A convolution blind source separation method for gas turbine related noise source identification

Abstract

A convolution blind source separation method for identifying a gas turbine related noise source is disclosed, in which an observation signal of a gas turbine noise is collected, and a corresponding time-frequency domain complex value signal X is obtained by performing a synchronous compression transformation on the observation signal; an instantaneous blind source separation is performed on the time-frequency domain complex value signal X at each frequency point, a complex value normalized boundary target function is established, and an iteration optimization is performed at each frequency point based on a sub-gradient optimization method, a complex value separation matrix and a separation signal are obtained by performing a plurality of blind extraction estimations, a power spectrum spectral density distance method is used to perform a permutation calibration on the separation signal, an amplitude calibration is performed using a minimum distortion algorithm to obtain a time-frequency domain separation signal, and the time-frequency domain separation signal is restored to a time domain by using a synchronous compression inverse transformation to obtain a reconstructed time domain separation signal.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of mechanical structure sound radiation signal processing, in particular to a convolution blind source separation method for gas turbine related noise source identification. BACKGROUND

[0002] Gas turbine is the core equipment of aviation, ship and energy. The noise monitoring signal of gas turbine is often the result of complex mixing process of various noise source signals, the mutual interference between multiple excitation source responses, the complex noise propagation path and the difficulty in establishing accurate model, which brings difficulties to the identification of noise source.

[0003] The blind source separation method can separate the source signals under the condition that the source signals and the transmission channel are unknown. The blind source separation method based on independent component analysis (ICA) is based on the independence assumption of source signals, however, the high and low pressure rotors of gas turbine are coupled with each other, the mechanical internal parts are coupled in cascade, and there is a certain causal relationship. The non-related source and the related source noise signal with some cross frequencies are mixed together. The noise source of gas turbine has both independent source and related source, and the blind source separation method based on the independence assumption is no longer applicable.

[0004] Bounded component analysis (BCA) provides a new way for the separation of related source signals. Therefore, the research on blind source separation of related noise source has important engineering significance and value for the identification of gas turbine noise source. Instantaneous mixing is the simplest model in blind source separation, such as blind source separation method based on minimum convex perimeter and maximum volume ratio. However, the objective function of these methods is complex to calculate, and the calculation efficiency is low. In addition, for gas turbine, the noise of gas turbine collected by the sensor is mixed by multi-stage parts and the transmission reflection of the inner wall of the cabin. In simple terms, the sound waves received by the sensor can be divided into two parts: one part is the direct sound emitted by the sound source; the other part is the reflected wave transmitted by the sound source through other ways. The effect produced by the reflected wave is called reverberation in acoustics, and a more accurate blind source separation model is urgently needed to identify the related noise source of gas turbine.

[0005] The above information disclosed in the background section is only used to enhance the understanding of the background of the present application, and therefore can contain information that is not prior art known to those of ordinary skill in the art. SUMMARY

[0006] In view of the problems in the prior art, the present application proposes a convolution blind source separation method for gas turbine related noise source identification, which uses a convolution mixing model to describe the blind source separation problem, which is more in line with the actual requirements, and can accurately identify the related noise source of gas turbine.

[0007] The object of the application is achieved by the technical solutions below, a convolution blind source separation method for identifying a gas turbine related noise source comprises,

[0008] Step 1, collecting an observation signal of a gas turbine noise, performing a synchronous compression transformation on the observation signal to obtain a corresponding time-frequency domain complex value signal X;

[0009] Step 2, performing instantaneous blind source separation on the time-frequency domain complex signal X at each frequency point, establishing a complex value normalized boundary target function, and iteratively optimizing based on a sub-gradient optimization method at each frequency point, and estimating a separation matrix by performing multiple blind extraction to obtain a complex value separation matrix and a separation signal;

[0010] Step 3, using a power spectrum spectral density distance method to calibrate the separation signal, and using a minimum distortion algorithm to perform amplitude calibration to obtain a time-frequency domain separation signal;

[0011] Step 4, restoring the time-frequency domain separation signal to a time domain by using a synchronous compression inverse transformation to obtain a reconstructed time domain separation signal.

