Hydroelectric generating set noise blind source separation method and system

Abstract

The invention discloses a noise blind source separation method and system for a hydroelectric generating set, belongs to the technical field of acoustic signal processing, and aims to solve the problem that the noise reduction effect of acoustic signals is affected due to the fact that all kinds of noise are difficult to effectively separate by an existing acoustic signal separation method. The method comprises the following steps: firstly, preprocessing actually measured sound signals of a bulb tubular hydroelectric generating set operation site through time domain, frequency domain and time frequency analysis, then carrying out feature analysis by adopting a YOLOv3 network, finally carrying out noise separation by adopting a blind source separation algorithm, and obtaining human voice signals of the operation site based on sound contribution. Therefore, human voice separation and noise reduction processing of actually measured sound signals in a working site are realized.

Description

Technical Field

[0001] The invention belongs to the technical field of acoustic signal processing, and in particular relates to a blind source separation method for noise of a hydropower unit and a design of a system. Background Art

[0002] The development of tubular turbine generator equipment began in 1892, and it has a history of more than 100 years. Since the 1980s, bulb tubular turbine units have become increasingly popular in my country, and have entered the practical stage below the 20m head section. Bulb tubular turbine generator units are a type of equipment that uses low head water in rivers to generate electricity. They look like incandescent lamps, and water flows into the bulb head from the bulb mouth to generate electricity. Bulb tubular turbine generator units have the advantages of relatively low unit capacity investment, relatively abundant water resources, high unit efficiency, and relatively low operating costs.

[0003] The acoustic signals measured at the bulb-type hydroelectric generator maintenance site contain various noises, so it is necessary to separate these acoustic signals, that is, to separate the human voice from the noise in the acoustic signals to achieve noise reduction processing of the acoustic signals. However, the existing acoustic signal separation methods are difficult to effectively separate all types of noise, which in turn affects the noise reduction effect of the acoustic signals. Summary of the invention

[0004] The purpose of the present invention is to solve the problem that the existing acoustic signal separation method is difficult to effectively separate all types of noise, thereby affecting the noise reduction effect of the acoustic signal, and proposes a blind source separation method and system for hydropower unit noise.

[0005] The technical solution of the present invention is as follows: In a first aspect, the present invention provides a method for blind source separation of noise from a hydropower unit, comprising the following steps:

[0006] S1. Perform data preprocessing on the measured acoustic signals at the bulb-type hydro-turbine generator set operation site to obtain acoustic signal characteristics.

[0007] S2. Perform feature analysis on the acoustic signal features to obtain a mixed signal.

[0008] S3. Perform blind source separation on the mixed signal to obtain the human voice signal at the work site, thereby achieving noise reduction processing on the measured sound signal at the work site.

[0009] Furthermore, the data preprocessing in step S1 includes time domain analysis, frequency domain analysis and time-frequency domain analysis.

[0010] Furthermore, the acoustic signal characteristics obtained in step S1 include an equivalent continuous sound level Leq, a sound pressure level SPL, and a peak sound pressure level Lpeak.

[0011] Furthermore, the calculation formula of the equivalent continuous sound level Leq is:

[0012]

[0013] Where L(t) represents the instantaneous sound level A at time t, and T represents the measurement time.

[0014] The calculation formula for sound pressure level SPL is:

[0015]

[0016] Where p(t) represents the instantaneous sound pressure at time t, and p0 represents the reference sound pressure.

[0017] The calculation formula of peak sound pressure level Lpeak is:

[0018]

[0019] where p peak Indicates the maximum sound pressure in the sound signal.

[0020] Furthermore, in step S2, a YOLOv3 network is used to perform feature analysis on the acoustic signal features to obtain a mixed signal including mechanical noise, white noise, personnel noise and human voice signal.

[0021] Further, step S3 includes the following sub-steps:

[0022] S31. Use a blind source separation algorithm to perform blind source separation on the mixed signal, remove various noises in the mixed signal, and obtain a separated signal.

[0023] S32. Calculate the acoustic contribution of the separated signal.

[0024] S33. In response to the acoustic contribution being greater than a preset threshold, the separated signal is output as a human voice signal at the work site, thereby achieving noise reduction processing on the measured acoustic signal at the work site.

[0025] Furthermore, the blind source separation algorithm in step S31 adopts the ICA algorithm, and its specific steps are:

[0026] A1. Get the mixed signal vector X = (x1, x2, ..., x n ) T , where x i represents the i-th component of the mixed signal vector, and n represents the dimension of the mixed signal.

[0027] A2. Find a coordinate system {ξ1,ξ2,…,ξ n}, so that when the components x1,x2,...,x n When projected under this coordinate system, it satisfies:

[0028] x i=s1ξ1+s2ξ2+…+s n ξ n ,i=1,2,…,n

[0029] where ξ i represents the i-th component in the coordinate system, s i represents the i-th projection coefficient, and the projection coefficients s1,s2,...,s n Independent of each other.

[0030] A3. According to the coordinate system {ξ1,ξ2,…,ξ n}Get the optimal linear transformation matrix W:

[0031] W -1 =[ξ1,ξ2,…,ξ n ]

[0032] A4. Perform a linear transformation on the mixed signal vector X through the optimal linear transformation matrix W to obtain a separated signal Y=WX with various noises removed, so that the mutual information of Y is 0.

[0033] Furthermore, the blind source separation algorithm in step S31 adopts the JADE algorithm, and its specific steps are:

[0034] B1. Perform centralization on the mixed signal vector X to obtain a whitened signal:

[0035]

[0036] Where z represents the whitening signal, E D represents the eigenvector of the covariance matrix of the mixed signal vector X, and D represents the eigenvalue of the covariance matrix of the mixed signal vector X.

[0037] B2. Calculate the cumulant matrix based on the whitened signal:

[0038] Q=E[zz T z T z]-I-2(zz T )

[0039] Where Q represents the cumulant matrix, E[·] represents the expected vector, and I represents the identity matrix.

