Bearing fault diagnosis method and system

Abstract

The present invention discloses a bearing fault diagnosis method and system, belonging to the field of fault diagnosis technology. Based on a collected bearing fault signal, the bearing fault period is obtained, and an optimization objective function for sparse representation of the bearing fault signal is determined. The adaptive parameter in the optimization objective function is used as the optimization target, and a comprehensive index is constructed. With the goal of minimizing the comprehensive index, the optimal adaptive parameter is determined, and the optimal sparse representation optimization objective function is determined. The bearing fault signal is subjected to sparse denoising. Based on the sparse denoised bearing fault signal, a fitness function of filter length and displacement number is constructed using correlation kurtosis. The filter length and displacement number parameter combination with the maximum correlation kurtosis is obtained. The sparsely represented bearing fault signal is filtered to obtain a signal-enhanced periodic fault component and perform envelope spectrum analysis to obtain the characteristic frequency and frequency multiplication of the fault signal. This method can achieve accurate bearing fault diagnosis in strong noise conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and more particularly to a bearing fault diagnosis method and system. Background Art

[0002] Bearings are essential components of rotating machinery. They provide support, guidance, and friction reduction during rotor motion, ensuring the proper functioning of mechanical systems. However, prolonged operation, high loads, and harsh environmental conditions can lead to bearing failures such as wear, fatigue, and cracks. These failures not only cause equipment downtime and production interruptions but can also lead to serious safety incidents. Therefore, developing efficient and accurate bearing fault diagnosis technologies is a pressing issue in the industrial sector.

[0003] When a bearing fails, the vibration signal of a localized bearing fault contains a periodic pulse response or a quasi-periodic pulse response with small fluctuations due to the periodic impact between the rolling elements and the inner and outer rings of the bearing. This signal can reflect the bearing's operating condition and the possible fault type. Therefore, vibration analysis methods are widely used in the field of bearing fault diagnosis. However, the bearing vibration signals collected by the equipment are relatively complex and are often obscured by strong background noise and other interfering components, making the extraction of weak features difficult.

[0004] Currently, sparse representation is widely used in the field of signal processing. The core idea of ​​sparse representation is to use a given overcomplete learning dictionary to represent the signal as a linear combination of a small number of basis elements to achieve the representation of the original signal. After sparse transformation, the signal displays the sparse characteristics of the original signal while removing redundant information. This has excellent feature extraction effects on vibration signals generated by local bearing faults. However, in practical applications, traditional sparse representation is complex in dictionary design and learning, has high computational complexity, is susceptible to noise interference, and relies on the theoretical value of the bearing fault for the periodic prior. This makes it impossible to effectively extract the characteristics of the fault signal, which affects the extraction and diagnosis of bearing fault information in the presence of strong noise. Summary of the Invention

[0005] In response to the problems existing in the above-mentioned fields, the present invention proposes a bearing fault diagnosis method and system. According to the characteristics of the fault signal, the fault signal is sparsely denoised and the fault information of the denoised fault signal is enhanced. This method has good noise resistance and fault feature extraction capabilities, and can achieve accurate diagnosis of bearing faults under strong noise.

[0006] To solve the above technical problems, the present invention discloses a bearing fault diagnosis method, comprising the following steps:

[0007] Based on the collected bearing fault signal, the frequency spectrum information of the signal is obtained and the fault frequency is filtered out to determine the estimated period value of the signal and obtain the bearing fault period;

[0008] Based on the bearing's fault cycle, the optimization objective function for the sparse representation of the bearing fault signal is determined. The adaptive parameters in the optimization objective function are used as the optimization target, and a comprehensive index is constructed. With the goal of minimizing the comprehensive index, the optimal adaptive parameters are determined. Based on the optimal adaptive parameters, the optimal sparse representation optimization objective function is determined, and sparse denoising is performed on the bearing fault signal. The comprehensive index is the ratio of the error between the sparse denoised signal and the noise-free original signal to the envelope spectrum peak factor, which is used to characterize the relationship between the adaptive parameters and the noise level.

[0009] Based on the sparse denoised bearing fault signal, the fitness function of the filter length and displacement number is constructed through the correlation kurtosis. The filter length and displacement number parameter combination with the maximum correlation kurtosis is obtained. The sparsely represented bearing fault signal is filtered to obtain the signal-enhanced periodic fault component.

[0010] The envelope spectrum analysis of the signal-enhanced periodic fault component is performed to obtain the characteristic frequency and frequency multiples of the fault signal.

[0011] Preferably, obtaining the failure cycle of the bearing specifically includes:

[0012] Signal Doing first-order difference operation yields , calculate the signal through Hilbert transform The analytical signal :

[0013] ;

[0014] right Perform FFT to obtain spectrum information. The obtained frequency domain information is bilaterally symmetrical. Only the right half is selected to obtain the frequency index after sorting the spectrum energy information in descending order.

[0015] According to the sorted index, select the top 20 largest frequencies as the active frequency set :

[0016] ;

[0017] in, f s is the sampling frequency, f c is the fault frequency, n is the index of the fault frequency set;

[0018] Get the active frequency set The frequency interval is less than The frequency individuals are removed and the current active frequency set is updated ;

[0019] The current set of active frequencies for updates Perform a cyclic search for each element in the , calculate the multiple relationship between them, and select the active frequency set Filter out active frequencies with multiple relationships;

[0020] Take a weighted average of the active frequencies that have a multiple relationship and calculate the failure frequency with the highest probability of occurrence ;

[0021] According to the fault frequency , determine the estimated period value of the signal P .

[0022] Preferably, determining the optimization objective function of the sparse representation of the bearing fault signal comprises the following steps:

[0023] The mathematical expression for obtaining the sparse representation of the signal without a dictionary is:

[0024] ;

[0025] in, is the original signal without noise, is the sparse denoised signal, is an adaptive parameter, P ( x ) is the penalty function;

[0026] Get a binary weight vector :

[0027] ;

[0028] in, is the estimated length of a fault impulse, is the interval length between adjacent fault impulses, is the group size, is the number of complete fault shocks contained in each group, is the length occupied by each complete fault impulse, that is, the estimated period value of the signal;

[0029] According to the sampling frequency and failure frequency ,parameter P Satisfies the equation:

[0030] ;

[0031] According to the binary weight vector , penalty function And the mathematical expression of sparse representation without dictionary, the optimization objective function of sparse representation of bearing fault signal is obtained as follows:

[0032] ;

[0033] in, i is the group index coefficient, For collection index, is the sparsity penalty function, , is a constant value that controls the degree of non-convexity of the penalty, is the balance parameter, To reweight l 1 penalty;

[0034] ,

[0035] ;

[0036] in, is a diagonal matrix with diagonal elements containing , Defined as:

[0037] ;

[0038] in, is a small positive constant used to avoid division by zero;

[0039] when and hour, It is convex.