[0012] In the convolution blind source separation method for identifying a gas turbine related noise source, it is assumed that n source signals of a gas turbine noise are mixed and m sensors measure observation signals, and a synchronous compression transformation is used to convert a convolution mixing model into an instantaneous mixing model at each frequency point:

[0013] X(f,t)=A(f)×S(f,t), wherein X(f,t)=[X1(f,t),…,X m (f,t)] T is a time-frequency vector of a mixed signal, f is a frequency point index, t is a time index, and T is a transposition operation; A(f)=[a1(f),…,a n (f)] is a frequency domain mixing filter; X(f,t)=[X1(f,t),…,X m (f,t)] T is a time-frequency vector of a mixed signal; S(f,t)=[S1(f,t),…,S n (f,t)] T is a time-frequency vector of a source signal, and a convolution mixing problem in a time domain is converted into an instantaneous mixing at each frequency point in a frequency domain.

[0014] In the convolution blind source separation method for identifying a gas turbine related noise source, when the number of n source signals and m sensors is the same, an inverse filter exists in the mixing filter, and the estimation signal is:

[0015] Y f (t)=W f ×X f (t),

[0016] wherein W f is a separation matrix at frequency f; Y f (t) is a separated signal Y f at frequency f and time t.

[0017] In the convolution blind source separation method for identifying the noise source of the gas turbine, the normalized boundary objective function of the complex value signal is:

[0018]

[0019] wherein: represents the boundary operation, Re(·) represents the real part operation of the complex value vector, Im(·) represents the imaginary part operation of the complex value vector, E(·) represents the expectation operation, the "*" in the formula represents the conjugate operator, and the separation vector is obtained by minimizing the objective function.

[0020] In the convolution blind source separation method for identifying the noise source of the gas turbine, the optimization method of the sub-gradient is used to optimize and deduce the separation vector w, and the optimization iteration format of the sub-gradient algorithm is:

[0021] w (p+1) =w (p) -u (p) g (p) ,

[0022] wherein g (p) represents the sub-gradient of the objective function J(y) at w (p) , and u (p) represents the step length of the pth iteration.

[0023] In the convolution blind source separation method for identifying the noise source of the gas turbine, the adaptive step length u adap in each iteration is:

[0024]

[0025] wherein J min is the minimum value of the objective function, is the non-negative measure of the distance between the current iteration and the minimum value.

[0026] In the convolution blind source separation method for identifying the noise source of the gas turbine, the value of Δ (p) is gradually reduced to 0 in the iteration process.

[0027] In the convolution blind source separation method for identifying the noise source of the gas turbine, the complex value separation matrix W f is obtained through multiple blind extractions, and Y f (t) = Wf X f (t) is an approximate estimation of the source signal S f (t) as a separated signal.

[0028] In the method, in step 4), the separated signals Y f (t) of all frequency points are recovered to time domain by using synchronous compression inverse transformation to obtain reconstructed time domain separated signals:

[0029] y(t) = [y1(t), y2(t), …, y n (t)] T , n represents the number of source signals.

[0030] In the method, the normalized cross-correlation coefficient between the source signal and the separated signal is used as a performance evaluation index, and the cross-correlation function of the two signals is calculated as follows:

[0031] C xy (τ) = E{[x(t)-E(x(t))][y(t+τ)-E(y(t))]}

[0032] In the formula, E(·) is a mean value operation, and τ is a moving time.

[0033] The normalized or standardized cross-correlation function is:

[0034]

[0035] In the formula, C xy (τ) is a cross-correlation function, σ x is a standard deviation of x(t), σ y is a standard deviation of y(t), and finally the normalized cross-correlation coefficient is

