PDC (Polycrystalline Diamond Compact) signal denoising method for iterative Wiener filtering based on maximum correlation entropy

Abstract

According to the PDC signal denoising method based on iteration Wiener filtering of the maximum correlation entropy, on the basis of a traditional Wiener filtering structure, the maximum correlation entropy criterion is introduced to serve as an optimization target, the statistical correlation between filter output and expected signals is enhanced, and the method has the good inhibition capacity for non-Gaussian and non-linear impact noise. According to the method, adaptive convergence can be realized under the condition that an accurate noise priori model is not needed, matrix inversion is avoided, the calculation amount is low, and the method has good engineering realizability and expansibility. The PDC signal preprocessing method is suitable for a PDC signal preprocessing task in the insulation state detection process of the hydro-generator, can serve as a preposed step for expanding Debye equivalent circuit model construction and polarization / depolarization current dynamic decoupling, provides a high-quality data basis for achieving accurate evaluation of major insulation degradation of the large hydro-generator, and has remarkable practical application value.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of signal denoising, and particularly relates to a PDC signal denoising method based on maximum correlation entropy iterative Wiener filtering. BACKGROUND

[0002] The polarization and depolarization current (PDC) method is a time-domain dielectric response test method widely used in the detection of the insulation state of power equipment in recent years. This method has the characteristics of non-destructive, high sensitivity, rich information, etc., and is particularly suitable for the aging state assessment of the main insulation of the stator winding of a hydro-generator. By monitoring the polarization and depolarization process current signals, the physical properties of the relaxation behavior and conductivity effect of the insulating medium inside can be obtained, and an extended Debye equivalent circuit model can be constructed for quantitative diagnosis.

[0003] However, in actual field tests, due to the influence of electromagnetic interference, equipment grounding, power supply noise and operation disturbance, the PDC signals obtained are often accompanied by strong background noise. This kind of noise often presents non-Gaussian distribution characteristics, has pulse burst characteristics and high amplitude and short time characteristics, and is often modeled as Bernoulli-Gaussian (BG) impulse noise. Especially when detecting early aging or slight degradation of insulation, the PDC signal amplitude itself is low, and it is more likely to be covered by noise, directly affecting the signal modeling accuracy and diagnosis reliability.

[0004] Traditional PDC signal denoising methods, such as low-pass filtering, sliding average, adaptive LMS or RLS filter, etc., are mainly aimed at stationary or Gaussian white noise, and cannot effectively suppress non-stationary, nonlinear impulse interference signals. While removing noise, these methods may also lose the dominant components with physical meaning in the signal, resulting in weakening or even distortion of the signal characteristics. The existing methods generally have slow convergence speed, poor robustness, and insufficient adaptability to complex noise, which restricts the stable application of the PDC method in engineering field. SUMMARY

[0005] The purpose of the present application is to solve the technical problems in the above background. The PDC signal denoising method based on maximum correlation entropy iterative Wiener filtering includes the following steps: S1, a Bernoulli-Gaussian noise signal is generated according to the non-stationary impulse noise characteristics in the PDC signal detection scene, and a corresponding reference input signal is constructed based on the Bernoulli-Gaussian noise signal; S2, a pure PDC signal is superimposed with a noise signal to form a noisy PDC signal, and the noisy PDC signal is sampled as an expected response signal; S3. Initialize the filter's weight coefficient vector; S4. Construct the input vector, calculate the output value based on the weight coefficient vector and the input vector, and then combine it with the expected response signal to obtain the residual signal; S5. Calculate the autocorrelation matrix of the input vector and the cross-correlation vector between the input vector and the desired response signal; S6. Calculate the iteration step size; S7. Update the weight coefficient vector at the next time step based on the autocorrelation matrix, cross-correlation vector, and iteration step size; S8. Repeat the above steps until the termination condition is met, and output the denoised PDC signal.

[0006] In the preferred embodiment, step S1 specifically includes: Generate Bernoulli-Gaussian (BG) noise signal Reference input in BG noise scenarios BG noise can be modeled as: ,in It is Gaussian white noise with a mean of 0 and a variance of 1. Represents a BG process, Let represent an independent and identically distributed binomial event. This indicates that the mean is 0 and the variance is . Gaussian white noise, Obtained through an autoregressive system The transfer function of the autoregressive system is .

