Acoustic signal separation and feature extraction method and system for main speed reducer

Abstract

The invention discloses an acoustic signal separation and feature extraction method and system for a main speed reducer. The method comprises the following steps: collecting acoustic signals of the main speed reducer; carrying out data sparse blind source preprocessing on the collected signal to obtain a de-mixing signal; partial coherence analysis is carried out on the unmixed signal, a main noise source is judged, and an analyzed signal is obtained; performing cepstrum coefficient calculation on the analyzed signal through a Mel filter and extracting acoustic signal features; noise existing in the acoustic signal acquired from the main speed reducer is avoided, and the feature extraction efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of main reducer fault analysis, and more specifically, to a method and system for separating and extracting acoustic signals of a main reducer. Background Art

[0002] The acoustic test of the main reducer is usually carried out under a strong noise background. The original signal collected by the sensor contains other external signals. It is difficult to separate the information useful for fault diagnosis from the weak signal state characteristic signal by directly extracting features from the original signal. Therefore, the original signal collected by the sensor needs to be analyzed and processed to remove the interference items in the original signal, and then the fault feature analysis and selection are carried out.

[0003] At present, the noise reduction measures for the main reducer are usually Fourier transform, wavelet transform and other measures, but there are defects when processing non-stationary signals. These measures can only obtain a part of the frequency components contained in a signal as a whole, and do not know the corresponding moment. Therefore, two signals with a large difference in time domain may have the same spectrum.

[0004] The main deficiencies of the existing main reducer noise reduction technology are as follows: the use of wavelet transform for noise reduction is not ideal when the frequency bands of useful signals and noise overlap (such as signals mixed with white noise); when extracting features from vibration signals, the use of deep belief networks makes the system complex and reduces efficiency, increasing the uncertainty of the system. Summary of the invention

[0005] In order to solve the problems in the prior art that the acoustic signals collected from the main reducer contain noise and the feature extraction efficiency is low, a method for separating and extracting acoustic signals of the main reducer and a system thereof are proposed to avoid noise in the acoustic signals collected from the main reducer and improve the efficiency of feature extraction.

[0006] The technical solution of the present invention is as follows:

[0007] On one hand, the present invention discloses a method for separating and extracting acoustic signals of a main reducer, comprising the following steps:

[0008] Collect the acoustic signal of the main reducer;

[0009] Perform data sparse blind source preprocessing on the collected signal to obtain a demixed signal;

[0010] Perform partial coherence analysis on the unmixed signal to determine the main noise source and obtain the analyzed signal;

[0011] The Mel filter is used to calculate the inverse spectrum coefficients of the analyzed signal and extract the acoustic signal features.

[0012] Preferably, the acoustic signal of the main reducer is collected by adopting a far-field coverage collection method and a near-field precise collection method;

[0013] The far-field coverage acquisition method and the near-field precise acquisition method both use a five-point method to collect acoustic signals.

[0014] Preferably, the data coefficient blind source preprocessing is to perform blind source separation on the mixed signal, which is expressed as the following linear mixed model:

[0015] Y=AX+V

[0016] Among them, Y and X are the observation matrix and source matrix respectively, and V is Gaussian noise;

[0017] According to the principle of blind source separation, the linear mixture model is optimized:

[0018]

[0019] Among them, ||·|| F is the F norm, A and X are unknown matrices; use the known observation matrix Y to input into the algorithm to predict X and A;

[0020] The sparse representation based on the analytical model is expressed as:

[0021]

[0022] The above formula is further optimized into Lagrangian form:

[0023]

[0024] The blind source separation process based on the analytical sparse model is expressed as the following joint optimization problem:

[0025]

[0026] The scalars λ and μ control the noise power and sparsity respectively; the dictionary, source matrix, and mixing matrix are estimated alternately in three stages:

[0027] Dictionary learning stage:

[0028]

[0029] Source estimation phase:

[0030]

[0031] Confusion matrix estimation stage:

[0032]

[0033] The demixed signal is obtained through the above three stages.