[0040] B3. Calculate the demixing matrix based on the cumulant matrix:

[0041]

[0042] Where W1 represents the demixing matrix, D k represents a k-dimensional diagonal matrix, and ||·|| represents the Frobenius norm of the matrix.

[0043] B4. Remove various noises in the mixed signal vector X by de-mixing the matrix to obtain a separated signal S=W1X.

[0044] Furthermore, the blind source separation algorithm in step S31 adopts the TDSEP algorithm, and its specific steps are:

[0045] C1. Calculate the delay covariance matrix of the mixed signal:

[0046] C=E τ {x(t)x(t-τ) T}

[0047] Where C represents the delay covariance matrix, E τ {·} means to obtain the covariance with a time delay of τ, x(t) represents the mixed signal at time t, τ represents the time delay constant, and the superscript T represents the transpose of the matrix.

[0048] C2. Objective function of constructing the transformation matrix based on the delay covariance matrix:

[0049]

[0050] Where J(W2) represents the objective function of the transformation matrix W2, S represents the set of delay constants, and off(·) represents the degree of diagonalization of the matrix.

[0051] C3. Obtain the transformation matrix W2 by minimizing the objective function, and use the transformation matrix W2 to remove various noises in the mixed signal vector to obtain the separated signal:

[0052] y(t)=W2x(t)

[0053] Where y(t) represents the separated signal at time t.

[0054] Furthermore, the acoustic contribution calculation formula of the separated signal in step S32 is:

[0055]

[0056] Where contri represents the acoustic contribution of the separated signal, P 2 (t) represents the energy of the separated signal, P U 2 (t) represents the energy of the mixed signal, t1 represents the start time of the separated signal, and t2 represents the end time of the separated signal.

[0057] In a second aspect, a blind source separation system for noise of a hydropower unit includes:

[0058] The data preprocessing module is used to perform data preprocessing on the measured sound signals at the operation site of the bulb tubular turbine generator set to obtain the characteristics of the sound signals.

[0059] The feature analysis module is used to perform feature analysis on the acoustic signal features to obtain a mixed signal.

[0060] The blind source separation module is used to perform blind source separation on the mixed signal to obtain the human voice signal at the work site and realize noise reduction processing on the measured sound signal at the work site.

[0061] Furthermore, the data preprocessing in the data preprocessing module includes time domain analysis, frequency domain analysis and time-frequency domain analysis.

[0062] Furthermore, the acoustic signal features obtained in the data preprocessing module include an equivalent continuous sound level Leq, a sound pressure level SPL, and a peak sound pressure level Lpeak.

[0063] Furthermore, the calculation formula of the equivalent continuous sound level Leq is:

[0064]

[0065] Where L(t) represents the instantaneous sound level A at time t, and T represents the measurement time.

[0066] The calculation formula for sound pressure level SPL is:

[0067]

[0068] Where p(t) represents the instantaneous sound pressure at time t, and p0 represents the reference sound pressure.

[0069] The calculation formula of peak sound pressure level Lpeak is:

[0070]

[0071] where p peak Indicates the maximum sound pressure in the sound signal.

[0072] Furthermore, the feature analysis module uses the YOLOv3 network to perform feature analysis on the acoustic signal features to obtain a mixed signal including mechanical noise, white noise, personnel noise and human voice signals.

[0073] Furthermore, the blind source separation module includes:

[0074] The blind source separation algorithm unit is used to perform blind source separation on the mixed signal by using the blind source separation algorithm, so as to remove various noises in the mixed signal and obtain a separated signal.

[0075] The acoustic contribution calculation unit is used to calculate the acoustic contribution of the separated signal.

[0076] The human voice signal output unit, in response to the sound contribution being greater than a preset threshold, outputs the separated signal as the human voice signal of the work site, thereby realizing noise reduction processing of the measured sound signal of the work site.

[0077] Furthermore, the blind source separation algorithm in the blind source separation algorithm unit adopts an ICA algorithm, and its specific steps are:

[0078] A1. Get the mixed signal vector X = (x1, x2, ..., x n ) T , where x i represents the i-th component of the mixed signal vector, and n represents the dimension of the mixed signal.

[0079] A2. Find a coordinate system {ξ1,ξ2,…,ξ n}, so that when the components x1,x2,...,x n When projected under this coordinate system, it satisfies:

[0080] x i =s1ξ1+s2ξ2+…+s n ξ n ,i=1,2,…,n

[0081] where ξ i represents the i-th component in the coordinate system, s i represents the i-th projection coefficient, and the projection coefficients s1,s2,...,s n Independent of each other.

[0082] A3. According to the coordinate system {ξ1,ξ2,…,ξ n}Get the optimal linear transformation matrix W:

[0083] W -1 =[ξ1,ξ2,…,ξ n ]

[0084] A4. Perform a linear transformation on the mixed signal vector X through the optimal linear transformation matrix W to obtain a separated signal Y=WX with various noises removed, so that the mutual information of Y is 0.

[0085] Furthermore, the blind source separation algorithm in the blind source separation algorithm unit adopts the JADE algorithm, and its specific steps are:

[0086] B1. Perform centralization on the mixed signal vector X to obtain a whitened signal:

[0087]

[0088] Where z represents the whitening signal, E D represents the eigenvector of the covariance matrix of the mixed signal vector X, and D represents the eigenvalue of the covariance matrix of the mixed signal vector X.

[0089] B2. Calculate the cumulant matrix based on the whitened signal:

[0090] Q=E[zz T z T z]-I-2(zz T )

[0091] Where Q represents the cumulant matrix, E[·] represents the expected vector, and I represents the identity matrix.

[0092] B3. Calculate the demixing matrix based on the cumulant matrix:

[0093]

[0094] Where W1 represents the demixing matrix, D k represents a k-dimensional diagonal matrix, and ||·|| represents the Frobenius norm of the matrix.