[0040] Preferably, determining the optimal adaptive parameters comprises the following steps:

[0041] Adaptive parameters in the optimization objective function of sparse representation of bearing fault signals , as the optimization target;

[0042] Based on the adaptive parameters , construct a comprehensive index , used to characterize the adaptive parameters Relationship with noise level;

[0043] The more periodic shocks there are in the vibration signal, the larger the signal envelope spectrum will be. The larger the peak value at the frequency of the impact, the larger the peak value at the frequency of the impact. The ratio of the maximum value to the effective value within the range is used as the envelope spectrum crest factor , expressed as:

[0044] ;

[0045] in, For frequency conversion;

[0046] The larger the value, the higher the signal-to-noise ratio and the greater the energy of the signal, and the stronger the periodic impact component. The envelope spectrum crest factor is used as a measure of the impact component in the signal.

[0047] The error between the sparse denoised signal and the original noise-free signal is:

[0048] ;

[0049] RMSE The smaller the value, the better the fit between the denoised signal and the original noise-free signal, indicating that the model is more accurate in estimation;

[0050] Will and RMSE Place them in the denominator and numerator of the fraction respectively to construct a comprehensive index :

[0051] ;

[0052] Comprehensive indicators Used to simultaneously reflect the denoising ability and estimation accuracy of the model;

[0053] Comprehensive indicators The smaller the value, the closer the signal is to the original fault pulse after denoising. When the value reaches the minimum, it means that the adaptive parameter Optimal;

[0054] Using comprehensive indicators Minimize and determine the optimal adaptive parameters The optimal value at different noise levels is determined by linear fitting curves. ,get The relationship between the noise level.

[0055] Preferably, the sparse denoising of the bearing fault signal specifically includes:

[0056] The optimal adaptive parameters Substitute the optimization objective function of the sparse representation of the bearing fault signal into the optimal sparse representation optimization objective function;

[0057] The iterative optimization process of the MM algorithm is used to solve the optimal sparse representation optimization objective function and obtain the bearing fault signal after sparse denoising.

[0058] Preferably, the step of obtaining the enhanced periodic fault component of the signal comprises the following steps:

[0059] Based on the maximum correlation kurtosis value of the bearing fault signal after sparse denoising, the fault signal is filtered and the fault features are extracted.

[0060] The essence of the MCKD algorithm is to find a filter FIR to restore the continuous impact signal submerged by noise; filter the signal through the correlation kurtosis. When it is maximum, the signal recovered by the solved filter satisfies the periodic impulse characteristic;

[0061] Correlation kurtosis The expression is:

[0062] ;

[0063] The condition that maximizes the correlation kurtosis is:

[0064] ;

[0065] Calculate the filter coefficients:

[0066] ;

[0067] in, is the deconvolution period, Shift number, m ∈[0, M ], is the filter length;

[0068] The coefficients in the formula are:

[0069] ;

[0070] ;

[0071] ;

[0072] ;

[0073] For the filter length L and displacement number M , construct the fitness function of filter length and displacement number through the relevant kurtosis;

[0074] By using a grid search method, a combination of filter length and displacement number parameters that maximizes the correlation kurtosis is searched to obtain an optimized filter length, displacement number, and deconvolution period, which are periodic fault components for signal enhancement.

[0075] Preferably, the step of obtaining the optimized filter length, shift number and deconvolution period comprises:

[0076] Determining the filter length L and displacement number M The search range and search step size;

[0077] Initialize parameter filter length according to search range L and displacement number M , the deconvolution period T is assigned according to the period estimation result;

[0078] According to the initialized parameter filter length L , displacement number M and the deconvolution period assigned according to the period estimation result T , calculate the corresponding correlation kurtosis as the correlation kurtosis A ;

[0079] Update the parameters according to the search step size, and calculate the relevant kurtosis corresponding to the updated parameters as the relevant kurtosis B ;

[0080] The kurtosis B Related kurtosis A For comparison; when the kurtosis B Less than the relevant kurtosis A When the relevant kurtosis is retained B ; When the relevant kurtosis B Greater than the relevant kurtosis A , preserving the relevant kurtosis A ;

[0081] After the search is completed, the parameter corresponding to the final retained correlation kurtosis is output as the optimal filter length L , displacement number M and deconvolution cycle T .

[0082] Preferably, the step of obtaining the characteristic frequency and frequency multiplication of the fault signal specifically includes:

[0083] According to the optimized filter length, displacement number and deconvolution period, the fault feature of the bearing fault signal after sparse denoising is enhanced by the MCKD algorithm. The enhanced result is subjected to envelope spectrum analysis to obtain the extraction and diagnosis results of the characteristic frequency and frequency multiplication of the bearing fault under strong noise.

[0084] Preferably, a bearing fault diagnosis system is also included, comprising:

[0085] The fault signal cycle acquisition module is used to obtain the frequency spectrum information of the collected bearing fault signal and filter out the fault frequency, thereby determining the estimated cycle value of the signal and obtaining the bearing fault cycle;

[0086] The fault signal sparse denoising module is used to determine the optimization objective function for the sparse representation of the bearing fault signal based on the bearing's fault cycle; the adaptive parameters in the optimization objective function are used as the optimization target, a comprehensive index is constructed, and the optimal adaptive parameters are determined with the goal of minimizing the comprehensive index; based on the optimal adaptive parameters, the optimal sparse representation optimization objective function is determined to perform sparse denoising on the bearing fault signal; the comprehensive index is the ratio of the error between the sparse denoised signal and the noise-free original signal to the envelope spectrum peak factor, which is used to characterize the relationship between the adaptive parameters and the noise level;

[0087] The fault signal enhancement module is used to construct a fitness function of filter length and displacement number based on the sparse denoised bearing fault signal through the correlation kurtosis, obtain the filter length and displacement number parameter combination with the maximum correlation kurtosis, filter the sparsely represented bearing fault signal, and obtain the signal-enhanced periodic fault component;

[0088] The fault signal diagnosis module is used to perform envelope spectrum analysis on the signal-enhanced periodic fault component to obtain the characteristic frequency and frequency multiples of the fault signal.

