Train bearing fault diagnosis method based on graph multi-scale complex entropy

Abstract

The present application belongs to the technical field of train fault detection, and discloses a train bearing fault diagnosis method based on graph multi-scale complex entropy, which converts the collected original vibration signal of the measured part from the time domain to the graph domain; the signal in the graph domain is coarsened to obtain a multi-scale coarse graph; according to the multi-scale graph, the graph multi-scale complex entropy of the measured part is calculated and used as the feature vector of the measured part; the selected features in the feature vector of the measured part are input into the train bearing fault classifier to obtain the train bearing fault classification result, and the train bearing fault diagnosis is completed. The present application can effectively reduce the interference of background noise and impurity frequency at the same time, has strong fault feature extraction capability, and has high robustness, can provide higher accuracy for the fault diagnosis result of the bearing, realizes accurate diagnosis of different fault types and different fault degrees of the train bearing, can effectively monitor the train bearing fault, and improves the safety and economy of train operation.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of train fault detection, and relates to train bearing fault detection technology, in particular to a train bearing fault diagnosis method based on graph multi-scale complex entropy. BACKGROUND

[0002] As a key component in rotating machinery, bearings play a crucial role in transmitting rotational speed and load, and their health status is directly related to the safe operation of the entire train system. In high-speed trains and other rail transportation equipment, bearings work in an environment of high speed, high load, and large temperature variation for a long time, which can easily cause failure faults such as fatigue spalling, pitting, and wear. When a bearing fails, the repeated collision between the damage point and other components will produce impact vibration, which may cause train operation failure, economic loss, and even casualties in severe cases. Therefore, effective fault diagnosis and monitoring of train bearings are key links to ensure the safety of railway transportation.

[0003] Traditional train bearing fault diagnosis methods mainly rely on the analysis of vibration signals. Vibration signals carry key information about the bearing system during operation, including the working state of the bearing, the load condition, the friction and wear degree, etc. Through in-depth analysis of these vibration signals, the abnormal state of the train bearing can be effectively identified, and predictive maintenance measures can be implemented to reduce unnecessary downtime and maintenance costs, thereby improving the reliability and safety of train operation.

[0004] However, in actual application, the vibration signals of bearings are often accompanied by vibration signals of other mechanical components and strong noise background. These signals are often a comprehensive representation of the vibration responses of multiple components, making it difficult to extract weak fault features and increasing the difficulty of diagnosis and the possibility of misjudgment. In addition, the complexity of the sensor transmission path causes signal energy attenuation or distortion, making it more difficult to distinguish between vibration signals containing bearing fault feature information and healthy vibration signals, further exacerbating the difficulty of bearing fault diagnosis. SUMMARY

[0005] In view of the above problems, the present application provides a train bearing fault diagnosis method based on graph multi-scale complex entropy, which extracts the features of the vibration signal accurately, reduces the interference of background noise, and reduces signal attenuation. The method processes the fault signal through graph signal processing to obtain the health status index of the bearing, and realizes effective diagnosis of the train bearing fault.

[0006] To achieve the above purpose, the present application adopts the following technical solutions.

[0007] The train bearing fault diagnosis method based on graph multi-scale complex entropy provided by the present application comprises the following steps:

[0008] S1 converts the collected original vibration signal of the test piece from the time domain to the image domain;

[0009] S2 performs coarsening processing on the signal in the image domain to obtain a multi-scale coarse image, which specifically includes the following steps:

[0010] S21 clusters the graph domain signal to obtain the graph segmentation matrix;

[0011] S22 calculates the adjacency matrix and graph signal of the coarsened graph based on the graph partition matrix;

[0012] Select multiple scale factors and repeat the above steps S21-S22 to coarsen the image domain signal into multiple coarse images of different scales, i.e., multi-scale coarse images;

[0013] S3 calculates the multi-scale complex entropy of the DUT graph based on the multi-scale graph obtained in S2, and uses it as the feature vector of the DUT to define the graph complex entropy ,The specific steps of calculating the complex entropy of the graph are as follows:

[0014] S31 reconstructs the phase space of the graph signal under each scale factor and constructs m dimensional vector sequence;

[0015] S32 calculation m The Manhattan distance between the dimensional vector sequences is used as the similarity, and the statistical similarity is less than the threshold r The number of similarity probabilities is calculated to obtain the logarithmic mean value. P m ( r ), and similarly calculate the embedding dimension as m The logarithmic mean of the similarity probability under +1 P m+1 ( r ), P m ( r )and P m+1 ( r ) is E 1;

[0016] S33 According to step S32 P m ( r )and P m+1 ( r ),calculate for E 2;

[0017] S34 is obtained according to step S32 mThe Manhattan distance between the dimensional vector sequences is estimated using the histogram method to estimate the empirical probability density function of the distance and calculated E 3. , P b represents the probability of falling into the b interval of the histogram, and B represents the number of intervals in the histogram;

[0018] S35 calculates adjacent m The Pearson correlation coefficient between dimensional vector sequences is in the range of [-1, 1]. [-1, 1] is evenly divided into A subintervals, divide all correlation coefficients into subintervals, count the probability of falling into each subinterval, and calculate E 4. , P a Indicates that it falls into the subinterval a probability;

[0019] S36 For each scale factor, calculate , that is, the multi-scale complex entropy of the graph is obtained and used as the feature vector of the device under test;

[0020] S4 inputs the features filtered out from the feature vector of the test piece into the train bearing fault classifier to obtain the train bearing fault classification result and complete the train bearing fault diagnosis.

[0021] In step S1 above, the specific implementation method is to map the collected original time series signal into a two-dimensional matrix to obtain an undirected weighted graph. The mapping method can be a visual graph, a horizontal visual graph, a k-nearest neighbor graph, etc. This method stores the information in the one-dimensional time series in a graph matrix containing a set of bidirectional relationships between objects and the strength of these relationships, enabling the analysis of time series signals in the form of a graph. Based on graph spectrum theory, a graph matrix representation is then obtained, including an adjacency matrix, a degree diagonal matrix, and a Laplacian matrix. The eigenvalues ​​and eigenvectors of the normalized Laplacian matrix are then obtained to achieve visual multidimensional analysis.

[0022] In the above step S2, a coarsening algorithm is proposed. , the original image signal is X , W is the adjacency matrix, U is the eigenvector matrix obtained by orthogonal decomposition of the graph Laplace matrix, s is the scale factor, n is the number of nodes in the coarsened graph, and .

[0023] In the above step S21, each column of the eigenvector matrix is ​​normalized to [-1, 1], and the eigenvectors with absolute values ​​less than 0.001 are regarded as 0. The non-zero rate of each eigenvector is calculated to obtain the standard deviation of the non-zero rate. The eigenvectors with non-zero rates greater than 3 times the standard deviation are selected, and the K-means clustering algorithm is applied to the rows of the eigenvectors. The parameters in the algorithm are k = n , N The rows are grouped into n Class, according to the clustering results, the graph segmentation matrix is ​​obtained , P No. i Rank j The column elements have the following values:

[0024] ;

[0025] In the above step S22, the adjacency matrix of the coarsened graph is calculated based on the graph partition matrix W c and image signal X c , the calculation formula is as follows:

[0026] W c =P T WP ;

[0027] X c =P T X .

[0028] In the above step S3, a graph complexity entropy is proposed algorithm.

[0029] In the above step S31, based on the image signal obtained in step S2, m The specific steps for constructing the vector sequence of dimension are as follows: X c (When the scale factor is 1, the signal is X ) to reconstruct the phase space and form n vector sequence , is the embedding dimension, that is, the dimension of the vector, where:

[0030] ;

[0031] ;

[0032] In the above step S32, the distance is defined is a vector x m ( i )and x m ( j ), that is:

[0033] ;

[0034] For each vector x m ( i ), we need to calculate the Manhattan distance between it and other vectors (including its own), so for each vector x m ( i )have n distance, given a threshold r , the statistical distance is less than the threshold r The number of P m ( r )for:

[0035] ;

[0036] Calculating embedding dimensions m +1 P m+1 ( r ),but E 1 is:

[0037] E 1= P m ( r )- P m+1 ( r ).

