A method and device for obtaining audible corona noise of contaminated insulators in windy areas

Abstract

The present invention provides a method and device for obtaining audible noise of polluted insulator corona in a windy area, which relates to the field of noise signal processing, including: training a neural network model through a training set; inputting an actual mixed noise frequency domain data set into the trained neural network model to obtain an evaluation result, and obtaining a wind-containing noise signal set through the evaluation result; performing signal decomposition on the wind-containing noise signal set and the reference acoustic signal set respectively to obtain a decomposed noisy signal set and a decomposed reference acoustic signal set; eliminating the decomposed noisy signal set and the decomposed reference acoustic signal set to obtain a eliminated acoustic signal set; reconstructing the eliminated acoustic signal set to obtain the audible noise of the insulator corona. The present invention obtains the audible noise of the insulator corona by decomposing, eliminating and reconstructing the wind-containing noise signal set and the reference acoustic signal set, which can more accurately extract complex patterns in the time series and improve the accuracy of prediction.

Description

Technical Field

[0001] The invention relates to the field of noise signal processing, and in particular to a method and device for obtaining corona audible noise of contaminated insulators in a strong wind area. Background Art

[0002] The judgment of insulator contamination level is usually obtained by collecting insulator surface contamination on site. The method of analyzing corona audible noise to obtain contamination level can reduce labor costs, but the impact of on-site wind noise cannot be ignored.

[0003] Currently, there is still a lack of effective means to obtain the audible noise of the polluted insulator corona in windy areas. The traditional method of using filters to remove wind noise has problems such as unstable performance, easy signal loss, and ineffective removal due to excessive wind noise. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a method and device for obtaining the audible noise of the corona of contaminated insulators in windy areas, so as to solve the problem that there is still a lack of effective means to obtain the audible noise of the corona of contaminated insulators in windy areas.

[0005] The present invention provides a method for obtaining corona audible noise of contaminated insulators in a strong wind area, comprising the following steps:

[0006] S1: Obtain a labeled mixed noise time domain data set, preprocess the labeled mixed noise time domain data set, and obtain a training set;

[0007] S2: Build a neural network model based on ECAPA-TDNN, train the neural network model through the training set, and obtain a trained neural network model;

[0008] S3: preprocessing the actual mixed noise time domain data set to obtain an actual mixed noise frequency domain data set, inputting the actual mixed noise frequency domain data set into the trained neural network model to obtain an evaluation result, and obtaining a wind noise signal set through the evaluation result;

[0009] S4: delaying the wind noise signal set to obtain a reference acoustic signal set, performing signal decomposition on the wind noise signal set and the reference acoustic signal set respectively, to obtain a decomposed noisy signal set and a decomposed reference acoustic signal set;

[0010] S5: removing the decomposed noisy signal set and the decomposed reference acoustic signal set to obtain a removed acoustic signal set;

[0011] S6: Reconstruct the acoustic signal set after elimination to obtain the insulator corona audible noise.

[0012] Preferably, step S1 specifically comprises:

[0013] S11: Obtain a sample mixed noise time domain data set, and convert each sample mixed noise time domain data into corresponding sample mixed noise frequency domain data by fast Fourier transform;

[0014] S12: Obtain the average amplitude of the frequency-frequency graph curve corresponding to the mixed noise frequency domain data of each sample, and set the average amplitude of the audible noise frequency domain curve;

[0015] S13: extracting the average amplitude of a certain amplitude-spectrum curve, and if the average amplitude of the amplitude-spectrum curve is higher than C% of the average amplitude of the audible noise frequency domain curve, marking the sample mixed noise frequency domain data corresponding to the average amplitude of the amplitude-spectrum curve as a non-removable signal;

[0016] If the average amplitude of the amplitude-frequency graph curve is lower than N% of the average amplitude of the audible noise frequency domain curve, the sample mixed noise frequency domain data corresponding to the average amplitude of the amplitude-frequency graph curve is marked as a removable signal, where N≤C;

[0017] S14: Repeat step S13 to label all sample mixed noise frequency domain data to obtain a training set.

