Mechanical fault diagnosis method based on sparse component analysis of density peak clustering
Through density peak clustering and orthogonal matching tracking algorithms, the complex fault signal is separated in mechanical equipment, and the problem of fault diagnosis in multi-fault signal coupling and strong noise environment is solved, and efficient fault identification and diagnosis is achieved.
Patent Information
- Application Number
- CN202210684400.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-06-17
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2042-06-17
Smart Images

Figure CN115017953B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a mechanical fault diagnosis method based on sparse component analysis of density peak clustering, belonging to the technical field of mechanical equipment state monitoring and fault diagnosis. Background Art
[0002] Rolling bearings are often installed in critical locations in rotating machinery, so even relatively subtle faults can be reflected in vibration signals. In actual industrial manufacturing, faults often originate from more than one source, meaning sensors often capture signals from multiple coupled sources rather than a single vibration source. This makes troubleshooting extremely difficult. Therefore, to address this issue and improve the accuracy of fault detection and diagnosis in modern equipment, the most critical step is to separate the fault signal from the mixed signal. This problem is known as blind source separation (BLS). Two common approaches to BLS are independent component analysis (ICA) and principal component analysis (SCA).
[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, performing SCA directly on the observed signal in the time domain is unwise. However, transforming it to the frequency domain or time-frequency domain makes it relatively easy to meet the sparsity requirement. Most existing SCA algorithms estimate 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 numerous interference sources, making the number of fault sources unknown in advance during actual testing. Summary of the Invention
[0005] In order to solve the above problems, the present invention provides a mechanical fault sparse component analysis algorithm based on density peak clustering.
[0006] Diagnostic method; this method can fully weaken the influence of other interference signals on the separated source signals for fault identification, and can realize blind separation of bearing composite faults when the number of sources is unknown, and use mechanical vibration signals for fault feature extraction and diagnosis.
[0007] The sparse component analysis algorithm based on density peak clustering of the present invention is implemented as follows: first, an acceleration sensor is installed on the shell surface of the mechanical equipment, and the acceleration sensor is used to pick up the observation signal of the mechanical vibration; secondly, a structural element is constructed using the sinC function, and a combination filter combining a morphological open-close filter and a morphological close-open filter is constructed based on the sinC type element; the combined filter is used to filter the picked-up mechanical vibration observation signal to obtain a filtered signal c h×t , the filtered signal c h×t Perform normalization processing to obtain the normalized signal And take the absolute value; the normalized signal is clustered by the density peak algorithm Perform clustering and obtain cluster centers. Finally, based on the cluster centers, use the orthogonal matching pursuit algorithm to reconstruct and separate k source signals containing single faults from the signal containing composite faults, where k is the dimension of the cluster centers. Perform FFT on the separated source signals, and then analyze the transformed signals in the frequency domain to determine the fault type and ultimately achieve fault diagnosis.
[0008] The specific steps of the above method are as follows:
[0009] A. First, install the acceleration sensor on the shell surface of the mechanical equipment, and pick up the observation signal x of the mechanical vibration through the acceleration sensor. h×t , where h is the number of sensors and t is the sampling time;
[0010] B. Use the sinC function to construct a structural element and construct a combined filter based on the sinC type element. The steps are as follows:
[0011] B1. Based on The function constructs a sinC-type element, then the structural element g(x) = sinC(x);
[0012] B2. Assume that the domains of the one-dimensional discrete input signal f(n) and the sequence structure element g(m) are D f ={x0,x1,…,x N-1} and D g ={y0,y1,…,y M-1}, and N>M, then the erosion (Θ) and expansion of f(n) with respect to g(m) They are defined as:
[0013] (fΘg)(n)=min{f(n+m)-g(m)}
[0014]
[0015] Where: (n+m), (nm)∈D f , m∈D g, N is the signal length, M is the length of the structural element;
[0016] B3, corresponding form opening Defined as The morphological closing (·) operation is defined as Then we get the morphological opening-closing operation and Close-Open (CO) operation
[0017] B4, then the combined filter is y(n) = [OC[f(n)] + CO[f(n)]] / 2;
[0018] C. Use the combined filter y(n) to filter the observation signal x of the mechanical vibration detected in step A. h×t Filter and obtain the filtered signal c h×t ;
[0019] D. Then the filtered signal c in step C is h×t Perform normalization processing to obtain the normalized signal And take the absolute value;
[0020] E. Normalized signal Clustering is performed using the density peak clustering algorithm to obtain the cluster center;
[0021] E1. First select the truncation error d c , and use it to define neighbor points; by calculating the distance from the sample point x in all data i The Euclidean distance is less than d c All data points, that is, for each data point x i The local density ρ i , the formula is as follows:
[0022] ρ i =∑ i≠j χ(d ij -d c )
[0023] Where χ(x) is a piecewise function. When x < 0, χ(x) = 1, otherwise χ(x) = 0; d ij Represents the distance between data i and data j, d c represents the cutoff distance;
[0024] E2. For each data point x i , find the local density ρ i The nearest point higher than it, and record the shortest distance δ to that point i , the formula is as follows:
[0025]
[0026] For the data i with the largest local density, we can take δ i =max j d ij ;
[0027] E3. Cluster centers are defined as clusters with high density ρ i and a larger distance δ i Point x i ; Let A i =ρ i δ i ,Pick is the cluster center, N is the total number of samples;
[0028] F. Based on the cluster centers, the orthogonal matching pursuit algorithm is used to reconstruct and separate k source signals containing single faults from the signal containing compound faults, where k is the dimension of the cluster centers:
[0029] F1, initialization parameters, initial residual r0, number of iterations L, Fourier orthogonal transform matrix E t×t , according to formula B ij =E t×t A ij Constructing a sensor matrix
[0030] F2. Use the inner product method to calculate the column vector of the sensing matrix W and the residual r i The projection coefficient, and record the position β corresponding to the maximum projection coefficient i , then set the column of the sensing matrix W corresponding to the maximum projection coefficient to 0;
[0031] F3, use the least squares method to calculate the estimated value x of the reconstructed signal of this iteration i =(β i T β i ) -1 β i T ·r i ;
[0032] F4, update residual r i+1 =r i -x i , and repeat step F2 until the iteration ends;
[0033] F5, use E t×t Perform inverse Fourier transform reconstruction to obtain a time domain signal x with a dimension of (kt×1), and according to the dimension n of the cluster center, split the time domain signal x with a dimension of (kt×1) into k time domain signals with a dimension of (t×1), thereby completing the blind separation of the signal;
[0034] G. Perform FFT on the separated source signal, and then analyze the frequency domain of the transformed signal to check whether the theoretically calculated fault characteristic frequency and its multiples of the mechanical components (including bearings, gears, etc.) are present on the spectrum graph;
[0035] If it exists, then it is determined that the frequency on the spectrum graph corresponds to the frequency of the actual mechanical component fault type, that is, the fault type is determined, and finally the fault diagnosis is achieved;
[0036] If it does not exist, it is considered that there is no fault and the spectrum diagram only shows the characteristic frequencies of the mechanical components.
[0037] (1) The present invention takes into account the special shape of the sinC function, which is very consistent with the impact characteristics of the vibration signal. Therefore, the sinC function is used for the first time to construct a structural element and applied to the morphological filtering algorithm, achieving a good noise reduction effect;
[0038] (2) The present invention uses the orthogonal matching pursuit algorithm to reconstruct the source signal, which greatly shortens the time cost of the algorithm and improves the clarity of the frequency domain. By comparing with the traditional SCA algorithm (fuzzy C-means clustering + linear programming), the experiment was repeated 50 times and the average value was taken. The time cost of the traditional SCA algorithm is about 90.2s, while the time cost of the algorithm of the present invention is about 22.05s. In comparison, the time cost is reduced by about 75%;
[0039] (3) The method of the present invention completes the blind separation and extraction of compound faults of mechanical components (including bearings, gears, etc.) without knowing the number of sources;
[0040] (4) The present invention converts the time domain signal into the 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
[0041] Figure 1 This is a diagram of the bearing fault simulation test bench and sensor locations in the present invention;
[0042] 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 1 of the present invention are shown in the figure above. The time domain waveforms of the vibration observation signals picked up by sensor I and the lower figure are shown in the figure below.
[0043] Figure 3 This is the spectrum of the vibration observation signal picked up by the sensor when the bearing fault simulation test bench of Example 1 of the present invention is running; the upper figure is the spectrum of the vibration observation signal picked up by sensor I, and the lower figure is the spectrum of the vibration observation signal picked up by sensor II;
[0044] Figure 4 Comparison diagram of the vibration observation signal picked up by the sensor before and after filtering when the bearing fault simulation test bench of Example 1 of the present invention is running; the upper figure is the spectrum of the vibration observation signal picked up by sensor I, and the lower figure is the spectrum of the vibration observation signal picked up by sensor II;
[0045] Figure 5 This is the result of density peak clustering of the filtered signal in Example 1 of the present invention;
[0046] Figure 6 The spectrum of the fault signal of Example 1 of the present invention is analyzed by the mechanical fault diagnosis method based on density peak clustering and frequency domain compressed sensing reconstruction. The upper figure is the spectrum of the outer race fault signal, and the lower figure is the spectrum of the inner race fault signal.