[0038] In the above step S33, according to the calculation in S32 P m ( r )and P m+1 ( r ),but E 2 is:

[0039] .

[0040] In the above step S34, all Manhattan distances obtained in step S32 (excluding i = j The distance at the time of the test is estimated by using the histogram method. Assume that the histogram has B intervals, and the probability of all distances falling into each interval is ,definition E 3 is:

[0041] .

[0042] In the above step S35, define is a vector x m ( i )and x m ( j ), that is:

[0043] ;

[0044] in, express x m ( i )middle m The average of the values, express x m ( j )middle m The average of the values; that is:

[0045] ;

[0046] ;

[0047] Calculate the Pearson correlation coefficient between all adjacent vectors and get n- 1 correlation coefficient, the range of the Pearson correlation coefficient is [-1, 1], [-1, 1] is evenly divided into A subintervals, divide all correlation coefficients into subintervals, and count the probability of falling into each subinterval ,definition E 4 is:

[0048] .

[0049] In the above step S36, for each scale factor, the graphs calculated in S31-S35 are As the multi-scale complex entropy at the corresponding scale; the multi-scale complex entropy of all scale factors is constructed as the characteristic vector of the test piece.

[0050] In the above step S4, a model such as a random forest model, a support vector machine, a logistic regression model or a decision tree model is used as a train bearing fault classifier.

[0051] The above train bearing fault classifier is trained through the following steps:

[0052] (1) Based on the training set, the random forest model is used to filter the features in the feature vector, and the filtered features constitute the updated feature vector;

[0053] (2) Reconstruct the training set based on the updated feature vector to obtain the reconstructed training set;

[0054] (3) Use the reconstructed training set to train the train bearing fault classifier.

[0055] In the above step (1), the training set is obtained according to the following steps: different types of train bearing fault history sample data are used as training sample data, and the training data are processed according to the steps S1-S3 given above to obtain the feature vector of each sample; all samples and their feature vectors constitute the training set.

[0056] When using a random forest model to learn a training set, the learning results can be used to calculate each feature's mean impurity reduction (MDI), mean decrease in accuracy (MDA), or permutation importance as the feature importance. This feature importance is then used as a metric for feature selection. Features that meet the requirements are selected (for example, by ranking the feature importances to select the top M (e.g., top 15 or top 20) features) to form an updated feature vector. The updated feature vector is then used to select the feature values ​​of each sample in the training set. The selected feature values ​​are then used to reconstruct the training set, resulting in a reconstructed training set.

[0057] In the above step S4, the feature vector of the test piece is screened based on the features screened in the previous step (1).

[0058] Compared with the existing technology, the train bearing fault diagnosis method based on graph multi-scale complex entropy provided by the present invention has the following beneficial effects:

[0059] (1) This invention addresses the problem of low accuracy in vibration signal feature extraction in the prior art and proposes a new feature extraction approach. Combining the entropy theory, this invention quantifies the complexity and irregularity of the signal at multiple angles and scales, and then proposes a graph multi-scale complex entropy. This multi-scale complex entropy is used as a feature vector to identify train bearing faults.

[0060] (2) The present invention also uses a tree-like classification algorithm as the basis of the classification model. First, features with higher feature importance are screened out from the multi-scale complex entropy of the graph, and then train bearing faults are identified based on these features, which can improve the accuracy and effectiveness of train fault identification;

[0061] (3) The present invention can effectively reduce the interference of background noise and impurity frequency at the same time, has a strong fault feature extraction capability, and has high robustness. It can provide a higher accuracy rate for the fault diagnosis results of bearings, and realizes the accurate diagnosis of different fault types and different fault degrees of train bearings. It can effectively monitor train bearing faults and improve the safety and economy of train operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a flow chart of the train bearing fault diagnosis method based on graph multi-scale complex entropy proposed in the present invention.

