Bearing fault diagnosis method based on morphological filtering and orthogonal matching pursuit

Abstract

The invention belongs to the technical field of rolling bearing fault detection, and particularly relates to a bearing fault diagnosis method based on morphological filtering and orthogonal matching pursuit. Comprising the steps of collecting rolling bearing vibration acceleration signals as fault diagnosis signals; firstly, multi-scale morphological filtering processing is carried out on signals, and different fault components are decomposed into different-scale filtering signals to realize separation of composite faults; then, the optimal scale component is screened out through a characteristic energy factor index to serve as an original signal of a K-SVD learning dictionary database, then a signal is reconstructed from the dictionary database through an orthogonal matching pursuit algorithm, then a residual signal is obtained by subtracting a reconstructed signal from the original signal in combination with an iterative differencing idea and serves as subsequent input, the steps are repeated, and the K-SVD learning dictionary database is obtained. And finally, carrying out envelope analysis on all the obtained reconstructed signals to determine the fault type of the rolling bearing. The method can successfully extract and separate the rolling bearing composite fault submerged by strong background noise.

Description

Technical Field

[0001] The invention belongs to the technical field of rolling bearing fault detection, and in particular relates to a bearing fault diagnosis method based on morphological filtering and orthogonal matching pursuit. Background Art

[0002] Rolling bearings are widely used in rotating machinery, and bearing failures often lead to mechanical failures. In recent years, rolling bearing fault diagnosis has received increasing attention in academic research and industrial applications. However, in actual production life, rolling bearing failures are not only manifested as single failures, but also as compound failures. When two or more failures occur at the same time, the fault sources often couple and interact with each other, aggravating the damage to the rolling bearing and causing far more damage to the machinery than a single failure. Therefore, in order to ensure the safe operation of rotating machinery, it is necessary to develop effective rolling bearing compound fault diagnosis technology.

[0003] Since the vibration responses generated when local faults occur in different parts of the bearing will interfere with each other, and the resonance frequency bands excited when different parts fail are also different, the vibration acceleration signals collected by the sensor are mostly nonlinear, non-stationary and multi-component. At present, scholars commonly use methods to process multi-component non-stationary signals, including wavelet transform (WT), empirical mode decomposition (EMD), variational mode decomposition (VMD), etc. Although the above methods have achieved certain results in rolling bearing fault diagnosis, there are some problems that need to be solved; the selection of wavelet basis in the wavelet transform method has always been a difficult problem, and the number of decomposition layers cannot be adaptively selected; the EMD method lacks a rigorous mathematical theoretical basis, and problems such as mode aliasing and endpoint effect will occur when adaptively decomposing signals; although the VMD method overcomes problems such as mode aliasing and endpoint effect, it is necessary to confirm parameters such as the number of mode decomposition k and the penalty factor α before use. Summary of the invention

[0004] In view of the shortcomings of the existing multi-component non-stationary signal processing methods, the purpose of the present invention is to provide a bearing fault diagnosis method based on morphological filtering and orthogonal matching pursuit, including collecting rolling bearing vibration acceleration signals as fault diagnosis signals; firstly, multi-scale morphological filtering is performed on the signal, and structural elements of different scales have different recognition effects on each fault component, so as to obtain filtered signals of corresponding scales, each filtered processed signal contains unique and different fault information, so as to achieve separation of composite faults; then, the best scale component is selected as the original signal of the K-SVD learning dictionary library through the characteristic energy factor index, and then the signal is reconstructed from the dictionary library through the orthogonal matching pursuit algorithm; then, in combination with the iterative difference idea, the residual signal is obtained by subtracting the reconstructed signal from the original signal as the subsequent input, and the above steps are repeated, and finally, all the obtained reconstructed signals are envelope analyzed to determine the fault type of the rolling bearing.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is: A bearing fault diagnosis method based on morphological filtering and orthogonal matching pursuit collects rolling bearing vibration acceleration signals as fault diagnosis signals; firstly, the signal is processed by multi-scale morphological filtering, and the recognition effects of structural elements of different scales on each fault component are different, so as to obtain filter signals of corresponding scales, and each filter processed signal contains unique and different fault information, so as to separate composite faults; then, the best scale component is selected by characteristic energy factor index as the original signal of K-SVD learning dictionary library, and then the signal is reconstructed from the dictionary library by orthogonal matching pursuit algorithm, and then combined with iterative difference idea, the residual signal is obtained by subtracting the reconstructed signal from the original signal as the subsequent input, and the above steps are repeated, and finally, all the obtained reconstructed signals are subjected to envelope analysis to determine the fault type of rolling bearing.

