A vibration signal feature extraction method based on high-pass filtering
By combining high-pass filtering and K-SVD dictionary construction algorithms in offshore wind turbines, the problem of low-frequency signal perturbation in offshore wind turbine fault diagnosis is solved, achieving more accurate fault diagnosis and reducing maintenance costs.
Patent Information
- Application Number
- CN202210376605.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-04-08
- Publication Date
- 2025-02-11
- Estimated Expiration
- 2042-04-08
AI Technical Summary
The planetary gearbox fault diagnosis of offshore wind turbines is difficult to effectively deal with low-frequency signal disturbances caused by external waves and wind, resulting in inaccurate fault diagnosis and high maintenance costs.
Using a vibration signal feature extraction method based on high-pass filter, the low-frequency signal perturbation is eliminated through a digital high-pass filter, and the K-SVD dictionary construction algorithm is used to convert complex vibration signals into simple atom combinations, thereby extracting characteristic signals and achieving more accurate Troubleshooting.
It effectively eliminates the disturbance of external low-frequency signals to the gearbox vibration signal, improves the accuracy and efficiency of fault diagnosis, and reduces maintenance costs.
Smart Images
Figure CN114707554B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of offshore wind power generation, and in particular to a vibration signal feature extraction method based on high-pass filtering. Background Art
[0002] In recent years, the annual installed capacity and cumulative installed capacity of offshore wind turbines around the world have achieved tremendous growth, and the global offshore wind power ratio has also increased year by year. As the installed capacity increases, the number of accidents of wind turbines is also on the rise, and repairs are extremely difficult, especially for offshore wind turbines. How to ensure the smooth operation of offshore wind turbines, improve the operating rate of the units, and reduce the cost of later maintenance has become an urgent problem to be solved. Usually, the failure of the planetary gearbox of an offshore wind turbine can often be attributed to the failure of major components such as gears and bearings.
[0003] Offshore wind turbines are subject to many external loads such as waves and wind, which have a great impact on the detection and processing of vibration signals. Therefore, in the analysis process, we must first consider eliminating the disturbance of external low-frequency signals, then extract the characteristics of the vibration signal, and use this as a basis to judge the operating status of the gearbox.
[0004] Traditional fault diagnosis of mechanical equipment such as gearboxes is mostly achieved by signal processing methods, and most of them do not take into account the low-frequency disturbances of waves and wind. For example, Chinese patent CN112629851 discloses a fault diagnosis method for offshore wind turbine gearboxes based on data enhancement methods and image recognition. The method directly processes the radial vibration signal of the gearbox output shaft bearing collected. This method does not take into account the impact of the special working environment of offshore wind turbine gearboxes. Summary of the invention
[0005] The present invention provides a vibration signal feature extraction method based on high-pass filtering. The high-pass filtering method is used to eliminate the disturbance of low-frequency signals of external wind and waves on the vibration signal of the gear box. On this basis, the complex vibration signal can be converted into a combination of a group of simple atoms through the K-SVD dictionary construction algorithm, so that the feature signal can be extracted more ideally and more accurate fault diagnosis can be achieved.
[0006] In order to achieve the above object, the present invention provides the following technical solutions:
[0007] At least one embodiment of the present disclosure provides a vibration signal feature extraction method based on high-pass filtering, the method comprising:
[0008] (1) The signal acquisition module collects and stores the vibration signal of the offshore wind turbine gearbox in real time, collects the original vibration signal x(n), and transmits it to the PC;
[0009] (2) The collected original vibration signal x(n) is filtered by a digital high-pass filter to eliminate the low-frequency disturbance of the gearbox monitoring signal caused by waves and wind, retain the high-frequency vibration part, and obtain the high-frequency vibration signal y(n);
[0010] (3) Analyze the high-frequency vibration signal y(n) through the K-SVD dictionary construction algorithm to solve the sparse matrix coefficient S;
[0011] (4) Fix the sparse matrix coefficient S, use the K-SVD algorithm to update the K atoms of the dictionary one by one, obtain the trained dictionary D, and perform sparse decomposition on the dictionary D to obtain the reconstructed unit;
[0012] (5) The K-SVD reconstructed signal is spliced and fast Fourier transform is performed on the K-SVD reconstructed signal. The time domain and frequency domain signal diagrams are analyzed to diagnose the health status of the gearbox.
