result depending on the signal characteristics and making it difficult to determine the number of IMFs, a large increase in the amount of calculation, not clearly considering the fault characteristics, and the input parameters affecting the performance efficiency.

[0006] To solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, an embodiment of the present invention provides a method for extracting multi-source features of the vibration signal of a on-load tap-changer of a transformer, including:

[0008] Optimizing the first parameter of the feature mode decomposition;

[0009] Based on the optimized first parameter, decomposing the vibration signal of the on-load tap-changer switching to obtain the first feature;

[0010] Based on the decomposed vibration signal of the on-load tap-changer switching, performing signal image conversion and image feature extraction to obtain the second feature;

[0011] Fusing the first feature and the second feature to form a high-dimensional feature space matrix; performing dimensionality reduction processing on the high-dimensional matrix to obtain the multi-source features for analysis and identification.

[0012] As a preferred scheme of the method for extracting multi-source features of the vibration signal of a on-load tap-changer of a transformer, wherein:

[0013] The step of decomposing the vibration signal of the on-load tap-changer switching based on the optimized first parameter to obtain the first feature includes:

[0014] Designing an adaptive filter to decompose the input signal, optimizing the target through correlation kurtosis, and screening modes using the correlation coefficient, only retaining the components highly correlated with the fault characteristics and removing the redundant modes;

[0015] Extracting one feature mode in each iteration and removing it from the signal until the signal residual energy is lower than the set threshold or the preset number of modes is reached.

[0016] As a preferred scheme of the method for extracting multi-source features of the vibration signal of a on-load tap-changer of a transformer, wherein:

[0017] The step of decomposing the vibration signal of the on-load tap-changer switching based on the optimized first parameter to obtain the first feature further includes:

[0018] Based on the optimized first parameter, the vibration signal of the OLTC switch is decomposed using empirical mode decomposition to obtain multiple IMF modal components, and set statistical feature parameters are extracted from each IMF.

[0019] As a preferred solution of the multi - feature extraction method for the vibration signal of the on - load tap - changer of a transformer, where:

[0020] Based on the decomposed vibration signal of the on - load tap - changer switching, signal image conversion and image feature extraction are performed to obtain the second feature, including:

[0021] The signal image is encoded, and the second feature is obtained by extracting local extrema of the image, extracting image edge features, and calculating contrast, homogeneity, and fractal dimension.

[0022] As a preferred solution of the multi - feature extraction method for the vibration signal of the on - load tap - changer of a transformer, where:

[0023] The encoding of the signal image includes:

[0024] Using the Gram matrix, the intercepted bearing vibration signal is used as a vector group, and the signal is encoded into a two - dimensional image by expressing the relationship between the vectors themselves and each other; specifically, the input vibration signal is normalized, and then the normalized data is mapped to angular values. The one - dimensional vibration signal sequence is converted into a two - dimensional matrix through the Gram - angle field, and the matrix contains the signal amplitude relationship and the time - series point correlation characteristics.

[0025] As a preferred solution of the multi - feature extraction method for the vibration signal of the on - load tap - changer of a transformer, where:

[0026] The calculation of the contrast, homogeneity, and fractal dimension includes:

[0027] Based on the gray - level co - occurrence matrix, the contrast and homogeneity are calculated to quantitatively describe the local gray - level change and uniformity of the signal two - dimensional image;

[0028] The GAF image is binarized, the optimal binarization threshold is determined using the Otsu algorithm, different - scale grids are divided, the number of non - empty grids is calculated, and the fractal dimension is estimated using a linear fitting method.

[0029] As a preferred solution of the multi - feature extraction method for the vibration signal of the on - load tap - changer of a transformer, where:

[0030] The fusion of the first feature and the second feature to form a high - dimensional feature space matrix includes:

[0031] The first feature and the second feature are feature - stitched and fused into a high - dimensional feature space matrix, where the rows of the matrix represent the number of samples and the columns represent the number of features;

[0032] The dimensionality reduction processing of the high-dimensional matrix includes: performing dimensionality reduction on the high-dimensional matrix using the UMAP algorithm: determining the similarity distribution of sample points and the local scale parameter in the high-dimensional space to control the neighborhood range; constructing a fuzzy neighborhood graph and defining a symmetric fuzzy neighborhood relationship; using a smooth distribution function in the low-dimensional space to define the neighborhood similarity of the points in the dimensionality-reduced space; measuring the difference between the neighborhood distribution in the high-dimensional space and the neighborhood distribution in the low-dimensional space through a cross-entropy loss function, and minimizing the loss function to find the optimal point distribution in the low-dimensional space, so as to obtain multivariate features for analysis and recognition.

[0033] In a second aspect, an embodiment of the present invention provides a multivariate feature extraction system for the vibration signal of a on-load tap-changer of a transformer, including:

[0034] A parameter optimization module for optimizing the first parameter of the feature mode decomposition;

[0035] A first feature acquisition module for decomposing the vibration signal of the on-load tap-changer switching based on the optimized first parameter to obtain a first feature;

[0036] A second feature acquisition module for performing signal image conversion and image feature extraction based on the decomposed vibration signal of the on-load tap-changer switching to obtain a second feature;

[0037] A multivariate feature acquisition module for fusing the first feature and the second feature to form a high-dimensional feature space matrix; performing dimensionality reduction processing on the high-dimensional matrix to obtain multivariate features for analysis and recognition.

