A Supervised Sparse Representation Bearing Fault Classification Method Based on Adaptive Feature Selection
Through adaptive feature selection and improved LC-KSVD algorithm, effective features are selected for rolling bearing failure classification, which solves the problem of inconsistent feature selection standards in the existing technology and improves the classification accuracy.
Patent Information
- Application Number
- CN202210829865.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-07-15
- Publication Date
- 2025-08-05
- Estimated Expiration
- 2042-07-15
AI Technical Summary
In the prior art, the characteristic selection criteria for rolling bearing fault diagnosis are not uniform, resulting in low classification accuracy and it is difficult to effectively distinguish different fault types.
Adaptive feature selection method is used to process signals based on features such as time domain, frequency domain and energy entropy. Effective features are filtered out by the objective function of maximizing inter-class distance and minimizing in-class distance, and fault classification is performed in combination with the improved LC-KSVD algorithm.
It improves the accuracy of rolling bearing fault classification and achieves better fault identification results.
Smart Images
Figure CN115204229B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of signal processing, and in particular relates to a supervised sparse representation bearing fault classification method based on adaptive feature selection. Background Art
[0002] Rolling bearings are among the most commonly used and most vulnerable parts in mechanical equipment. Failures can cause significant losses, leading to significant attention paid to bearing fault diagnosis technology. Feature extraction is a critical step in supervised dictionary learning for bearing fault diagnosis. If the extracted features accurately reflect the fault category, they provide high accuracy for bearing fault classification.
[0003] There are many common features used to characterize faults, such as amplitude, phase, frequency, time-domain signals, and time-frequency signals. Because time-domain features contain a large amount of fault information in the signal, and time-frequency features can better distinguish different fault types through the relationship between time and frequency, these two features are primarily used for fault classification. Common time-frequency domain features such as kurtosis and root mean square (RMS) can effectively characterize fault components, but lack a unified feature selection standard for sparse representation methods using dictionary learning, resulting in low classification accuracy. Summary of the Invention
[0004] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a supervised sparse representation method based on adaptive feature selection. Signals processed with common time domain, frequency domain, and energy entropy features are used as training data for the sparse representation classification method. Based on the adaptive feature selection method, the features with the best fault classification effect are extracted as the input training set. The improved LC-KSVD fault classification method is used to achieve rolling bearing fault classification.
[0005] To achieve the above-mentioned object of the invention, a supervised sparse representation bearing fault classification method based on adaptive feature selection comprises the following steps:
[0006] Step S1, obtaining an initial signal training set and a test set;
[0007] Step S2: Establish an adaptive objective function for the training set composed of feature matrices. Different features in the same dictionary will correspond to different sparse representation coefficients. Different sparse representation coefficients can be used to classify faults. Therefore, by maximizing the inter-class distance and minimizing the intra-class distance, the index screening can be effectively achieved.
[0008] Step S3: feature screening;
[0009] The scores of each indicator obtained by the objective function are sorted from small to large, and the eigenvalues corresponding to the minimum values in the first stage are selected as the filtered feature data;
[0010] Step S4: Re-divide the training set and test set;
[0011] Select the feature after adaptive feature filtering new Extract bearing vibration signals and generate new training set Y and test set Y train ;
[0012] Step S5: Initialization of LC-KSVD (label consistency K-fold singular value decomposition):
[0013] Initialize the initial dictionary D, the classification sparse code Q, the linear transformation matrix A (the purpose is to make the transformed AX and Q similar, so that the sparse representation coefficients of the same type of signals are similar and the classification accuracy is improved), and the classification parameter W. Use the KSVD (K-fold singular value decomposition) algorithm to learn the training data set for 20 iterations to obtain the initial dictionary D (0) , use the multivariate ridge regression model to initialize the linear transformation matrix A (0) and classification parameter W (0) :
[0014]
[0015] A (0) =QX T (XX T +λ2I) -1
[0016] W (0) =HX T (XX T +λ1I) -1
[0017] Among them, argmin A represents the variable value that minimizes the objective function A, X is the sparse coefficient matrix, I is the identity matrix, Q is the categorical sparse coding for classifying the input signal (the purpose is to generate a distinct sparse representation coefficient even if the number of dictionaries is large or small), λ1 and λ2 are the multivariate ridge regression coefficients, and H represents the label;
[0018] Step S6, LC-KSVD dictionary learning;
[0019]
[0020] Where Y represents the initial signal, D is the initial dictionary, X is the sparse coefficient matrix, W is the classification parameter, x i represents the elements in the sparse coefficient matrix X, Represents the constraint condition, any i, x iThe absolute value of is less than or equal to T, T represents the upper limit of sparsity, and is also the upper limit of non-zero terms in the sparse representation coefficient. α and β represent weight coefficients, and their sizes affect the influence of each term on the entire optimization function.