[0013] Preferably, the data acquisition module uses FPGA to complete the data acquisition and preprocessing process. After the system is initialized, AD sampling is performed, and the converted data is cached in the FIFO memory, and then the cached data is sent to the host computer through the UART serial port, completing the entire sampling process.
[0014] At least one embodiment of the present disclosure provides a vibration signal feature extraction method based on high-pass filtering, wherein the digital high-pass filter adopts a FIR digital filter;
[0015] The main technical indicators include passband cutoff frequency ω p , stopband cutoff frequency ω s and the maximum attenuation δ allowed in the passband p ,ω s -ω p is the transition zone Δω.
[0016] When operating in the special environment of the ocean, the vibration monitoring signal of the gearbox is greatly affected by the low-frequency signals from waves and sea breeze. The FIR digital filter has high stability, can effectively suppress low-frequency signals, remove interference, and can change the filtering conditions by changing technical indicators according to the external environment, and has strong adjustability.
[0017] At least one embodiment of the present disclosure provides a vibration signal feature extraction method based on high-pass filtering, and the FIR digital filter is designed using a window function method. The specific process is as follows:
[0018] (21) Select the window function according to the minimum attenuation requirement of the stop band, and then determine the size of the FIR digital filter order N by the transition bandwidth to obtain the window function ω(n), n=0,1,2,…,N-1;
[0019] (22) Calculate the cutoff frequency ω of the ideal high-pass filter c :
[0020]
[0021] Among them, ω s is the stopband cutoff frequency;
[0022] ω p is the passband cutoff frequency;
[0023] (23) Calculate the corresponding ideal filter unit impulse response h d (n):
[0024]
[0025] Where N is the order of the digital filter, which is usually an odd number;
[0026] a is the delay constant, a = (N-1) / 2;
[0027] h d (n) is even symmetric;
[0028] (24) Obtain the unit impulse response of the designed FIR digital filter h(n) = h d (n)ω(n),n=0,1,2,…,N-1;
[0029] (25) By finding H(e jω )=DTFT[h(n)] to verify whether the design result meets the design requirements. If not, redesign is required.
[0030] At least one embodiment of the present disclosure provides a vibration signal feature extraction method based on high-pass filtering, wherein the unit impulse response h of the ideal high-pass filter is d (n) is:
[0031]
[0032] Where, N is the order of the digital filter, and N is an odd number;
[0033] n is an ordinal number, n = 0, 1, 2, ..., N-1;
[0034] a is the time delay constant, a=(N-1) / 2.
[0035] Preferably, the unit impulse response of the ideal high-pass filter is even symmetrical.
[0036] At least one embodiment of the present disclosure provides a vibration signal feature extraction method based on high-pass filtering, wherein the relationship between the original vibration signal x(n) and the filtered high-frequency vibration signal y(n) is a convolution sum relationship, that is:
[0037] y(n)=x(n)*h(n).
[0038] At least one embodiment of the present disclosure provides a vibration signal feature extraction method based on high-pass filtering, and the specific solution process of the sparse matrix coefficient S in step (3) is:
[0039] The filtered high-frequency vibration signal y(n) is divided into units to obtain a training sample library for the dictionary learning algorithm. K samples are selected from the sample library to form the initial dictionary D0, and the sparse matrix coefficient S is solved by the orthogonal matching pursuit (OMP) algorithm. The solution steps are as follows:
[0040] Define the signal to be decomposed x, dictionary D, sparsity T, and number of loops t;
[0041] When there is no iteration, that is, when t = 0, the residual R 0 = x, the index set of atoms Selected Atom Set Sparse coefficient s 0 =0;
[0042] Preferably, the signal to be decomposed x, ie, the target signal to be processed, is a high-frequency signal y(n) when the OMP algorithm is used for the first time.