[0006] The filter of morphological filtering adopts CMFH filter, and the structural element is a straight line with an amplitude of 1. CMFH filter is composed of basic morphological filters, namely dilation filter, erosion filter, opening filter and closing filter, and its expression is as follows: Dilation filter:

[0007] Corrosion Filter:

[0008] Open operation filter:

[0009] Closed operation filter:

[0010] in, are defined in and One-dimensional discrete signal on , and ( , is the input signal, is the structural element. Since the opening and closing operations can remove the positive pulse sequence and the negative pulse sequence respectively, the combination morphological filter of opening and closing and closing and opening and closing is constructed according to the combination order, which can take into account all the filtering performance of the opening and closing operation filters. Its expression is as follows: Switching filter:

[0011] Close the filter:

[0012] Although the on-off filter and the closed-open filter can remove the positive and negative pulses of the signal at the same time, there will be a statistical bias defect in the actual application process, that is, the amplitude of the on-off filter signal will become smaller, while the amplitude of the closed-open filter signal will increase, which directly affects the calculation result. Therefore, using any combination of filtering alone cannot achieve the ideal filtering effect. In order to overcome this defect, subsequent scholars took the arithmetic mean of the closed-open and open-close filtering as the final filtering result, and constructed the closed-open-open-close combined average filter, also known as the combination morphological filter (CMF), which is expressed as follows:

[0013] Inspired by the self-complementary Top-hat transform, the CMF filtered signal is subtracted from the original signal to obtain the CMFH morphological filter, which has the advantage of Top-hat transform enhancing detail features while extracting positive and negative pulse sequences. Its expression is as follows:

[0014] The scale selection range of multi-scale morphological filtering is as follows:

[0015] Where: is the maximum value of the scale, is the signal sampling frequency, is the target failure frequency of the bearing. Among bearing failures, the most common ones are the outer ring and inner ring failures of the bearing, while the rolling element and cage failures occur less frequently. The theoretical calculation formulas for the four failure frequencies are as follows: Outer race fault:

[0016] Inner race fault:

[0017] Rolling element failure:

[0018] Cage failure:

[0019] in, is the number of rolling elements, is the shaft speed, is the pitch diameter of the rolling bearing, is the rolling element diameter (mm), is the contact angle of rolling bearing (°).

[0020] Screening out the best scale component through characteristic energy factor index includes the following steps: The characteristic energy factors of different components obtained by morphological filtering at different scales are calculated using the following formula:

[0021] Where: is the amplitude energy of the fault characteristic frequency and its multiples in the envelope spectrum, is the total energy of the signal envelope spectrum amplitude, is the fault characteristic frequency When the fault signal is expanded using Fourier series, a combination of sinusoidal signals expanded to 3 to 5 times the fault frequency can fully describe the main frequency components of the fault signal. Therefore, in the present invention, FEF describes the proportion of the fault frequency and its frequency multiple information in the entire fault signal. The maximum FEF indicates that the fault information in the filtered signal is the most prominent. The scale corresponding to the maximum FEF value is selected as the input signal for subsequent processing. Y is the target signal.

[0022] The adaptive learning of the dictionary based on K-SVD includes the following steps: (1) The objective optimization function is:

[0023] Where: is the target signal, ( ) is a sparse representation dictionary, is the coefficient, Indicates that the error is estimated using the Frobenius norm, is the sparsity constraint factor; (2) The core of the K-SVD dictionary D update process is to update each column of the dictionary D in turn; update the kth column , and its corresponding sparse coefficient vector is , the process is as follows:

[0024] Where: is the kth column of dictionary D; is the kth row of the sparse matrix X; is the updated residual matrix; (3) Use singular value decomposition method to process the residual matrix The components corresponding to different singular values ​​are obtained as follows:

[0025] In the formula: jSingular Components , Indicates j singular values, represents a left singular matrix U No. j List, represents a left singular matrix V No. j Column, Δ is the singular value diagonal matrix with the singular values ​​arranged in descending order, using U The columns are updated in sequence D , get the updated new dictionary.

[0026] The signal is reconstructed from the dictionary library through the orthogonal matching pursuit algorithm: each seeks a best matching atom in the updated dictionary library , that is, using the optimal scale signal and dictionary The inner product operation is performed on each atom in the is the best matching atom, based on which the reconstructed signal is obtained.