[0089] Compared with the prior art, the present invention has the following beneficial effects:

[0090] The present invention proposes a bearing fault diagnosis method that constructs an optimization objective function for a sparse representation of the bearing fault signal, uses the adaptive parameters in the optimization objective function as optimization targets, constructs a comprehensive index, and determines the optimal adaptive parameters with the goal of minimizing the comprehensive index. The method then performs sparse denoising on the bearing fault signal. Based on the characteristics of the fault signal, without requiring any information about the faulted bearing, the method dynamically estimates the possible fault cycles in the fault signal and constructs a binary vector embedded in a penalty function to promote intra-group and inter-group sparsity of the fault signal, thereby achieving fault signal denoising and obtaining a sparsely represented denoised fault signal. Based on the sparsely represented denoised bearing fault signal, a fitness function for filter length and displacement number is constructed using correlation kurtosis. The filter length and displacement number parameter combination with the maximum correlation kurtosis is obtained, and the sparsely represented bearing fault signal is filtered to obtain a signal-enhanced periodic fault component. Fault information is enhanced on the sparsely represented bearing fault signal, effectively combining the advantages of pulse extraction and pulse feature enhancement of the adaptive sparse periodic group lasso, resulting in excellent noise immunity and fault feature extraction capabilities for vibration signals. This method has good noise resistance and fault feature extraction capabilities, and can achieve accurate diagnosis of bearing faults under strong noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 This is a flow chart of the bearing fault diagnosis method AdaSRMD proposed in the present invention;

[0092] Figure 2 Several traditional penalty functions provided by the embodiment of the present invention are L 0 the relationship between norms;

[0093] Figure 3 The embodiment of the present invention provides and noise level The relationship diagram between

[0094] Figure 4 The fault pulse component provided by the embodiment of the present invention;

[0095] Figure 5 The period estimation results under different noise levels provided by the embodiment of the present invention;

[0096] Figure 6 Comparison of fault pulse extraction performance of analog signals provided by embodiments of the present invention;

[0097] Figure 7 Comparison of simulation results under different noise levels provided by the embodiment of the present invention;

[0098] Figure 8 The processing result of the bearing inner ring fault provided by the embodiment of the present invention;

[0099] Figure 9 The results of processing bearing inner race faults using AdaSRMD and AdaSR respectively provided in the embodiments of the present invention;

[0100] Figure 10 The processing result of the bearing outer ring fault provided by the embodiment of the present invention;

[0101] Figure 11 The extraction result of the bearing inner ring fault provided by the embodiment of the present invention;

[0102] Figure 12 The update curve of the failure cycle versus the number of iterations provided by the embodiment of the present invention;

[0103] Figure 13 This is the extraction result of the bearing outer ring fault provided by the embodiment of the present invention. DETAILED DESCRIPTION

[0104] The following is a combination of the embodiments of the present invention Figures 1-13 , the technical solutions in the embodiments of the present invention are clearly and completely described. It should be understood that the terms used in the present invention are only used to describe specific implementation methods and are not intended to limit the present invention.

[0105] like Figure 1 As shown, the present invention proposes a bearing fault diagnosis method, comprising the following steps:

[0106] Determine the sparse denoising model

[0107] Define a binary weight vector :

[0108] ;

[0109] in, is the estimated length of a fault impulse, is the interval length between adjacent fault impulses, is the group size, is the number of complete fault shocks contained in each group, is the length taken by each complete fault impulse, that is, the estimated period value of the signal.

[0110] According to the sampling frequency and fault characteristic frequency ,parameter P Satisfies the equation:

[0111] ;

[0112] According to the binary weight vector , penalty function And the mathematical expression of sparse representation without dictionary, it can be concluded that the mathematical expression of the optimization problem, that is, the optimization objective function of the sparse representation of bearing fault signals, is:

[0113] ;

[0114] in, i is the group index coefficient, For collection index, is the sparsity penalty function, , is a constant value that controls the degree of non-convexity of the penalty, is the balance parameter, To reweight l 1 penalty.

[0115] ,

[0116] ;

[0117] in, is a diagonal matrix with diagonal elements containing , Defined as:

[0118] ;

[0119] in, is a small positive constant used to avoid division by zero.

[0120] when and hour, It is convex, so the optimization problem can be solved by MM algorithm.

[0121] The core idea of ​​the MM algorithm is to gradually approximate the original non-convex optimization objective function by constructing a simple convex function (called the master function) and finding the minimum point of the master function at each step. The optimization iteration formula is derived as follows:

[0122] ;

[0123] ;

[0124] ;

[0125] in, , is the number of algorithm iterations.

[0126] The penalty function is based on the sparsity between groups (SWAG) of the fault signal containing periodic pulses.

[0127] A penalty function ESGL proposed in the prior art that is both non-convex and reweighted divides the periodic fault signal into groups, assuming that the size of each group has the same value , and the maximum overlap is chosen to preserve the most information. , if the length of the signal is , the size of each group is .

[0128] Define each group as:

[0129] ;

[0130] in, is the group index coefficient.

[0131] Non-convex and reweighted penalty terms to enhance SWAG properties The expression is as follows:

[0132] ;

[0133] in, is a balancing parameter that controls the sparsity within each group.

[0134] Table 1 gives the penalty functions for some attributes ,like Figure 2 As shown, it is the visualization corresponding to each penalty function. is the input of punishment, is a constant value that controls the degree of non-convexity of the penalty, is the function used in the derivation of the MM algorithm. Compared to punish, 、 The penalty is more inclined to promote sparsity while effectively preserving the magnitude.

[0135] Table 1 Sparsity penalty function

[0136]

[0137] also, Reweighting The penalty is defined as follows:

[0138] ;

[0139] in, is a diagonal matrix with diagonal elements containing , Defined as:

[0140] ;

[0141] The mathematical expression using sparse representation without dictionary is as follows:

[0142] ;

[0143] in, is the original signal, is a sparse signal, is the regularization parameter, P ( x ) is the penalty function.

[0144] The expression consists of two parts: the fidelity term and the regularization term. The fidelity term retains the original information of the signal, and the regularization term makes the signal sparser by penalizing the fault features. The classic dictionary-free SR model uses The norm is used as a penalty function to promote the sparsity of the signal, but The norm has the disadvantage of underestimating the magnitude. The magnitude underestimation defect of the norm can be overcome by using a non-convex penalty function.

[0145] The basic idea of ​​the Maximum Correlation Kurtosis Deconvolution (MCKD) method is to find a finite impulse response filter (FIR) to restore the continuous impulse signal submerged by noise; filter the signal through the correlation kurtosis. When the maximum value is reached, the signal recovered by the filter satisfies the periodic impact characteristics, and the signal is filtered to extract the fault characteristics.

[0146] Among them, the relevant kurtosis The expression is as follows:

[0147] ;

[0148] The condition that maximizes the correlation kurtosis is:

[0149] ;

[0150] Calculate the filter coefficients:

[0151] ;

[0152] in, is the deconvolution period, Shift number, m ∈[0, M ], is the filter length.

[0153] The coefficients in the above formula are:

[0154] ;

[0155] ;

[0156] ;

[0157] .

[0158] The present invention proposes a bearing fault diagnosis method based on adaptive sparse periodic group lasso and maximum correlation kurtosis deconvolution, namely a bearing fault diagnosis method based on (Adaptive Sparse periodic group lasso and Maximum correlation kurtosis Deconvolution, AdaSRMD)

[0159] MCKD uses deconvolution to highlight continuous pulses buried in noise, improving the kurtosis of the original signal. However, when the noise is high, the noise component of the signal has a greater kurtosis than the fault-related pulse component, resulting in poor signal filtering. Previous adaptive sparse periodic group lasso models relied on theoretical values ​​of bearing faults for their periodicity priors and also had limitations in maintaining the amplitude of the processed signal.