[0043] (31) Select the atom that best matches the current residual, calculate the dot product of the residual and all atoms, and the largest one is indexed as λ = argmax| <R t-1 ,d k >|;
[0044] Among them, R t-1 is the residual obtained from the previous cycle;
[0045] d k is a dictionary atom;
[0046] (32) Calculate the atomic index set Λ at the cycle number t t (t):
[0047] Λ t (t) = λ,
[0048] Update the selected atom set D at the time of loop number t t :
[0049] D t =[D t-1 ,d λ ];
[0050] (33) Calculate the sparse coefficient s at the number of cycles t t :
[0051]
[0052] (34) Calculate the residual R at the number of cycles t t :
[0053] R t =xD t s t ;
[0054] (35) Cycle number t+1;
[0055] (36) Determine whether to stop the loop. If the number of loops t> the sparsity T, stop the loop. Otherwise, repeat steps (31) to (35). The number of loops t at this time is t+1 in step (35).
[0056] Preferably, the orthogonal matching pursuit algorithm performs Schmidt orthogonalization on the selected atoms in each cycle, so that the convergence is faster and the computational efficiency is higher.
[0057] At least one embodiment of the present disclosure provides a vibration signal feature extraction method based on high-pass filtering, wherein the method for updating K atoms in step (4) is:
[0058] Define the training sample library Y = [y1, y2, y3, ...y n ], the training sample library Y is a matrix with p rows and n columns; the dictionary D = [d1, d2, d3, ... d K ], the dictionary D is a matrix with p rows and K columns; the corresponding sparse coefficient matrix S = [s1, s2, s3, ...s n ], the sparse coefficient matrix S is a matrix with K rows and n columns, the maximum number of loops is M, and the initial number of loops is t = 1;
[0059] (41) Determine the dictionary atom d of the kth atom k The sample set ω k :
[0060] ω k ={j|1≤j≤n,s j,k ≠0};
[0061] Among them, j is a natural number, which only represents the count;
[0062] (42) Calculate the error matrix E of the kth atom k :
[0063]
[0064] Wherein, i is a natural number, 1≤i≤k;
[0065] d i is the i-th atom in the dictionary;
[0066] is the sparse coefficient matrix S and d i The row that multiplies;
[0067] (43) Settings And Ω k Middle (ω k (j),j) elements are 1, and the other elements are 0; only the error matrix E is retained. k Non-zero value Keep the value after non-zero value
[0068] Among them, Ω k is an n×|ω k |-order matrix, and all elements in the matrix are real numbers, which have no practical meaning. It is used to multiply other matrices and remove the common elements;
[0069] is the sparse coefficient matrix S and the dictionary atom d k The row that multiplies;
[0070] (44) Only the error matrix E is retained k Non-zero value Do SVD singular value decomposition:
[0071] make
[0072] make
[0073]
[0074] Among them, U and V are unitary matrices obtained by SVD singular value decomposition;
[0075] Δ is a semi-positive definite diagonal matrix;
[0076] V T is the transpose of V;
[0077] Δ(1,1) represents the element corresponding to the first row and first column of the Δ matrix;
[0078] is the first column of the unitary matrix U, which is the kth atom d in the dictionary k Updated results;
[0079] U(:,1) represents the first column vector of the matrix;
[0080] (45) The number of cycles t increases by one;
[0081] (46) Check whether the loop should be terminated. If the loop number t is greater than the maximum loop number M, stop the loop; otherwise, repeat steps (42) to (45), where the loop number t is t in step (45) plus one.
[0082] Preferably, the K-SVD algorithm performs singular value decomposition on an error matrix, thereby achieving alternating changes between a dictionary and a sparse matrix.