[0018] Preferably, step S3 specifically comprises:

[0019] S31: converting each actual mixed noise time domain data into actual mixed noise frequency domain data by fast Fourier transform to obtain an actual mixed noise frequency domain data set;

[0020] S32: inputting the actual mixed noise frequency domain data set into the trained neural network model for recognition, and obtaining evaluation results corresponding to each actual mixed noise frequency domain data;

[0021] S33: extracting a certain actual mixed noise frequency domain data, and if the evaluation result of the actual mixed noise frequency domain data is 0, identifying the actual mixed noise frequency domain data as a non-removable signal;

[0022] If the evaluation result of the actual mixed noise frequency domain data is 1, the actual mixed noise frequency domain data is identified as a removable signal;

[0023] S34: Repeat step S33 to obtain all removable signals, and all removable signals constitute a signal set containing wind noise.

[0024] Preferably, step S5 is specifically:

[0025] S51: Construct Wiener adaptive filter;

[0026] S52: taking the decomposed noisy signal set and the decomposed reference acoustic signal set as the initial data set to be eliminated;

[0027] S53: adjusting the parameters of the filter, inputting the data set to be eliminated into the Wiener adaptive filter for elimination, and obtaining the final filter coefficients and the eliminated data set;

[0028] S54: using the eliminated data set as a new data set to be eliminated;

[0029] S55: Repeat steps S53-S54 until the final filter coefficient is less than a preset value or reaches a predetermined number of iterations, and use the final eliminated data set as the eliminated acoustic signal set.

[0030] Preferably, step S53 is specifically as follows:

[0031] S531: setting the initial value of time point n to 0;

[0032] S532: Adjust the filter parameters, input the to-be-eliminated data x[n] corresponding to time point n into the Wiener adaptive filter, and obtain the data after elimination

[0033] S533: Calculate and obtain the error signal e[n] at time point n;

[0034] S534: Calculate and obtain the filter coefficient w[n+1] at time point n+1;

[0035] S535: let n=n+1;

[0036] S536: Repeat steps S532-S535 until all the data to be eliminated are input into the Wiener adaptive filter to obtain the final filter coefficients.

[0037] Preferred:

[0038] The expression of the filter coefficient w[n+1] is:

[0039]

[0040] w[n+1]=w[n]+2μe[n]x[n]

[0041] Where μ is the step size factor, d[n] is the reference signal at time point n, and w[n] is the filter coefficient at time point n.

[0042] A storage medium, wherein the storage medium stores instructions and data for implementing a method for obtaining audible corona noise of contaminated insulators in a high-wind area.

[0043] A device for acquiring audible corona noise of contaminated insulators in a high-wind area comprises: a processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement a method for acquiring audible corona noise of contaminated insulators in a high-wind area.

[0044] The present invention has the following beneficial effects:

[0045] Using frequency domain data to construct a training set to train the neural network model makes it easier to screen out mixed noise graphs that meet the accuracy standards, thereby improving the evaluation accuracy of the trained neural network model; by delaying the wind noise signal set to obtain a reference acoustic signal set, additional reference signal collection is avoided; by decomposing, eliminating and reconstructing the wind noise signal set and the reference acoustic signal set to obtain the insulator corona audible noise, complex patterns in the time series can be extracted more accurately, thereby improving the accuracy of the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a flow chart of a method according to an embodiment of the present invention;

[0047] Figure 2 It is a schematic diagram of the decomposition of the wind noise signal set and the reference acoustic signal set;

[0048] Figure 3 This is a structural diagram of the device according to an embodiment of the present invention;

[0049] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0050] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0051] Reference Figure 1 The present application provides a method for obtaining corona audible noise of contaminated insulators in a windy area, comprising the following steps:

[0052] S1: Obtain a labeled mixed noise time domain data set, preprocess the labeled mixed noise time domain data set, and obtain a training set;