[0027] Combined with the idea of ​​iterative difference, the residual signal obtained by subtracting the reconstructed signal from the original signal is used to filter the first fault signal. The above steps are repeated for the residual signal until the FEF value is less than 0.001 at all scales for two consecutive times, and the iteration stops.

[0028] All reconstructed signals obtained in the above process are processed by Hilbert envelope transform, and the prominent frequency components in the envelope spectrum are extracted and compared with the theoretical fault characteristic frequencies of various bearing components to determine the type of bearing fault.

[0029] The beneficial effects achieved by the present invention are: (1) Morphological filtering is based on mathematical morphology and set theory. It uses structural elements of different shapes and sizes to perform corrosion and expansion operations on signals. Due to its principle characteristics, there will be no modal aliasing and endpoint effects. Small-scale structural elements can capture subtle changes and remove high-frequency noise; large-scale smooth signals and remove low-frequency noise. Different scales respond differently to signals of different frequencies, and can identify fault components in different resonance bands and achieve complex fault separation. It has obvious advantages in vibration signal processing for mechanical fault diagnosis.

[0030] (2) Multiscale morphological filtering is based on the former, using multiple scale structural elements to process signals. The characteristics of the filtered signals at each scale are different, and the characteristic energy factor can measure the information content. Fault information is concentrated in a specific frequency. Multiscale morphological filtering separates signals by frequency. The signals selected based on the characteristic energy factor can accurately reflect the fault information and have a significant effect in power system fault detection.

[0031] (3) K-SVD is a dictionary learning algorithm based on sparse representation. By iteratively updating the dictionary atoms and sparse representation coefficients of the denoised signal, the fault impact atoms can be learned. After separating different faults, interference is eliminated to ensure the accuracy of dictionary atom learning, which helps in the diagnosis of bearing faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 Flowchart of the algorithm in the present invention.

[0033] Figure 2 It is the original time domain waveform and envelope spectrum of the experimental signal in the present invention.

[0034] Figure 3 It is the amplitude-frequency response diagram of the CMFH filter used in the present invention.

[0035] Figure 4 It is the time domain waveform and envelope spectrum of the first fault signal reconstructed after being processed by the algorithm in the present invention.

[0036] Figure 5 It is the time domain waveform and envelope spectrum of the second fault signal reconstructed after being processed by the algorithm in the present invention. DETAILED DESCRIPTION

[0037] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0038] The present invention provides a bearing fault diagnosis method based on morphological filtering and orthogonal matching pursuit, comprising collecting a rolling bearing vibration acceleration signal as a fault diagnosis signal; firstly performing multi-scale morphological filtering on the signal, using structural elements of different scales to have different recognition effects on each fault component, obtaining a filtering signal of a corresponding number of scales, each filtering processing signal containing unique and different fault information, and realizing separation of composite faults; then selecting the best scale component as the original signal of a K-SVD learning dictionary library through a characteristic energy factor index, reconstructing the signal from the dictionary library through an orthogonal matching pursuit algorithm, and then combining the iterative difference idea, using the original signal to subtract the reconstructed signal to obtain a residual signal as a subsequent input, repeating the above steps, and finally performing envelope analysis on all the obtained reconstructed signals to determine the fault type of the rolling bearing.

[0039] The filter of the morphological filtering adopts a CMFH filter, and the structural element is a straight line with an amplitude of 1; the scale selection range of the multi-scale morphological filtering is as follows:

[0040] Where: is the maximum value of the scale, is the signal sampling frequency, is the target failure frequency of the bearing; Screening out the best scale component through characteristic energy factor index includes the following steps: The characteristic energy factors of different components obtained by morphological filtering at different scales are calculated using the following formula:

[0041] Where: is the amplitude energy of the fault characteristic frequency and its multiples in the envelope spectrum, is the total energy of the signal envelope spectrum amplitude, is the fault characteristic frequency Considering the computational efficiency, we take , select the scale corresponding to the maximum value of FEF as the input signal for subsequent processing; The adaptive learning of the dictionary based on K-SVD includes the following steps: (1) The objective optimization function is:

[0042] Where: is the target signal, ( ) is a sparse representation dictionary, is the coefficient, Indicates that the error is estimated using the Frobenius norm, is the sparsity constraint factor; (2) The core of the K-SVD dictionary D update process is to update each column of the dictionary D in turn. To update the kth column For example, the corresponding sparse coefficient vector is , the process is as follows:

[0043] Where: is the kth column of dictionary D; is the kth row of the sparse matrix X; is the updated residual matrix; (3) Use singular value decomposition method to process the residual matrix The components corresponding to different singular values ​​are obtained as follows:

[0044] In the formula: j Singular Components , Indicates j singular values, represents a left singular matrix U No. j List, represents a left singular matrix V No. jColumns. Δ is a diagonal matrix of singular values ​​arranged in descending order. U The columns are updated in sequence D , get the updated new dictionary.

