A rolling bearing compound fault diagnosis method based on enhanced harmonic vector analysis

Abstract

The application provides a rolling bearing compound fault diagnosis method based on enhanced harmonic vector analysis, vibration signals of rolling bearings in different fault and normal states are acquired in advance as source signals, and are convolved and mixed; the mixed signals are denoised by a wavelet threshold value method, and the denoised mixed signals are subjected to harmonic structure enhancement by using a cepstrum threshold value method; a two-dimensional time-frequency mask function similar to Wiener is constructed, the mixed signals subjected to the harmonic structure enhancement are subjected to blind source separation in a time-frequency domain, and the frequency domain scale of the separated signals is aligned by using reverse projection; for each separated signal, a fast spectral kurtosis method based on a 1 / 3 binary tree structure is used to calculate the kurtosis diagram of the signal, the frequency band with the maximum kurtosis value is found, and band-pass filtering is performed on the frequency band; the filtered signals are subjected to Hilbert envelope spectrum analysis, fault characteristics are obtained, and compound fault diagnosis is completed. The application improves the diagnosis precision of rolling bearing compound faults under strong noise.

Description

Technical Field

[0001] This invention belongs to the field of rolling bearing fault diagnosis, and specifically relates to a method for diagnosing composite faults in rolling bearings based on enhanced harmonic vector analysis. Background Technology

[0002] Rolling bearings are an indispensable component of modern industry, playing a vital role in mechanical equipment such as wind turbines, petrochemicals, metallurgy, transportation, and aerospace. Because rolling bearings frequently operate in harsh environments with heavy loads, high temperatures, and high pressures, failure is inevitable. Furthermore, bearing failures are often not singular but rather occur in compound forms, resulting in complex fault signals. The mixing of different fault signals, coupled with the influence of noise, makes diagnosing complex faults a challenging task.

[0003] Commonly used signal decomposition methods include Empirical Mode Decomposition (EMD), Wavelet Transform (WT), and Variational Mode Decomposition (VMD). However, these methods are less effective at processing complex fault signals with similar frequencies and are easily affected by environmental noise. This can lead to the inability to accurately separate independent fault signals, thus impacting fault diagnosis results.

[0004] Blind source separation, which estimates source signals with almost no prior knowledge of the signal source and transmission channel, is widely used in signal separation and extraction. Hierarchical Virtualization (HVA) was proposed to solve the blind source separation problem for convolutional mixed audio signals. HVA defines a time-frequency masking of audio signals with cepstral sparsity to enhance harmonic structure and implicitly model the source signal, enabling the algorithm to achieve state-of-the-art performance in both speech and music signal processing. However, direct application of HVA to rolling bearing fault diagnosis often yields unsatisfactory results.

[0005] Wavelet thresholding is an effective method for denoising impact signals and has been widely applied to the denoising of rolling bearing vibration signals in recent years. The concept of spectral kurtosis was initially proposed by Dwyer, and Antoni systematized and standardized it, proposing a fast kurtosis plot calculation method based on filter banks, namely the fast spectral kurtosis method based on 1 / 3 binary trees, which improves computational speed. In the field of rolling bearing fault diagnosis, using spectral kurtosis to select the optimal demodulation band for filtering, followed by envelope demodulation, remains one of the most effective fault feature extraction methods, effectively removing the influence of background noise as much as possible. Summary of the Invention

[0006] Purpose of the invention: The purpose of this invention is to address the shortcomings of existing technologies by providing a method for diagnosing complex faults in rolling bearings based on enhanced harmonic vector analysis. This invention introduces and improves upon the traditional harmonic vector analysis method, making it applicable to the processing of complex bearing fault signals. Furthermore, it employs wavelet threshold denoising to preprocess the complex fault signals, and utilizes spectral kurtosis and bandpass filtering for post-processing of the separated signals to enhance fault features and achieve accurate diagnosis of complex faults in rolling bearings under strong noise conditions.