[0095] B4. Remove various noises in the mixed signal vector X by de-mixing the matrix to obtain a separated signal S=W1X.

[0096] Furthermore, the blind source separation algorithm in the blind source separation algorithm unit adopts the TDSEP algorithm, and its specific steps are:

[0097] C1. Calculate the delay covariance matrix of the mixed signal:

[0098] C=E τ {x(t)x(t-τ) T}

[0099] Where C represents the delay covariance matrix, E τ {·} means to obtain the covariance with a time delay of τ, x(t) represents the mixed signal at time t, τ represents the time delay constant, and the superscript T represents the transpose of the matrix.

[0100] C2. Objective function of constructing the transformation matrix based on the delay covariance matrix:

[0101]

[0102] Where J(W2) represents the objective function of the transformation matrix W2, S represents the set of delay constants, and off(·) represents the degree of diagonalization of the matrix.

[0103] C3. Obtain the transformation matrix W2 by minimizing the objective function, and use the transformation matrix W2 to remove various noises in the mixed signal vector to obtain the separated signal:

[0104] y(t)=W2x(t)

[0105] Where y(t) represents the separated signal at time t.

[0106] Furthermore, the acoustic contribution calculation formula of the separated signal in the acoustic contribution calculation unit is:

[0107]

[0108] Where contri represents the acoustic contribution of the separated signal, P 2 (t) represents the energy of the separated signal, P U 2 (t) represents the energy of the mixed signal, t1 represents the start time of the separated signal, and t2 represents the end time of the separated signal.

[0109] The beneficial effects of the present invention are:

[0110] (1) The present invention targets the measured sound signals at the operation site of a bulb-type cross-flow turbine generator set. The present invention first pre-processes the sound signals through time domain, frequency domain and time-frequency analysis. Then, the YOLOv3 network is used for feature analysis. Finally, a blind source separation algorithm is used for noise separation. Based on the sound contribution, the human voice signal at the operation site is obtained, thereby realizing the human voice separation and noise reduction processing of the measured sound signals at the operation site.

[0111] (2) The present invention adopts the YOLOv3 network for feature analysis, adopts upsampling with a step size of 2 between feature layers, and adopts a 1×1 convolution kernel after feature fusion to eliminate the aliasing effect. It can accurately identify the scale change of the sound signal features and effectively extract the key feature parameters of mechanical noise, white noise, personnel noise and human voice signals.

[0112] (3) The present invention can use three different blind source separation algorithms, namely the ICA algorithm, the JADE algorithm and the TDSEP algorithm, to perform human voice separation and noise reduction on the sound signals measured at the work site, thereby effectively improving the robustness of the noise reduction of the sound signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0113] Figure 1 The figure shows a flow chart of a method for blind source separation of noise of a hydropower unit provided in the first embodiment of the present invention.

[0114] Figure 2 The figure shows the YOLOv3 network feature pyramid structure diagram provided by Embodiment 1 and Embodiment 2 of the present invention.

[0115] Figure 3 The figure shows a flow chart of blind source separation of noise of a hydropower unit provided in the second embodiment of the present invention. DETAILED DESCRIPTION

[0116] Now, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. It should be understood that the embodiments shown and described in the accompanying drawings are only exemplary and are intended to explain the principles and spirit of the present invention, rather than to limit the scope of the present invention.

[0117] Embodiment 1:

[0118] The embodiment of the present invention provides a method for blind source separation of noise from a hydropower unit. Figure 1 As shown, the following steps S1 to S3 are included:

[0119] S1. Perform data preprocessing on the measured acoustic signals at the bulb-type hydro-turbine generator set operation site to obtain acoustic signal characteristics.

[0120] In the embodiment of the present invention, data preprocessing includes time domain analysis, frequency domain analysis and time-frequency domain analysis.

[0121] Among them, time domain analysis is used to analyze the periodicity and repetitive characteristics of the acoustic signal, and the autocorrelation function is used to reveal the periodic structure and repetitive pattern of the signal by calculating the similarity between the signal and its delayed version. In actual operation, the periodic characteristics and delay structure of the signal can be identified by plotting the autocorrelation function graph.

[0122] Frequency domain analysis converts the time domain signal to the frequency domain through transformation to observe the frequency components of the signal. In the embodiment of the present invention, fast Fourier transform (FFT) is used for frequency domain analysis to convert the time domain signal into the frequency domain signal and analyze its spectrum characteristics. FFT reveals the frequency distribution of the signal by decomposing the time domain signal into sinusoidal components of different frequencies. The spectrum diagram obtained by FFT can show the intensity of various frequency components in the signal, helping to identify the main frequency components and the frequency distribution of noise sources.

[0123] In the embodiment of the present invention, the time-frequency domain analysis can adopt short-time Fourier transform (STFT) or wavelet transform. STFT divides the signal into short time periods and performs Fourier transform on each time period to obtain time-frequency representation. Wavelet transform performs multi-scale analysis on the signal through wavelet function, providing frequency information of the signal at different times and scales. Wavelet transform can provide signal representation with high time resolution and high frequency resolution at the same time.

[0124] In the embodiment of the present invention, the acoustic signal characteristics include an equivalent continuous sound level Leq, a sound pressure level SPL, and a peak sound pressure level Lpeak.

[0125] Among them, the calculation formula of equivalent continuous sound level Leq is:

[0126]

[0127] Where L(t) represents the instantaneous sound level A at time t, and T represents the measurement time.

[0128] The calculation formula for sound pressure level SPL is:

[0129]

[0130] Where p(t) represents the instantaneous sound pressure at time t, and p0 represents the reference sound pressure (usually 20 μPa).

[0131] The calculation formula of peak sound pressure level Lpeak is:

[0132]

[0133] where p peak Indicates the maximum sound pressure in the sound signal.

[0134] S2. Analyze the characteristics of the acoustic signal to obtain a mixed signal.