[0007] In the preferred scheme, the transfer function of the autoregressive system The transfer function is used to adjust the impact characteristics and distribution of the noise signal to adapt to the detection environment of PDC signal.

[0008] In the preferred embodiment, the specific expression of the desired response signal in step S2 is as follows: ,in It is a PDC signal. It's a noisy PDC signal. (Regarding...) Sampling: The sampling frequency for noisy PDC signals is 10Hz~1kHz, and the sampling duration is consistent with the polarization-depolarization period of the PDC signal to ensure complete capture of the relaxation characteristics of the signal.

[0009] In the preferred embodiment, the weight coefficient vector in step S3 is: Where M is the filter order; initial time; current time hour, This ensures stability during the initial stage of filtering.

[0010] In the preferred embodiment, in step S4, the reference input is... Sampling , and by Calculate the current time input vector The filter is based on the current time. weight coefficient vector For the current moment input vector Filtering is performed to obtain the output value at the current time. , ;Depend on and Get the current time residual signal .in, The calculation formula is: With continuous updates, It's getting closer and closer. At this point, the noise is eliminated.

[0011] In the preferred embodiment, the formulas for calculating the autocorrelation matrix and cross-correlation vector in step S5 are as follows: Autocorrelation matrix of the input vector The cross-correlation vector between the input vector and the desired response The calculation is as follows: The filter generates the current time. , and Calculated separately as follows: ,and .

[0012] In the formula, , λ represents the kernel width, and λ represents the forgetting factor, which ranges from 0.9 to 1.

[0013] In the preferred embodiment, the formula for calculating the iteration step size in step S6 is: iteration step size The calculation is as follows: ,in This is the gradient vector.

[0014] In the preferred scheme, the weight vector is updated according to steps S1, S4, S5, and S6, and the filter is used to calculate the next time step. weight vector , .

[0015] In the preferred scheme, the mixed noise pre-classification and separation step is also included: before step S1, the PDC signal containing noise is decomposed into different frequency band components by using the wavelet packet decomposition combined with the empirical mode decomposition method, the power frequency interference, the BG impact noise and the random noise type are identified through the spectral characteristics, the peak factor and the kurtosis value, and each noise component set and the PDC signal dominant component set are obtained respectively, so that the mixed multi-type noise scene is adapted.

[0016] In the preferred scheme, the kernel width sigma in step S5 is adaptively adjusted, and the specific formula is , wherein =0.8, is the real-time BG noise energy proportion; a pulse noise enhancement and suppression step is newly added after step S4, the strong pulse noise is detected based on the dynamic threshold , and the residual signal is replaced by the threshold and locally smoothed, so that the mixed noise suppression capability is improved.

[0017] Compared with the prior art, the present application has the following beneficial effects: (1) The present application adopts the iterative Wiener filtering algorithm, which converges fast and has good steady-state performance. Therefore, the present application has a better removal effect on the PDC signal noise, and has strong generalization ability.

[0018] (2) The present application iteratively updates the weight coefficient based on the iterative Wiener filtering algorithm of the maximum correlation entropy, avoids the matrix inversion operation, thereby reducing the high computational complexity problem of the algorithm, and makes the algorithm easy to implement; the PDC signal denoising has more accurate effect. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is a PDC signal before and after containing noise and denoising effect comparison chart.

[0020] Figure 2 is a double exponential oscillation simulation signal denoising effect comparison chart.

[0021] Figure 3 is a PDC signal before and after containing another noise and denoising effect comparison chart.

[0022] Figure 4 is another oscillation simulation signal denoising effect comparison chart. DETAILED DESCRIPTION

[0023] The present application discloses a PDC signal denoising method based on the maximum correlation entropy iterative Wiener filtering, and the steps are mainly: A, generating a Bernoulli-Gaussian (BG) noise signal and a reference input under the BG noise scene ; B, generating an expected response signal C. Generation of weight vectors; D. Calculation of output values ​​and residual signals obtained from steps A and B; E. Autocorrelation matrix of input vectors. The cross-correlation vector between the input vector and the desired response F, iteration step size Calculation: G. Update the weight vector obtained from steps A, D, E, and F; H. Iterate, letting This continues until the noise is completely eliminated.

