Bearing compound fault sparse classification method and system based on adaptive feature selection

Abstract

The invention provides a bearing compound fault sparse classification method and system based on adaptive feature selection, and the method comprises the steps: firstly, enhancing sample features through intra-class and inter-class distance measurement, and removing a redundant feature set to achieve the adaptive fault feature screening, and effectively improving the fault recognition performance; constructing a single-class fault initialization dictionary by using the feature training set, and enhancing sparsity and keeping model convexity through a norm constraint term; and iteratively training the over-complete dictionary by using the corresponding sample set of each fault category, and finally realizing the discriminant classification of the bearing compound fault through a discriminant vector. Compared with a traditional method in simulation and test, the method can effectively solve the feature adaptive selection problem in the LC-KSVD algorithm, improves the accuracy and robustness of fault classification under the small dictionary condition and the strong noise condition, reduces the time cost, and provides valuable technical reference for engineering application.

Description

Technical Field

[0001] The invention belongs to the field of fault diagnosis and state monitoring, and in particular relates to a bearing composite fault sparse classification method and system with adaptive feature selection. Background Art

[0002] Gears, bearings and other important rotating parts are widely used in various mechanical transmission systems, and their operating conditions have a significant impact on the safety of the entire equipment. Due to the complex operating conditions in actual working conditions, multiple faults may exist simultaneously in the rotating mechanical system, resulting in difficult-to-distinguish composite faults. Therefore, accurate classification of composite faults is crucial for effective fault diagnosis. In dictionary learning (DL) methods, the label consistent singular value decomposition (LC-KSVD) algorithm has been proven to effectively reduce the sparsity of the corresponding coefficient matrix in the dictionary, and has the ability to classify different feature data, so it has become a widely used tool in the fields of image and signal classification. It uses the l0 norm regularization term as a sparse penalty function, but due to its discreteness and non-convexity, it is often difficult to find the optimal solution, and its resistance to outliers is insufficient, and the classification accuracy is low under strong noise conditions. In the dictionary construction process of the LC-KSVD method, the prior knowledge of the fault signal is often not fully utilized to construct an initial dictionary that adapts to various fault types, and adaptive feature selection cannot be achieved. Therefore, it is necessary to construct a sparse dictionary learning method for adaptive feature selection for bearing composite fault classification. Summary of the invention

[0003] In view of the above technical problems, the present invention provides a sparse classification method and system for bearing composite faults with adaptive feature selection. First, the sample features are enhanced by intra-class and inter-category distance measurement and redundant feature sets are removed to achieve adaptive fault feature screening, effectively improving fault identification performance; a single-class fault initialization dictionary is constructed using a feature training set, and sparsity is enhanced and model convexity is maintained by a norm constraint term; the corresponding sample set of each fault category is used to iteratively train an overcomplete dictionary, and finally the discriminant classification of bearing composite faults is achieved through a discriminant vector. Compared with the traditional method in simulation and experiment, the proposed method can effectively solve the feature adaptive selection problem in the LC-KSVD algorithm, improve the accuracy and robustness of fault classification under small dictionary conditions and strong noise conditions, reduce time costs, and provide valuable technical references for engineering applications.

[0004] The present invention achieves the above technical objectives through the following technical means.

[0005] A sparse classification method for bearing composite faults based on adaptive feature selection includes the following steps:

[0006] Step S1, adaptive feature selection: divide the bearing composite fault signal samples into i categories, calculate the intra-class Euclidean distance and inter-class Euclidean distance feature indicators of the bearing fault signal samples, enhance the sample features through an adaptive feature selection algorithm, remove redundant feature sets, implement adaptive feature selection on the fault signals through the sorting of the adaptive feature selection algorithm, and construct an initial training set and test set of bearing fault signal samples of each category;

[0007] Step S2, model construction: constructing the LC-KSVD label consistency singular value decomposition algorithm model, adding the l1 norm constraint term to obtain a new objective function;

[0008] Step S3, dictionary initialization: construct the initial dictionary D using the training set obtained in step S1 *(0) , and then obtain the initial linear transformation matrix A (0) and the classifier parameter W (0) ;

[0009] Step S4, dictionary training: Based on the objective function obtained in step S2, the initial dictionary D is trained using the test set obtained in step S1. *(0) Perform iterative training until the specified number of iterations j is reached and the feature dictionary D is obtained *(j) And the corresponding linear transformation matrix A and classifier parameters W;

[0010] Step S5, classification stage: based on the feature dictionary D obtained in step S4 *(j) , and obtain the sparse representation coefficient matrix. The discriminant vector l is obtained through the sparse representation coefficient matrix, and the bearing composite fault signal is discriminated and classified through the index with the largest value in the discriminant vector l.

[0011] In the above scheme, the specific steps of step S1 include:

[0012] Step S11: Calculate and maximize the intra-class Euclidean distance S1 of the fault signal. The calculation formula of the intra-class Euclidean distance S1 is as follows:

[0013]

[0014] Among them, Md(·) represents the median, c is the Euclidean distance, a m 、a n It is the characteristic data extracted from the bearing composite fault signal;

[0015] Step S12: Calculate and minimize the inter-category Euclidean distance S2. The calculation formula of the inter-category Euclidean distance S2 is as follows:

[0016]

[0017] Among them, QU(·) represents the upper quartile, QL(·) represents the lower quartile, and a i is the characteristic data of the i-th type bearing fault signal;

[0018] Step S13: Using the intra-class Euclidean distance S1 and inter-class Euclidean distance S2 obtained in steps S11 and S12, the eigenvalues ​​S of the bearing fault signals are calculated and sorted using an adaptive feature selection algorithm. The feature data sets with higher eigenvalue S scores are retained, and the feature data sets with lower scores are deleted. The retained feature data sets are used to construct a new training set. With the test set

[0019] In the above scheme, the objective function of the adaptive feature selection algorithm in step S1 is as follows:

[0020]

[0021] Among them, S represents the eigenvalue; S1 is the Euclidean distance within the class; S2 is the Euclidean distance between categories; MAX(·) means finding the maximum value, and MIN(·) means finding the maximum value.