[0160] In order to solve these problems, the present invention proposes a bearing fault diagnosis method based on AdaSRMD, the flow chart of which is as follows: Figure 1 As shown. Based on the characteristics of bearing fault signals, this method studies a new method for estimating bearing fault cycles without any fault bearing information, and estimates the possible cycles in the signal; the period sequence Dynamically embedding the fault signal into a non-convex penalty function for iterative processing reveals the sparsity of the fault signal within and between groups. A sparse representation regularization parameter adaptive optimization method is studied, and an effective denoising method for bearing fault signals is proposed. The key parameters of the maximum correlation kurtosis deconvolution based on grid search are studied. The best optimization method is proposed, and a bearing periodic fault component enhancement method is proposed to achieve accurate diagnosis of bearing faults under strong noise.

[0161] Among them, the adaptive period estimation method (AdaFP)

[0162] In actual work, when the bearing parameters are unknown, it is impossible to obtain the theoretical fault frequency of each part, which makes it difficult to extract and diagnose the fault. Therefore, this paper proposes a method for estimating the fault period, which is of certain significance for finding the fault frequency.

[0163] Since the first-order difference can highlight the changes or trends in the signal, it is Doing first-order difference operation yields , calculate the signal through Hilbert transform The analytical signal :

[0164] ;

[0165] right Perform FFT to obtain spectrum information. The frequency domain information obtained is bilaterally symmetrical. Subsequent calculations select only the right half of the spectrum. Obtain the frequency index after sorting the spectrum energy information in descending order. Select the top 20 largest frequencies as the active frequency set based on the sorted index. :

[0166] ;

[0167] In order to avoid interference, The frequency interval is less than The frequency individuals are removed and the current active frequency set is updated .

[0168] The current set of active frequencies for updates Perform a cyclic search for each element in the , calculate the multiple relationship between them, and select the active frequency set Filter out active frequencies with multiple relationships.

[0169] Take a weighted average of the active frequencies that have a multiple relationship and calculate the failure frequency with the highest probability of occurrence ; According to the fault frequency , determine the estimated period value of the signal P .

[0170] Fault signal denoising method based on optimized sparse representation

[0171] Sparse representation mainly involves several parameters For the parameter Mainly to construct binary weight vector To enhance the fault information. Therefore, the present invention focuses on the period of each pulse rather than the specific duration. Based on the experience set to , used to roughly estimate the length of each pulse. Parameters Indicates the number of cycles in the sequence. For fast implementation, set .according to Figure 2 You can see for The effect of promoting sparsity while maintaining amplitude is closest to Norm. At the same time, in order to best promote this property, set . Existing technology has been verified affects the sparsity within each group, regardless of the noise level. Therefore, we set . Number of algorithm iterations The solution of the optimization problem has little impact, and the early stopping strategy is used to set For the adaptive regularization parameter , also known as adaptive parameters, needs to be further discussed based on the signal being under different noise levels.

[0172] By constructing a Comprehensive indicator of relationship with noise level To evaluate the performance of the algorithm.

[0173] The more periodic shocks there are in the vibration signal, the larger the signal envelope spectrum will be. The larger the peak value at the impact frequency, the smaller the envelope spectrum of the filtered signal will be. The ratio of the maximum value to the effective value within the range is called the envelope spectrum crest factor Expressed as:

[0174] ;

[0175] in, For frequency conversion, n The value 2 can be used to eliminate the frequency conversion The impact of value.

[0176] The larger the value, the higher the signal-to-noise ratio and the greater the energy, and the stronger the periodic impulse component. Therefore, the envelope spectrum crest factor can be used as a measure of the impulse component in the signal.

[0177] The error between the sparse denoised signal and the original noise-free signal is:

[0178] ;

[0179] in, represents the denoised signal, Represents a noise-free analog signal.

[0180] RMSEThe smaller the value of , the better the fit between the sparse denoised signal and the noise-free original signal, indicating that the model is more accurate in estimation.

[0181] because The larger the value, the better. RMSE The smaller the value, the better, so and RMSE Place them in the denominator and numerator of the fraction to form a comprehensive indicator :

[0182] ;

[0183] Comprehensive indicators It can simultaneously reflect the denoising ability and estimation accuracy of the model.

[0184] Comprehensive indicators The smaller the index value is, the closer the signal is to the original fault pulse after denoising. When the value reaches the minimum, it means that the parameter Therefore, the comprehensive index can be used Choose based on the principle of minimization .

[0185] To explore the best and noise levels A set of simulations were performed using simulated signals:

[0186] ;

[0187] in, median (·) means taking the median, are the orthogonal wavelet (sym8) coefficients of the optimal scale.

[0188] use Minimize to determine the optimal Parameters, perform 100 independent experiments at different noise levels, calculate the average of the 100 experimental results, and plot these results as shown in Figure 3 shown.

[0189] from Figure 3 It can be clearly observed that the error between the experimental average and the individual experiments is very small. In addition, the difference between the average and the value obtained by fitting is also negligible. Therefore, the linear fitting curve can be used to determine the optimal , the specific formula is as follows:

[0190] .

[0191] Periodic fault enhancement method based on optimized MCKD algorithm

[0192] To solve the problem that the amplitude of the fault signal after adaptive sparse periodic group lasso processing still has certain defects, a grid search-based maximum correlation kurtosis deconvolution key parameter combination optimization method is proposed to achieve effective enhancement and accurate diagnosis of bearing fault information.

[0193] The MCKD algorithm contains the deconvolution cycle , filter length , displacement number The strict requirements of the important parameters of the MCKD algorithm limit the performance improvement of the MCKD algorithm in extracting transient characteristics from noisy signals and identifying mechanical faults. Only by ensuring the accuracy and proper use of these parameters can the superiority of the MCKD algorithm be highlighted. A more accurate value can be obtained by using the AdaFP method.

[0194] for Parameters, the present invention first uses the correlation kurtosis to construct the fitness function of the filter length and the number of displacements; then, the grid search method is used to search for the maximum correlation kurtosis. The specific implementation steps of parameter combination and parameter optimization are as follows:

[0195] Step 1: Determine the filter length and displacement number The search range and search step size;

[0196] Step 2: Initialize parameters based on search scope 、 , for the deconvolution cycle Then the value can be assigned according to the period estimation result;

[0197] Step 3: Filter length according to the initialized parameters L , displacement number M and the deconvolution period assigned according to the period estimation result T , calculate the corresponding correlation kurtosis as the correlation kurtosis A ;

[0198] Step 4: Update the parameters according to the search step size and calculate the relevant kurtosis corresponding to the updated parameters as the relevant kurtosis B ; The relevant kurtosis B Related kurtosis A For comparison; when the kurtosis B Less than the relevant kurtosis A When the relevant kurtosis is retained B ; When the relevant kurtosis B Greater than the relevant kurtosisA , preserving the relevant kurtosis A ;

[0199] Step 5: After the search is completed, the parameter corresponding to the final retained correlation kurtosis is output as the optimal filter length L , displacement number M and deconvolution cycle T .