[0036] Compared with the prior art, the application has the following advantages: the convolution blind source separation method for gas turbine related noise source identification provided by the application converts a time domain convolution problem into an instantaneous complex blind source separation problem in a time-frequency domain by using a synchronous compressive transform (SST); secondly, a normalized boundary target function in a real number domain is extended to a complex number domain, a subgradient optimization method is used for optimization, and an optimization step is adaptively selected, so that the separation precision is ensured; thirdly, a permutation calibration method is used to sort the separated signals at each frequency point, and a minimum distortion algorithm is used to perform amplitude correction; finally, the time-frequency domain signals are converted into time domain signals by using a synchronous compressive inverse transform, and the related source signals are reconstructed. The bounded component analysis reduces the independent assumption requirement of the existing independent component analysis method on the source signals, and the frequency domain convolution blind source separation ensures the accurate separation of the convolution mixed signals. Simulation tests show that the method still has good separation precision for the convolution mixed related source signals. BRIEF DESCRIPTION OF DRAWINGS

[0037] Various other advantages and benefits of the present application will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The accompanying drawings are intended to only illustrate preferred embodiments of the application and are consequently not to be considered limiting of its scope. Obviously, other drawings than those described below can be derived from the drawings described below without paying creative labor, and they are also within the scope of the present application. Moreover, the same reference numbers are used throughout the drawings for the same or similar components.

[0038] In the drawings:

[0039] Figure 1 is a single blind extraction algorithm flow chart of the convolution blind source separation method for gas turbine related noise source identification according to an embodiment of the application;

[0040] Figure 2 is a flow chart of the convolution blind source separation method for gas turbine related noise source identification according to an embodiment of the application;

[0041] Figure 3a 、 Figure 3b is a source signal waveform and spectrum diagram of the convolution blind source separation method for gas turbine related noise source identification according to an embodiment of the application;

[0042] Figure 4 is a mixed filter waveform diagram of the convolution blind source separation method for gas turbine related noise source identification according to an embodiment of the application;

[0043] Figure 5a 、 Figure 5bThe mixed signal waveform and spectrum chart of the convolution blind source separation method for gas turbine related noise source identification according to one embodiment of the application has a signal-to-noise ratio of 10 dB;

[0044] Figure 6a 、 Figure 6b The separated signal waveform and spectrum chart of the convolution blind source separation method for gas turbine related noise source identification according to one embodiment of the application;

[0045] Figure 7 The separated performance comparison chart.

[0046] The application will be further explained below in conjunction with the accompanying drawings and embodiments. DETAILED DESCRIPTION

[0047] The application will be further explained below in conjunction with the accompanying drawings and embodiments. Figures 1 to 7 The specific embodiments of the application will be described in more detail with reference to the drawings. Although the specific embodiments of the application are shown in the drawings, it should be understood that the application can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided so that the application can be more thoroughly understood and the scope of the application can be fully conveyed to those skilled in the art.

[0048] It should be noted that certain terms are used throughout the specification and claims which have particular meanings. Those skilled in the art will appreciate that the same component can be referred to by different terms. The specification and claims do not distinguish components based on differences in terminology, but rather on differences in functionality. As used throughout the specification and claims, "comprising" or "including" is an open term, which should be interpreted to mean "including but not limited to". The subsequent description describes preferred embodiments of the application for the purpose of illustrating the general principles of the application, and not for the purpose of limiting the scope of the application. The scope of the application is defined by the appended claims.

[0049] For the purpose of facilitating the understanding of the embodiments of the application, the following will further explain and describe specific embodiments with reference to the accompanying drawings, and each drawing does not constitute a limitation on the embodiments of the application.

[0050] For a better understanding, Figures 1 to 6b As shown in the figure, the convolution blind source separation method for gas turbine related noise source identification includes:

[0051] Step 1, collecting an observation signal of gas turbine noise, and performing synchronous compression transformation on the observation signal to obtain a corresponding time-frequency domain complex value signal X;

[0052] Step 2, performing instantaneous blind source separation on the time-frequency domain complex signal X at each frequency point, establishing a complex value normalized boundary target function, and iteratively optimizing based on a sub-gradient optimization method at each frequency point, and estimating a separation matrix by performing multiple blind extraction to obtain a complex value separation matrix and a separation signal;

[0053] Step 3, using a power spectrum spectral density distance method to calibrate the separation signal, and using a minimum distortion algorithm to perform amplitude calibration to obtain a time-frequency domain separation signal;

[0054] Step 4, restoring the time-frequency domain separation signal to the time domain by using a synchronous compression inverse transform to obtain a reconstructed time domain separation signal.

