Partial discharge signal denoising method

Abstract

The application discloses a partial discharge signal denoising method, comprising the following steps: building a partial discharge signal detection platform and a partial signal collection platform; measuring and detecting characteristic signals in the partial discharge signals; screening and classifying the measured signals through a Gaussian test to obtain the characteristic signals; processing and denoising the signals through a fast S transform algorithm and a singular value decomposition SVD algorithm; testing and diagnosing the denoising effect by using two coefficients of noise suppression ratio and amplitude attenuation ratio; and displaying the diagnostic analysis result on a platform display interface. The application discloses a partial discharge signal diagnostic system based on a fast S transform and an adaptive singular value, and solves the problems that the existing technical methods are difficult to balance the denoising effect and extraction speed in the characteristic extraction of the partial discharge signals and are difficult to extract the characteristics.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electric signal noise reduction, and particularly relates to a partial discharge signal noise reduction method. BACKGROUND

[0002] Partial discharge (PD) refers to a discharge form occurring between electrodes but not penetrating the electrodes, a weak point existing in the interior of device insulation or a defect caused by a production process, and a phenomenon of repeated breakdown and extinction under the action of high electric field intensity.

[0003] Common PD signal noise reduction methods include a wavelet noise reduction method, which has improved noise reduction effect compared with traditional noise reduction algorithms, but it is difficult to select a suitable wavelet basis function and decomposition layer number for wavelet transform; and an empirical mode decomposition (EMD) for PD signal noise reduction, which has a certain adaptivity, but the algorithm itself adds new white noise to the processed signal, resulting in incomplete white noise filtering and mode aliasing. SUMMARY

[0004] The present application aims to at least partially solve one of the technical problems in the related art. To this end, one object of the present application is to provide a partial discharge signal noise reduction method capable of improving the noise reduction processing effect of partial discharge noise and improving the noise reduction time.

[0005] According to the partial discharge signal noise reduction method provided by the present application, the method steps are as follows:

[0006] S1: Build a partial discharge signal noise reduction experimental platform;

[0007] S2: Complete simulation of the partial discharge signal by using a double exponential oscillation decay model and a single exponential oscillation decay model, set multiple groups of partial discharge noise-containing signals as experimental groups, and each group of partial discharge noise-containing signals is different;

[0008] S3: Train each experimental group according to a predetermined number of control groups, determine the kurtosis test, set the confidence interval of the Gaussian test, improve the Gaussian test preprocessing, perform Gaussian test preprocessing on the noise-containing partial discharge signal of each experimental group, and obtain the frequency band range of the characteristic signal;

[0009] S4: After preprocessing, the signal is subjected to fast S transform, and a noise reduction algorithm combining time-frequency spectrum analysis filtering and threshold noise reduction is used to remove most of the periodic narrowband interference signals in the signal;

[0010] S5: After fast S transform, the signal is subjected to a fitting threshold approximation singular value noise reduction algorithm, analyzes the two-dimensional modal time-frequency matrix of the signal, and filters out the Gaussian white noise in the signal;

[0011] S6: Perform noise suppression ratio and amplitude attenuation coefficient test on the noise-reduced signal, analyze the test results, and display the test results on the display interface.

[0012] Preferably, the method steps of Gaussian test preprocessing of the noisy partial discharge signal in step S3 are as follows:

[0013] Given the random variable X with mean μ and variance σ, Gaussian test is performed on the random variable X, and according to Chebyshev inequality, the distribution interval of the random variable X satisfies the probability inequality:

[0014]

[0015] In the above formula: q = 1-(σ 2 / ε 2 ), ε>0, assuming that there is a set of observation samples x(t) with mean 0 and number N in the random variable X conforming to Gaussian distribution, when σ 2 / ε 2 is small enough, the estimation of kurtosis value and variance of the random variable X can be represented as:

[0016]

[0017]

[0018]

[0019] wherein, is the mean of the kurtosis value, the variance of the kurtosis value, is the kurtosis value estimate of the random variable sequence x(t), and the kurtosis coefficient Early failure of the reaction system is determined by the number of observation points N and the confidence level q, if the estimated kurtosis value of the variable sequence x(t) is in the interval then it indicates that the measured signal has obvious Gaussian characteristics and needs to be excluded.

