A mechanical fault diagnosis method and system based on blind deconvolution

Abstract

The present invention discloses a mechanical fault diagnosis method and system based on blind deconvolution. The method comprises: picking up a mechanical vibration observation signal through an acceleration sensor; constructing a weighted energy threshold coefficient, and performing single source point detection on the mechanical vibration observation signal collected by the sensor based on the weighted energy threshold coefficient to obtain feature data and an estimated number of sources H; normalizing the filtered feature data to obtain feature data represented by directional angles; clustering the feature data represented by directional angles to obtain cluster centers and membership degrees; reconstructing and separating H source signals containing single faults from a signal containing composite faults based on the cluster centers and membership degrees, performing an inverse short-time Fourier transform on the separated source signals to obtain time-domain source signals; performing feature enhancement on the time-domain source signals; and performing envelope analysis on the enhanced signals to determine the fault type. The present invention can achieve blind separation of composite faults when the number of sources is unknown, and use mechanical vibration signals for fault feature extraction and diagnosis.

Description

Technical Field

[0001] The present invention relates to a mechanical fault diagnosis method and system based on blind deconvolution, belonging to the technical field of mechanical equipment state monitoring and fault diagnosis. Background Art

[0002] The operating condition of rolling bearings significantly impacts the operation of the entire equipment. Furthermore, due to the complex operating environment, a single fault in a local area can quickly develop into multiple faults. Therefore, separating and extracting multiple fault signatures from multiple fault signals is highly valuable for bearing fault diagnosis. Blind source separation (BBS) technology uses observed signals to estimate the source signal when the source signal and signal channel are unknown.

[0003] The ICA algorithm assumes statistical independence of the source signals, meaning that each independent component must conform to a non-Gaussian distribution. However, modern machinery struggles to meet this assumption. In contrast, the sparsity assumption of SCA is relatively easy to meet. Overall, SCA is a more suitable approach for addressing BSS. However, due to the complexity of actual production machinery and the diversity of field noise, the acquired time-domain signals often fail to meet the sparsity requirements of SCA.

[0004] Therefore, it is unwise to perform SCA directly on the observed signal in the time domain. Furthermore, most existing SCA algorithms recover the source signals using a mixing matrix (i.e., the cluster centers of the present invention) when the number of sources is known. However, industrial sites have strong background noise and many interference sources, making the number of fault sources unknown in advance during actual testing. Summary of the Invention

[0005] The present invention provides a mechanical fault diagnosis method based on blind deconvolution, which can realize blind separation of composite faults when the number of sources is unknown, and use mechanical vibration signals to extract and diagnose fault features.

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

[0007] According to one aspect of the present invention, a mechanical fault diagnosis method based on blind deconvolution is provided, comprising:

[0008] Step A: picking up an observation signal of mechanical vibration through an acceleration sensor;

[0009] Step B: constructing a weighted energy threshold coefficient, and performing single source point detection on the mechanical vibration observation signal collected by the sensor based on the weighted energy threshold coefficient to obtain feature data and an estimated number of sources H; normalizing the filtered feature data to obtain feature data represented by the direction angle;

[0010] Step C, clustering the feature data represented by the direction angle and obtaining the cluster center and membership degree;

[0011] Step D: reconstructing and separating H source signals containing single faults from the signal containing the composite fault based on the cluster center and the membership degree, and performing inverse short-time Fourier transform on the separated source signals to obtain time domain source signals;

[0012] Step E: performing feature enhancement on the time domain source signal;

[0013] Step F: Perform envelope analysis on the enhanced signal to determine the fault type.

[0014] The step B comprises:

[0015] B1. Observation signal x Y×t Perform short-time Fourier transform to transform it into the time-frequency domain to obtain the time-frequency signal X(t,f); where Y is the number of sensors and t is the sampling time;

[0016] B2. Calculate the energy of the time-frequency signal X(t,f) at point I as E I , the expression is:

[0017] E I =∑[|X(t I ,f)|] 2 ,I=1,2,...L

[0018] Among them, X(t I ,f) represents the time-frequency signal at the moment of point I, L is the signal length;

[0019] B3. Based on energy E I , the average energy of the time-frequency signal X(t,f) is obtained as The calculated energy is greater than The proportion of data points in all data points is called the weighted energy threshold E e The calculation formula is as follows:

[0020]

[0021] B4. Use a Y×2Y sliding window centered at the current time-frequency point in the time-frequency signal X(t,f). Each time the window is slid, a Y×2Y real number matrix X(Ω) is selected. PCA is used to extract the local confidence T(Ω) from the matrix X(Ω). Feature data is extracted using the local confidence and stored in the matrix Z.