[0007] Technical Solution: This invention provides a method for diagnosing composite faults in rolling bearings based on enhanced harmonic vector analysis, specifically including the following steps:

[0008] (1) The vibration signals of the rolling bearing under different fault and normal conditions are obtained in advance as source signals and then convolved and mixed to obtain a mixed signal;

[0009] (2) Perform wavelet threshold denoising on the mixed signal to reduce the impact of noise on the fault source signal;

[0010] (3) Harmonic structure enhancement of the noise-reduced mixed signal is performed using the cepstral thresholding method;

[0011] (4) Construct a Wiener-like two-dimensional time-frequency masking function, perform blind source separation on the obtained harmonic structure-enhanced hybrid signal in the time-frequency domain, and use back projection to align the frequency domain scale of the obtained separated signal.

[0012] (5) For each separated signal, calculate their kurtosis map using the fast spectral kurtosis method based on the 1 / 3 binary tree structure, find the frequency band with the largest kurtosis value, and perform bandpass filtering on the frequency band;

[0013] (6) Perform Hilbert envelope spectrum analysis on the filtered signal to obtain fault characteristics and complete the diagnosis of composite faults.

[0014] Furthermore, the implementation process of step (1) is as follows:

[0015] Vibration signals of bearings in inner race fault, outer race fault, and normal state are collected at a fixed sampling frequency. An FIR filter is constructed to convolve and mix the vibration signals to obtain a mixed signal with the same number of source signals.

[0016] Furthermore, the wavelet threshold denoising process for the mixed signal described in step (2) is as follows:

[0017] The original signal is transformed to the wavelet domain, and thresholding is performed in the wavelet domain to suppress wavelet coefficients containing random noise. The processed wavelet coefficients are then used to reconstruct the signal to obtain the signal after suppressing random noise.

[0018] If the function φ(t) satisfies the following condition:

[0019]

[0020] Then φ(t) is the generating function of the wavelet or the basic wavelet. It is the Fourier transform of φ(t);

[0021] The wavelet function obtained by stretching and translating the basic wavelet function is expressed as:

[0022]

[0023] Where a and b represent the stretching and translation parameters of the wavelet transform, respectively;

[0024] The wavelet transform pairs of the signal are:

[0025]

[0026]

[0027] Where * denotes conjugation, W f (a,b) represents the wavelet transform of signal f(t), and equation (4) is the inverse transform, which represents signal reconstruction.

[0028] Furthermore, the blind source separation process described in step (4) is as follows:

[0029] For a signal time-frequency domain convolution mixing process of the form X[t,f]=C[f]S[t,f], its demixing process is W[f]X[t,f]=Y[t,f], where X(t)=[x1,x2,…,x M ] T For a mixed signal, C is an M×N mixing matrix, and S(t) = [s1, s2, ..., s]. N ] T Let Y(t) be the source signal, W be the M×N separation matrix to be solved, and Y(t) = [y1, y2, ..., y]. N ] T For separate signals;

[0030] HVA solves the positive definite blind source separation problem for M≥N by estimating an unmixing matrix W to recover the source signals; handling the DBSS problem requires assuming that the source signals are statistically independent, and the measure of independence reduces to a minimization problem of the following form:

[0031]

[0032] The minimization problem is simplified to the following form using the primal-dual splitting algorithm:

[0033]

[0034] Where w is the quantity to be optimized, I and J are the first and second terms of the objective function, and L is a bounded linear operator; since I and J are both non-differentiable functions, the following proximity operator is used to solve the problem:

[0035]

[0036] Let δ represent the variable value that minimizes [·]; β is a temporary variable in the iteration process, and μ is the step size;

[0037] The proximity operator of the mixed norm of the 1-norm and 2-norm used in the HVA algorithm is expressed using the group threshold operator as follows:

[0038]

[0039] Where λ>0 is the threshold parameter, (·) + =max{0,·} represents a half-wave rectifier that replaces negative values ​​with zeros, (·) n [t,f] represents the (n,t,f)th element of the N×T×F array; the threshold shrinkage operator can be expressed as:

[0040] (H λ [β]) n [t,f]=(M(β)) n [t,f]β n [t,f] (9)

[0041] Where, 0≤(M(β)) n [t,f]≤1 is a nonnegative scalar dependent on the input β; the group threshold operator in equation (8) is time-frequency masked using the mask shown in equation (10):

[0042]

[0043] The PDS algorithm incorporates general time-frequency masking and defines an implicit source model, resulting in the HVA algorithm.