[0034] Furthermore, in the dictionary learning stage, by fixing A to estimate the dictionary Ω, the joint optimization problem in the dictionary learning stage is expressed as:

[0035]

[0036] For each signal, a corresponding dictionary needs to be learned. K-SVD is used to analyze the joint optimization problem of the above dictionary learning stage, which can be expressed as:

[0037]

[0038] Furthermore, in the source estimation stage, the source x is estimated by solving a convex optimization problem i ;

[0039] The convex optimization problem is expressed as:

[0040]

[0041] Among them, Ω i and i,f Obtained during the dictionary learning phase; the solution is obtained by setting the gradient of the convex optimization problem to 0:

[0042]

[0043] Where I is the identity matrix.

[0044] Furthermore, in the confusion matrix estimation stage, all unknowns except A are fixed, and the confusion matrix estimation problem is simplified to:

[0045]

[0046] The residual of xi is defined as This minimization problem is solved by setting its gradient to 0, yielding:

[0047]

[0048] In order to maintain the column norm, the mixing matrix A needs to be normalized after each update.

[0049] Preferably, the partial coherence analysis is specifically as follows:

[0050] The unmixed signal is framed and windowed, and the framed signal is transformed through Fourier transform to obtain the spectrum, from which the power spectral density function can be obtained:

[0051]

[0052] Where R x(τ) is the autocorrelation function of x(t), τ is the time difference, T is the observation time; and the correlation function R x (τ) is defined as:

[0053]

[0054] The coherence function represents the correlation between two random signals in the frequency domain, and converts the cross-correlation function in the time domain into the cross-spectral density function in the frequency domain;

[0055]

[0056] In the formula, S xy (f) is the cross-spectral density function of the random signal, x(t) and y(t);

[0057] The expression of the coherence function is:

[0058]

[0059] Where 0≤γ 2 xy (f)≤1,S xx (f) and S yy (f) are the autospectral density functions of x(t) and y(t), respectively.

[0060] Preferably, the specific steps of calculating the cepstral coefficients and extracting the acoustic signal features through the Mel filter are:

[0061] To select the center frequency:

[0062] Set several bandpass filters H within the signal spectrum m (k), 1≤m≤M, M is the number of filters; each filter has a triangular filtering characteristic, where the center frequency is f(m), 1≤m≤M; within the Mel frequency range, these filters have equal width:

[0063]

[0064] Where f l is the lowest frequency in the filter frequency range, f h is the highest frequency in the filter frequency range, N is the Fourier length, f s is the sampling frequency;

[0065] Fmel's inverse function:

[0066]

[0067] When M = n, f l , f h , f s, N, M values ​​are fixed after determination, and F on the Mel frequency scale is mel (f l ) to F mel (f h ) is divided into M+1 segments, generating M+2 Mel frequency values; wherein n is an arbitrarily determined positive integer;

[0068] The actual frequency value corresponding to the Mel frequency is:

[0069]

[0070] The first set of data for the features is obtained:

[0071] feat(i,j)=J(i,j)·K(j)

[0072] Get the second and third sets of data of Mel frequency cepstral coefficients. The second set of parameters is the first-order differential operation based on the first set of basic parameters, and the third set of parameters is the first-order differential operation based on the second set of parameters:

[0073] dtfeat[i][j]=feat[i+1][j]-feat[i-1][j]+2feat[i+2][j]-feat[i-2][j]

[0074] The dynamic characteristics of the audio are reflected by the data of the Mel-frequency cepstral coefficients, and the acoustic signal characteristics are completely extracted.

[0075] On the other hand, the present invention discloses an acoustic signal separation and feature extraction system for a main reducer, which adopts the acoustic signal separation and feature extraction method for a main reducer, and comprises: a signal acquisition module, a signal separation module, a partial coherence module, and a feature extraction module;

[0076] The signal acquisition module is used to collect the acoustic signal of the main reducer;

[0077] The signal separation module is used to perform data sparse blind source preprocessing on the collected signal to obtain an unmixed signal;

[0078] The partial coherence module performs partial coherence analysis on the unmixed signal, determines the main noise source, and obtains the analyzed signal;

[0079] The feature extraction module calculates the cepstral coefficients through a Mel filter and extracts the acoustic signal features.

[0080] Furthermore, the signal acquisition module includes two microphones, a microphone array, and a sound intensity probe;

[0081] The two microphones are in phase with each other and are arranged face to face with a gap between them. When the sound wave is transmitted along the normal plane of the microphone, there is a gradient between the two sound pressure measurement values, and the sound intensity value of the measurement point is obtained according to the imaginary part of the cross spectrum of the two microphones.