[0022] B5. Adjust the weighted energy threshold coefficient ε so that the proportion of the filtered feature data reaches the weighted energy threshold E e , the expression is:

[0023]

[0024] Where ||Z||2 represents the l2 norm of all time-frequency vectors in the extracted feature data, represents the average value of the l2 norm of all time-frequency vectors in the extracted feature data, and ε is the weighted energy threshold coefficient;

[0025] B6. Normalize the feature data to obtain the direction angle β k The characteristic data z k =(cosβ k ,sinβ k ); Estimate the number of sources H based on the normalized characteristic data.

[0026] The feature enhancement of time domain source signal is performed using maximum correlation kurtosis deconvolution based on improved autocorrelation function.

[0027] The envelope analysis of the enhanced signal to determine the fault type includes: performing Hilbert transform on the feature-enhanced signal, and then performing envelope analysis using fast Fourier transform to check whether the fault characteristic frequency and its multiples calculated by bearing theory are present on the envelope spectrum;

[0028] If it exists, then it is determined that the frequency on the envelope spectrum corresponds to the frequency of the actual mechanical component type, that is, the fault type is determined, and finally the fault diagnosis is achieved;

[0029] If not present, it is assumed that no fault exists.

[0030] According to another aspect of the present invention, there is provided a mechanical fault diagnosis device, comprising:

[0031] A pickup module, used for picking up observation signals of mechanical vibration through an acceleration sensor;

[0032] The first acquisition module is used to construct a weighted energy threshold coefficient and perform single source point detection on the mechanical vibration observation signal collected by the sensor based on the weighted energy threshold coefficient to obtain characteristic data and an estimated number of sources H; the filtered characteristic data is normalized to obtain characteristic data represented by the direction angle;

[0033] The second acquisition module is used to perform clustering processing on the feature data represented by the direction angle and obtain the cluster center and membership degree;

[0034] The third acquisition module is used to reconstruct and separate H source signals containing single faults from the signal containing composite faults based on the cluster centers and membership degrees, where H is the dimension of the cluster centers, and perform inverse short-time Fourier transform on the separated source signals to obtain time domain source signals;

[0035] An enhancement module, used to enhance the characteristics of the time domain source signal;

[0036] The judgment module is used to perform envelope analysis on the enhanced signal to determine the fault type.

[0037] The beneficial effects of the present invention are:

[0038] (1) Taking into account the low energy value of noise in engineering practice, the present invention proposes a weighted energy threshold coefficient for the first time and applies it to the single source point detection method, thereby improving the accuracy of feature data selection;

[0039] (2) The present invention not only shortens the time cost of the algorithm but also accurately extracts the fault frequency in the envelope spectrum; compared with the traditional SCA algorithm (fuzzy C-means clustering + linear programming), the time cost is reduced;

[0040] (3) The method of the present invention completes the blind separation and extraction of bearing composite faults without knowing the number of sources;

[0041] (4) The present invention converts the time domain signal into the time-frequency domain for analysis, so the signal is not required to fully meet the sparsity. This can fully weaken the influence of other interference signals on the fault identification of the separated source signal and realize the blind separation of bearing composite faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a diagram of the bearing fault simulation test bench and sensor locations in the present invention;

[0043] Figure 2 The time domain waveforms of the vibration observation signals picked up by the sensors during the operation of the bearing fault simulation test bench in Example 2 of the present invention are shown in the figure above. The time domain waveforms of the vibration observation signals picked up by sensor I are shown in the figure below.