[0200] The optimal filter length obtained by optimizing the above parameters is L , displacement number M and deconvolution cycle T , as the periodic fault component of signal enhancement.

[0201] The optimized parameters of the fault signal after sparse denoising are obtained through the above steps. and Finally, the fault signal is processed by the MCKD algorithm to enhance the fault characteristics. The final result is then subjected to envelope spectrum analysis to extract and diagnose bearing fault information under strong noise.

[0202] The performance of the proposed fault diagnosis model is evaluated using simulated fault signals. The simulated fault signals consist of pulse components and noise caused by rolling bearing faults:

[0203] .

[0204] The vibration signal caused by rolling bearing defects x ( i ) is simulated as a series of pulse transients, each transient consists of 10 samples, the pulse signal z ( t )for:

[0205] ;

[0206] in, , is noise, yes A random integer in , t is the time variable, is the sampling interval, is random slip, is the amplitude of the pulse signal is the natural frequency of the pulse signal, is the phase of the pulse signal, 、 and They obey the normal distribution of N(1, 1), N(2000, 10) and N(0, 1), respectively, and the impulse signal Combination composition Figure 4 The sampling rate in (a) is Quasi-periodic fault sequence , the sampling time is , fault frequency In order to simulate the period uncertainty of real faults, the fault cycle is set to have a random variation of 0.8%, and the previous Set to zero to simulate a healthy bearing state. Figure 4 (b) Adding Gaussian noise ( ) fault signal.

[0207] AdaFP cycle estimation effect

[0208] In order to verify the effect of period estimation, eliminate the period random sliding phenomenon, and compare with the method proposed in the prior art, 100 independent experiments were conducted under different noise levels. The average value of the results of these 100 experiments was calculated as a reference, as shown in Figure 5 shown.

[0209] from Figure 5 It can be seen that in When , both methods can estimate the accurate results p = 80 is the same as the theoretical value, but as the noise level increases, the (Adaptive Sparse Periodic Group Lasso, AdaESPGL) estimation method produces the largest estimate p = 82 and the minimum estimate p = 74, with an error percentage of 8.8%. The maximum estimated value of AdaFP proposed by the present invention is p = 81, the minimum estimate p = 79.8, and the error percentage is 1.3%, indicating that the period estimation method proposed in the present invention has better estimation accuracy.

[0210] In addition, it is worth noting that the AdaESPGL period estimation method requires the precalculation of the theoretical bearing fault frequency before the period can be estimated. The proposed AdaFP method can estimate the possible fault period without prior information, which is of certain significance in practical engineering.

[0211] The present invention also proposes a bearing fault diagnosis system, comprising:

[0212] The fault signal cycle acquisition module is used to obtain the frequency spectrum information of the collected bearing fault signal and filter out the fault frequency, thereby determining the estimated cycle value of the signal and obtaining the bearing fault cycle;

[0213] The fault signal sparse denoising module is used to determine the optimization objective function for the sparse representation of the bearing fault signal based on the bearing's fault cycle; the adaptive parameters in the optimization objective function are used as the optimization target, a comprehensive index is constructed, and the optimal adaptive parameters are determined with the goal of minimizing the comprehensive index; based on the optimal adaptive parameters, the optimal sparse representation optimization objective function is determined to perform sparse denoising on the bearing fault signal; the comprehensive index is the ratio of the error between the sparse denoised signal and the noise-free original signal to the envelope spectrum peak factor, which is used to characterize the relationship between the adaptive parameters and the noise level;

[0214] The fault signal enhancement module is used to construct a fitness function of filter length and displacement number based on the sparse denoised bearing fault signal through the correlation kurtosis, obtain the filter length and displacement number parameter combination with the maximum correlation kurtosis, filter the sparsely represented bearing fault signal, and obtain the signal-enhanced periodic fault component;

[0215] The fault signal diagnosis module is used to perform envelope spectrum analysis on the signal-enhanced periodic fault component to obtain the characteristic frequency and frequency multiples of the fault signal.

[0216] The present invention proposes a bearing fault diagnosis method, which can realize accurate diagnosis of bearing faults under strong noise.

[0217] Simulation signal results

[0218] Comparison results of different methods

[0219] In order to verify the effectiveness of the method proposed in this invention, the method proposed in this invention was compared with three methods: (Periodic Group-Sparsity Learning, PGSL), (Basis Pursuit Denoising, BPD) and AdaESPGL.

[0220] Among them, the PGSL method belongs to a sparse Bayesian learning framework, so no parameter adjustment is required. k The sparse parameter is set to true. Key parameter choices for AdaESPGL , in order to observe the contrast effect, set the same number of iterations The relevant parameter settings in AdaSRMD are shown in Table 2. The results of the above method on the above simulation signal processing are as follows Figure 6 shown.

[0221] Table 2 AdaSRMD algorithm parameter settings

[0222]

[0223] from Figure 6As can be seen in (a), despite the complete knowledge of k-sparsity, the fault pulse extraction performance of BPD is still the worst. Figure 6 (b) It is observed that although PGSL has a certain accuracy in fault pulse extraction, it makes misjudgments in fault extraction within the first 0.4s. Figure 6 As shown in (c), AdaESPGL did not make any misjudgment within 0.4s, but its performance in maintaining the fault pulse amplitude was poor. Figure 6 As can be seen in (d), the proposed method performs well in both the accuracy of fault pulse extraction and the ability to maintain the fault pulse amplitude.

[0224] Robustness evaluation of AdaSRMD algorithm

[0225] The estimation accuracy is evaluated by RMSE index, and the envelope spectrum peak factor is used The fault feature detection capability is evaluated by the indicator. 100 independent simulation experiments are conducted at each noise level, and the average value is calculated as the analysis result. The results are compared with the PGSL, BPD and AdaESPGL methods. Figure 7 shown.

[0226] from Figure 7 As can be seen in (a), the RMSE values ​​obtained by the four methods all increase with the increase of noise level. However, the proposed method is better than the other methods in restoring the fault signal amplitude. AdaESPGL has little difference in maintaining amplitude compared with the proposed method when the noise level is low, but as the noise increases, the proposed method has a significantly better ability to maintain amplitude after MCKD fault enhancement. Figure 7 (b) The envelope spectrum peak factors of AdaESPGL and BPD methods can be seen It gradually decreases with the increase of noise level, indicating that the anti-interference ability of the two methods is weak.