[0020] Preferably, in the preprocessing stage, the frequency range of the useful PD signal is obtained, the envelope extremum algorithm is used to set the characteristic frequency reservation mechanism and the non-characteristic signal elimination mechanism, and the envelope spectrum is used to screen out the characteristic frequency for separate processing;

[0021] Because the fast S transform is reversible and linear, a time-frequency filter is designed according to the time-frequency signal after the transform, the data in the time range t∈[t1,t2] and the frequency range f∈[f1,f2] are reserved, and the data in the time range t∈[t3,t4] and the frequency range f∈[f3,f4] are suppressed, the fast S transform is discretized, and f = n / NT and τ = jT:

[0022]

[0023]

[0024]

[0025] Where f is the frequency, t and τ are the time, w(t-τ,f) is the Gaussian window function, λ and p are adjustment factors, with λ=0.55 and p=0.8, S[jT,n / NT] is the fast S-transform in the discrete case, m and n are constants, T is the sampling time interval, N is the total number of sampling points, and X[(m+n) / NT] is the discrete Fourier transform of the signal x(nT);

[0026] In step S4, the preprocessed signal undergoes a fast S-transform. The steps of the fast S-transform are as follows:

[0027] S41: Based on the preprocessed envelope spectrum, useful frequency points k are selected through Gaussian testing. i , i = 1, 2, 3, ..., n;

[0028] S42: For the selected frequency band, the window function W′ for performing a Fourier transform is:

[0029]

[0030] S43: Perform an inverse Fourier transform to obtain point k. i Fast S-transform;

[0031] S44: Repeat step S43 until the fast S-transform is completed at all frequency points.

[0032] Preferably, the method for analyzing the two-dimensional modulus time-frequency matrix of the signal and filtering out Gaussian white noise in step S5 is as follows:

[0033] S51: Select the first n non-zero singular values ​​of the singular value matrix Λ obtained by singular value decomposition and transform it into a singular value vector;

[0034] S52: Select m data points from the singular value vector and divide them into n / m segments to be processed. Based on the property that singular values ​​appear in pairs, the number of m must be even.

[0035] S53: Calculate the regression coefficient of the treatment interval, determine the point with the largest second-order difference coefficient, select the two adjacent intervals of that point, and perform curve fitting.

[0036] S54: Determine the intersection point of the curve fitting, and at the same time, select data slightly smaller than the intersection point in the singular value vector as the threshold of the singular value denoising algorithm.

[0037] Preferably, in order to detect the filtering effect in step S6, the noise suppression ratio p is introduced NRR and the amplitude attenuation ratio p ARR Two coefficients are used to comprehensively evaluate the denoising performance:

[0038]

[0039]

[0040] In the formula, s1 is the standard deviation of the signal before denoising, s2 is the standard deviation of the signal after denoising, A m1 is the maximum amplitude of the signal before denoising, A m2 is the maximum amplitude of the signal after denoising.

[0041] The beneficial effects in the present application are:

[0042] (1) The present application proposes a combination method of pre-processing and fast S transform based on time-frequency spectrum analysis filtering method and threshold denoising method of Gaussian test, which can effectively identify whether the measured signal has research significance, and the fast S transform is an improvement of the algorithm operation efficiency of the generalized S transform. Through pre-processing of the signal, the frequency range of the PD signal is obtained, and then the fast S transform is performed according to the frequency range, which can realize the fast optimization of the algorithm, solve the problem of low feature extraction efficiency when the data amount of partial discharge signal is too large, and is suitable for noise feature extraction in transformer partial discharge signal. SVD is a matrix decomposition method for transforming a matrix into its components, which decomposes the vector space of the noisy signal into a "signal subspace" and a "noise subspace", and uses orthogonal matrix decomposition technology in linear algebra to simplify matrix calculation.

[0043] (2) The traditional singular value algorithm determines the threshold value of the singular value by artificial selection or empirical prediction method, which can cause large error. If cubic spline interpolation is used to determine the threshold value of the singular value, correct singular value threshold can be selected when the data is less and the noise is simple. When the data points are too many and the noise is complex, the threshold selection effect is not good, and the fitting threshold improves the denoising method of singular value decomposition to solve the problem of non-adaptive threshold.

