High-speed train bogie bearing fault diagnosis method and device

By combining non-overlapping sampling and graph convolutional neural network in bearing fault diagnosis, adaptive graph structure and framelet transformation are solved, and the problem of low fault recognition accuracy in traditional methods in strong noise environments is realized, and the high-speed train bogie is realized The multiple types of fault diagnosis of bearings improve diagnostic accuracy and stability.

CN119066487APending Publication Date: 2024-12-03GUANGXI UNIV

Patent Information

Application Number
CN202411091899.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-09
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

Traditional signal-based bearing fault diagnosis methods are flooded in the fault characteristic information in a strong noise environment, resulting in low fault recognition accuracy and difficulty in characterizing high-dimensional and complex nonlinear relationships, which limits the accuracy of fault diagnosis.

Method used

Using overlapping sampling to divide the vibration signal into multi-segment sub-samples, build a learning-adaptive graph structure, combine framelet transformation and graph convolution neural network to form an AGFCN model, extract node features and classify them through graph convolution layer to reduce noise Interference and improve fault recognition accuracy.

Benefits of technology

Effectively reduce noise interference, adaptively learn the feature information of the graph, realize accurate diagnosis of multiple types of faults of high-speed train bogie bearings under complex working conditions, improve fault recognition accuracy, and realize intelligent fault diagnosis of bearings.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119066487A_ABST
    Figure CN119066487A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of bearing fault diagnosis, in particular to a high-speed train bogie bearing fault diagnosis method and device, and the method comprises the steps: collecting a vibration signal in the operation process of a train bogie bearing, and dividing the original vibration signal into a plurality of sections of subsamples with fixed lengths through non-overlapping sampling; a subsample is regarded as a node on the graph, a time domain signal is regarded as a node feature, a fault type is regarded as a node label, and a graph structure adapting to the complex and changeable working conditions of the high-speed train bogie bearing is constructed by adopting a learnable and adaptive graph adjacency matrix; the method comprises the following steps of: constructing graph convolution with framelet transformation to learn node representation on a graph, extracting node features through two graph convolution layers, performing node classification by using two full-connection layers, and combining a self-adaptive graph, a framelet transformation technology and a graph convolution neural network to obtain an AGFCN model; and finally performing training and optimization by using the training set, and inputting the test set into the trained AGFCN model for fault diagnosis.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault diagnosis, and in particular to a method and device for diagnosing bearing faults of a high-speed train bogie. Background Art

[0002] High-speed trains have become the first choice for alleviating congestion and green travel in major cities around the world due to their advantages such as large capacity, high speed, safety and punctuality. Bearings are the support of mechanical transmission shafts and an important guarantee for realizing the functions of high-speed train bogies. Their health is closely related to the safe and reliable operation of trains. Due to the complex operating conditions of bogie bearings and long-term operation in harsh conditions such as high speed, heavy load and strong coupling impact, the occurrence of bearing failures is aggravated, which seriously threatens the safety of high-speed train operation. Therefore, the health of bogie bearings under strong noise and variable load conditions is of great significance.

[0003] Traditional signal-based bearing fault diagnosis methods require manual extraction of feature information from signals, and then use such methods to extract feature information from vibration signals, construct feature vectors, input them into machine learning algorithms, and output bearing fault diagnosis results. In fact, fault feature information is submerged in strong noise environments, limiting the method's feature extraction capabilities. Shallow machine learning methods have difficulty representing high-dimensional, complex nonlinear relationships, which restricts the accuracy of fault diagnosis. Summary of the invention

[0004] The purpose of the present invention is to provide a high-speed train bogie bearing fault diagnosis method and device, aiming to solve the problem of low fault identification accuracy caused by strong noise conditions in existing signal-based bearing fault diagnosis technology.

[0005] To achieve the above object, the present invention provides a high-speed train bogie bearing fault diagnosis method and device, comprising the following steps:

[0006] Step 1: collecting vibration signals during the operation of the train bogie bearing, wherein the vibration signals include a normal state vibration signal and a fault state vibration signal of the bogie bearing;

[0007] Step 2: Divide the vibration signal into a plurality of sub-samples of fixed length using non-overlapping sampling;

[0008] Step 3: Consider the subsample as a node on the graph, use the time domain signal as the node feature, and the fault type as the node label. Use a learnable and adaptive graph adjacency matrix to construct the graph structure.