[0038] In a third aspect, an embodiment of the present invention provides a computing device, including:

[0039] A memory and a processor;

[0040] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the one or more programs are executed by the one or more processors, the one or more processors implement the method for extracting multivariate features of the vibration signal of the on-load tap-changer of a transformer as described in any embodiment of the present invention.

[0041] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, which stores computer-executable instructions. When the computer-executable instructions are executed by a processor, the method for extracting multivariate features of the vibration signal of the on-load tap-changer of a transformer is implemented.

[0042] Advantages of the present invention: The present invention collects vibration signals during the OLTC switching process and optimizes the FMD decomposition parameters through the Gazelle Optimization Algorithm (GOA) to improve the accuracy and reliability of signal decomposition. After the signal is processed by Feature Mode Decomposition (FMD), it is decomposed into multiple Intrinsic Mode Functions (IMFs). Calculate the statistical features of each IMF component, including mean, variance, peak value, kurtosis, root mean square value, peak factor, impulse factor, waveform factor, and margin factor, to form a high-dimensional feature vector. On this basis, perform Gramian Angular Field (GAF) encoding on the IMF components to generate GAF images, and extract geometric features from them, including edge features, local extrema, contrast, homogeneity, and fractal dimension, to further describe the geometric characteristics of the signal. Finally, through feature fusion and UMAP dimensionality reduction, a feature space matrix is obtained to characterize the complex patterns of vibration signals. The method of the present invention can efficiently extract signal features, improve the accuracy and robustness of OLTC fault diagnosis, has important engineering application value, and can be widely applied to the field of OLTC condition monitoring and fault detection. Description of the Drawings

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0044] Figure 1 is the overall flowchart of the method for extracting multi-dimensional features of the vibration signal of the on-load tap-changer of the transformer described in the present invention;

[0045] Figure 2 is the time-domain diagram of the vibration signal during the OLTC switching process of the method for extracting multi-dimensional features of the vibration signal of the on-load tap-changer of the transformer described in the present invention;

[0046] Figure 3 is the time-domain signal diagram of each IMF component after the vibration signal of the method for extracting multi-dimensional features of the vibration signal of the on-load tap-changer of the transformer described in the present invention is decomposed by FMD;

[0047] Figure 4 is the frequency-domain signal diagram of each IMF component after the vibration signal of the method for extracting multi-dimensional features of the vibration signal of the on-load tap-changer of the transformer described in the present invention is decomposed by FMD;

[0048] Figure 5 is the GAF-encoded image of each IMF component of the method for extracting multi-dimensional features of the vibration signal of the on-load tap-changer of the transformer described in the present invention. Detailed Embodiments

[0049] To make the above objects, features, and advantages of the present invention more apparent and understandable, the following provides a detailed description of the specific embodiments of the present invention in conjunction with the accompanying drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present invention.

[0050] Embodiment 1

[0051] Referring to Figure 1 , which is the first embodiment of the present invention. This embodiment provides a method for extracting multi - feature of the vibration signal of on - load tap - changer of a transformer, including:

[0052] S1: Optimize the first parameter of the feature mode decomposition;

[0053] S2: Based on the optimized first parameter, decompose the vibration signal of the on - load tap - changer switching to obtain the first feature;

[0054] S3: Based on the decomposed vibration signal of the on - load tap - changer switching, perform signal - image conversion and image feature extraction to obtain the second feature;

[0055] S4: Fuse the first feature and the second feature to form a high - dimensional feature space matrix; perform dimensionality reduction on the high - dimensional matrix to obtain multi - features for analysis and recognition.

[0056] It should be noted that through steps S1 - S4, a complete and systematic feature extraction process is constructed. First, the first parameter of the feature mode decomposition is optimized, laying a good foundation for subsequent signal decomposition; then, based on the optimized parameter, the vibration signal of the on - load tap - changer switching is decomposed to obtain the first feature. After that, starting from the decomposed vibration signal, the second feature is obtained through signal - image conversion and image feature extraction operations. Finally, the first feature and the second feature are fused to form a high - dimensional feature space matrix, and dimensionality reduction is performed on it to obtain multi - features that are convenient for analysis and recognition. This method effectively integrates different types of feature information through multi - step processing, reduces the data dimension, improves the availability and pertinence of the features, and helps to analyze and identify the operating state of the on - load tap - changer of the transformer more accurately and efficiently.

[0057] Embodiment 2

[0058] Referring to Figures 1 - 5 , which is an embodiment of the present invention. Based on the previous embodiment, a method for extracting multi - feature of the vibration signal of on - load tap - changer of a transformer is provided, including:

[0059] In the embodiments of the present application, the first parameter in the above step S1 includes filter coefficients, filter length, correlation delay, and the number of modes;

[0060] In another possible implementation, the first parameter may further include filter-related parameters:

[0061] Filter type: Different types of filters (such as Butterworth filters, Chebyshev filters, etc.) have different frequency response characteristics, which will have a significant impact on the signal decomposition quality. For example, Butterworth filters have flat passband characteristics, while Chebyshev filters can achieve steeper attenuation in the passband or stopband.

[0062] Filter cut-off frequency: The cut-off frequency determines the signal frequency range allowed to pass through the filter. Reasonably setting the cut-off frequency can better match the characteristic frequency of the signal and avoid interference from unnecessary frequency components on the decomposition result.

[0063] In another possible implementation, the first parameter may further include signal processing-related parameters:

[0064] Sampling frequency: The sampling frequency affects the degree of signal discretization and plays an important role in subsequent empirical mode decomposition. If the sampling frequency is too low, signal information may be lost; if the sampling frequency is too high, the data processing volume will increase.

