High-speed train traction motor bearing fault diagnosis method

Abstract

The invention discloses a high-speed train traction motor bearing fault diagnosis method. The method comprises the following steps: step 1, acquiring state data of a high-speed train traction motor bearing; 2, performing multi-level decomposition on the preprocessed state data according to a dynamic decomposition rule through a Deep-PCA method to obtain a plurality of sub-data sets; 3, projecting the sub-data set to a high-dimensional space through a KICA method; 4, performing double-space feature extraction on the subset data in the high-dimensional space through an ICA-PCA method, and decomposing each sub-data set into a Gaussian part and a non-Gaussian part; and 5, respectively calculating the statistical magnitude of the Gaussian part and the non-Gaussian part, and judging the occurrence of a fault according to a preset threshold value: when the statistical magnitude of a sub-data set exceeds the threshold value, determining that the fault occurs. According to the method, the sensitivity, noise immunity and adaptability of fault detection are improved, and the problem that the detection effect of a traditional method on non-Gaussian, nonlinear and strong noise data is not ideal is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a KICA-DPCA-based high-speed train traction motor bearing fault diagnosis method. Background Art

[0002] High-speed trains, thanks to their convenience, comfort, and efficiency, have become a major means of transportation. The electric traction system, as the core of a high-speed train, provides power for the entire train. The traction motor is a key component of the high-speed train traction system, converting electrical energy into mechanical energy. Among all motor failure types, bearing failures have the highest incidence, accounting for 40% to 50% of all motor failures. If a bearing failure is not detected and addressed promptly, the problem will worsen over time, becoming a serious failure that can affect the motor's normal operation and even lead to the collapse of the traction system. Therefore, fault detection for traction motor bearings is of great research value.

[0003] Existing data-driven methods for high-speed train traction motor bearing fault diagnosis rarely comprehensively consider the characteristics of all data dimensions. The traction system is a complex electromechanical system, and the data collected by sensors in various components often has non-Gaussian, nonlinear, and highly noisy characteristics. Ignoring these characteristics or only considering a subset of them can severely impact fault detection accuracy. Therefore, fault diagnosis of high-speed train traction motor bearings based on the different data characteristics has become a research hotspot.

[0004] Currently, PCA-based methods are the most widely used among existing motor bearing fault detection methods. However, this method can only process data with a Gaussian distribution. While ICA can handle non-Gaussian data, its target data type is relatively limited. When encountering signals with strong noise, the noise data can obscure the fault information in the original data, making it difficult to extract fault information features. Especially for nonlinear data, the fault features extracted using these linear detection methods will be significantly offset, ultimately resulting in suboptimal overall fault detection performance. Summary of the Invention

[0005] In view of the deficiencies in the prior art, the present invention provides a method for diagnosing faults of traction motor bearings of high-speed trains to solve the technical problem that the prior art cannot well process traction motor bearing status data with non-Gaussian, nonlinear and strong noise data characteristics.

[0006] The present invention provides a method for diagnosing a bearing fault of a high-speed train traction motor, comprising the following steps:

[0007] Step 1: Obtain the status data of the high-speed train traction motor bearing and pre-process the status data;

[0008] Step 2: Use the Deep-PCA method to perform multi-level decomposition on the pre-processed state data according to the dynamic decomposition rules to obtain several sub-datasets;

[0009] Step 3: Project the sub-dataset into a high-dimensional space using the KICA method;

[0010] Step 4: Use the ICA-PCA method to extract dual-space features from the subset data in the high-dimensional space, and decompose each subset data set into Gaussian and non-Gaussian parts;

[0011] Step 5: Calculate the statistics of the Gaussian part and the non-Gaussian part respectively, and judge the occurrence of the fault based on the preset threshold: when the statistics of a sub-data set exceeds the threshold, a fault occurs.

[0012] Furthermore, in step 1, preprocessing the status data includes:

[0013] The state data is converted into a data matrix, and the mean and standard deviation of each column vector are calculated, and then standardized.

