Train bearing fault diagnosis method based on improved joint distribution adaptation

Abstract

The application provides a train bearing fault diagnosis method based on improved joint distribution adaptation, and the method comprises the following steps: considering the fault mechanism of the bearing and the requirements of the machine learning algorithm on the training samples, extracting common fault feature components from three fields of time domain, frequency domain and entropy domain; adopting a method of quantitative cluster analysis and correlation analysis, and constructing an effective and comprehensive feature vector group as the input of the bearing fault diagnosis algorithm; adopting a K nearest neighbor algorithm based on weighted chi-square distance as an embedded classifier of the joint distribution adaptation algorithm, finally narrowing the distance between the source domain and the target domain samples in the feature space at the beginning of the transfer learning, reducing the difficulty of the transfer learning, and more accurately realizing the transfer of bearing fault diagnosis of different models and different working conditions, thereby expanding the transfer range of the model, effectively solving the problem of small sample and lack of sample in the target domain, and improving the accuracy of the fault diagnosis.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of fault diagnosis of rail transit bearings, and particularly relates to a train bearing fault diagnosis method based on improved joint distribution adaptation. BACKGROUND

[0002] Due to the characteristics of high speed, large passenger capacity, high road priority and small interference from other traffic participants, rail transit has become one of the most popular transportation modes. At the same time, the safety of trains has attracted the attention of people from all walks of life. According to relevant reports, whether any part of the train fails, the accidents and disasters it brings are devastating. As a basic component of urban rail transit trains, the health status of rolling bearings is closely related to many major accidents of trains. Relevant data show that the mechanical component failure accounts for as high as 78% of the train running part failures. In the drive system failure of motor trains, about 30% of the failures are caused by rolling bearings. Whether the health status of rolling bearings is normal and stable is very important for the normal operation of trains.

[0003] In the actual operation of trains, it is often difficult to obtain a large amount of bearing fault data for model construction and training of diagnostic algorithms, and there is a small sample problem. At the same time, due to the change of the structure and size of the bearing, the running conditions and the use conditions often change with the application scene, and the feature distribution of the collected signal data will also change, which does not meet the basic assumption that the training sample set and the test sample set are independent and identically distributed, and it is difficult to use traditional machine learning models for modeling and fault diagnosis. In addition, time, application scene or other dynamic factors will affect the feature distribution of the vibration signal, and the out-of-date data needs to be collected again, and the model needs to be rebuilt, which consumes a lot of human and material resources.

[0004] Therefore, how to solve the problem of bearing fault detection model construction without samples, how to use laboratory data to complete model construction, and how to improve the precision and accuracy of the existing diagnostic model are great challenges in the field of bearing fault diagnosis.

[0005] At present, some researches have introduced transfer learning into the field of bearing fault diagnosis to provide a solution to the above problems. The most widely used method is the joint distribution adaptation method. Compared with the edge distribution adaptation which is prone to overfitting and the conditional distribution adaptation which is prone to underfitting, the joint distribution adaptation method considers the differences in the edge distribution and the conditional distribution of the source domain and the target domain samples, and considers more comprehensively.

[0006] In the joint distribution adaptation method, the measurement selection of the classifier KNN for constructing the target domain pseudo label is the key. Good measurement criteria can improve the performance of the classifier, thereby increasing the correctness of the constructed target domain pseudo label, and then obtaining the optimal feature mapping before the migration learning iteration. It reduces the difficulty of migration learning iteration, enhances the migratability of samples, thereby expanding the migration range of the model; it also improves the computational efficiency of the model, which provides a possibility for real-time health detection of bearings. The traditional joint distribution adaptation method uses Euclidean distance as the measurement criterion for measuring the distance between samples. The following defects exist: (1) Euclidean distance calculates the absolute distance between sample feature components, and when there is a correlation between feature components, it cannot accurately depict the distance between samples. However, in actual engineering cases, the features extracted from bearing signals often have certain correlations. (2) At the same time, Euclidean distance does not consider the influence of each feature attribute on the classification result, and treats each feature component equally, which does not meet the actual fault diagnosis requirements and expands the system error caused by feature selection. Wang Yuzhe improved the joint distribution adaptation method by using correlation entropy induced distance as the distance measurement criterion, but because the correlation entropy induced distance weakens the effective measurement range of the classifier in the sample space, it cannot capture all the information in the sample space, resulting in the locality of the classifier measurement, so it cannot accurately capture the attributes of the source domain and target domain samples. (3) At the same time, in the feature extraction stage, there are many feature components related to bearing faults, and scholars continuously propose new related variables from various angles. Most experiments only select feature components from the perspective of model classification effect, ignoring the correlation analysis of feature components, so they cannot contain comprehensive fault information, thereby failing to diagnose early bearing faults and causing greater disasters after the fault worsens. SUMMARY

[0007] The purpose of the present application is to provide a train bearing fault diagnosis method based on improved joint distribution adaptation, which comprehensively considers the influence of data processing and migration model on diagnosis accuracy. The proposed data preprocessing method can effectively extract more fault information from the vibration signal with fewer feature components. The weighted chi-square distance introduced in the model-embedded classifier as a distance measurement can better adapt to the migration of bearing fault signal data under different working conditions and different types, improve the sensitivity of the model to subtle differences in features, and thus realize timely and accurate judgment of early bearing faults, avoid serious harm caused by fault deterioration, and provide protection for rail transit safety. In order to achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0008] A train bearing fault diagnosis method based on improved joint distribution adaptation, comprising the following steps:

[0009] Step S1, based on the original vibration signal data of the rolling bearings with different health conditions in the laboratory conditions, determine the source domain original sample points and the target domain original sample points involved in the joint distribution adaptation method;

[0010] Step S2, obtain the optimal feature vector group as the input sample of the classifier involved in the joint distribution adaptation:

[0011] 21) Obtain the basic feature vector group X base :

[0012] Extract a plurality of feature components from the source domain sample points and the target domain sample points respectively, and attach the corresponding health condition label m to the source domain sample points; m is a positive integer; wherein the feature components extracted from the source domain sample points and the target domain sample points are one-to-one corresponding;

[0013] Then, perform quantitative cluster analysis on the sample points of each feature component extracted from the source domain sample points, and select the basic feature vector group X base corresponding to the source domain sample points according to the clustering effect of the label respectively;

[0014] 22) Based on the correlation analysis method, analyze the correlation between the feature components X other not included in X base and each feature component in X base , and X base , and include X other with low correlation with each feature component in X base in X base to obtain the optimal feature vector group X;

[0015] Select the same feature components as the optimal feature vector group X of the source domain to form the optimal feature vector group Y of the target domain;

[0016] Step S3, based on X and Y, calculate the weighted chi-square distance between each input sample point of the target domain and each input sample point of the source domain;

[0017] Step S4, select the K-nearest neighbor algorithm as the classifier of the joint distribution adaptation method, use the weighted chi-square distance obtained in step S3 as the distance measurement criterion of the K-nearest neighbor algorithm, and construct the pseudo label of the target domain classifier input sample point Y obtained in step S2;

[0018] Step S5, use the labeled sample to perform transfer learning to determine the health condition of the rolling bearing represented by the target domain sample point in step S1.

[0019] Preferably, in step 21), 25 feature components representing vibration signal data are extracted from the time domain, the frequency domain and the entropy domain; the bearing health conditions include normal, inner ring fault and outer ring fault.

[0020] Preferably, in step 21), the quantitative cluster analysis method is specifically:

[0021] The average values of the feature components corresponding to different labels are calculated, and the average values are compared in pairs, and the label with a larger average value is recorded as a temporary label m large ; and the label with a smaller average value is recorded as a temporary label m small ;

[0022] If, for a certain feature component, the average values of the feature components corresponding to any two sample labels satisfy: Then the feature component is effective for label classification, and the classification of different states of the same bearing can be directly completed, and is included in the basic feature vector group X base ;

[0023] Where, σ m - the standard deviation of all sample points corresponding to the label m, n- the number of sample points corresponding to the label m.

[0024] Preferably, in step 22), the correlation analysis method is specifically:

[0025] X base and the source domain sample points, each feature component in X base , the feature component X base not included in X other are input into the correlation analysis model, and the correlation analysis model outputs the Pearson correlation coefficient and the test p value of each component in X other and X base ;

[0026] If there are 0.8l feature components in X base and the p value of the feature component X other is greater than 0.1, then the feature component X other is included in X base ; wherein, l is the number of feature components in X base ;

[0027] The best feature vector group X of the source domain is finally obtained

[0028] Wherein, l is the number of feature components in X; n is the number of sample points contained in the source domain input sample set;

[0029] Each row in the X matrix represents a sample point input into the classifier, and each sample point contains l feature components. The i-th source domain input sample can be represented as: X i ={x 1,1 ,x 1,2 ,…,x 1,l};

[0030] Wherein, xi,a is the i-th input sample point X i of the source domain;

[0031] optimal feature vector set of the target domain wherein, l is the number of feature components in Y; c is the number of sample points contained in the input sample set of the target domain;

[0032] Each row of the Y matrix represents an input sample point of the classifier, and each sample point contains l feature components. The j-th input sample of the target domain can be expressed as: Y j = {y j,1 , y j,2 , …, y j,a , …, y j,l};

[0033] wherein, y j,a is the a-th feature component value of the j-th input sample point Y j of the target domain.

[0034] Preferably, in step S3, the calculation method of the weighted chi-square distance is:

[0035] 31) Calculate the influence weight coefficient of each feature component on the classification result:

[0036] Divide the source domain input sample X into two groups X G and X C in the ratio of 3:1, wherein X G is used as a control sample to analyze the influence weight of the feature component on the classification result, and X C is used as a test sample; use the chi-square distance as the measurement criterion of the KNN algorithm to classify the test sample, compare the label classified by the classifier with the existing label, and count the number of samples with classification errors, denoted as μ all .

[0037] Remove the a-th feature component in turn, and then use the same method to count the number of samples with label classification errors, denoted as μ a .

[0038] is the influence factor of each feature component on the error of the classifier. The greater the influence factor, the more obvious the effect of the feature component on the classification result of the classifier;

[0039] is the influence weight coefficient of the a-th feature component on the classification result, and satisfies

[0040] 32) Calculate the weighted chi-square distance:

[0041]

[0042] wherein x i,a is the a-th feature component value of the i-th input sample point X i of the source domain, y j,a is the a-th feature component value of the j-th input sample point Y j of the target domain.