[0065] Window function type and length: In signal processing, the window function is used to intercept a part of the signal for analysis. Different types of window functions (such as Hanning window, Hamming window, etc.) have different spectral characteristics, and the length of the window function also affects the time-frequency resolution of the signal, thus affecting the effect of empirical mode decomposition.

[0066] In another possible implementation, the first parameter may further include decomposition result evaluation-related parameters:

[0067] Decomposition termination condition: In addition to using envelope entropy as the fitness function to evaluate the decomposition result, other decomposition termination conditions can also be set, such as the number of modes obtained by decomposition reaching a certain threshold, the correlation between adjacent modes being lower than a certain value, etc.

[0068] Noise threshold: During the decomposition process, a noise threshold can be set. When the noise content in the mode obtained by decomposition is lower than this threshold, the mode is considered a valid characteristic mode; otherwise, it is treated as a noise mode.

[0069] Optimizing the first parameter for empirical mode decomposition includes:

[0070] Optimize the filter coefficient h[k], filter length L, correlation delay τ, and number of modes K in the feature mode decomposition (FMD) using the Gazelle Optimization Algorithm (GOA);

[0071] Specifically, initialize a group of gazelles, where each gazelle represents a combination of a set of FMD parameters;

[0072] For each gazelle, perform FMD decomposition on the signal using the parameter settings it represents, and calculate the fitness function of the decomposition result; use the envelope entropy as the fitness function, and set the fitness function to the minimum value. The envelope entropy represents the sparse characteristics of the original signal. When there is more noise and less feature information in the IMF, the envelope entropy value is larger; conversely, the envelope entropy value is smaller.

[0073] The mathematical definition of the envelope entropy is:

[0074]

[0075] In the formula, a(j) represents the envelope spectrum signal obtained after the mode function is demodulated by Hilbert; p j represents the normalization processing of a(j); m represents the number of signal features.

[0076] It should be noted that the first parameters selected in this embodiment include the filter coefficient h[k], filter length L, correlation delay τ, and number of modes K. These input parameters have a relatively obvious impact on the performance and efficiency of FMD. The filter coefficient h[k] determines the frequency response characteristics of the filter and is the core parameter for controlling the signal decomposition quality. Reasonable h[k] can ensure that the modal frequency range matches the signal characteristics, avoiding modal aliasing or missing important components. Unreasonable design of h[k] will lead to distorted modal extraction and even inability to reflect the actual characteristics of the signal. The filter length L is the length of the finite impulse response (FIR) filter, which controls the time window size of the filter; if L is set too small, low-frequency information will be lost; if L is set too large, high-frequency and short-time impact signals will be lost. The correlation delay τ is the time delay parameter used when calculating the correlation kurtosis CK, which is used to measure the impulsiveness and periodicity of the signal and reflects the specific period of the signal; if τ is set reasonably, the periodic characteristics in the signal can be amplified. If τ is set too large, the correlation of the characteristic signal will be weakened, affecting modal extraction, and too small τ will result in the inability to capture the periodic characteristics. The number of modes K determines how many characteristic modes the signal is decomposed into. If K is set too large, the decomposition result will contain too many noise modes. If K is set too small, the decomposition result cannot cover all important components in the signal.

[0077] The parameters involved in FMD are not independent but interact with each other and jointly determine the decomposition effect. The size of the filter length L affects the selection range of the correlation delay τ. When L is small and the signal impact period is long, it is impossible to accurately match the correlation delay τ. If the number of modes K is large and the filter length L is too small, it will lead to mode frequency overlap. The initial parameter settings of the filter in FMD are all set according to experience, which greatly affects the decomposition effect.

[0078] In the embodiment of the present application, in step S2 above, based on the optimized first parameter, the vibration signal of the on-load tap-changer switching is decomposed, and the first features obtained include:

[0079] As Figure 2 shown, collect the vibration signal during the OLTC switching process;

[0080] Design an adaptive filter; use a finite impulse response FIR filter to decompose the signal;

[0081] Specifically, for the input signal x[n], the output signal y[n] of the filter is expressed as:

[0082]

[0083] Optimize the objective through the correlation kurtosis CK; the calculation formula of the correlation kurtosis CK is:

[0084]

[0085] where E is the expectation operator.

[0086] Screen the modes through the correlation coefficient CC, and only retain the components highly correlated with the fault characteristics;

[0087] The calculation formula of the correlation coefficient CC is:

[0088]

[0089] If CC exceeds the set threshold, retain the mode; otherwise, it is regarded as a redundant mode and eliminated.

[0090] In each iteration, extract a characteristic mode and remove it from the signal until the signal residual energy is lower than the set threshold or the preset number of modes is reached. After extracting the i-th characteristic mode, the residual signal is updated by the following formula:

[0091] r[n] = x[n] - y i [n]

[0092] where y i [n] represents the extracted characteristic mode.

[0093] Based on the optimized first parameter, calculate the statistical features of each IMF component: Bring the optimized first parameter back into FMD to decompose the vibration signal of OLTC switch switching, as Figure 3 , Figure 4 shown. Finally, K IMF modal components are obtained, which is the first feature.

[0094] Specifically, the present invention uses FMD to decompose the vibration signal and extracts 9 feature parameters from each IMF. These feature parameters include Mean, Variance, Peak Value, Kurtosis, RMS, Peak Factor, Impulse Factor, Shape Factor, and Margin Factor. Each feature can reflect different dynamic characteristics of the signal and reveal potential fault modes in the vibration signal. By calculating these time-domain features, a feature vector of each sample is constructed to provide input data for subsequent fault classification.