[0044] Figure 3 This is the envelope spectrum of the vibration observation signal picked up by the sensor when the bearing fault simulation test bench of Example 2 of the present invention is running; the upper figure is the envelope spectrum of the vibration observation signal picked up by sensor I, and the lower figure is the envelope spectrum of the vibration observation signal picked up by sensor II;

[0045] Figure 4 This is a time domain scatter plot of the vibration observation signal picked up by the sensor when the bearing fault simulation test bench of Example 2 of the present invention is running;

[0046] Figure 5 This is a time-frequency domain scatter plot of the vibration observation signal picked up by the sensor after single-source point detection when the bearing fault simulation test bench of Example 2 of the present invention is running;

[0047] Figure 6This is the envelope spectrum of the vibration observation signal picked up by the sensor during operation of the bearing fault simulation test bench in Example 2 of the present invention after preliminary separation through sparse component analysis; the upper figure is the preliminary separated envelope spectrum of the bearing outer ring fault, and the lower figure is the preliminary separated envelope spectrum of the bearing inner ring fault;

[0048] Figure 7 The time domain waveforms of the outer race fault after the initial separation and IMCKD of Example 2 of the present invention are shown in the upper figure. The time domain waveforms of the outer race fault after the initial separation and IMCKD are shown in the lower figure.

[0049] Figure 8 The time domain waveforms of the inner race fault initially separated in Example 2 of the present invention after IMCKD are shown in the upper figure. The time domain waveforms of the inner race fault initially separated after IMCKD are shown in the lower figure.

[0050] Figure 9 This is the envelope spectrum of the fault signal analyzed by the mechanical fault diagnosis method based on blind deconvolution using sparse component analysis based on single-source point detection combined with improved maximum correlation kurtosis deconvolution in Example 2 of the present invention. The upper figure shows the envelope spectrum of the outer ring fault signal, and the lower figure shows the envelope spectrum of the inner ring fault signal.

[0051] Figure 10 The results of Example 2 of the present invention using morphological filtering combined with sparse component analysis;

[0052] Figure 11 The four-axis machining center in Example 3 of the present invention;

[0053] Figure 12 This is the time domain waveform of the vibration observation signal picked up by the sensor when the Y-axis of the four-axis machining center of Example 3 of the present invention is running. The upper figure is the time domain waveform of the vibration observation signal picked up by the axial sensor, and the lower figure is the time domain waveform of the vibration observation signal picked up by the radial sensor;

[0054] Figure 13 This is the envelope spectrum of the vibration observation signal picked up by the sensor when the Y-axis of the four-axis machining center of Example 3 of the present invention is running; wherein, the upper figure is the envelope spectrum of the vibration observation signal picked up by the axial sensor, and the lower figure is the envelope spectrum of the vibration observation signal picked up by the radial sensor;

[0055] Figure 14 This is a time domain scatter plot of the vibration observation signal picked up by the sensor when the Y-axis of the four-axis machining center is running in Example 3 of the present invention;

[0056] Figure 15 This is a time-frequency domain scatter plot of the vibration observation signal picked up by the sensor after single source point detection when the Y-axis of the four-axis machining center in Example 3 of the present invention is running;

[0057] Figure 16 This is the envelope spectrum of the vibration observation signal picked up by the sensor during the operation of the Y-axis of the four-axis machining center in Example 3 of the present invention after preliminary separation through sparse component analysis; the upper figure is the preliminary separated envelope spectrum of the bearing outer ring fault, and the lower figure is the preliminary separated envelope spectrum of the bearing inner ring fault;

[0058] Figure 17 The time domain waveforms of the outer race fault initially separated in Example 3 of the present invention after IMCKD; the upper figure is the time domain waveform of the outer race fault initially separated before IMCKD, and the lower figure is the time domain waveform of the outer race fault initially separated after IMCKD;

[0059] Figure 18 The time domain waveforms of the inner race fault initially separated in Example 3 of the present invention after IMCKD are shown. The upper figure shows the time domain waveform of the inner race fault initially separated before IMCKD, and the lower figure shows the time domain waveform of the inner race fault initially separated after IMCKD.

[0060] Figure 19 This is the envelope spectrum of the fault signal analyzed by the mechanical fault diagnosis method based on blind deconvolution using sparse component analysis based on single-source point detection combined with improved maximum correlation kurtosis deconvolution in Example 3 of the present invention. The upper figure shows the envelope spectrum of the outer ring fault signal, and the lower figure shows the envelope spectrum of the inner ring fault signal.