[0063] Figure 2 is the confusion matrix of the fault detection result in the embodiment of the present invention. DETAILED DESCRIPTION

[0064] The following will clearly and completely describe the technical solutions of various embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention. Example

[0065] This embodiment provides a train bearing fault diagnosis method based on graph multi-scale complex entropy. Figure 1 As shown, the following steps are included:

[0066] S1 converts the collected original vibration signal of the test piece from the time domain to the image domain;

[0067] S2 performs coarsening processing on the signal in the image domain to obtain a multi-scale coarse image;

[0068] S3 calculates the multi-scale complex entropy of the DUT graph based on the multi-scale graph obtained in S2 and uses it as the feature vector of the DUT;

[0069] S4 inputs the features filtered out from the feature vector of the test piece into the train bearing fault classifier to obtain the train bearing fault classification result and complete the train bearing fault diagnosis.

[0070] In the step S1, during the process of collecting the original vibration signal, the acceleration sensor is installed on the surface of the train axle box, the acceleration sensor on the surface of the axle box obtains the train bearing fault signal, and the vibration signal is taken as the original signal for fault diagnosis. The original time sequence signal collected is mapped into a two-dimensional matrix to obtain an undirected weighted graph, and the mapping method adopts a visual graph algorithm. Through this method, the information in the one-dimensional time sequence is stored in a graph matrix form containing the bidirectional relationship between a group of objects and the strength of the relationship, and the time sequence signal is analyzed in the form of a visual graph. Based on the graph theory, the graph matrix representation form is obtained, including the adjacency matrix, the degree diagonal matrix and the Laplacian matrix, the eigenvalue and eigenvector of the normalized Laplacian matrix are obtained, and the visual multi-dimensional analysis is realized.

[0071] In the embodiment, the number of sampling points of the sample is set to 2048, each element in the sample is taken as a node in the graph, an undirected weighted graph is constructed according to the visibility relationship between the nodes, and the Euclidean distance between two nodes is taken as the weight of the edge between the two nodes. The calculation formula is as follows:

[0072] ;

[0073] In the formula, x i and x j are the values of the nodes i and j , respectively, the defined weight is the Euclidean distance between the two nodes, and the difference between the vertices is truly reflected.

[0074] Based on the graph theory, the matrix representation of the graph is obtained, including the adjacency matrix W , the degree diagonal matrix D and the graph Laplacian matrix L . In the degree diagonal matrix, for each node i in the graph, ; and L=DW , then the standard orthogonal decomposition is performed on L to obtain the eigenvalue and eigenvector U .

[0075] The standard orthogonal decomposition is performed on L by the following formula to obtain the eigenvalue and eigenvector:

[0076] ;

[0077] In the formula, the elements on the diagonal are the eigenvalues of U , and the columns of L are the eigenvectors of .

[0078] In the step S2, a roughing algorithm is proposed to rough the graph signals to obtain multi-scale rough graphs. s is a scale factor, n is the number of nodes of the rough graph, and The scale factor in the embodiment is s= 1, 2, …, 20.

[0079] For each scale factor, the roughing of the graph signals includes the following steps:

[0080] S21 clustering the graph signals to obtain a graph partition matrix.

[0081] In this step, first, each column of the feature vector matrix is normalized to [-1, 1], and then the absolute value of the feature vector less than 0.001 is regarded as 0. The non-zero rate (i.e. the ratio of non-zero feature values in the feature vector) of each feature vector is calculated, the standard deviation of the non-zero rates of all feature vectors is obtained, the feature vectors with non-zero rates higher than 3 times the standard deviation are selected, and the K-means clustering algorithm is used on the rows of the feature vectors. In the algorithm, the parameters k n , N The rows are clustered into n classes, and the graph partition matrix P is obtained according to the clustering result. i The element in the i-th row and the j-th column of the matrix j is as follows:

[0082] .

[0083] S22 calculating the adjacency matrix and the graph signal of the rough graph according to the graph partition matrix.

[0084] In this step, the adjacency matrix W c and the graph signal X c of the rough graph are calculated according to the graph partition matrix, and the calculation formula is as follows:

[0085] W c =P T WP ;

[0086] X c =P T X .