[0095] The mean Mean represents the average value of a data set or signal, which is the sum of all data points divided by the number of data points. For a discrete signal x1, x2, …, x n , the mean is expressed as:

[0096]

[0097] The mean of an ideal fault-free vibration signal should be close to zero. If the mean deviates from zero, it indicates the existence of some persistent abnormality, such as imbalance, misalignment, or eccentricity.

[0098] The variance Variance is used to reflect the degree of fluctuation of a data set, indicating the degree of deviation of data points from the mean, which is the average of the squares of the deviations of each data point from the mean. For a discrete signal x1, x2, …, x n , the variance is expressed as:

[0099]

[0100] where μ is the mean. If the variance of the signal is large, it indicates that the vibration changes violently; if the variance is small, the signal changes smoothly. In mechanical faults or impact events, the peak value of the vibration often increases significantly.

[0101] The peak Peak Value is the difference between the maximum and minimum values in the signal, referring to the maximum amplitude of the signal. The peak is expressed as:

[0102] PeakValue = max(|x i |)

[0103] where, |xi | is the absolute value of the signal.

[0104] Kurtosis is a statistic with a well-defined formula used to describe the peakedness of a signal and measure the frequency of extreme values or outliers in the data distribution. For a discrete signal x1, x2, …, x n , kurtosis is expressed as:

[0105]

[0106] For a normal distribution, the kurtosis is 3. When the kurtosis is greater than 3, it indicates that the data distribution has more extreme values and is called a peaked distribution; when it is less than 3, it indicates a relatively flat distribution. A high kurtosis indicates that the signal has more extreme values and is related to sudden events such as impacts and slider jams.

[0107] The Root Mean Square (RMS) value is a measure of the effective energy or power of a signal. For a discrete signal x1, x2, …, x n , the RMS value is expressed as:

[0108]

[0109] The larger the RMS value, the greater the vibration energy, which means that the equipment has a heavier load or a malfunction.

[0110] The Peak Factor is the ratio between the peak value of a signal and its RMS value. The Peak Factor is expressed as:

[0111]

[0112] It reflects the extremity of the signal. A larger Peak Factor means that there are sudden events in the signal, such as impacts or collisions.

[0113] The Impulse Factor is the ratio between the maximum amplitude in a signal and the RMS value of the signal, used to measure the impulsiveness of the signal. The Impulse Factor is expressed as:

[0114]

[0115] For impact faults in mechanical equipment, such as gear impacts or bearing pitting, the Impulse Factor will increase significantly.

[0116] The Shape Factor is the ratio between the RMS value of a signal and the average absolute value. The Shape Factor is expressed as:

[0117]

[0118] It reflects the waveform characteristics of the signal, is used to measure the symmetry of the signal or whether there is a large offset, and is used to distinguish periodic faults and aperiodic faults. Periodic faults such as gear meshing faults.

[0119] The margin factor is the ratio of the maximum amplitude to the effective value in the signal, reflects whether the signal is close to its maximum amplitude, and is used to describe the stability of the system or signal. This index describes whether the signal is close to its limit operating state. If the margin factor is large, it indicates that there are large peaks in the signal, indicating the existence of sudden events or abnormal conditions.

[0120] In another possible implementation, appropriate wavelet bases (such as db wavelets, sym wavelets, etc.) and the number of decomposition levels can also be selected, and then the vibration signal is wavelet decomposed to obtain a series of wavelet coefficients at different scales. According to actual needs, the wavelet coefficients related to the fault characteristics can be selected to reconstruct the signal, thereby obtaining the first feature.

[0121] It should be noted that in this step, the vibration signal of the on-load tap changer switching is decomposed based on the optimized first parameter. An adaptive filter and a finite impulse response FIR filter are used, and the relevant kurtosis CK optimization target and the correlation coefficient CC are combined to screen the modes, which can accurately extract the components highly related to the fault characteristics and effectively eliminate redundant modes, thereby obtaining the first feature that accurately reflects the potential fault mode of the vibration signal; 9 characteristic parameters are extracted from each IMF component, including mean value, variance, peak value, etc. These characteristics can reflect the dynamic characteristics of the signal from different angles. For example, the mean value can detect persistent anomalies, the variance reflects the degree of fluctuation, and the kurtosis reflects the frequency of extreme values, etc., providing rich and comprehensive input data for subsequent fault classification and helping to analyze the fault type and degree more deeply.

[0122] In the embodiment of the present application, the signal image conversion in the above step S3 includes:

[0123] Perform Gramian angular field image coding on the signal image;

[0124] It should be noted that the basic principle of the Gramian angular field image coding method is: Use the Gram matrix Gramin to take a part of the bearing vibration signal as a group of vectors, and encode this part of the vibration signal into a two-dimensional image by expressing the characteristics of each vector itself and the relationship with other vectors. This part of the signal is encoded in the polar coordinate system, retaining the local time-related information, and is re-calibrated with the data through trigonometric functions, such as Figure 5 as shown, to generate a unique feature map. The specific steps are as follows:

[0125] Perform normalization:

[0126] Mark the input vibration signal as X = {x1, x2,..., xn}, x n It is indicated that there are n data points in total, and the normalized form of each data point is as follows:

[0127]

[0128] Among them, min(X) represents the minimum value of the signal amplitude, and max(X) represents the maximum value of the signal amplitude. Through normalization, the data is scaled into [-1, 1].