[0021] Step S7: Combine and simplify the formula S6 to obtain a new dictionary D new , update D, A and W by KSVD algorithm;
[0022]
[0023]
[0024] make
[0025]
[0026]
[0027] where Y new Indicates the new signal formed by the combination, D new Represents the new dictionary formed by the combination, α and β represent weight coefficients, Indicates that the objective function D new The variable value when X takes the minimum value, st represents the constraint condition;
[0028] Step S8: Since (D, A, W) is in D in LC-KSVD new It is l2-normed and the expected value of (D, A, W) can be calculated. Determine parameters
[0029] Step S9: Use the learned linear classifier W to estimate the label vector Represents the expected corresponding element of the sparse coefficient matrix X, y i The label corresponds to the index of the maximum value in the label vector l, so the category of the input signal is determined and classification is achieved.
[0030] Furthermore, the specific process of step S1 is:
[0031] The n time domain, frequency domain, and energy entropy indicators are used to extract the initial signal features. Assuming that the lengths of the vibration signals of the bearing outer ring fault, inner ring fault, and rolling element fault are x respectively, and the total signal length is 3x, the extracted initial signal feature matrix constitutes the initial signal training set F = {f ij |i∈[1,m],j∈[1,n]}, i is the row of the training set, and each row is A eigenvalue is obtained from each point, j represents the column of the training set, m represents the number of indicators, and the same method is used to obtain the original signal test set.
[0032] Furthermore, the objective function of step S2 is:
[0033]
[0034] Among them: a is the characteristic data after one of the indicators extracts the signal, a1, a2, and a3 refer to the characteristic data extracted from the bearing inner ring fault signal, the bearing outer ring fault signal, and the bearing rolling element fault signal, respectively, Md represents the median, QU represents the upper quartile, QL represents the lower quartile, MAX represents the maximum value, and MIN represents the minimum value.
[0035] Furthermore, the detailed steps of the feature screening in step S4 are as follows:
[0036] By maximizing the inter-class distance and minimizing the intra-class distance, the objective function is calculated:
[0037]
[0038] Where: a is the characteristic data extracted from the signal by one of the indicators, a1, a2, and a3 refer to the characteristic data extracted from the bearing inner race fault signal, bearing outer race fault signal, and bearing rolling element fault signal, respectively; Md represents the median, QU represents the upper quartile, QL represents the lower quartile, MAX represents the maximum value, and MIN represents the minimum value;
[0039] Minimize the objective function, sort the scores of each indicator obtained by the objective function from small to large, select several points before the first change point, and the indicators corresponding to these minimum scores are the indicators screened by adaptive feature selection.
[0040] Further, in step S8, the expected value of the available (D, A, W) is calculated. Determine parameters
[0041]
[0042]
[0043]
[0044] where x k,i is the kth row and ith column of the sparse coefficient matrix X, d k is the kth element in dictionary D, a k is the kth element in the linear transformation matrix A, w k is the kth element in the classification parameter, They are and The kth column of .
[0045] The present invention proposes a supervised sparse representation method with adaptive feature selection, extracts the most effective features for fault classification as input training sets, and uses an improved LC-KSVD fault classification method to classify rolling bearing faults. Simulations and case studies demonstrate that this method achieves better classification results than traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 The invention is a supervised sparse representation bearing fault classification based on adaptive feature selection.
[0047] Figure 2 It is the score of each indicator after being processed by the adaptive feature objective function.
[0048] Figure 3 This is a table showing the classification accuracy of the supervised sparse representation bearing classification method based on adaptive feature selection for Case Western Reserve University (CRWU) and homemade test bench data. DETAILED DESCRIPTION
[0049] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the conclusions of the invention are not limited thereto.
[0050] A supervised sparse representation bearing fault classification method based on adaptive feature selection includes the following steps:
[0051] Step S1: Obtain an initial signal training set and a test set.