[0044] Furthermore, the process implemented in step (5) is as follows:

[0045] The spectral kurtosis is calculated as follows:

[0046]

[0047] Where S(t,f) is the complex envelope of signal x(t) at frequency f, <|S(t,f) n |> represents the nth order spectral moment of the signal, <·> represents the average value, and |·| represents the modulus operator;

[0048] Construct a cutoff frequency fc A low-pass filter h(t) with the expression 1 / 8 + δ, where δ > 0, was constructed; quasi-analytic low-pass filter h0(t) and quasi-analytic high-pass filter h1(t) were also constructed:

[0049]

[0050] The filtering results are downsampled layer by layer using low-pass and high-pass filters respectively, and their lengths are shortened by the same factor to ensure that the length of each layer of data in the filter is the same as that of the original data.

[0051] The kurtosis of each of the above filtering results is calculated according to equation (11) to obtain the spectral kurtosis:

[0052]

[0053] in, This represents the filtering result of the i-th filter at the k-th level of the decomposition tree; by combining all the spectral kurtosis, a fast spectral kurtosis map is obtained.

[0054] Beneficial effects: Compared with the prior art, the beneficial effects of the present invention are as follows: This method is aimed at the diagnosis of composite faults in rolling bearings; it solves the problem that the traditional blind source separation method has poor ability to separate convolutional mixed signals in strong noise environment, resulting in low accuracy of bearing composite fault diagnosis. By combining wavelet threshold denoising, harmonic vector analysis and fast spectral kurtosis, the diagnostic accuracy of rolling bearing composite faults under strong noise is improved. Attached Figure Description

[0055] Figure 1 This is a flowchart of the method of the present invention;

[0056] Figure 2 The following is a physical diagram of the rolling bearing failure vibration test equipment in an embodiment of the present invention; wherein, (a) is a rolling bearing failure test bench; and (b) is the test bearing;

[0057] Figure 3 The following are time-domain plots and envelope spectra of two fault source signals and one normal source signal collected for the experiment of this invention; wherein, (a) is the time-domain plot and envelope spectrum of the normal bearing source signal; (b) is the time-domain plot and envelope spectrum of the inner ring fault bearing source signal; and (c) is the time-domain plot and envelope spectrum of the outer ring fault bearing source signal.

[0058] Figure 4 The images show the time-domain plot and envelope spectrum of the mixed signal obtained after convolutional mixing of the source signals according to the present invention; wherein, (a) is the time-domain plot and envelope spectrum of mixed signal 1; (b) is the time-domain plot and envelope spectrum of mixed signal 2; and (c) is the time-domain plot and envelope spectrum of mixed signal 3.

[0059] Figure 5The envelope spectrum of the separated signal obtained by HVA separation according to the present invention is shown below; wherein, (a) is the envelope spectrum of separated signal 1; (b) is the envelope spectrum of separated signal 2; and (c) is the envelope spectrum of separated signal 3.

[0060] Figure 6 Here are the kurtosis diagrams of the separated signals obtained by HVA separation according to the present invention; wherein, (a) is the kurtosis diagram of separated signal 1; (b) is the kurtosis diagram of separated signal 2; and (c) is the kurtosis diagram of separated signal 3.