[0055] In the preferred embodiment of the convolution blind source separation method for identifying the gas turbine related noise source, after n source signals of the gas turbine noise are mixed, m sensors measure the observation signal, and the convolution mixing model is converted into an instantaneous mixing model at each frequency point by using a synchronous compression transform:

[0056] X(f,t)=A(f)×S(f,t), wherein X(f,t)=[X1(f,t),…,X m (f,t)] T is a mixed signal time-frequency vector, f is a frequency point index, t is a time index, and T is a transpose operation; A(f)=[a1(f),…,a n (f)] is a frequency domain mixing filter; X(f,t)=[X1(f,t),…,X m (f,t)] T is a mixed signal time-frequency vector; S(f,t)=[S1(f,t),…,S n (f,t)] T is a source signal time-frequency vector, and the convolution mixing problem in the time domain is converted into an instantaneous mixing at each frequency point in the frequency domain.

[0057] In the preferred embodiment of the convolution blind source separation method for identifying the gas turbine related noise source, when the number of n source signals and m sensors is the same, the mixing filter has an inverse filter, and the estimated signal is:

[0058] Y f (t)=W f ×X f (t),

[0059] In the formula, W f is a separation matrix at the frequency point f; Y f (t) is a separation signal Y f (f,t) at the frequency point f.

[0060] In the preferred embodiment of the convolution blind source separation method for identifying the noise source of the gas turbine, the normalized boundary objective function of the complex value signal is:

[0061]

[0062] In the formula: The boundary operation is represented by Re(·) represents the real part operation of the complex value vector, Im(·) represents the imaginary part operation of the complex value vector, E(·) represents the expectation operation, the "*" in the formula represents the conjugate operator, and the separation vector is obtained by minimizing the objective function.

[0063] In the preferred embodiment of the convolution blind source separation method for identifying the noise source of the gas turbine, the optimization method of the sub-gradient is used to optimize and derive the separation vector w, and the optimization iteration format of the sub-gradient algorithm is:

[0064] w (p+1) =w (p) -u (p) g (p) ,

[0065] In the formula, g (p) represents the sub-gradient of the objective function J(y) at w (p) , and u (p) represents the step size of the pth iteration.

[0066] In the preferred embodiment of the convolution blind source separation method for identifying the noise source of the gas turbine, the adaptive step size u adap in each iteration is:

[0067]

[0068] In the formula, J min is the minimum value of the objective function, is the non-negative measure of the distance between the current iteration and the minimum value.

[0069] In the preferred embodiment of the convolution blind source separation method for identifying the noise source of the gas turbine, the value of Δ (p) is gradually reduced to 0 in the iteration process.

[0070] In the preferred embodiment of the convolution blind source separation method for identifying the noise source of the gas turbine, the complex value separation matrix W f is obtained by multiple blind extraction, Y f (t) = W f X f (t) is an approximate estimation of the source signal S f (t) as a separation signal.

[0071] In the preferred embodiment of the convolution blind source separation method for identifying the noise source of the gas turbine, in step 4), the separated signals Y f (t) of all frequency points are recovered to the time domain by using the synchronous compression inverse transform to obtain the reconstructed time-domain separated signals:

[0072] y(t) = [y1(t), y2(t), …, y n (t)] T , n represents the number of source signals.

[0073] In the preferred embodiment of the convolution blind source separation method for identifying the noise source of the gas turbine, the normalized cross-correlation coefficient between the source signals and the separated signals is used as the performance evaluation index, and the cross-correlation function of the two signals is calculated as follows:

[0074] C xy (τ) = E{[x(t) - E(x(t))][y(t + τ) - E(y(t))]}

[0075] In the formula, E(·) is the mean value operation, and τ is the moving time.