[0022] In the above scheme, the objective function of the LC-KSVD label consistency singular value decomposition algorithm model in step S2 is<D,W,A,X> for:

[0023]

[0024] Here, argmin(·) means finding the combination of independent variable values ​​that minimizes the function value, X = [x1,…,x N ] represents the sparse representation coefficient, x z is the Zth column of X (Z = 0, 1, ..., N), represents the two-norm, is the i-type test data set, represents a set of real numbers, K is the number of columns of D, N is the number of columns of Y and K>N, D is an overcomplete dictionary, α represents the weights of the reconstruction term and the label consistency term, refers to the error of sparse representation classification, Q represents the sparse classification of the input signal, represents the linear transformation matrix, β represents the weight of the reconstruction error term, is the reconstruction error term, Represents the classifier parameters, label matrix represents the classification label assigned to the input signal, T represents the coefficient constraint, and ||x z ||0 means zero norm, and st means constraint.

[0025] Furthermore, the step S2 adds the l1 norm constraint term, and then obtains a new objective function <σ,X>:

[0026]

[0027] Among them, ||·||1 represents a norm, y * is the training set, τ i It is a threshold parameter used to adjust the output sparsity.

[0028] In the above scheme, the initial linear transformation matrix A in step S3 is (0) and the classifier parameter W (0) as follows:

[0029] A (0) =QX *T (X * X *T +λI) -1

[0030] W (0) =HX *T (X * X *T +λI) -1

[0031] Among them, Q represents the sparse classification of the input signal, X * To solve the sparse coefficients for each initial feature dictionary, we obtain the sparse representation coefficient matrix, X * =[X1,…,X i ], λ is the unit parameter, I is the unit matrix, and the label matrix represents the classification label assigned to the input signal, Represents the set of real numbers.

[0032] In the above scheme, the specific steps of step S4 include:

[0033] Step S41: solving the sparse representation coefficient matrix X according to the fixed dictionary σ of the new objective function in step S2;

[0034] Step S42: Update dictionary atoms column by column according to the solved sparse representation coefficient matrix, and slowly include the representation error into the original dictionary while satisfying the sparsity, that is, update and optimize row by row and column by column to reduce the overall error. The objective function of updating the dictionary is:

[0035]

[0036] Among them, d k is the kth atom in the dictionary, For missing d in the dictionary k and the sparse representation coefficient matrix is ​​missing x k The sparse representation coefficient after Ek For missing d in the dictionary k and the sparse representation coefficient matrix is ​​missing x k The sparse representation residual after ;

[0037] Step S43: Repeat steps S41 and S42 until the specified number of iterations j is reached to obtain the optimal feature matrix and its sparse representation coefficient matrix From this, the linear transformation matrix A and classifier parameters W can be calculated.

[0038] Furthermore, the sparse representation residual E in step S42 k :

[0039]

[0040] Where, d j is the jth atom in the dictionary, x j represents the jth coefficient of the sparse representation coefficient matrix, y * For the training set.

[0041] In the above scheme, the specific steps of step S5 include:

[0042] Step S51: Dictionary D obtained in step S4 *(j) And the corresponding linear transformation matrix A and classifier parameters W are normalized:

[0043]

[0044] in, Respectively represent D *(d) , A, W after normalization calculation result, a K is the Kth column of matrix A, for a K Normalized result, d K is the matrix D *(j) The Kth column of is d K Normalized result, w K is the Kth column of matrix W, w K Normalized result, ||·||2 represents the second norm.

[0045] Step S52: According to the normalization result of step S51, solve the test set as follows: The sparse representation coefficient matrix x test :

[0046]

[0047] Here, argmin(τ) represents the combination of independent variable values ​​that minimizes the function value. is the test set, represents the two norm, st represents the constraint, X=[x1,…,x N ] represents the sparse representation coefficient, x z is the Zth column of X (Z=0,1,…,N), τ i is a threshold parameter used to adjust the output sparsity, and ||·||1 represents a norm;

[0048] Step S53: Solve the discriminant vector l as follows, and identify the index with the largest value in l as The corresponding category is used to achieve discriminative classification:

[0049]

[0050] Here, argmax(·) represents the combination of independent variable values ​​that maximizes the function value, and l is the coefficient x with sparse representation. test Linear classifier Generated exist The discriminant vector on .