[0227] PGSL The fluctuation is not large with the increase of noise level, which proves that the PGSL method has certain robustness. However, compared with AdaSRMD, it can be obviously seen that the proposed method has better fault extraction ability.

[0228] In summary, it is proved that AdaSRMD has certain advantages over PGSL, BPD and AdaESPGL.

[0229] Case Analysis

[0230] In order to verify the feasibility of the method proposed in the present invention, real bearing fault signals are used to verify the effectiveness of the method proposed in the present invention. The data sets are respectively from the bearing fault data set of Western Reserve University and the bearing fault data collected by individuals.

[0231] Setting up the dataset experiment

[0232] Data Source

[0233] The rolling bearing failure simulation test bench of Case Western Reserve University in the United States is mainly composed of components such as motors, torque sensors, power testers and electronic controllers.

[0234] Experimental environment

[0235] The experimental bearings were electrospark machined to create single-point damage to simulate fault conditions on the inner and outer races and rolling elements. The damage used in this experiment had a diameter of 0.007 inches. The machined bearings were then installed in a test motor, and vibration acceleration signals were recorded while the motor operated under various loads. The signals were acquired using a 16-channel data logger with a sampling frequency of 12 kHz and a motor speed of 1797 rpm.

[0236] Based on various bearing parameters, the characteristic frequency of the inner race fault in the Western Reserve University bearing failure experiment was calculated to be 162.2 Hz, and the characteristic frequency of the outer race fault was 107.3 Hz. Since the k-sparseness parameter of the BPD is unknown, it was set to 10% of the total coefficients. The remaining parameters were the same as those in the simulation experiment in Section 4. In this study, the system was equipped with an NVIDIA GeForce RTX 4060 and a 3.4 GHz Gen Intel Core i7-13700K processor, and the computation software was MATLAB 2024a.

[0237] Experimental results analysis

[0238] like Figure 8 Figure 2 shows the extraction results and envelope spectra of different strategies when a bearing inner race fault occurs. To reflect the proportion of fault frequency and frequency multiples, the amplitude of the envelope spectrum is normalized.

[0239] from Figure 8 (a) and Figure 8 (b) It is difficult to find effective fault information in the original vibration signal (the fault frequency is interfered by the sideband). Figure 8 In (c), we can see that the time domain signal after BPD processing has poor performance in fault information extraction and amplitude preservation. Figure 8 As can be seen from (d), the BPD algorithm fails to effectively separate the fault frequency from the sidebands it generates because it does not fully utilize the group sparsity and periodic structure of the fault pulse. Figure 8(e) Figure 8 (f) and Figure 8 (g) Figure 8 (h) shows the results of the PGSL algorithm and the AdaESPGL algorithm extraction. It can be seen that both methods have good results in extracting fault information. The fault frequency and its multiples can also be observed in the frequency domain. However, both methods have different degrees of attenuation and loss in the extraction of time domain fault information, and the fault information is not retained to the greatest extent. Figure 8 (i) and Figure 8 (j) shows the proposed method. In the time domain, AdaSRMD fully extracts the fault pulse component. In the frequency domain, compared to the extraction results of the PGSL and AdaESPGL algorithms, AdaSRMD also highlights the fault information at and , demonstrating the superiority of the proposed method.

[0240] In order to demonstrate the superiority of AdaSRMD, AdaSRMD and AdaSR are used to handle the bearing inner ring fault. Figure 9 (a) and Figure 9 (b) It can be observed that the result after sparse representation processing has the ability to extract fault pulses, but the restoration of fault information and the amplitude preservation effect are not as good as those of Figure 9 (c) and Figure 9 The effects after AdaSRMD processing in (d) are quite different.

[0241] like Figure 10 As shown in Figure 1, the extraction results and envelope spectra of different strategies when the bearing outer ring fails are shown. In order to reflect the strong noise background, Gaussian white noise ( ).

[0242] Figure 10 (a) and Figure 10 (b) is the original vibration signal when the bearing outer ring fails. Figure 10 (c) and Figure 10 As can be seen in (d), BPD suffers from significant performance degradation because the sparse structure of the pulses is almost masked by the strong background noise. Figure 10 (e) and Figure 10 (f) shows the results after PGSL algorithm extraction. Due to the robustness of PGSL, the extraction of fault information is relatively complete. However, compared with the proposed method, the AdaSRMD algorithm is 、 、 、 There are different degrees of improvement. Figure 10 (g) and Figure 10(h) is the extraction result of AdaESPGL. Due to the strong noise background, there are different degrees of attenuation and loss at the fault frequency and its multiples, and the fault information is not retained to the greatest extent. Figure 10 (i) and Figure 10 (j) is the method proposed by the present invention. It can be seen that when a fault occurs on the bearing outer ring, AdaSRMD completely extracts the fault pulse component in the time domain.

[0243] TYS1-8 bearing vibration dataset

[0244] Data Source

[0245] This experiment uses a private dataset to verify the effectiveness of the proposed method. In the experiment, a Liens B06A00 uniaxial acceleration sensor is installed on the vertical axis of the test bearing seat. The sensor is installed in a magnetic way.

[0246] Experimental environment

[0247] To simulate the actual operating environment of the bearing, a radial load of 2.5 kN and an axial load of 5 kN were applied to the test bearing seat. The data acquisition card used was a PXIe-69529 dynamic signal acquisition module with a sampling frequency of 12 kHz. The experimental faulty bearing was an SKF 6206-2Z rolling bearing. The bearing parameters are shown in Table 3.

[0248] Table 3 Main parameters of bearings

[0249]

[0250] The inner and outer rings of the bearing are artificially processed with crack faults, and the fault size is 1.0×1.0mm. According to the theoretical fault frequency calculation formula, the characteristic frequency of the bearing inner ring fault can be obtained as , the characteristic frequency of the outer race fault is Hz.

[0251] Experimental results analysis

[0252] like Figure 11 As shown in the figure, the extraction results and envelope spectra of different strategies when the bearing inner ring fails based on the TYS1-8 bearing vibration dataset are shown. Figure 11 (a) and Figure 11 (b) is the original vibration signal when the bearing inner ring fails. Figure 11 It can be seen that the fault pulse signal of the inner race fault is masked by the strong background noise. In this case, Figure 11 (c) and Figure 11 (d) It can be seen that the BPD algorithm always has the worst extraction effect. Figure 11 (e) and Figure 11(f) It can be seen that the PGSL algorithm performs better in extracting faults, but compared with the proposed algorithm, there are gaps in both time and frequency domains. Figure 11 (g) and Figure 11 (h) It can be seen that AdaESPGL captures some important pulses, but the key periodic pulses corresponding to the bearing inner race fault are ignored. This is due to the error between the fault period value obtained by the AdaESPGL method and the theoretical value. Figure 11 (i) and Figure 11 (j) is the method proposed by the present invention. It can be seen that when a fault occurs in the bearing inner ring, AdaSRMD can completely extract the fault pulse component in the time domain.

