Harmonic modal detection and removal method based on blind source separation
Through the blind source separation method, the problems of low efficiency and susceptibility to near-frequency interference of traditional harmonic detection methods are solved, and efficient separation and reconstruction of rotating mechanical vibration signals are realized, improving the accuracy and efficiency of modal parameter recognition.
Patent Information
- Application Number
- CN202311160536.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-09-08
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2043-09-08
AI Technical Summary
Traditional harmonic detection methods are inefficient and susceptible to near-frequency interference when processing vibration response signals of rotating machinery, and may lose modal information during the removal process, making it difficult to accurately identify the real structural modal parameters of rotating machinery.
Harmonic mode detection and removal method based on blind source separation is adopted, and harmonic modes are automatically separated through technologies such as polynomial trend term elimination, short-time Fourier transform, energy maximum method, L1 norm minimum method and probability density curve analysis. The state and system modal signals realize clear decomposition and reconstruction of the signals, avoiding human subjective judgment and near-frequency interference.
Improve the accuracy and efficiency of rotating mechanical modal parameters recognition, can effectively separate harmonic modes with few measurement points, retain the integrity of the system modal signal, avoid near-frequency interference of the spectrum curve, and significantly improve detection efficiency and accuracy.
Smart Images

Figure CN119004244A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of vibration signal analysis of mechanical structures, and in particular to a method for processing vibration response signals of a rotating mechanical structure in a running state. Background Art
[0002] Rotating mechanical structures are usually excited by periodic forces during operation. The structural vibration under gas turbine operation mainly comes from multi-harmonic excitation caused by rotor imbalance and critical speed, which will induce harmonic modal interference in the structural mode. The structural vibration during operation is a combination of the response to random disturbances and harmonic excitation caused by rotating parts, and the complex periodic aerodynamic effects during gas turbine operation will also introduce harmonic interference. These harmonics usually appear as false resonance peaks in the response spectrum to disrupt the modal parameter identification of the real structure. Therefore, the detection and removal of mechanical structure harmonic modes is a prerequisite for the correct identification of modal parameters under operation.
[0003] Traditional harmonic detection methods are mainly divided into two categories: time domain and frequency domain. The time domain method requires single-frequency filtering of the frequency components to be detected, which has low detection and calculation efficiency and requires subjective judgment. Although the frequency domain method does not require digital filtering to extract frequency components, has high calculation efficiency, and has the characteristics of synchronous detection of full-band components, when processing response signals containing multiple harmonics, the spectrum curve of harmonic detection will have close-frequency interference, which is difficult to identify, and the modal information of the original signal may be lost during the harmonic removal process.
[0004] Based on this, a harmonic mode detection and removal method based on blind source separation is studied. This method does not require prior information on harmonic frequencies, and can automatically separate harmonic modes from system modal signals. Moreover, the blind source separation method used can still achieve effective signal separation under underdetermined conditions, that is, when the number of measurement points is less than the number of modes; harmonic detection is performed on each separated source signal, avoiding the process of subjective selection of frequency components to be detected and filtering, and the detection efficiency is high; and the blind separation process actually decomposes complex signals into simple signals, which is clearer and more accurate when performing frequency domain detection, avoiding the phenomenon of near-frequency interference; in the process of source signal reconstruction, the complete information of the system mode is retained, which is conducive to the subsequent modal parameter identification.
[0005] The above information disclosed in this Background section is only for enhancement of understanding of the background of the invention and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the invention
[0006] In view of the problems existing in the prior art, the present invention proposes a harmonic modal detection and removal method based on blind source separation, which is based on the high similarity between modal expansion and blind source separation principles to overcome the problems of low detection efficiency and the need for human subjective judgment in harmonic modal detection and removal.
[0007] The purpose of the present invention is to achieve the following technical solution: a harmonic mode detection and removal method based on blind source separation includes:
[0008] Step 1, collecting vibration response acceleration signals of various measuring points under the running state of the rotating machinery;
[0009] Step 2, performing polynomial trend term elimination and smoothing processing on the vibration response acceleration signal to obtain a signal sequence;
[0010] Step 3, obtaining the time-frequency domain by short-time Fourier transform of the signal sequence;
[0011] Step 4, clustering the time-frequency scattered points in the time-frequency domain based on the maximum energy method to obtain an estimated value of the mixing matrix;
[0012] Step 5, based on the L1 norm minimization method, the source recovery problem is transformed into an optimization problem to achieve source signal recovery;
[0013] Step 6, performing probability density curve calculation on the source signal to find the source signal with harmonic characteristics;
[0014] Step 7, performing spectral kurtosis detection on the source signal with harmonic characteristics to obtain the harmonic modal frequency;
[0015] Step 8, directly remove the source signal with obvious harmonic characteristics in the probability density curve, and filter and remove the source signal with unclear harmonic characteristics based on the harmonic modal frequency, wherein when the probability density curve has two peaks, the harmonic characteristics are obvious, and if only one peak can be seen, the harmonic characteristics are not obvious;
[0016] Step 9, combining other source signals containing system characteristics and source signals after removing harmonic frequencies, and combining with a mixing matrix to achieve signal reconstruction, the reconstructed signal is a signal after removing harmonics;
[0017] Step 10, comparing the vibration response acceleration signal with the signal after harmonic removal, the difference before and after harmonic removal can be clearly and intuitively seen, and the signal after harmonic removal is the output result.
[0018] In the harmonic modal detection and removal method based on blind source separation, in step 1, the number of measurement points arranged in the rotating machinery is less than the sum of the harmonic modal number and the system modal number, wherein the harmonic modal number is obtained by harmonic detection and the system modal number is obtained by modal analysis.