[0051] A system for implementing the above-mentioned sparse classification method for bearing composite faults by adaptive feature selection, comprising a feature selection module, a model building module, a dictionary initialization module, a dictionary training module and a classification module;

[0052] The feature selection module is used to classify bearing composite fault signal samples into i categories, calculate the intra-class Euclidean distance and inter-class Euclidean distance feature indicators of bearing fault signal samples, enhance sample features through an adaptive feature selection algorithm, remove redundant feature sets, implement adaptive feature selection of fault signals through the sorting of the adaptive feature selection algorithm, and construct an initial training set and test set of bearing fault signal samples of each category;

[0053] The model building module is used to build the LC-KSVD label consistency singular value decomposition algorithm model, adding the l1 norm constraint term to obtain a new objective function;

[0054] The dictionary initialization module is used to construct an initial dictionary D through the training set obtained by the feature selection module. *(0) , and then obtain the initial linear transformation matrix A (0) and the classifier parameter W (0) ;

[0055] The dictionary training module is used to train the initial dictionary D using the test set obtained by the feature selection module according to the objective function obtained by the model building module. *(0)Perform iterative training until the specified number of iterations j is reached and the feature dictionary D is obtained *(j) And the corresponding linear transformation matrix A and classifier parameters W;

[0056] The classification module is used to obtain the feature dictionary D based on the dictionary training module. *(j) , and obtain the sparse representation coefficient matrix. The discriminant vector l is obtained through the sparse representation coefficient matrix, and the bearing composite fault signal is discriminated and classified through the index with the largest value in the discriminant vector l.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] According to one embodiment of the present invention, the present invention can effectively solve the feature adaptive selection problem in the LC-KSVD algorithm, improve the accuracy and robustness of fault classification under small dictionary conditions and strong noise conditions, reduce time costs, and provide valuable technical references for engineering applications.

[0059] According to one embodiment of the present invention, adaptive fault feature screening is achieved by enhancing sample features through intra-class and inter-class distance metrics and removing redundant feature sets, thereby effectively improving fault identification performance.

[0060] According to one embodiment of the present invention, each category of fault corresponds to a training set, and each category of fault has a corresponding training set, test set and dictionary, which can increase the specificity of the dictionary and improve the accuracy of the dictionary in identifying fault types.

[0061] According to one embodiment of the present invention, sparsity can be enhanced and model convexity can be maintained through a norm constraint term. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 It is a schematic diagram of a process of one embodiment of the present invention.

[0063] Figure 2 is a loss function diagram under different noise conditions according to an embodiment of the present invention, wherein Figure 2 (a) is the loss function diagram when the noise level is 10dB. Figure 2 (b) is the loss function diagram when the noise level is 5dB. Figure 2 (c) is the loss function diagram when the noise level is 0dB. Figure 2 (d) is the loss function diagram when the noise level is -5dB. Figure 2 (e) Loss function graph when the noise level is -10dB.

[0064] Figure 3 : is a comparison chart of the effect of dictionary size on classification accuracy and operation time under different noise conditions in one embodiment of the present invention and the traditional method, wherein Figure 3 (a) is a comparison chart showing the effect of dictionary size on classification accuracy and operation time when the noise level is 10dB. Figure 3 (b) is a comparison chart showing the effect of dictionary size on classification accuracy and operation time when the noise level is 5dB. Figure 3 (c) is a comparison chart showing the effect of dictionary size on classification accuracy and operation time when the noise level is 0dB. Figure 3 (d) is a comparison chart showing the effect of dictionary size on classification accuracy and operation time when the noise level is -5dB. Figure 3 (e) is a comparison chart of the impact of dictionary size on classification accuracy and calculation time when the noise level is -10dB.

[0065] Figure 4 This is a comparison chart of the classification accuracy of an embodiment of the present invention and a machine learning method under different noise conditions.

[0066] Figure 5 is a structural diagram of a test bench 1 and a test bench 2 according to an embodiment of the present invention, wherein Figure 5 (a) is a structural diagram of the test bench 1 in the present invention, wherein Figure 5 (b) is a structural diagram of the test bench 2 in the present invention.

[0067] Figure 6 This is a comparison chart of classification accuracy and operation time based on LC-KSVD to increase the improvement measures in the present invention in one embodiment of the present invention, wherein Figure 6 (a) is a comparison chart of the classification accuracy and operation time of the test bench 1 data by various improvement measures in the present invention, Figure 6 (b) is a comparison chart of the classification accuracy and operation time of the test bench 2 data for each improvement measure in the present invention. DETAILED DESCRIPTION

[0068] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0069] In this embodiment, preferably, in order to verify the effectiveness and practicality of the present invention, composite fault detection experiments are carried out using simulation data, data collected during actual bearing operation, and operation data of a homemade test bench. These three sets of data sets all contain seven types of bearing composite fault signals, including normal bearing status, outer ring fault, inner ring fault, rolling element fault, outer rolling fault, inner rolling fault, and inner and outer ring faults.

[0070] Example 1

[0071] Figure 1 The figure shows a preferred implementation of the bearing composite fault sparse classification method based on adaptive feature selection according to the present invention, which includes the following steps:

[0072] Step S1, adaptive feature selection: divide the bearing composite fault signal samples into i categories, calculate the intra-class Euclidean distance and inter-class Euclidean distance feature indicators of the bearing fault signal samples, enhance the sample features through an adaptive feature selection algorithm, remove redundant feature sets, implement adaptive feature selection on the fault signals through the sorting of the adaptive feature selection algorithm, and construct an initial training set and test set of bearing fault signal samples of each category;

[0073] Step S2, model construction: constructing the LC-KSVD label consistency singular value decomposition algorithm model, adding the l1 norm constraint term to obtain a new objective function;

[0074] Step S3, dictionary initialization: construct the initial dictionary D using the training set obtained in step S1 *(0) , and then obtain the initial linear transformation matrix A (0) and the classifier parameter W (0) ;

[0075] Step S4, dictionary training: Based on the objective function obtained in step S2, the initial dictionary D is trained using the test set obtained in step S1. *(0) Perform iterative training until the specified number of iterations j is reached and the feature dictionary D is obtained *(j) And the corresponding linear transformation matrix A and classifier parameters W;

[0076] Step S5, classification stage: based on the feature dictionary D obtained in step S4 *(j) , and obtain the sparse representation coefficient matrix. The discriminant vector l is obtained through the sparse representation coefficient matrix, and the bearing composite fault signal is discriminated and classified through the index with the largest value in the discriminant vector l.