[0009] Step 4: Build a graph convolution model with framelet transformation to learn the node representation on the graph, extract node features through two graph convolution layers, use two fully connected layers to classify nodes, and obtain the AGFCN model by combining the adaptive graph, framelet transformation technology with the graph convolutional neural network;

[0010] Step 5: Divide the non-overlapping vibration signal data into a training set and a test set in proportion, use the training set to train and optimize the parameters of the AGFCN model, and then input the test set into the trained AGFCN model for fault judgment.

[0011] Optionally, the subsample in step 2 is represented by Multiple fixed-length subsample sets are recorded as:

[0012]

[0013] Where Π is the obtained sample set, X is the subsample, Y is the label, and N is the number of subsamples.

[0014] Optionally, the execution process of step 3 includes the following steps:

[0015] Step 3.1: Define a non-negative function A ij =g(X i ,X j )(i,j∈{1,2,...,τ}) to represent two nodes X i and X j Based on the feature matrix X i connection relationship;

[0016] g(X i ,X j ) is implemented through a layer of neural network with a learnable weight vector w=(w1,w2,...,w s ) T ∈R s ×1 ;

[0017] The learnable adaptive graph adjacency matrix can be defined as:

[0018]

[0019] Among them, ReLU is to ensure A ij Non-negative activation function, softmax operation normalizes each row of A;

[0020] Step 3.2: Update the weight vector w by the following loss function:

[0021]

[0022] Optionally, the execution process of step 4 includes the following steps:

[0023] Step 4.1: The framelet of the undirected weighted graph G = (V, E, w) ∈ l2 (G) is defined by a filter bank and the spectrum of the graph Laplacian matrix L;

[0024] For ξ∈R, the filter bank satisfies the following extraction formula:

[0025]

[0026] Where a and b(r) are the low-pass and high-pass filters of the framelet transform, respectively, which are used to extract and represent the approximate and detailed information of the image signal, and the scaling function Ψ={α;β (1) ,...,β (n)} is a set of frame generators corresponding to the filter bank η, used to generate framelets;

[0027] For the eigenvalue and eigenvector pairs of the graph Laplacian matrix L Define the framelets of graph G at scale level j=1,...,J according to the scaling function:

[0028]

[0029]

[0030] in and are the low-pass and high-pass framelets of the node p, respectively. j and p are the dilation at scale j and the translation at node p∈V, respectively.

[0031] Step 4.2: Define the low-pass and high-pass framelet coefficients of the signal f on the graph G as:

[0032]

[0033] where v0 and are respectively low-pass and high-pass filters for the signal f;

[0034] For the eigenpairs of the graph Laplacian matrix L Let U=[u1,u2,...,u N ] is the eigenvector matrix, Λ=diag([λ1,λ2,...,λ N ]) is the diagonal of the eigenvalue, then there is the following formula:

[0035]

[0036]

[0037] in and are the filter diagonal matrices with low-pass and high-pass filters respectively;

[0038] The framelet coefficients are expressed as follows:

[0039]

[0040] Step 4.3: Construct the framelet and framelet transformation of the graph signal f, for a set of coefficients on the graph G The framelet reconstruction can be expressed as:

[0041]

[0042] Step 4.4: Derive the tensor-based framelet transformation and reconstruct it as:

[0043] f J =W T c=Vc

[0044] Step 4.5: For a graph G with N nodes, its framelet convolution is defined as follows:

[0045] θ*X=V(Shrinkage(diag(θ)(WX′))),X′=XW θ

[0046] where θ is the network filter, X∈R N×d is the input feature of the graph, W θ ∈R d×d′ is the learnable graph filter parameter matrix, d′ is the dimension reduction, WX′ is the framelet coefficient matrix after X′ is transformed by the decomposition operator W, Shrinkage is the signal compression activation function, and V is the reconstruction operator;

[0047] The Shrinkage activation function is defined as follows:

[0048] Shrinkage(x)=sgn(x)(|x|-λ) +

[0049] Where sgn() is the symbolic function and x is the input feature;

[0050] Step 4.6: After the framelet convolution learns the node representation, it is input into the fully connected layer for node classification, which is expressed as:

[0051] Z = softmax(FC(θ*X))

[0052] Where Z is the set of predicted labels, and the cross entropy loss is used for classification. The cross entropy loss on all labeled nodes is expressed as:

[0053]

[0054] Where I is the number of samples, C is the number of fault categories, y is the true label, is the model prediction label;

[0055] Combining adaptive graph construction and graph framelet convolution, the loss function of the final model can be expressed as:

[0056] L loss =L grah_learning +L CE .

[0057] Optionally, in step 5, a sliding window is used to truncate non-overlapping original signals, and each obtained fault data set is divided into a training set and a test set in a ratio of 7:3.