[0019] In the harmonic mode detection and removal method based on blind source separation, in step 2, in the elimination of polynomial trend terms, the measured vibration response acceleration signal sampling data is {x k}(k=1,2,3,…,n), n represents the number of sampling points, and the polynomial function is set to:
[0020]
[0021] In the formula, a j (j=0, 1, ..., m) are the coefficients of the polynomial to be determined, m is the highest order of the polynomial, and m=1 to 6 is used to determine the function The unknown coefficients of With x k The sum of squared errors is the smallest, that is:
[0022]
[0023] E represents the sum of squares of the errors of the two functions. To minimize E, the extreme value condition is:
[0024]
[0025] Taking the partial derivative of E with respect to a in turn, a system of linear equations with m+1 variables is generated:
[0026]
[0027] Solve the system of equations and find the unknown coefficients;
[0028] The formula for eliminating the polynomial trend term is:
[0029]
[0030] Five-point multiple smoothing includes:
[0031]
[0032]
[0033]
[0034]
[0035] In the above formula, x i(i=1, ..., n) is the measured vibration response acceleration signal sequence, n represents the number of sampling points, y i (i=1, ..., n) is the signal sequence after smoothing.
[0036] In the harmonic mode detection and removal method based on blind source separation, in the short-time Fourier transform, the formula of the frequency energy function is first defined as:
[0037]
[0038] Where: E(f) is the frequency energy value; X(t, f) is the signal after time-frequency transformation; t represents time, f represents frequency, and ∞ represents infinity; in the time-frequency domain, the energy distribution of a single channel in the frequency domain is first calculated, and then the energy of multiple channels is added at the same frequency point, that is:
[0039]
[0040] Where: E s (f) is the energy sum of the signals received by all sensors in the frequency domain; Real (X i (t, f)) is the real part of the i-th sensor signal after time-frequency transformation; Img(X i (t, f)) is the imaginary part of the i-th sensor signal after time-frequency transformation; m is the number of measurement channels, t represents time, f represents frequency, and the peak frequency points are extracted using the peak detection method.
[0041] In the harmonic mode detection and removal method based on blind source separation, in step 5, based on the L1 norm minimization method, for all points in the sparse time-frequency domain:
[0042]
[0043] Where: s i (t, f) are the source signals; A is the mixing matrix, x(t, f) is the measurement channel signal; st is the constraint condition; t represents time, f represents frequency; the L1 norm of the solution is required to be the smallest, where s(t, f) is the optimal estimate of the source signal.
[0044] In the harmonic mode detection and removal method based on blind source separation, in step 6, it is assumed that the mixed signal x(n) is composed of a Gaussian white noise signal and a harmonic signal, that is:
[0045] x(n)=g(n)+h(n),
[0046] Where n = 1, 2, ..., N, N is the number of sampling points; g(n) is the Gaussian signal sequence and h(n) is the harmonic signal sequence. The sparse component analysis method is used to separate the g(n) and h(n) signals. The ideal Gaussian white noise signal conforms to the normal distribution and its probability density is:
[0047]
[0048] The probability densities of the two random variables X and T are f X (x) and f T (t), and x=h(t), according to the joint probability density function theory, we can get:
[0049] f X (x) = f t [h(x)]|h′(x)|, α<x<β,
[0050] Where h′(x) is the first-order derivative of h(x), α and β are the range of values of x, and for harmonic signals, it is considered as:
[0051] x = h(t) = a sin(2πft),
[0052] The time t belongs to uniform distribution, and its probability density function is a constant value f t , then the probability density function of the harmonic response is expressed as:
[0053]
[0054] For any frequency f, when the harmonic has a constant amplitude a, the above formula is always valid. When the mean value of the harmonic component is 0, that is, x→a or x→-a, the probability density function tends to infinity, and there are two obvious peaks in the probability density curve.
[0055] In the harmonic mode detection and removal method based on blind source separation, in step 7, the spectral kurtosis can be defined as:
[0056]
[0057] Where: E{|X(f)| 4} is the fourth-order cumulant; E[|X(f)| 2 ] is the second-order cumulant,
[0058] For a mixed signal of Gaussian white noise signal and harmonic signal, its Fourier transform result is:
[0059] X(f)=G(f)+H(f),
[0060] The kurtosis value of the mixed signal spectrum is:
[0061]
[0062] In the formula, f0 is the harmonic frequency value, and the spectral kurtosis value at the corresponding frequency is -1, so it is determined to be a harmonic component.
[0063] In the harmonic modal detection and removal method based on blind source separation, the rotating machinery includes a gas turbine or a wind turbine.
[0064] In the harmonic mode detection and removal method based on blind source separation, multiple accelerator sensors are used to collect vibration response acceleration signals of various measuring points under the operating state of the rotating machinery.
[0065] Compared with the prior art, the present invention has the following advantages: the present invention combines sparse component analysis, probability density and spectral kurtosis to detect and remove harmonic modes, thereby improving the accuracy of operating modal parameter identification. The method is based on the high similarity between modal expansion theory and blind source separation theory, and can detect harmonic modes with only a few measurement points, without any prior information on harmonics, and retains the integrity of system modal signals during the detection and removal of harmonics. BRIEF DESCRIPTION OF THE DRAWINGS
[0066] By reading the detailed description of the preferred specific embodiments below, various other advantages and benefits of the present invention will become clear to those of ordinary skill in the art. The drawings in the specification are only for the purpose of illustrating the preferred embodiments and are not considered to be limitations of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without creative work. Moreover, the same reference numerals are used to represent the same components throughout the drawings.
