A method and device for extracting dynamic change characteristics of signals in a strong noise environment

Abstract

This invention discloses a method and apparatus for extracting dynamic features of signals under strong noise environments. The method includes determining the optimal decomposition level for a wind turbine generator sound signal using wavelet entropy weights after wavelet decomposition; calculating the lower and upper thresholds of the wavelet coefficients at each level under the optimal decomposition level, as well as the wavelet estimation coefficients, and reconstructing the denoised wind turbine generator sound signal; uniformly dividing the wavelet estimation coefficients of each level in the denoised wind turbine generator sound signal into time sub-intervals to calculate the total wavelet energy and spectral centroid; and constructing the frequency domain feature matrix of the denoised wind turbine generator sound signal using the total wavelet energy and spectral centroid of each time sub-interval as the extraction result output. This invention aims to eliminate the interference of background noise on wind turbine generator sound signals and achieve accurate and effective extraction of the dynamic features of wind turbine generator sound signals with strong background noise.

Description

Technical Field

[0001] This invention relates to the field of fault feature extraction technology for wind turbine generator sets, specifically to a method and apparatus for extracting dynamic change features of signals under strong noise conditions. Background Technology

[0002] With the continuous development of wind power generation and technological advancements, mechanical equipment is constantly evolving towards higher power, higher speeds, and continuous operation. This means that most wind turbine generators now operate in environments with strong noise, making early sound fault signals difficult to detect. Therefore, extracting the sound characteristics of wind turbine generator operation under strong noise conditions is crucial for enterprise safety, preventing employee accidents and avoiding potential safety hazards. Weak signal extraction technology is an important branch of signal processing, primarily studying the characteristic frequencies of target signals, the variation patterns and characteristics of noise, and thus extracting weak target features submerged in strong noise. Wavelet threshold denoising of wind turbine generator sound in strong noise environments is an effective denoising method. Numerous experiments have shown that selecting the correct wavelet basis, designing a suitable threshold function, and providing a quantitative determination of the number of wavelet decomposition levels are important factors affecting the threshold denoising results of wind turbine generator sound. Traditional wavelet thresholding denoising algorithms mainly rely on soft and hard thresholding functions proposed by Dohono et al. However, the traditional hard thresholding function is discontinuous at the threshold, resulting in oscillating effects in the reconstructed signal. While the soft thresholding function offers good continuity at the threshold, the deviation between the wavelet coefficients and the estimated wavelet coefficients is significant, reducing the similarity to the original signal and thus limiting denoising effectiveness. Traditional methods for quantitatively determining the number of wavelet decomposition levels are primarily based on noise testing and singular value analysis. However, noise testing requires prior knowledge of the noise feature sequence, which is often difficult to meet in the actual environment of wind turbine power generation. Singular value analysis methods require calculating the corresponding singular value slope based on the singular value characteristics of the signal's wavelet decomposition coefficients and comparing it with a threshold slope to determine the optimal number of levels, resulting in high computational complexity and difficulty in setting a suitable threshold slope. In summary, existing wavelet thresholding denoising methods for wind turbine power generation signals in high-noise environments are insufficient to accurately separate the wind turbine power generation signal from the background noise, still exhibiting certain errors. Furthermore, the computational complexity for quantitatively determining the number of wavelet decomposition levels is too high to meet the needs of practical environments. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a method and apparatus for extracting dynamic change features of signals under strong noise environment, in view of the above-mentioned problems of the prior art. The present invention aims to eliminate the interference of background noise on the sound signal of wind turbine power generation and realize the accurate and effective extraction of the dynamic features of the sound signal of wind turbine power generation with strong background noise.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0005] A method for extracting dynamic change features of signals under strong noise environment, comprising the following steps:

[0006] S101, For the acquired wind turbine power generation sound signal under strong noise environment, starting from the initial wavelet decomposition level q, perform wavelet decomposition on the original data of the wind turbine power generation sound signal at different wavelet decomposition levels q, and calculate the wavelet entropy weight η after q-level wavelet decomposition. q If η q If the value is less than or equal to the wavelet entropy weight setting value, then the current wavelet decomposition level q is taken as the optimal decomposition level, and the process jumps to step S102.

[0007] S102, using wavelet entropy weights to calculate the lower and upper thresholds of wavelet coefficients for each layer under the optimal decomposition layer number, given the threshold value;

[0008] S103, calculate the wavelet estimation coefficients of each layer by combining the lower and upper thresholds of the wavelet coefficients of each layer;

[0009] S104, the denoised wind turbine power generation sound signal is reconstructed using the wavelet estimation coefficients of each layer under the optimal decomposition layer number;

[0010] S105, the wavelet estimation coefficients of each layer in the denoised wind turbine power generation sound signal are evenly divided into r time sub-intervals, the total wavelet energy and spectral centroid of each time sub-interval are calculated, and the frequency domain feature matrix of the denoised wind turbine power generation sound signal is constructed by the total wavelet energy and spectral centroid of each time sub-interval as the extraction result output.