[0058] Furthermore, the present invention also proposes a high-speed train bogie bearing fault diagnosis device, which is used in the high-speed train bogie bearing fault diagnosis method, comprising a signal acquisition module, a signal processing and feature selection module and a fault diagnosis module, wherein the signal acquisition module, the signal processing and feature selection module and the fault diagnosis module are electrically connected;

[0059] The signal acquisition module is used to collect vibration signals during the rotation of the bearing and store the vibration signals in a computer device through serial communication;

[0060] The signal processing and feature selection module is used to extract feature information, construct feature vectors, and obtain new feature representations through reconstruction operators;

[0061] The fault diagnosis module is used to train and optimize target samples, and use the trained target samples to perform fault diagnosis on high-speed train bogie bearings, thereby achieving fault classification of high-speed train bogie bearings.

[0062] Among them, the signal acquisition module includes a signal collector, a vibration acceleration sensor and a signal converter. The vibration acceleration sensor is installed at the upper end and side end of the bearing support frame. When the bearing starts to move, the vibration acceleration sensor converts the vibration of the bearing into an electrical signal. The electrical signal is amplified by the signal collector, and then converted into vibration data using the signal converter and stored in a computer.

[0063] The present invention provides a high-speed train bogie bearing fault diagnosis method and device. First, the vibration signal of the train bogie bearing during operation is collected by a vibration acceleration sensor, and the original vibration signal is divided into multiple sub-samples of fixed length by non-overlap sampling; the sub-sample is regarded as a node on the graph, the time domain signal is used as the node feature, the fault type is used as the node label, and a learnable adaptive graph adjacency matrix is ​​used to construct a graph structure adapted to the complex and changeable working conditions of the high-speed train bogie bearing; a graph convolution model with framelet transformation is built to learn the node representation on the graph, the node features are extracted by two graph convolution layers, and two fully connected layers are used to classify the nodes, and the AGFCN model is obtained by combining the adaptive graph and framelet transformation technology with the graph convolution neural network; finally, the data set is divided into a training set and a test set in proportion, the parameters of the AGFCN model are trained and optimized by using the training set, and then the test set is input into the trained AGFCN model for fault diagnosis. The present invention can effectively reduce the interference of noise, adaptively learn the characteristic information of the graph, can effectively realize the diagnosis of multiple types of faults of the high-speed train bogie bearing under complex working conditions, improve the fault identification accuracy, and realize the intelligent fault diagnosis of the bearing. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0065] Figure 1 The present invention is a schematic flow chart of the steps of a high-speed train bogie bearing fault diagnosis method.

[0066] Figure 2 It is a schematic diagram of the calculation process of the adaptive graph framelet convolutional neural network proposed in the present invention.

[0067] Figure 3 It is a structural schematic diagram of a high-speed train bogie bearing fault diagnosis device of the present invention.

[0068] Figure 4 It is a schematic diagram of the fault diagnosis results of 5 experiments of a specific embodiment of the present invention.

[0069] Figure 5 It is a time domain waveform diagram and a spectrum diagram of an original signal and a SNR=-10 signal according to a specific embodiment of the present invention.

[0070] Figure 6It is a schematic diagram of the fault diagnosis results obtained by each method in different noise environments in the specific embodiments of the present invention. DETAILED DESCRIPTION

[0071] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0072] The following is an explanation of the English abbreviations used in this invention:

[0073] AGFCN: Adaptive Graph Framelet Convolutional Network;

[0074] GCN: Graph Convolutional Neural Network;

[0075] GAT: Graph Attention Network;

[0076] GIN: Graph Isomorphism Network;

[0077] MRF-GCN: Multi-receptive field graph convolutional neural network;

[0078] MSCNN: Multi-branch Multi-scale Convolutional Network.

[0079] See also Figure 1 The present invention provides a high-speed train bogie bearing fault diagnosis method, comprising the following steps:

[0080] S1: collecting vibration signals during the operation of the train bogie bearing, wherein the vibration signals include a normal state vibration signal and a fault state vibration signal of the bogie bearing;

[0081] S2: Divide the vibration signal into a plurality of sub-samples of fixed length using non-overlapping sampling;

[0082] S3: The subsample is regarded as a node on the graph, the time domain signal is used as the node feature, the fault type is used as the node label, and the graph structure is constructed using a learnable and adaptive graph adjacency matrix;

[0083] S4: Build a graph convolution model with framelet transformation to learn node representation on the graph, extract node features through two graph convolution layers, use two fully connected layers to classify nodes, and obtain the AGFCN model by combining adaptive graph, framelet transformation technology with graph convolutional neural network;

[0084] S5: Divide the non-overlapping vibration signal data into a training set and a test set in proportion, use the training set to train and optimize the parameters of the AGFCN model, and then input the test set into the trained AGFCN model for fault judgment.