[0053] Furthermore, step S1 is specifically as follows:

[0054] S11: Obtain a sample mixed noise time domain data set, and convert each sample mixed noise time domain data into corresponding sample mixed noise frequency domain data by fast Fourier transform;

[0055] S12: Obtain the average amplitude of the frequency-frequency graph curve corresponding to the mixed noise frequency domain data of each sample, and set the average amplitude of the audible noise frequency domain curve;

[0056] S13: extracting the average amplitude of a certain amplitude-spectrum curve, and if the average amplitude of the amplitude-spectrum curve is higher than C% of the average amplitude of the audible noise frequency domain curve, marking the sample mixed noise frequency domain data corresponding to the average amplitude of the amplitude-spectrum curve as a non-removable signal;

[0057] If the average amplitude of the amplitude-frequency graph curve is lower than N% of the average amplitude of the audible noise frequency domain curve, the sample mixed noise frequency domain data corresponding to the average amplitude of the amplitude-frequency graph curve is marked as a removable signal, where N≤C;

[0058] S14: Repeat step S13 to label all sample mixed noise frequency domain data to obtain a training set.

[0059] S2: Build a neural network model based on ECAPA-TDNN, train the neural network model through the training set, and obtain a trained neural network model;

[0060] Specifically, the neural network model based on ECAPA-TDNN can be divided into five main components: one-dimensional squeeze-excitation residual module, multi-layer feature fusion, statistical pooling, dilated convolution and residual connection.

[0061] The one-dimensional squeeze excitation residual module first compresses the "Squeeze" feature map into a channel descriptor through global average pooling, and then learns the weight of each channel through the "Excitation" module composed of two fully connected layers to achieve channel dependency modeling and re-weighting. At the same time, the Res2Net structure is introduced, which introduces a multi-branch structure inside the residual unit, and each branch has a different dilation rate, so as to extract features in parallel on different receptive fields.

[0062] ECAPA-TDNN fuses multiple layers of features through feature aggregation, which includes but is not limited to concatenation, weighted summation, and attention mechanism.

[0063] Dilated convolution introduces holes (or "holes") in the standard convolution kernel, so that the elements in the convolution kernel are spaced apart from each other when performing the convolution operation. The size of this spacing is called the dilation rate. When the dilation rate is greater than 1, the convolution kernel will slide jumpily on the input data, covering more space or time range, so that even if the convolution kernel size remains unchanged, each unit in the output feature map can perceive a larger input area.

[0064] The process of neural network model training is as follows:

[0065] The cross entropy loss function is used as the training loss function of the ECAPA-TDNN model as follows:

[0066]

[0067] In this formula, y i is the value of the i-th element in the true label vector (1 for the true category element and 0 for the rest), is the probability of the i-th category predicted by the model. This loss function encourages the model to predict high probability values ​​corresponding to the true category, while reducing the predicted probabilities of other categories. When the model's prediction is completely accurate (i.e. For the true category i, the rest ), the cross entropy loss will reach a minimum value of 0.

[0068] The present invention adopts cross-validation to select the best model, specifically: the sample set is randomly divided into k parts, k-1 parts are used as training sets, 1 part is used as verification set, the training set and verification set are rotated k times in turn, and the model with the smallest verification error is selected as the best model and saved.

[0069] S3: preprocessing the actual mixed noise time domain data set to obtain an actual mixed noise frequency domain data set, inputting the actual mixed noise frequency domain data set into the trained neural network model to obtain an evaluation result, and obtaining a wind noise signal set through the evaluation result;

[0070] Furthermore, step S3 is specifically as follows:

[0071] S31: converting each actual mixed noise time domain data into actual mixed noise frequency domain data by fast Fourier transform to obtain an actual mixed noise frequency domain data set;

[0072] S32: inputting the actual mixed noise frequency domain data set into the trained neural network model for recognition, and obtaining evaluation results corresponding to each actual mixed noise frequency domain data;

[0073] S33: extracting a certain actual mixed noise frequency domain data, and if the evaluation result of the actual mixed noise frequency domain data is 0, identifying the actual mixed noise frequency domain data as a non-removable signal;

[0074] If the evaluation result of the actual mixed noise frequency domain data is 1, the actual mixed noise frequency domain data is identified as a removable signal;

[0075] S34: Repeat step S33 to obtain all removable signals, and all removable signals constitute a signal set containing wind noise.