[0082] The microphone array is composed of a group of microphones arranged in a geometric structure, which simultaneously receives signals from different spatial directions, and determines the position of the main reducer noise source based on the phased array principle by comparing the phase differences of the noise signals reaching each microphone;

[0083] The sound intensity probe is perpendicular to the envelope surface and is P cm away from the measurement surface. The sound intensity measurement result of the center point of the grid is processed to obtain the sound intensity distribution cloud map of each frequency band of each measurement surface.

[0084] Beneficial effects of the present invention:

[0085] 1. The noise reduction method of data sparse blind source preprocessing is used to eliminate the interference of noise signals on acoustic signals.

[0086] 2. Through partial coherence analysis, the fault analysis of the signal is made more accurate, and the efficiency of feature extraction is improved by calculating the inverse spectrum coefficients through the Mel filter. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] Figure 1 A flowchart of an acoustic signal separation and feature extraction method for a main reducer is provided in an embodiment.

[0088] Figure 2 A schematic diagram of a far-field coverage acquisition method provided in an embodiment.

[0089] Figure 3 A schematic diagram of a near-field precise acquisition method provided in an embodiment. DETAILED DESCRIPTION

[0090] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0091] Example 1

[0092] In this embodiment, if Figure 1 As shown, the following steps are included:

[0093] Collect the acoustic signal of the main reducer;

[0094] Perform data sparse blind source preprocessing on the collected signal to obtain a demixed signal;

[0095] Perform partial coherence analysis on the unmixed signal to determine the main noise source and obtain the analyzed signal;

[0096] The Mel filter is used to calculate the inverse spectrum coefficients of the analyzed signal and extract the acoustic signal features.

[0097] In this embodiment, the far-field coverage acquisition method and the near-field precise acquisition method are used to collect the acoustic signal of the main reducer;

[0098] The far-field coverage acquisition method and the near-field precise acquisition method both use a five-point method to collect acoustic signals.

[0099] In this embodiment, if Figure 2 As shown, a far-field coverage acquisition device is set up, and five microphone measuring points are arranged around the experimental bench, and the distance between each measuring point and the experimental bench is 1m.

[0100] like Figure 3 As shown, a near-field precise acquisition device is set up to further collect the acoustic signal of the main reducer. Since the sound propagates in the air with the characteristics of being larger near and smaller far away, the acoustic signal collected by the near-field method is basically the acoustic signal of the main reducer. When collecting by the near-field precise method, a total of five microphone measurement points are arranged around the main reducer, and the distance between each measurement point and the main reducer is 0.3m.

[0101] The blind source preprocessing of data coefficients is to perform blind source separation on the mixed signal, which is expressed as the following linear mixed model:

[0102] Y=AX+V

[0103] Among them, Y and X are the observation matrix and source matrix respectively, and V is Gaussian noise;

[0104] According to the principle of blind source separation, the linear mixture model is optimized:

[0105]

[0106] Among them, ||·|| F is the F norm, A and X are unknown matrices; use the known observation matrix Y to input into the algorithm to predict X and A;

[0107] The sparse representation based on the analytical model is expressed as:

[0108]

[0109] The above formula is further optimized into Lagrangian form:

[0110]

[0111] The blind source separation process based on the analytical sparse model is expressed as the following joint optimization problem:

[0112]

[0113] The scalars λ and μ control the noise power and sparsity respectively; the dictionary, source matrix, and mixing matrix are estimated alternately in three stages:

[0114] Dictionary learning stage:

[0115]

[0116] Source estimation phase:

[0117]

[0118] Confusion matrix estimation stage:

[0119]

[0120] The demixed signal is obtained through the above three stages.

[0121] More specifically, in the dictionary learning stage, by fixing A to estimate the dictionary Ω, the joint optimization problem in the dictionary learning stage is expressed as:

[0122]

[0123] For each signal, a corresponding dictionary needs to be learned. K-SVD is used to analyze the joint optimization problem of the above dictionary learning stage, which can be expressed as:

[0124]

[0125] More specifically, in the source estimation stage, the source x is estimated by solving a convex optimization problem i ;

[0126] The convex optimization problem is expressed as:

[0127]

[0128] Among them, Ω i and i,f Obtained during the dictionary learning phase; the solution is obtained by setting the gradient of the convex optimization problem to 0:

[0129]

[0130] Where I is the identity matrix.