[0085] The following is further described in conjunction with specific embodiments and execution steps:

[0086] In step S2, the original vibration signal is divided into multiple sub-samples of fixed length using non-overlapping sampling, each sub-sample is represented as Multiple fixed-length subsample sets are recorded as:

[0087]

[0088] Where Π is the obtained sample set, X is the subsample, Y is the label, and N is the number of subsamples.

[0089] In step S3, using a learnable and adaptive graph adjacency matrix to construct a graph structure that adapts to the complex and variable working conditions of high-speed train bogie bearings includes the following sub-steps:

[0090] S0301: Define a non-negative function A ij =g(X i ,X j )(i,j∈{1,2,...,τ}) represents the connection relationship between two nodes Xi and Xj based on the feature matrix Xi. g(Xi,Xj) is implemented by a layer of neural network, which has a learnable weight vector w=(w1,w2,...,w s )T∈R s×1 The learnable adaptive graph adjacency matrix can be defined as:

[0091]

[0092] Among them, ReLU is the activation function that ensures that Aij is non-negative, and the softmax operation normalizes each row of A;

[0093] S0302: Update the weight vector w through the following loss function:

[0094]

[0095] In step S4, a GFCN model with framelet transformation is built to learn the node representation on the graph, node features are extracted through two GFCN layers, and node classification is performed using two fully connected layers, including the following sub-steps:

[0096] S0401: The framelet of an undirected weighted graph G = (V, E, w) ∈ l2 (G) is defined by a filter bank and the spectrum of the graph Laplacian matrix L. For a framelet system, its filter bank η = {a; b (1) ,...,b (n)}∈l0(Z) is composed of a tight support sequence. a and b(r) are the low-pass and high-pass filters of the framelet transform, respectively, which are used to extract and represent the approximate and detailed information of the image signal. The scaling function Ψ={α;β (1) ,...,β (n)} is a set of frame generators corresponding to the filter bank η, which is used to generate framelets. At the same time, α, β(r) and their Fourier transforms belong to the l2(R) space. For ξ∈R, the filter bank satisfies the following extraction formula:

[0097]

[0098] For the eigenvalue and eigenvector pairs of the graph Laplacian matrix L According to the above scaling function, we can define the framelets of graph G at scale level j=1,...,J:

[0099]

[0100]

[0101] in and are the low-pass and high-pass framelets of the node p, respectively, and j and p are the dilation at scale j and the translation at node p∈V, respectively.

[0102] S0402: According to the above graph framework, the low-pass and high-pass framelet coefficients of the signal f on the graph G can be defined as:

[0103]

[0104] where v0 and They are low-pass and high-pass filters of signal f respectively. The framelet transform is the graph signal f and its framelet coefficients For the eigenvalues ​​of the graph Laplacian matrix L Let U = [u1, u2, ..., uN] be the eigenvector matrix, Λ = diag ([λ1, λ2, ..., λ N ]) is the diagonal of the eigenvalue, then there is the following formula:

[0105]

[0106]

[0107] in and are the filter diagonal matrices with low-pass and high-pass filters respectively. Then the framelet coefficients can be expressed as follows:

[0108]

[0109] S0403: The framelet and framelet transform of the graph signal f can be constructed. In addition, the decomposition and reconstruction of the framelet are reversible. (1) ,...,b (n)} to achieve decomposition and reconstruction. In the experiment, a Haar type filter bank was used. By repeatedly using the filter extraction formula, the following framelet transformation matrix can be obtained:

[0110]

[0111] where K is a real value such that the maximum eigenvalue of the graph Laplacian matrix L satisfies λ max ≤2 K π. However, the complexity of matrix decomposition is very high, and it is difficult to effectively perform decomposition and reconstruction when framelet transform is used. Drawing on the approach in GCN, the Chebyshev polynomial T0,...,Tn is used to approximate the filter bank, that is, a≈T0, b (r) ≈T r . Then the framelet transformation matrix above can be expressed as:

[0112]

[0113] According to the properties of matrix polynomials and the characteristic decomposition formula of the graph Laplace matrix UΛU T =L can be converted into:

[0114]

[0115] The decomposition and reconstruction of the framelet can be expressed by the transformation matrix Wr,j. The framelet decomposition can be expressed as:

[0116]