[0076] The normalized or standardized cross-correlation function is:

[0077]

[0078] In the formula, C xy (τ) is the cross-correlation function, σ x is the standard deviation of x(t), and σ y is the standard deviation of y(t), and the normalized cross-correlation coefficient is finally taken as

[0079] In one embodiment, the method comprises the following steps,

[0080] Step 1), performing synchronous compression transform on the collected observation signals to obtain corresponding time-frequency domain complex-valued signals X;

[0081] Step 2), performing instantaneous blind source separation on the time-frequency domain complex signals at each frequency point, establishing a complex-valued normalized boundary objective function, and iteratively optimizing based on the sub-gradient optimization method at each frequency point, estimating the separation matrix by performing multiple blind extraction, and then preliminarily obtaining the complex-valued separation matrix;

[0082] Step 3), using the power spectrum spectral density distance method to calibrate the separated data, and using the minimum distortion algorithm to perform amplitude calibration;

[0083] Step 4), recovering the time-frequency domain separated signals to the time domain by using the synchronous compression inverse transform, i.e., obtaining the reconstructed time-domain separated signals.

[0084] In step 1), a Synchro Squeezing Transform (SST) is used to transform the mixed signals from time domain to time-frequency domain. Assuming that there are n source signals and m sensors are used to measure the observed signals after the n source signals pass through an unknown mixing system, the convolution mixing model can be approximated as an instantaneous mixing model at each frequency point:

[0085] X(f,t) = A(f) x S(f,t)

[0086] where X(f,t) = [X1(f,t),…,Xm(f,t)]T is the time-frequency vector of the mixed signals, f is the frequency index, t is the time index, T is the transpose operation; A(f) = [a1(f),…,an(f)] is the frequency-domain mixing filter; X(f,t) = [X1(f,t),…,Xm(f,t)]T is the time-frequency vector of the mixed signals; S(f,t) = [S1(f,t),…,Sn(f,t)]T is the time-frequency vector of the source signals, and the convolution mixing problem in time domain is converted into an instantaneous mixing problem at each frequency point in frequency domain. m T n m T n T

[0087] In step 1), when m = n, assuming that the inverse filter of the mixing filter exists, the estimated signal can be written as:

[0088] Y f (t) = W f x X f (t)

[0089] where Wf is the separation matrix at frequency point f; Y(f,t) is the separated signal Y(f,t) at frequency point f. Therefore, the main goal of the frequency-domain convolution blind source separation is to solve the separation matrix Wf at each frequency point f. f f f f

[0090] In step 2), the normalized boundary objective function of the complex-valued signal can be defined as:

[0091]

[0092] where represents the boundary operation, ​​​​​​​​​​​​Re(·) represents the real part operation of a complex value vector, Im(·) represents the imaginary part operation of a complex value vector, E(·) represents the expectation operation, "*" in the formula represents the conjugate operator, and the separation vector is obtained by minimizing the objective function.

[0093] In step 2), the boundary operation of the separation vector in the objective function is not a continuous function, and the objective function is not differentiable, so a subgradient optimization method is considered to optimize the derivation of the separation vector w. The optimization iteration format of the subgradient algorithm is:

[0094] w (p+1) = w (p) - u (p) g (p)

[0095] where g (p) represents the subgradient of the objective function J(y) at w (p) , and u (p) represents the step size of the pth iteration.

[0096] In step 2), the adaptive setting of the step size u in each iteration can ensure the convergence speed and convergence performance of the algorithm. The adaptive step size u adap is:

[0097]

[0098] where J min is the minimum value of the objective function, a non-negative measure of the distance between the current iteration and the minimum value. Applying the above formula to the iteration update process can improve the convergence speed and stability of the algorithm. However, since J min is unknown, the initial value of Δ can be roughly selected and gradually reduced to 0 during the iteration process. The objective function will undergo oscillation and rough descent in the initial stage of iteration. As the iteration continues, the decrease of Δ will make the separation vector oscillate around the minimum value until convergence. The algorithm flowchart of single blind extraction based on bounded component analysis is shown in Figure 1 .