[0011] Optionally, in step S101, the wavelet entropy weight η after the q-layer wavelet decomposition is calculated. q The function expression is:

[0012]

[0013] In the above formula, W q The wavelet entropy is the result of wavelet decomposition at the q-level.

[0014] Optionally, the expression for calculating the wavelet entropy after the q-layer wavelet decomposition is:

[0015]

[0016] In the above formula, m is the number of subintervals into which the wavelet coefficients are evenly divided after q-layer wavelet decomposition, and W q,i Let be the wavelet entropy of the i-th subinterval after q-level wavelet decomposition, and we have:

[0017]

[0018] In the above formula, P q,iLet E be the relative wavelet energy of the i-th subinterval. q,i E represents the wavelet energy of the i-th sub-interval after q-level wavelet decomposition, where N is the number of sampling points. q D represents the wavelet energy during q-level wavelet decomposition. q,k Let be the k-th wavelet coefficient after q-level wavelet decomposition, and we have:

[0019]

[0020] In the above formula, w(n) is the nth discrete signal of the wind turbine's sound signal, n = 1, 2, ..., N.

[0021] Optionally, when calculating the lower and upper thresholds of wavelet coefficients at each level using the given threshold value of wavelet entropy weight in step S102, calculating the lower and upper thresholds of any j-th level wavelet coefficient using the given threshold value of wavelet entropy weight includes:

[0022] S201, based on the wavelet entropy weight η of the j-th layer j Determine the threshold adjustment factor β under the wavelet coefficients of the j-th layer. j,L and upper threshold adjustment factor β j,H :

[0023] If η j >a1, then:

[0024] β j,L =β j,Lmax ,β j,H =β j,Hmax ;

[0025] If a2 < η j If ≤a1, then:

[0026] β j,L =αβ j,Lmax ,β j,H =αβ j,Hmax ;

[0027] If a3 < η j If ≤a2, then:

[0028]

[0029] If η j If ≤ a3, then all wavelet coefficients of the j-th layer are retained, where α is a parameter less than 1, and β... j,Lmax and β j,Hmax Let be the maximum values ​​of the lower and upper adjustment factors, respectively, and we have:

[0030]

[0031] Where, μ j and s jLet D be the mean and standard deviation of the wavelet coefficients of the j-th layer. j,k Let a be the k-th wavelet coefficient after the j-th layer wavelet decomposition, a3 be the wavelet entropy weight setting value, and a1 and a2 be a set of threshold parameters arranged from largest to smallest in the range (100% to a3).

[0032] S202, after obtaining the wavelet coefficients of the j-th layer, the upper threshold adjustment factor β j,L and β j,H Based on this, the lower and upper thresholds of the wavelet coefficients of the j-th layer are determined according to the following formula:

[0033] λ j,L =μ j -β j,L s j , λ j,H =μ j +β j,H s j ,

[0034] In the above formula, λ j,L λ is the lower threshold of the wavelet coefficients at the j-th layer. j,H is the upper threshold of the wavelet coefficients of the j-th layer.

[0035] Optionally, in step S103, when calculating the wavelet estimation coefficients of each layer by combining the lower and upper thresholds of the wavelet coefficients of each layer, the calculation function expression for the wavelet estimation coefficient of any j-th layer is:

[0036]

[0037] In the above formula, Let D be the estimated coefficient of the k-th wavelet in the j-th layer, sign be the sign function, and D be the value of D. j,k Let λ be the k-th wavelet coefficient after j-level wavelet decomposition. j,L λ is the lower threshold of the wavelet coefficients at the j-th layer. j,H λ is the upper threshold of the wavelet coefficients of the j-th layer. j,avg The lower threshold λ of the wavelet coefficients of the j-th layer j,L Upper threshold λ j,H The absolute mean between them, where M is the smoothing factor.

[0038] Optionally, the function expression for the denoised wind turbine generator sound signal reconstructed using the wavelet estimation coefficients of each layer in step S104 is as follows:

[0039]

[0040] In the above formula, This represents the nth denoised sound signal from the wind turbine. Let ψ(2) be the wavelet estimation coefficient of the j-th layer, and let ψ(2) be the wavelet estimation coefficient of the k-th layer. -jnk) represents the k-th wavelet function of the j-th layer, n is the sampling point index, k is the index of the wavelet estimation coefficient, and N is the number of sampling points.

[0041] Optionally, in step S105, when calculating the total wavelet energy and spectral centroid of each time sub-interval, the calculation function expression for the total wavelet energy and spectral centroid of any i-th time sub-interval is:

[0042]

[0043]

[0044] In the above formula, E i Let SC be the total wavelet energy of the i-th time sub-interval. i Let f be the spectral centroid of the i-th time subinterval. s Sampling frequency, Let be the estimated coefficients of the k-th wavelet in the j-th layer.