[0061] Figure 20 The results of Example 3 of the present invention using morphological filtering combined with sparse component analysis;

[0062] Figure 21 Flowchart of the present invention;

[0063] Figure 1 Middle: a-bearing seat, b-sensor I, c-sensor II, d-motor;

[0064] Figure 11 Middle: e-CNC machine tool Y axis, f-CNC machine tool X axis. DETAILED DESCRIPTION

[0065] The invention will be further described below with reference to the accompanying drawings and embodiments, but the content of the present invention is not limited to the scope of the drawings.

[0066] Example 1: Figure 21As shown, according to one aspect of an embodiment of the present invention, a mechanical fault diagnosis method based on blind deconvolution is provided, including: step A, picking up the observation signal of mechanical vibration through an acceleration sensor; step B, constructing a weighted energy threshold coefficient, and performing single source point detection on the mechanical vibration observation signal collected by the sensor based on the weighted energy threshold coefficient to obtain feature data and an estimated number of sources H; normalizing the screened feature data to obtain feature data represented by the direction angle; step C, clustering the feature data represented by the direction angle and obtaining the cluster center and membership; step D, reconstructing and separating H source signals containing a single fault from the time-frequency signal containing a composite fault based on the cluster center and the membership, where H is the dimension of the cluster center, and performing an inverse short-time Fourier transform on the separated source signal to obtain a time domain source signal; step E, performing feature enhancement on the time domain source signal; and step F, performing envelope analysis on the enhanced signal to determine the fault type.

[0067] Furthermore, the step A includes: firstly, installing an acceleration sensor on the housing surface of the mechanical equipment, and the mechanical vibration observation signal picked up by the acceleration sensor is x Y×t , where Y is the number of sensors and t is the sampling time;

[0068] Furthermore, the step B comprises:

[0069] B1. Observation signal x Y×t Perform short-time Fourier transform to transform it into the time-frequency domain to obtain the time-frequency signal X(t,f); where Y is the number of sensors and t is the sampling time;

[0070] B2. Calculate the energy of the time-frequency signal X(t,f) at point I as E I , the expression is:

[0071] E I =∑[|X(t I ,f)|] 2 ,I=1,2,...L

[0072] Among them, X(t I ,f) represents the time-frequency signal at the moment of point I, L is the signal length;

[0073] B3. Based on energy E I , the average energy of the time-frequency signal X(t,f) is obtained as The calculated energy is greater than The proportion of data points in all data points is called the weighted energy threshold E e The calculation formula is as follows:

[0074]

[0075] The calculated energy is greater than The proportion of data points in all data points is called the weighted energy threshold E e ; The energy in the time-frequency signal X(t,f) is greater than The data points are taken as feature data x(t,f); that is, the present invention determines the proportion of all feature data x(t,f) in the total according to the weighted energy threshold;

[0076] B4. Use a Y×2Y sliding window centered at the current time-frequency point in the time-frequency signal X(t,f). Since the mixing matrix of the instantaneous mixing model is real, take the real part of each element and select a Y×2Y real matrix X(Ω) for each sliding. Use PCA to extract the local confidence T(Ω) from the matrix X(Ω). When the region is a single-source domain, that is, when there is only one active source signal in the region, T(Ω) will obtain a larger value. If the region is a multi-source domain or a passive domain, T(Ω) will obtain a smaller value. The local confidence T(Ω) is defined as:

[0077]

[0078] where λ1(Ω)>λ2(Ω>...>λ Y (Ω) is the descending eigenvalue of the covariance matrix of X(Ω); the feature data is extracted using the local confidence and stored in the matrix Z.

[0079] B5. Since there are many points with small amplitude near zero, these low energy points can be removed by the following formula to further reduce the error of the mixing matrix estimation. Specifically, by adjusting the weighted energy threshold coefficient ε, the proportion of the filtered feature data reaches E e , the expression is:

[0080]

[0081] Where ||Z||2 represents the l2 norm of all time-frequency vectors in the extracted feature data, represents the average value of the l2 norm of all time-frequency vectors in the extracted feature data, and ε is the weighted energy threshold coefficient;

[0082] B6. Normalize the feature data to obtain the direction angle β k The characteristic data z k =(cosβ k ,sinβ k ); Estimate the number of sources H based on the normalized characteristic data.