[0083] At least one embodiment of the present disclosure provides a vibration signal feature extraction method based on high-pass filtering, wherein the diagnosis basis in step (5) is:
[0084] If the frequency in the frequency domain signal diagram is f m ±af p ±bf c If an obvious peak appears at the position of the gear, it indicates that there is a local fault in the planetary gear teeth of the gearbox;
[0085] Among them, a and b are natural numbers;
[0086] f m is the meshing frequency of the planetary gear set;
[0087] f p is the characteristic frequency of the planetary gear that causes damage;
[0088] f c is the rotation frequency of the planet carrier itself.
[0089] The local faults include broken teeth, chipped corners, pitting and the like.
[0090] The present invention has the following beneficial effects:
[0091] (1) The present invention fully considers the influence of low-frequency signals from waves and sea breeze on the vibration signal of the wind turbine gearbox in the special environment where the offshore wind turbine gearbox is located. The disturbance of low-frequency signals from external wind and waves on the vibration signal of the gearbox is eliminated by high-pass filtering. On this basis, the complex vibration signal can be converted into a combination of a group of simple atoms by the K-SVD dictionary construction algorithm, so that the characteristic signal can be extracted more ideally and more accurate fault diagnosis can be achieved.
[0092] (2) The filter of the present invention adopts an FIR digital filter, which has good system stability, is easy to achieve strict linear phase characteristics, and its unit impulse response is finite length. The fast Fourier transform algorithm can be used to implement the filtering process, thereby greatly improving the operation efficiency;
[0093] (3) The present invention adopts a method combining the K-SVD dictionary construction algorithm and the OMP sparse matrix solution algorithm as the main signal feature extraction method, which can realize the reconstruction of the gearbox vibration signal. Compared with the general method, the time domain diagram of the K-SVD reconstructed signal is clearer, and the peaks at special positions in the spectrum diagram are more obvious, so it is easier to identify whether there is a gear fault. BRIEF DESCRIPTION OF THE DRAWINGS
[0094] Figure 1 It is a flow chart of the signal feature extraction method of the present invention.
[0095] Figure 2 It is a design flow chart of the FIR digital filter of the present invention.
[0096] Figure 3 It is a schematic diagram of the amplitude-frequency response of the high-pass filter of the present invention.
[0097] Figure 4 It is a phase-frequency response diagram of the high-pass filter of the present invention.
[0098] Figure 5 It is the original signal spectrum diagram of the pitting planetary gearbox of the present invention.
[0099] Figure 6 It is the Gabor reconstructed signal spectrum diagram of the pitting planetary gearbox of the present invention.
[0100] Figure 7 It is the K-SVD reconstructed signal spectrum diagram of the pitting planetary gearbox of the present invention. DETAILED DESCRIPTION
[0101] The specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings. It should be pointed out that the embodiments are only specific explanations of the invention and should not be regarded as limitations of the invention. The purpose of the embodiments is to enable those skilled in the art to better understand and reproduce the technical solutions of the present invention. The protection scope of the present invention shall still be based on the scope defined by the claims.
[0102] like Figure 1 As shown, this embodiment provides a vibration signal feature extraction method based on high-pass filtering, the method comprising:
[0103] S1, the signal acquisition module collects and stores the vibration signal of the offshore wind turbine gearbox in real time, collects the original vibration signal x(n), and transmits it to the PC;
[0104] Preferably, the data acquisition module uses FPGA to complete the data acquisition and preprocessing process. After the system is initialized, AD sampling is performed, and the converted data is cached in the FIFO memory, and then the cached data is sent to the host computer through the UART serial port, completing the entire sampling process.
[0105] S2, filtering the collected original vibration signal x(n) through a digital high-pass filter to eliminate the low-frequency disturbance of the gearbox monitoring signal caused by waves and wind, retain the high-frequency vibration part, and obtain the high-frequency vibration signal y(n);
[0106] The digital high-pass filter adopts a FIR digital filter;
[0107] The main technical indicators include passband cutoff frequency ω p , stopband cutoff frequency ω s and the maximum attenuation δ allowed in the passband p ,ω s -ω p is the transition zone Δω.