[0061] Figure 7 The envelope spectrum of the enhanced separated signal obtained by bandpass filtering the resonant frequency band of the separated signal according to the present invention is shown in (a), which is the envelope spectrum of the enhanced separated signal 1; (b) is the envelope spectrum of the enhanced separated signal 2; and (c) is the envelope spectrum of the enhanced separated signal 3. Detailed Implementation

[0062] The present invention will now be described in further detail with reference to the accompanying drawings.

[0063] This invention proposes a method for diagnosing complex faults in rolling bearings based on Enhanced Harmonic Vector Analysis (EHVA), addressing the problem of low accuracy in diagnosing complex faults in rolling bearings under strong noise conditions. First, wavelet thresholding (WT) is used to denoise the acquired vibration signal to reduce the impact of noise. Second, Harmonic Vector Analysis (HVA) is used to remove the convolution effect of the signal transmission path, performing blind separation of the fault signal. HVA employs cepstral thresholding to enhance the harmonic structure of the signal and constructs a Wiener-like two-dimensional time-frequency mask, making the separated signals more independent in each iteration. Then, back projection technology is used to align the frequency scales of the separated signals, obtaining individual fault signals from the complex fault diagnosis signal. Finally, to make the fault characteristics more prominent, the spectral kurtosis (SK) of the separated signals is calculated to find the resonant frequency band, which is then bandpass filtered to enhance the fault characteristics. Figure 1 As shown, the specific implementation process is as follows:

[0064] S1: Vibration signals of different fault types of rolling bearings are collected at a fixed sampling frequency Fs as source signals. An FIR filter is constructed to convolve and mix the source signals to obtain a mixed signal with the same number of source signals.

[0065] This invention uses faulty bearing test data as the source signal. The test bench is as follows: Figure 2As shown in (a), the test bench mainly consists of a base plate, drive motor, bearing housing, circular tube shell, shell support, and control cabinet. It has three bearings installed and uses a piezoelectric accelerometer. The motor's rated speed is 3000 rpm. The experimental bearings are as follows... Figure 2 As shown in (b), the bearing failure is caused by crack damage from electrical discharge machining. The crack width is 1.2 mm. There are two types of failures: inner ring failure and outer ring failure. The bearing parameters are shown in Table 1.

[0066] Table 1 Parameters of the test rolling bearings

[0067]

[0068] The selected motor speed was 2012 r / min. The calculated characteristic frequencies for inner race faults were 181 Hz and outer race faults were 121 Hz. The sampling frequency was 25.6 kHz. The time-domain waveforms and envelope spectra of the various source signals acquired in the experiment are shown below. Figure 3 As shown, where, Figure 3 (a) shows the time-domain plot and envelope spectrum of a normal bearing source signal; Figure 3 (b) shows the time-domain plot and envelope spectrum of the bearing source signal with inner race fault; Figure 3 (c) shows the time-domain plot and envelope spectrum of the bearing source signal with outer ring fault.

[0069] Source signals from bearings in inner race fault, outer race fault, and normal state, each lasting 5 seconds, were convolved and mixed. Gaussian white noise with an SNR of 0dB was added to generate three mixed signals to simulate the acquisition signals from three accelerometers. The time-domain waveforms and envelope spectra of each mixed signal are shown below. Figure 4 As shown, where, Figure 4 (a) shows the time-domain plot and envelope spectrum of the mixed signal 1; Figure 4 (b) shows the time-domain plot and envelope spectrum of the mixed signal 2; Figure 4 In the middle (c), the time domain plot and envelope spectrum of the mixed signal 3 are shown.

[0070] S2: Perform wavelet decomposition on the mixed signal, then threshold processing, and then wavelet reconstruction to complete the wavelet threshold noise reduction of the mixed signal, thereby reducing the impact of noise on the fault source signal.

[0071] The specific steps for wavelet thresholding noise reduction are as follows:

[0072] If the function φ(t) satisfies the following condition:

[0073]

[0074] Then φ(t) is the generating function of the wavelet or the basic wavelet. It is the Fourier transform of φ(t).