[0135] In an embodiment of the present invention, a YOLOv3 network is used to perform feature analysis on acoustic signal features to obtain a mixed signal including mechanical noise, white noise, personnel noise and human voice signals, which can accurately locate and separate different noise sources and provide more accurate initial conditions for blind source separation.

[0136] The feature pyramid used by the YOLOv3 network is as follows Figure 2 As shown in the figure, the feature pyramid is a network-level structure that comprehensively utilizes the information of multi-scale feature maps to detect target objects. In order to obtain multi-scale features, the YOLOv3 network uses upsampling with a step size of 2 between feature layers and uses a 1×1 convolution kernel after feature fusion to eliminate the aliasing effect, thereby identifying whether the scale of the feature map has changed. Figure 2 As shown in Figure 1, the output layer of the YOLOv3 network consists of three feature layers, located in the middle layer, the lower middle layer, and the bottom layer. The output layer of the YOLOv3 network is the location of the bounding box and the target category therein.

[0137] S3. Perform blind source separation on the mixed signal to obtain the human voice signal at the work site, thereby achieving noise reduction processing on the measured sound signal at the work site.

[0138] Step S3 includes the following sub-steps S31 to S33:

[0139] S31. Use a blind source separation algorithm to perform blind source separation on the mixed signal, remove various noises in the mixed signal, and obtain a separated signal.

[0140] In the embodiment of the present invention, the blind source separation algorithm adopts an independent component analysis (ICA) algorithm, a joint diagonalization (JADE) algorithm or a TDSEP algorithm.

[0141] Among them, the purpose of the ICA algorithm is to determine a nonlinear transformation so that the output components are as independent as possible. Therefore, the maximization of independence between the output components is generally used as the optimization criterion of the algorithm. The degree of independence between signal components can be measured by non-Gaussianity, high-order statistics, entropy, mutual information, probability density function (pdf), and Kullback-Leibler divergence of two pdf distances. The specific steps of the ICA algorithm are:

[0142] A1. Get the mixed signal vector X = (x1, x2, ..., x n ) T , where x i represents the i-th component of the mixed signal vector, and n represents the dimension of the mixed signal.

[0143] Assume that the mixed signal vector X consists of m unknown source signals (or independent sources) r1, r2, ..., r m Linear combination, let R = (r1, r2, ..., r m ) T is an m-dimensional unknown source signal, then the ICA linear model can be expressed as:

[0144]

[0145] Where A=[a1,a2,...a m ] is a full-rank n*m matrix, called a mixing matrix, a j is the basis vector of the mixing matrix. From the above model, we can see that each observation data x i is from an independent source j Through different j Linear weighted. Independent source r j It is an implicit variable and cannot be measured directly; the mixing matrix A is also an unknown matrix, and the only available information is the mixed signal vector X. If there are no restrictions, to estimate R and A only from X, the solution of the equation must be multiple solutions. And ICA gives a unique solution to the equation under certain restrictions based on the statistical characteristics of X, and realizes the extraction of independent components. As mentioned above, an important basic assumption of ICA is the requirement for the independence of unknown source signals. For the specific ICA model, the unknown source signals are independent of each other. In the ICA model, in addition to requiring the source signals to be independent of each other, they must also meet the characteristics of non-Gaussian distribution.

[0146] A2. Find a coordinate system {ξ1,ξ2,…,ξ n}, so that when the components x1,x2,...,x n When projected under this coordinate system, it satisfies:

[0147] xi =s1ξ1+s2ξ2+…+s n ξ n ,i=1,2,…,n

[0148] where ξ i represents the i-th component in the coordinate system, s i represents the i-th projection coefficient, and the projection coefficients s1,s2,...,s n Independent of each other.

[0149] A3. According to the coordinate system {ξ1,ξ2,…,ξ n}Get the optimal linear transformation matrix W:

[0150] W -1 =[ξ1,ξ2,…,ξ n ]

[0151] A4. Linearly transform the mixed signal vector X through the optimal linear transformation matrix W to obtain the separated signal Y=WX with various noises removed, so that the mutual information of Y is 0, that is, the component y in the separated signal Y i Statistically independent of each other.

[0152] Based on the fact that there is no mutual cumulant between independent signals, the JADE algorithm constructs a fourth-order cumulant matrix of multiple multivariate data, and performs joint diagonalization on them to complete the solution of the separation matrix. The specific steps of the JADE algorithm are:

[0153] B1. Perform centralization on the mixed signal vector X to obtain a whitened signal:

[0154]

[0155] Where z represents the whitening signal, E D represents the eigenvector of the covariance matrix of the mixed signal vector X, and D represents the eigenvalue of the covariance matrix of the mixed signal vector X.

[0156] B2. Calculate the cumulant matrix based on the whitened signal:

[0157] Q=E[zz T z T z]-I-2(zz T )

[0158] Where Q represents the cumulant matrix, E[·] represents the expected vector, and I represents the identity matrix.

[0159] B3. Calculate the demixing matrix based on the cumulant matrix:

[0160]

[0161] Where W1 represents the demixing matrix, D k represents a k-dimensional diagonal matrix, and ||·|| represents the Frobenius norm of the matrix.

[0162] B4. Remove various noises in the mixed signal vector X by de-mixing the matrix to obtain a separated signal S=W1X.

[0163] The TDSEP algorithm is based on multiple time-delay covariance matrices and Jacobi rotation technology. It also proposes the assumption that the source signals are independent of each other, but it is significantly different from the general random signal mixing model. The requirement of Gaussian signal no longer exists because the time series signal has autocovariance, which can replace the non-Gaussianity. The specific steps of the TDSEP algorithm are:

[0164] C1. Calculate the delay covariance matrix of the mixed signal:

[0165] C=E τ {x(t)x(t-τ) T}

[0166] Where C represents the delay covariance matrix, E τ {·} means to obtain the covariance with a time delay of τ, x(t) represents the mixed signal at time t, τ represents the time delay constant, and the superscript T represents the transpose of the matrix.