[0099] In step 2), the energy minimization of the subtracted mixed signal is used as the source contribution extraction criterion, and the subtraction method is used to separate multiple source signals. The main idea of the subtraction method is to subtract the separated source signal from the mixed signal, and then continue to separate other source signals from the subtracted mixed signal, and repeat until all source signals are separated. Let h ij represent the contribution coefficient of the ith source signal in the jth mixed signal, and the contribution of the ith source signal in the jth mixed signal is:

[0100]

[0101] In step 2), the accurate separation matrix W can be obtained through multiple blind samplings. f Then Y f (t)=W f ×X f (t) represents the source signal S f An approximate estimate of (t).

[0102] In step 2), the Y values ​​between the separated data are obtained. f (t) There are inconsistencies in order and amplitude. In step 3), the power spectral density cosine distance method and the minimum distortion method are used to correct the order and amplitude of all the data to obtain accurate separation data Y. f (t).

[0103] In step 4), the separated signals Y of all frequency points are... f (t), obtained by restoring the separated signal to the time domain using synchronous compression inverse transform:

[0104] y(t)=[y1(t),y2(t),…,y n (t)] T , where n represents the number of source signals.

[0105] In step 4), to evaluate the performance of this method in estimating the separated signal, the normalized cross-correlation coefficient between the source signal and the separated signal is used as the performance evaluation index. The cross-correlation function of the two signals is calculated as follows:

[0106] C xy (τ)=E{[x(t)-E(x(t))][y(t+τ)-E(y(t))]}

[0107] In the formula: τ is the moving time, and E(·) is the mean value. The normalized or standardized cross-correlation function is:

[0108]

[0109] In the formula: C xy (τ) — Cross-correlation function; σ x —Standard deviation of x(t); σ y —The standard deviation of y(t). Finally, the normalized cross-correlation coefficient is taken as

[0110] The convolutional mixture model used in this invention is: i = 1, ..., n, j = 1, ..., m. Where s(t) = [s1(t), s2(t), ..., s...]. n (t)] T —Source signal; x(t)=[x1(t),x2(t),…,x m (t)]T - mixed signal; A(k) = [a ji (k)] m×n—— Mixed filter coefficients. The goal of multichannel deconvolutive blind source separation is to find a set of separating filters W, such that y(t) is deconvolved. The permutation ambiguity (also known as order uncertainty) is an important factor affecting the performance of frequency domain methods, since the complex-valued BSS operation is performed independently at each frequency bin, resulting in an inconsistent order of separated signals at each frequency bin. The separating matrix W f at each frequency bin satisfies the following equation:

[0111] W f = D f P f

[0112] where D f is a diagonal matrix and P f is a permutation matrix. The ordering ambiguity causes the permutation matrix P f to be inconsistent at each frequency bin, which will cause severe distortion when the time-frequency domain signal is directly converted back to the time domain. Only when the same column of the estimated mixing matrix at all frequency bins corresponds to the same source signal, i.e., the permutation matrix P f is consistent at each frequency bin, can the source signal be correctly separated. The ordering calibration is performed by using the power spectral density (PSD) distance method, which uses the PSD of a signal as a similarity sequence and uses the cosine distance as an optimization index. The cosine distance between vectors x and y can be expressed as:

[0113]

[0114] where <·> is the inner product operation of vectors and ||·||2is the modulus operation. Obviously, the more similar x and y are, the smaller the value of dc(x, y) is. The criterion for determining the permutation matrix is to minimize the cosine distance of the similarity sequence between adjacent frequency bins.

[0115] The minimal distortion principle (MDP) is used to minimize the distortion of the signal. According to the description of the minimal distortion principle, the output signal Y(f, t) should be as close to the input signal X(f, t) as possible, and the separating matrix that meets the minimal distortion principle is:

[0116]

[0117] Finally, the modified signal is synchronized and compressed to obtain the separated signal in time domain. The application does not need to know the prior knowledge of the noise in the signal, and does not need to assume that all source signals are independent of each other. When the source signals are correlated, the application still has good estimation performance and good robustness. The application provides effective support for actual vibration noise source separation. The flow chart of the method of the application is shown in Figure 2 .