[0129] Map the normalized data value to an angle And represent the sequence in the polar coordinate system as:

[0130]

[0131] Among them, is the data point in the normalized form; is the vibration signal in the normalized form. Through this method, the normalized data is converted into angle information to obtain a set of angle values

[0132] After being mapped to angle values, the one-dimensional vibration signal sequence is converted into a two-dimensional matrix G through the Gram angle field:

[0133]

[0134] Through the above calculation, the original one-dimensional vibration signal is mapped into a two-dimensional matrix of N×N. The elements in the matrix include not only the amplitude relationship of the signal amplitude but also the correlation characteristics between time sequence points.

[0135] In the embodiment of the present application, the image feature extraction in the above step S3 includes:

[0136] Extract the local extreme values of the Gram angle field image of the signal image;

[0137] It should be noted that the local extreme value means that the value of the signal at a certain point is larger than all the local maximum values of its neighborhood values or smaller than all the local minimum values of its neighborhood values. The physical meaning of local extreme value extraction is to capture the key features reflecting the mechanical state change in the signal. During the normal operation of the OLTC, the vibration signal usually shows stable periodic characteristics, and the distribution of extreme values in its GAF image is regular and uniform. Mechanical faults of the OLTC will introduce non-stationary signals or mutation signals, resulting in significant changes in the distribution of local extreme values: periodic faults will make the distribution of extreme values show periodicity and enhancement, and irregular extreme points will appear in the GAF image of non-periodic faults.

[0138] In a two-dimensional image, such as a GAF image, local extrema are detected through neighborhood search. If GAF[i,j] > GAF[i±1,j±1], then GAF[i,j] is a local maximum. If GAF[i,j] < GAF[i±1,j±1], then GAF[i,j] is a local minimum.

[0139] Perform Gramian Angular Field (GAF) image edge feature extraction on the signal image;

[0140] It should be noted that edge feature extraction can extract key texture change regions from the GAF image, reflecting the dynamic change characteristics of the vibration signal. Edge features measure regions in the image where the gray value changes significantly. After Gaussian filtering and smoothing processing through the Canny algorithm, edge detection is performed on the regions with significant amplitude mutations in the signal, thereby providing important features for the recognition and classification of fault patterns. The specific steps are as follows:

[0141] Perform smoothing processing on the input image I(x,y) using Gaussian filtering, expressed as:

[0142]

[0143] I′(x,y) = I(x,y) * G(x,y)

[0144] Where (x,y) represents the coordinates of a certain pixel in the filter; σ represents the standard deviation of the Gaussian kernel, which can control the degree of smoothing; * represents the convolution operation. Gaussian filtering is used to smooth the GAF image, removing noise and fine interference in the signal while retaining the main features of the signal.

[0145] Use the Sobel operator to calculate the gradient intensity G(x,y) and gradient direction θ of the image to determine the region with the largest change in gray value in the image, expressed as:

[0146] G x = I′(x,y) * S x

[0147] G y = I′(x,y) * S y

[0148] Where:

[0149]

[0150] The calculation formulas for the gradient intensity G(x,y) and gradient direction θ are:

[0151]

[0152] OLTC faults, such as stuck contacts and loose drive gears, usually lead to sudden changes in the amplitude of vibration signals, and these sudden changes appear as edge features in GAF images.

[0153] Only retain the pixel points with the maximum value along the gradient direction θ, and suppress non-edge points. Check the gradient intensity G(x,y) of each pixel point. Compare it with its two adjacent points before and after along the gradient direction θ. If the current point is not a local maximum, set its intensity to 0. In this way, eliminate the points that do not conform to the edge characteristics and ensure the accuracy of the extracted edges.

[0154] Set a high threshold T high and a low threshold T low to judge the gradient intensity, so as to distinguish strong edges, weak edges, and non-edges. If G(x,y) > T high then this pixel is a strong edge and is directly retained; if T low ≤ G(x,y) ≤ T high then this pixel is a weak edge and needs to be connected to the strong edge, otherwise it is discarded; if G(x,y) < T low then this pixel is a non-edge and is directly discarded. Non-maximum suppression and double-threshold detection are used to accurately locate edge points and remove the influence of redundant weak edges and noise.

[0155] It should be noted that OLTC faults often cause sudden changes in the amplitude of vibration signals, and these sudden changes appear as regions with significant changes in gray values in GAF images. Edge feature extraction can accurately capture these changes. Different types of OLTC faults will show different edge features in GAF images: if the contacts are stuck, the edges are densely distributed and the gradient directions are uneven; if the gears are loose, the edges are sparsely distributed and the gradient intensity is weak; if the contacts are ablated, the edges are concentrated in specific regions.