[0117] Where W 0,j f and W r,j f are the low-pass and high-pass coefficients of signal f, j represents the scale level, and r=1,...,n is the number of high-pass. For a set of coefficients on graph G The framelet reconstruction can be expressed as:

[0118]

[0119] S0404: Based on the above framelet transformation, we can derive the tensor-based framelet transformation. First, define a decomposition operator W∈R (nJ+1)N×N , which can be expressed as follows:

[0120] W=[W 0,J ,W 1,J ,...,W n,J ,...,W 1,J ,...,W n,J ] T

[0121] Then, set a column vector of length (nJ+1)N. Let c be the coefficient It is associated with a framelet system with n high-pass and N node graphs. The framelet decomposition can be expressed as:

[0122] c=Wf

[0123] Finally, a reconstruction operator V is defined, which is the permutation and reorganization of the decomposition operator W in the framelet transformation, that is, V = WT. The framelet reconstruction can be expressed as:

[0124] f J =W T c=Vc

[0125] S0405: Based on the graph framelet transformation, we can define a graph framelet convolution similar to the spectral domain graph convolution. For a graph G with N nodes, its framelet convolution is defined as follows:

[0126] θ*X=V(Shrinkage(diag(θ)(WX′))),X′=XW θ

[0127] where θ is the network filter, X∈R N×d is the input feature of the graph, W θ ∈R d×d′is a learnable graph filter parameter matrix, d′ is the dimension reduction dimension, WX′ is the framelet coefficient matrix after X′ is transformed by the decomposition operator W, Shrinkage is the signal compression activation function, and V is the reconstruction operator. Unlike the ReLU and Sigmoid activation functions commonly used in general graph convolutions, the graph framelet convolution uses the Shrinkage activation function. Since the model in this article needs to analyze data from multiple scales, the "one-size-fits-all" principle in most activation functions will destroy the multi-scale characteristics of the framelet representation. The Shrinkage threshold can adapt to variable coefficient scales and provide more accurate truncation values. The Shrinkage activation function is defined as follows:

[0128] Shrinkage(x)=sgn(x)(|x|-λ) +

[0129] Where sgn() is the sign function and x is the input feature. For N coefficients, is a threshold value, which only affects the smaller high-pass framelet coefficients and controls the signal component value, further compressing the signal. In addition, the Shrinkage function needs to be used before the reconstruction operator V as the Shrinkage threshold of the high-pass coefficients in the framelet. The Shrinkage function can reduce the noise of the framelet coefficients and compress the frequency signal, thereby assisting the graph framelet convolution. The constructed graph framelet convolution layer is as follows Figure 2 shown.

[0130] S0406: After the framelet convolution learns the node representation, it is input into the fully connected layer for node classification, which can be expressed as:

[0131] Z = softmax(FC(θ*X))

[0132] Where Z is the set of predicted labels. Using cross entropy loss for classification, the cross entropy loss on all labeled nodes can be expressed as:

[0133]

[0134] Where I is the number of samples, C is the number of fault categories, y is the true label, is the model prediction label. Combining adaptive graph construction and graph framelet convolution, we can get the loss function of the final model, which can be expressed as:

[0135] L loss =L grah_learning +L CE

[0136] In step S5, the total data set is divided into a training set and a test set in proportion, including the following sub-steps: using a sliding window to truncate non-overlapping original signals. After data segmentation, each fault data set obtained is divided into a training set and a test set in a ratio of 7:3.

[0137] Furthermore, in step S5, when the AGFCN model is trained and optimized, if the AGFCN model shows good fault diagnosis accuracy on both the training set and the test set, it indicates that the constructed high-speed train bogie bearing fault diagnosis method based on graph convolution learning can be used to realize bearing fault diagnosis.

[0138] See also Figure 3 , Figure 3 It is a structural schematic diagram of a high-speed train bogie bearing fault diagnosis device of the present invention, which specifically includes a signal acquisition module, a signal processing and feature extraction module and a fault diagnosis module;

[0139] The signal acquisition module collects vibration signals during the bearing rotation process by installing vibration acceleration sensors vertically and radially on the bearing support frame, and stores the vibration signals in the computer device through serial communication;

[0140] Signal processing and feature selection module, according to the graph framelet transformation, for a graph G with N nodes, its framelet convolution is defined as follows:, where θ is the network filter, is the input feature of the graph, is the learnable graph filter parameter matrix, is the dimension reduction dimension, is the framelet coefficient matrix after the decomposition operator W transformation, Shrinkage is the signal compression activation function, and V is the reconstruction operator. The decomposition operator W transforms the features with low-pass and high-pass filter operators. These operators can be reorganized into an N×N sparse matrix, and the coefficients can be obtained by matrix multiplication. By applying the filter θ to the framelet coefficients, the graph learning is propagated in the frequency domain. After the signal passes through the filter θ, the Shrinkage activation function is used to truncate the high-frequency components to activate the filter coefficients. Finally, the new feature representation is obtained through the reconstruction operator;

[0141] The fault diagnosis module is used to train and optimize the target samples, and use the trained target samples to diagnose the faults of the high-speed train bogie bearings, so as to realize the fault classification of the high-speed train bogie bearings.