[0067] In the attached picture:
[0068] Figure 1 It is a flow chart of a harmonic mode detection and removal method based on blind source separation provided by an embodiment of the present invention;
[0069] Figure 2 is a schematic diagram of a three-degree-of-freedom mass spring damping system provided by an embodiment of the present invention;
[0070] Figure 3 is a time-domain three-dimensional scatter plot of a collected vibration response signal provided by an embodiment of the present invention;
[0071] Figure 4 A three-dimensional scatter plot in the time-frequency domain after sparse transformation provided by an embodiment of the present invention;
[0072] Figure 5is a three-dimensional scatter plot of frequency domain energy provided by an embodiment of the present invention;
[0073] FIG. 6( a ) and FIG. 6( b ) are probability density curve harmonic detection diagrams provided by one embodiment of the present invention;
[0074] Figure 7 is a spectral kurtosis curve harmonic detection diagram of source signal 1 provided by an embodiment of the present invention;
[0075] Figure 8 is a spectral kurtosis curve harmonic detection diagram of the source signal 3 provided by an embodiment of the present invention;
[0076] Fig. 9 is an energy peak diagram of a harmonic mode provided by an embodiment of the present invention;
[0077] Fig.10 This is an energy peak diagram after removing the harmonic mode provided by an embodiment of the present invention.
[0078] The present invention is further explained below in conjunction with the accompanying drawings and embodiments. DETAILED DESCRIPTION
[0079] The following will refer to the attached Figures 1 to 10 Specific embodiments of the present invention are described in more detail. Although specific embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.
[0080] It should be noted that certain words are used in the specification and claims to refer to specific components. Those skilled in the art should understand that technicians may use different nouns to refer to the same component. This specification and claims do not use the difference in nouns as a way to distinguish components, but use the difference in the functions of the components as the criterion for distinction. As mentioned throughout the specification and claims, "including" or "comprising" is an open term, so it should be interpreted as "including but not limited to". The subsequent description of the specification is a preferred embodiment of the present invention, but the description is based on the general principles of the specification and is not intended to limit the scope of the present invention. The scope of protection of the present invention shall be determined by the attached claims.
[0081] To facilitate understanding of the embodiments of the present invention, further explanation will be given below by taking specific embodiments as examples in conjunction with the accompanying drawings, and each of the accompanying drawings does not constitute a limitation on the embodiments of the present invention.
[0082] like Figures 1 to 10 As shown, the harmonic mode detection and removal method based on blind source separation includes:
[0083] Step 1, collecting vibration response acceleration signals of various measuring points under the running state of the rotating machinery;
[0084] Step 2, performing polynomial trend elimination and smoothing processing on the vibration signal;
[0085] Step 3, performing time-frequency transformation on the time-domain vibration signal to convert it into a time-frequency domain with better sparsity;
[0086] Step 4, clustering each time-frequency scatter point based on the maximum energy method to obtain an estimated value of the mixing matrix;
[0087] Step 5: Based on the L1 norm minimization method, the source recovery problem is transformed into an optimization problem to achieve source signal recovery;
[0088] Step 6, calculate the probability density curve of each source signal to find the source signal with harmonic characteristics;
[0089] Step 7, focusing on performing spectral kurtosis detection on the source signal with harmonic characteristics, so as to obtain the harmonic modal frequency;
[0090] Step 8, directly remove the source signal with obvious harmonic characteristics in the probability density curve, and filter and remove the source signal with unclear harmonic characteristics according to the harmonic mode frequency;
[0091] Step 9, combining other source signals containing system characteristics and source signals after removing harmonic frequencies, combining with a mixing matrix, to achieve signal reconstruction, and the reconstructed signal is the signal after removing harmonics;
[0092] Step 10, compare the original signal with the signal after removing harmonics.
[0093] The harmonic mode detection and removal method based on blind source separation can handle the situation of underdetermined blind source separation. The number of measurement points is usually less than the sum of the number of harmonic modes and the number of system modes. This method is suitable for the situation of few measurement points.
[0094] In the harmonic mode detection and removal method based on blind source separation, in step 2, in terms of eliminating the polynomial trend term, the measured vibration response acceleration signal sampling data is {x k}(k=1,2,3,…,n), n represents the number of sampling points, and the polynomial function is set to:
[0095]
[0096] In the formula, a j (j=0, 1, ..., m) are the coefficients of the polynomial to be determined, m is the highest order of the polynomial, and m=1 to 6 is used to determine the function The unknown coefficients of With x k The sum of squared errors is the smallest, that is:
[0097]
[0098] E represents the sum of squares of the errors of the two functions. To minimize E, the extreme value condition must be met:
[0099]
[0100] Taking the partial derivative of E with respect to a in turn, a system of linear equations with m+1 variables is generated:
[0101]
[0102] Solve the system of equations and find the unknown coefficients;
[0103] The formula for eliminating the polynomial trend term is:
[0104]
[0105] Five-point multiple smoothing includes:
[0106]
[0107]
[0108]
[0109]
[0110] In the above formula, x i (i=1, ..., n) is the original vibration response acceleration signal sequence, n represents the number of sampling points, y i (i=1, ..., n) is the signal sequence after smoothing.