[0077] According to this embodiment, preferably, the data source in step S1 is a plurality of simulation signals, and the sampling frequency is assumed to be fs=500 Hz, the sampling point is n=200000, and t is the time, then the signal X(t) representing the bearing outer ring fault can be expressed as:

[0078]

[0079] Where: Indicates rounding down, A k represents the k-th fault impulse amplitude, U(·) is the unit impulse response function, ξ is the damping ratio, k is the number of impulse responses, T0 is the impulse response period, is the initial phase value, f is the resonant frequency, fs is the sampling frequency, n is the sample length, and τ0 represents the signal delay. The values ​​of each parameter are shown in the following table:

[0080] Table 1 Outer ring fault parameter values

[0081]

[0082] The bearing inner ring fault signal X(t) can be expressed as follows:

[0083]

[0084] Where: a k represents the kth fault impulse amplitude, f r Indicates the frequency of rotation, d k is the amplitude modulation amount, c k is the amplitude offset, and the other parameter definitions are consistent with the rolling bearing outer ring fault vibration signal model. The values ​​of each parameter are shown in the following table:

[0085] Table 2 Inner ring fault parameter values

[0086]

[0087] The mathematical expression of the rolling element fault simulation signal is as follows:

[0088]

[0089] Where: R represents the impact between the rolling element and the inner ring and outer ring. When R=1, it means that there is an impact between the rolling element and the inner ring. When R=2, it means that there is an impact between the rolling element and the outer ring. s(t) represents the natural frequency resonance caused by the fault. s1 means the inner ring raceway collides, and s2 means the outer ring raceway collides. The definitions of the remaining parameters are consistent with the vibration signal model of the outer ring fault of the rolling bearing. The values ​​of each parameter are shown in the following table:

[0090] Table 3 Rolling element fault parameter values

[0091]

[0092] The remaining composite fault simulation signals are set as the superposition of each single fault signal, and the normal state signal is a periodic signal that obeys the normal distribution.

[0093] According to this embodiment, preferably, in step S1, based on the Euclidean distance, the intra-class and inter-category distances are used as feature metrics to measure the separability of the classes, and an adaptive feature selection algorithm integrating the intra-class and inter-category distance information is proposed. The objective function S of the algorithm is defined as follows:

[0094]

[0095] Among them, S1 is the Euclidean distance between classes, S2 is the Euclidean distance within a class, MAX(·) means finding the maximum value, and MIN(·) means finding the maximum value.

[0096] According to this embodiment, preferably, in step S1, there are seven composite faults in the bearing, and the characteristic data extracted from the composite fault signal are recorded as a1, a2, a3, a4, a5, a6 and a7. The specific steps in step S1 are as follows:

[0097] (1) Calculate and maximize the Euclidean distance between fault classes to quantify the degree of separation between different fault statistical characteristics.

[0098]

[0099] Among them, Md(·) represents the median, c refers to the Euclidean distance, a m 、a n It is the characteristic data extracted from the bearing composite fault signal.

[0100] (2) Calculate and minimize the Euclidean distance S2 between categories.

[0101]

[0102] Among them, QU(·) represents the upper quartile, QL(·) represents the lower quartile, and a i is the characteristic data of the i-th type bearing fault signal.

[0103] (3) Calculate the S value of each feature and sort it by using the intra-class Euclidean distance S1 and inter-class Euclidean distance S2 obtained in steps (1) and (2), and construct a new training set by retaining the selected indicators. With the test set In this embodiment, 14 time-frequency domain features are screened out. Two non-overlapping signals with a length of 100 are selected from the 7 composite fault signals for feature extraction to form a training set y with a size of 14×700. * With the test set

[0104] According to this embodiment, preferably, the LC-KSVD original objective function used in step S2 is<D,W,A,X> for:

[0105]

[0106] Here, argmin(·) means finding the combination of independent variable values ​​that minimizes the function value, X = [x1,…,x N ] represents the sparse representation coefficient, x z is the Zth column of X (Z = 0, 1, ..., N), represents the two-norm, is the i-type test data set, represents a set of real numbers, K is the number of columns of D, N is the number of columns of Y and K>N, D is an overcomplete dictionary, α represents the weights of the reconstruction term and the label consistency term, refers to the error of sparse representation classification, Q represents the sparse classification of the input signal. represents the linear transformation matrix, β represents the weight of the reconstruction error term, is the reconstruction error term, Represents classification parameters and label matrix represents the classification label assigned to the input signal, ||x z ||0 represents zero norm, T represents coefficient constraint, and st represents constraint. The above formula is simplified to:

[0107]

[0108] Let the two merge matrices And adding the l1 norm constraint term, we get the objective function <σ,X>:

[0109]

[0110] Among them, ||·||1 represents a norm, τ i is a threshold parameter used to adjust the output sparsity. In this embodiment, τ i = 1, α = 1, β = 1. The one-norm constraint term can enhance the sparsity and maintain the convexity of the model.