[0083] Furthermore, the step C includes: using the fuzzy C mean based on the cosine distance to normalize the feature data z k Perform clustering to obtain cluster centers and membership degrees;

[0084] C1. The distance between the normalized feature data and the cluster center can be expressed by the cosine distance

[0085]

[0086] Among them, h is the number of clusters, c h is the cluster center, z k is the characteristic data, is the angle between the cluster center and the feature data. Cosine distance measures the similarity between two feature vectors by measuring the cosine value of the angle between them;

[0087] C2. Use cosine distance instead of the Euclidean distance in the traditional fuzzy C-means to update the cluster center and membership calculation formula. The new cluster center and membership calculation formula are as follows:

[0088]

[0089]

[0090] Where η is the update step size, c h + is the updated cluster center, w hk is the updated membership degree.

[0091] D. When all the filtered feature data x(t,f) point to H sources, reconstruct the sources according to the following steps. is the sample assigned to the hth source, L is the length of the time-frequency signal X(t,f) containing the composite fault, c h is the corresponding cluster center, that is, the column corresponding to the mixing matrix. Initialization And h=1,...,H can be used to calculate the characteristic data x according to the following formula Sh Reconstruct and separate H source signals: Use inverse short-time Fourier transform to transform the initially separated source signal into the time domain; where u h is the reconstructed h-th source signal.

[0092] Furthermore, the step E comprises: h Improved maximum correlation kurtosis deconvolution (IMCKD) is performed to enhance the fault characteristics. The steps are as follows:

[0093] E1, initialize the maximum number of iterations Gen, filter length L1, filter coefficient F = [01...00] T , number of iterations d = 1;

[0094] E2. Calculate the source signal u hThe envelope autocorrelation function R u (τ)=∫u(t)u(t+τ)dt, when the above formula takes the maximum value, the translation coefficient τ is recorded as τ max , in the iterative process T s =τ max ;

[0095] E3, N is the number of signal sampling points, calculate the signal u h The autocorrelation matrix of and its inverse matrix calculate r=mT s ,m=1,2,...,M;

[0096] E4 and M are displacement constants, according to To u h Filter and calculate

[0097] E5. Calculate the autocorrelation function of the filtered signal and update T s 、 X r , Filter again and repeat E2-E4 until the maximum number of iterations is reached;

[0098] E6. In all cycles, the filtered signal corresponding to the maximum kurtosis value is selected as the optimal filtered signal, thereby completing feature enhancement.

[0099] Furthermore, step F includes: performing Hilbert transform on the feature-enhanced signal, and then performing envelope analysis using fast Fourier transform to check whether the fault characteristic frequency and its multiples calculated by bearing theory exist on the envelope spectrum; if so, then determining whether the frequency on the envelope spectrum corresponds to the frequency of the actual mechanical component type, that is, determining the fault type, and ultimately achieving fault diagnosis; if not, it is considered that no fault exists.

[0100] According to another aspect of an embodiment of the present invention, there is provided a machine fault diagnosis device, comprising:

[0101] A pickup module, used for picking up observation signals of mechanical vibration through an acceleration sensor;

[0102] The first acquisition module is used to construct a weighted energy threshold coefficient and perform single source point detection on the mechanical vibration observation signal collected by the sensor based on the weighted energy threshold coefficient to obtain characteristic data and an estimated number of sources H; the filtered characteristic data is normalized to obtain characteristic data represented by the direction angle;

[0103] The second acquisition module is used to perform clustering processing on the feature data represented by the direction angle and obtain the cluster center and membership degree;

[0104] The third acquisition module is used to reconstruct and separate H source signals containing single faults from the signal containing composite faults based on the cluster centers and membership degrees, where H is the dimension of the cluster centers, and perform inverse short-time Fourier transform on the separated source signals to obtain time domain source signals;

[0105] An enhancement module, used to enhance the characteristics of the time domain source signal;

[0106] The judgment module is used to perform envelope analysis on the enhanced signal to determine the fault type.