[0142] The signal acquisition module includes a signal collector, a vibration acceleration sensor and a signal converter. The vibration acceleration sensor is installed at the upper end and side end of the bearing support frame. When the bearing starts to move, the vibration acceleration sensor converts the vibration of the bearing into an electrical signal. The electrical signal is amplified by the signal collector and then converted into vibration data using the signal converter and stored in a computer.

[0143] Furthermore, in order to verify the effectiveness of the fault diagnosis of the present invention, in the embodiments, the method of the present invention is compared with five advanced algorithms in the field of bearing fault diagnosis, and the comparison methods include graph convolutional neural network (GCN), graph attention network (GAT), graph isomorphism network (GIN), multi-receptive field graph convolutional neural network (MRF-GCN) and multi-branch multi-scale convolutional network (MSCNN).

[0144] Specifically, in order to ensure the reliability of the fault diagnosis results of AGFCN and the comparison algorithm, 5 experiments were conducted for each method and the average and standard deviation of the five experimental results were calculated. The results are shown in Table 1. In order to clearly show the advantages of AGFCN, a bar chart is drawn, as shown in Table 1. Figure 4 . The standard deviation of the diagnosis results of the AGFCN algorithm in 5 fault diagnosis experiments was 0.1%, which was 0.23%, 0.08%, 0.07%, 0.06% and 0.45% smaller than the standard deviation of the diagnosis results of the other 5 comparison methods, respectively. This shows that the stability of AGFCN is the best. At the same time, it can be seen that the four comparison graph neural network models also have good stability, while the stability of MSCNN is poor. The average value of the 5 fault diagnosis results of the AGFCN algorithm is 99.92%, and the results of two diagnoses reached 100%, indicating that AGFCN has good diagnostic performance. Compared with the other 5 comparison algorithms, the accuracy of AGFCN has increased by 0.42%, 0.18%, 0.67%, 0.11% and 2.66% respectively. Based on the above analysis, compared with the 5 advanced fault diagnosis models, AFGCN has a better advantage, which proves the reliability of the proposed method.

[0145] Table 1. Each method was tested 5 times. The average and standard deviation of the results of the five experiments were calculated.

[0146] Methods first / % Second time / % The third time / % Fourth time / % Fifth time / % average value / % Standard Deviation / % AGFCN 100 99.89 100 99.79 99.87 99.91 0.09 GCN 99.89 99.57 99.79 99.68 99.57 99.7 0.14 GAT 99.47 99.79 99.89 99.68 99.36 99.64 0.22 GIN 99.36 99.26 99.36 99.47 99.79 99.45 0.21 MRF-GCN 99.15 99.57 99.68 99.25 99.47 99.42 0.22 MSCNN 98.62 98.72 98.83 98.94 98.51 98.72 0.17

[0147] In order to verify the advantages of AGFCN in fault diagnosis in a strong noise environment, five signal-to-noise ratio (SNR) scenarios were set to simulate the operation of the bearing under the condition of gradually increasing noise. Figure 5 The time domain waveform and spectrum diagram of the original signal and the signal with SNR = -10. In the original signal, the time domain signal is simple and the pulse signal in the frequency domain signal is obvious, which helps to accurately diagnose the bearing fault. When SNR = -10, the time domain signal is more complex than the original signal, and some pulse signals in the frequency domain are submerged by noise, which increases the difficulty of bearing fault diagnosis. Each method was run 5 times in different scenarios, and the fault diagnosis results were as follows: Figure 6As shown. When SNR = -2, the diagnostic accuracy of AGFCN decreased by 0.57%, and the diagnostic accuracy of GCN, GAT, GIN, MRF-GCN and MSCNN decreased by 4.59%, 3.79%, 5.52%, 0.93% and 15.53% respectively. It can be determined that in a noisy environment, the diagnostic accuracy of AGFCN is the highest, and the difference in diagnostic results compared with the original signal is small. Except for MRF-GCN, the diagnostic accuracy of other methods decreased by more than 1%. When SNR = -4, the diagnostic accuracy of AGFCN is 98.04, which is 1.3% lower than that of AGFCN when SNR = -2. Among the other five comparison methods, MRF-GCN has the highest diagnostic accuracy of 95.42%, and the diagnostic accuracy of other comparison methods is less than 95%. The worst is that the diagnostic accuracy of MSCNN is less than 80%. As the SNR changes, the diagnostic accuracy of each method gradually decreases. When SNR = -10, the diagnostic accuracy of AFFCN still reaches 85.65% and the stability is the best. The diagnostic accuracy of the other five comparison methods is less than 80%. The diagnostic accuracy of AFFCN is 20.54%, 8.67%, 29.8%, 7.2% and 18.99% higher than that of the other five comparison methods, respectively. The difference in fault diagnosis results is more prominent, which also proves that AFFCN has significant fault diagnosis advantages in strong noise environments, indicating that the framelet transformation graph convolution proposed in this invention plays a very critical role in reducing noise interference.