[0108] When operating in the special environment of the ocean, the vibration monitoring signal of the gearbox is greatly affected by the low-frequency signals from waves and sea breeze. The FIR digital filter has high stability, can effectively suppress low-frequency signals, remove interference, and can change the filtering conditions by changing technical indicators according to the external environment, and has strong adjustability.
[0109] like Figure 2 As shown, the FIR digital filter is designed using the window function method. The specific process is:
[0110] S21, select the window function according to the minimum attenuation requirement of the stop band, and then determine the size of the FIR digital filter order N according to the transition bandwidth to obtain the window function ω(n), n=0,1,2,…,N-1;
[0111] S22, calculate the cutoff frequency ω of the ideal high-pass filter c :
[0112]
[0113] Among them, ω s is the stopband cutoff frequency;
[0114] ω p is the passband cutoff frequency;
[0115] S23, calculate the unit impulse response h of the corresponding ideal filter d (n):
[0116]
[0117] Where N is the order of the digital filter, which is usually an odd number;
[0118] a is the delay constant, a = (N-1) / 2;
[0119] h d (n) is even symmetric;
[0120] The amplitude and phase characteristics are obtained from h(n) as follows Figure 3-Figure 4 shown.
[0121] S24, obtain the unit impulse response h(n)=h of the designed FIR digital filter d (n)ω(n),n=0,1,2,…,N-1;
[0122] S25, by finding H(e jω )=DTFT[h(n)] to verify whether the design result meets the design requirements. If not, redesign is required.
[0123] The unit impulse response h of the ideal high-pass filter d (n) is:
[0124]
[0125] Where, N is the order of the digital filter, and N is an odd number;
[0126] n is an ordinal number, n = 0, 1, 2, ..., N-1;
[0127] a is the time delay constant, a=(N-1) / 2.
[0128] Preferably, the unit impulse response of the ideal high-pass filter is even symmetrical.
[0129] The relationship between the original vibration signal x(n) and the filtered high-frequency vibration signal y(n) is a convolution sum relationship, that is:
[0130] y(n)=x(n)*h(n).
[0131] S3, analyzing the high-frequency vibration signal y(n) by using the K-SVD dictionary construction algorithm to solve the sparse matrix coefficient S;
[0132] The filtered high-frequency vibration signal y(n) is divided into units to obtain a training sample library for the dictionary learning algorithm.
[0133] When dividing the signal unit, the collected data should be divided into several signal units of equal length, and the units are allowed to overlap, and it is ensured that in each signal unit, there is only one impact point between the gear and the ring gear, which is reflected as a pulse in the time domain. Therefore, through this method, an attempt is made to re-establish the pulse signal reflecting the fault impact in the signal unit, and the regular harmonic vibration signal is partially retained, which can better reflect the characteristics of the signal unit.
[0134] The orthogonal matching pursuit algorithm performs Schmidt orthogonalization on the selected atoms in each cycle, so it converges faster and has higher computational efficiency.
[0135] Select K samples from the sample library to form the initial dictionary D0, and use the orthogonal matching pursuit (OMP) algorithm to solve the sparse matrix coefficient S. The solution steps are:
[0136] Define the signal to be decomposed x, dictionary D, sparsity T, and number of loops t;
[0137] When there is no iteration, that is, when t = 0, the residual R 0 = x, the index set of atoms Selected Atom Set Sparse coefficient s 0 =0;
[0138] Preferably, the signal to be decomposed x, ie, the target signal to be processed, is a high-frequency signal y(n) when the OMP algorithm is used for the first time.
[0139] S31, select the atom that best matches the current residual, calculate the dot product of the residual and all atoms, and the largest one has an index of λ=argmax| <R t-1 ,d k >|;
[0140] Among them, R t-1 is the residual obtained from the previous cycle;
[0141] d k is a dictionary atom;
[0142] S32, calculate the atomic index set Λ at the cycle number t t (t):
[0143] Λ t (t) = λ,
[0144] Update the selected atom set D at the time of loop number t t :
[0145] D t =[D t-1 ,d λ ];
[0146] S33, calculate the sparse coefficient s when the number of cycles is t t :
[0147]
[0148] S34, calculate the residual R at the cycle number t t :
[0149] R t =xD t s t ;
[0150] S35, cycle number t+1;
[0151] S36, determine whether to stop the loop. If the number of loops t> the sparsity T, stop the loop. Otherwise, repeat steps S31 to S35. The number of loops t at this time is t+1 in step S35.