[0118] The performance of the application is illustrated by simulation experiments. Artificial convolution mixed mechanical signals are used for blind source separation experiments to show the separation effect. The source signals used in the experiment have correlation between them.

[0119] Three typical signals are set to simulate the multi-rotor harmonic noise signals of a gas turbine, and the three source signals are mixed with the convolution filter to generate a convolution mixed signal. The expression of the source signal is as follows:

[0120]

[0121] In the simulation experiment, s2(t) and s3(t) both contain the main frequency 213Hz, and the two signals are source correlated signals. The correlation coefficient between each source signal is calculated as:

[0122]

[0123] Among them, the correlation coefficient between s1(t) and s2(t), s3(t) is close to 0, which indicates that the correlation between s1(t) and s2(t), s3(t) is low, and the correlation coefficient between s2(t) and s3(t) is 0.4168, which indicates that s2(t) and s3(t) are correlated source signals. The sampling frequency is set to 12800Hz, and the sampling time length is T=1s. The waveform graph and the spectrum graph of the source signal s are shown in Figure 3a 、 Figure 3b In order to intuitively show the waveform and spectrum characteristics of the signal, only the 0-0.1s part is shown in the time domain waveform graph, and only the 0-800Hz part is shown in the spectrum graph.

[0124] The number of observations m=3, and the number of sources n=3. The mixing filter uses a single-sided oscillation attenuation FIR filter with order R=20, Figure 4 The mixing filter waveform graph is shown in

[0125] Figure 5a 、 Figure 5b The mixed signal waveform and spectrum at a signal-to-noise ratio of 3dB are shown in

[0126] The normalized cross-correlation coefficient between the source signal and the separated signal is used as the performance evaluation index. The cross-correlation function of two signals is calculated as follows:

[0127] C xy (τ) = E{[x(t) - E(x(t))][y(t + τ) - E(y(t))]}

[0128] where E(·) is the operation of taking the mean value, and τ is the moving time. The normalized or standardized cross-correlation function is:

[0129]

[0130] where C xy (τ) is the cross-correlation function, σ x——x(t) is the standard deviation of x(t), and σ y—— is the standard deviation of y(t). Finally, the normalized cross-correlation coefficient is:

[0131] The separated signal waveform and spectrum obtained by the method of the present application are shown in Figure 6a , Figure 6b The cross-correlation coefficient of the separated signal and the source signal is:

[0132]

[0133] It can be seen preliminarily that, when the source signals are correlated, the three source signals can be well estimated by the method, and a good separation effect can be achieved. It should be noted that the estimated source signals are with noise, and therefore the separated signal waveform has burrs compared with the source signal.

[0134] In order to objectively describe the separation performance of the method, the method is compared with the IVA method and the BCA method based on volume ratio, and the waveform correlation coefficient (p) is used as a performance index to evaluate the separation performance. The correlation coefficient curves of the three separation methods under different signal-to-noise ratios are shown in Figure 7 , and the values are the average values of 10 experiments. It can be seen from Figure 7 that, under different signal-to-noise ratios, the separation performance of the traditional IVA method is the worst. Under different signal-to-noise ratios, the separation performance of the method is better than that of the BCA method based on volume ratio, indicating that the method has good robustness and separation performance.

[0135] Although the embodiments of the present application are described above in combination with the drawings, the present application is not limited to the above specific embodiments and application fields, and the above specific embodiments are only illustrative and guiding, but not limiting. Those skilled in the art can make many forms under the guidance of the present application and without departing from the scope protected by the claims of the present application, and these all belong to the protection of the present application.