[0111] According to this embodiment, preferably, the label matrix H is a diagonal matrix of size 7×700, and the remaining elements outside the diagonal are 0, as shown below:

[0112]

[0113] According to this embodiment, preferably, in the initialization phase in step S3, In each, a column vector is taken as the initial dictionary In this embodiment, a=10, that is, the dictionary size is 70. The sparse coefficients of each initial feature dictionary are solved to obtain the sparse representation coefficient matrix X * =[X1,…,X i ], initialize A (0) ,W (0) As follows:

[0114] A (0) =QX *T (X * X *T +λI)-1

[0115] W (0) =HX *T (X * X *T +λI) -1

[0116] Among them, Q represents the sparse classification of the input signal, X * To solve the sparse coefficients for each initial feature dictionary, we obtain the sparse representation coefficient matrix, X * =[X1,…,X i ], λ is the unit parameter, I is the unit matrix, and the label matrix represents the classification label assigned to the input signal. According to this embodiment, preferably, λ=2.

[0117] According to this embodiment, preferably, the error matrix in step S4 is defined as follows:

[0118] E k =y * -∑ j≠k d j x j

[0119] Where, d j is the jth atom in the dictionary, x j Represents the j-th coefficient of the sparse representation coefficient matrix.

[0120] According to this embodiment, preferably, the specific steps of gradually approximating in steps to obtain the optimal solution in step S4, that is, finding the global optimal dictionary, are as follows:

[0121] (1) Solve the sparse representation coefficient matrix X based on the fixed dictionary σ;

[0122] (2) Update the dictionary atoms column by column according to the solved sparse representation coefficient matrix, and slowly include the representation error into the original dictionary while satisfying the sparsity, that is, update and optimize row by row and column by column to reduce the overall error. The objective function of updating the dictionary is:

[0123]

[0124] Among them, d k is the kth atom in the dictionary, represents the two-norm, For missing d in the dictionary k and the sparse representation coefficient matrix is ​​missing x k The sparse representation coefficient after E k For missing d in the dictionary k and the sparse representation coefficient matrix is ​​missing x kThe sparse representation residual after ;

[0125] (3) Repeat steps (1) and (2) until the specified number of iterations j is reached to obtain the optimal feature matrix and its sparse representation coefficient matrix From this, A and W can be calculated.

[0126] Figure 2 is a loss function diagram of this embodiment under different noise conditions, where Figure 2 (a) is the loss function diagram when the noise level is 10dB. Figure 2 (b) is the loss function diagram when the noise level is 5dB. Figure 2 (c) is the loss function diagram when the noise level is 0dB. Figure 2 (d) is the loss function diagram when the noise level is -5dB. Figure 2 (e) Loss function diagram when the noise level is -10 dB; it can be seen from the figure that the present invention can achieve fast convergence within five iterations in this embodiment, and j=10 in this embodiment.

[0127] According to this embodiment, preferably, the classification process in step S5 first obtains D from step S4. *(j) , A, and W are normalized to obtain:

[0128]

[0129] in, Respectively represent D *(j) , A, W are normalized and calculated. K is the Kth column of matrix A, for a K Normalized result, d K is the matrix D *(j) The Kth column of is d K Normalized result, w K is the Kth column of matrix W, w K Normalized result, ||·||2 represents the second norm.

[0130] According to this embodiment, preferably, in step S5, in order to obtain the test set obtained in step S1 The sparse representation coefficient matrix x test , the optimization problem to be solved is as follows:

[0131]

[0132] According to this embodiment, preferably, the discrimination and classification of the test data in step S5 is achieved by the following objective function:

[0133]

[0134] Here, argmax(·) represents the combination of independent variable values ​​that maximizes the function value, and l is the coefficient x with sparse representation. test Linear classifier Generated exist The discriminant vector on . Excluding the index equal to zero, the index with the largest value in l is identified as The corresponding category is determined to achieve discriminative classification.

[0135] This example tests the effect of different dictionary sizes on classification accuracy and operation time under noise conditions and compares it with the traditional method, that is, the original LC-KSVD and KSVD classification algorithms. The dictionary size is changed from 70, 140, 210 to 630 in sequence. The results are as follows: Figure 3 As shown, Figure 3 (a) is a comparison chart showing the effect of dictionary size on classification accuracy and operation time when the noise level is 10dB. Figure 3 (b) is a comparison chart showing the effect of dictionary size on classification accuracy and operation time when the noise level is 5dB. Figure 3 (c) is a comparison chart showing the effect of dictionary size on classification accuracy and operation time when the noise level is 0dB. Figure 3 (d) is a comparison chart showing the effect of dictionary size on classification accuracy and operation time when the noise level is -5dB. Figure 3 (e) is a comparison diagram of the influence of dictionary size on classification accuracy and operation time when the noise level is -10dB. It can be seen from the figure that in the original LC-KSVD and KSVD classification algorithms, the accuracy will gradually increase with the dictionary size, while the method of the present invention can still maintain a high accuracy when the dictionary is small. In addition, in the original LC-KSVD and KSVD classification algorithms, the size of the dictionary will affect the operation time, which is specifically reflected in that as the dictionary increases, the operation time will also increase, while the method of the present invention can still maintain a stable operation time when the dictionary increases, and will not increase with the increase of the dictionary. That is, compared with the two traditional methods, the present invention is less sensitive to changes in dictionary size under various noise levels, and is also superior in terms of the stability and accuracy of the diagnostic results. In terms of computational efficiency, the present invention is also substantially superior to the two traditional methods.