[0087] ​​For multiple scale factors, the above steps S21-S22 are repeated, thereby coarsening the image domain signal into multiple coarse images of different scales, that is, multi-scale coarse images.

[0088] In the above step S3, a graph complex entropy algorithm is proposed to extract the features of the test piece.

[0089] Defining graph complexity entropy ,The specific steps of calculating the complex entropy of the graph are as follows:

[0090] S31 reconstructs the phase space of the graph signal under each scale factor and constructs m dimensional vector sequence.

[0091] In this step, based on the image signal obtained in step S2, m The specific steps for constructing the vector sequence of dimension are as follows: X c (When the scale factor is 1, the signal is X ) to reconstruct the phase space and form n A vector sequence is the embedding dimension, that is, the dimension of the vector. In this embodiment, the embedding dimension is 2, where:

[0092] ;

[0093] .

[0094] S32 calculation m The Manhattan distance between the dimensional vector sequences is used as the similarity, and the statistical similarity is less than the threshold r The number of similarity probabilities is calculated to obtain the logarithmic mean value. P m ( r ), and similarly calculate the embedding dimension as m The logarithmic mean of the similarity probability under +1 P m+1 ( r ), P m ( r )and P m+1 ( r ) is E 1。

[0095] In this step, define the distance is a vector x m ( i )and x m ( j ) is the Manhattan distance, that is:

[0096] ;

[0097] For each vector x m ( i ), we need to calculate the Manhattan distance between it and other vectors (including its own), so for each vector x m ( i )have n distance, given a threshold r (Take 0.25 times the standard deviation of the time series), the statistical distance is less than the threshold r The number of P m ( r )for:

[0098] ;

[0099] Then calculate the embedding dimension m=3 Next P 3( r ),but E 1 is recorded as:

[0100] E 1= P 2( r )- P 3( r ).

[0101] S33 According to step S32 P m ( r )and P m+1 ( r ),calculate for E 2.

[0102] According to the calculation in S32 P 2( r )and P 3( r ),but E 2 is recorded as:

[0103] .

[0104] S34 uses the histogram method to estimate the empirical probability density function of the distance based on the Manhattan distance between the dimensional vector sequences obtained in step S32, and calculates E 3. , P bIt represents the probability of falling into the b interval of the histogram, and B represents the number of intervals in the histogram.

[0105] In this step, all Manhattan distances obtained in step S32 (excluding i = j The distance at the time of the test is estimated by using the histogram method. Assume that the histogram has B intervals (the number of intervals in this embodiment is 500), and the probability of all distances falling into each interval (that is, the percentage of distances falling into each interval to all distances) is: ,definition E 3 is:

[0106] .

[0107] S35 calculates the Pearson correlation coefficient between adjacent dimensional vector sequences. The correlation coefficient range is [-1, 1]. [-1, 1] is evenly divided into A subintervals, divide all correlation coefficients into subintervals, count the probability of falling into each subinterval, and calculate E 4. , P a Indicates that it falls into the subinterval a probability.

[0108] In this step, define is a vector x m ( i )and x m ( j ), that is:

[0109] ;

[0110] in, express x m ( i )middle m The average of the values, express x m ( j )middle m The average of the values.

[0111] Calculate the Pearson correlation coefficient between all adjacent vectors and get n- 1 correlation coefficient, the range of the Pearson correlation coefficient is [-1, 1], [-1, 1] is evenly divided into Asubintervals (the number of intervals in this embodiment is 50), all correlation coefficients are divided into subintervals, and the probability of falling into each subinterval is calculated. ,definition E 4 is:

[0112] .

[0113] S36 For each scale factor graph, calculate and get the multi-scale complex entropy of the graph , and use it as the feature vector of the device under test.

[0114] In this embodiment, for each scale factor, the values ​​calculated by S31-S35 are As the multi-scale complex entropy at the corresponding scale; the multi-scale complex entropy of all scale factors is constructed as the characteristic vector of the test piece.