[0045] Reconstruct the signal from the dictionary library using the orthogonal matching pursuit algorithm: Find the best matching atom in each updated dictionary , that is, using the optimal scale signal and dictionary The inner product operation is performed on each atom in the is the best matching atom, based on which the reconstructed signal is obtained; Combined with the idea of ​​iterative difference, the residual signal obtained by subtracting the reconstructed signal from the original signal is used to filter the first fault signal. The above steps 2-6 are repeated for the residual signal until the FEF value is less than 0.001 at all scales for two consecutive times, and the iteration stops; All reconstructed signals obtained in the above process are processed by Hilbert envelope transform, and the prominent frequency components in the envelope spectrum are extracted and compared with the theoretical fault characteristic frequencies of various bearing components to determine the type of bearing fault.

[0046] The specific implementation steps of the present invention are as follows: The target rolling bearing is detected by the vibration acceleration sensor to obtain the corresponding vibration acceleration signal ; The filter of the morphological filter adopts CMFH filter, and the structural element is a linear type with an amplitude of 1; the collected vibration acceleration signal Perform multi-scale morphological filtering to obtain filter models at different scales ,in is the length of the structural element scale. Multiscale morphological filter scale By formula Determine, where: is the maximum value of the scale, is the signal sampling frequency, is the target fault frequency of the bearing; the optimal scale component is selected through the characteristic energy factor index and made into The characteristic energy factor is given by the formula Determine, where: is the amplitude energy of the fault characteristic frequency and its multiples in the envelope spectrum, is the total energy of the signal envelope spectrum amplitude, is the fault characteristic frequency Considering the computational efficiency, we use .

[0047] The FEF maximum filtered signal As the original signal for K-SVD dictionary learning. Determine the length n and number K of atoms in the initial dictionary. Divide the original signal according to the length n of the atom, and randomly select K atoms from it to form the initial dictionary D. Update the atoms in the dictionary D column by column. To update the kth column For example, the corresponding sparse coefficient vector is , the process is as follows: , where: is the kth column of dictionary D; is the kth row of the sparse matrix X; is the updated residual matrix; Using singular value decomposition to process the residual matrix Get the components corresponding to different singular values. By formula Determine, where: j Singular Components , Indicates j singular values, represents a left singular matrix U No. j List, represents a left singular matrix V No. j Columns. Δ is a diagonal matrix of singular values ​​arranged in descending order. U The columns are updated in sequence D , get the updated new dictionary .

[0048] The original signal and the updated dictionary library D are used as the input of the orthogonal matching pursuit algorithm to calculate the coefficient sparse matrix X, based on which the reconstructed signal is obtained The reconstructed signal is given by the formula Solve, where: To reconstruct the signal ( I is the number of iterations), is the updated dictionary, is the reconstruction coefficient. Combined with the idea of ​​iterative difference, the residual signal is obtained by subtracting the reconstructed signal from the original signal ,Right now , the residual signal As the new original signal, the above steps are iteratively performed until the iteration termination condition is met, that is, when the FEF value at all scales is less than 0.001, the iteration stops; all reconstructed signals obtained in the above process are processed by Hilbert envelope transform, and the prominent frequency components in the envelope spectrum are extracted and compared with the theoretical fault characteristic frequencies of various bearing components to determine the type of bearing fault.

[0049] The flow chart of the bearing fault diagnosis method based on morphological filtering and orthogonal matching pursuit of the present invention is as follows: Figure 1 As shown, the specific steps include the following steps.

[0050] The embodiment uses the composite fault vibration acceleration signal of a cylindrical roller bearing used in a certain type of asynchronous motor as experimental data for analysis. The fault type is the inner ring and rolling element fault of the bearing. The signal sampling frequency is 12800Hz, the number of sampling points is 102400, and the speed is 600r / min. The formula calculates that at this speed, the characteristic frequencies of the inner ring, outer ring, roller and cage faults of the bearing are 80.8Hz, 59.5Hz, 34.4Hz and 4.3Hz respectively.