[0043] Preferably, step S4 is specifically:

[0044] Firstly, a suitable k value is set by using cross-validation and grid search;

[0045] Then, the k source domain sample points closest to the target domain sample are selected;

[0046] According to the label of the source domain input sample point and the distribution of its feature components, a pseudo label of the target domain input sample point is constructed for subsequent implementation of fault diagnosis based on transfer learning;

[0047] Thus, the target domain input sample set (Y, L t ) with pseudo labels is obtained, Y is the set of target domain input sample points, and L t is the pseudo label set constructed by the classifier KNN;

[0048] L t takes a value representing the health status label m of the bearing predicted by the classifier; m is a positive integer; and c is the number of sample points in the target domain sample set;

[0049] wherein Y and L t correspond one-to-one.

[0050] Preferably, step S5 is specifically: the target domain input sample set (Y, L t ) with pseudo labels and the source domain input sample set (X, L) are brought into a joint distribution adaptation algorithm for transfer learning, a transformation matrix A is obtained, and memory training is performed;

[0051] wherein L is the health status label attached to the source domain initial sample point in step S2;

[0052] L takes a value representing the health status label m of the bearing predicted by the classifier; m is a positive integer; n is the number of sample points in the source domain sample set, X and L correspond one-to-one;

[0053] Then, the distance between the source domain and the target domain input sample points in step S2 is measured by using the maximum mean difference, and it is judged whether the error precision requirement is met; iterative operation is performed until the precision requirement is met.

[0054] Compared with the prior art, the advantages of the present application are:

[0055] (1) The method based on quantitative clustering analysis and correlation analysis extracts the optimal input vector of the effective feature component model, avoids the information redundancy brought by invalid feature components, ensures the comprehensive information amount contained in the feature vector group, and more effectively plays the ability of machine learning to mine data information. Thus, the optimal lightweight feature vector group based on actual data is determined, which is brought into the fault diagnosis model to improve the accuracy and computational efficiency of fault diagnosis.

[0056] (2) Considering that the distance measurement method used by the joint distribution adaptation method in the existing research does not consider the influence of the correlation of feature components and the difference between feature components on the classification result, the K-nearest neighbor algorithm based on weighted chi-square is used, and the chi-square distance and the iterative weighting according to the classification result can effectively identify the sensitivity of each feature to fault diagnosis, thereby improving the fault diagnosis accuracy after migration, expanding the migration range of the model, and increasing the applicability of the model.

[0057] (3) The method provided by the present application comprehensively solves the problems of actual fault data missing, multiple bearing models and variable working conditions in the field of traditional fault diagnosis algorithm research, effectively solves the problems of small sample of bearing fault and large difference of sample features in different scenes, can more accurately and quickly capture the early slight fault of the in-service bearing, reduces the casualties and losses caused by bearing failure, and can ensure the safety of rail transit. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 is the flowchart of the train bearing fault diagnosis method based on the improved joint distribution adaptation of the present embodiment;

[0059] Figure 2 is a two-dimensional schematic diagram before sample migration of the present embodiment;

[0060] Figure 3 is a schematic diagram of the sample after migration learning by the improved method of the present embodiment. DETAILED DESCRIPTION

[0061] The train bearing fault diagnosis method based on the improved joint distribution adaptation of the present application will be described in more detail below in conjunction with the schematic diagram, wherein the preferred embodiment of the present application is represented, and it should be understood that the present application described herein can be modified by those skilled in the art, while still achieving the advantageous effects of the present application. Therefore, the following description should be understood as extensive knowledge for those skilled in the art, and not as a limitation on the present application.

[0062] A train bearing fault diagnosis method based on improved joint distribution adaptation, comprising the following steps:

[0063] Step S1, based on the original vibration signal data of the rolling bearing with different health conditions under laboratory conditions, determine the source domain sample points and target domain sample points involved in the joint distribution adaptation method.

[0064] The effectiveness of the method of the present application will be verified below by using the fault bearing data of a certain institute in Qingdao. The bearings are numbered according to the fault type and placed in a laboratory environment to collect vibration signals in a simulation experiment designed based on the real train running environment.

[0065] This embodiment selects five groups of bearings with normal, inner ring fault (numbered 1822, 1425, representing two groups of different types of inner ring fault), outer ring fault (numbered 2067, 1430, representing two groups of different types of outer ring fault) as research objects to complete the sample transfer learning under different rotating speeds.

[0066] 11) First, obtain the original vibration signal data of rolling bearings with various health conditions under laboratory conditions.

[0067] Specifically, the research objects of the present example can be artificially damaged rolling bearings, or real fault bearings obtained by disassembling the running trains due to natural wear. The rolling bearings will be damaged to varying degrees in the inner ring, outer ring and rolling elements due to vibration and collision between internal parts, connection and impact with other external mechanical equipment such as tracks, and corrosion of the external environment.

[0068] The rolling bearing faults are mainly divided into three types according to the location: inner ring fault, outer ring fault and rolling element fault. Due to the different rotating speeds of the bearings, bearing types, severity of the fault and shape of the damage, they can be further subdivided.