[0044] (3) Because pure partial discharge signal cannot be measured in actual test process, the present application introduces noise suppression ratio and amplitude attenuation ratio two coefficients to comprehensively evaluate the denoising performance. BRIEF DESCRIPTION OF DRAWINGS

[0045] In the drawings:

[0046] Figure 1 is the Gaussian test detection diagram of the partial discharge signal proposed by the present application;

[0047] Figure 2 Flow chart of the experimental platform proposed in the present application;

[0048] Figure 3 Partial discharge waveform chart proposed in the present application;

[0049] Figure 4 Fast S transform and S transform comparison chart proposed in the present application;

[0050] Figure 5 Time-frequency analysis waveform chart proposed in the present application;

[0051] Figure 6 Singular value curve chart proposed in the present application;

[0052] Figure 7 Linear regression function simulation chart proposed in the present application;

[0053] Figure 8 Denoising effect comparison chart proposed in the present application;

[0054] Figure 9 Wavelet transform denoising effect chart proposed in the present application;

[0055] Figure 10 Empirical mode decomposition denoising effect chart proposed in the present application;

[0056] Figure 11 Singular value denoising effect chart proposed in the present application;

[0057] Figure 12 Real-time signal denoising effect chart proposed in the present application;

[0058] Figure 13 Measured results display chart proposed in the present application;

[0059] Figure 14 Real-time monitoring interface of the partial discharge diagnosis system proposed in the present application;

[0060] Figure 15 Historical fault query interface of the partial discharge diagnosis system proposed in the present application. DETAILED DESCRIPTION

[0061] As shown in the accompanying drawings, Figure 1 a partial discharge signal denoising method, the method steps are as follows:

[0062] (1) Build a partial discharge signal denoising experimental platform;

[0063] (2) Complete the simulation of partial discharge signals by using double exponential oscillation decay model and single exponential oscillation decay model. At the same time, four groups of different partial discharge noise signals are set respectively;

[0064] (3) Set four kinds of noise signals of 50 groups for training, determine the kurtosis test, set the confidence interval of Gaussian test, improve the Gaussian test preprocessing, preprocess four groups of noisy partial discharge signals by Gaussian test, and obtain the frequency band range of the characteristic signal;

[0065] (4) After the preprocessed signal is subjected to fast S transform, a denoising algorithm combining time-frequency spectrum analysis filtering method and threshold denoising method is adopted to remove most of the periodic narrowband interference signals in the signal;

[0066] (5) After fast S transform, the signal is subjected to a fitting threshold approximation singular value denoising algorithm, and the two-dimensional signal matrix is analyzed to filter out the Gaussian white noise in the signal.

[0067] (6) The noise suppression ratio and amplitude attenuation coefficient of the denoised signal are tested, the test results are analyzed, and the test results are displayed on the display interface.

[0068] In the step (3), the signal to be processed is subjected to Gaussian test preprocessing, and the specific operation of the preprocessing is as follows:

[0069] A random variable X with known mean μ and variance σ can be subjected to Gaussian test. According to Chebyshev inequality, the distribution interval satisfies the probability inequality:

[0070]

[0071] In the above formula: q = 1-(σ 2 / ε 2 ), ε>0, assuming that there is a group of observation samples x(t) with mean value 0 and number N in the random variable X conforming to Gaussian distribution, when σ 2 / ε 2 is small enough, the kurtosis value and the variance estimate can be represented as:

[0072]

[0073]

[0074]

[0075] is the mean value of the kurtosis value, the variance of the kurtosis value, is the kurtosis value estimate of the random variable sequence x(t), and the kurtosis coefficient Early failure of the reaction system is determined by the number of observation points N and the confidence q. If the estimated kurtosis value of the variable sequence x(t) is in the interval, it indicates that the measured signal has obvious Gaussian characteristics and needs to be excluded.

[0076] The step (4) carries out fast S transform on the signal after pretreatment, and the specific processing procedure of fast S transform is as follows:

[0077] In the pretreatment stage, the frequency range of useful PD signal is obtained, the characteristic frequency reservation mechanism and the non-characteristic signal elimination mechanism are set through the envelope extremum algorithm, and the characteristic frequency is screened out through the envelope spectrum so as to be processed individually.

[0078] Because fast S transform is reversible and linear, a time-frequency filter is designed according to the time-frequency signal after transform, data in the time range t∈[t1,t2] and the frequency range f∈[f1,f2] are reserved, and data in the time range t∈[t3,t4] and the frequency range f∈[f3,f4] are suppressed, the fast S transform is discretized, and f=n / NT and τ=jT are set:

[0079]

[0080]

[0081]

[0082] Wherein, f is frequency, t and τ are time, w(t-τ,f) is a Gaussian window function, λ and p are adjustment factors, λ=0.55 and p=0.8 are taken, S[jT,n / NT] is fast S transform in the discrete case, m and n are constants, T is a sampling time interval, N is a total sampling point number, and X[(m+n) / NT] is a discrete Fourier transform of the signal x(nT).