[0115] For each scale factor s , get the rough graph through step S2, and repeat the above steps S31 to S36 to calculate four entropy values, namely the graph complex entropy. s Taking 1, 2, ... 20, one sample can obtain 80 entropy values, and these 80 entropy values ​​are constructed as the feature vector of the test piece.

[0116] In the above step S4, during the process of collecting the original vibration signal, the acceleration sensor is installed on the surface of the train axle box, and the train bearing fault signal is obtained by the acceleration sensor located on the surface of the axle box, and the vibration signal is used as the original signal for fault diagnosis.

[0117] In this embodiment, a random forest model is used as a train bearing fault classifier.

[0118] The above train bearing fault classifier is trained through the following steps:

[0119] (1) Based on the training set, the random forest model is used to filter the features in the feature vector, and the filtered features constitute the updated feature vector.

[0120] The training set is obtained by following the steps below: using different types of train bearing fault history sample data as training sample data, processing the training data according to the steps S1-S3 given above to obtain the feature vector of each sample; all samples and their feature vectors constitute the training set.

[0121] In this step, the random forest model is used to learn the training set. Based on the learning results, the mean impurity reduction (MDI) of each feature can be calculated as the feature importance of the feature.

[0122] The calculation process of mean impurity reduction (MDI) is as follows: Gini Impurity is used as the impurity indicator. The reduction in impurity of the feature in the decision tree before and after the data is split is calculated. The average of the impurity reduction of the feature in all decision trees is the mean impurity reduction.

[0123] The MDI calculation formula is:

[0124] ;

[0125] Where T represents the number of decision trees, Indicates the i The Gini impurity of the node before splitting in the tree, Indicates the i Gini impurity of the nodes after splitting in the tree.

[0126] The feature importance is set as the feature screening indicator, and the features that meet the requirements are screened out to form the updated feature vector. In this embodiment, as previously given, each sample has 80 entropy values ​​as feature values. The feature importance of these 80 entropy values ​​is sorted from large to small, and the top 20 features are screened out to form the updated feature vector.

[0127] Then, based on the updated eigenvector, the eigenvalues ​​of each sample in the training set are screened, and the training set is reconstructed using the screened eigenvalues ​​to obtain the reconstructed training set.

[0128] (3) Use the reconstructed training set to train the train bearing fault classifier.

[0129] In this step, the reconstructed training set is used to further train the random forest model trained in step (1) to obtain the final train bearing fault classifier. The random forest model is trained using conventional methods disclosed in the art. The training process is described in Diesel Engine Turbocharger Fault Diagnosis Based on Random Forest Method, Jia Zheyu, Wen Huabing, Zhu Junchao, Zhao Zhenyu, Ship Science and Technology, Vol. 45, No. 6, 2023.

[0130] For the part to be tested, the feature vector of the part to be tested is first screened based on the features screened in the previous step (1). The screened features are then input into the trained train bearing fault classifier to obtain the fault type of the part to be tested.

[0131] This embodiment uses the bearing vibration signal collected by the Railway Research and Education Center of the University of Birmingham for example verification.

[0132] Several samples of different fault types are collected. Seven different fault types are given here, as shown in Table 1. For each health state, 30 vibration signal samples are collected, for a total of 210 sample data. The training set and test set are divided in a 1:1 ratio. The 210 sample data are randomly divided according to this ratio.

[0133] Table 1 Classification of train bearing health status

[0134]

[0135] According to the steps S1-S3 given above, the multi-scale complex entropy of the graph of the training set and the test set samples is calculated.

[0136] According to the previous step (1), based on the training set, the random forest model is used to filter the features in the feature vector, and then the filtered features are used to reconstruct the training set to train the random forest model as the train bearing fault classifier to obtain the final train bearing fault classifier.