[0075] The wavelet function obtained by stretching and translating the basic wavelet function is expressed as:

[0076]

[0077] Where a and b represent the stretching and translation parameters of the wavelet transform, respectively.

[0078] The wavelet transform pairs of the signal are:

[0079]

[0080]

[0081] Where * denotes conjugation, W f (a,b) represents the wavelet transform of signal f(t), and equation (4) is the inverse transform, representing signal reconstruction. From the perspective of signal processing, wavelet threshold denoising preserves the useful high-frequency features of the signal, making it superior to ordinary low-pass filters.

[0082] S3: Harmonic structure enhancement of the noise-reduced mixed signal is performed using the cepstral thresholding method.

[0083] S4: A Wiener-like two-dimensional time-frequency masking function is constructed and introduced into the PDS algorithm to solve the objective function for blind source separation. This performs blind source separation on the harmonic-enhanced mixed signal in the time-frequency domain and uses back projection to align the frequency domain scale of the obtained separated signals. The time-domain waveforms and envelope spectra of each separated signal are shown below. Figure 5 As shown, where, Figure 5 In the middle (a), the envelope spectrum of the separated signal 1 is shown. Figure 5 (b) shows the envelope spectrum of the separated signal 2; Figure 5 In the middle (c), the envelope spectrum of the separated signal 3 is shown.

[0084] The specific steps for blind source separation are as follows:

[0085] For a signal time-frequency domain convolution mixing process of the form X[t,f]=C[f]S[t,f], its demixing process is W[f]X[t,f]=Y[t,f], where X(t)=[x1,x2,…,x M ] T For a mixed signal, C is an M×N mixing matrix, and S(t) = [s1, s2, ..., s]. N ] T Let Y(t) be the source signal, W be the M×N separation matrix to be solved, and Y(t) = [y1, y2, ..., y]. N ] T This is a separated signal.

[0086] HVA solves the positive definite blind source separation problem for M ≥ N by recovering the source signals through estimation of an unmixing matrix W. Addressing the DBSS problem requires the assumption that the source signals are statistically independent. The measure of independence can be reduced to a minimization problem of the following form:

[0087]

[0088] The minimization problem of equation (5) can be solved using the primal-dual splitting (PDS) algorithm, which simplifies the minimization problem of equation (5) to the following form:

[0089]

[0090] Where w is the quantity to be optimized, I and J are the first and second terms of the objective function, and L is a bounded linear operator. Since I and J are both non-differentiable functions, gradient-based optimization methods are not applicable; instead, a proximity operator of the following form can be used to solve the problem:

[0091]

[0092] Let δ represent the variable value that minimizes [·]. β is a temporary variable in the iteration process, and μ is the step size.

[0093] Since the proximity operator of the sparse induced penalty function shown in equation (7) can also be represented by a threshold (or shrinkage) operator, the proximity operator of the mixed norm of the 1-norm and 2-norm used in the HVA algorithm is expressed by the group threshold operator as follows:

[0094]

[0095] Where λ>0 is the threshold parameter, (·) + =max{0,·} represents a half-wave rectifier that replaces negative values ​​with zeros, (·) n [t,f] represents the (n,t,f)th element of the N×T×F array. The threshold shrinkage operator can be expressed as:

[0096] (H λ [β]) n [t,f]=(M(β)) n [t,f]β n [t,f] (9)

[0097] Where, 0≤(M(β)) n [t,f]≤1 is a nonnegative scalar dependent on the input β. This process can be interpreted as time-frequency masking using a data-dependent masking M(β). The group threshold operator in equation (8) uses the mask shown in equation (10) for time-frequency masking:

[0098]

[0099] Based on this result, a general time-frequency masking was introduced into the PDS algorithm, and an implicit source model was defined, resulting in the HVA algorithm.

[0100] S5: For each separated signal obtained in S4, calculate their kurtosis map using the fast spectral kurtosis method based on a 1 / 3 binary tree structure, find the frequency band with the largest kurtosis value, and perform bandpass filtering on this frequency band.