[0047] Figure 7 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.
[0048] Figure 8 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;
[0049] Figure 9 This is the result of density peak clustering of the filtered signal in Example 2 of the present invention;
[0050] Figure 10 The spectrum of the fault signal of Example 2 of the present invention is analyzed by the mechanical fault diagnosis method based on density peak clustering and frequency domain compressed sensing reconstruction. The upper figure is the spectrum of the inner race fault signal, and the lower figure is the spectrum of the outer race fault signal;
[0051] Figure 1 Middle: 1-motor, 2-gearbox, 3-bearing seat, 4-sensor I, 5-faulty bearing, 6-sensor II, 7-drive shaft, 8-turntable. DETAILED DESCRIPTION
[0052] Example 1: 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 D = 39 mm, rolling element diameter d = 7.5 mm, number of rolling elements Z = 12, and contact angle α = 0. The inner ring of the bearing rotates on the main shaft, while the outer ring is fixed. The speed is 800 r / min, that is, the rotational frequency is 13.33 Hz. Based on the above parameters, the characteristic frequency of the bearing outer ring fault is calculated to be 64.61 Hz, and the characteristic frequency of the inner ring fault is 95.38 Hz.
[0053] The diagnostic steps using the method of the present invention are as follows:
[0054] (1) Install two acceleration sensors (sensor I4 and sensor II6) perpendicular to each other on the housing surface of the bearing on the test bench (see Figure 1 ), using the NI Signal Express acquisition module and the NI-9234 four-channel acquisition card to acquire the signal and obtain the bearing vibration observation signal x picked up by the acceleration sensor h×t , where m is the number of sensors, 2, t is the sampling time, sampling frequency fs = 8192 Hz, number of sampling points N = 8192, x h×t The time domain waveform is as follows Figure 2 The corresponding spectrum is shown as Figure 3 As shown in the spectrum, it is clear that the characteristic frequency components of the outer and inner race faults are completely mixed together, making it impossible to identify their faults.
[0055] (2) Based on The function constructs a sinC-type element, then the structural element g(x) = sinC(x);
[0056] (3) Assume that the domains of the one-dimensional discrete input signal f(n) and the sequence structure element g(m) are D f ={x0,x1,…,x N-1} and D g ={y0,y1,…,y M-1}, and N>M, then the erosion (Θ) and expansion of f(n) with respect to g(m) They are defined as:
[0057] (fΘg)(n)=min{f(n+m)-g(m)}
[0058]
[0059] Where: (n+m), (nm)∈D f , m∈D g , N is the signal length, M is the length of the structural element;
[0060] (4) Open the corresponding form Defined as The morphological closing (·) operation is defined as Then we get the morphological opening-closing operation and close-open operations Then the combined filter is y(n) = [OC[f(n)] + CO[f(n)]] / 2;
[0061] (5) Use the combined filter y(n) to filter the vibration signal x detected in step (1) h×t Filter and obtain the filtered signal c h×t ,like Figure 4 As shown; the filtered signal is normalized to obtain the normalized signal
[0062] (6) Normalized signal Clustering is performed using the density peak clustering algorithm to obtain the cluster center. The steps are as follows:
[0063] a. First select the truncation error d c , and use it to define neighbor points by calculating the distance from the sample point x in all data i The Euclidean distance is less than d c All data points, that is, for each data point x i The local density ρ i , the formula is as follows:
[0064] ρ i =∑ i≠j χ(d ij -d c )
[0065] Where χ(x) is a piecewise function. When x < 0, χ(x) = 1, otherwise χ(x) = 0. ij Represents the distance between data i and data j, d c represents the cutoff distance;
[0066] b. For each data point x i , find the local density ρ i The nearest point higher than it, and record the shortest distance δ to that point i , the formula is as follows:
[0067]
[0068] For the data i with the largest local density, we can take δ i =max j d ij ;
[0069] c. The cluster center is defined as having high density ρ i and a larger distance δ i Point x i ; Let A i =ρ i δ i ,Pick is the cluster center, and the results are shown in Figure 5 ;
[0070] (7) Based on the cluster centers and their dimensions, the orthogonal matching pursuit algorithm is used to reconstruct and separate two source signals containing single faults from the signal containing composite faults;
[0071] a. Initialization parameters, initial residual r0, number of iterations L, Fourier orthogonal transform matrix E t×t , according to formula B ij =E t×t A ij Constructing a sensor matrix
[0072] b. Use the inner product method to calculate the column vector of the sensing matrix W and the residual r i The projection coefficient, and record the position β corresponding to the maximum projection coefficient i , then set the column of the sensing matrix W corresponding to the maximum projection coefficient to 0;