[0014] Furthermore, in step 2, the specific process of multi-level decomposition is as follows:

[0015] Step 21: Use the pre-processed state data as the data to be decomposed;

[0016] Step 22: Obtain the covariance matrix of the data to be decomposed, and perform singular value decomposition on the covariance matrix to obtain the singular value vectors. Decompose the data to be decomposed according to the singular value vectors to obtain several sub-data;

[0017] Step 23: Repeat the decomposition process of step 22 with the sub-data as the data to be decomposed until the decomposition times are reached to obtain the sub-data set:

[0018]

[0019] Where, P j,k is the kth load matrix after the data matrix is ​​decomposed j times; k is the number of principal elements.

[0020] Furthermore, in step 3, the KICA method projects the sub-dataset into a high-dimensional space through a Gaussian kernel function and standardizes the kernel matrix.

[0021] Furthermore, in step 4, the specific process of performing dual-space feature extraction by the ICA-PCA method is as follows:

[0022] Step 41: Construct a whitening matrix of the sub-dataset;

[0023] Step 42: Perform ICA decomposition on the whitened matrix to obtain key independent components and residual parts, and take the key independent components as the non-Gaussian part;

[0024] Step 43: First, perform non-normalization processing on the residual part, and then perform PCA decomposition to obtain the principal component and residual, and use the principal component as the Gaussian part.

[0025] Furthermore, in step 5, the statistics include: squared prediction error statistics and Hotelling T-squared statistics, and the specific formula is:

[0026]

[0027] Where n is the number of samples; is the normalized kernel matrix; V j,k is the eigenvector matrix; j,k is the eigenvalue matrix; is the whitening matrix after extracting independent components; t j,k is the sample mean vector.

[0028] Furthermore, the preset threshold of the squared prediction error statistic is:

[0029]

[0030] Furthermore, the preset threshold of the Hotelling T-squared statistic is:

[0031]

[0032] Where q j,k For Gaussian data The number of principal components for principal component analysis; α is the significance level, 1-α is the confidence level, then It obeys the F distribution with the first degree of freedom being α and the second degree of freedom being n-α.

[0033] Beneficial effects of the present invention:

[0034] The present invention significantly improves the sensitivity, noise resistance and adaptability of fault detection through the collaborative optimization of deep decomposition and kernel space mapping, combined with dual-space feature separation and partial statistics monitoring mechanism, and effectively solves the problem that traditional methods have unsatisfactory detection effects on non-Gaussian, nonlinear and strong noise data.

[0035] The present invention adopts dynamic decomposition rules combined with Deep-PCA, and can extract fault information that cannot be identified by PCA by performing multi-layer iterative decomposition on the data. This has obvious advantages in processing strong noise information.

[0036] The present invention connects the output data of the Deep-PCA method to KICA, forming a deep combination of KICA and Deep-PCA. On this basis, KICA further extracts nonlinear features and enhances the sensitivity of fault detection.

[0037] This method performs ICA-PCA dual-space feature separation on sub-datasets to obtain Gaussian and non-Gaussian components, enabling simultaneous monitoring of systematic deviations and sudden failures. Furthermore, the method independently designs statistics and thresholds for the two components, allowing for independent judgments. This dual-threshold determination provides dual protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The features and advantages of the present invention will be more clearly understood by referring to the accompanying drawings, which are schematic and should not be construed as limiting the present invention in any way. In the accompanying drawings:

[0039] Figure 1 is a flow chart of a specific embodiment of the present invention;

[0040] Figure 2 is a flowchart of Deep-PCA decomposition in a specific embodiment of the present invention;

[0041] Figure 3 This is a flow chart of dual-space decomposition in a specific embodiment of the present invention;

[0042] Figure 4 This is a flowchart of fault detection in a specific embodiment of the present invention;

[0043] Figure 5 is the fault detection result of the TDCS-FIB data set in a specific embodiment of the present invention;

[0044] Figure 6 This is the fault detection result of the Western Reserve University dataset in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0045] To make the purpose, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0046] The present invention will be further described below with reference to specific examples. Those skilled in the art will appreciate that these examples are intended only to illustrate the present invention and are not intended to limit the scope of the present invention, and that modifications to various equivalent forms of the present invention fall within the scope defined by the appended claims.