[0156] Obtaining the second feature also includes calculating the contrast and homogeneity based on the gray-level co-occurrence matrix;

[0157] It should be noted that contrast and homogeneity are image texture features extracted based on the gray-level co-occurrence matrix (GLCM), which are used to quantitatively describe the local gray-level change characteristics and uniformity of signals in two-dimensional image representations. The higher the contrast and homogeneity, the greater the gray-level changes in the signal, usually corresponding to mutation points or energy concentration regions in mechanical vibration signals. Contrast reflects the change intensity of gray-level values in the GAF image and can quantify the degree of change between different time points of the signal. In a normal vibration signal, the gray-level change in the GAF image is relatively stable and the contrast is low. OLTC faults will cause sudden changes in the signal amplitude, which are manifested as regions with significant gray-level differences in the GAF image, thereby increasing the contrast. The more severe the fault, the more obvious the signal mutation, the greater the difference in gray-level values in the GAF image, and the contrast increases accordingly, which can be used as a quantitative index to evaluate the severity of the fault. Homogeneity measures the local uniformity degree of gray-level values in the GAF image. The higher the homogeneity value, the more uniform the gray-level distribution and the gentler the change; the lower the homogeneity value, the more uneven the gray-level changes in the image, and there may be mutations or complex textures, that is, corresponding to mutations and impacts caused by faults.

[0158] Specifically, the calculation steps for contrast and homogeneity are as follows:

[0159] Normalize the values of the G matrix to the range of [0, 1] and convert them into gray-level values:

[0160]

[0161] In the formula, the round[] function rounds the input value to the nearest integer; β is the number of target gray levels, and β takes the value of 256. G(x, y) represents the gradient value of the image at the point (x, y); max(G) represents the maximum gradient value in the G matrix, which is the strongest edge or texture region in the image; min(G) represents the minimum gradient value in the G matrix, which is the flat region in the image with almost no edges or changes.

[0162] Generate the gray-level co-occurrence matrix:

[0163]

[0164] In the formula, G norm (x + Δx, y + Δy) represents the gray-level value of the adjacent pixel to the current pixel in a specific direction. The definition of the indicator function δ() is as follows:

[0165]

[0166] The offset direction selects the main diagonal direction of 45°, i.e., (Δx, Δy) = (-1, 1), which is applicable to the analysis of periodic signals. x and y represent positions on the coordinate axes and are spatial position parameters; Δx and Δy are the offsets in the coordinate axis directions; p and q represent the relative positions between adjacent pixel sample points in the 45° direction of the main diagonal; i and j are the row and column indices of the matrix, used to represent different positions of the data; G norm is the result obtained by normalizing each element in the G matrix.

[0167] Calculate the contrast through the gray-level co-occurrence matrix:

[0168]

[0169] Calculate the homogeneity:

[0170]

[0171] where P(i, j) is the similarity between two samples.

[0172] Obtaining the second feature also includes calculating the fractal dimension through the Box-counting algorithm;

[0173] Binarize the GAF image and convert G(x, y) into a binary matrix B(x, y):

[0174]

[0175] In the formula, T is the binarization threshold.

[0176] Use Otsu to maximize the between-class variance to determine the optimal binarization threshold T:

[0177]

[0178]

[0179] where ω1(T) is the proportion of foreground pixels segmented by the binarization threshold T; ω2(T) is the proportion of background pixels segmented by the binarization threshold T; P(i) represents the probability that a pixel with a gray value of i appears; μ1(T) is the average gray value of the foreground; μ2(T) is the average gray value of the background.

[0180] Assume the image size is M×N, the grid side length is ε, the image is divided into ε×ε grids, and the number of grids in the row direction is:

[0181] N x = M / ε

[0182] The number of grids in the column direction is:

[0183] N y = N / ε

[0184] For each network (i, j), determine whether it is a non-empty network. The grid coverage area of the i-th row and j-th column is:

[0185] R ij = {B(x, y) | (i - 1)ε < x ≤ iε, (j - 1)ε < y ≤ jε}

[0186] If at least one pixel value in the region R ij is 1, then the network is a non-empty network. Furthermore, calculate the number of non-empty networks:

[0187]

[0188] where the definition of the indicator function is:

[0189]

[0190] Select multiple scales ε, such as ε = 2, 4, 8,.... For each scale ε, calculate the corresponding N(ε).

[0191] For each scale ε, calculate log N(ε) and Estimate the fractal dimension D using the linear fitting method:

[0192]

[0193] where K is the number of different grid scales, ε k is the k-th grid scale, and are the means.

[0194] In summary, through the extraction of local extrema of the Gram angular field image, the extraction of edge features of the Gram angular field image, and the calculation of contrast, homogeneity, and fractal dimension, the second feature is obtained.

[0195] In another possible implementation, the second feature can also be extracted based on the fast Fourier transform (FFT);

[0196] Specifically, step 1: Prepare the vibration signal

[0197] Ensure that the vibration signal of the on-load tap-changer switching is a discrete time series and the sampling frequency is known.

[0198] Step 2: Perform FFT transformation

[0199] Use the FFT function in a programming language (such as Python) to transform the vibration signal. In Python, the numpy.fft.fft function can be used.

[0200] Step 3: Calculate the spectral amplitude

[0201] Since the FFT result is a complex number, its amplitude needs to be calculated to obtain the spectrum.

[0202] Step 4: Calculate the frequency axis

[0203] Calculate the frequency axis according to the sampling frequency and the signal length.

[0204] Step 5: Extract features

[0205] Extract features from the spectrum, such as the main frequency, the amplitudes of each frequency component, etc. For example, find the frequency with the largest spectral amplitude as the main frequency;

[0206] The extracted main frequency, the amplitudes of each frequency component, etc. are the second features.