[0073] c. Use the least squares method to calculate the estimated value x of the reconstructed signal for this iteration i =(β i T β i ) -1 β i T ·r i ;
[0074] d. Update residual r i+1 =r i -x i , and repeat step F2 until the iteration ends;
[0075] e. Use E t×t Perform inverse Fourier transform reconstruction to obtain a time domain signal x with a dimension of (kt×1), and according to the dimension k of the cluster center, split the time domain signal x with a dimension of (kt×1) into k time domain signals with a dimension of (t×1), thereby completing the blind separation of the signal;
[0076] (8) The FFT result of the separated source signal is as follows Figure 6 As shown, from Figure 6 The first separated signal clearly identifies the amplitudes of the X1, X2, and X3 multiples of the outer race fault characteristic frequency of 65Hz. The second separated signal can distinguish the spectrum line of the rotation frequency of 13Hz and its multiples, as well as the more sidebands on both sides of 95Hz with an interval of 13Hz of the rotation frequency, which corresponds to the inner race fault frequency of 95.38Hz, and is diagnosed as an inner race fault. The frequency error is caused by the frequency resolution and is generally within ±2Hz. There are almost no other interference components in the two component diagrams, indicating that the algorithm can completely separate the fault. The frequency spectrum diagram shows the frequency and multiples of the fault characteristics calculated for the actual faulty bearing, which determines the fault type and ultimately achieves fault diagnosis.
[0077] Example 2: The relevant parameters of the faulty bearing model NU205 used in this example are the same as those in Example 1, except that the spindle speed in this example is 1200 r / min and the rotation frequency is 20 Hz. Based on the above parameters, the characteristic frequency of the bearing outer ring fault is calculated to be 96.92 Hz, and the characteristic frequency of the inner ring fault is 143 Hz.
[0078] The steps of the method of the present invention are as follows:
[0079] 1. Install two accelerometers (sensor I4 and sensor II6) perpendicular to each other on the bearing housing surface on the test bench. Use the NI Signal Express acquisition module and the NI-9234 four-channel acquisition card to acquire the signal and obtain the observation signal x of the mechanical vibration picked up by the accelerometer. h×t ; Sampling frequency fs = 8192 Hz, number of sampling points N = 8192, x h×t The time domain waveform is as follows Figure 7 As shown, the corresponding envelope spectrum is as follows Figure 8 As shown in the spectrum, it is clear that the characteristic frequency components of the outer and inner race faults are completely mixed together, making it impossible to identify their faults.
[0080] Where, h is the number of sensors, which is 2, and t is the sampling time;
[0081] 2. Use the sinC function to construct a structural element, and construct a combined filter based on the sinC type element, as in Example 1;
[0082] 3. Use the combined filter y(n) to filter the observed signal x of the mechanical vibration detected in step A. h×t Filter and obtain the filtered signal c h×t ; The filtered signal c h×t Perform normalization processing to obtain the normalized signal And take the absolute value;
[0083] 4. Normalized signal Clustering is performed using the density peak clustering algorithm to obtain cluster centers, similar to the steps in Example 1. The results are shown in Figure 9 ;
[0084] 5. Based on the cluster centers and their dimensions, the orthogonal matching pursuit algorithm is used to reconstruct and separate two source signals containing single faults from the signal containing compound faults;
[0085] 6. Perform FFT on the separated source signal and estimate the envelope spectrum of the source signal as follows Figure 10 ,from Figure 10The first separated signal can distinguish the 20Hz rotation frequency spectrum and its multiples, as well as the sidebands on either side of 141Hz, separated by 20Hz of the rotation frequency. This corresponds to the inner race fault frequency of 143Hz, confirming the inner race fault diagnosis. The second separated signal clearly identifies the outer race fault characteristic frequency of 97Hz and its multiple of 193Hz, corresponding to the inner race fault frequency of 96.92Hz, confirming the outer race fault diagnosis. The frequency error, caused by frequency resolution, is generally within ±2Hz. The two component plots are virtually free of other interference components, demonstrating that the algorithm can completely separate the faults. The frequency spectrum shows the calculated fault characteristic frequency and its multiples of the actual faulty bearing, confirming the fault type and ultimately achieving fault diagnosis.