[0069] During the actual operation of the train, the service life of the bearing is often disturbed by various uncontrollable factors in the running environment, and the rotating speed and type of the bearing will vary due to different needs, making it difficult to use laboratory data and traditional fault diagnosis techniques to timely detect early faults of the bearing. At the same time, the sample data of different fault types is small, which is difficult to meet the requirements of model training.

[0070] This embodiment can solve these problems to some extent, improve the feature component selection method, introduce weighted chi-square as the distance measure of the joint distribution adaptation classifier, effectively improve the sensitivity of information acquisition, and is conducive to improving the accuracy of bearing diagnosis and timely detecting early faults of the bearing; it is conducive to expanding the range of transfer learning, providing a solution to small sample problems, and has practical application value.

[0071] 12) Based on the mature fault bearing theoretical model and the original vibration signal data, the source domain input sample points and the target domain input sample points involved in the joint distribution adaptation method are determined.

[0072] The fault bearing theoretical model is: a sample set (original vibration signal data) in which the vehicle speed is 120 km / h (test environment speed) is selected and the source domain original data set is extracted therefrom, and a sample set (original vibration signal data) in which the vehicle speed is 150 km / h is selected and the target domain original data set is extracted therefrom.

[0073] The above five groups of bearings are respectively placed on the instruments with test speeds of 120 km / h and 150 km / h for rotation to complete the running speed test, and a total of 10 running speed tests are performed.

[0074] The sampling frequency of the data acquisition device is 12.8 kHz*2.56. It is specified that 1024 sampling points are taken for each original sample point of each health condition, and a total of 60 original sample points are taken. Among them, the sampling point refers to the vibration signal data point obtained by the data acquisition device, the original sample point indicates that 1024 sampling points constitute an initial sample, and the input sample point indicates a set of feature components after feature extraction (for subsequent classifiers and transfer learning) of each original sample point. A sample point indicates a point, and a sample set indicates a set of sample points.

[0075] Step S2, obtaining the best feature vector group as the input sample of the classifier involved in the joint distribution adaptation.

[0076] When the vibration signal is processed, if a single feature component is selected for fault diagnosis, the vibration information description is insufficient, and it is not sufficient to effectively summarize all the features of the rolling bearing fault. Therefore, the original vibration signal data set obtained by the present application is subjected to multi-field feature extraction, so that the feature information of the bearing fault is more obviously manifested.

[0077] Meanwhile, in view of the different sensitivities of different feature components to different types of bearing data sets, too few feature components may not be able to capture enough effective information, and too many feature components may lead to problems such as data redundancy and information repetition in the model. The present embodiment adopts a combination of quantitative cluster analysis and similarity analysis to construct a feature vector group most suitable for experimental data, in order to obtain high-quality classifier input samples.

[0078] 21) Obtain the basic feature vector group X base .

[0079] 25 kinds of feature components are extracted from the source domain original sample points and the target domain original sample points respectively, and the source domain sample points are attached with corresponding health status labels m; m is a positive integer. By quantifying the clustering of each feature component according to the label, the most effective feature component is selected as the basic feature vector group, laying the foundation for constructing a lightweight early fault diagnosis model.

[0080] Among them, the feature components extracted from the source domain original sample points and the target domain original sample points are one-to-one correspondence. Specifically, 25 kinds of feature components representing vibration signal data are extracted from the time domain, the frequency domain and the entropy domain; the bearing health status includes normal, inner ring fault and outer ring fault.

[0081] In order to facilitate operation, the health status of the bearing is defined as follows:

[0082] Bearing condition Normal Inner ring fault Outer ring fault Label m 1 2 3

[0083] 25 kinds of feature components are extracted from the time domain, the frequency domain and the target domain, wherein x i is a vibration signal, as follows:

[0084]

[0085]

[0086]

[0087] 16 kinds of common features are selected from the time domain as the feature vector group to construct the alternative feature components:

[0088] (1) Effective value is used to describe the size of the work done by the vibration signal in a period.

[0089] (2) Maximum value and minimum value are used to describe the extreme degree of the peak value of the vibration signal in the waveform.

[0090] (3) Average value is used to describe the average intensity of the vibration signal.

[0091] (4) Peak-to-peak value is used to describe the variation range of the vibration signal.

[0092] (5) Root mean square is used to describe the energy of the vibration signal.

[0093] (6) Variance is used to describe the dispersion degree of the vibration signal, representing the dynamic component of the signal energy.

[0094] (7) Standard deviation is used to describe the dispersion degree of the vibration signal, which is conducive to the unified dimension.

[0095] (8) Rectified average value is used to describe the overall characteristics of the vibration signal.

[0096] (9) Kurtosis is used to describe the distribution characteristics of the vibration signal.

[0097] (10) Absolute value is used to describe the energy of the vibration signal.

[0098] (11) Waveform factor is used to detect the wear condition of the mechanical equipment.

[0099] (12) Peak factor is used to detect whether there is an impact in the signal.

[0100] (13) Kurtosis factor is used to describe the flatness of the waveform and the distribution of the variable.

[0101] (14) Pulse factor is used to detect whether there is an impact in the signal.

[0102] (15) Margin factor is used to detect the wear condition of the mechanical equipment.

[0103] Five common features are selected from the frequency domain as the alternative feature components for constructing the feature vector group:

[0104] (1) Frequency center is used to describe the geometric mean value of the vibration signal in the frequency domain.