[0111] In the harmonic mode detection and removal method based on blind source separation, in step 3, selecting a suitable sparse transformation method is a prerequisite for executing the sparse component analysis method. The short-time Fourier transform in the time-frequency transform is the simplest and most effective method to achieve sparseness, which can be written as the following formula:
[0112]
[0113] In the formula, w(τ-t) represents the window function moving with time t, x(τ) is the truncated signal, X(t, f) is the signal after time-frequency transformation, and the short-time Fourier transform considering the phase difference is:
[0114]
[0115] According to Parseval's theorem, this formula can be written as:
[0116]
[0117] In the formula, is the Fourier transform of the window function, and the frequency holding range is [-Δ, Δ], is the Fourier transform of the initial signal. For two harmonic signals of different frequencies:
[0118]
[0119]
[0120] In the formula, f1 and f2 represent the corresponding harmonic frequencies; A1 and A2 represent the corresponding amplitudes. The corresponding short-time Fourier transform can be written as:
[0121]
[0122]
[0123] The two source signals are linearly and instantaneously mixed and measured by the two measurement channels:
[0124] y2(t)=a 11 ·x1(t)+a 12 ·x2(t)
[0125] y2(t)=a 21 ·x1(t)+a 22 ·x2(t)
[0126] Where y1(t) and y2(t) represent the mixed measurement channel signals, and a 11 、a 12 、a 21 、a 22 Represents the mixing coefficient. Based on the linear characteristics of short-time Fourier transform, the short-time Fourier transform of the two measurement channels can be written as:
[0127] Y1(t,f)=α 11 ·X1(t,f)+a 12 X2(t, f)
[0128] Y2(t,f)=a 21 ·X1(t,f)+a 22 X2(t, f)
[0129] It should be noted that the interval of harmonic frequencies should be larger than the frequency retention range of the window function, that is, |f1-f2|>2Δ, so a window function with a larger window length is selected to ensure that the frequency retention range is small enough. There is no overlap between the two harmonic signals in the time-frequency domain. For f∈[f1-Δ, f1+Δ], the following formula can be constructed:
[0130]
[0131] For f∈[f2-Δ, f2+Δ], the following formula can be constructed:
[0132]
[0133] In the formula, Re represents the real part of the complex number. From the above formula, it can be seen that when drawing the real part scatter plot of the short-time Fourier transform of the measurement channel signal in this case, for the frequency point, it is actually equivalent to drawing a series of special Lissajous figures, which are straight lines in this example, and the direction of the straight line is the column vector direction of the mixing matrix, such as or These cluster lines appear in the scatter plot, which is also the theoretical basis for estimating the mixing matrix.
[0134] In the harmonic mode detection and removal method based on blind source separation, in step 4, the mixing matrix is estimated based on the maximum energy method. For estimating the mixing matrix of the source signal, there are mainly potential function method and clustering method. The potential function method is to expand the angle of the clustering line to the polar coordinate axis, and determine the angle between the clustering line and the coordinate axis by calculating the peak point, so as to estimate each column vector of the mixing matrix, but it can only be used for two-channel mixed signals, which has certain limitations; while estimating the center of the clustering line to determine the mixing matrix is not limited by the number of channels, so it is widely used. Whether it is at the harmonic frequency or the structural resonance frequency, the energy amplitude corresponding to the frequency is very large, so it is reasonable to separate the harmonic mode and system mode by clustering through the maximum energy method. From the perspective of signal energy, the signal energy in the frequency band is estimated. Based on the short-time Fourier transform of the signal, the formula of the frequency energy function is first defined as:
[0135]
[0136] Where: E(f) is the frequency energy value; X(t, f) is the signal after time-frequency transformation, t represents time, f represents frequency, and ∞ represents infinity; in the time-frequency domain, the energy distribution of a single channel in the frequency domain is first calculated, and then the energy of multiple channels is added at the same frequency point, that is:
[0137]
[0138] Where: E s(f) is the energy sum of the signals received by all sensors in the frequency domain; Real (X i (t, f)) is the real part of the i-th sensor signal after time-frequency transformation; Img(X i (t, f)) is the imaginary part of the i-th sensor signal after time-frequency transformation; m is the number of measurement channels, t represents time, and f represents frequency. The peak frequency points are extracted using the peak detection method. If the commonly used zero derivative method is used in the peak detection method, the first-order derivative will accidentally cross zero due to the presence of noise in the measured signal, resulting in false detection. If some low-pass filters are used to smooth the curve, the effective information of the signal will be lost to a certain extent. Therefore, the method of finding peaks and valleys is adopted here. A peak is defined by having lower points around it, and a peak threshold e is set. It is required that there is at least a difference of the threshold level between the peak and the surrounding points to define it as a peak.
[0139] In the harmonic mode detection and removal method based on blind source separation, in step 5, the source signal is recovered using the L1 norm minimum method. Since the number of unknown sources is greater than the number of equations, the equation has no unique solution. A constraint condition is introduced to require the L1 norm of the solution to be the minimum, and the source recovery is converted into an optimization problem. For all points in the sparse domain:
[0140]
[0141] Where: s i (t, f) are the source signals; A is the mixing matrix, x(t, f) is the measurement channel signal, t represents time, f represents frequency, and st is the constraint condition; this formula converts source recovery into an optimization problem. Since the number of unknown sources is greater than the number of equations, the equation has no unique solution. The constraint condition is introduced to require the L1 norm of the solution to be the smallest, where s(t, f) is the optimal estimate of the source signal. Theoretically, it is more intuitive to use the L0 norm for this equation. The L0 norm represents the number of non-zero elements in the vector, and naturally, a sparse solution with the least non-zero elements among all feasible solutions can be found. However, this constraint equation is non-convex and highly discrete, and it is difficult to solve it by numerical calculation. However, the sparse representation theory points out that if the solution is indeed sparse, the L1 norm, as the optimal convex approximation of the L0 norm, can obtain the same sparse solution as the L0 norm and is easier to optimize and solve.
[0142] Taking the signals of two measurement channels as an example, the constraints show that x(t) must be linearly decomposed in the direction of any two column vectors in the mixing matrix in order to find the shortest path from O to the observed signal x(t). a1a2a3 are the three column vectors of the mixing matrix, and point E is any point of the observed signal. At this time, the mixing matrix has no inverse matrix, and the solution with the smallest L1 norm is the solution that satisfies the shortest distance from point E to the origin. Reduce the dimension of the mixing matrix to generate three 2×2 sub-matrices (that is, three combinations of the column vectors of the mixing matrix). Assuming that point E is between vectors a2a3, the sub-matrix composed of a1a2 is the optimal solution matrix, and OBOC is the optimal solution for point Z in the direction of a2a3. The inverse of the sub-matrix is multiplied by the observation point to obtain an estimate of the source signal. The specific steps of the L1 norm minimization algorithm are as follows:
[0143] (1) Find the mixing matrix There are m×m dimensional sub-matrices, let B k ,
[0144] (2) Find all possible solutions for a point X in the sparse domain.