[0167] C2. Objective function of constructing the transformation matrix based on the delay covariance matrix:

[0168]

[0169] Where J(W2) represents the objective function of the transformation matrix W2, S represents the set of delay constants, and off(·) represents the degree of diagonalization of the matrix. The TDSEP algorithm uses multi-step delays. In general, it is always hoped that the delay covariance matrix can also be diagonalized at the same time. However, simultaneous diagonalization is sometimes impossible, so it is necessary to use a function to represent the degree of diagonalization achieved.

[0170] C3. Obtain the transformation matrix W2 by minimizing the objective function, and use the transformation matrix W2 to remove various noises in the mixed signal vector to obtain the separated signal:

[0171] y(t)=W2x(t)

[0172] Where y(t) represents the separated signal at time t. The separated signal y(t) needs to satisfy the transient covariance equal to 0, and the delay covariance must also be equal to 0, that is:

[0173] E{y(t)y(t-τ)}=0

[0174] S32. Calculate the acoustic contribution of the separated signal.

[0175] In the embodiment of the present invention, the acoustic contribution calculation formula of the separated signal is:

[0176]

[0177] Where contri represents the acoustic contribution of the separated signal, P 2 (t) represents the energy of the separated signal, P U 2 (t) represents the energy of the mixed signal, t1 represents the start time of the separated signal, and t2 represents the end time of the separated signal.

[0178] S33. In response to the acoustic contribution being greater than a preset threshold, the separated signal is output as a human voice signal at the work site, thereby achieving noise reduction processing on the measured acoustic signal at the work site.

[0179] In an embodiment of the present invention, when the acoustic contribution of the separated signal is greater than a preset threshold, it means that the separated signal has sufficient contribution of human voice signal energy, that is, the current sound information contains communication information between maintenance personnel. Therefore, the separated signal is retained as the human voice signal output at the work site to achieve noise reduction processing of the measured sound signal at the work site.

[0180] Embodiment 2:

[0181] The embodiment of the present invention provides a blind source separation system for noise of a hydropower unit. Figure 3 As shown, including:

[0182] The data preprocessing module is used to perform data preprocessing on the measured sound signals at the operation site of the bulb tubular turbine generator set to obtain the characteristics of the sound signals.

[0183] The feature analysis module is used to perform feature analysis on the acoustic signal features to obtain a mixed signal.

[0184] The blind source separation module is used to perform blind source separation on the mixed signal to obtain the human voice signal at the work site and realize noise reduction processing on the measured sound signal at the work site.

[0185] In the embodiment of the present invention, the data preprocessing in the data preprocessing module includes time domain analysis, frequency domain analysis and time-frequency domain analysis.

[0186] Among them, time domain analysis is used to analyze the periodicity and repetitive characteristics of the acoustic signal, and the autocorrelation function is used to reveal the periodic structure and repetitive pattern of the signal by calculating the similarity between the signal and its delayed version. In actual operation, the periodic characteristics and delay structure of the signal can be identified by plotting the autocorrelation function graph.

[0187] Frequency domain analysis converts the time domain signal to the frequency domain through transformation to observe the frequency components of the signal. In the embodiment of the present invention, fast Fourier transform (FFT) is used for frequency domain analysis to convert the time domain signal into the frequency domain signal and analyze its spectrum characteristics. FFT reveals the frequency distribution of the signal by decomposing the time domain signal into sinusoidal components of different frequencies. The spectrum diagram obtained by FFT can show the intensity of various frequency components in the signal, helping to identify the main frequency components and the frequency distribution of noise sources.

[0188] In the embodiment of the present invention, the time-frequency domain analysis can adopt short-time Fourier transform (STFT) or wavelet transform. STFT divides the signal into short time periods and performs Fourier transform on each time period to obtain time-frequency representation. Wavelet transform performs multi-scale analysis on the signal through wavelet function, providing frequency information of the signal at different times and scales. Wavelet transform can provide signal representation with high time resolution and high frequency resolution at the same time.

[0189] In the embodiment of the present invention, the acoustic signal features obtained in the data preprocessing module include the equivalent continuous sound level Leq, the sound pressure level SPL, and the peak sound pressure level Lpeak.

[0190] Among them, the calculation formula of equivalent continuous sound level Leq is:

[0191]

[0192] Where L(t) represents the instantaneous sound level A at time t, and T represents the measurement time.

[0193] The calculation formula for sound pressure level SPL is:

[0194]

[0195] Where p(t) represents the instantaneous sound pressure at time t, and p0 represents the reference sound pressure (usually 20 μPa).

[0196] The calculation formula of peak sound pressure level Lpeak is:

[0197]

[0198] where p peak Indicates the maximum sound pressure in the sound signal.

[0199] In an embodiment of the present invention, a YOLOv3 network is used in a feature analysis module to perform feature analysis on acoustic signal features, thereby obtaining a mixed signal including mechanical noise, white noise, personnel noise and human voice signals, which can accurately locate and separate different noise sources, and provide more accurate initial conditions for blind source separation.

[0200] The feature pyramid used by the YOLOv3 network is as follows Figure 2 As shown in the figure, the feature pyramid is a network-level structure that comprehensively utilizes the information of multi-scale feature maps to detect target objects. In order to obtain multi-scale features, the YOLOv3 network uses upsampling with a step size of 2 between feature layers and uses a 1×1 convolution kernel after feature fusion to eliminate the aliasing effect, thereby identifying whether the scale of the feature map has changed. Figure 2 As shown in Figure 1, the output layer of the YOLOv3 network consists of three feature layers, located in the middle layer, the lower middle layer, and the bottom layer. The output layer of the YOLOv3 network is the location of the bounding box and the target category therein.

[0201] In the embodiment of the present invention, Figure 3 As shown, the blind source separation module includes:

[0202] The blind source separation algorithm unit is used to perform blind source separation on the mixed signal by using the blind source separation algorithm, so as to remove various noises in the mixed signal and obtain a separated signal.