[0024] A PDC signal denoising method based on iterative Wiener filtering with maximum correlation entropy, the steps of which are as follows: A. Generate Bernoulli-Gaussian (BG) noise signal Reference input in BG noise scenarios

[0025] BG noise can be modeled as: ,in It is Gaussian white noise with a mean of 0 and a variance of 1. Represents a BG process, Let represent an independent and identically distributed binomial event. This indicates that the mean is 0 and the variance is . Gaussian white noise, Obtained through an autoregressive system The transfer function of the autoregressive system is .

[0026] B. Generate expected response signal

[0027] ,in It is a PDC signal. It's a noisy PDC signal. (Regarding...) sampling.

[0028] C. Generation of weight vectors The filter generates the current time. The weight coefficients of the first M-1 time steps , , ..., And these M weight coefficients constitute the current time step. weight coefficient vector , Current moment hour .

[0029] D. Calculate the output value and residual signal obtained from steps A and B. For reference input Sampling and by the input vector at the current time is calculated The filter filters the input vector at the current time according to the weight coefficient vector at the current time to obtain the output value at the current time , The residual signal at the current time is obtained by and The calculation formula of is as follows: With continuous updating, is closer and closer to , and at this time the noise is eliminated. E、The autocorrelation matrix of the input vector The cross-correlation vector of the input vector and the expected response is calculated as follows:

[0030] The filter generates the weight vector at the current time , and is calculated as follows: , and . , and .

[0031] In the formula, , represents the kernel width, and λ represents the forgetting factor, which is in the range of 0.9-1.

[0032] F、The iteration step size is calculated as follows: wherein is the gradient vector.

[0033] G、According to steps A, D, E, F, the weight vector is calculated to update and obtain The filter calculates the weight vector at the next time , .

[0034] H、Iteration Let , repeat steps A, B, C, D, E, F, G, H until the algorithm ends.

[0035] 2. Simulation experiment: In order to verify the effectiveness of the present application, a comparative simulation experiment is carried out. The order M of the filter in the simulation experiment is set to 5. In the simulation experiment, the input vector​​Figure 1 In the case of PDC signal, a double exponential oscillation signal model is used, and the PDC signal is completely submerged in the impact noise signal In the case of PDC signal, a double exponential oscillation signal model is used, and the PDC signal is completely submerged in the impact noise signal ; the first line is the original PDC signal without noise, the second line is the PDC signal with noise, and the third line is the signal obtained after removing the noise, and it can be seen that the method proposed in the application has good denoising effect. Figure 2 In the case of PDC signal, a double exponential oscillation signal model is used, and the PDC signal is completely submerged in the impact noise signal , the parameter setting of the GsLMF method is , and the parameter setting of the RLS method is . It can be seen that the method (MCC-IWF) proposed in the application has the lowest MSE and the best denoising effect. Figure 3 In the case of PDC signal, a double exponential oscillation signal model is used, and the PDC signal is completely submerged in the impact noise signal In the case of PDC signal, a double exponential oscillation signal model is used, and the PDC signal is completely submerged in the impact noise signal ; it can be seen that the method proposed in the application has good denoising effect. Figure 4 In the case of PDC signal, a double exponential oscillation signal model is used, and the PDC signal is completely submerged in the impact noise signal , the parameter setting of the GsLMF method is , and the parameter setting of the RLS method is .

[0036] It can be found that the method of the application has better noise elimination performance. Figure 1 , Figure 2 , Figure 3 , Figure 4 It can be found that the method of the application has better noise elimination performance.

[0037] The above examples are only used to illustrate the technical solutions of the application, and not to limit them; although the application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that they can modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the application.​​

Claims

1. A PDC signal denoising method based on maximum correlation entropy iterative Wiener filtering, characterized by: The method comprises the following steps: S1, generating a Bernoulli-Gaussian noise signal for detecting non-stationary impact noise characteristics in a PDC signal scene, and constructing a corresponding reference input signal based on the Bernoulli-Gaussian noise signal; S2, superimposing a pure PDC signal and a noise signal to form a noisy PDC signal, taking the noisy PDC signal as an expected response signal, and sampling the noisy PDC signal; S3, initializing a weight coefficient vector of the filter; S4, constructing an input vector, calculating an output value according to the weight coefficient vector and the input vector, and obtaining a residual signal in combination with the expected response signal; S5, calculating an autocorrelation matrix of the input vector and a cross-correlation vector of the input vector and the expected response signal; S6, calculating an iteration step; S7, updating the weight coefficient vector at the next time according to the autocorrelation matrix, the cross-correlation vector and the iteration step; S8, repeating the above steps until a termination condition is met, and outputting a denoised PDC signal.