[0145] (3) Find the L1 norm of each solution and take the one with the smallest norm. As the optimal estimate of the source signal,
[0146] (4) Repeat the above steps to find the optimal solution for all points in the sparse domain.
[0147] (5) Perform a sparse inverse transform on the source signal to obtain a time domain estimate of the source signal.
[0148] In the harmonic mode detection and removal method based on blind source separation, in step 6, a probability density function is calculated by performing a probability density curve on each source signal. The probability density function represents the probability that the signal amplitude falls within a specified area. The detection method is based on the estimation of the probability density curve of the harmonic signal and the Gaussian white noise signal. The core idea is to estimate the probability density curve of the separated single-frequency time domain signal and distinguish the harmonic response and the random response according to the different statistical characteristics.
[0149] Assume that the mixed signal x(n) consists of Gaussian white noise signal and harmonic signal, that is:
[0150] x(n)=g(n)+h(n)
[0151] Where n = 1, 2, ..., N, N is the number of sampling points; g(n) is the Gaussian signal sequence and h(n) is the harmonic signal sequence. The sparse component analysis method mentioned above can be used to separate the g(n) and h(n) signals, so that they can be studied separately. The ideal Gaussian white noise signal conforms to the normal distribution and its probability density is:
[0152]
[0153] Although in actual engineering, the structure will be stimulated by multiple independent Gaussian white signals with different means and variances, and may not be an ideal Gaussian white noise signal, its probability density curve is still approximately normal distribution, and even if a series of signal processing is performed on the signal, the distribution form of its probability density function will not change.
[0154] Assume that the probability densities of the two random variables X and T are f x (x) and f T (t), and x=h(t), according to the joint probability density function theory, we can get:
[0155] f X (x) = f t [h(x)]|h′(x)|, α<x<β
[0156] Where h′(x) is the first-order derivative of h(x), and α and β are the range of values of x. For harmonic signals, consider:
[0157] x=h(t)=asin(2πft)
[0158] The time t belongs to uniform distribution, and its probability density function is a constant value f t , then the probability density function of the harmonic response is expressed as:
[0159]
[0160] For any frequency f, when the harmonic has a constant amplitude a, the above formula always holds. When the mean value of the harmonic component is 0, that is, x→a or x→-a, the probability density function tends to infinity, and there will be two obvious peaks in the probability density curve.
[0161] In the harmonic mode detection and removal method based on blind source separation, in step 7, a spectral kurtosis curve is calculated for each source signal, and the harmonic frequency is detected.
[0162] The spectral kurtosis can be defined as:
[0163]
[0164] Where: E{|X(f)| 4}-fourth-order cumulant; E[|X(f)| 2 ]-Second-order cumulant.
[0165] Divide the time series into M segments and perform Fourier transform X on each segment i Substituting (f) (i = 1, 2, ..., M) into the above formula, we can obtain the unbiased estimate of the fourth-order cumulant:
[0166]
[0167] Similarly, the unbiased estimator of the second-order cumulant is obtained:
[0168]
[0169] Finally, the unbiased estimator of spectral kurtosis is obtained:
[0170]
[0171] For a linear system, its response will not change the input signal characteristics. For a Gaussian white noise signal g(n) after Fourier transform is G(f), the spectral kurtosis value SK is calculated by the above formula g (f)=0.
[0172] Assume the frequency is f0 and the phase is The harmonic signal is:
[0173]
[0174] After Fourier transform:
[0175]
[0176] Among them, δ(f0) is the Dirac function, and the spectral kurtosis value SK at f0 is calculated by the formula h (f0) = -1.
[0177] The linear property of Fourier transform, for the mixed signal of Gaussian white noise signal and harmonic signal, its Fourier transform result is:
[0178] X(f)=G(f)+H(f)
[0179] In summary, the kurtosis value of the mixed signal spectrum is:
[0180]
[0181] Therefore, the spectral kurtosis value at the corresponding frequency can be -1, thereby determining that it is a harmonic component.
[0182] In the harmonic mode detection and removal method based on blind source separation, in step 9, the detected harmonic signal is removed. If the probability density indicates that the entire source signal is a harmonic signal, the entire source signal is directly removed. For example, the i-th source signal is a harmonic signal S i (t, f), the S in the source signal i (t, f) is removed to obtain S R (t, f), and the corresponding a in the mixing matrix i Remove to get A R , and obtain the reconstructed signal X R (t, f) = A R S R (t, f), and then convert the signal to the time domain to get x R (t). If the harmonic energy accounts for too small a proportion of the entire source signal, the probability density cannot clearly indicate that it is a harmonic, but the spectral kurtosis method indicates the harmonic frequency in the source signal, then there is no need to remove the entire source signal. Instead, the stopband filter can be used to remove the corresponding harmonic frequency components before participating in the subsequent signal reconstruction.
[0183] In step 10, the initial signal is compared with the signal after the harmonics are removed to obtain a more intuitive result.
[0184] In one embodiment, Figures 1 to 10 As shown, a harmonic mode detection and removal method based on blind source separation includes:
[0185] 1) Measure the vibration response acceleration signals of all measuring points. In order to explain the process and results of this method more clearly and concisely, a three-degree-of-freedom mass spring damping simulation system is established here, as shown in the attached figure. Figure 2 As shown in the figure, white noise excitation is applied to it, and the vibration response acceleration signal is measured. The method is explained based on this simulation data. The system parameters are M1 = M2 = M3 = 1 kg, K1 = K2 = 1 × 10 6 N / m, K3=1.5×10 6 N / m, C=M+10 -6 K, then the theoretical values of its modal parameters can be obtained through modal analysis theory, as shown in Appendix 1.