[0083] The fast S transform algorithm procedure is as follows:

[0084] 1) According to the envelope spectrum obtained by pretreatment, useful frequency points k are screened out through Gaussian test i , i=1,2,3,...,n.

[0085] The Fourier transform of the selected frequency points of the frequency band is carried out with a Gauss self-adaptive optimization window, and the window function W′ is:

[0086]

[0087] 3) Fourier inverse transform is carried out, so as to obtain the fast S transform of the point k i

[0088] 4) Repeat 3) until the fast S transform of all frequency points is completed.

[0089] The step (5) carries out singular value decomposition on the two-dimensional mode-time frequency matrix of the signal, and the specific algorithm procedure of the threshold approximation singular value denoising algorithm is as follows: ​

[0090] Singular Value Decomposition (SVD) is a widely used algorithm in the field of machine learning. It can be used not only for feature decomposition in dimensionality reduction algorithms, but also in recommendation systems and natural language processing.

[0091] The basic formula of SVD is:

[0092] SVD: A = UΛV

[0093] In the above formula, A is an m x n real trajectory matrix

[0094] A∈R m*n , U∈R m*m , V∈R n*n , Λ∈R m*n All elements except those on the main diagonal are zero, and each element on the main diagonal is called a singular value, which is sorted in descending order. The column vectors of U are the eigenvectors of AA T , and each eigenvector in U is generally called a left singular vector of A. The column vectors of V are the eigenvectors of A T A, and each eigenvector in V is generally called a right singular vector of A.

[0095] Since all other positions are zero except the diagonal line, which is the singular value, finding the singular value matrix Λ is equivalent to finding each singular value.

[0096]

[0097]

[0098] Similar to the eigenvalues in feature decomposition, the diagonal elements in the singular value matrix are also sorted in descending order, and the number of singular values decreases rapidly. In most cases, the sum of the top 1% to 10% singular values accounts for most of the total singular value sum. Therefore, k singular values and their corresponding left and right singular vectors can be used to approximate the matrix. That is, the real trajectory matrix can be truncated, and the eigenvalues can be reduced in dimension.

[0099] Before performing SVD denoising on the signal, the signal is transformed by the Fast S Transform to obtain a two-dimensional modulus time-frequency matrix of the signal. This compensates for the time defect of traditional singular value denoising algorithms, which can only process one-dimensional time-frequency signals. The two-dimensional modulus time-frequency matrix increases the amount of information such as time and power compared to the one-dimensional time-frequency matrix, which is beneficial for further research and analysis of partial discharge signals and avoids the distortion of some characteristic quantities of PD signals after denoising due to the lack of time signals. Therefore, in this paper, the original signal is transformed by the Fast S Transform to obtain a two-dimensional modulus time-frequency matrix, and the two-dimensional modulus time-frequency matrix with more information is selected as the trajectory matrix for singular value denoising algorithm for denoising analysis.

[0100] The gap between the approximate matrix fitted by singular value and the actual matrix is crucial to the selection of k value. The limitation of singular value algorithm lies in how to select a suitable threshold value.

[0101] The fitting threshold approximation method is used to select the threshold value of singular value adaptively. The singular value algorithm is used for noise reduction because there is a large difference between the singular values of the main signal and the noise signal. The singular value of the main signal is generally much larger than that of the noise signal, so it can be processed by singular value for noise reduction. By fitting threshold approximation, the maximum point of the change of the curve slope is found, thereby determining the threshold function of the singular value.

[0102] Fitting threshold approximation method steps;

[0103] 1) Select the first n non-zero singular values from the singular value matrix Λ obtained by singular value decomposition, and convert them into a singular value vector.

[0104] 2) Select m data from the singular value vector, and divide them into n / m intervals to be processed. According to the property of singular value pairs, m must be an even number.

[0105] 3) Calculate the regression coefficient of the processing interval, determine the point with the maximum second-order difference coefficient, select the adjacent two intervals of the point, and perform curve fitting.

[0106] 4) Determine the intersection point of the curve fitting, and select data slightly smaller than the intersection point in the singular value vector as the threshold value of the singular value noise reduction algorithm.