[0086] 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 sparse component analysis of density peak clustering, characterized in that: Firstly, the acceleration sensor is installed on the shell surface of the mechanical equipment, and the observation signal of the mechanical vibration is picked up by the acceleration sensor; secondly, the structural element is constructed using the sinC function, and a combined filter combining the morphological open-closed filter and the morphological closed-open filter is constructed based on the sinC type element; the combined filter is used to filter the picked-up mechanical vibration observation signal to obtain the filtered signal c h×t , the filtered signal c h×t Perform normalization processing to obtain a normalized signal And take the absolute value; the normalized signal is clustered by the density peak algorithm Perform clustering processing and obtain cluster centers; finally, based on the cluster centers, use the orthogonal matching pursuit algorithm to reconstruct and separate k source signals containing single faults from the signal containing composite faults, where k is the dimension of the cluster center. Perform FFT on the separated source signals, and then analyze the transformed signals in the frequency domain to determine the fault type, thereby achieving fault diagnosis. The steps of using the sinC function to construct the structural element and constructing the combined filter based on the sinC type element are as follows: (1) Based on The function constructs a sinC-type element, then the structural element g(x) = sinC(x); (2) Assume that the domains of the one-dimensional discrete input signal f(n) and the sequence structure element g(m) are D f ={x0,x1,…,x N-1 } and D g ={y0,y1,…,y M-1 }, and N>M, then the erosion (Θ) and expansion of f(n) with respect to g(m) They are defined as: (fΘg)(n)=min{f(n+m)-g(m)} Where: (n+m), (nm)∈D f , m∈D g , N is the signal length, M is the length of the structural element; (3) Open the corresponding form Defined as The morphological closing (·) operation is defined as Then we get the morphological opening-closing operation and close-open operation Then the combined filter is y(n) = [OC[f(n)] + CO[f(n)]] / 2; The normalized signal is clustered by the density peak algorithm The steps for clustering are as follows: (1) First, select the truncation error d c , and use it to define neighbor points by calculating the distance from the sample point x in all data i The Euclidean distance is less than d c All data points, that is, for each data point x i The local density ρ i , the formula is as follows: r i =∑ i≠j x(d ij -d c ) Where χ(x) is a piecewise function. When x < 0, χ(x) = 1, otherwise χ(x) = 0; d ij Represents the distance between data i and data j, d c represents the cutoff distance; (2) For each data point x i , find the local density ρ i The nearest point higher than it, and record the shortest distance to that point δ i , the formula is as follows: For the data i with the largest local density, we take δ i =max j d ij ; (3) The cluster center is defined as having high density ρ i and the larger distance δ i Point x i ; Let A i =ρ i δ i ,Pick is the cluster center, and N is the total number of samples.
2. The mechanical fault diagnosis method based on sparse component analysis of density peak clustering according to claim 1 is characterized in that: The steps of using the orthogonal matching pursuit algorithm to reconstruct and separate k source signals containing single faults from a signal containing composite faults are as follows: F1, initialization parameters, initial residual r0, number of iterations L, Fourier orthogonal transform matrix E t×t , according to formula B ij =E t×t A ij Constructing the sensor matrix F2. Use the inner product method to calculate the column vector of the sensor matrix W and the residual r i The projection coefficient, and record the position β corresponding to the maximum projection coefficient i , and then set the column of the sensor matrix W corresponding to the maximum projection coefficient to 0; F3, use the least squares method to calculate the estimated value x of the reconstructed signal of this iteration i =(β i T β i ) -1 β i T ·r i ; F4, update residual r i+1 =r i -x i , and repeat F2 until the iteration ends; F5, use E t×t The inverse Fourier transform is performed to reconstruct the time domain signal x with a dimension of (kt×1), and according to the dimension k of the cluster center, the time domain signal x with a dimension of (kt×1) is divided into k time domain signals with a dimension of (t×1), thereby completing the blind separation of the signal.
3. The mechanical fault diagnosis method based on sparse component analysis of density peak clustering according to claim 1 is characterized in that: The mechanical vibration observation signal picked up by the acceleration sensor is x h×t , where h is the number of sensors and t is the sampling time.
4. The mechanical fault diagnosis method based on sparse component analysis of density peak clustering according to claim 1 is characterized in that: Analyze the signal after FFT in the frequency domain to check whether there is the theoretical characteristic frequency and its multiple frequency of mechanical component failure on the spectrum diagram; If it exists, then it is determined that the frequency on the spectrum diagram 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 it does not exist, it is considered that there is no fault and the spectrum diagram only shows the characteristic frequency of the mechanical parts.