[0253] like Figure 12 As shown, the update curve of the period estimation with the number of iterations is displayed. The method proposed in the present invention, namely the AdaFP period estimation method, has an estimation error close to 2, while the error between the convergence value of the AdaESPGL method and the theoretical value reaches 6, which further proves the superiority of the method proposed in the present invention.

[0254] like Figure 13 As shown in the figure, the extraction results and envelope spectra of different strategies when the bearing outer ring fails based on the TYS1-8 bearing vibration dataset are shown. Figure 13 (a) and Figure 13 (b) is the original vibration signal when the bearing outer ring fails. Figure 13 (c) and Figure 13 (d) It can be seen that BPD's ability to extract faults in the presence of strong noise is still the worst. Figure 13 (g) and Figure 13 (h) It can be seen that the AdaESPGL algorithm has poor denoising performance in the time domain and can only extract 、 . Figure 13 (e) and Figure 13 (f) PGSL can achieve good performance in this case. However, it is worth noting that the AdaSRMD method can provide the fault frequency more clearly. Figure 13 (i) and Figure 13 (j) is the method proposed by the present invention. It can be seen that when a fault occurs on the bearing outer ring, AdaSRMD completely extracts the fault pulse component in the time domain.

[0255] Table 4 Comparison results of different methods

[0256]

[0257] Table 4 shows the results of different methods for the TYS1-8 platform dataset. BPD fails to extract inner-ring faults and performs poorly for outer-ring faults. AdaESPGL performs poorly due to a large error between its estimated period and the theoretical value. AdaSRMD and PGSL methods have better extraction capabilities than the other methods. However, it is worth noting that AdaSRMD is more capable of fault restoration.

[0258] In summary, the present invention proposes an AdaSRMD bearing fault detection method. First, based on the characteristics of the fault signal, without the need for any fault bearing information, the possible fault cycles in the fault signal are dynamically estimated, and constructed into a binary vector embedded in the penalty function to promote the intra-group and inter-group sparsity of the fault signal, thereby achieving denoising of the fault signal. Then, the denoised signal is subjected to the MCKD algorithm to enhance the fault information, effectively combining the respective advantages of the pulse extraction of the adaptive sparse periodic group lasso and the pulse feature enhancement of the MCKD algorithm, so that it has good noise resistance and fault feature extraction capabilities for vibration signals. Finally, the effectiveness and applicability of the method are demonstrated through a simulation study and two experimental cases. The effectiveness of the bearing fault diagnosis method proposed in the present invention is further verified by comparison with the BPD, PGSL and AdaESPGL methods.

[0259] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

[0260] In addition, unless otherwise specified, all technical and scientific terms used in the present invention have the same meaning as commonly understood by those skilled in the art to which the present invention belongs. All documents mentioned in this specification are incorporated by reference to disclose and describe the methods related to the documents. In the event of any conflict with any incorporated document, the content of this specification shall prevail.

Claims

1. A bearing fault diagnosis method, characterized in that: The following steps are involved: Based on the collected bearing fault signal, the frequency spectrum information of the signal is obtained and the fault frequency is filtered out to determine the estimated period value of the signal and obtain the bearing fault period; The obtaining of the bearing failure cycle specifically includes: Perform a first-order difference operation on the signal x to obtain Calculate the signal by Hilbert transform The analytical signal right Perform FFT to obtain spectrum information. The obtained frequency domain information is bilaterally symmetrical. Only the right half is selected to obtain the frequency index after sorting the spectrum energy information in descending order. According to the sorted index, select the top 20 largest frequencies as the active frequency set Among them, f s is the sampling frequency, f c is the fault frequency, n is the index of the fault frequency set; Get the active frequency set The frequency individuals with a medium frequency interval less than 10Hz are removed and the current active frequency set is updated The current set of active frequencies for updates Perform a cyclic search for each element in the , calculate the multiple relationship between them, and select the active frequency set Filter out active frequencies with multiple relationships; Take a weighted average of the active frequencies that have a multiple relationship and calculate the failure frequency f with the highest probability of occurrence. c ; According to the fault frequency f c , determine the estimated period value P of the signal; Based on the bearing's fault cycle, the optimization objective function for the sparse representation of the bearing fault signal is determined. The adaptive parameters in the optimization objective function are used as the optimization target, and a comprehensive index is constructed. With the goal of minimizing the comprehensive index, the optimal adaptive parameters are determined. Based on the optimal adaptive parameters, the optimal sparse representation optimization objective function is determined, and sparse denoising is performed on the bearing fault signal. The comprehensive index is the ratio of the error between the sparse denoised signal and the noise-free original signal to the envelope spectrum peak factor, which is used to characterize the relationship between the adaptive parameters and the noise level. Based on the sparse denoised bearing fault signal, the fitness function of the filter length and displacement number is constructed through the correlation kurtosis. The filter length and displacement number parameter combination with the maximum correlation kurtosis is obtained. The sparsely represented bearing fault signal is filtered to obtain the signal-enhanced periodic fault component. The envelope spectrum analysis of the signal-enhanced periodic fault component is performed to obtain the characteristic frequency and frequency multiples of the fault signal.

2. The bearing fault diagnosis method according to claim 1, characterized in that: The method of determining the optimization objective function for the sparse representation of the bearing fault signal comprises the following steps: The mathematical expression for obtaining the sparse representation of the signal without a dictionary is: Where y∈R N is the original noise-free signal, x∈R N is the sparse denoised signal, λ>0 is the adaptive parameter, and P(x) is the penalty function; Get a binary weight vector b: Where N1 is the estimated length of a fault impulse, N0 is the interval length between adjacent fault impulses, K is the group size, M is the number of complete fault impulses contained in each group, and P is the length occupied by each complete fault impulse, that is, the estimated period value of the signal; According to the sampling frequency f s and fault frequency f c , the parameter P satisfies the equation: P=N1+N0=K / M=f s / f c ; According to the mathematical expression of binary weight vector b, penalty function P(x) and sparse representation without dictionary, the optimization objective function of sparse representation of bearing fault signal is obtained as follows: Where i is the group index coefficient, Ω is the set index, φ(x,a) is the sparsity penalty function, θ(x,b,i)=||b⊙X i,K ||2, a is a constant value that controls the degree of non-convexity of the penalty, μ is a balance parameter, ||x|| 1,w is the reweighted l1 penalty; ||x|| 1,W =||Wx||1,W=diag(w); Where w is a diagonal matrix, and the diagonal elements contain w=[w1,…,w n ],w i Defined as: where ε is a small positive constant used to avoid division by zero; When \(0 < a\leq\frac{1}{MN1\lambda}\) and , \(F(x)\) is convex.