[0186] Appendix 1 Theoretical values of natural frequency and damping ratio of three-degree-of-freedom system:
[0187]
[0188] 2) Select a signal for a certain period of time, perform polynomial trend elimination and five-point multiple smoothing on the signal; the purpose of polynomial trend elimination is that due to zero drift caused by temperature changes, unstable low-frequency performance outside the sensor frequency range, and interference from the sensor's surrounding environment, the signal often deviates from the baseline, and even the deviation from the baseline changes with time. The whole process is called the trend term of the signal, which directly affects the correctness of the signal and should be removed;
[0189] The specific steps of eliminating polynomial trend terms are:
[0190] The measured vibration signal sampling data is {x k}(k=1,2,3,…,n), set a polynomial function as:
[0191]
[0192] Usually it depends on the situation, and m=1~6 can be used;
[0193] Determine the function Each undetermined coefficient makes the function With x k The sum of squared errors is the smallest, that is:
[0194]
[0195] The conditions for E to have an extreme value are:
[0196]
[0197] Taking the partial derivatives of E with respect to a in turn, we can generate a system of m+1 linear equations:
[0198]
[0199] Solve the system of equations to find the unknown coefficients;
[0200] The formula for eliminating the polynomial trend term is:
[0201]
[0202] The purpose of smoothing is that the collected vibration signals are often superimposed with noise. In addition to some periodic interference signals, there are also some irregular random interference signals. Since the interference signal has a wide frequency band, sometimes the high-frequency component accounts for a large proportion, making the collected signal present many burrs and very rough. In order to weaken the influence of the interference signal and improve the smoothness of the vibration curve, the signal needs to be smoothed.
[0203] The specific steps of five-point multiple smoothing are:
[0204]
[0205]
[0206]
[0207]
[0208] In the above formula, x i (i=1, ..., n) is the original vibration response acceleration signal sequence, n represents the number of sampling points, y i (i=1, ..., n) is the signal sequence after smoothing.
[0209] 3) The time domain signal is transformed into the time-frequency domain through short-time Fourier transform. Short-time Fourier transform in time-frequency transform is the simplest and most effective method to achieve sparseness, which can be written as follows:
[0210]
[0211] Wherein, w(τ-t) represents the window function moving with time t, x(τ) is the truncated signal, and X(t, ff) is the signal after time-frequency transformation. The response signals of the three measurement channels in the embodiment are selected, and a three-dimensional scatter plot in the time domain is drawn. It can be found that the scattered point signals are mixed together in the time domain, and the sparsity is very poor, such as Figure 3 As shown. Using the above-mentioned sparse transformation method based on short-time Fourier transform, the time domain signal is converted to the time-frequency domain, and a three-dimensional scatter plot of the time-frequency domain is drawn. It can be found that there is an obvious clustering straight line in the three-dimensional plot, which meets the premise assumption of good sparsity. Figure 4 shown.
[0212] 4) Estimate the mixing matrix based on the maximum energy method. First, define the formula of the frequency energy function as:
[0213]
[0214] Where: E(f)-frequency energy value; X(t, f)-signal after time-frequency transformation; t represents time, f represents frequency, and ∞ represents infinity. For modal response signals, the energy is concentrated at certain frequency points, and the clustering direction at the local energy maximum frequency point represents the clustering direction of the source signal. The mixing matrix can be estimated only by calculating the clustering direction at these frequency points. The maximum value of each clustering line is at the end of the line, and the other points have the same cosine distance from the end. The specific process is: first calculate the energy distribution of a single channel in the frequency domain in the time-frequency domain, and then add the energy of multiple channels at the same frequency point, that is:
[0215]
[0216] Where: Es (f)-The energy sum of the signals received by all sensors in the frequency domain; Real (X i (t, f))-the real part of the i-th sensor signal after time-frequency transformation; Img(X i (t, f))-is the imaginary part of the ith sensor signal after time-frequency transformation; t represents time, f represents frequency, and m represents the number of measurement channels. The peak detection method is used to extract the peak frequency points. If the commonly used zero derivative method is used in the peak detection method, the first-order derivative will accidentally cross zero due to the presence of noise in the measured signal, resulting in false detection. If some low-pass filters are used to smooth the curve, the effective information of the signal will be lost to a certain extent. Therefore, the method of finding peaks and valleys is adopted here. A peak is defined by having lower points around it, and a peak threshold is set. It is required that the peak value has a difference of at least the threshold level from the surrounding points to define it as a peak. Each point in the cluster line corresponds to the frequency band of the source in the time-frequency domain representation, such as Figure 5 shown.
[0217] 5) Source signal recovery based on L1 norm minimization. The source signal is recovered using the L1 norm minimization method. Since the number of unknown sources is greater than the number of equations, the equation has no unique solution. A constraint condition is introduced to require the L1 norm of the solution to be minimized, and the source recovery is transformed into an optimization problem. For all points in the sparse time-frequency domain:
[0218]
[0219] Where: s i (t, f) is each source signal; A is a mixing matrix, x(t, f) is a measurement channel signal, t represents time, f represents frequency, and st represents a constraint; this formula converts source recovery into an optimization problem. Since the number of unknown sources is greater than the number of equations, the equation has no unique solution, and a constraint is introduced to require the L1 norm of the solution to be minimum, where s(t, f) is the optimal estimate of the source signal. Theoretically, the equation is more intuitive with the L0 norm, which represents the number of non-zero elements in the vector, and naturally can find a sparse solution with the least non-zero elements in all feasible solutions. However, the constraint equation is non-convex and highly discrete, and it is difficult to solve by numerical calculation, but the sparse representation theory points out that if the solution is indeed sparse, the L1 norm, as the optimal convex approximation of the L0 norm, can obtain the same sparse solution as the L0 norm and is easier to optimize and solve. In this embodiment, a total of 5 source signals are separated by this method.