[0105] (2) Root mean square frequency is used to describe the position change of the main frequency band of the vibration signal in the frequency domain.

[0106] (3) Mean frequency is used to describe the average intensity of the vibration signal in the frequency domain.

[0107] (4) Hoyer coefficient is used to describe the distribution condition of the signal.

[0108] (5) Norm is used to describe the signal energy size.

[0109] Four common features are selected from the entropy domain as the alternative feature components for constructing the feature vector group: (1) Approximate entropy is used to describe the regularity of the vibration fluctuation; and (2) Shannon entropy is used to describe the uncertainty of the vibration signal.

[0110] Then, the quantization clustering analysis is performed on each feature component extracted from the source domain sample points, and the basic feature vector group X base corresponding to the source domain sample points is selected according to the clustering effect of the labels respectively.

[0111] The quantization clustering analysis method is as follows:

[0112] The average values of the feature components corresponding to different labels are calculated, and the average values are compared with each other, the label with the larger average value is recorded as the temporary label m large , and the label with the smaller average value is recorded as the temporary label m small .

[0113] If for a certain feature component, the average value of the feature component corresponding to the sample labels of any two classes satisfies: Then the feature component is effective for label classification, and can directly complete the classification of different states of the same bearing. Therefore, it is included in the basic feature vector group X base ; Where, σ m - The standard deviation of all sample points corresponding to the label m, n- The number of sample points corresponding to the label m.

[0114] The acquisition process of the basic feature vector group X base is described as an example:

[0115] Taking the feature component standard deviation as an example, σ1=0.025, n1=80 are calculated. σ2=0.21, n2=80 are calculated. Therefore, for any m=1, 2, 3, satisfies

[0116] First, compare the labels m=1 and m=2, Therefore, the temporary label is set as follows: m large =2, m small =1.

[0117]

[0118]

[0119] satisfies the screening rule of

[0120] Second, compare the labels m=1 and m=3, Therefore, the label is set as follows: m large =3, m small =1.

[0121]

[0122]

[0123] satisfies the screening rule of

[0124] Finally, compare the labels m=2 and m=3, Therefore, the label is set as follows: m large =3, m small =2.

[0125]

[0126] ​​​

[0127] satisfy The filtering rules.

[0128] In summary, for any m = 1, 2, 3, the following condition is satisfied: It is believed that standard deviation is effective as a feature component for label classification, and it is used as a feature component in the basic feature vector set.

[0129] The mean and standard deviation of the samples under each label (3 types) were calculated using the above method for the remaining 24 feature components. The samples were then selected according to the aforementioned screening rules, and the final extracted basic feature vector set was X. base The characteristic components it contains include: standard deviation, minimum value, effective value, kurtosis factor, mean, root mean square, mean of absolute values, absolute value, approximate entropy, peak factor, variance, and waveform factor.

[0130] 22) Based on this, considering that some feature components, although not directly usable for bearing health status classification, contain effective information that other feature vectors cannot express, we further analyzed the features not included in X. base Characteristic component X other The Pearson correlation coefficients and p-values ​​with each component in the basic feature vector set are as follows:

[0131] For X base Based on correlation analysis, the unincluded feature components X were analyzed from the source domain sample points. other With X base The correlation between the various characteristic components in X will be related to X. base1 X with low correlation among various feature components other Include X base To obtain the best feature vector set in the source domain Where l is the number of feature components in X, l = 14; n is the number of sample points in the source domain, n = 300. The feature components included in the optimal feature vector are standard deviation, minimum value, effective value, kurtosis factor, mean, root mean square, mean of absolute values, absolute value, approximate entropy, peak factor, variance, waveform factor, peak-to-peak value, and Hoyer coefficient.

[0132] For the optimal feature vector set Y of the target domain, since the labeled sample dataset may be small or even non-existent in practical applications, the same feature components as the optimal feature vector set X of the source domain are selected to form the optimal feature vector set of the target domain. Where l is the number of feature components in Y, l = 14; c is the number of sample points in the target domain, c = 300.

[0133] Finally, X and Y are normalized to eliminate the influence of different dimensions of the data.

[0134] The optimal feature vector set contains comprehensive fault information while avoiding data redundancy and information duplication, thus improving the model's computational efficiency.

[0135] The correlation analysis method specifically includes:

[0136] For X base And source domain sample points, put X base The characteristic components in, not included in X base Characteristic component X in other All inputs are into a correlation analysis model (such as SPSS), and the correlation analysis model outputs X. other With X base The Pearson correlation coefficients and p-values ​​of each component are calculated.

[0137] If X base There are 0.81 characteristic components that are related to characteristic component X. other If the p-value is greater than 0.1, then the feature component X is... other Include X base Where, l is X base The number of characteristic components in the data.

[0138] Step S3: Based on X and Y, calculate the weighted chi-square distance between the same feature components.

[0139] This embodiment improves upon the traditional joint distribution adaptation algorithm by introducing a weighted chi-square distance as a metric for the classifier.

[0140] Weighted chi-square distance can effectively observe the relative distance changes of different feature vectors. It can accurately capture all the information in the sample space and assign different weights to different sample feature vectors, thereby improving the classifier's sensitivity to sample features, buffering the impact of overall differences, effectively reducing error fluctuations and accumulation, and narrowing the distance between the source and target domain sample spaces before transfer, thus expanding the model's transfer range and improving the accuracy of fault diagnosis.