Claims

1. A method of convolutional blind source separation for gas turbine related noise source identification, characterized in that, It comprises, Step 1, collecting the observation signal of the gas turbine noise, and performing synchronous compression transformation on the observation signal to obtain a corresponding time-frequency domain complex value signal X; Step 2, performing instantaneous blind source separation on the time-frequency domain complex signal X at each frequency point, establishing a complex value normalized boundary target function, and iteratively optimizing based on a subgradient optimization method at each frequency point, and obtaining a complex value separation matrix and a separation signal by performing multiple blind extraction estimation; Step 3, using a power spectrum spectral density distance method to calibrate the separation signal, and using a minimum distortion algorithm to perform amplitude calibration to obtain a time-frequency domain separation signal; Step 4, the time-frequency domain separation signal is recovered to the time domain by using the synchronous compression inverse transform to obtain the reconstructed time domain separation signal, the separated source signal is subtracted from the mixed signal, and other source signals are continuously separated from the subtracted mixed signal, and the cycle is repeated until all source signals are separated, and h ij represents the contribution coefficient of the i-th source signal in the j-th mixed signal, and the contribution of the i-th source signal in the j-th mixed signal: The optimization method of the subgradient is used to optimize and derive the separation vector w, and the optimization iteration format of the subgradient algorithm is: w (p+1) = w (p) - u (p) g (p) , where g (p) represents the subgradient of the objective function J(y) at w (p) , u (p) represents the step size at the pth iteration, and the adaptive step size u adap is given by: where J min is the minimum of the objective function, Δ (p) is a non-negative measure of the distance of the current iteration from the minimum, and Y f (t) is the separated signal for all frequencies, the separated signals are recovered to the time domain using a synchronized inverse compression transform to obtain the reconstructed time-domain separated signals: y(t) = [yl(t), y2(t),..., yn(t)]T n (t)] T , n denotes the number of source signals.

2. The convolutive blind source separation method for gas turbine related noise source identification according to claim 1, characterized in that, After the n source signals of the gas turbine noise are mixed, m sensors measure the observation signal, and the convolution mixing model is converted into an instantaneous mixing model at each frequency point f by synchronous compression transformation: X(f, t) = A(f) x S(f, t), where X(f, t) = [X1(f, t),..., XN(f, t)]T, A(f) = [a1(f),..., aN(f)]T, S(f, t) = [S1(f, t),..., SN(f, t)]T, and N is the number of sources. m (f, t) T , which is a mixed signal time-frequency vector, f is a frequency index, t is a time index, T is a transpose operation; A(f) = [a1(f),..., aN(f)]T, which is a frequency domain mixed filter; S(f, t) = [S1(f, t),..., SN(f, t)]T, which is a source signal time-frequency vector, and N is the number of sources. n (f, t) n , which is a mixed signal time-frequency vector, f is a frequency index, t is a time index, T is a transpose operation; A(f) = [a1(f),..., aN(f)]T, which is a frequency domain mixed filter; S(f, t) = [S1(f, t),..., SN(f, t)]T, which is a source signal time-frequency vector, and N is the number of sources. T (f, t) T , which is a mixed signal time-frequency vector, f is a frequency 3. The convolutive blind source separation method for gas turbine related noise source identification according to claim 2, characterized in that, When the number of n source signals and m sensors is the same, the inverse filter of the mixing filter exists, and the estimated signal is: Y f (t) = W f x X f (t), where: W f is the separation matrix at frequency f; X f (t) is the separated signal at frequency f.

4. The method of claim 1, wherein, The normalized boundary target function of the complex value signal is: wherein: denotes a boundary operation, Re(·) denotes a real part operation on a complex valued vector, Im(·) denotes an imaginary part operation on a complex valued vector, E(·) denotes an expectation operation, the "*" in the formulas denotes a conjugate operator, and the separated vectors are obtained by minimizing an objective function.

5. The method of claim 4, wherein the method further comprises: Δ (p) the values of the are gradually reduced to 0 during the iteration process.

6. The method of claim 1, wherein, The normalized cross-correlation coefficient between the source signal and the separation signal is used as a performance evaluation index, and the cross-correlation function of the two signals is calculated as follows: C xy (r) = E{[x(t) - E(x(t))][y(t + r) - E(y(t))]}, In the formula, E(·) is a mean value operation, τ is a moving time, The normalized or standardized cross-correlation function is: where: C xy (τ) is the cross-correlation function; σ x is the standard deviation of x(t); σ y is the standard deviation of y(t), and finally the normalized cross-correlation coefficient is taken as

Images