[0207] It should be noted that in this step, the one-dimensional vibration signal is transformed into a two-dimensional image through Gramian Angular Field (GAF) image coding, which not only retains the local time-related information, but also can reflect the signal amplitude relationship and the correlation characteristics of time series points through the matrix elements, presenting the complex signal information in an intuitive image form for subsequent analysis; it can capture the key features reflecting the mechanical state changes. When operating normally, the extreme value distribution of the GAF image of the vibration signal is regular and uniform, and the distribution will change significantly when a fault occurs, and periodic and aperiodic faults can be identified based on this; by using the Canny algorithm combined with steps such as Gaussian filtering, Sobel operator, non-maximum suppression, and double-threshold detection, the edge features reflecting the amplitude mutation of the vibration signal in the GAF image can be accurately extracted. Different types of on-load tap-changer (OLTC) faults show different edge features in the GAF image, which helps in the identification and classification of fault modes; calculating the contrast and homogeneity based on the gray-level co-occurrence matrix can quantitatively describe the local gray-level change characteristics and uniformity of the signal in the two-dimensional image representation. The contrast can evaluate the severity of the fault, and the homogeneity can reflect the mutation and impact caused by the fault; calculating the fractal dimension through the Box-counting algorithm further explores the feature information of the GAF image, enriching the extraction of the on-load tap-changer fault features from different angles and improving the accuracy of fault diagnosis; the second features obtained by comprehensively calculating the local extreme value extraction, edge feature extraction, contrast and homogeneity calculation, and fractal dimension calculation of the GAF image can reflect the operating state of the on-load tap-changer from multiple dimensions, providing a comprehensive and accurate basis for the fault diagnosis of the on-load tap-changer.

[0208] In the embodiment of the present application, in step S4 above, fusing the first feature and the second feature to form a high-dimensional feature space matrix includes:

[0209] Perform feature splicing on the first feature and the second feature, and fuse them into a high-dimensional feature space matrix:

[0210]

[0211] Among them, the row represents the number of samples N, and the column represents the number of features M.

[0212] In the embodiment of the present application, the dimensionality reduction process of the high-dimensional matrix in the above step S4 to obtain multivariate features for analysis and recognition includes:

[0213] Using UMAP for dimensionality reduction, the specific steps include:

[0214] For the sample point x in the high-dimensional space i and x j , the similarity distribution is:

[0215]

[0216] σ i is the local scale parameter of x i , used to control the neighborhood range, and its determination method is:

[0217]

[0218] Among them, k is the number of nearest neighbors; x ik represents the value of the sample point x i on the k-th feature dimension; x jk represents the value of the sample point x j on the k-th feature dimension; d represents the dimension of the feature space.

[0219] Construct a fuzzy neighborhood graph. UMAP defines a symmetric fuzzy neighborhood relationship by combining the similarity p ij from point i to j and p ji :

[0220] P ij = p ij + p ji - p ij · p ji

[0221] In the low-dimensional space, UMAP uses a smooth distribution function to define the neighborhood similarity of the points y i and y j after dimensionality reduction:

[0222]

[0223] Among them, a and b are hyperparameters used to control the distribution shape. y ik represents the value of the sample point y i on the k-th feature dimension; y jk represents the value of the sample point yj The value on the k-th feature dimension.

[0224] The ultimate goal of UMAP is to make the neighborhood distribution P in the high-dimensional space ij and the neighborhood distribution Q in the low-dimensional space ij as consistent as possible, and the difference between the two is measured by the cross-entropy loss function:

[0225]

[0226] where P ij log(Q ij ) encourages the points that are close in the high-dimensional space to remain close in the low-dimensional space; (1 - P ij )log(1 - Q ij ) encourages the points that are far apart in the high-dimensional space to also remain separated in the low-dimensional space. By minimizing L, UMAP finds the optimal point distribution in the low-dimensional space, making the low-dimensional space as faithful as possible to the structure of the high-dimensional space.

[0227] Embodiment 3

[0228] The above is a schematic solution of the method for extracting multi-source features of the on-load tap-changer vibration signal of the transformer in this embodiment. It should be noted that the technical solution of the system for extracting multi-source features of the on-load tap-changer vibration signal of the transformer belongs to the same concept as the technical solution of the above method for extracting multi-source features of the on-load tap-changer vibration signal of the transformer. For the details not described in detail in the technical solution of the system for extracting multi-source features of the on-load tap-changer vibration signal of the transformer in this embodiment, reference can be made to the description of the technical solution of the above method for extracting multi-source features of the on-load tap-changer vibration signal of the transformer.

[0229] This embodiment also provides a system based on the method for extracting multi-source features of the on-load tap-changer vibration signal of the transformer, including:

[0230] A parameter optimization module for optimizing the first parameter of the feature mode decomposition;

[0231] A first feature acquisition module for decomposing the vibration signal of the on-load tap-changer switching based on the optimized first parameter to obtain the first feature;

[0232] A second feature acquisition module for performing signal image conversion and image feature extraction on the decomposed vibration signal of the on-load tap-changer switching to obtain the second feature;

[0233] A multi-source feature acquisition module for fusing the first feature and the second feature to form a high-dimensional feature space matrix; performing dimensionality reduction processing on the high-dimensional matrix to obtain the multi-source features for analysis and recognition.

[0234] This embodiment also provides a computing device, which is applicable to the case of the multi-feature extraction method for the vibration signal of on-load tap-changer of transformer, and includes:

[0235] A memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the multi-feature extraction method for the vibration signal of on-load tap-changer of transformer as proposed in the above embodiment.

[0236] This embodiment also provides a storage medium, on which a computer program is stored, and when the program is executed by a processor, it implements the multi-feature extraction method for the vibration signal of on-load tap-changer of transformer as proposed in the above embodiment.