[0141] Specifically, the selected 14 feature components (from the optimal feature vector group) are used to calculate the chi-square distance between each input sample point in the target domain and each input sample point in the source domain.

[0142] The weighted chi-square distance is calculated as follows:

[0143] 31) Calculate the weighting coefficient of each feature component's influence on the classification result:

[0144] The source domain samples X are randomly divided into two groups X1 and X2 in a 3:1 ratio. G and X C , where XG X is the source domain sample point, Y is the target domain sample point, a is the feature component, and μ is the number of samples with label classification errors C As a test sample. The test sample is classified by using the chi-square distance as the measurement criterion of the KNN algorithm, the label classified by the classifier is compared with the existing label, and the number of samples with classification errors is counted, denoted as μ all ;

[0145] The a-th feature component is sequentially removed, and the number of samples with label classification errors is counted by using the same method, denoted as μ a ;

[0146] is the influence factor of each feature component on the error of the classifier, and the greater the influence factor, the more obvious the feature component is on the classification effect of the classifier;

[0147] is the influence weight coefficient of the a-th feature component on the classification result, and satisfies

[0148] 32) Calculate the weighted chi-square distance:

[0149]

[0150] Wherein, x i,a is the a-th feature component value of the i-th sample point of the source domain, y j,a is the a-th feature component value of the j-th sample point of the target domain.

[0151] Step S4, select the K nearest neighbor algorithm as the classifier of the joint distribution adaptation method, use the weighted chi-square distance obtained in step S3 as the distance measurement criterion of the K nearest neighbor algorithm, and construct the pseudo label of the target domain classifier input sample point Y obtained in step S2; the target domain classifier input sample point Y is the best feature vector group Y of the target domain.

[0152] Wherein, the joint distribution adaptation method, the weighted chi-square distance and the pseudo label are prior art.

[0153] Specifically: first, set a suitable k value by using cross-validation and gridding search. In the example of the present application, the k value is taken as 10, and the K nearest neighbor algorithm based on chi-square distance is used to predict the pseudo label of the target domain sample.

[0154] Then select the 10 source domain sample points X j closest to each sample point Y i in the target domain vector group Y According to the label and feature component distribution of the source domain input sample point, the pseudo label of the target domain input sample point is constructed for subsequent implementation of fault diagnosis based on transfer learning.

[0155] In this embodiment, N m is the number of sample points with health status m in the k nearest neighbor labels, m = {1, 2, 3}.

[0156] If max(N m ) = N1, the pseudo label corresponding to the target domain is obtained

[0157] If max(N m ) = N2, the pseudo label corresponding to the target domain is obtained

[0158] If max(N m ) = N3, the pseudo label corresponding to the target domain is obtained

[0159] Thus, the input sample set (Y, L t ) of the target domain with pseudo labels is obtained, Y is the set of input sample points of the target domain, L t is the pseudo label set constructed by the classifier KNN, the value of the element in L t represents the health status label m of the bearing predicted by the classifier; m = {1, 2, 3}; c is the number of sample points in the sample set of the target domain, c = 300. Wherein, Y and L t are in one-to-one correspondence.

[0160] Step S5, using the sample with labels to perform transfer learning to determine the health status of the rolling bearing represented by the sample points in the target domain in step S1. This step is prior art.

[0161] Specifically: the labeled sample is brought into the joint distribution adaptation algorithm to perform transfer learning, obtain the transformation matrix A, and perform transfer learning; memory training is performed to map the feature vector groups of the two fields to the feature space, and the distance between the samples is measured using the maximum mean difference to determine whether the error precision requirement is met; iterative operation is performed until the precision requirement is met.

[0162] Step S5 specifically: the input sample set (Y, L t ) of the target domain with pseudo labels and the input sample set (X, L) of the source domain are brought into the joint distribution adaptation algorithm to perform transfer learning, obtain the transformation matrix A, and perform memory training;

[0163] Wherein, L is the health status label attached to the initial sample points of the source domain in step S2;

[0164] The value of L represents the corresponding health status label m of the bearing predicted by the classifier; m is a positive integer; n is the number of sample points in the source domain sample set, and X corresponds one-to-one with L.

[0165] Then, the distance between the input sample points in the source and target domains in step S2 is measured using the maximum mean difference to determine whether the error accuracy requirement is met; iterative calculation is performed until the accuracy requirement is met.

[0166] Specifically, experiments were conducted to perform a gridded search for the dimensionality reduction dimension dim and the regularization parameter λ for the maximum mean difference in the joint distribution adaptation domain.

[0167] Based on empirical values, the dimensionality reduction dimension of the model is set to an integer between [5, 25], while keeping other variables the same. Ten independent experiments are conducted, and the average accuracy is used as the criterion to obtain the optimal dimensionality reduction dimension.

[0168] Using the same method, the model's regularization parameter is set to a value within the interval λ, while keeping other variables the same. Ten independent and repeated experiments are conducted, and the λ with the highest accuracy is taken as the model parameter for subsequent experiments. Where λ∈{0.0001,0.001,0.01,0.1,0.5,1,10,100}.