[0203] The acoustic contribution calculation unit is used to calculate the acoustic contribution of the separated signal.

[0204] The human voice signal output unit, in response to the sound contribution being greater than a preset threshold, outputs the separated signal as the human voice signal of the work site, thereby realizing noise reduction processing of the measured sound signal of the work site.

[0205] In the embodiment of the present invention, the blind source separation algorithm adopts an independent component analysis (ICA) algorithm, a joint diagonalization (JADE) algorithm or a TDSEP algorithm.

[0206] Among them, the purpose of the ICA algorithm is to determine a nonlinear transformation so that the output components are as independent as possible. Therefore, the maximization of independence between the output components is generally used as the optimization criterion of the algorithm. The degree of independence between signal components can be measured by non-Gaussianity, high-order statistics, entropy, mutual information, probability density function (pdf), and Kullback-Leibler divergence of two pdf distances. The specific steps of the ICA algorithm are:

[0207] A1. Get the mixed signal vector X = (x1, x2, ..., x n ) T , where x i represents the i-th component of the mixed signal vector, and n represents the dimension of the mixed signal.

[0208] Assume that the mixed signal vector X consists of m unknown source signals (or independent sources) r1, r2, …, r m Linear combination, let R = (r1, r2, ..., r m ) Tis an m-dimensional unknown source signal, then the ICA linear model can be expressed as:

[0209]

[0210] Where A=[a1,a2,…a m ] is a full-rank n*m matrix, called a mixing matrix, a j is the basis vector of the mixing matrix. From the above model, we can see that each observation data x i is from an independent source j Through different j Linear weighted. Independent source r j It is an implicit variable and cannot be measured directly; the mixing matrix A is also an unknown matrix, and the only available information is the mixed signal vector X. If there are no restrictions, to estimate R and A only from X, the solution of the equation must be multiple solutions. And ICA gives a unique solution to the equation under certain restrictions based on the statistical characteristics of X, and realizes the extraction of independent components. As mentioned above, an important basic assumption of ICA is the requirement for the independence of unknown source signals. For the specific ICA model, the unknown source signals are independent of each other. In the ICA model, in addition to requiring the source signals to be independent of each other, they must also meet the characteristics of non-Gaussian distribution.

[0211] A2. Find a coordinate system {ξ1,ξ2,…,ξ n}, so that when the components x1,x2,...,x n When projected under this coordinate system, it satisfies:

[0212] x i =s1ξ1+s2ξ2+…+s n ξ n ,i=1,2,…,n

[0213] where ξ i represents the i-th component in the coordinate system, s i represents the i-th projection coefficient, and the projection coefficients s1,s2,...,s n Independent of each other.

[0214] A3. According to the coordinate system {ξ1,ξ2,…,ξ n}Get the optimal linear transformation matrix W:

[0215] W -1 =[ξ1,ξ2,…,ξ n ]

[0216] A4. Linearly transform the mixed signal vector X through the optimal linear transformation matrix W to obtain the separated signal Y=WX with various noises removed, so that the mutual information of Y is 0, that is, the component y in the separated signal Yi Statistically independent of each other.

[0217] Based on the fact that there is no mutual cumulant between independent signals, the JADE algorithm constructs a fourth-order cumulant matrix of multiple multivariate data, and performs joint diagonalization on them to complete the solution of the separation matrix. The specific steps of the JADE algorithm are:

[0218] B1. Perform centralization on the mixed signal vector X to obtain a whitened signal:

[0219]

[0220] Where z represents the whitening signal, E D represents the eigenvector of the covariance matrix of the mixed signal vector X, and D represents the eigenvalue of the covariance matrix of the mixed signal vector X.

[0221] B2. Calculate the cumulant matrix based on the whitened signal:

[0222] Q=E[zz T z T z]-I-2(zz T )

[0223] Where Q represents the cumulant matrix, E[·] represents the expected vector, and I represents the identity matrix.

[0224] B3. Calculate the demixing matrix based on the cumulant matrix:

[0225]

[0226] Where W1 represents the demixing matrix, D k represents a k-dimensional diagonal matrix, and ||·|| represents the Frobenius norm of the matrix.

[0227] B4. Remove various noises in the mixed signal vector X by de-mixing the matrix to obtain a separated signal S=W1X.

[0228] The TDSEP algorithm is based on multiple time-delay covariance matrices and Jacobi rotation technology. It also proposes the assumption that the source signals are independent of each other, but it is significantly different from the general random signal mixing model. The requirement of Gaussian signal no longer exists because the time series signal has autocovariance, which can replace the non-Gaussianity. The specific steps of the TDSEP algorithm are:

[0229] C1. Calculate the delay covariance matrix of the mixed signal:

[0230] C=E τ {x(t)x(t-τ) T}

[0231] Where C represents the delay covariance matrix, E τ {·} means to obtain the covariance with a time delay of τ, x(t) represents the mixed signal at time t, τ represents the time delay constant, and the superscript T represents the transpose of the matrix.

[0232] C2. Objective function of constructing the transformation matrix based on the delay covariance matrix:

[0233]

[0234] Where J(W2) represents the objective function of the transformation matrix W2, S represents the set of delay constants, and off(·) represents the degree of diagonalization of the matrix. The TDSEP algorithm uses multi-step delays. In general, it is always hoped that the delay covariance matrix can also be diagonalized at the same time. However, simultaneous diagonalization is sometimes impossible, so it is necessary to use a function to represent the degree of diagonalization achieved.