[0131] In the confusion matrix estimation stage, all unknowns except A are fixed, and the confusion matrix estimation problem is simplified to:

[0132]

[0133] The residual of xi is defined as This minimization problem is solved by setting its gradient to 0, yielding:

[0134]

[0135] In order to maintain the column norm, the mixing matrix A needs to be normalized after each update.

[0136] Example 2

[0137] In this embodiment, the partial coherence analysis is specifically as follows:

[0138] The unmixed signal is framed and windowed, and the framed signal is transformed through Fourier transform to obtain the spectrum, from which the power spectral density function can be obtained:

[0139]

[0140] Where R x (τ) is the autocorrelation function of x(t), τ is the time difference, T is the observation time; and the correlation function R x (τ) is defined as:

[0141]

[0142] The coherence function represents the correlation between two random signals in the frequency domain, and converts the cross-correlation function in the time domain into the cross-spectral density function in the frequency domain;

[0143]

[0144] In the formula, S xy (f) is the cross-spectral density function of random signal, x(t) and y(t);

[0145] The expression of the coherence function is:

[0146]

[0147] Where 0≤γ 2 xy (f)≤1,S xx (f) and S yy (f) are the autospectral density functions of x(t) and y(t), respectively.

[0148] The coherence function value is calculated using the autospectral density and cross-spectral density. The noise source and the effective signal source in the unmixed signal are distinguished by the coherence function, and the Mel-cepstral coefficients are processed for the effective signal source.

[0149] In this embodiment, the specific steps of calculating the cepstral coefficients and extracting the acoustic signal features through the Mel filter are:

[0150] To select the center frequency:

[0151] Set several bandpass filters H within the signal spectrum m (k), 1≤m≤M, M is the number of filters; each filter has a triangular filtering characteristic, where the center frequency is f(m), 1≤m≤M; within the Mel frequency range, these filters have equal width:

[0152]

[0153] Where f l is the lowest frequency in the filter frequency range, f h is the highest frequency in the filter frequency range, N is the Fourier length, f s is the sampling frequency; Mel frequency refers to the transformation to the Mel frequency scale, and the transformation formula is:

[0154] Fmel(f)=2595*log10(1+700f);

[0155] Fmel's inverse function:

[0156]

[0157] When M = n, f l , f h , f s , N, M values ​​are fixed after determination, and F on the Mel frequency scale is mel (f l ) to F mel (f h ) is divided into M+1 segments, generating M+2 Mel frequency values; wherein n is an arbitrarily determined positive integer;

[0158] The actual frequency value corresponding to the Mel frequency is:

[0159]

[0160] The first set of data for the features is obtained:

[0161] feat(i,j)=J(i,j)·K(j)

[0162] Get the second and third sets of data of Mel frequency cepstral coefficients. The second set of parameters is the first-order differential operation based on the first set of basic parameters, and the third set of parameters is the first-order differential operation based on the second set of parameters:

[0163] dtfeat[i][j]=feat[i+1][j]-feat[i-1][j]+2feat[i+2][j]-feat[i-2][j]

[0164] The dynamic characteristics of the audio are reflected by the data of the Mel-frequency cepstral coefficients, and the acoustic signal characteristics are completely extracted.

[0165] Because the Mel frequency cepstral coefficient reflects the static characteristics of the audio, the first-order difference and second-order difference are calculated using the difference equation to reflect the dynamic characteristics of the audio, making the speech feature extraction more complete.

[0166] Example 3

[0167] An acoustic signal separation and feature extraction system for a main reducer, using an acoustic signal separation and feature extraction method for a main reducer as described in embodiments 1 and 2, comprising: a signal acquisition module, a signal separation module, a partial coherence module, and a feature extraction module;

[0168] The signal acquisition module is used to collect the acoustic signal of the main reducer; non-contact acoustic testing technology is adopted; the far-field acoustic signal will contain a large amount of noise signal during the acquisition process, which may cause the signal of the main reducer to be completely covered by the noise signal, and the acoustic signal of the main reducer is collected through the near-field precise acquisition method.