[0169] By linking the marginal and conditional distribution probabilities using c = 0...C, the two problems of minimizing the distance are merged, resulting in the overall optimization objective:

[0170]

[0171] in As a regularization term, it can ensure that the model has a good definition effect.

[0172] Since the problem may have no solution under unrestricted conditions, a constraint is set to ensure that the variance of the matrix before and after the transformation remains unchanged, i.e., A. T XHX T A = I.

[0173] Where H is the central matrix and I is the identity matrix.

[0174] Combining this with the original optimization objective, we get:

[0175] Since A can be stretched without changing the final result, for ease of calculation, we assume the denominator is a fixed value, resulting in the final optimization problem:

[0176]

[0177] stA T XHX TA = I;

[0178] Finally, the solution was obtained using the Lagrange method: Where Φ is a Lagrange multiplier. Transfer learning is performed on labeled samples to obtain a transformation matrix, which is the feature space where the two samples are close to each other. Samples are projected into this space, the distance between samples is calculated, and it is determined whether the accuracy requirement is met. Iterative training is then performed to finally obtain a classification result that meets the requirements.

[0179] In summary, addressing the issues of missing fault data and significant differences in the distribution of bearing fault samples under different operating conditions and models in practical engineering applications, an improved joint distribution adaptation algorithm is proposed. This algorithm utilizes a K-nearest neighbor algorithm based on weighted chi-square distance as the embedded classifier, overcoming the limitation of existing algorithms in accurately characterizing the distance between correlated feature component samples. Furthermore, it better captures the differences in sensitivity of different feature components to classification, effectively reducing error fluctuations and accumulation, and obtaining more accurate pseudo-labels for the target domain. This, in turn, shortens the distance between source and target domain samples in the feature space during the initial stage of transfer learning, reducing the difficulty of transfer learning and enabling more accurate transfer of fault diagnosis for bearings of different models and operating conditions. This expands the model's transfer range and improves the accuracy of fault diagnosis. The method provided by this invention comprehensively solves the problems of missing practical fault data, diverse bearing models, and variable operating conditions in traditional fault diagnosis algorithm research. It effectively addresses the issues of small bearing fault samples and significant differences in sample features under different scenarios, enabling more accurate and rapid capture of early minor faults in in-service bearings, reducing casualties and losses caused by bearing faults, and ensuring rail transit safety.

[0180] Figure 2 , Figure 3 The t-SNE algorithm was used to perform dimensionality reduction and visualization analysis on the samples before and after migration, and the two-dimensional clustering of sample data with different labels was plotted.

[0181] A represents the two-dimensional clustering of fault samples in the outer ring of the source domain, and B represents the two-dimensional clustering of fault samples in the outer ring of the target domain. (Source domain before migration) Figure 2 (A) and target domain ( Figure 2 The outer ring fault samples (B) are distributed in different regions. The two-dimensional clustering space mapped by the correspondence between the best feature vector set of the source domain and the label (outer ring fault) is region A, and the two-dimensional clustering space mapped by the correspondence between the best feature vector set of the target domain and the label (outer ring fault) is region B. Regions A and B do not overlap. If the best feature vector set of the outer ring fault samples of the source domain and the corresponding label are used to train the classifier, it cannot correctly diagnose the outer ring faults of the target domain samples. After transfer, the outer ring fault samples of the source domain and the target domain are distributed in the same region (B). Figure 3The best feature vector group of the two-domain sample extraction and the label (outer ring fault) correspond to the same two-dimensional clustering space, and if the best feature vector group of the source domain outer ring fault sample and the label (outer ring fault) are used to train the classifier, the outer ring fault of the unknown label target domain sample can be correctly diagnosed. By comparing before and after migration, it can be found that after migration of the improved joint distribution adaptation method, the source domain sample points and target domain sample points under the same outer ring fault at different rotating speeds are overlapped in the clustering area of the feature space, effectively verifying that the improved method can predict the bearing fault at unknown rotating speed, solving the small problem of the labeled sample set caused by the multiple types of bearings and the changeable running environment in practical application, and therefore can be applied to actual industrial production.

[0182] The above is only the preferred embodiment of the present application, and does not limit the present application in any way. Any person skilled in the art can make any form of equivalent replacement, modification or change to the technical solutions and technical contents disclosed by the present application without departing from the scope of the technical solutions of the present application, and still belongs to the protection scope of the present application.