[0047] like Figure 1 As shown, the present invention provides a high-speed train traction motor bearing fault diagnosis method, comprising the following steps:

[0048] Step 1: Obtain the status data of the high-speed train traction motor bearings. To reduce the impact of data of different magnitudes on fault detection accuracy, standardize the data to the same scale.

[0049] Taking the traction motor bearing as the research object, the bearing data set X∈R is obtained from the sensor. m×n , where m is the number of variables and n is the number of samples; to eliminate the influence of magnitude and analyze the data on the same scale, the data matrix is ​​standardized:

[0050] Calculate the data matrix X m×n The mean and standard deviation of each column vector

[0051]

[0052] Where, is the average value of the data in the jth column; S j is the standard deviation of the data in the jth column;

[0053] Get the standardized data matrix X:

[0054]

[0055] Step 2: Use Deep-PCA method to perform multi-level decomposition on the pre-processed state data according to the dynamic decomposition rules to obtain several sub-datasets, such as Figure 2 As shown, the specific process is:

[0056] Step 21: Use the pre-processed state data as the data to be decomposed;

[0057] Step 22: Obtain the covariance matrix of the data to be decomposed, and perform singular value decomposition on the covariance matrix to obtain the singular value vectors. Decompose the data to be decomposed according to the singular value vectors to obtain several sub-data;

[0058] Among them, the covariance matrix of the decomposed data is subjected to singular value decomposition, and the singular value vector obtained is:

[0059]

[0060] Where X is the data to be decomposed; m represents the number of sensors; Λ 0,1 ∈R m×m , Λ 0,1 =diag(λ 0,1 ,...λ 0,m), λ 0,m represents the mth singular value of S; P1∈R m×m is the singular value vector of S. According to the number of main elements, P1 can be decomposed into P 1，1 and P 1，2 , that is, P1=[P 1,1 ,P 1,2 ];

[0061] Decompose the data to be decomposed according to the singular value vector to obtain several sub-data:

[0062] Decompose the data to be decomposed X into two parts X 1.1 and X 1.2 ,Right now:

[0063] X=X 1.1 +X 1.2

[0064]

[0065] Where I is the identity matrix;

[0066] Step 23: Repeat the decomposition process of step 22 with the sub-data as the data to be decomposed until the decomposition times are reached.

[0067] The decomposed X 1.1 and X 1.2 As the new X, calculate X separately 1.1 and X 1.2 The covariance matrix S 1.1 and S 1.2 , and the covariance matrix S 1.1 and S 1.2 Perform singular value decomposition and transform X 1.1 and X 1.2 It is divided into two parts:

[0068] X 1.1 =X 2.1 +X 2.2

[0069] X 1.2 =X 2.3 +X 2.4

[0070] X=X 2.1 +X 2.2 +X 2.3 +X 2.4

[0071] Repeat the above steps and decompose the data X j times to get 2 j non-overlapping subsets, then the k-th dataset in layer j is:

[0072]

[0073] Where, P j,k is the kth load matrix after the data matrix is ​​decomposed j times; k is the number of principal elements.

[0074] Step 3: Project the sub-dataset into a high-dimensional space using the KICA method;

[0075] In order to process nonlinear data with linear methods, the Gaussian kernel function is used to project the data into a high-dimensional space, where the Gaussian kernel function is:

[0076]

[0077] Among them, x i and x r is the i-th row and r-th column vector of matrix X.