[0101] The steps of the fast spectral kurtosis method are as follows:

[0102]

[0103] Where S(t,f) is the complex envelope of signal x(t) at frequency f, <|S(t,f) n |> represents the nth-order spectral moment of the signal, <·> represents the average value, and |·| represents the modulus operator.

[0104] First, construct a cutoff frequency f. c A low-pass filter h(t) = 1 / 8 + δ, where δ > 0, is given. Based on this, quasi-analytic low-pass filters h0(t) and h1(t) are constructed respectively.

[0105]

[0106] Then, the filtering results are downsampled layer by layer using low-pass and high-pass filters respectively, and their lengths are shortened by the same factor to ensure that the length of each layer of data in the filter is the same as that of the original data.

[0107] Finally, the kurtosis of each of the above filtering results is calculated according to equation (11) to obtain the spectral kurtosis:

[0108]

[0109] in, This represents the filtering result of the i-th filter at the k-th level of the decomposition tree. Combining all the spectral kurtosis yields a fast spectral kurtosis map.

[0110] Kurtosis plots of each separated signal are as follows Figure 6 As shown, where, Figure 6 (a) is the kurtosis plot of the separated signal 1; Figure 6 (b) is the kurtosis plot of the separated signal 2; Figure 6 (c) shows the kurtosis plot of the separated signal 3. From the kurtosis plot, we can find the layer number k where the frequency band with the largest kurtosis value, i.e., the resonant frequency band, is located, and the center frequency F... c and bandwidth B wThe resonant frequency band contains the richest fault information. Filtering the resonant frequency band can significantly improve the signal-to-noise ratio of the separated signals.

[0111] S6: Perform Hilbert envelope spectrum analysis on the filtered separated signals, and obtain the envelope spectra of each separated signal as follows: Figure 7 As shown, where Figure 7 (a) shows the envelope spectrum of the enhanced separated signal 1; Figure 7 (b) shows the envelope spectrum of the enhanced separated signal 2; Figure 7 (c) shows the envelope spectrum of the enhanced separated signal 3. Comparing the envelope spectrum analysis of the source signal with the fault diagnosis results, it can be seen that... Figure 7 In the envelope spectra of the two separated signals (b) and (c), there are obvious outer and inner ring fault characteristic frequencies, respectively. In addition, the distribution of octaves and sidebands is also highly similar to the source signal, which confirms the fault diagnosis result: the bearing composite fault includes inner ring fault and outer ring fault.

[0112] Experimental results show that the fault features in the envelope spectrum of the separated signal are clear and have a high similarity to the source signal, resulting in accurate fault diagnosis. Furthermore, the signal-to-noise ratio of the separated signal is high, effectively reducing the influence of noise, demonstrating that this method has a good effect on the diagnosis of composite faults in rolling bearings.