[0220] 6) 5 source signals are calculated probability density curves to carry out preliminary harmonic detection, as shown in Fig. 6 (a) and Fig. 6 (b), the abscissa is the amplitude, and the ordinate is the probability density function value, and the probability density function represents the probability that the signal amplitude falls on the specified area. This detection method is based on the estimation of the probability density curve of harmonic signal and Gaussian white noise signal, and its core idea is to estimate its probability density curve by the separated single-frequency time-domain signal, distinguish harmonic response and random response according to the difference of statistical characteristics, and the harmonic signal can present two obvious peak values, while the white noise signal has only one peak value. In the present embodiment, it can be seen that the probability density curves of source signal 1 and source signal 3 present two obvious peak values, so they are determined to be harmonic signals.
[0221] 7) Calculate the spectral kurtosis curve for source signal 1 and source signal 3 and locate the frequency corresponding to the harmonic. Spectral kurtosis can be defined as:
[0222]
[0223] Where: E{|X(f)| 4}-fourth-order cumulant; E[|X(f)| 2 ]-Second-order cumulant.
[0224] Divide the time series into M segments and perform Fourier transform X on each segment i Substituting (f) (i = 1, 2, ..., M) into the above formula, we can obtain the unbiased estimate of the fourth-order cumulant:
[0225]
[0226] Similarly, the unbiased estimator of the second-order cumulant is obtained:
[0227]
[0228] Finally, the unbiased estimator of spectral kurtosis is obtained:
[0229]
[0230] For a linear system, its response will not change the input signal characteristics. For a Gaussian white noise signal g(n) after Fourier transform is G(f), the spectral kurtosis value SK is calculated by the above formula g (f)=0.
[0231] Assume the frequency is f0 and the phase is The harmonic signal is:
[0232]
[0233] After Fourier transform:
[0234]
[0235] Among them, δ(f0) is the Dirac function, and the spectral kurtosis value SK at f0 is calculated by the formula h (f0) = -1.
[0236] The linear property of Fourier transform, for the mixed signal of Gaussian white noise signal and harmonic signal, its Fourier transform result is:
[0237] X(f)=G(f)+H(f)
[0238] In summary, the kurtosis value of the mixed signal spectrum is:
[0239]
[0240] Therefore, the spectral kurtosis value at the corresponding frequency is -1, which can be used to determine that it is a harmonic component. Figure 7 and Figure 8 As shown, the frequency spectra of source signal 1 and source signal 3 and the corresponding spectral kurtosis curves are plotted. It can be found that the spectral kurtosis value of source signal 1 at 50 Hz is close to -1, and the spectral kurtosis value of source signal 3 at 150 Hz is close to -1, so it can be determined that the harmonic frequencies are 50 Hz and 150 Hz respectively.
[0241] 8) Harmonic mode removal and signal reconstruction. The detected harmonic signal is removed. If the probability density indicates that the entire source signal is a harmonic signal, the entire source signal is directly removed. For example, the i-th source signal is a harmonic signal S i (t, f), the S in the source signal i (t, f) is removed to obtain S R (t, f), and the corresponding a in the mixing matrix i Remove to get A R , and obtain the reconstructed signal x R (t, f) = A R S R (t, f), and then convert the signal to the time domain to get x R (t). If the harmonic energy accounts for too small a proportion of the entire source signal, the probability density cannot clearly indicate that it is a harmonic, but the spectral kurtosis method indicates the harmonic frequency in the source signal, then there is no need to remove the entire source signal. Instead, the stopband filter can be used to remove the corresponding harmonic frequency components before participating in the subsequent signal reconstruction.
[0242] In this embodiment, the probability density can clearly indicate that source signal 1 and source signal 3 are harmonics, so these two source signals are directly removed, and the corresponding mixing coefficients in the mixing matrix are removed, and then the signal is reconstructed based on the reconstruction formula.
[0243] 9) Compare the initial signal with the reconstructed signal after deharmonicization. Plot the frequency energy diagrams of the initial signal and the reconstructed signal after deharmonicization, and observe the peaks corresponding to the harmonic frequencies. It can be found that in the initial signal, the peaks corresponding to 50Hz and 150Hz are very obvious, which are harmonic modes, while the peaks at 50Hz and 150Hz in the reconstructed signal after deharmonicization are not obvious. The remaining three frequency peaks are all peaks of the system mode, and finally the detection and removal of the harmonic mode is achieved.
[0244] Although the embodiments of the present invention are described above in conjunction with the accompanying drawings, the present invention is not limited to the above specific embodiments and application fields, and the above specific embodiments are only illustrative and instructive, rather than restrictive. A person of ordinary skill in the art can also make many forms under the guidance of this specification and without departing from the scope of protection of the claims of the present invention, all of which belong to the protection of the present invention.
Claims
1. A harmonic mode detection and removal method based on blind source separation, It is characterized in that It includes the following steps, Step 1, collecting vibration response acceleration signals of various measuring points under the running state of the rotating machinery; Step 2, performing polynomial trend term elimination and smoothing processing on the vibration response acceleration signal to obtain a signal sequence; Step 3, obtaining the time-frequency domain by short-time Fourier transform of the signal sequence; Step 4, clustering the time-frequency scattered points in the time-frequency domain based on the maximum energy method to obtain an estimated value of the mixing matrix; Step 5, based on the L1 norm minimization method, the source recovery problem is transformed into an optimization problem to achieve source signal recovery; Step 6, performing probability density curve calculation on the source signal to find the source signal with harmonic characteristics; Step 7, performing spectral kurtosis detection on the source signal with harmonic characteristics to obtain the harmonic modal frequency; Step 8, directly remove the source signal with obvious harmonic characteristics in the probability density curve, and filter and remove the source signal with unclear harmonic characteristics based on the harmonic modal frequency, wherein when the probability density curve has two peaks, the harmonic characteristics are obvious, and if only one peak can be seen, the harmonic characteristics are not obvious; Step 9, combining other source signals containing system characteristics and source signals after removing harmonic frequencies, and combining with a mixing matrix to achieve signal reconstruction, the reconstructed signal is a signal after removing harmonics; Step 10, comparing the vibration response acceleration signal with the signal after harmonic removal, the difference before and after harmonic removal can be clearly and intuitively seen, and the signal after harmonic removal is the output result.