Claims

1. A train bearing fault diagnosis method based on improved joint distribution adaptation, characterized in that, The method comprises the following steps: Step S1, based on the original vibration signal data of the rolling bearing with different health conditions under laboratory conditions, determine the source domain original sample points and target domain original sample points involved in the joint distribution adaptation method; Step S2, obtain the optimal feature vector group as the input sample of the classifier involved in the joint distribution adaptation: 21) Acquiring base feature vector set : A plurality of feature components are extracted from the source domain sample points and the target domain sample points respectively, and the source domain sample points are labeled with corresponding health status ; is a positive integer; wherein the feature components extracted from the source domain sample points and the target domain sample points are one-to-one corresponding Then, the extracted feature components in the source domain sample points are subjected to quantitative cluster analysis, and the basic feature vector groups corresponding to the source domain sample points are selected according to the cluster effect of the labels respectively ; 22) for and source domain sample points, based on correlation analysis method, respectively analyze the feature components not included in the basic vector group and the correlation of each feature component in , the correlation of each feature component in is low , and , to obtain the best feature vector group ; Selecting and source domain optimal eigenvector set The same feature component constitutes the target domain optimal eigenvector set ; Step S3, based on and , the weighted chi-square distance between each input sample point of the target domain and each input sample point of the source domain is calculated; Step S4, selecting K-nearest neighbor algorithm as the classifier of the joint distribution adaptation method, using the weighted chi-square distance obtained in step S3 as the distance measurement criterion of the K-nearest neighbor algorithm, and constructing the pseudo-labels of the input sample points of the target domain classifier obtained in step S2 . Step S5, use the labeled sample to perform transfer learning, and determine the health condition of the rolling bearing represented by the target domain sample point in step S1; In step 21), the quantitative cluster analysis method is specifically: Calculate the average value of the feature components corresponding to different labels, and compare the size two by two, the label with larger average value is recorded as temporary label ; the label with smaller average value is recorded as temporary label m small ; If for a certain feature component, the average value of the feature component corresponding to the labels of any two classes of samples satisfies: , the feature component is effective for label classification, and the classification of different states of the same bearing can be directly completed. Therefore, it is included in the basic feature vector group . wherein, - the label the standard deviation of all the corresponding sample points, - the label the number of the corresponding sample points.

2. The train bearing fault diagnosis method based on improved joint distribution adaptation according to claim 1, characterized in that, In step 21), 25 characteristic components representing the vibration signal data are extracted from the time domain, frequency domain and entropy domain; the bearing health state includes normal, inner ring fault and outer ring fault.

3. The train bearing fault diagnosis method based on improved joint distribution adaptation according to claim 1, characterized in that, In step 22), the correlation analysis method is specifically: against and source domain sample points, The characteristic components in, not included Characteristic components All inputs are used in the correlation analysis model, and the correlation analysis model outputs... and Pearson correlation coefficients and tests for each component value; like There exists Each feature component and this feature component of If the value is greater than 0.1, then this feature component... Included ;in, for The number of characteristic components in; The optimal feature vector set of the source domain is finally obtained ; wherein, is the number of feature components in is the number of sample points contained in the source domain input sample set; Each row in the matrix represents a sample point input to the classifier, each sample point contains feature components, the th source domain input sample can be represented as: ; wherein, is a value of a first feature component of a source domain first input sample point ; and is a value of a second feature component of the source domain first target domain optimal feature vector group ; wherein, is the number of feature components in the number of sample points contained in the target domain input sample set Each row in the matrix represents a sample point input to the classifier, each sample point contains feature components, the input sample of the target domain can be expressed as: ; wherein, is a value of a feature component of a target domain th input sample point th feature component of a target domain th input sample point.

4. The train bearing fault diagnosis method based on improved joint distribution adaptation according to claim 1, characterized in that, In step S3, the calculation method of the weighted chi-square distance is: 31) Calculate the influence weight coefficient of each characteristic component on the classification result: The source domain input sample Randomly divided into two groups in the ratio of 3:1 And Wherein, As a control sample of the analysis feature component for the influence weight of the classification result, As a test sample; use the chi-square distance as the measurement criterion of the KNN algorithm to classify the test sample, compare the label classified by the classifier with the existing label, and count the number of samples with classification errors, denoted as ; The first feature component is removed in sequence The number of samples with label classification errors is counted by using the same method, and is denoted as ; The impact factor of each feature component on the classifier error, the greater the impact factor, the more obvious the feature component on the classification effect of the classifier; The influence weight coefficient of the first feature component on the classification result satisfies ;​ 32) Calculate the weighted chi-square distance: ； in, It is the source domain number input sample points The Each feature component value, Is the target domain number input sample points The Each feature component value.

5. The train bearing fault diagnosis method based on improved joint distribution adaptation according to claim 1, characterized in that, Step S4 is specifically: Firstly, use cross-validation and grid search to set the appropriate k value; Then select the k source domain input sample points closest to the target domain input sample points; According to the label of the source domain input sample point and the feature component distribution, a pseudo label of the target domain input sample point is constructed for subsequent implementation of fault diagnosis based on transfer learning. Thus, the input sample set with target domain pseudo-labels is obtained , is a set of target domain input sample points, is a pseudo-label set constructed for the classifier KNN; , The value represents the label of the corresponding health condition of the bearing predicted by the classifier ; is a positive integer; is the number of sample points in the target domain sample set; wherein, corresponds to one-to-one.

6. The train bearing fault diagnosis method based on improved joint distribution adaptation according to claim 1, characterized in that, Step S5 is specifically: inputting the target domain input sample set with pseudo labels into the joint distribution adaptation algorithm for transfer learning to obtain a transformation matrix A and perform memory training. and the source domain input sample set ) into the joint distribution adaptation algorithm for transfer learning to obtain a transformation matrix A and perform memory training. wherein, is the health condition label attached to the source domain initial sample point in step S2; ; the value of the classifier represents the corresponding health condition label of the bearing predicted by the classifier ; is a positive integer; is the number of sample points in the source domain sample set, and correspond one-to-one; Then use the maximum mean difference to measure the distance between the source domain and target domain input sample points in step S2 to determine whether the error precision requirement is met; iterative operation until the precision requirement is met.

Images