[0137] The trained train bearing fault classifier is then tested using the test set. According to the features selected in the previous step (1), the feature vectors composed of the multi-scale complex entropy of the graph corresponding to each sample in the test set are selected, and then the selected features are input into the trained train bearing fault classifier to obtain the prediction results of the fault type of each sample. The confusion matrix of the final classification result is as follows Figure 2 As shown in the figure, the results show that five sample fault types are classified incorrectly, and the classification accuracy is 95.2%. Therefore, this method can effectively and accurately diagnose different types and degrees of bearing faults, provide a higher accuracy for the bearing fault diagnosis results, realize the accurate diagnosis of train bearing fault types, effectively monitor train bearing faults, and improve the safety and economy of train operation.

[0138] Those skilled in the art will appreciate that the embodiments herein are intended to help readers understand the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific descriptions and embodiments. Those skilled in the art may make various other specific variations and combinations based on the technical teachings disclosed herein without departing from the essence of the present invention, and such variations and combinations are still within the scope of protection of the present invention.

Claims

1. A train bearing fault diagnosis method based on graph multi-scale complex entropy, comprising the following steps: S1 converts the collected original vibration signal of the test piece from the time domain to the image domain; S2 performs coarsening processing on the signal in the image domain to obtain a multi-scale coarse image, which specifically includes the following steps: S21 clusters the graph domain signal to obtain a graph segmentation matrix; S22 calculates the adjacency matrix and graph signal of the coarsened graph based on the graph partition matrix; Select multiple scale factors and repeat the above steps S21-S22 to coarsen the image domain signal into multiple coarse images of different scales, i.e., multi-scale coarse images; S3 calculates the multi-scale complex entropy of the DUT graph based on the multi-scale graph obtained in S2, and uses it as the feature vector of the DUT to define the graph complex entropy ,The specific steps of calculating the complex entropy of the graph are as follows: S31 reconstructs the phase space of the graph signal under each scale factor and constructs an m-dimensional vector sequence; S32 Calculate the Manhattan distance between the m-dimensional vector sequences as the similarity, count the number of similarities less than the threshold r, and calculate the logarithmic mean P of the similarity probability m (r), similarly calculate the logarithmic mean P of the similarity probability when the embedding dimension is m+1 m+1 (r), P m (r) and P m+1 The difference of (r) is E1; S33 Based on the P in step S32 m (r) and P m+1 (r), calculate is E2; S34 uses the histogram method to estimate the empirical probability density function of the distance based on the Manhattan distance between the m-dimensional vector sequences obtained in step S32, and calculates E3. , P b represents the probability of falling into the b interval of the histogram, and B represents the number of intervals in the histogram; S35 calculates the Pearson correlation coefficient between adjacent m-dimensional vector sequences. The correlation coefficient range is [-1, 1]. [-1, 1] is evenly divided into A subintervals. All correlation coefficients are divided into subintervals, and the probability of falling into each subinterval is counted. E4 is calculated. , P a represents the probability of falling into subinterval a; S36 For each scale factor, calculate , that is, the multi-scale complex entropy of the graph is obtained and used as the feature vector of the device under test; S4 inputs the features filtered out from the feature vector of the test piece into the train bearing fault classifier to obtain the train bearing fault classification result and complete the train bearing fault diagnosis.

2. The train bearing fault diagnosis method based on graph multi-scale complex entropy according to claim 1 is characterized in that: In step S1, the specific implementation method is to map the collected original time series signal into a two-dimensional matrix to obtain an undirected weighted graph; the mapping method is a visible graph, a horizontal visible graph or a k-nearest neighbor graph.

3. The train bearing fault diagnosis method based on graph multi-scale complex entropy according to claim 1 is characterized in that: In step S2, let the original image , W is the adjacency matrix, the graph Laplacian matrix L=DW, D is the degree diagonal matrix, for each node i in the graph, , taking the Euclidean distance between two nodes as the weight of the edge between the two nodes, , x i and x j are the values ​​of node i and node j respectively, and L is subjected to standard orthogonal decomposition to obtain the eigenvalue λ and eigenvector U; In step S21, each column of the eigenvector matrix U is normalized to [-1, 1], and the eigenvectors with absolute values ​​less than 0.001 are regarded as 0. The non-zero rate of each eigenvector is calculated, and the standard deviation of the non-zero rate is obtained. The eigenvectors with a non-zero rate greater than 3 times the standard deviation are selected, and the K-means clustering algorithm is applied to the rows of the eigenvectors. The parameter k=n in the algorithm is used, and N rows are clustered into n classes, where n is the number of nodes in the coarsened graph, and , s is the scale factor, and the graph segmentation matrix is ​​obtained according to the clustering results , the values ​​of the elements in the i-th row and j-th column of P are as follows: ； In step S22, the adjacency matrix W of the coarsened graph is calculated based on the graph partition matrix. c and image signal X c , the calculation formula is as follows: IN c =P T WP; X c =P T X; Among them, X is the original image signal.