[0152] The number of cycles t of sparse decomposition needs to be obtained through a large number of data samples. If t is set too small, the basic waveform of the signal cannot be reconstructed, resulting in the inability to observe the frequency domain signal; on the contrary, if t is too large, the reconstructed signal almost overlaps with the original signal in the time-frequency domain, which violates the original intention of feature extraction.
[0153] Preferably, the orthogonal matching pursuit algorithm performs Schmidt orthogonalization on the selected atoms in each cycle, so that the convergence is faster and the computational efficiency is higher.
[0154] S4, fix the sparse matrix coefficient S, use the K-SVD algorithm to update the dictionary K atoms one by one, obtain the trained dictionary D, and perform sparse decomposition on the dictionary D to obtain the reconstructed unit;
[0155] The method for updating K atoms is:
[0156] Define the training sample library Y = [y1, y2, y3, ...y n ], the training sample library Y is a matrix with p rows and n columns; the dictionary D = [d1, d2, d3, ... d K ], the dictionary D is a matrix with p rows and K columns; the corresponding sparse coefficient matrix S = [s1, s2, s3, ...s n ], the sparse coefficient matrix S is a matrix with K rows and n columns, the maximum number of loops is M, and the initial number of loops is t = 1;
[0157] S41, determine the dictionary atom d of the kth atom k The sample set ω k :
[0158] ω k ={j|1≤j≤n,sj,k ≠0};
[0159] Among them, j is a natural number, which only represents the count;
[0160] S42, calculate the error matrix E of the kth atom k :
[0161]
[0162] Wherein, i is a natural number, 1≤i≤k;
[0163] d i is the i-th atom in the dictionary;
[0164] is the sparse coefficient matrix S and d i The row that multiplies;
[0165] S43, Settings And Ω k Middle (ω k (j),j) elements are 1, and the other elements are 0; only the error matrix E is retained. k Non-zero value Keep the value after non-zero value
[0166] Among them, Ω k is an n×|ω k |-order matrix, and all elements in the matrix are real numbers, which have no practical meaning. It is used to multiply other matrices and remove the common elements;
[0167] is the sparse coefficient matrix S and the dictionary atom d k The row that multiplies;
[0168] S44, only the error matrix E is retained k Non-zero value Do SVD singular value decomposition:
[0169] make
[0170] make
[0171]
[0172] Among them, U and V are unitary matrices obtained by SVD singular value decomposition;
[0173] Δ is a semi-positive definite diagonal matrix;
[0174] V T is the transpose of V;
[0175] Δ(1,1) represents the element corresponding to the first row and first column of the Δ matrix;
[0176] is the first column of the unitary matrix U, which is the kth atom d in the dictionary k Updated results;
[0177] U(:,1) represents the first column vector of the matrix;
[0178] S45, the number of cycles t is increased by one;
[0179] S46, check whether the loop should be terminated. If the loop number t is greater than the maximum loop number M, stop the loop, otherwise repeat steps S42-S45, where the loop number t is t in step S45 plus one.
[0180] Preferably, the K-SVD algorithm performs singular value decomposition on an error matrix, thereby achieving alternating changes between a dictionary and a sparse matrix.
[0181] S5, splicing to obtain K-SVD reconstructed signal, performing fast Fourier transform on the K-SVD reconstructed signal, and analyzing the time domain and frequency domain signal diagrams to diagnose the health status of the gearbox;
[0182] If the frequency in the frequency domain signal diagram is f m ±af p ±bf c If an obvious peak appears at the position of the gear, it indicates that there is a local fault in the planetary gear teeth of the gearbox;
[0183] Among them, a and b are natural numbers;
[0184] f m is the meshing frequency of the planetary gear set;
[0185] f p is the characteristic frequency of the planetary gear that causes damage;
[0186] f c is the rotation frequency of the planet carrier itself.