[0045] Optionally, the functional expression of the frequency domain feature matrix of the denoised wind turbine power generation sound signal constructed in step S105 is:

[0046]

[0047] In the above formula, G is the frequency domain characteristic matrix, E1~E r The total wavelet energy of the 1st to rth time sub-intervals is SC1 to SC2. r Let be the spectral centroid of the first to r-th time sub-intervals.

[0048] In addition, the present invention also provides a device for extracting dynamic change features of signals under strong noise environment, including a microprocessor and a memory connected to each other, wherein the microprocessor is programmed or configured to execute the method for extracting dynamic change features of signals under strong noise environment.

[0049] Furthermore, the present invention also provides a computer-readable storage medium storing a computer program, the computer program being programmed or configured by a microprocessor to execute the method for extracting dynamic signal change features under strong noise conditions.

[0050] Compared with the prior art, the present invention has the following main advantages:

[0051] 1. Wind turbines often generate strong background noise during operation, which easily obscures the sound features related to their operating status, making accurate extraction difficult. Wavelet thresholding denoising, with its advantages of fast computation speed and flexible basis selection, is widely used in signal denoising. In wavelet thresholding denoising, insufficient decomposition levels result in inadequate noise removal, while excessive decomposition levels lead to the loss of some effective information. Therefore, determining the optimal number of wavelet decomposition levels is crucial. The technical challenge lies in how to provide a suitable and easily calculable index that can distinguish the characteristics of strong background noise signals and wind turbine generation sound signals after multi-level wavelet decomposition, thereby quantitatively determining the optimal number of wavelet decomposition levels. This invention utilizes wavelet entropy weighting η... q It can quantitatively and accurately find the appropriate wavelet decomposition level for the sound signal of wind turbine power generation in a high-noise environment, thus achieving sufficient removal of noise signals while retaining the effective sound signal of wind turbine power generation.

[0052] 2. After determining the optimal wavelet decomposition level, it is necessary to select appropriate thresholds and threshold functions to correct the wavelet coefficients obtained from the decomposition. Traditional methods often use fixed thresholds and soft / hard threshold functions. As the wavelet decomposition level increases, the signal length increases significantly, and the threshold becomes significantly larger. Traditional hard threshold functions have poor continuity, introducing oscillation factors into the wavelet estimation coefficients. Traditional soft threshold functions show a large deviation between the estimated and actual wavelet coefficients, making it difficult to reconstruct an effective wind turbine generator sound signal. This invention includes using wavelet entropy weights to calculate the lower and upper thresholds of the wavelet coefficients at each level under the optimal decomposition level, achieving adaptive lower and upper thresholds. By providing effective thresholds and threshold functions to the wavelet coefficients obtained after wavelet decomposition, accurate wavelet estimation coefficients are obtained. This achieves quantization by providing appropriate adaptive thresholds and improved threshold functions to obtain accurate wavelet estimation coefficients, enabling effective extraction of dynamic features from wind turbine generator sound signals with strong background noise.

[0053] 3. This invention includes segmenting the signal into different time series based on the denoised wavelet estimation coefficients, calculating the wavelet energy and spectral centroid of each segment signal as the overall frequency domain feature of the denoised wind turbine power generation sound signal, constructing the frequency domain feature matrix of the denoised wind turbine power generation sound signal based on this, and then obtaining the time series of the frequency domain feature matrix of the denoised wind turbine power generation sound signal, which can be used to identify and compare its dynamic change features, and can accurately characterize the dynamic change features in the frequency domain.

[0054] By combining the above methods, the present invention can effectively eliminate the interference of background noise on the sound signal of wind turbine power generation, and achieve accurate and effective extraction of the dynamic characteristics of the sound signal of wind turbine power generation containing strong background noise. Attached Figure Description

[0055] Figure 1This is a schematic diagram of the basic process of the method in an embodiment of the present invention.

[0056] Figure 2 This is a waveform diagram of the fan sound signal with strong background noise used in an embodiment of the present invention.

[0057] Figure 3 The waveform diagram shows the wavelet coefficients of each layer under the optimal decomposition layer number in this embodiment of the invention.

[0058] Figure 4 These are the original wavelet coefficients of the first layer in the embodiment of the present invention.

[0059] Figure 5 These are the original wavelet coefficients of the second layer in the embodiment of the present invention.

[0060] Figure 6 These are the original wavelet coefficients of the third layer in the embodiment of the present invention.

[0061] Figure 7 These are the original wavelet coefficients of the fourth layer in the embodiment of the present invention.

[0062] Figure 8 These are the wavelet estimation coefficients of the first layer in this embodiment of the invention.

[0063] Figure 9 These are the wavelet estimation coefficients for the second layer in this embodiment of the invention.

[0064] Figure 10 These are the wavelet estimation coefficients for the third layer in this embodiment of the invention.