[0107] Example 2: This example diagnoses a composite bearing fault in a rotating test bench. The bearing model used in this example is NU205, and the relevant parameters are: pitch circle diameter 39 mm, rolling element diameter 7.5 mm, number of rolling elements 12, contact angle 0 degrees; the inner ring of the bearing rotates on the main shaft, and the outer ring is fixed;

[0108] The diagnostic steps using the method of the present invention are as follows:

[0109] (1) Number of sampling points N = 40960, sampling frequency fs = 8192 Hz. The data collected at 800 rpm are analyzed, i.e., the rotation frequency fs = 8192 Hz. r =13.33Hz. The characteristic frequency of the bearing components and the characteristic frequency of the outer ring f are further obtained by calculation. o =64.61Hz, characteristic frequency f of the inner ring i =95.38Hz. The data acquisition hardware is NI9234 sensor, and the rolling bearing vibration signal is picked up by a pair of PCB acceleration sensors installed perpendicularly to each other on the bearing seat (see Figure 1 ), obtain the observation signal x of the bearing vibration picked up by the acceleration sensor Y×t , where Y is the number of sensors, which is 2, t = N / fs is the sampling time, x Y×t The time domain waveform is as follows Figure 2 As shown, the corresponding envelope spectrum is as follows Figure 3 As shown in the figure, it is clear from the envelope spectrum that the characteristic frequency components of the outer and inner race faults are completely mixed together, making it impossible to identify their faults.

[0110] (2) For the observed signal x Y×t Perform single source point detection to obtain feature data z k , the observed signal x Y×t The time domain scatter plot is as follows Figure 4 As shown, the characteristic data z k The time-frequency domain scatter plot is as follows Figure 5 As shown, two linearly distributed straight lines are presented, and it is determined that there are two source signals;

[0111] (3) For the characteristic data zk Clustering is performed using the DFCM clustering algorithm to obtain cluster centers and membership degrees;

[0112] (4) Based on the cluster center and membership degree, H source signals containing single faults are reconstructed and separated from the signal containing composite faults, where H is the dimension of the cluster center. The separated source signals are subjected to inverse short-time Fourier transform to obtain the time domain source signal;

[0113] The inverse short-time Fourier transform is used to transform the preliminary reconstructed source signal into the time domain. The preliminary reconstructed source signal u h Envelope spectrum Figure 6 As shown, Figure 6 The envelope spectrum in the figure above only contains the outer ring fault frequency (f o ) and its frequency multiple (2f o , 3f o ),but Figure 6 The envelope spectrum in the figure below contains not only the inner ring fault frequency f i and its frequency multiple (2f i , 3f i ) and also contains the outer race fault frequency f o Therefore, it can be seen that directly performing envelope spectrum analysis on the initially reconstructed source signal cannot effectively distinguish faults; in view of the shortcomings of directly performing envelope spectrum analysis, the present invention further uses IMCKD to analyze the initially reconstructed source signal u h Perform feature enhancement;

[0114] (5) Use IMCKD to reconstruct the source signal u h Perform feature enhancement, and the time domain waveforms before and after feature enhancement are as follows Figure 7 Figure 8 As shown;

[0115] (6) The envelope analysis results of the signal after feature enhancement are as follows Figure 9 As shown, from Figure 9 It can be clearly seen that the outer ring fault frequency (f o ) and its frequency multiples, which are consistent with the fault characteristics of the bearing outer ring. After feature enhancement, the rotation frequency (f r ) and inner race fault characteristic frequency (f i ) and its frequency doublings, and (f i +f r , f i -f r ) is consistent with the inner race fault characteristics. In summary, the outer race and inner race fault characteristics are clearly separated and extracted. Figure 10 Only the outer race fault frequency (f o) and its 2nd and 3rd frequency, the inner race fault envelope spectrum is separated, which contains not only the inner race fault frequency f i Also includes the outer ring fault frequency f o Fault diagnosis is not implemented.