[0076] S4: delaying the wind noise signal set to obtain a reference acoustic signal set, performing signal decomposition on the wind noise signal set and the reference acoustic signal set respectively, to obtain a decomposed noisy signal set and a decomposed reference acoustic signal set;

[0077] Specifically, ICEEMDAN is used to decompose the wind noise signal set and the reference acoustic signal set into PNLMS easy-to-process signals. By improving the noise addition method and decomposition strategy, the complex patterns in the time series can be more accurately extracted and the prediction accuracy can be improved.

[0078] The decomposition of the wind noise signal set and the reference acoustic signal set is as follows: Figure 2 As shown, the following process is included:

[0079] S41, collect noisy signals X1(t), X2(t)...X n (t).

[0080] S42, then for each noisy signal X n (t) Perform ICEEMDAN empirical mode decomposition, which is an iterative process that aims to decompose the signal into a series of intrinsic mode functions (IMFs) and a residual term. The core of ICEEMDAN is to find local extreme points, define the upper and lower envelopes by averaging the envelopes, and then take the median between the two as the new trial IMF. This process is repeated until the conditions defined by the IMF are met (the number of local minima is equal to the number of local maxima, and the median between the envelopes crosses zero no more than twice).

[0081] S43, then perform modal averaging and optimization, and perform average or median processing on the IMF set obtained from all noisy signals to reduce the inconsistency introduced by noise. For the jth IMF component, there can be:

[0082]

[0083] Here, N is the total number of noisy signals, and IMF i，j represents the jth IMF component of the ith signal.

[0084] S44, then iteration and convergence judgment are performed, and the noise addition and ICEEMDAN process are repeated until the preset upper limit of the number of iterations is reached or the decomposition quality of the signal is judged to be good enough (SNR>Thresho).

[0085] S45. Finally, adaptive strategies and optimization are performed. In actual operation, such as the AEMD and AEMD-H strategies, the noise level will be dynamically adjusted according to the results of the previous rounds of iterations to further optimize the decomposition effect and reduce pseudo-modes and modal aliasing.

[0086] S5: removing the decomposed noisy signal set and the decomposed reference acoustic signal set to obtain a removed acoustic signal set;

[0087] Furthermore, step S5 is specifically as follows:

[0088] S51: Construct Wiener adaptive filter;

[0089] S52: taking the decomposed noisy signal set and the decomposed reference acoustic signal set as the initial data set to be eliminated;

[0090] S53: adjusting the parameters of the filter, inputting the data set to be eliminated into the Wiener adaptive filter for elimination, and obtaining the final filter coefficients and the eliminated data set;

[0091] S54: using the eliminated data set as a new data set to be eliminated;

[0092] S55: Repeat steps S53-S54 until the final filter coefficient is less than a preset value or reaches a predetermined number of iterations, and use the final eliminated data set as the eliminated acoustic signal set.

[0093] Furthermore, the Wiener adaptive filter automatically adjusts the filter coefficients according to the characteristics of the input signal and the reference signal. The goal of the adjustment is to make the signal output by the filter as close to the actual noise signal as possible. Once the filter parameters are optimized, the noise signal output by the filter will be aligned with the noise phase in the original signal and superimposed in reverse phase, thereby eliminating the noise to the greatest extent. The original signal minus the noise estimate output by the filter can obtain a relatively pure signal.