[0065] Figure 11 These are the wavelet estimation coefficients for the fourth layer in this embodiment of the invention. Detailed Implementation

[0066] This invention addresses the challenge of effectively extracting the dynamic characteristics of wind turbine sound signals with strong background noise. It provides a solution for extracting dynamic change features of signals under strong noise conditions, specifically for the fault diagnosis of wind turbine generators. The sound signal of a wind turbine generating power under strong noise conditions can be represented as:

[0067] w(t)=u(t)+ξ(t),

[0068] Wherein, w(t) represents the sound signal of wind turbine power generation under strong noise conditions, u(t) represents the effective sound signal of wind turbine power generation, and ξ(t) represents the background noise signal. This invention aims to eliminate the influence of background noise signals and achieve dynamic feature extraction from the effective sound signal of wind turbine power generation. The invention will be further described in detail below with reference to the accompanying drawings.

[0069] like Figure 1As shown, the method for extracting dynamic signal change features under strong noise environment in this embodiment includes the following steps:

[0070] S101, For the acquired wind turbine power generation sound signal under strong noise environment, starting from the initial wavelet decomposition level q, perform wavelet decomposition on the original data of the wind turbine power generation sound signal at different wavelet decomposition levels q, and calculate the wavelet entropy weight η after q-level wavelet decomposition. q If η q If the value is less than or equal to the wavelet entropy weight setting value, then the current wavelet decomposition level q is taken as the optimal decomposition level, and the process jumps to step S102.

[0071] S102, using wavelet entropy weights to calculate the lower and upper thresholds of wavelet coefficients for each layer under the optimal decomposition layer number, given the threshold value;

[0072] S103, calculate the wavelet estimation coefficients of each layer by combining the lower and upper thresholds of the wavelet coefficients of each layer;

[0073] S104, the denoised wind turbine power generation sound signal is reconstructed using the wavelet estimation coefficients of each layer under the optimal decomposition layer number;

[0074] S105, the wavelet estimation coefficients of each layer in the denoised wind turbine power generation sound signal are evenly divided into r time sub-intervals, the total wavelet energy and spectral centroid of each time sub-interval are calculated, and the frequency domain feature matrix of the denoised wind turbine power generation sound signal is constructed by the total wavelet energy and spectral centroid of each time sub-interval as the extraction result output.

[0075] In this embodiment, the sound signal of the wind turbine generating electricity under strong noise environment is measured at a sampling frequency f. s The discrete sequence of the wind turbine generator sound signal under high noise environment is represented by w(n), where n = 1, 2, ..., N, and N is the number of sampling points. For ease of testing, this embodiment uses a simulated wind turbine generator sound signal.

[0076] w(t)=cos(100πt)+0.5cos(200πt)+0.3cos(300πt)+0.2cos(400πt)

[0077] Simultaneously, multi-dimensional Gaussian white noise is added, with a sampling rate of f. s =1000Hz, obtaining the sound signal of wind turbine power generation under strong noise environment, such as Figure 2 As shown.

[0078] Step S101: Starting from the initial wavelet decomposition level q, the raw data of the wind turbine power generation sound signal is sequentially decomposed into wavelet decompositions of different levels q. The initial wavelet decomposition level q = 1. Then, for each q-level wavelet decomposition, the wavelet entropy weight η after one q-level wavelet decomposition is calculated.q If η q If the wavelet entropy weight is less than or equal to 5% of the set value (which can be adjusted according to actual needs), the wavelet coefficients can be considered to have been iteratively stabilized. The current wavelet decomposition level q is then taken as the optimal decomposition level, and the process jumps to step S102. Otherwise, the wavelet decomposition level q is incremented by 1, and the iteration continues (after performing q-level wavelet decomposition, the wavelet entropy weight η after q-level wavelet decomposition is calculated). q And the wavelet entropy weight η q (To make a judgment). In this embodiment, the wavelet entropy weight η for different wavelet decomposition levels q q The calculation results are shown in Table 1.

[0079] Table 1 Wavelet entropy weights η for different wavelet decomposition levels q q The calculation results.

[0080] <![CDATA[Wavelet entropy weight η q > 0.998 0.734 0.623 0.048 0.047

[0081] As shown in Table 1, the optimal decomposition level is 4. In this embodiment, wavelet decomposition is performed on the original data of the wind turbine power generation sound signal using an optimal decomposition level of 4. The wavelet coefficients obtained after decomposition are as follows: Figure 3 As shown, different colors represent wavelet coefficients of different layers.

[0082] Wavelet decomposition with q levels is a well-known method. The db wavelet function possesses excellent time and frequency localization properties, enabling fine-grained signal analysis in both time and frequency. Since only signals reconstructed by wavelet functions of db5 or higher exhibit first-order differentiability, and this method ensures both accurate frequency band division and real-time dynamic analysis requirements, the db5 wavelet is employed. The wavelet coefficients D after q-level wavelet decomposition can then be obtained. q,k The wavelet energy E at scale q q , can be represented as:

[0083]

[0084] In the above formula, q is also called the scaling factor, and k is the translation factor. ψ q,k (n) is the wavelet basis function.