[0078] The mapping relationship between the original space and the high-dimensional space can be expressed as:

[0079] Φ i =Φ(X i )

[0080] Finally, the projected kernel matrix is ​​calculated as:

[0081] k ir =Φ(x i ) T Φ(x r )=Φ(x i )·Φ(x r )=k(x i ,x r )

[0082] Center the kernel matrix:

[0083]

[0084] in,

[0085] Normalize the kernel matrix to the same scale to obtain for:

[0086]

[0087] Step 4: Use the ICA-PCA method to extract dual-space features from the subset data in the high-dimensional space, and decompose each subset data set into Gaussian and non-Gaussian parts;

[0088] like Figure 3 As shown in Figure 2, the specific steps of dual-space feature extraction are:

[0089] Step 41: Right Perform eigenvalue decomposition to obtain the first u largest eigenvalues ​​λ1≥λ2≥λ3≥…≥λ u and the corresponding eigenvectors v1v2v3…v u Among them, Λ j,k =diag(λ1,λ2,...,λ u ), V j,k =(v1,v2,...,v u );

[0090] Calculate the covariance matrix of the sub-dataset The eigenvectors of are:

[0091]

[0092] Where, Can be get.

[0093] Reduce the eigenvectors to normalized form and obtain the whitening matrix:

[0094]

[0095] The whitening matrix Z j,k Normalize and get its vector expression:

[0096]

[0097] Step 42: Perform ICA decomposition on the whitened matrix to obtain the key independent components and the residual part, and take the key independent components as the non-Gaussian part:

[0098]

[0099] Where, represents a non-Gaussian signal; represents the residual part;

[0100] Step 43: Due to It has been normalized to have unit variance and zero mean and cannot be directly applied to principal component analysis, so the residual part is first denormalized:

[0101]

[0102] Then perform PCA decomposition to obtain the principal components and residuals:

[0103]

[0104] Take the principal components as Gaussian parts.

[0105] After retaining p independent components of the non-Gaussian part and not normalizing it, we can get:

[0106]

[0107] Then the whitening matrix after extracting independent components becomes:

[0108]

[0109] like Figure 4 As shown, step 5: calculate the statistics of the Gaussian part and the non-Gaussian part respectively, and judge the occurrence of the fault according to the preset threshold: when the statistics of a sub-data set exceeds the threshold, a fault occurs.

[0110] Statistics include: squared prediction error statistic and Hotelling T squared statistic, the specific formula is:

[0111]

[0112] The preset threshold value of Hotelling's T-squared statistic is:

[0113]

[0114] Using the kernel density estimation method, the threshold of the statistic SPE is calculated as:

[0115]

[0116] The method of the present invention is simulated and verified below:

[0117] Step A1: Use the TDCS-FIB platform to simulate train operation. The simulation time is 3 seconds, and the motor bearing fault is injected at 1 second. Select 6 sensors in the high-speed train traction system (transformer 3-phase current I sa , I sb , I sc , train speed S p , motor speed R s , motor electromagnetic torque T e ) obtain the original data set;

[0118] Step A2: Use Matlab to write a program to implement the content of the present invention;

[0119] Step A3: Using Matlab / TDCS-FIB co-simulation, the data generated in step 1 is imported into the Matlab program for fault detection;

[0120] Step A4: Using the Western Reserve University bearing dataset, further simulation is performed to verify the effect of fault detection.

[0121] Figure 5is the detection result of the TDCS-FIB dataset, which includes T 2 The curves of the two statistics of and SPE, the dotted line represents the threshold. In order to reduce the redundancy of calculation and improve the efficiency of fault detection, only consider expanding the data set X to the second order and selecting the statistics corresponding to some sub-data sets SPE 3，2 、SPE 3，3 .

[0122] From the four sub-graphs, we can see that at the fault setting point, the statistics quickly exceed the threshold. 2 The detection effect is good, no false positives occur for normal data, and the detection effect of fault data is obvious. The method of the present invention improves the sensitivity of SPE to faults, so that when a fault occurs, the fault can be easily detected through SPE statistics.

[0123] Figure 6 This is the test result of the bearing data set of Western Reserve University. This data is measured data and can be used as a good supplement to the simulation data. The statistics selected are the same as the TDCS-FIB data. SPE 3，2 、SPE 3，3 Where T 2 Statistics can detect periodic failures of bearings, and T 2 The statistic can detect normal data without error, and the detection results at the fault point are very obvious. The SPE statistic is not sensitive to this data set, but as long as SPE and T 2 If there is a statistic that can effectively detect a fault, it means that a fault has occurred, so the method of the present invention is still effective for the bearing data detection of Western Reserve.