[0107] The step (6) is to detect the filtering effect, and introduce noise suppression ratio ρ NRR and amplitude attenuation ratio ρ ARR Two coefficients to comprehensively evaluate the denoising performance:

[0108] During the actual test process, it is impossible to measure the pure local discharge signal without noise, and it is also impossible to simulate the pure local discharge signal in the field environment, so two parameters need to be introduced to evaluate the denoising performance. Therefore, two coefficients of noise suppression ratio ρ NRR and amplitude attenuation ratio ρ ARR are introduced to comprehensively evaluate the denoising performance.

[0109]

[0110]

[0111] In the formula, σ1 is the standard deviation of the signal before denoising, σ2 is the standard deviation of the signal after denoising, A m1 is the maximum amplitude of the signal before denoising, and A m2 is the maximum amplitude of the signal after denoising.

[0112] The present application adopts a double exponential oscillation decay model and a single exponential oscillation decay model to simulate pure partial discharge signals, periodic narrowband interference noise is simulated by a periodic function, and Gaussian white noise is automatically generated by a mathematical algorithm. The signal is preprocessed by Gaussian testing, and denoising is achieved by fast S transform and fitting threshold singular value approximation algorithm.

[0113] The above method can complete the denoising processing of the transformer partial discharge noise signal.

[0114] A partial discharge signal denoising experiment platform is built. First, partial discharge noise is simulated to verify the partial discharge signal denoising experiment platform. A large number of transformer partial discharge noises are simulated, and four typical partial discharge noises are selected for experimental detection. The specific partial discharge noise parameters are shown in Table 1

[0115] Table 1 Four kinds of partial discharge noise parameters

[0116]

[0117]

[0118] The processed signal is subjected to Gaussian testing, which requires calculating the kurtosis value of the measured signal. According to the kurtosis interval, it is judged whether the measured signal can pass the Gaussian test, so as to distinguish between useful signals and useless signals. When performing Gaussian testing, the selection of the confidence parameter required for testing becomes particularly important. The greater the confidence value, the greater the confidence interval, and the more rigorous the Gaussian test. If the confidence interval is too large, useful signals may be mistakenly excluded. In order to select the appropriate confidence, so as to strictly distinguish between useful signals and noise signals, the measured data is shown in the following table:

[0119] Table 2 Kurtosis value estimation

[0120] Signal type Signal type Number Kurtosis value range One 0-50dB Gaussian white noise 50 0～4 Two 0.3Ghz-1.5Ghz periodic narrowband interference 50 0～4 Three Gaussian white noise + periodic narrowband interference 50 0～4 Four Noisy partial discharge signal 50 20～70

[0121] All points that cannot pass the Gaussian test are selected from the point with the maximum envelope curve by Gaussian testing, so as to perform fast S transform. The generalized S transform and the fast S transform are compared.

[0122] The generalized S transform is generally composed of two denoising methods, threshold denoising method and time-frequency spectrum analysis filtering method. The threshold denoising method is to quantize the data according to the maximum and minimum values of the two-dimensional matrix distribution, and set appropriate threshold for each quantization interval for denoising. The time-frequency analysis filtering method is to filter and denoise according to the noise range and time range of the signal. The time-frequency spectrum analysis filtering method and the threshold denoising method are combined, the frequency range of the useful signal is filtered out in the Gaussian test process, the time range of the PD signal is determined according to the energy distribution of the signal, and the noise signals outside the range are filtered out, and the signals in the specific time range and frequency range are reserved. The signals in other time-frequency regions are all filtered out. Figure 4 Comparison waveform diagram of fast S transform and S transform.

[0123] The gap between the approximate matrix fitted by singular value and the actual matrix and the selection of k value are crucial. The limitation of the singular value algorithm lies in how to select a suitable threshold. The singular value algorithm is used for denoising, because there is a large difference between the singular values of the main signal and the noise signal, and the singular value of the main signal is generally much larger than that of the noise signal, so the singular value can be used for denoising. By fitting the threshold approximation, the maximum point of the change of the curve slope can be found, so as to determine the threshold function of the singular value. By using the fitting threshold approximation method, the two adjacent intervals with the maximum change of the second-order difference spectrum of the slope are found, the regression coefficients of the two intervals are linearly regressed to fit the two curves, the intersection point between the two curves is found, and finally the singular value slightly smaller than the intersection point value is found in the singular value sequence as the threshold value of the singular value denoising algorithm this time, so as to reconstruct the singular value. Figure 6 、 Figure 7 Waveform diagram of singular value denoising effect by fitting threshold approximation.