[0187] The local faults include broken teeth, chipped corners, pitting and the like.
[0188] like Figure 5-7 As shown in the figure, it is the spectrum diagram of the original signal of the pitting planetary gearbox and the two sets of reconstructed signals. By observation, it can be found that the original signal is m 、f m ±f c 、f m ±2f c 、f m ±3f cThere are peaks at the positions of , but no peaks are seen near the linear combination of the meshing frequency and the fault characteristic frequency. Therefore, based on the original signal alone, it is impossible to determine whether the load imbalance is caused by a local fault in the gear teeth, or whether the peak is caused by the amplitude modulation effect of the distributed fault and the pass effect. In the spectrum of the Gabor reconstructed signal, due to incomplete reconstruction, the spectrum diagram fails to reflect the characteristics of the original signal, the effect does not meet expectations, and it is impossible to determine whether a fault exists. In the K-SVD reconstructed signal spectrum, an obvious peak also appears at f m 、f m ±f c 、f m ±2f c 、f m ±3f c At the same time, at f m -f c -f p 、f m -3f p Peaks also appear at the same positions, which are manifestations of local gear tooth faults. Since laser pitting is a weak fault signal, its gear tooth damage is lighter than the first two fault modes, so the peak value at the peak is smaller. Based on the peak position in the K-SVD reconstructed signal spectrum, it can be judged that the planetary gear has a local fault.
[0189] Therefore, the K-SVD reconstructed signal can observe the peak representing the local gear tooth damage in the frequency domain. Compared with the original signal and the Gabor reconstructed signal, it is easier to confirm the existence of planetary gear tooth damage and has strong adaptability to experimental data.
[0190] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.
Claims
1. A vibration signal feature extraction method based on high-pass filtering, characterized in that: The method comprises: (1) The signal acquisition module collects and stores the vibration signal of the offshore wind turbine gearbox in real time, collects the original vibration signal x(n), and transmits it to the PC; (2) The collected original vibration signal x(n) is filtered by a digital high-pass filter to obtain a high-frequency vibration signal y(n); (3) The high-frequency vibration signal y(n) is analyzed by the K-SVD dictionary construction algorithm, and the sparse matrix coefficient S is solved by the orthogonal matching pursuit algorithm; (4) Fix the sparse matrix coefficient S, use the K-SVD algorithm to update the K atoms of the dictionary one by one, obtain the trained dictionary D, and perform sparse decomposition on the dictionary D to obtain the reconstructed unit; (5) splicing to obtain K-SVD reconstructed signal, performing fast Fourier transform on the K-SVD reconstructed signal, and analyzing the time domain and frequency domain signal diagrams to diagnose the health status of the gearbox; The digital high-pass filter adopts a FIR digital filter; The FIR digital filter is designed using the window function method, and the specific process is as follows: (21) Select the window function according to the minimum attenuation requirement of the stop band, and then determine the size of the FIR digital filter order N by the transition bandwidth to obtain the window function ω(n), n=0,1,2,…,N-1; (22) Calculate the cutoff frequency ω of the ideal high-pass filter c : Among them, ω s is the stopband cutoff frequency; ω p is the passband cutoff frequency; (23) Calculate the corresponding ideal filter unit impulse response h d (n): Where N is the order of the digital filter, which is an odd number; a is the delay constant, a = (N-1) / 2; h d (n) is even symmetric; (24) Obtain the unit impulse response of the designed FIR digital filter h(n) = h d (n)ω(n),n=0,1,2,…,N-1; (25) By finding H(e jω )=DTFT[h(n)] to verify whether the design result meets the design requirements. If not, redesign is required.
2. The vibration signal feature extraction method based on high-pass filtering according to claim 1 is characterized in that: The relationship between the original vibration signal x(n) and the filtered high-frequency vibration signal y(n) is a convolution sum relationship, that is: y(n)=x(n)*h(n).