[0052] The n time domain, frequency domain, and energy entropy indicators are used to extract the initial signal features. Assume that the lengths of the vibration signals of the bearing outer ring fault, inner ring fault, and rolling element fault are x respectively, and the total signal length is 3x. The extracted initial signal feature matrix constitutes the initial signal training set F = {f ij |i∈[1,m],j∈[1,n]}, i is the row of the training set, and each row is The original signal test set is obtained in the same way.
[0053] Step S2: Establish an adaptive objective function for the training set composed of feature matrices.
[0054] For the same dictionary, different features will correspond to different sparse representation coefficients. Different sparse representation coefficients can be used to classify faults. Therefore, by maximizing the inter-class distance and minimizing the intra-class distance, the index screening can be effectively achieved. The objective function is:
[0055]
[0056] Where a is the feature data extracted from one of the indicators, a1, a2, and a3 are the feature data extracted from the bearing inner race fault signal, bearing outer race fault signal, and bearing rolling element fault signal, respectively. Md represents the median, QU represents the upper quartile, QL represents the lower quartile, MAX represents the maximum value, and MIN represents the minimum value.
[0057] Step S3: Feature screening.
[0058] The scores of each indicator obtained by the objective function are sorted from small to large, and the eigenvalues corresponding to the minimum values in the first stage are selected as the filtered feature data.
[0059] Step S4: Re-divide the training set and test set.
[0060] Select the feature after adaptive feature filtering new Extract bearing vibration signals and generate new training set Y and test set Y train .
[0061] Step S5, initialization of LC-KSVD:
[0062] Initialize the initial dictionary D, linear transformation matrix A, and classification parameter W, and use the KSVD algorithm to perform 20 iterations on the training data set to obtain the initial dictionary D (0) , use the multivariate ridge regression model to initialize the linear transformation matrix A (0) and classification parameter W (0) .
[0063]
[0064] A (0) =QX T (XX T +λ2I) -1
[0065] W (0) =HX T (XX T +λ1I) -1
[0066] Where X is the sparse coefficient matrix, Q is the categorical sparse coding for classifying the input signal (the purpose is to generate a distinct sparse representation coefficient even if the number of dictionaries is large or small), and λ1 and λ2 are the multivariate ridge regression coefficients respectively.
[0067] Step S6: LC-KSVD dictionary learning.
[0068]
[0069]
[0070] Where Y represents the initial signal, D is the initial dictionary, X is the sparse coefficient matrix, W is the classification parameter, x i represents the elements in the sparse coefficient matrix X, Represents the constraint condition, any i, x i The absolute value of is less than or equal to T, T represents the upper limit of sparsity (that is, the upper limit of non-zero terms in the sparse representation coefficient), α and β represent weight coefficients, and their sizes affect the influence of each term on the entire optimization function.
[0071] Step S7: Combine and simplify the formula in step S6 to obtain D new , update D, A and W through the KSVD algorithm.
[0072]
[0073]
[0074] make
[0075]
[0076]
[0077] where Y new Indicates the new signal formed by the combination, D new Represents the new dictionary formed by the combination, α and β represent weight coefficients, Indicates that the objective function D new The variable value when X and X take the minimum value, st represents the constraint condition.
[0078] Step S8: Since (D, A, W) is in D in LC-KSVD new It is l2-normed and the expected value of (D, A, W) can be calculated. Determine parameters
[0079]
[0080]
[0081]
[0082] in They are and The kth column of .
[0083] Step S9: Use the learned linear classifier W to estimate the label vector y i The label corresponds to the index of the maximum value in the label vector l, so the category of the input signal is determined and classification is achieved.
[0084] To further verify the effectiveness of the supervised sparse representation bearing classification method with adaptive feature selection, we investigated a case with outer race fault, inner race fault, and rolling element fault to validate the proposed method. The first set of data was collected from the University of Western Reserve University, and the second set of data was collected on a self-made test bench.
[0085] The first set of data comes from Case Western Reserve University (CRWU). The signal lengths for outer race fault, inner race fault, and rolling element fault are all 121,265, respectively. The acceleration sampling rate is 12 kHz. The first 32,000 points are used as training data, and the last 32,000 points are used as test data. A supervised sparse representation bearing classification method with adaptive feature selection is used, and the classification accuracy is 100%.
[0086] The second set of data comes from a homemade test bench. The signal lengths for outer ring fault, inner ring fault, and rolling element fault are all 322,560, respectively. The acceleration sampling rate is 12.8kHz. The first 32,000 points are used as training data, and the last 32,000 points are used as test data. A supervised sparse representation bearing classification method with adaptive feature selection is used, and the classification accuracy is also 100%.