[0136] In this embodiment, a=10 is selected, that is, the dictionary size is 70, and the classification accuracy comparison diagram of the machine learning method, specifically SVM, decision tree algorithm and random forest classification algorithm, under different noise conditions is shown in the figure below: Figure 4 As shown, the noise levels are 10dB, 5dB, 0dB, -5dB and -10dB. From the analysis of the figure, it can be seen that the classification accuracy of the present invention is higher under various noise levels.

[0137] In order to further verify the effectiveness and practicality of the bearing composite fault sparse classification method and system proposed in the present invention based on adaptive feature selection, the composite fault detection experiment is further verified using the operating data of the test bench. The two test bench structures used are as follows: Figure 5 As shown in the figure, test bench 1 is a SKF6205 deep groove ball bearing, and 7 bearing composite fault operation modes are set, namely outer ring fault, inner ring fault, rolling element fault, inner and outer ring fault, inner ring rolling element fault, outer ring rolling element fault and normal mode. Test bench 1 device includes motor, encoder, test bearing, buffer device, loading bolt, signal acquisition instrument and other devices, such as Figure 5 (a) shows the relevant characteristics and experimental parameters of the bearings used are shown in Table 4. Test bench 2 is a FGA30206 tapered roller bearing with the same seven bearing operating conditions as test bench 1. It consists of a motor, a support frame, a photoelectric sensor, an acceleration sensor, a gearbox and other components. Test bench 2 is as follows Figure 5 (b), the bearing specifications and experimental parameters used are shown in Table 4. Both the experimental training set and the test set consist of a data set of size 16×700.

[0138] Table 4 Hardware platform configuration

[0139]

[0140] Therefore, through the above steps, the sparse classification method for bearing composite faults using adaptive feature selection described in this application has higher accuracy and faster calculation speed in composite fault classification, especially when the dictionary size is small. In addition, it shows strong robustness even in the case of high levels of noise. Preliminary verification through simulation experiments has confirmed that this method is suitable for composite fault classification of bearings, can monitor process anomalies faster, and provides a new method for condition monitoring.

[0141] The effects of different dictionary sizes on classification accuracy and operation time in the present invention and the original LC-KSVD, LC-KSVD+Score with adaptive feature selection, LC-KSVD+Score+Class with class dictionary construction, and LC-KSVD+Score+L1 with a norm sparse constraint term are compared respectively, and the dictionary size is changed from 70, 140, 210 to 630 in sequence. The results are as follows Figure 6 As shown. Figure 6(a) is a comparison chart of the classification accuracy and operation time of the test bench 1 data for various improvement measures in the present invention, Figure 6 (b) is a comparison chart of the classification accuracy and operation time of the test bench 2 data for various improvement measures in the present invention. Figure 6 As shown in (a), the diagnostic accuracy trends of several methods remain consistent, which means that a large dictionary can bring higher accuracy. However, under actual working conditions, it is very difficult to obtain sufficient computing time and samples for real-time diagnosis. The traditional LC-KSVD classification accuracy is low and fluctuates greatly with changes in the dictionary size. The present invention adds adaptive feature selection to improve the accuracy. In addition, since the present invention adds standard constraints and single-class feature dictionary construction, the diagnostic accuracy is also greatly improved. Even with a smaller dictionary size, the present invention can achieve a higher diagnostic accuracy with less overall fluctuation. As shown in Figure 6 As shown in (b), based on the diagnostic calculation time of the two sets of experimental data, it can be seen that increasing the dictionary size will lead to a gradual increase in calculation time and a decrease in calculation efficiency. However, the overall calculation time of the method proposed in this paper is relatively short and remains stable, and is less affected by changes in dictionary size.

[0142] Example 2

[0143] A system for implementing the sparse classification method for bearing composite faults by adaptive feature selection described in Example 1, characterized in that it includes a feature selection module, a model building module, a dictionary initialization module, a dictionary training module and a classification module;

[0144] The feature selection module is used to classify bearing composite fault signal samples into i categories, calculate the intra-class Euclidean distance and inter-class Euclidean distance feature indicators of bearing fault signal samples, enhance sample features through an adaptive feature selection algorithm, remove redundant feature sets, implement adaptive feature selection of fault signals through the sorting of the adaptive feature selection algorithm, and construct an initial training set and test set of bearing fault signal samples of each category;

[0145] The model building module is used to build the LC-KSVD label consistency singular value decomposition algorithm model, adding the l1 norm constraint term to obtain a new objective function;

[0146] The dictionary initialization module is used to construct an initial dictionary D through the training set obtained by the feature selection module. *(0) , and then obtain the initial linear transformation matrix A (0) and the classifier parameter W (0) ;

[0147] The dictionary training module is used to train the initial dictionary D using the test set obtained by the feature selection module according to the objective function obtained by the model building module. *(0)Perform iterative training until the specified number of iterations j is reached and the feature dictionary D is obtained *(j) And the corresponding linear transformation matrix A and classifier parameters W;

[0148] The classification module is used to obtain the feature dictionary D based on the dictionary training module. *(j) , and obtain the sparse representation coefficient matrix. The discriminant vector l is obtained through the sparse representation coefficient matrix, and the bearing composite fault signal is discriminated and classified through the index with the largest value in the discriminant vector l.