[0124] By the above attachment Figure 5 、 6 It can be seen that the method of the present invention can effectively realize the fault detection of traction motor bearings of high-speed trains, and effectively solves the diagnosis and engineering application problems of motor bearing faults in electric traction systems, which is of great significance to the safe operation of high-speed trains.

[0125] Although the embodiments of the present invention have been described with reference to the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention. Such modifications and variations are all within the scope defined by the appended claims.

Claims

1. A method for diagnosing bearing faults of a high-speed train traction motor, characterized in that: The steps include: Step 1: Obtain the status data of the high-speed train traction motor bearing and pre-process the status data; Step 2: Use the Deep-PCA method to perform multi-level decomposition on the pre-processed state data according to the dynamic decomposition rules to obtain several sub-datasets; Step 3: Project the sub-dataset into a high-dimensional space using the KICA method; Step 4: Use the ICA-PCA method to extract dual-space features from the subset data in the high-dimensional space, and decompose each subset data set into Gaussian and non-Gaussian parts; Step 5: Calculate the statistics of the Gaussian part and the non-Gaussian part respectively, and judge the occurrence of the fault based on the preset threshold: when the statistics of a sub-data set exceeds the threshold, a fault occurs.

2. The high-speed train traction motor bearing fault diagnosis method according to claim 1, characterized in that: In step 1, preprocessing the status data includes: The state data is converted into a data matrix, and the mean and standard deviation of each column vector are calculated, and then standardized.

3. The high-speed train traction motor bearing fault diagnosis method according to claim 1, characterized in that: In step 2, the specific process of multi-level decomposition is: Step 21: Use the pre-processed state data as the data to be decomposed; Step 22: Obtain the covariance matrix of the data to be decomposed, and perform singular value decomposition on the covariance matrix to obtain the singular value vectors. Decompose the data to be decomposed according to the singular value vectors to obtain several sub-data; Step 23: Repeat the decomposition process of step 22 with the sub-data as the data to be decomposed until the decomposition times are reached to obtain the sub-data set: Where, P j,k is the kth load matrix after the data matrix is ​​decomposed j times; k is the number of principal elements; I is the identity matrix.

4. The high-speed train traction motor bearing fault diagnosis method according to claim 1, characterized in that: In step 3, the KICA method projects the sub-dataset into a high-dimensional space through a Gaussian kernel function and standardizes the kernel matrix.

5. The high-speed train traction motor bearing fault diagnosis method according to claim 1, characterized in that: In step 4, the specific process of performing dual-space feature extraction by the ICA-PCA method is as follows: Step 41: Construct a whitening matrix of the sub-dataset; Step 42: Perform ICA decomposition on the whitened matrix to obtain key independent components and residual parts, and take the key independent components as the non-Gaussian part; Step 43: First, perform non-normalization processing on the residual part, and then perform PCA decomposition to obtain the principal component and residual, and use the principal component as the Gaussian part.

6. The high-speed train traction motor bearing fault diagnosis method according to claim 1, characterized in that: In step 5, the statistics include: squared prediction error statistics and Hotelling squared statistics, and the specific formula is: Where n is the number of samples; is the normalized kernel matrix; V j,k is the eigenvector matrix; j,k is the eigenvalue matrix; is the whitening matrix after extracting independent components; t j,k is the sample mean vector.

7. The high-speed train traction motor bearing fault diagnosis method according to claim 6, characterized in that: The preset threshold of the squared prediction error statistic is:

8. The high-speed train traction motor bearing fault diagnosis method according to claim 6 or 7, characterized in that: The preset threshold value of the Hotelling T-squared statistic is: Where q j,k For Gaussian data The number of principal components for principal component analysis; α is the significance level, 1-α is the confidence level, then It obeys the F distribution with the first degree of freedom being α and the second degree of freedom being n-α.

Images