2. According to the method for harmonic mode detection and removal based on blind source separation according to claim 1, It is characterized in that Preferably, in step 1, the number of measurement points arranged in the rotating machinery is less than the sum of the harmonic modal number and the system modal number, wherein the harmonic modal number is obtained by harmonic detection, and the system modal number is obtained by modal analysis.
3. According to the method for harmonic mode detection and removal based on blind source separation of claim 1, It is characterized in that In step 2, in the elimination of polynomial trend terms, the measured vibration response acceleration signal sampling data is {x k }(k=1,2,3,…,n), n represents the number of sampling points, and the polynomial function is set to: In the formula, a j (j=0, 1, ..., m) are the coefficients of the polynomial to be determined, m is the highest order of the polynomial, and m=1 to 6 is used to determine the function The unknown coefficients of With x k The sum of squared errors is the smallest, that is: E represents the sum of squares of the errors of the two functions. To minimize E, the extreme value condition is: Taking the partial derivative of E with respect to a in turn, a system of linear equations with m+1 variables is generated: Solve the system of equations and find the unknown coefficients; The formula for eliminating the polynomial trend term is: Five-point multiple smoothing includes: In the above formula, x i (i=1, ..., n) is the measured vibration response acceleration signal sequence, n represents the number of sampling points, y i (i=1, ..., n) is the signal sequence after smoothing.
4. According to the method for harmonic mode detection and removal based on blind source separation of claim 1, It is characterized in that In the short-time Fourier transform, the formula for the frequency energy function is first defined as: Where: E(f) is the frequency energy value; x(t, f) is the signal after time-frequency transformation; t represents time, f represents frequency, and ∞ represents infinity. In the time-frequency domain, the energy distribution of a single channel in the frequency domain is first calculated, and then the energy of multiple channels is added at the same frequency point, that is: Where: E z (f) is the energy sum of the signals received by all sensors in the frequency domain; Real (X i (t, f)) is the real part of the i-th sensor signal after time-frequency transformation; Img(X i (t, f)) is the imaginary part of the i-th sensor signal after time-frequency transformation; m is the number of measurement channels, t represents time, f represents frequency, and the peak frequency points are extracted using the peak detection method.
5. The harmonic mode detection and removal method based on blind source separation according to claim 1, It is characterized in that In step 5, based on the L1 norm minimization method, for all points in the sparse time-frequency domain: Where: s i (t, f) are the source signals; A is the mixing matrix, x(t, f) is the measurement channel signal; st is the constraint condition; t represents time, f represents frequency; the L1 norm of the solution is required to be the smallest, where s(t, f) is the optimal estimate of the source signal.
6. The harmonic mode detection and removal method based on blind source separation according to claim 1, It is characterized in that In step 6, it is assumed that the mixed signal x(n) is composed of a Gaussian white noise signal and a harmonic wave signal, that is: x(n)=g(n)+h(n), Where n = 1, 2, ..., N, N is the number of sampling points; g(n) is the Gaussian signal sequence and h(n) is the harmonic signal sequence. The sparse component analysis method is used to separate the g(n) and h(n) signals. The ideal Gaussian white noise signal conforms to the normal distribution and its probability density is: The probability densities of the two random variables X and T are f x (x) and f T (t), and x=h(t), according to the joint probability density function theory, we can get: f X (x)=f t [h(x)]|h′(x)|,α<x<β, Where h′(x) is the first-order derivative of h(x), α and β are the range of values of x, and for harmonic signals, it is considered as: x=h(t)=asin(2πft), The time t belongs to uniform distribution, and its probability density function is a constant value f t , then the probability density function of the harmonic response is expressed as: For any frequency f, when the harmonic has a constant amplitude a, the above formula is always valid. When the mean value of the harmonic component is 0, that is, x→a or x→-a, the probability density function tends to infinity, and there are two obvious peaks in the probability density curve.
7. The harmonic mode detection and removal method based on blind source separation according to claim 1, It is characterized in that In step 7, the spectral kurtosis can be defined as: Where: E{|X(f)| 4 } is the fourth-order cumulant; E[|X(f)| 2 ] is the second-order cumulant, For a mixed signal of Gaussian white noise signal and harmonic signal, its Fourier transform result is: X(f)=G(f)+H(f), The kurtosis value of the mixed signal spectrum is: In the formula, f 0 is the harmonic frequency value, and the spectral kurtosis value at the corresponding frequency is -1, so it is determined to be a harmonic component.
8. The harmonic mode detection and removal method based on blind source separation according to claim 1, It is characterized in that The rotating machine includes a gas turbine or a wind turbine.
9. The harmonic mode detection and removal method based on blind source separation according to claim 1, It is characterized in that Multiple accelerometer sensors are used to collect the vibration response acceleration signals of each measuring point under the operating state of the rotating machinery.
Citation Information
Patent Citations
Method for identifying operation mode under underdetermined condition based on blind source separation technology
CN111241904A
Underdetermined blind source separation method, system and device based on minimization and maximization
CN112201274A
Gear vibration signal source underdetermined blind source separation method based on density and compressed sensing
CN112560699A
Rolling bearing composite fault diagnosis method based on enhanced harmonic vector analysis
CN116628422A