[0116] Example 3: The data collected in this example is from a 4-axis machining center. The Y axis of the machine tool has a fault and periodic impact can be clearly heard. Data collection and analysis are performed on the Y axis and X axis motor end support bearings. Figure 11 As shown. The sensor is installed on the Y axis. The bearing model is 30TAC62B, with a pitch diameter of 47mm, a rolling element diameter of 6.85mm, 19 rollers, a contact angle of 60 degrees, and a screw lead of 10 (pitch of 10mm).

[0117] The steps of the method of the present invention are as follows:

[0118] 1. (Single-wire screw pitch is equal to lead), the rotation frequency is equal to the feed speed divided by the lead, and the feed speed is selected as 4m / s, which is the rotation frequency f r =6.67Hz, number of sampling points N = 25600, sampling frequency fs = 25600Hz, and the characteristic frequency of the bearing components and the characteristic frequency of the outer ring f are further obtained by calculation. o =58.72Hz, characteristic frequency f of the inner ring i =67.95Hz. Get the observation signal x of the bearing vibration picked up by the acceleration sensor Y×t , where Y is the number of sensors, which is 2, and t = N / f s is the sampling time, x Y×t The time domain waveform is as follows Figure 12 As shown, the corresponding envelope spectrum is as follows Figure 13 As shown in the figure, it is clear from the envelope spectrum that the characteristic frequency components of the outer and inner race faults are completely mixed together, making it impossible to identify their faults.

[0119] 2. Observe the signal x Y×t Perform single source point detection to obtain feature data z k , the observed signal x Y×t The time domain scatter plot is as follows Figure 14 As shown, the characteristic data z k The time-frequency domain scatter plot is as follows Figure 15 As shown, two linearly distributed straight lines are presented, and it is determined that there are two source signals;

[0120] 3. For feature data z k Perform DFCM to obtain the cluster center c l , membership degree w lk ;

[0121] 4. Based on the cluster center and membership degree, H source signals containing single faults are reconstructed and separated from the signal containing composite faults. H is the dimension of the cluster center. The separated source signals are subjected to inverse short-time Fourier transform to obtain the time domain source signal. The preliminary reconstructed source signal u h Envelope spectrum Figure 16 As shown; Figure 16 The envelope spectrum in the figure above only contains the outer ring fault frequency f o ,but Figure 16 The inner ring fault frequency (f i ) and its frequency multiple (2f i ) is relatively weak, so it can be seen that directly performing envelope spectrum analysis on the initially reconstructed source signal cannot effectively distinguish faults;

[0122] 5. Use IMCKD to reconstruct the source signal u i Perform feature enhancement, and the time domain waveforms before and after feature enhancement are as follows Figure 17 Figure 18 As shown;

[0123] 6. The envelope analysis results of the signal after feature enhancement are as follows Figure 19 As shown, from Figure 19 It can be clearly seen that the outer ring fault frequency (f o ) and its 2 times frequency, which is consistent with the fault characteristics of the bearing outer ring. After feature enhancement, the envelope spectrum of the inner ring signal can be clearly seen to have 3 times the rotation frequency (3f r ), 6 times the frequency (6f r ) and inner race fault characteristic frequency (f i ) spectrum, which is consistent with the inner race fault characteristics. Therefore, the outer and inner race fault characteristics are clearly separated and extracted. It should be noted that the frequency error is caused by the frequency resolution. The rotation frequency obtained by the present invention is 6.08Hz, which has an error of 0.59Hz compared with the theoretical rotation frequency. Therefore, when the present invention performs fault diagnosis on the envelope spectrum, it determines the frequency of the fault characteristics calculated for the actual faulty bearing by considering this error. By considering the above error and combining it with the above analysis, it can be seen that Figure 19 There are almost no other interference components in the two component graphs, indicating that the algorithm can completely separate the fault; the frequency of the fault characteristics calculated by the actual faulty bearing exists on the envelope spectrum, which determines the fault type and ultimately realizes fault diagnosis. Figure 20 Only the outer race fault frequency (f o ) and its double frequency, the inner race fault frequency is not separated and the fault diagnosis is not achieved.

[0124] The above technical solution has been verified not only in a laboratory environment, but also based on actual engineering, further proving that the present invention can be applied to more complex engineering practices and improving the practical basis of the algorithm.

[0125] The specific embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in this field without departing from the scope of the present invention.