[0235] C3. Obtain the transformation matrix W2 by minimizing the objective function, and use the transformation matrix W2 to remove various noises in the mixed signal vector to obtain the separated signal:

[0236] y(t)=W2x(t)

[0237] Where y(t) represents the separated signal at time t. The separated signal y(t) needs to satisfy the transient covariance equal to 0, and the delay covariance must also be equal to 0, that is:

[0238] E{y(t)y(t-τ)}=0

[0239] In the embodiment of the present invention, the acoustic contribution calculation formula of the separated signal in the acoustic contribution calculation unit is:

[0240]

[0241] Where contri represents the acoustic contribution of the separated signal, P 2 (t) represents the energy of the separated signal, P U 2 (t) represents the energy of the mixed signal, t1 represents the start time of the separated signal, and t2 represents the end time of the separated signal.

[0242] In an embodiment of the present invention, when the acoustic contribution of the separated signal is greater than a preset threshold, it means that the separated signal has sufficient contribution of human voice signal energy, that is, the current sound information contains communication information between maintenance personnel. Therefore, the separated signal is retained as the human voice signal output at the work site to achieve noise reduction processing of the measured sound signal at the work site.

[0243] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A method for blind source separation of noise from a hydropower unit, characterized in that: The following steps are involved: S1. Preprocess the measured acoustic signals at the bulb-type hydroelectric generator set operation site to obtain acoustic signal characteristics; S2, performing feature analysis on the acoustic signal features to obtain a mixed signal; S3. Perform blind source separation on the mixed signal to obtain the human voice signal at the work site, thereby achieving noise reduction processing on the measured sound signal at the work site.

2. The method for blind source separation of noise from hydropower units according to claim 1 is characterized in that: The data preprocessing in step S1 includes time domain analysis, frequency domain analysis and time-frequency domain analysis.

3. The method for blind source separation of noise from hydropower units according to claim 1, characterized in that: The acoustic signal characteristics obtained in step S1 include an equivalent continuous sound level Leq, a sound pressure level SPL, and a peak sound pressure level Lpeak.

4. The method for blind source separation of noise from a hydropower unit according to claim 3 is characterized in that: The calculation formula of the equivalent continuous sound level Leq is: Where L(t) represents the instantaneous A sound level at time t, and T represents the measurement time; The calculation formula of the sound pressure level SPL is: Where p(t) represents the instantaneous sound pressure at time t, and p0 represents the reference sound pressure; The calculation formula of the peak sound pressure level Lpeak is: where p peak Indicates the maximum sound pressure in the sound signal.

5. The method for blind source separation of noise from hydropower units according to claim 1, characterized in that: In step S2, the YOLOv3 network is used to perform feature analysis on the acoustic signal features to obtain a mixed signal including mechanical noise, white noise, human noise and human voice signal.

6. The method for blind source separation of noise from a hydropower unit according to claim 1, characterized in that: The step S3 comprises the following sub-steps: S31, performing blind source separation on the mixed signal using a blind source separation algorithm to remove various noises in the mixed signal to obtain a separated signal; S32, calculating the acoustic contribution of the separated signal; S33. In response to the acoustic contribution being greater than a preset threshold, the separated signal is output as a human voice signal at the work site, thereby achieving noise reduction processing on the measured acoustic signal at the work site.

7. The method for blind source separation of hydropower unit noise according to claim 6 is characterized in that: The blind source separation algorithm in step S31 adopts the ICA algorithm, and its specific steps are: A1. Get the mixed signal vector X = (x1, x2, ..., x n ) T , where x i represents the i-th component of the mixed signal vector, and n represents the dimension of the mixed signal; A2. Find a coordinate system {ξ1,ξ2,…,ξ n }, so that when the components x1,x2,...,x n When projected in this coordinate system, it satisfies: x i =s1ξ1+s2ξ2+…+s n x n ,i=1,2,…,n where ξ i represents the i-th component in the coordinate system, s i represents the i-th projection coefficient, and the projection coefficients s1,s2,...,s n Independent of each other; A3. According to the coordinate system {ξ1,ξ2,…,ξ n }Get the optimal linear transformation matrix W: W -1 =[ξ1,ξ2,…,ξ n ] A4. Perform a linear transformation on the mixed signal vector X through the optimal linear transformation matrix W to obtain a separated signal Y=WX with various noises removed, so that the mutual information of Y is 0.

8. The method for blind source separation of noise from a hydropower unit according to claim 6, characterized in that: The blind source separation algorithm in step S31 adopts the JADE algorithm, and its specific steps are: B1. Perform centralization on the mixed signal vector X to obtain a whitened signal: Where z represents the whitening signal, E D represents the eigenvector of the covariance matrix of the mixed signal vector X, and D represents the eigenvalue of the covariance matrix of the mixed signal vector X; B2. Calculate the cumulant matrix based on the whitened signal: Q=E[zz T With T z]-I-2(zz T ) Where Q represents the cumulant matrix, E[·] represents the expected vector, and I represents the identity matrix; B3. Calculate the demixing matrix based on the cumulant matrix: Where W1 represents the demixing matrix, D k represents a k-dimensional diagonal matrix, ||·|| represents the Frobenius norm of the matrix; B4. Remove various noises in the mixed signal vector X by de-mixing the matrix to obtain a separated signal S=W1X.

9. The method for blind source separation of noise from a hydropower unit according to claim 6, characterized in that: The blind source separation algorithm in step S31 adopts the TDSEP algorithm, and its specific steps are: C1. Calculate the delay covariance matrix of the mixed signal: C=E τ {x(t)x(t-τ) T } Where C represents the delay covariance matrix, E τ {·} means to obtain the covariance with a time delay of τ, x(t) represents the mixed signal at time t, τ represents the time delay constant, and the superscript T represents the transpose of the matrix; C2. Objective function of constructing the transformation matrix based on the delay covariance matrix: Where J(W2) represents the objective function of the transformation matrix W2, S represents the set of delay constants, and off(·) represents the degree of diagonalization of the matrix; C3. Obtain the transformation matrix W2 by minimizing the objective function, and use the transformation matrix W2 to remove various noises in the mixed signal vector to obtain the separated signal: y(t)=W2x(t) Where y(t) represents the separated signal at time t.