[0085] Divide the wavelet coefficients after q-level decomposition into m sub-intervals, and calculate the relative wavelet energy P of the i-th sub-interval. q,i With wavelet entropy W q,i :

[0086]

[0087] In the above formula, E q,i Let be the wavelet energy of the i-th subinterval at scale q, where i = 0, 1, ..., m-1.

[0088] Therefore, the wavelet entropy W of the wavelet coefficients after q-level decomposition q for:

[0089]

[0090] In step S101 of this embodiment, the wavelet entropy weight η after q-layer wavelet decomposition is calculated. q The function expression is:

[0091]

[0092] In the above formula, W q Let be the wavelet entropy after q-level wavelet decomposition. The expression for the wavelet entropy after q-level wavelet decomposition is:

[0093]

[0094] In the above formula, m is the number of subintervals into which the wavelet coefficients are evenly divided after q-layer wavelet decomposition, and W q,i Let be the wavelet entropy of the i-th subinterval after q-level wavelet decomposition, and we have:

[0095]

[0096] In the above formula, P q,i Let E be the relative wavelet energy of the i-th subinterval. q,i E represents the wavelet energy of the i-th sub-interval after q-level wavelet decomposition, where N is the number of sampling points. q D represents the wavelet energy during q-level wavelet decomposition. q,k Let be the k-th wavelet coefficient after q-level wavelet decomposition, and we have:

[0097]

[0098] In the above formula, w(n) is the nth discrete signal of the wind turbine's sound signal, n = 1, 2, ..., N.

[0099] In step S102 of this embodiment, when calculating the lower and upper thresholds of wavelet coefficients at each layer using the given threshold value of wavelet entropy weight, the calculation of the lower and upper thresholds of any j-th layer wavelet coefficient using the given threshold value of wavelet entropy weight includes:

[0100] S201, based on the wavelet entropy weight η of the j-th layer j Determine the threshold adjustment factor β under the wavelet coefficients of the j-th layer. j,L and upper threshold adjustment factor β j,H :

[0101] If η j >a1, then:

[0102] β j,L =βj,Lmax ,β j,H =β j,Hmax ;

[0103] If a2 < η j If ≤a1, then:

[0104] β j,L =αβ j,Lmax ,β j,H =αβ j,Hmax ;

[0105] If a3 < η j If ≤a2, then:

[0106]

[0107] If η j If ≤a3, then all wavelet coefficients of the j-th layer are retained, where α is a parameter less than 1 (the value can be chosen according to actual needs; for example, as an optional implementation, the value is 0.9 in this embodiment), β j,Lmax and β j,Hmax Let be the maximum values ​​of the lower and upper adjustment factors, respectively, and we have:

[0108]

[0109] Where, μ j and s j Let D be the mean and standard deviation of the wavelet coefficients of the j-th layer. j,k Let a3 be the wavelet coefficient after j-level wavelet decomposition, a3 be the wavelet entropy weight setting value, and a1 and a2 be a set of threshold parameters arranged from largest to smallest within the range (100% to a3); max(|D j,k <0|) represents the absolute value of the maximum peak value of the negative half-axis of the wavelet coefficients in the j-th layer, max(|D j,k >0|) is the absolute value of the maximum peak value of the positive half-axis of the wavelet coefficients of the j-th layer;

[0110] S202, after obtaining the wavelet coefficients of the j-th layer, the upper threshold adjustment factor β j,L and β j,H Based on this, the lower and upper thresholds of the wavelet coefficients of the j-th layer are determined according to the following formula:

[0111] λ j,L =μ j -β j,L s j , λ j,H =μ j +β j,H s j ,

[0112] In the above formula, λ j,Lλ is the lower threshold of the wavelet coefficients at the j-th layer. j,H is the upper threshold of the wavelet coefficients of the j-th layer.

[0113] The commonly used fixed threshold method has poor adaptability, and the fixed threshold becomes significantly larger as the signal length N increases. In step S102 of this embodiment, when calculating the lower and upper thresholds of wavelet coefficients at each layer using the given threshold value of wavelet entropy weight, adaptive threshold processing of wavelet coefficients at each layer is achieved through the above steps S201 and S202, which is beneficial to improve the adaptability of the threshold and reduce the threshold deviation. It should be noted that the threshold parameters a1 and a2 can be set according to actual needs. For example, as an optional implementation, a1 is set to 95% and a2 is set to 70% in this embodiment. Combined with the wavelet entropy weight setting value a3 set to 5% mentioned above, we have: (1) When η j When the percentage is >95%, the wavelet coefficients of this layer contain only noise information, β j,L =β j,Lmax ,β j,H =β j,Hmax (2) When 70% < η j When ≤95%, the wavelet coefficients of this layer contain a large amount of noise information and a small amount of effective wind turbine power generation sound information. In order to fully retain the small amount of useful information, β j,L and β j,H Slightly less than the maximum value, i.e., the adjustment factor β j,L =αβ j,Lmax ,β j,H =αβ j,Hmax ,α=0.9;(3)When 5%<η j When the efficiency is ≤70%, the wavelet coefficients of this layer contain a large amount of effective information about wind turbine power generation sound and a small amount of noise information. The adjustment factor is:

[0114]

[0115] (4) Because when η j If the value is ≤5%, the optimal wavelet decomposition level is satisfied. Therefore, it is considered that the wavelet coefficients of this level only contain effective information about the sound of wind turbine power generation, and the wavelet coefficients of this level should be retained.

[0116] In step S103 of this embodiment, when calculating the wavelet estimation coefficients of each layer by combining the lower and upper thresholds of the wavelet coefficients of each layer, the calculation function expression for the wavelet estimation coefficient of any j-th layer is as follows:

[0117]

[0118] In the above formula, Let D be the estimated coefficient of the k-th wavelet in the j-th layer, sign be the sign function, and D be the value of D. j,kLet λ be the k-th wavelet coefficient after j-level wavelet decomposition. j,L λ is the lower threshold of the wavelet coefficients at the j-th layer. j,H λ is the upper threshold of the wavelet coefficients at the j-th layer, M is the smoothing factor, and λ is the lower threshold. j,avg The lower threshold λ of the wavelet coefficients of the j-th layer j,L Upper threshold λ j,H The absolute mean between them, that is:

[0119]

[0120] Traditional soft and hard threshold functions are difficult to accurately estimate wavelet coefficients. In this embodiment, the key to the calculation function expression of wavelet estimation coefficients at any j-th layer lies in the addition of the following adjustment amount:

[0121]

[0122] This adjustment utilizes the approximation property of the exponential function and its rapid change within a fixed interval to achieve accurate estimation of wavelet coefficients. M is a smoothing factor, typically set to M=2, which controls the approximation speed of the threshold function. Increasing M makes the denoised wind turbine sound signal smoother, while decreasing M retains more effective information from the wind turbine sound signal.

[0123] This embodiment constructs an adaptive threshold and threshold function. Based on wavelet entropy weighting, it assigns values ​​to adaptive threshold adjustment factors for different decomposition levels. Furthermore, by improving the threshold function under the premise of the adaptive threshold and adding an adjustable smoothing factor, more accurate wavelet estimation coefficients are obtained, enabling the extraction of effective sound signals from wind turbine power generation in high-noise environments. The original wavelet coefficients in this embodiment are as follows: Figures 4-7 As shown, where Figure 4 These are the original wavelet coefficients of the first layer. Figure 5 These are the original wavelet coefficients of the second layer. Figure 6 These are the original wavelet coefficients of the third layer. Figure 7 These are the original wavelet coefficients of the 4th layer. The wavelet estimated coefficients after adaptive thresholding and thresholding function are as follows: Figures 8-11 As shown, where Figure 8 The coefficients are estimated for the first layer of wavelet. Figure 9 For the second layer wavelet estimation coefficients, Figure 10 For the third layer wavelet estimation coefficients, Figure 11 These are the estimated coefficients for the fourth layer of wavelets.

[0124] In step S104 of this embodiment, the function expression for reconstructing the denoised wind turbine generator sound signal using the wavelet estimation coefficients of each layer is as follows:

[0125]

[0126] In the above formula, This represents the nth denoised sound signal from the wind turbine. Let ψ(2) be the wavelet estimation coefficient of the j-th layer, and let ψ(2) be the wavelet estimation coefficient of the k-th layer. -j (nk) represents the k-th wavelet function of the j-th layer, where n is the sampling point index, k is the index of the wavelet estimation coefficient, N is the number of sampling points, and j = 1, 2, ..., J op J op This represents the optimal number of decomposition layers.

[0127] In step S105 of this embodiment, a definition of the frequency domain feature matrix of the denoised wind turbine generator sound signal is proposed. Based on the wavelet estimation coefficients after denoising, different time-series segments are performed, and the wavelet energy and spectral centroid of each segment signal are calculated as the overall frequency domain feature of the denoised wind turbine generator sound signal. Based on this, the frequency domain feature matrix of the denoised wind turbine generator sound signal is constructed, thereby obtaining the time series of the frequency domain feature matrix of the denoised wind turbine generator sound signal, thus realizing the identification and comparison of its dynamic change characteristics. In step S105, the wavelet estimation coefficients of each layer in the denoised wind turbine generator sound signal are uniformly divided into r time sub-intervals. Then, the wavelet energy E at the wavelet scale j in the i-th time sub-interval can be calculated. j,i :

[0128]