[0124] The denoising processing of partial discharge signal should follow two principles: one is to meet the specified signal-to-noise ratio requirement, and the other is to have small distortion of the denoised signal. The signal-to-noise ratio (SNR), root mean square error (RMSE), waveform similarity coefficient (NCC), and variation trend parameter (VTP) are introduced to represent the denoising effect of partial discharge, and the parameters are set as shown in the following table:

[0125] Table 3 Denoising effect parameter table

[0126]

[0127] Comprehensive comparison. By comparing the method of this paper with wavelet transform denoising method, empirical mode decomposition denoising method, singular value denoising method, the denoising effect is as follows

[0128] Table 4 RMSE data table

[0129] De-noising method Waveform 1 Waveform 2 Waveform 3 Waveform 4 Method in this paper 0.093 0.105 0.137 0.143 Hard threshold 0.321 2.024 0.397 0.274 Soft threshold 0.414 1.513 0.469 0.358 Fixed threshold 0.463 1.426 0.574 0.390 EMD 0.421 0.457 0.506 0.442 SVD 0.336 0.205 0.529 0.449

[0130] Table 5 SNR data table

[0131]

[0132]

[0133] Table 6 NCC data table

[0134] De-noising method Waveform 1 Waveform 2 Waveform 3 Waveform 4 Method in this paper 0.995 0.993 0.990 0.995 Hard threshold 0.955 0.363 0.453 0.342 Soft threshold 0.907 0.324 0.467 0.357 Fixed threshold 0.875 0.309 0.469 0.359 EMD 0.554 0.696 0.731 0.632 SVD 0.833 0.357 0.381 0.229

[0135] Table 7 VTP data table

[0136] De-noising method Waveform 1 Waveform 2 Waveform 3 Waveform 4 Method in this paper 1.004 0.995 1.078 1.060 Hard threshold 1.157 1.357 0.751 0.684 Soft threshold 1.297 1.533 1.174 0.750 Fixed threshold 1.284 1.235 0.784 0.822 EMD 0.846 1.324 1.452 0.855 SVD 0.885 0.950 0.702 0.642

[0137] The insulation system of the transformer is complex, involves a large number of materials, and the electric field distribution is uneven, so there are many types of partial discharge in the transformer. According to the position, phenomenon and mechanism of partial discharge, the partial discharge in the transformer can be roughly divided into three basic types: ① partial discharge in the internal insulation medium; ② partial discharge on the surface of the insulation medium; ③ corona discharge at the tip of the high-voltage electrode. Therefore, it is necessary to measure the transformer on site, and the on-site measurement site is the transformer field space of a certain transformer substation as shown in the figure, and an antenna is used to receive the UHF signal of partial discharge.

[0138] Table 8 Actual filtering parameters

[0139] NRR ]]> ​ ARR ]]> ​ Time / s 23.60 0.052 0.341

[0140] In summary, the application realizes the denoising processing of partial discharge signal by Gaussian test preprocessing, fast S transform and singular value fitting threshold approximation denoising method, and filters out Gaussian white noise and periodic narrowband interference noise. In view of the problem of long time consumption of local signal feature extraction, the fast S transform algorithm is proposed to reduce the operation speed of the algorithm and improve the time-frequency resolution of the algorithm. The traditional singular value denoising algorithm can only process one-dimensional data, and the improved singular value algorithm is used to process two-dimensional time-frequency matrix, which makes up for the signal distortion. In the field test, it can effectively filter out the noise in the partial discharge signal, reduce the feature extraction time, and also reduce the distortion occurring in the denoising process, and maximize the preservation of the original characteristics of the signal.

Claims

1. A partial discharge signal denoising method, characterized in that, The method steps are as follows: S1: build a partial discharge signal denoising experiment platform; S2: complete the simulation of partial discharge signals by adopting a double exponential oscillation decay model and a single exponential oscillation decay model, set multiple groups of partial discharge noise signals as experimental groups, and each group of partial discharge noise signals is different; S3: each experimental group sets a control group for training according to a predetermined number, determines the kurtosis test, sets the confidence interval of the Gaussian test, improves the Gaussian test preprocessing, pre-processes the noise partial discharge signals of each experimental group, and obtains the frequency band range of the characteristic signals; S4: after the pre-processed signals are subjected to fast S transformation, a denoising algorithm combining time-frequency spectrum analysis filtering and threshold denoising is adopted to remove most of the periodic narrowband interference signals in the signals; S5: after the fast S transformation, the signal two-dimensional modulus time-frequency matrix is analyzed, and the Gaussian white noise in the signal is filtered out by adopting a fitting threshold approximation singular value denoising algorithm; S6: test the noise suppression ratio and amplitude attenuation coefficient of the denoised signal, analyze the test results, and display the test results on the display interface.