4. The train bearing fault diagnosis method based on graph multi-scale complex entropy according to claim 1 is characterized in that: In step S31, based on the graph signal obtained in step S2, the m-dimensional vector sequence is constructed as follows: c Perform phase space reconstruction to form n vector sequences , is the embedding dimension, that is, the dimension of the vector, where: ； ； In step S32, the distance is defined is the vector x m (i) and x m The Manhattan distance of (j) is: ； For each vector x m (i), we need to calculate the Manhattan distance between it and other vectors, so for each vector x m (i) There are n distances, given a threshold r, and the number of distances less than the threshold r is counted. Define P m (r) is: ； Calculate P under the embedding dimension m+1 m+1 (r), then E1 is: E1=P m (r)-P m+1 (r)。 5. The train bearing fault diagnosis method based on graph multi-scale complex entropy according to claim 1 is characterized in that: In step S34, all Manhattan distances obtained in step S32 are not counted, and the distance when i=j is estimated using the histogram method. Assume that the histogram has B intervals, and the probability of all distances falling into each interval is calculated as , define E3 as: ； In step S35, define is the vector x m (i) and x m The Pearson correlation coefficient between (j), that is: ； in, Represents x m The average of the m values ​​in (i), Represents x m (j) The average of the m values; Calculate the Pearson correlation coefficient between all adjacent vectors and obtain n-1 correlation coefficients. The range of the Pearson correlation coefficient is [-1, 1]. Divide [-1, 1] into A subintervals on average. Divide all correlation coefficients into subintervals and calculate the probability of falling into each subinterval. , define E4 as: 。 6. The train bearing fault diagnosis method based on graph multi-scale complex entropy according to any one of claims 2 to 5, characterized in that: In step S36, for each scale factor, the graphs calculated in steps S31-S35 are As the multi-scale complex entropy at the corresponding scale; the multi-scale complex entropy of all scale factors is constructed as the characteristic vector of the test piece.

7. The train bearing fault diagnosis method based on graph multi-scale complex entropy according to claim 1 is characterized in that: In step S4, a random forest model is used as a train bearing fault classifier.

8. The train bearing fault diagnosis method based on graph multi-scale complex entropy according to claim 1 or 7, characterized in that: The train bearing fault classifier is trained through the following steps: (1) Based on the training set, the random forest model is used to filter the features in the feature vector, and the filtered features constitute the updated feature vector; (2) Reconstruct the training set based on the updated feature vector to obtain the reconstructed training set; (3) Use the reconstructed training set to train the train bearing fault classifier.

9. The train bearing fault diagnosis method based on graph multi-scale complex entropy according to claim 8 is characterized in that: In step (1), the training set is obtained according to the following steps: different types of train bearing fault history sample data are used as training sample data, and the training data are processed according to the steps S1-S3 given above to obtain the feature vector of each sample; all samples and their feature vectors constitute the training set.

10. The train bearing fault diagnosis method based on graph multi-scale complex entropy according to claim 9 is characterized in that: In step (2), based on the training results, the average impurity reduction (MDI), average accuracy reduction (MeanDecreaseAccuracy) or permutation importance of each feature is calculated as the feature importance of the feature, and the feature importance is set as the feature screening indicator to screen out the features that meet the requirements and form an updated feature vector; then, based on the updated feature vector, the feature value of each sample in the training set is screened, and the screened feature values ​​are used to reconstruct the training set to obtain a reconstructed training set.

Images