[0129] It should be noted that the number of time sub-intervals r can be chosen as needed. As an optional implementation, in step S105 of this embodiment, the wavelet estimation coefficients of each layer in the denoised wind turbine power generation sound signal are uniformly divided into 10 time sub-intervals. In step S105 of this embodiment, when calculating the total wavelet energy and spectral centroid of each time sub-interval, the calculation function expression for the total wavelet energy and spectral centroid of any i-th time sub-interval is:

[0130]

[0131]

[0132] In the above formula, E i Let SC be the total wavelet energy of the i-th time sub-interval. i Let f be the spectral centroid of the i-th time subinterval. s Sampling frequency, f is the estimated coefficient of the k-th wavelet in the j-th layer. s / 2 j The corresponding frequency under the scale factor j.

[0133] The functional expression of the frequency domain feature matrix of the denoised wind turbine power generation sound signal constructed in step S105 of this embodiment is as follows:

[0134]

[0135] In the above formula, G is the frequency domain characteristic matrix, E1~E r The total wavelet energy of the 1st to rth time sub-intervals is SC1 to SC2. r For the spectral centroids of the 1st to rth time sub-intervals, the frequency domain characteristic matrix can also be expressed as:

[0136]

[0137] Specifically, the frequency domain feature matrix obtained in this embodiment is:

[0138]

[0139] The results show that the time series of the frequency domain feature matrix obtained by the signal dynamic change feature extraction method under strong noise environment in this embodiment can accurately extract the dynamic features of the denoised wind turbine power generation sound signal, thus enabling effective diagnosis of wind turbine fault signals. For example, as an optional implementation, the frequency domain feature matrix can be used as input to a pre-trained machine learning classification model (e.g., using a Long Short-Term Memory network LSTM) (either alone or in combination with other parameter features of the wind turbine), thereby classifying whether the wind turbine is in a normal or faulty state.

[0140] In summary, the method for extracting dynamic features of signals under strong noise conditions in this embodiment is based on the quantitative determination of the optimal wavelet decomposition level of the wind turbine generator sound signal under strong noise conditions using wavelet entropy weights. This achieves sufficient removal of noise signals while retaining the effective sound signal of the wind turbine generator. This embodiment constructs an adaptive threshold and threshold function. Using wavelet entropy weights as a benchmark, it provides values ​​for adaptive threshold adjustment factors for different decomposition levels. By improving the threshold function under the premise of adaptive thresholds and adding an adjustable smoothing factor, more accurate wavelet estimation coefficients are obtained, thus realizing the extraction of the effective sound signal of the wind turbine generator under strong noise conditions. This embodiment proposes the definition of the frequency domain feature matrix of the denoised wind turbine generator sound signal. Based on the denoised wavelet estimation coefficients, different time-series segments are performed, and the wavelet energy and spectral centroid of each segment are calculated as the overall frequency domain features of the denoised wind turbine generator sound signal. Based on this, the frequency domain feature matrix of the denoised wind turbine generator sound signal is constructed, thereby obtaining the time series of the frequency domain feature matrix of the denoised wind turbine generator sound signal, thus realizing the identification and comparison of its dynamic change features. Based on the improvements in the above three aspects, the method for extracting dynamic features of signals under strong noise environment in this embodiment can effectively eliminate the interference of background noise on the sound signal of wind turbine power generation, and realize the accurate and effective extraction of the dynamic features of the sound signal of wind turbine power generation with strong background noise.

[0141] Furthermore, this embodiment also provides a device for extracting dynamic signal change features under strong noise conditions, including a microprocessor and a memory interconnected thereto. The microprocessor is programmed or configured to execute the method for extracting dynamic signal change features under strong noise conditions. This embodiment also provides a computer-readable storage medium storing a computer program for being programmed or configured by the microprocessor to execute the method for extracting dynamic signal change features under strong noise conditions.