[0148] In summary, the present invention has the following beneficial technical effects:

[0149] (1) The present invention utilizes the idea of ​​dynamic learning to construct an adaptive graph structure learning method, which can dynamically learn the structural characteristics of the graph, thereby better adapting to the complex working conditions of HSTB bearings.

[0150] (2) The present invention combines the framelet transformation technology with graph convolution and proposes a graph convolutional neural network based on framelet transformation, which can effectively reduce the noise information of the graph and improve the ability of model feature extraction under strong noise conditions.

[0151] (3) The present invention constructs a high-speed train bogie bearing fault diagnosis framework under complex working conditions, which can effectively reduce noise interference and adaptively learn the feature information of the graph, and can effectively realize the diagnosis of multiple types of high-speed train bogie bearing faults under complex working conditions.

[0152] (4) The bearing fault diagnosis device designed in the present invention has the characteristics of fast diagnosis speed, high fault diagnosis accuracy, light and efficient model, and strong anti-noise interference ability. It can meet the practical application needs of bearing fault diagnosis and also has great application potential in other application fields.

[0153] What is disclosed above is only a preferred embodiment of the present invention, and it certainly cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of the above embodiment and equivalent changes made according to the claims of the present invention still fall within the scope of the invention.

Claims

1. A high-speed train bogie bearing fault diagnosis method, characterized in that: The following steps are involved: Step 1: collecting vibration signals during the operation of the train bogie bearing, wherein the vibration signals include a normal state vibration signal and a fault state vibration signal of the bogie bearing; Step 2: Divide the vibration signal into a plurality of sub-samples of fixed length using non-overlapping sampling; Step 3: Consider the subsample as a node on the graph, use the time domain signal as the node feature, and the fault type as the node label. Use a learnable and adaptive graph adjacency matrix to construct the graph structure. Step 4: Build a graph convolution model with framelet transformation to learn the node representation on the graph, extract node features through two graph convolution layers, use two fully connected layers to classify nodes, and obtain the AGFCN model by combining the adaptive graph, framelet transformation technology with the graph convolutional neural network; Step 5: Divide the non-overlapping vibration signal data into a training set and a test set in proportion, use the training set to train and optimize the parameters of the AGFCN model, and then input the test set into the trained AGFCN model for fault judgment.

2. The high-speed train bogie bearing fault diagnosis method according to claim 1, characterized in that: The subsample described in step 2 is represented by Multiple fixed-length subsample sets are recorded as: Π=[(X1,Y1),(X2,Y2),...,(X N ,Y N )], Where Π is the obtained sample set, X is the subsample, Y is the label, and N is the number of subsamples.

3. The high-speed train bogie bearing fault diagnosis method according to claim 2, characterized in that: The execution process of step 3 includes the following steps: Step 3.1: Define a non-negative function A ij =g(X i ,X j )(i,j∈{1,2,...,τ}) to represent two nodes X i and X j Based on the feature matrix X i connection relationship; g(X i ,X j ) is implemented through a layer of neural network with a learnable weight vector w=(w1,w2,...,w s ) T ∈R s×1 ; The learnable adaptive graph adjacency matrix can be defined as: Among them, ReLU is to ensure A ij Non-negative activation function, softmax operation normalizes each row of A; Step 3.2: Update the weight vector w by the following loss function:

4. The high-speed train bogie bearing fault diagnosis method according to claim 3, characterized in that: The execution process of step 4 includes the following steps: Step 4.1: The framelet of the undirected weighted graph G = (V, E, w) ∈ l2 (G) is defined by a filter bank and the spectrum of the graph Laplacian matrix L; For ξ∈R, the filter bank satisfies the following extraction formula: Where a and b(r) are the low-pass and high-pass filters of the framelet transform, respectively, which are used to extract and represent the approximate and detailed information of the image signal, and the scaling function Ψ={α;β (1) ,...,β (n) } is a set of frame generators corresponding to the filter bank η, used to generate framelets; For the eigenvalue and eigenvector pairs of the graph Laplacian matrix L Define the framelets of graph G at scale level j=1,...,J according to the scaling function: in and are the low-pass and high-pass framelets of the node p, respectively. j and p are the dilation at scale j and the translation at node p∈V, respectively. Step 4.2: Define the low-pass and high-pass framelet coefficients of the signal f on the graph G as: where v0 and are respectively low-pass and high-pass filters for the signal f; For the eigenpairs of the graph Laplacian matrix L Let U=[u1,u2,...,u N ] is the eigenvector matrix, Λ=diag([λ1,λ2,...,λ N ]) is the diagonal of the eigenvalue, then there is the following formula: in and are the filter diagonal matrices with low-pass and high-pass filters respectively; The framelet coefficients are expressed as follows: Step 4.3: Construct the framelet and framelet transformation of the graph signal f, for a set of coefficients on the graph G The framelet reconstruction can be expressed as: Step 4.4: Derive the tensor-based framelet transformation and reconstruct it as: f J =W T c=You Step 4.5: For a graph G with N nodes, its framelet convolution is defined as follows: θ*X=V(Shrinkage(diag(θ)(WX′))),X′=XW θ where θ is the network filter, X∈R N×d is the input feature of the graph, W θ ∈R d×d′ is the learnable graph filter parameter matrix, d′ is the dimension reduction, WX′ is the framelet coefficient matrix after X′ is transformed by the decomposition operator W, Shrinkage is the signal compression activation function, and V is the reconstruction operator; The Shrinkage activation function is defined as follows: Shrinkage(x)=sgn(x)(|x|-λ) + Where sgn() is the symbolic function and x is the input feature; Step 4.6: After the framelet convolution learns the node representation, it is input into the fully connected layer for node classification, which is expressed as: Z = softmax(FC(θ*X)) Where Z is the set of predicted labels, and the cross entropy loss is used for classification. The cross entropy loss on all labeled nodes is expressed as: Where I is the number of samples, C is the number of fault categories, y is the true label, is the model prediction label; Combining adaptive graph construction and graph framelet convolution, the loss function of the final model can be expressed as: L loss =L grah_learning +L CE .

5. The high-speed train bogie bearing fault diagnosis method according to claim 4, characterized in that: In step 5, a sliding window is used to truncate the non-overlapping original signal, and each fault data set obtained is divided into a training set and a test set in a ratio of 7:

3.

6. A high-speed train bogie bearing fault diagnosis device, used in the high-speed train bogie bearing fault diagnosis method as claimed in claim 5, characterized in that: It includes a signal acquisition module, a signal processing and feature selection module and a fault diagnosis module, wherein the signal acquisition module, the signal processing and feature selection module and the fault diagnosis module are electrically connected; The signal acquisition module is used to collect vibration signals during the rotation of the bearing and store the vibration signals in a computer device through serial communication; The signal processing and feature selection module is used to extract feature information, construct feature vectors, and obtain new feature representations through reconstruction operators; The fault diagnosis module is used to train and optimize target samples, and use the trained target samples to perform fault diagnosis on high-speed train bogie bearings, thereby achieving fault classification of high-speed train bogie bearings.

7. The high-speed train bogie bearing fault diagnosis method and device according to claim 6, characterized in that: The signal acquisition module includes a signal collector, a vibration acceleration sensor and a signal converter. The vibration acceleration sensor is installed at the upper end and side end of the bearing support frame. When the bearing starts to move, the vibration acceleration sensor converts the vibration of the bearing into an electrical signal. The electrical signal is amplified by the signal collector and then converted into vibration data using the signal converter and stored in a computer.

Citation Information

Patent Citations

  • Fault diagnosis method based on adaptive graph neural network multi-source data fusion

    CN117332818A

  • Rail transit train bearing fault diagnosis method and device based on meta-learning algorithm

    CN117668689A

  • Bearing fault diagnosis method and system based on hybrid graph neural network of GAT and GCN

    CN118260532A

  • Bearing fault diagnosis method based on fuzzy broad learning model

    US11928613B1

Cited By

  • Railway bogie bearing fault diagnosis method and device based on feature extraction network

    CN120873693A

  • A Fault Diagnosis Method and Equipment for Railway Bogie Bearings Based on Feature Extraction Networks

    CN120873693B