2. The partial discharge signal denoising method of claim 1, wherein, The method steps for Gaussian test preprocessing of the noise partial discharge signal in step S3 are as follows: Given a random variable X with mean μ and variance σ, the random variable X is subjected to Gaussian test, and according to Chebyshev inequality, the distribution interval of the random variable X satisfies the probability inequality: In the above equation: q = 1 - (σ 2 / ε 2 ), ε > 0, assuming that there is a set of observation samples x(t) with mean value 0 and number N in the random variable X that conforms to Gaussian distribution, when σ 2 / ε 2 is small enough, the estimation of the kurtosis value and the variance of the random variable X can be represented as: wherein is the mean value of the kurtosis value, is the variance of the kurtosis value, is the kurtosis value estimate of the random variable sequence x(t), the kurtosis coefficient Early failure of the reaction system is determined by the number of observation points N and the confidence level q. If the estimated kurtosis value of the variable sequence x(t) is in the interval, it indicates that the measured signal has obvious high-order statistical properties and needs to be excluded.

3. The partial discharge signal denoising method according to claim 1, characterized in that: In the preprocessing stage, the frequency range of the useful PD signal is obtained, the envelope extremum algorithm is used to set the characteristic frequency retention mechanism and the non-characteristic signal rejection mechanism, and the envelope spectrum is used to screen out the characteristic frequency for separate processing; Because the fast S transformation is reversible and linear, a time-frequency filter is designed according to the time-frequency signal after the transformation, the data in the time range t∈[t1,t2] and the frequency range f∈[f1,f2] are retained, and the data in the time range t∈[t3,t4] and the frequency range f∈[f3,f4] are suppressed, the fast S transformation is discretized, and f=n / NT and τ=jT are set: Wherein, f is the frequency, t and τ are the time, w(t-τ,f) is the Gaussian window function, λ and p are the adjustment factors, λ=0.55 and p=0.8 are taken, S[jT,n / NT] is the fast S transformation in the discrete case, m and n are constants, T is the sampling time interval, N is the total number of samples, and X[(m+n) / NT] is the discrete Fourier transform of the signal x(nT); The method steps of the fast S transformation of the pre-processed signal in step S4 are as follows: S41: According to the envelope spectrum obtained by preprocessing, the useful frequency points k are screened out by Gaussian test i , i = 1, 2, 3,..., n; S42: Fourier transform is performed on the selected frequency points of the frequency band, and the window function W' is: S43: perform an inverse Fourier transform to obtain the point k i of the fast S transform; S44: repeat step S43 until the fast S transformation of all frequency points is completed.

4. The partial discharge signal denoising method of claim 1, wherein: The method steps of analyzing the signal two-dimensional modulus time-frequency matrix and filtering out the Gaussian white noise in the signal in step S5 are as follows: S51: select the first n singular values of the singular value matrix Λ obtained by singular value decomposition, and convert them into a singular value vector; S52: Select m data from the singular value vector, divide them into n / m intervals, and select the maximum second-order difference coefficient point in each interval. The m must be even number; S53: Calculate the regression coefficient of the processing interval, determine the point with the maximum second-order difference coefficient, and select the adjacent two intervals of the point for curve fitting; S54: Determine the intersection point of the curve fitting, and select the data slightly smaller than the intersection point in the singular value vector as the threshold value of the singular value denoising algorithm.

5. The partial discharge signal denoising method of claim 1, wherein, In step S6, in order to detect the filtering processing effect, a noise suppression ratio p is introduced NRR and an amplitude attenuation ratio p ARR Two coefficients to comprehensively evaluate the denoising performance: where σ1 is the standard deviation of the signal before denoising, σ2 is the standard deviation of the signal after denoising, A m1 is the maximum amplitude of the signal before denoising, A m2 is the maximum amplitude of the signal after denoising.

Images