[0149] In summary, the present invention provides a sparse classification method and system for bearing composite faults with adaptive feature selection, including the following steps: first, adaptive fault feature screening is achieved by enhancing sample features through intra-category and inter-category distance metrics and removing redundant feature sets, thereby effectively improving fault identification performance; a single-class fault initialization dictionary is constructed using a feature training set, and sparsity is enhanced and model convexity is maintained through a norm constraint; the corresponding sample set of each fault category is used to iteratively train an overcomplete dictionary, and finally the discriminant classification of bearing composite faults is achieved through a discriminant vector. The experimental results show that compared with the traditional method, the proposed method can effectively solve the feature adaptive selection problem in the LC-KSVD algorithm in simulation and experiment, improve the accuracy and robustness of fault classification under small dictionary conditions and strong noise conditions, such as -10dB conditions, and reduce time costs. Finally, the present invention uses two sets of test bench data collection as an example to verify the effectiveness of the proposed method.

[0150] The above is a detailed introduction to the sparse classification method of bearing composite faults with adaptive feature selection provided by the present invention. This article uses specific embodiments to briefly explain the basic principles and implementation methods of the present invention, but the protection scope of the present invention is not limited to this. For ordinary technicians in this field, it can be understood that without departing from the principles and spirit of the present invention, various changes, modifications, substitutions and variations can be made to the embodiments in this article. The scope of the present invention is defined by the attached claims and their equivalents.

[0151] It should be understood that although this specification is described according to various embodiments, not every embodiment contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0152] The series of detailed descriptions listed above are only specific descriptions of feasible embodiments of the present invention. They are not intended to limit the scope of protection of the present invention. All equivalent embodiments or changes that do not deviate from the technical spirit of the present invention should be included in the scope of protection of the present invention.

Claims

1. A sparse classification method for bearing composite faults based on adaptive feature selection, characterized in that: The following steps are involved: Step S1, adaptive feature selection: divide the bearing composite fault signal samples into i categories, calculate the intra-class Euclidean distance and inter-class Euclidean distance feature indicators of the bearing fault signal samples, enhance the sample features through an adaptive feature selection algorithm, remove redundant feature sets, implement adaptive feature selection on the fault signals through the sorting of the adaptive feature selection algorithm, and construct an initial training set and test set of bearing fault signal samples of each category; Step S2, model construction: constructing the LC-KSVD label consistency singular value decomposition algorithm model, adding the l1 norm constraint term to obtain a new objective function; Step S3, dictionary initialization: construct the initial dictionary D using the training set obtained in step S1 *(0) , and then obtain the initial linear transformation matrix A (0) and the classifier parameter W (0) ; Step S4, dictionary training: Based on the objective function obtained in step S2, the initial dictionary D is trained using the test set obtained in step S1. *(0) Perform iterative training until the specified number of iterations j is reached and the feature dictionary D is obtained *(j) And the corresponding linear transformation matrix A and classifier parameters W; Step S5, classification stage: based on the feature dictionary D obtained in step S4 *(j) , and obtain the sparse representation coefficient matrix. The discriminant vector l is obtained through the sparse representation coefficient matrix, and the bearing composite fault signal is discriminated and classified through the index with the largest value in the discriminant vector l.

2. The bearing composite fault sparse classification method based on adaptive feature selection according to claim 1 is characterized in that: The specific steps of step S1 include: Step S11: Calculate and maximize the intra-class Euclidean distance S1 of the fault signal. The calculation formula of the intra-class Euclidean distance S1 is as follows: Among them, Md(·) represents the median, c is the Euclidean distance, a m 、a n It is the characteristic data extracted from the bearing composite fault signal; Step S12: Calculate and minimize the inter-category Euclidean distance S2. The calculation formula of the inter-category Euclidean distance S2 is as follows: Among them, QU(·) represents the upper quartile, QL(·) represents the lower quartile, and a i is the characteristic data of the i-th type bearing fault signal; Step S13: Using the intra-class Euclidean distance S1 and inter-class Euclidean distance S2 obtained in steps S11 and S12, the eigenvalues ​​S of the bearing fault signals are calculated and sorted using an adaptive feature selection algorithm. The feature data sets with higher eigenvalue S scores are retained, and the feature data sets with lower scores are deleted. The retained feature data sets are used to construct a new training set. With the test set 3. The bearing composite fault sparse classification method based on adaptive feature selection according to claim 1 is characterized in that: The objective function of the adaptive feature selection algorithm in step S1 is as follows: Among them, S represents the eigenvalue; S1 is the Euclidean distance within a class; S2 is the Euclidean distance between categories; MAX(·) means finding the maximum value, and MIN(·) means finding the minimum value.

4. The bearing composite fault sparse classification method based on adaptive feature selection according to claim 1 is characterized in that: The objective function of the LC-KSVD label consistency singular value decomposition algorithm model in step S2<D,W,A,X> for: Here, argmin(·) means finding the combination of independent variable values ​​that minimizes the function value, X = [x1,…,x N ] represents the sparse representation coefficient, x z is the Zth column of X (Z = 0, 1, ..., N), represents the two-norm, is the i-type test data set, represents a set of real numbers, K is the number of columns of D, N is the number of columns of Y and K>N, D is an overcomplete dictionary, α represents the weights of the reconstruction term and the label consistency term, refers to the error of sparse representation classification, Q represents the sparse classification of the input signal, represents the linear transformation matrix, β represents the weight of the reconstruction error term, is the reconstruction error term, Represents the classifier parameters, label matrix represents the classification label assigned to the input signal, ||x z ||0 represents zero norm, T represents coefficient constraint, and st represents constraint.