[0169] The signal separation module is used to perform data sparse blind source preprocessing on the collected signal to obtain an unmixed signal;

[0170] The partial coherence module performs partial coherence analysis on the unmixed signal, determines the main noise source, and obtains the analyzed signal; converts the multi-input single-output system into a multi-input single-output system with conditional input, and calculates the impact of different parts of the acoustic signal on the output.

[0171] The feature extraction module calculates the inverse spectrum coefficients through the Mel filter and extracts the acoustic signal features; the Mel spectrum is subjected to discrete cosine transformation to remove the correlation between the signals of each dimension, and then the recognition performance is improved by adding differential parameters.

[0172] The signal acquisition module includes two microphones, a microphone array, and a sound intensity probe.

[0173] The two microphones are in phase with each other and are arranged face to face with a small gap between them. When the sound wave is transmitted along the normal plane of the microphone, there is a gradient between the two sound pressure measurement values. The sound intensity value of the measurement point is obtained according to the imaginary part of the cross spectrum of the two microphones.

[0174] The microphone array is composed of a group of microphones arranged in a geometric structure, which simultaneously receive signals from different spatial directions, and determine the position of the main reducer noise source based on the phased array principle by comparing the phase differences of the noise signals reaching each microphone.

[0175] The sound intensity probe is perpendicular to the envelope surface and 50 cm away from the measurement surface. The measurement personnel are away from the axis position of the probe. The test is carried out after reaching the required working conditions and running stably. The sound intensity measurement results of the center point of the grid are processed to obtain the sound intensity distribution cloud map of each frequency band of each measurement surface.

[0176] Obviously, the above embodiments of the present invention are only examples for clearly explaining the present invention, and are not intended to limit the implementation methods of the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A method for separating and extracting acoustic signals of a main reducer, characterized in that: The following steps are involved: Collect the acoustic signal of the main reducer; Perform data sparse blind source preprocessing on the collected signal to obtain a demixed signal; Perform partial coherence analysis on the unmixed signal to determine the main noise source and obtain the analyzed signal; The Mel filter is used to calculate the inverse spectrum coefficients of the analyzed signal and extract the acoustic signal features.

2. The method for separating and extracting acoustic signals of a main reducer according to claim 1, characterized in that: The acoustic signals of the main reducer are collected using the far-field coverage collection method and the near-field precise collection method; The far-field coverage acquisition method and the near-field precise acquisition method both use a five-point method to collect acoustic signals.

3. The method for separating and extracting acoustic signals of a main reducer according to claim 1, characterized in that: Data sparse blind source preprocessing is to perform blind source separation on mixed signals, which can be expressed as the following linear mixed model: Y=AX+V Among them, Y and X are the observation matrix and source matrix respectively, and V is Gaussian noise; According to the principle of blind source separation, the linear mixture model is optimized: Among them, ||·|| F is the F norm, A and X are unknown matrices; use the known observation matrix Y to input into the algorithm to predict X and A; The sparse representation based on the analytical model is expressed as: The above formula is further optimized into Lagrangian form: The blind source separation process based on the analytical sparse model is expressed as the following joint optimization problem: The scalars λ and μ control the noise power and sparsity respectively; the dictionary, source matrix, and mixing matrix are estimated alternately in three stages: Dictionary learning stage: Source estimation phase: Confusion matrix estimation stage: The demixed signal is obtained through the above three stages.

4. The method for separating and extracting acoustic signals of a main reducer according to claim 3, characterized in that: In the dictionary learning stage, the dictionary Ω is estimated by fixing A. The joint optimization problem of the dictionary learning stage is expressed as: For each signal, a corresponding dictionary needs to be learned. K-SVD is used to analyze the joint optimization problem of the above dictionary learning stage, which can be expressed as:

5. The method for separating and extracting acoustic signals of a main reducer according to claim 4, characterized in that: In the source estimation stage, the source xi is estimated by solving a convex optimization problem; The convex optimization problem is expressed as: Among them, Ω i and i,f Obtained during the dictionary learning phase; the solution is obtained by setting the gradient of the convex optimization problem to 0: Where I is the identity matrix.