[0094] Step S53 is specifically as follows:

[0095] S531: setting the initial value of time point n to 0;

[0096] S532: Adjust the filter parameters, input the to-be-eliminated data x[n] corresponding to time point n into the Wiener adaptive filter, and obtain the data after elimination

[0097] S533: Calculate and obtain the error signal e[n] at time point n;

[0098] S534: Calculate and obtain the filter coefficient w[n+1] at time point n+1;

[0099] S535: let n=n+1;

[0100] S536: Repeat steps S532-S535 until all the data to be eliminated are input into the Wiener adaptive filter to obtain the final filter coefficients.

[0101] Furthermore, the expression of the filter coefficient w[n+1] is:

[0102]

[0103] w[n+1]=w[n]+2μe[n]x[n]

[0104] Where μ is the step size factor, d[n] is the reference signal at time point n, and w[n] is the filter coefficient at time point n.

[0105] Specifically, the error signal e[n] is used to adjust the filter coefficient, and the formula is as follows:

[0106]

[0107] Where G(n) is an N-dimensional diagonal gain matrix, which is used to assign step weights to each weight coefficient of the filter, expressed as:

[0108] G(n+1)=diag{g1(n+1), g2(n+1), .., g N (n+1)}

[0109] in

[0110] γ k (n+1)=max{ρ·max[δ p , |w1(n)|, |w1(n)|,...,|w N (n)|],|w k (n)|}

[0111] In the formula

[0112]

[0113] PNLMS introduces new variables based on NLMS, and its stability is better than NLMS. ρ is generally set to a small integer to prevent the inner product of the input data vector from being too small, which will cause the stability performance to decrease.

[0114] Then the elimination capability of the adaptive filtering algorithm is evaluated by the ERLE return loss gain:

[0115]

[0116] The larger the ERLE value is, the more prominent the inhibitory ability is.

[0117] S6: Reconstruct the acoustic signal set after elimination to obtain the insulator corona audible noise.

[0118] See also Figure 3 , Figure 3 It is a working diagram of the hardware device of an embodiment of the present invention, and the hardware device specifically includes: a device 401 for acquiring audible corona noise of contaminated insulators in a strong wind area, a processor 402 and a storage medium 403.

[0119] A device 401 for acquiring audible noise from corona of contaminated insulators in a windy area: The device 401 for acquiring audible noise from corona of contaminated insulators in a windy area implements a method for acquiring audible noise from corona of contaminated insulators in a windy area.

[0120] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the method for obtaining the corona audible noise of contaminated insulators in a high wind area.

[0121] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the method for obtaining the corona audible noise of contaminated insulators in a strong wind area.

[0122] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0123] The serial numbers of the embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In a unit claim that lists several means, several of these means may be embodied by the same hardware item. The use of the words first, second, and third, etc. does not indicate any order and these words may be interpreted as identifiers.

[0124] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for obtaining corona audible noise of contaminated insulators in a strong wind area, characterized in that: The following steps are involved: S1: Obtain a labeled mixed noise time domain data set, preprocess the labeled mixed noise time domain data set, and obtain a training set; S2: Build a neural network model based on ECAPA-TDNN, train the neural network model through the training set, and obtain a trained neural network model; S3: preprocessing the actual mixed noise time domain data set to obtain an actual mixed noise frequency domain data set, inputting the actual mixed noise frequency domain data set into the trained neural network model to obtain an evaluation result, and obtaining a wind noise signal set through the evaluation result; S4: delaying the wind noise signal set to obtain a reference acoustic signal set, performing signal decomposition on the wind noise signal set and the reference acoustic signal set respectively, to obtain a decomposed noisy signal set and a decomposed reference acoustic signal set; S5: removing the decomposed noisy signal set and the decomposed reference acoustic signal set to obtain a removed acoustic signal set; S6: Reconstruct the acoustic signal set after elimination to obtain the insulator corona audible noise.