5. The bearing composite fault sparse classification method based on adaptive feature selection according to claim 4 is characterized in that: In step S2, the l1 norm constraint term is added to obtain a new objective function <σ,X>: Among them, ||·||1 represents a norm, y * is the training set, τ i It is a threshold parameter used to adjust the output sparsity.

6. The bearing composite fault sparse classification method based on adaptive feature selection according to claim 1 is characterized in that: The initial linear transformation matrix A in step S3 (0) and the classifier parameter W (0) as follows: A (0) =QX *T (X * X *T +λI) -1 W (0) =HX *T (X * X *T +λI) -1 Among them, Q represents the sparse classification of the input signal, X * To solve the sparse coefficients for each initial feature dictionary, we obtain the sparse representation coefficient matrix, X * =[X1,…,X i ], λ is the unit parameter, I is the unit matrix, and the label matrix represents the classification label assigned to the input signal, Represents the set of real numbers.

7. The bearing composite fault sparse classification method based on adaptive feature selection according to claim 1 is characterized in that: The specific steps of step S4 include: Step S41: solving the sparse representation coefficient matrix X according to the fixed dictionary σ of the new objective function in step S2; Step S42: Update dictionary atoms column by column according to the solved sparse representation coefficient matrix, and slowly include the representation error into the original dictionary while satisfying the sparsity, that is, update and optimize row by row and column by column to reduce the overall error. The objective function of updating the dictionary is: Among them, d k is the kth atom in the dictionary, represents the two-norm, For missing d in the dictionary k and the sparse representation coefficient matrix is ​​missing x k The sparse representation coefficient after E k For missing d in the dictionary k and the sparse representation coefficient matrix is ​​missing x k The sparse representation residual after ; Step S43: Repeat steps S41 and S42 until the specified number of iterations j is reached to obtain the optimal feature matrix and its sparse representation coefficient matrix From this, the linear transformation matrix A and classifier parameters W can be calculated.

8. The bearing composite fault sparse classification method based on adaptive feature selection according to claim 7 is characterized in that: In step S42, the sparse representation residual E k for: E k =y * -∑ j≠k d j x j Where, d j is the jth atom in the dictionary, x j represents the jth coefficient of the sparse representation coefficient matrix, y * For the training set.

9. The bearing composite fault sparse classification method based on adaptive feature selection according to claim 1 is characterized in that: The specific steps of step S5 include: Step S51: Dictionary D obtained in step S4 *(j) And the corresponding linear transformation matrix A and classifier parameters W are normalized: in, Respectively represent D *(j) , A, W are normalized and calculated. K is the Kth column of matrix A, for a K Normalized result, d K is the matrix D *(j) The Kth column of is d K Normalized result, w K is the Kth column of matrix W, w K Normalized result, ||·||2 represents the second norm; Step S52: According to the normalization result of step S51, solve the test set as follows: The sparse representation coefficient matrix x test : Here, argmin(·) means finding the combination of independent variable values ​​that minimizes the function value. is the test set, represents the two norm, st represents the constraint, X=[x1,…,x N ] represents the sparse representation coefficient, x z is the Zth column of X (Z=0,1,…,N), τ i is a threshold parameter used to adjust the output sparsity, and ||·||1 represents a norm; Step S53: Solve the discriminant vector l as follows, and identify the index with the largest value in l as The corresponding category is used to achieve discriminative classification: Here, argmax(·) represents the combination of independent variable values ​​that maximizes the function value, and l is the coefficient x with sparse representation. test Linear classifier Generated exist The discriminant vector on .

10. A system for implementing the sparse classification method for bearing composite faults by adaptive feature selection as described in any one of claims 1 to 9, characterized in that: It includes feature selection module, model building module, dictionary initialization module, dictionary training module and classification module; The feature selection module is used to classify bearing composite fault signal samples into i categories, calculate the intra-class Euclidean distance and inter-class Euclidean distance feature indicators of bearing fault signal samples, enhance sample features through an adaptive feature selection algorithm, remove redundant feature sets, implement adaptive feature selection of fault signals through the sorting of the adaptive feature selection algorithm, and construct an initial training set and test set of bearing fault signal samples of each category; The model building module is used to build the LC-KSVD label consistency singular value decomposition algorithm model, adding the l1 norm constraint term to obtain a new objective function; The dictionary initialization module is used to construct an initial dictionary D through the training set obtained by the feature selection module. *(0) , and then obtain the initial linear transformation matrix A (0) and the classifier parameter W (0) ; The dictionary training module is used to train the initial dictionary D using the test set obtained by the feature selection module according to the objective function obtained by the model building module. *(0) Perform iterative training until the specified number of iterations j is reached and the feature dictionary D is obtained *(j) And the corresponding linear transformation matrix A and classifier parameters W; The classification module is used to obtain the feature dictionary D based on the dictionary training module. *(j) , and obtain the sparse representation coefficient matrix. The discriminant vector l is obtained through the sparse representation coefficient matrix, and the bearing composite fault signal is discriminated and classified through the index with the largest value in the discriminant vector l.

Images