3. The bearing fault diagnosis method according to claim 2, characterized in that: The step of determining the optimal adaptive parameters comprises the following steps: The adaptive parameter λ in the optimization objective function of the sparse representation of the bearing fault signal is used as the optimization target; Based on the adaptive parameter λ, a comprehensive index RC is constructed E , used to characterize the relationship between the adaptive parameter λ and the noise level; The more periodic shocks there are in the vibration signal, the larger the signal envelope spectrum E m The larger the peak value at the frequency of the impact, the larger the peak value at the frequency of the impact. r ,f s / 2] is the ratio of the maximum value to the effective value, which is the envelope spectrum peak factor C. E , expressed as: Among them, f r For frequency conversion; C E The larger the value, the higher the signal-to-noise ratio and the greater the energy of the signal, and the stronger the periodic impact component. The envelope spectrum crest factor is used as a measure of the impact component in the signal. The error between the sparse denoised signal and the original noise-free signal is: The smaller the RMSE, the better the fit between the denoised signal and the original noise-free signal, indicating that the model is more accurate in estimation; C E and RMSE are placed in the denominator and numerator of the fraction respectively to construct the comprehensive index RC E : Comprehensive indicator RC E Used to simultaneously reflect the denoising ability and estimation accuracy of the model; Comprehensive indicator RC E The smaller the value, the closer the signal is to the original fault pulse after denoising. E When the value reaches the minimum, it means that the adaptive parameter λ is optimal; Using the comprehensive indicator RC E The optimal adaptive parameter λ is minimized and the optimal λ under different noise levels is determined by linear fitting curve, and the relationship between λ and noise level is obtained.

4. The bearing fault diagnosis method according to claim 3, characterized in that: The sparse denoising of the bearing fault signal specifically includes: Substituting the optimal adaptive parameter λ into the optimization objective function of the sparse representation of the bearing fault signal, the optimal sparse representation optimization objective function is obtained; The iterative optimization process of the MM algorithm is used to solve the optimal sparse representation optimization objective function and obtain the bearing fault signal after sparse denoising.

5. The bearing fault diagnosis method according to claim 4, characterized in that: The method of obtaining the signal-enhanced periodic fault component comprises the following steps: Based on the maximum correlation kurtosis value of the bearing fault signal after sparse denoising, the fault signal is filtered and the fault features are extracted. The essence of the MCKD algorithm is to find a filter FIR to restore the continuous impact signal submerged by noise; filter the signal through the correlation kurtosis. M When (T) is maximum, the signal recovered by the solved filter satisfies the periodic impulse characteristic; Correlation kurtosis CK M The expression of (T) is: The condition that maximizes the correlation kurtosis is: Calculate the filter coefficients: Where T is the deconvolution period, M is the number of shifts, m∈[0,M], L is the filter length, and f represents the finite impulse response filter vector; The coefficients in the formula are: r = [0 T 2T … mT]; For the filter length L and the displacement number M, the fitness function of the filter length and the displacement number is constructed by the related kurtosis; By using a grid search method, a combination of filter length and displacement number parameters that maximizes the correlation kurtosis is searched to obtain an optimized filter length, displacement number, and deconvolution period, which are periodic fault components for signal enhancement.

6. The bearing fault diagnosis method according to claim 5, characterized in that: The step of obtaining the optimized filter length, displacement number and deconvolution period comprises: Determine the search range and search step size of the filter length L and the displacement number M; Initialize the parameter filter length L and displacement number M according to the search range, and assign the deconvolution period T according to the period estimation result; According to the initialized parameter filter length L, the displacement number M and the deconvolution period T assigned according to the period estimation result, the corresponding correlation kurtosis is calculated as the correlation kurtosis A; Update the parameters according to the search step size, and calculate the relevant kurtosis corresponding to the updated parameters as the relevant kurtosis B; Compare the relevant kurtosis B with the relevant kurtosis A. When the relevant kurtosis B is smaller than the relevant kurtosis A, retain the relevant kurtosis B; when the relevant kurtosis B is larger than the relevant kurtosis A, retain the relevant kurtosis A. After the search is completed, the parameters corresponding to the finally retained correlation kurtosis are output as the optimal filter length L, displacement number M and deconvolution period T.

7. The bearing fault diagnosis method according to claim 6, characterized in that: The obtaining of the characteristic frequency and frequency multiplication of the fault signal specifically includes: According to the optimized filter length, displacement number and deconvolution period, the fault feature of the bearing fault signal after sparse denoising is enhanced by the MCKD algorithm. The enhanced result is subjected to envelope spectrum analysis to obtain the extraction and diagnosis results of the characteristic frequency and frequency multiplication of the bearing fault under strong noise.

8. A bearing fault diagnosis system, characterized in that: include: The fault signal cycle acquisition module is used to obtain the frequency spectrum information of the collected bearing fault signal and filter out the fault frequency, thereby determining the estimated cycle value of the signal and obtaining the bearing fault cycle; The obtaining of the bearing failure cycle specifically includes: Perform a first-order difference operation on the signal x to obtain Calculate the signal by Hilbert transform The analytical signal right Perform FFT to obtain spectrum information. The obtained frequency domain information is bilaterally symmetrical. Only the right half is selected to obtain the frequency index after sorting the spectrum energy information in descending order. According to the sorted index, select the top 20 largest frequencies as the active frequency set Among them, f s is the sampling frequency, f c is the fault frequency, n is the index of the fault frequency set; Get the active frequency set The frequency individuals with a medium frequency interval less than 10Hz are removed and the current active frequency set is updated The current set of active frequencies for updates Perform a cyclic search for each element in the , calculate the multiple relationship between them, and select the active frequency set Filter out active frequencies with multiple relationships; Take a weighted average of the active frequencies that have a multiple relationship and calculate the failure frequency f with the highest probability of occurrence. c ; According to the fault frequency f c , determine the estimated period value P of the signal; The fault signal sparse denoising module is used to determine the optimization objective function for the sparse representation of the bearing fault signal based on the bearing's fault cycle; the adaptive parameters in the optimization objective function are used as the optimization target, a comprehensive index is constructed, and the optimal adaptive parameters are determined with the goal of minimizing the comprehensive index; based on the optimal adaptive parameters, the optimal sparse representation optimization objective function is determined to perform sparse denoising on the bearing fault signal; the comprehensive index is the ratio of the error between the sparse denoised signal and the noise-free original signal to the envelope spectrum peak factor, which is used to characterize the relationship between the adaptive parameters and the noise level; The fault signal enhancement module is used to construct a fitness function of filter length and displacement number based on the bearing fault signal after sparse denoising through correlation kurtosis, obtain the filter length and displacement number parameter combination with the maximum correlation kurtosis, filter the bearing fault signal after sparse representation, and obtain the signal-enhanced periodic fault component; The fault signal diagnosis module is used to perform envelope spectrum analysis on the signal-enhanced periodic fault component to obtain the characteristic frequency and frequency multiples of the fault signal.

Images