[0051] First, the target device is detected by the vibration acceleration sensor to obtain the corresponding vibration acceleration signal ; Figure 2 The collected vibration acceleration signals are From the time domain waveform and Hilbert envelope spectrum, we can see that due to the interference of strong background noise, the fault impulse component is submerged and the fault characteristic frequency cannot be identified. The operator used for morphological filtering is selected. The filter of the morphological filtering adopts CMFH filter, and the structural element is a straight line with an amplitude of 1; Figure 3 The figure shows the amplitude-frequency response of the CMFH operator under different scale structural elements. As can be seen from the figure, as the scale decreases, the cutoff frequency gradually increases, that is, more high-frequency components are retained. This shows that the CMFH operator can separate the fault features step by step according to the scale change law.

[0052] The collected vibration acceleration signal Perform multi-scale morphological filtering to obtain filter models at different scales The multi-scale morphological filter scale is given by the formula Determine, where: is the maximum value of the scale, is the signal sampling frequency, is the target fault frequency of the bearing; the optimal scale component in the current filter signal is selected through the characteristic energy factor index and made into .

[0053] The characteristic energy factor is given by the formula Determine, where: is the amplitude energy of the fault characteristic frequency and its multiples in the envelope spectrum, is the total energy of the signal envelope spectrum amplitude, is the fault characteristic frequency Considering the computational efficiency, we use .

[0054] The FEF maximum filtered signal As the original signal for K-SVD dictionary learning. Determine the length n and number K of atoms in the initial dictionary. Divide the original signal according to the length n of the atom, and randomly select K atoms from it to form the initial dictionary D. Update the atoms in the dictionary D column by column. To update the kth column For example, the corresponding sparse coefficient vector is , the process is as follows: , where: is the kth column of dictionary D; is the kth row of the sparse matrix X; is the updated residual matrix; the residual matrix is ​​processed using the singular value decomposition method Get the components corresponding to different singular values.

[0055] in By formula Determine, where: j Singular Components , Indicates j singular values, represents a left singular matrix U No. j List, represents a left singular matrix V No. j Columns. Δ is a diagonal matrix of singular values ​​arranged in descending order. U The columns are updated in sequence D , get the updated new dictionary .

[0056] The original signal and the updated dictionary library As the input of the orthogonal matching pursuit algorithm, the coefficient sparse matrix X is calculated, based on which the reconstructed signal is obtained The reconstructed signal is given by the formula Sure, Where: To reconstruct the signal ( I is the number of iterations), is the updated dictionary, is the reconstruction coefficient. Combined with the idea of ​​iterative difference, the residual signal is obtained by subtracting the reconstructed signal from the original signal ,in . The residual signal As the new original signal, the above operation is iteratively performed until the iteration termination condition is met, that is, when the FEF value at all scales is less than 0.001, the iteration stops; all reconstructed signals obtained in the above process are processed by Hilbert envelope transform, and the prominent frequency components in the envelope spectrum are extracted and compared with the theoretical fault characteristic frequencies of various bearing components to determine the type of bearing fault.

[0057] Figure 4 This is the first fault signal separated from the original vibration acceleration signal by the algorithm in this paper. The obvious impact component can be seen from the time domain waveform, and the frequency conversion appears in its Hilbert envelope spectrum. and its frequency multiple 2 , 3 , 4 , and successfully extracted the bearing inner ring fault characteristic frequency .

[0058] Figure 5 This is the second fault signal separated from the original vibration acceleration signal by the algorithm in this paper. The obvious impact component can be seen from the time domain waveform, and the cage fault characteristic frequency appears in the Hilbert envelope spectrum. and its frequency multiple 2 , 3 , and successfully extracted the roller fault characteristic frequency and its frequency multiple 2 .

[0059] In summary, the bearing fault diagnosis method based on morphological filtering and orthogonal matching pursuit of the present invention can successfully extract and separate the rolling bearing compound faults submerged by strong background noise.

Claims

1. A bearing fault diagnosis method based on morphological filtering and orthogonal matching pursuit, characterized in that: The vibration acceleration signal of the rolling bearing is collected as the signal for fault diagnosis. First, the signal is processed by multi-scale morphological filtering, and different fault components are decomposed into different scale filter signals to achieve the separation of composite faults. Then, the optimal scale component is screened out through the characteristic energy factor index as the original signal of the K-SVD learning dictionary library, and then the signal is reconstructed from the dictionary library through the orthogonal matching pursuit algorithm. Then, combined with the iterative difference idea, the residual signal is obtained by subtracting the reconstructed signal from the original signal as the subsequent input, and the above steps are repeated. Finally, all the reconstructed signals are subjected to envelope analysis to determine the fault type of the rolling bearing.