3. The vibration signal feature extraction method based on high-pass filtering according to claim 1 is characterized in that: The specific solution process of the sparse matrix coefficient S in step (3) is: The filtered high-frequency vibration signal y(n) is divided into units to obtain a training sample library for the dictionary learning algorithm. K samples are selected from the sample library to form the initial dictionary D0, and the sparse matrix coefficient S is solved by the orthogonal matching pursuit (OMP) algorithm. The solution steps are as follows: Define the signal to be decomposed x, dictionary D, sparsity T, and number of loops t. When there is no iteration, that is, when t = 0, the residual R 0 = x, the index set of atoms Selected Atom Set Sparse coefficient s 0 =0; (31) Select the atom that best matches the current residual, calculate the dot product of the residual and all atoms, and the largest one is indexed as λ = argmax| <R t-1 ,d k >|; Among them, R t-1 is the residual obtained from the previous cycle; d k is a dictionary atom; (32) Calculate the atomic index set Λ at the cycle number t t (t): L t (t)=λ, Update the selected atom set D at the time of loop number t t : D t =[D t-1 ,d λ ]; (33) Calculate the sparse coefficient s at the number of cycles t t : (34) Calculate the residual R at the cycle number t t : R t =x-D t s t ; (35) Cycle number t+1; (36) Determine whether to stop the loop. If the number of loops t> the sparsity T, stop the loop. Otherwise, repeat steps (31) to (35). The number of loops t at this time is t+1 in step (35).
4. The vibration signal feature extraction method based on high-pass filtering according to claim 1 is characterized in that: The method for updating K atoms in step (4) is: Define the training sample library Y = [y1, y2, y3, ...y n ], the training sample library Y is a matrix with p rows and n columns; the dictionary D = [d1, d2, d3, ... d K ], the dictionary D is a matrix with p rows and K columns; the corresponding sparse coefficient matrix S = [s1, s2, s3, ...s n ], the sparse coefficient matrix S is a matrix with K rows and n columns, the maximum number of loops is M, and the initial number of loops is t = 1; (41) Determine the dictionary atom d of the kth atom k The sample set ω k : oh k ={j|1≤j≤n,s j,k ≠0}; Among them, j is a natural number, which only represents the count; (42) Calculate the error matrix E of the kth atom k : Where i is a natural number, 1≤i≤k; d i is the i-th atom in the dictionary; is the sparse coefficient matrix S and d i The row that multiplies; (43) Settings And Ω k Middle (ω k (j),j) elements are 1, and the other elements are 0; only the error matrix E is retained. k Non-zero value Keep the value after non-zero value Among them, Ω k is an n×|ω k |-order matrix, and all elements in the matrix are real numbers, which have no practical meaning. It is used to multiply other matrices and remove the common elements; is the sparse coefficient matrix S and the dictionary atom d k The row that multiplies; (44) Only the error matrix E is retained k Non-zero value Do SVD singular value decomposition: make make Among them, U and V are unitary matrices obtained by SVD singular value decomposition; Δ is a semi-positive definite diagonal matrix; V T is the transpose of V; Δ(1,1) represents the element corresponding to the first row and first column of the Δ matrix; is the first column of the unitary matrix U, which is the kth atom d in the dictionary k Updated results; U(:,1) represents the first column vector of the matrix; (45) The number of cycles t increases by one; (46) Check whether the loop should be terminated. If the loop number t is greater than the maximum loop number M, stop the loop; otherwise, repeat steps (42) to (45), where the loop number t is t in step (45) plus one.
5. The vibration signal feature extraction method based on high-pass filtering according to claim 1 is characterized in that: The diagnosis in step (5) is based on: If the frequency in the frequency domain signal diagram is f m ±af p ±bf c If an obvious peak appears at the position of the gear, it indicates that there is a local fault in the planetary gear teeth of the gearbox; Among them, a and b are natural numbers; f m is the meshing frequency of the planetary gear set; f p is the characteristic frequency of the planetary gear that is damaged; f c is the rotation frequency of the planet carrier itself.