2. The PDC signal denoising method based on the maximum correlation entropy iterative Wiener filtering according to claim 1, characterized in that, Step S1 is specifically: Generating a bernoulli-gaussian (BG) noise signal and reference input under a BG noise scenario , the BG noise is modeled as: where is a Gaussian white noise with mean 0 and variance 1, denotes a BG process, denotes an independent and identically distributed binomial distributed event, denotes a Gaussian white noise with mean 0 and variance , is obtained by passing through an autoregressive system with transfer function .

3. The PDC signal denoising method based on the maximum correlation entropy iterative Wiener filtering according to claim 2, characterized in that, Transfer function of an autoregressive system Impulse characteristics and distribution laws of the noise signal are adjusted by the transfer function, and the detection environment of the PDC signal is adapted.

4. The PDC signal denoising method based on the maximum correlation entropy iterative Wiener filtering according to claim 1, characterized in that, The specific expression of the expected response signal in step S2 is wherein is the PDC signal, is the PDC signal with noise; the sampling of is the sampling of the PDC signal with noise; the sampling frequency of the PDC signal with noise is 10 Hz to 1 kHz, and the sampling duration is consistent with the polarization-depolarization period of the PDC signal, so as to ensure complete capture of the relaxation characteristics of the signal.

5. The PDC signal denoising method based on the maximum correlation entropy iterative Wiener filtering according to claim 1, characterized in that, The weight coefficient vector in step S3 is where M is the filter order. initial time instant current time instant time, , to ensure stability of the filter start-up phase.

6. The PDC signal denoising method based on the maximum correlation entropy iterative Wiener filtering according to claim 1, characterized in that, In step S4, the reference input is... Sampling , and by Calculate the current time input vector The filter is based on the current time. weight coefficient vector For the current moment input vector Filtering is performed to obtain the output value at the current time. , ;Depend on and Get the current time residual signal ,in, The calculation formula is: ； As the update continues, The closer it gets At this point the noise is eliminated.

7. The PDC signal denoising method based on the maximum correlation entropy iterative Wiener filtering according to claim 1, characterized in that, The calculation formulas of the autocorrelation matrix and the cross-correlation vector in step S5 are respectively: autocorrelation matrix of the input vector cross-correlation vector of the input vector and the desired response is computed as follows: Filter generates the current time , and are calculated as: ； ； wherein , denotes the kernel width, and λ denotes a forgetting factor, which takes a value in the range between 0.9 and 1.

8. The PDC signal denoising method based on the maximum correlation entropy iterative Wiener filtering according to claim 1, characterized in that, The formula for calculating the iteration step length in step S6 is iteration step length The calculation is as follows: where is the gradient vector.

9. The PDC signal denoising method based on the maximum correlation entropy iterative Wiener filtering according to claim 1, characterized in that, According to the weight vector calculation of steps S1, S4, S5, S6, the weight vector update is obtained and the filter is calculated for the next time , .​ 10. The PDC signal denoising method based on the maximum correlation entropy iterative Wiener filtering according to claim 1, characterized in that, Further comprising a hybrid noise pre-classification and separation step: before step S1, a noisy PDC signal is decomposed into different frequency band components by using a wavelet packet decomposition and empirical mode decomposition combined method, frequency interference, BG impact noise and random noise types are identified through spectral characteristics, peak factor and kurtosis value, and each noise component set and a PDC signal dominant component set are obtained, thereby adapting to a mixed multi-type noise scene.

11. The PDC signal denoising method based on the maximum correlation entropy iterative Wiener filtering according to claim 10, characterized in that, The nuclear width σ in step S5 is adaptively adjusted, and the specific formula is wherein = 0.8, is the real-time BG noise energy proportion; a step of enhancing and suppressing impulse noise is newly added after step S4, and a dynamic threshold is used to detect strong impulse noise, replace the residual signal by threshold, and perform local smoothing processing to improve the mixed noise suppression capability.

Images