10. The method for blind source separation of noise from a hydropower unit according to claim 6, characterized in that: The acoustic contribution calculation formula of the separated signal in step S32 is: Where contri represents the acoustic contribution of the separated signal, P 2 (t) represents the energy of the separated signal, P U 2 (t) represents the energy of the mixed signal, t1 represents the start time of the separated signal, and t2 represents the end time of the separated signal.

11. A blind source separation system for noise of hydropower units, characterized in that: include: A data preprocessing module is used to preprocess the measured acoustic signals at the bulb-type hydro-turbine generator set operation site to obtain acoustic signal characteristics; A feature analysis module is used to perform feature analysis on the acoustic signal features to obtain a mixed signal; The blind source separation module is used to perform blind source separation on the mixed signal to obtain the human voice signal at the work site and realize noise reduction processing on the measured sound signal at the work site.

12. The blind source separation system for noise from hydropower units according to claim 11, characterized in that: The data preprocessing in the data preprocessing module includes time domain analysis, frequency domain analysis and time-frequency domain analysis.

13. The blind source separation system for noise from hydropower units according to claim 11, characterized in that: The acoustic signal features obtained in the data preprocessing module include equivalent continuous sound level Leq, sound pressure level SPL and peak sound pressure level Lpeak.

14. The blind source separation system for noise from hydropower units according to claim 13, characterized in that: The calculation formula of the equivalent continuous sound level Leq is: Where L(t) represents the instantaneous A sound level at time t, and T represents the measurement time; The calculation formula of the sound pressure level SPL is: Where p(t) represents the instantaneous sound pressure at time t, and p0 represents the reference sound pressure; The calculation formula of the peak sound pressure level Lpeak is: where p peak Indicates the maximum sound pressure in the sound signal.

15. The blind source separation system for noise from hydropower units according to claim 11, characterized in that: The feature analysis module uses the YOLOv3 network to perform feature analysis on the acoustic signal features to obtain a mixed signal including mechanical noise, white noise, personnel noise and human voice signals.

16. The blind source separation system for noise from hydropower units according to claim 11, characterized in that: The blind source separation module comprises: A blind source separation algorithm unit is used to perform blind source separation on the mixed signal by using a blind source separation algorithm, thereby removing various noises in the mixed signal and obtaining a separated signal; an acoustic contribution calculation unit, for calculating the acoustic contribution of the separated signal; The human voice signal output unit, in response to the sound contribution being greater than a preset threshold, outputs the separated signal as the human voice signal of the work site, thereby realizing noise reduction processing of the measured sound signal of the work site.

17. The blind source separation system for noise from hydropower units according to claim 16, characterized in that: The blind source separation algorithm in the blind source separation algorithm unit adopts the ICA algorithm, and its specific steps are: A1. Get the mixed signal vector X = (x1, x2, ..., x n ) T , where x i represents the i-th component of the mixed signal vector, and n represents the dimension of the mixed signal; A2. Find a coordinate system {ξ1,ξ2,…,ξn} so that when the components x1,x2,…,xn in the mixed signal vector X n When projected in this coordinate system, it satisfies: x i =s1ξ1+s2ξ2+…+s n x n ,i=1,2,…,n where ξ i represents the i-th component in the coordinate system, s i represents the i-th projection coefficient, and the projection coefficients s1,s2,...,s n Independent of each other; A3. According to the coordinate system {ξ1,ξ2,…,ξ n }Get the optimal linear transformation matrix W: W -1 =[ξ1,ξ2,…,ξ n ] A4. Perform a linear transformation on the mixed signal vector X through the optimal linear transformation matrix W to obtain a separated signal Y=WX with various noises removed, so that the mutual information of Y is 0.

18. The blind source separation system for noise from hydropower units according to claim 16, characterized in that: The blind source separation algorithm in the blind source separation algorithm unit adopts the JADE algorithm, and its specific steps are: B1. Perform centralization on the mixed signal vector X to obtain a whitened signal: Where z represents the whitening signal, E D represents the eigenvector of the covariance matrix of the mixed signal vector X, and D represents the eigenvalue of the covariance matrix of the mixed signal vector X; B2. Calculate the cumulant matrix based on the whitened signal: Q=E[zz T With T z]-I-2(zz T ) Where Q represents the cumulant matrix, E[·] represents the expected vector, and I represents the identity matrix; B3. Calculate the demixing matrix based on the cumulant matrix: Where W1 represents the demixing matrix, D k represents a k-dimensional diagonal matrix, ||·|| represents the Frobenius norm of the matrix; B4. Remove various noises in the mixed signal vector X by de-mixing the matrix to obtain a separated signal S=W1X.

19. The blind source separation system for noise from hydropower units according to claim 16, characterized in that: The blind source separation algorithm in the blind source separation algorithm unit adopts the TDSEP algorithm, and its specific steps are: C1. Calculate the delay covariance matrix of the mixed signal: C=E τ {x(t)x(t-τ) T } Where C represents the delay covariance matrix, E τ {·} means to obtain the covariance with a time delay of τ, x(t) represents the mixed signal at time t, τ represents the time delay constant, and the superscript T represents the transpose of the matrix; C2. Objective function of constructing the transformation matrix based on the delay covariance matrix: Where J(W2) represents the objective function of the transformation matrix W2, S represents the set of delay constants, and off(·) represents the degree of diagonalization of the matrix; C3. Obtain the transformation matrix W2 by minimizing the objective function, and use the transformation matrix W2 to remove various noises in the mixed signal vector to obtain the separated signal: y(t)=W2x(t) Where y(t) represents the separated signal at time t.

20. The blind source separation system for noise from hydropower units according to claim 16, characterized in that: The acoustic contribution calculation formula of the separated signal in the acoustic contribution calculation unit is: Where contri represents the acoustic contribution of the separated signal, P 2 (t) represents the energy of the separated signal, P U 2 (t) represents the energy of the mixed signal, t1 represents the start time of the separated signal, and t2 represents the end time of the separated signal.

Images