Claims

1. A mechanical fault diagnosis method based on blind deconvolution, characterized in that: include: Step A: picking up an observation signal of mechanical vibration through an acceleration sensor; Step B: constructing a weighted energy threshold coefficient, and performing single source point detection on the mechanical vibration observation signal collected by the sensor based on the weighted energy threshold coefficient to obtain feature data and an estimated number of sources H; normalizing the filtered feature data to obtain feature data represented by the direction angle; Step C, clustering the feature data represented by the direction angle and obtaining the cluster center and membership degree; Step D: reconstructing and separating H source signals containing single faults from the signal containing the composite fault based on the cluster center and the membership degree, and performing inverse short-time Fourier transform on the separated source signals to obtain time domain source signals; Step E: performing feature enhancement on the time domain source signal; Step F: Perform envelope analysis on the enhanced signal to determine the fault type; The step B comprises: B1. Observation signal x Y×t Perform short-time Fourier transform to transform it into the time-frequency domain to obtain the time-frequency signal X(t,f); where Y is the number of sensors and t is the sampling time; B2. Calculate the energy of the time-frequency signal X(t,f) at point I as E I , the expression is: E I =∑[|X(t I ,f)|] 2 ,I=1,2,...L Among them, X(t I ,f) represents the time-frequency signal at the moment of point I, L is the signal length; B3. Based on energy E I , the average energy of the time-frequency signal X(t,f) is obtained as The calculated energy is greater than The proportion of data points in all data points is called the weighted energy threshold E e The calculation formula is as follows: B4. Use a Y×2Y sliding window centered at the current time-frequency point in the time-frequency signal X(t,f). Each time the window is slid, a Y×2Y real number matrix X(Ω) is selected. PCA is used to extract the local confidence T(Ω) from the matrix X(Ω). Feature data is extracted using the local confidence and stored in the matrix Z. B5. Adjust the weighted energy threshold coefficient ε so that the proportion of the filtered feature data reaches the weighted energy threshold E e , the expression is: Where ||Z||2 represents the l2 norm of all time-frequency vectors in the extracted feature data, represents the average value of the l2 norm of all time-frequency vectors in the extracted feature data, and ε is the weighted energy threshold coefficient; B6. Normalize the feature data to obtain the direction angle β k The characteristic data z k =(cosβ k ,sinβ k ); Estimate the number of sources H based on the normalized characteristic data.

2. The mechanical fault diagnosis method based on blind deconvolution according to claim 1, characterized in that: The feature enhancement of time domain source signal is performed using maximum correlation kurtosis deconvolution based on improved autocorrelation function.

3. The mechanical fault diagnosis method based on blind deconvolution according to claim 1, characterized in that: The envelope analysis of the enhanced signal to determine the fault type includes: performing Hilbert transform on the feature-enhanced signal, and then performing envelope analysis using fast Fourier transform to check whether the fault characteristic frequency and its multiples calculated by bearing theory are present on the envelope spectrum; If it exists, then it is determined that the frequency on the envelope spectrum corresponds to the frequency of the actual mechanical component type, that is, the fault type is determined, and finally the fault diagnosis is achieved; If not present, it is assumed that no fault exists.

4. A mechanical fault diagnosis device using the mechanical fault diagnosis method based on blind deconvolution according to any one of claims 1 to 3, characterized in that: include: A pickup module, used for picking up observation signals of mechanical vibration through an acceleration sensor; The first acquisition module is used to construct a weighted energy threshold coefficient and perform single source point detection on the mechanical vibration observation signal collected by the sensor based on the weighted energy threshold coefficient to obtain characteristic data and an estimated number of sources H; the filtered characteristic data is normalized to obtain characteristic data represented by the direction angle; The second acquisition module is used to perform clustering processing on the feature data represented by the direction angle and obtain the cluster center and membership degree; The third acquisition module is used to reconstruct and separate H source signals containing single faults from the signal containing composite faults based on the cluster centers and membership degrees, where H is the dimension of the cluster centers, and perform inverse short-time Fourier transform on the separated source signals to obtain time domain source signals; An enhancement module, used to enhance the characteristics of the time domain source signal; The judgment module is used to perform envelope analysis on the enhanced signal to determine the fault type.

Images