[0087] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "illustrative embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, illustrative uses of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.
[0088] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.
Claims
1. A supervised sparse representation bearing fault classification method based on adaptive feature selection, characterized in that: The following steps are involved: Step S1, obtaining an initial signal training set and a test set; Step S2: Establish an adaptive objective function for the training set composed of feature matrices. Different features in the same dictionary will correspond to different sparse representation coefficients. Different sparse representation coefficients can be used to classify faults. Therefore, by maximizing the inter-class distance and minimizing the intra-class distance, the index screening can be effectively achieved. Step S3: feature screening; The scores of each indicator obtained by the objective function are sorted from small to large, and the eigenvalues corresponding to the minimum values in the first stage are selected as the filtered feature data; Step S4: Re-divide the training set and test set; Select the feature after adaptive feature filtering new Extract bearing vibration signals and generate new training set Y and test set Y train ; Step S5, initialization of LC-KSVD: Initialize the initial dictionary D, classification sparse code Q, linear transformation matrix A, and classification parameter W, and use the KSVD algorithm to perform 20 iterations on the training data set to obtain the initial dictionary D (0) , use the multivariate ridge regression model to initialize the linear transformation matrix A (0) and classification parameter W (0) : A (0) =QX T (XX T +λ2I) -1 W (0) =HX T (XX T +λ1I) -1 Among them, argmin A represents the variable value when the objective function A takes the minimum value, X is the sparse coefficient matrix, I is the identity matrix, Q is the categorical sparse coding for input signal classification, λ1 and λ2 are the multivariate ridge regression coefficients, and H represents the label; Step S6, LC-KSVD dictionary learning; Where Y represents the initial signal, D is the initial dictionary, X is the sparse coefficient matrix, W is the classification parameter, x i represents the elements in the sparse coefficient matrix X, Represents the constraint condition, any i, x i The absolute value of is less than or equal to T, T represents the upper limit of sparsity, and is also the upper limit of non-zero terms in the sparse representation coefficient. α and β represent weight coefficients, and their sizes affect the influence of each term on the entire optimization function. Step S7: Combine and simplify the formula S6 to obtain a new dictionary D new , update D, A and W by KSVD algorithm; make where Y new Indicates the new signal formed by the combination, D new Represents the new dictionary formed by the combination, α and β represent weight coefficients, Indicates that the objective function D new The variable value when X takes the minimum value, st represents the constraint condition; Step S8: Since (D, A, W) is in D in LC-KSVD new It is l2-normed and the expected value of (D, A, W) can be calculated. Determine parameters Step S9: Use the learned linear classifier W to estimate the label vector Represents the expected corresponding element of the sparse coefficient matrix X, y i The label corresponds to the index of the maximum value in the label vector l, so the category of the input signal is determined and classification is achieved; The specific process of step S1 is: The n time domain, frequency domain, and energy entropy indicators are used to extract the initial signal features. Assuming that the lengths of the vibration signals of the bearing outer ring fault, inner ring fault, and rolling element fault are x respectively, and the total signal length is 3x, the extracted initial signal feature matrix constitutes the initial signal training set F = {f ij |i∈[1,m],j∈[1,n]}, i is the row of the training set, and each row is Calculate an eigenvalue for each point, j represents the column of the training set, m represents the number of indicators, and the same method is used to obtain the original signal test set; The detailed steps of the feature screening in step S4 are as follows: By maximizing the inter-class distance and minimizing the intra-class distance, the objective function is calculated: Where: a is the characteristic data extracted from the signal by one of the indicators, a1, a2, and a3 refer to the characteristic data extracted from the bearing inner race fault signal, bearing outer race fault signal, and bearing rolling element fault signal, respectively; Md represents the median, QU represents the upper quartile, QL represents the lower quartile, MAX represents the maximum value, and MIN represents the minimum value; Minimize the objective function, sort the scores of each indicator obtained by the objective function from small to large, select several points before the first change point, and the indicators corresponding to these minimum scores are the indicators screened by adaptive feature selection.
2. A supervised sparse representation bearing fault classification method based on adaptive feature selection according to claim 1, characterized in that: In step S8, the expected value of (D, A, W) that can be used is calculated. Determine parameters where x k,i is the kth row and ith column of the sparse coefficient matrix X, d k is the kth element in dictionary D, a k is the kth element in the linear transformation matrix A, w k is the kth element in the classification parameter, They are and The kth column of .