2. The method for obtaining corona audible noise of contaminated insulators in high wind areas according to claim 1 is characterized in that: Step S1 is specifically as follows: S11: Obtain a sample mixed noise time domain data set, and convert each sample mixed noise time domain data into corresponding sample mixed noise frequency domain data by fast Fourier transform; S12: Obtain the average amplitude of the frequency-frequency graph curve corresponding to the mixed noise frequency domain data of each sample, and set the average amplitude of the audible noise frequency domain curve; S13: extracting the average amplitude of a certain amplitude-spectrum curve, and if the average amplitude of the amplitude-spectrum curve is higher than C% of the average amplitude of the audible noise frequency domain curve, marking the sample mixed noise frequency domain data corresponding to the average amplitude of the amplitude-spectrum curve as a non-removable signal; If the average amplitude of the amplitude-frequency graph curve is lower than N% of the average amplitude of the audible noise frequency domain curve, the sample mixed noise frequency domain data corresponding to the average amplitude of the amplitude-frequency graph curve is marked as a removable signal, where N≤C; S14: Repeat step S13 to label all sample mixed noise frequency domain data to obtain a training set.

3. The method for obtaining corona audible noise of contaminated insulators in high wind areas according to claim 1 is characterized in that: Step S3 is specifically as follows: S31: converting each actual mixed noise time domain data into actual mixed noise frequency domain data by fast Fourier transform to obtain an actual mixed noise frequency domain data set; S32: inputting the actual mixed noise frequency domain data set into the trained neural network model for recognition, and obtaining evaluation results corresponding to each actual mixed noise frequency domain data; S33: extracting a certain actual mixed noise frequency domain data, and if the evaluation result of the actual mixed noise frequency domain data is 0, identifying the actual mixed noise frequency domain data as a non-removable signal; If the evaluation result of the actual mixed noise frequency domain data is 1, the actual mixed noise frequency domain data is identified as a removable signal; S34: Repeat step S33 to obtain all removable signals, and all removable signals constitute a signal set containing wind noise.

4. The method for obtaining corona audible noise of contaminated insulators in high wind areas according to claim 1 is characterized in that: Step S5 is specifically as follows: S51: Construct Wiener adaptive filter; S52: taking the decomposed noisy signal set and the decomposed reference acoustic signal set as the initial data set to be eliminated; S53: adjusting the parameters of the filter, inputting the data set to be eliminated into the Wiener adaptive filter for elimination, and obtaining the final filter coefficients and the eliminated data set; S54: using the eliminated data set as a new data set to be eliminated; S55: Repeat steps S53-S54 until the final filter coefficient is less than a preset value or reaches a predetermined number of iterations, and use the final eliminated data set as the eliminated acoustic signal set.

5. The method for obtaining corona audible noise of contaminated insulators in high wind areas according to claim 4 is characterized in that: Step S53 is specifically as follows: S531: setting the initial value of time point n to 0; S532: Adjust the filter parameters, input the to-be-eliminated data x[n] corresponding to time point n into the Wiener adaptive filter, and obtain the data after elimination S533: Calculate and obtain the error signal e[n] at time point n; S534: Calculate and obtain the filter coefficient w[n+1] at time point n+1; S535: let n=n+1; S536: Repeat steps S532-S535 until all the data to be eliminated are input into the Wiener adaptive filter to obtain the final filter coefficients.

6. The method for obtaining corona audible noise of contaminated insulators in a strong wind area according to claim 5, characterized in that: The expression of the filter coefficient w[n+1] is: w[n+1]=w[n]+2μe[n]x[n] Where μ is the step size factor, d[n] is the reference signal at time point n, and w[n] is the filter coefficient at time point n.

7. A storage medium, characterized in that: The storage medium stores instructions and data for implementing a method for acquiring audible corona noise of contaminated insulators in a windy area as described in any one of claims 1 to 6.

8. A device for obtaining corona audible noise of dirty insulators in windy areas, characterized by: include: Processor and storage medium; the processor loads and executes instructions and data in the storage medium to implement a method for obtaining audible corona noise of contaminated insulators in a windy area as described in any one of claims 1 to 6.

Images