2. The bearing fault diagnosis method based on morphological filtering and orthogonal matching pursuit according to claim 1 is characterized in that: The filter of morphological filtering adopts CMFH filter, and the structural element is a straight line with an amplitude of 1. The CMFH filter is composed of basic morphological filters, namely, dilation filter, corrosion filter, opening operation filter and closing operation filter. Its expression is as follows: Dilation filter: ， Corrosion Filter: ， Open operation filter: ， Closed operation filter: ， in, are defined in and One-dimensional discrete signal on , and N ≥ M, is the input signal, is the structural element; a combination of open-close and close-open morphological filters is constructed, and its expression is as follows: Switching filter: = ， Close the filter: = ， The arithmetic mean of the closed-open and open-close filtering is taken as the final filtering result, and a closed-open-open combined average filter is constructed, also known as a combined morphological filter CMF, which is expressed as follows: ， Subtract the CMF filtered signal from the original signal to obtain the CMFH morphological filter, which is expressed as follows: 。 3. The bearing fault diagnosis method based on morphological filtering and orthogonal matching pursuit according to claim 2 is characterized in that: The scale selection range of multi-scale morphological filtering is as follows: ， Where: is the maximum value of the scale, is the signal sampling frequency, is the target failure frequency of the bearing.

4. The bearing fault diagnosis method based on morphological filtering and orthogonal matching pursuit according to claim 3 is characterized in that: Screening out the best scale component through characteristic energy factor index includes the following steps: The characteristic energy factor energy of the components decomposed at different scales by morphological filtering is calculated using the following formula: ， Where: is the amplitude energy of the fault characteristic frequency and its multiples in the envelope spectrum, is the total energy of the signal envelope spectrum amplitude, is the fault characteristic frequency Frequency doubling, select the scale corresponding to the maximum value of FEF as the input signal for subsequent processing, Y is the target signal.

5. The bearing fault diagnosis method based on morphological filtering and orthogonal matching pursuit according to claim 4 is characterized in that: 。 6. The bearing fault diagnosis method based on morphological filtering and orthogonal matching pursuit according to claim 5 is characterized in that: The adaptive learning of the dictionary based on K-SVD includes the following steps: (1) The objective optimization function is: , Where: is the target signal, ( ) is a sparse representation dictionary, is the coefficient, Indicates that the error is estimated using the Frobenius norm, is the sparsity constraint factor; (2) The core of the K-SVD dictionary D update process is to update each column of the dictionary D in turn; update the kth column , and its corresponding sparse coefficient vector is , the process is as follows: , Where: is the kth column of dictionary D; is the kth row of the sparse matrix X; is the updated residual matrix; (3) Use singular value decomposition method to process the residual matrix The components corresponding to different singular values ​​are obtained as follows: , In the formula: j Singular Components , Indicates j singular values, represents a left singular matrix U No. j List, represents a left singular matrix V No. j Column, Δ is the singular value diagonal matrix with the singular values ​​arranged in descending order, using U The columns are updated in sequence D , get the updated new dictionary.

7. The bearing fault diagnosis method based on morphological filtering and orthogonal matching pursuit according to claim 6 is characterized in that: The signal is reconstructed from the dictionary library through the orthogonal matching pursuit algorithm: each seeks a best matching atom in the updated dictionary library , that is, using the optimal scale signal and dictionary The inner product operation is performed on each atom in the is the best matching atom, based on which the reconstructed signal is obtained.

8. The bearing fault diagnosis method based on morphological filtering and orthogonal matching pursuit according to claim 7 is characterized in that: Combined with the idea of ​​iterative difference, the residual signal obtained by subtracting the reconstructed signal from the original signal is used to filter the first fault signal. The above steps are repeated for the residual signal until the FEF value is less than 0.001 at all scales for two consecutive times, and the iteration stops.

9. The bearing fault diagnosis method based on morphological filtering and orthogonal matching pursuit according to claim 8 is characterized in that: All reconstructed signals obtained in the above process are processed by Hilbert envelope transform, and the prominent frequency components in the envelope spectrum are extracted and compared with the theoretical fault characteristic frequencies of various bearing components to determine the type of bearing fault.

Images