[0113] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style of the specification is merely for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A rolling bearing complex fault diagnosis method based on enhanced harmonic vector analysis, characterized in that, It comprises the following steps: (1) pre-acquire vibration signals of different faults and normal states of the rolling bearing as source signals, and perform convolution mixing to obtain mixed signals; (2) wavelet threshold denoising is performed on the mixed signals to reduce the influence of noise on the fault source signals; (3) the cepstrum threshold method is used to enhance the harmonic structure of the denoised mixed signals; (4) a two-dimensional time-frequency masking function similar to Wiener is constructed, and the mixed signals with the enhanced harmonic structure are subjected to blind source separation in the time-frequency domain, and the frequency domain scale of the separated signals is aligned by using back projection; (5) the fast spectral kurtosis method based on the 1 / 3 binary tree structure is used to calculate the kurtosis diagram of each separated signal, find the frequency band with the maximum kurtosis value, and perform band-pass filtering on the frequency band; (6) Hilbert envelope spectrum analysis is performed on the filtered signals to obtain fault features and complete compound fault diagnosis; The wavelet threshold denoising of the mixed signals in step (2) is implemented as follows: The original signal is transformed into the wavelet domain, the threshold is processed in the wavelet domain, the wavelet coefficients containing random noise are suppressed, the signal is reconstructed by using the processed wavelet coefficients, and the signal after suppressing random noise is obtained; If the function φ(t) satisfies the condition: φ(t) is the mother wavelet or generating function of the wavelet, is the Fourier transform of φ(t); The basic wavelet function is stretched and translated to obtain the wavelet function, which is represented as: Wherein, a and b represent the stretching and translation parameters of the wavelet transform respectively; The wavelet transform pair of the signal is: where * denotes conjugation, W f (a,b) denotes the wavelet transform of the signal f(t), and equation (4) is the inverse transform, which represents the signal reconstruction. The process implemented in step (5) is as follows: The calculation method of spectral kurtosis is: where S(t,f) is the complex envelope of the signal x(t) at frequency f, <|S(t,f) n is the n-th spectral moment of the signal, <·> is the average value and |·| is the modulus operator. A low-pass filter h(t) with a cut-off frequency f c = 1 / 8 + δ, where δ > 0; a quasi-analytical low-pass filter h0(t) and a quasi-analytical high-pass filter h1(t) are constructed, respectively: The low-pass filter and the high-pass filter are used to down-sample the filtering results layer by layer, and the length is shortened by the same multiple, so that the length of each layer of the filter data is the same as that of the original data; The kurtosis of the above filtering results is calculated according to formula (11), and the spectral kurtosis is obtained: wherein, represents the filtering result of the kth layer of the decomposition tree, the ith filter; all the spectral kurtosis are combined to obtain a fast spectral kurtogram.

2. The method according to claim 1, characterized in that, The implementation process of step (1) is as follows: The vibration signals of the inner ring fault, the outer ring fault and the normal state bearing are collected at a fixed sampling frequency, an FIR filter is constructed to perform convolution mixing on the vibration signals, and mixed signals with the same number of source signals are obtained.

3. The method according to claim 1, characterized in that, The implementation process of blind source separation in step (4) is as follows: For a signal time-frequency domain convolution mixing process of the form X[t,f]=C[f]S[t,f], its demixing process is W[f]X[t,f]=Y[t,f], where X(t)=[x1,x2,…,x M ] T For a mixed signal, C is an M×N mixing matrix, and S(t) = [s1, s2, ..., s]. N ] T Let Y(t) be the source signal, W be the M×N separation matrix to be solved, and Y(t) = [y1, y2, ..., y]. N ] T For separate signals; HVA solves the positive definite blind source separation problem of M>N, and restores the source signals by estimating a demixing matrix W; the DBSS problem needs to assume that the source signals are statistically independent, and the measure of independence is reduced to the following minimization problem: The original dual splitting algorithm is used to solve the problem, and the minimization problem is simplified as follows: Wherein, w is the quantity to be optimized, I and J are the first and second terms of the objective function, and L is a bounded linear operator; since I and J are not differentiable functions, the following approximation operator is used to solve them: denotes the variable delta value that minimizes [·]; β is a temporary variable of the iteration process, and μ is the step size; The approximation operator of the mixed norm of 1 norm and 2 norm used in the HVA algorithm is expressed as a group threshold operator: where λ > 0 is a threshold parameter, (·) + = max{0, ·} is a half-wave rectification that replaces negative values by zero, (·) n [t, f] denotes the (n, t, f)-th element of the N x T x F array; the threshold shrinkage operator is denoted by: (H λ [β]) n [t,f] = (M(β)) n [t,f]β n [t,f] (9) where 0 ≤ (M(β)) n [t,f] ≤ 1 is a non-negative scalar depending on the input β; the group threshold operator in equation (8) is time-frequency masked with a mask shown in equation (10): In the PDS algorithm, a general time-frequency mask is introduced, and an implicit source model is defined to obtain the HVA algorithm.

Images