[0142] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0143] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for extracting dynamic change features of signals under strong noise environment, characterized in that, Includes the following steps: S101, for the collected wind turbine power generation sound signal under strong noise environment, from the initial wavelet decomposition level... q The raw data of the wind turbine's acoustic signal were then processed sequentially at different wavelet decomposition levels. q Wavelet decomposition and calculation q Wavelet entropy weights after layer wavelet decomposition η q ,like η q If the value is less than or equal to the wavelet entropy weight setting, then the current wavelet decomposition level will be adjusted. q As the optimal number of decomposition levels, proceed to step S102; S102, using wavelet entropy weights to calculate the lower and upper thresholds of wavelet coefficients for each layer under the optimal decomposition layer number, given the threshold value; S103, calculate the wavelet estimation coefficients of each layer by combining the lower and upper thresholds of the wavelet coefficients of each layer; S104, the denoised wind turbine power generation sound signal is reconstructed using the wavelet estimation coefficients of each layer under the optimal decomposition layer number; S105, the wavelet estimation coefficients of each layer in the denoised wind turbine power generation sound signal are evenly divided into r time sub-intervals, the total wavelet energy and spectral centroid of each time sub-interval are calculated respectively, and the frequency domain feature matrix of the denoised wind turbine power generation sound signal is constructed by the total wavelet energy and spectral centroid of each time sub-interval as the extraction result output. The calculation in step S101 q Wavelet entropy weight after wavelet decomposition of layer η q The function expression of the above formula is: ， In the above formula, W q To q Wavelet entropy after wavelet decomposition of the layer When calculating the lower and upper thresholds of wavelet coefficients at each level using the wavelet entropy weight given threshold value in step S102, the calculation of the lower and upper thresholds of any j-th level wavelet coefficient using the wavelet entropy weight given threshold value includes: S201, based on the wavelet entropy weight of the j-th layer η j Determine the first j Threshold adjustment factor under layer wavelet coefficients and upper threshold adjustment factor : like Then we have: ， ； like Then we have: ， ； If then there is: ， like Then all wavelet coefficients of the j-th layer are retained, where For parameters less than 1, and Let be the maximum values ​​of the lower and upper adjustment factors, respectively, and we have: ， ， in, and For the first j Mean and standard deviation of layer wavelet coefficients for j The k-th wavelet coefficient after layer wavelet decomposition. Set values ​​for wavelet entropy weights. and For (100%~ A set of threshold parameters arranged from largest to smallest within a range; S202, after obtaining the first j Lower and upper threshold adjustment factors of layer wavelet coefficients and Based on this, the lower and upper thresholds of the wavelet coefficients of the j-th layer are determined according to the following formula: ， ， In the above formula, The lower threshold of the wavelet coefficients of the j-th layer. is the upper threshold of the wavelet coefficients of the j-th layer.

2. The method according to claim 1, wherein, The q The expression for the wavelet entropy calculation function after layer wavelet decomposition is: ， In the above formula, m for q The number of subintervals into which the wavelet coefficients are evenly divided after layer wavelet decomposition. for q The wavelet entropy of the i-th subinterval after layer wavelet decomposition is given by: ， ， ， In the above formula, For the first i The relative wavelet energy of each sub-interval, for q The wavelet energy of the i-th subinterval after layer wavelet decomposition. The number of sampling points. for q Wavelet energy during layer wavelet decomposition for q The k-th wavelet coefficients after layer wavelet decomposition are: ， In the above formula, is the n-th discrete signal of the wind turbine sound signal, .

3. The method according to claim 1, wherein, In step S103, when calculating the wavelet estimation coefficients of each layer by combining the lower and upper thresholds of the wavelet coefficients of each layer, the calculation function expression for the wavelet estimation coefficient of any j-th layer is: In the above formula, Here, represents the estimated coefficient of the k-th wavelet in the j-th layer, and sign is the sign function. for j The k-th wavelet coefficient after layer wavelet decomposition. The lower threshold of the wavelet coefficients at the j-th layer is given by [the threshold value]. The upper threshold of the wavelet coefficients of the j-th layer is... The lower threshold of the wavelet coefficients of the j-th layer Upper threshold The absolute mean between This is a smoothing factor.

4. The method for extracting dynamic change features of signals under strong noise environment according to claim 1, characterized in that, In step S104, the function expression for reconstructing the denoised wind turbine generator sound signal using the wavelet estimation coefficients of each layer is as follows: ， In the above formula, This represents the nth denoised sound signal from the wind turbine. For the estimated coefficients of the k-th wavelet in the j-th layer, This represents the k-th wavelet function of the j-th layer. The sampling point number, Here are the indexes of the wavelet-estimated coefficients. This represents the number of sampling points.

5. The method of claim 1, wherein, In step S105, when calculating the total wavelet energy and spectral centroid of each time sub-interval, the function expression for calculating the total wavelet energy and spectral centroid of any i-th time sub-interval is: ， ， In the above formula, is the total wavelet energy of the i-th time sub-interval, is the spectral centroid of the i-th time sub-interval, is the sampling frequency, is the k-th wavelet estimation coefficient of the j-th layer.

6. The method of claim 1, wherein, The functional expression of the frequency domain feature matrix of the denoised wind turbine power generation sound signal constructed in step S105 is as follows: ， In the above formula, is a frequency domain feature matrix, is the wavelet total energy of the 1st to rth time sub-interval, is the spectral centroid of the 1st to rth time sub-interval.​​ 7. A device for extracting dynamic change features of signals under strong noise environment, comprising a microprocessor and a memory interconnected, characterized in that, The microprocessor is programmed or configured to perform the signal dynamic change feature extraction method under strong noise environment as described in any one of claims 1 to 6.

8. A computer-readable storage medium having stored therein a computer program, characterized in that, The computer program is used to be programmed or configured by a microprocessor to execute the signal dynamic change feature extraction method under strong noise environment as described in any one of claims 1 to 6.

Images