6. The method for separating and extracting acoustic signals of a main reducer according to claim 5, characterized in that: In the confusion matrix estimation stage, all unknowns except A are fixed, and the confusion matrix estimation problem is simplified to: The residual of xi is defined as This minimization problem is solved by setting its gradient to 0, yielding: In order to maintain the column norm, the mixing matrix A needs to be normalized after each update.

7. The method for separating and extracting acoustic signals of a main reducer according to claim 1, characterized in that: The partial coherence analysis is as follows: The unmixed signal is framed and windowed, and the framed signal is transformed through Fourier transform to obtain the spectrum, from which the power spectral density function can be obtained: Where R x (τ) is the autocorrelation function of x(t), τ is the time difference, T is the observation time; and the correlation function R x (τ) is defined as: The coherence function represents the correlation between two random signals in the frequency domain, and converts the cross-correlation function in the time domain into the cross-spectral density function in the frequency domain; In the formula, S xy (f) is the cross-spectral density function of the random signal, x(t) and y(t); The expression of the coherence function is: Where 0≤γ 2 xy (f)≤1,S xx (f) and S yy (f) are the autospectral density functions of x(t) and y(t), respectively.

8. The method for separating and extracting acoustic signals of a main reducer according to claim 1, characterized in that: The specific steps of calculating the cepstral coefficients of the analyzed signal through the Mel filter and extracting the acoustic signal features are: To select the center frequency: Set several bandpass filters H within the signal spectrum m (k), 1≤m≤M, M is the number of filters; each filter has a triangular filtering characteristic, where the center frequency is f(m), 1≤m≤M; In the Mel frequency range, these filters are of equal width: Where f l is the lowest frequency in the filter frequency range, f h is the highest frequency in the filter frequency range, N is the Fourier length, f s is the sampling frequency; Fmel's inverse function: When M = n, f l , f h , f s , N, M values ​​are fixed after determination, and F on the Mel frequency scale is mel (f l ) to F mel (f h ) is divided into M+1 segments, generating M+2 Mel frequency values; wherein n is an arbitrarily determined positive integer; The actual frequency value corresponding to the Mel frequency is: The first set of data for the features is obtained: feat(i,j)=J(i,j)·K(j) Get the second and third sets of data of Mel frequency cepstral coefficients. The second set of parameters is the first-order differential operation based on the first set of basic parameters, and the third set of parameters is the first-order differential operation based on the second set of parameters: dtfeat[i][j]=feat[i+1][j]-feat[i-1][j]+2feat[i+2][j]-feat[i-2][j] The dynamic characteristics of the audio are reflected by the data of the Mel-frequency cepstral coefficients, and the acoustic signal characteristics are completely extracted.

9. An acoustic signal separation and feature extraction system for a main reducer, using the acoustic signal separation and feature extraction method for a main reducer as claimed in any one of claims 1 to 8, characterized in that: include: Signal acquisition module, signal separation module, partial coherence module, feature extraction module; The signal acquisition module is used to collect the acoustic signal of the main reducer; The signal separation module is used to perform data sparse blind source preprocessing on the collected signal to obtain an unmixed signal; The partial coherence module performs partial coherence analysis on the unmixed signal, determines the main noise source, and obtains the analyzed signal; The feature extraction module calculates the cepstral coefficients through a Mel filter and extracts the acoustic signal features.

10. The acoustic signal separation and feature extraction system of a main reducer according to claim 9, characterized in that: The signal acquisition module includes two microphones, a microphone array, and a sound intensity probe; The two microphones are in phase with each other and are arranged face to face with a gap between them. When the sound wave is transmitted along the normal plane of the microphone, there is a gradient between the two sound pressure measurement values, and the sound intensity value of the measurement point is obtained according to the imaginary part of the cross spectrum of the two microphones. The microphone array is composed of a group of microphones arranged in a geometric structure, which simultaneously receives signals from different spatial directions, and determines the position of the main reducer noise source based on the phased array principle by comparing the phase differences of the noise signals reaching each microphone; The sound intensity probe is perpendicular to the envelope surface and is P cm away from the measurement surface. The sound intensity measurement result of the center point of the grid is processed to obtain the sound intensity distribution cloud map of each frequency band of each measurement surface.

Images