[0237] The storage medium proposed in this embodiment and the multi-feature extraction method for the vibration signal of on-load tap-changer of transformer proposed in the above embodiment belong to the same inventive concept. For the technical details not described in detail in this embodiment, reference can be made to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0238] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A method for extracting multi - feature of on - load tap - changer vibration signal of a transformer, characterized in that Including: Optimizing the first parameter of the characteristic mode decomposition; Based on the optimized first parameter, decomposing the vibration signal of the on-load tap-changer switching to obtain the first feature; Based on the decomposed vibration signal of the on-load tap-changer switching, performing signal image conversion and image feature extraction to obtain the second feature; Fusing the first feature and the second feature to form a high-dimensional feature space matrix; Performing dimensionality reduction processing on the high-dimensional matrix to obtain multivariate features for analysis and recognition.

2. The method for extracting multi - feature of on - load tap - changer vibration signal of a transformer according to claim 1, wherein The decomposing the vibration signal of the on-load tap-changer switching based on the optimized first parameter to obtain the first feature includes: Designing an adaptive filter to decompose the input signal, optimizing the target through correlation kurtosis, and using the correlation coefficient to screen modes, only retaining the components highly correlated with the fault features and removing the redundant modes; Extracting one characteristic mode in each iteration and removing it from the signal until the signal residual energy is lower than the set threshold or reaches the preset mode number.

3. The method for extracting multi - feature of on - load tap - changing switch vibration signal of a transformer according to claim 2, wherein, The decomposing the vibration signal of the on-load tap-changer switching based on the optimized first parameter to obtain the first feature further includes: Based on the optimized first parameter, using characteristic mode decomposition to decompose the vibration signal of the OLTC switch switching to obtain multiple IMF mode components, and extracting the set statistical feature parameters from each IMF.

4. The method for extracting multi - feature of on - load tap - changer vibration signal of a transformer according to claim 3, wherein, The performing signal image conversion and image feature extraction based on the decomposed vibration signal of the on-load tap-changer switching to obtain the second feature includes: Performing image coding on the signal image, and obtaining the second feature through image local extreme value extraction, image edge feature extraction, and calculation of contrast, homogeneity, and fractal dimension.

5. The method for extracting multi - feature of vibration signal of on - load tap - changer of transformer according to claim 4, wherein, The performing image coding on the signal image includes: Using the Gram matrix to take the intercepted bearing vibration signal as a vector group, and encoding the signal into a two-dimensional image by expressing the relationship between the vector itself and each other; specifically, normalizing the input vibration signal, then mapping the normalized data to angular values, and converting the one-dimensional vibration signal sequence into a two-dimensional matrix through the Gram angular field, and the matrix contains the signal amplitude relationship and the time series point correlation characteristics.

6. The method for extracting multi - feature of on - load tap - changer vibration signal of transformer according to claim 5, characterized in that, The calculation of the contrast, homogeneity, and fractal dimension includes: Calculating the contrast and homogeneity based on the gray-level co-occurrence matrix to quantitatively describe the local gray change and uniformity of the signal two-dimensional image; Binarizing the GAF image, determining the optimal binarization threshold using the Otsu algorithm, dividing different scale grids, calculating the number of non-empty grids, and estimating the fractal dimension using the linear fitting method.

7. The method for extracting multi - feature of on - load tap - changer vibration signal of transformer according to claim 6, characterized in that The fusing the first feature and the second feature to form a high-dimensional feature space matrix includes: Performing feature splicing on the first feature and the second feature to fuse them into a high-dimensional feature space matrix, where the rows of the matrix represent the number of samples and the columns represent the number of features; The dimensionality reduction processing of the high-dimensional matrix includes: using the UMAP algorithm to perform dimensionality reduction on the high-dimensional matrix: determining the similarity distribution of sample points and the local scale parameter in the high-dimensional space to control the neighborhood range; constructing a fuzzy neighborhood graph and defining a symmetric fuzzy neighborhood relationship; using a smooth distribution function in the low-dimensional space to define the neighborhood similarity of the points in the dimensionality-reduced space; measuring the difference between the neighborhood distribution in the high-dimensional space and the neighborhood distribution in the low-dimensional space through a cross-entropy loss function, and minimizing the loss function to find the optimal point distribution in the low-dimensional space, so as to obtain multivariate features for analysis and recognition.

8. A multi - feature extraction system for on - load tap - changer vibration signals of a transformer, which applies the multi - feature extraction method for on - load tap - changer vibration signals of a transformer as described in any one of claims 1 to 7, is characterized in that, including: a parameter optimization module for optimizing the first parameter of the feature mode decomposition; a first feature acquisition module for decomposing the vibration signal of the on-load tap-changer switching based on the optimized first parameter to obtain a first feature; a second feature acquisition module for performing signal image conversion and image feature extraction based on the decomposed vibration signal of the on-load tap-changer switching to obtain a second feature; a multivariate feature acquisition module for fusing the first feature and the second feature to form a high-dimensional feature space matrix; performing dimensionality reduction processing on the high-dimensional matrix to obtain multivariate features for analysis and recognition.

9. A computing device, including: a memory and a processor; the memory is used for storing computer-executable instructions, and the processor is used for executing the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the method for extracting multivariate features of the on-load tap-changer vibration signal of the transformer according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing computer-executable instructions, and when the computer-executable instructions are executed by a processor, the steps of the method for extracting multivariate features of the on-load tap-changer vibration signal of the transformer according to any one of claims 1 to 7 are implemented.

Images