Rolling bearing fault diagnosis method of digital twin drive CNN-BIGRU learning model

Through the digital twin-driven CNN-BIGRU learning model, the three-dimensional modeling and generation adversarial network are used to generate samples, solve the problem of difficulty in collecting bearing failure data, realize efficient diagnosis of rolling bearing failures, and improve equipment reliability and safety.

CN119066382APending Publication Date: 2024-12-03ZHENGZHOU UNIVERSITY OF LIGHT INDUSTRY
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202411177121.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-26
Publication Date
2024-12-03

Smart Images

  • Figure CN119066382A_ABST
    Figure CN119066382A_ABST
Patent Text Reader

Abstract

The invention discloses a rolling bearing fault diagnosis method of a digital twin driven CNN-BIGRU learning model. Firstly, a deep groove ball bearing is used as a research object to establish a rolling bearing digital twinborn model; secondly, on the basis of the established digital twinning model, kinetic analysis is carried out through finite element analysis software and an explicit kinetic algorithm, and preset fault bearing twinning data are obtained; and finally, inputting twin fault data into the CNN-BIGRU fault classification model, dividing a training set and a test set, establishing a rolling bearing fault diagnosis intelligent model, and judging the fault type of the rolling bearing. Therefore, the data obtained from the twin model can be used for fault diagnosis of real situations. The method provided by the invention can effectively solve the problem of bearing data sample missing under different working conditions, and the diagnosis accuracy of the fault bearing is remarkably improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a rolling bearing fault diagnosis method of a digital twin driven CNN-BIGRU learning model, and relates to the technical field of rolling bearing fault diagnosis. Background Art

[0002] Bearings are key components in mechanical equipment. During operation, they are often in a high-speed, high-load working environment, and are prone to defects such as pitting, spalling, and wear. Studying the mechanism of bearing failure is an important topic in the engineering field, and is crucial to improving equipment reliability, reducing costs, and ensuring safety. Studies have shown that bearings are the most prone to failure in machinery. Only about 10% of bearings can run to the designed length of time. Bearing failure may cause equipment failure and accidents, posing potential dangers to workers and the environment. Severe bearing failure may cause the entire production line to shut down, causing huge economic losses. In addition, the actual working conditions of bearings are extremely complex, and fault label data is difficult to collect. Summary of the invention

[0003] In view of the problem that it is difficult to obtain sample data labels and the sample size is small, the present invention discloses a rolling bearing fault diagnosis method based on a digital twin driven CNN-BIGRU learning model, comprising the following steps:

[0004] S1: Output twin data;

[0005] S2: enhanced data sample;

[0006] S3: data classification and partitioning labels;

[0007] S4: Define the CNN-BIGRU network model;

[0008] S5: Diagnose fault.

[0009] Furthermore, step S1 is specifically as follows:

[0010] S1-1: Constructing a 3D bearing model

[0011] In the modeling software UG, the three-dimensional bearing model of the rolling bearing in a healthy state is drawn according to the physical parameters of the SKF6205 bearing.

[0012] S1-2: Grid division

[0013] Import the 3D model of the bearing into Hypermesh for meshing. Use SOLID MAP to clean and segment each part, obtain a series of SOLIDs that meet the MAP conditions, and then perform 3D meshing separately. After meshing, the number of units in the 3D model of the bearing is 450816, and the number of nodes is 501136. Hexahedral unit mesh is used for each part of the bearing. For the key spheres in the bearing, the size is adjusted and the mesh is refined to more accurately reflect the actual situation. In the software, shell164 shell meshes are established for the inner and outer rings respectively, and set as rigid body materials to facilitate the subsequent application of loads and constraints.

[0014] S1-3: Setting materials

[0015] In view of the dynamic characteristics of rolling bearings, when selecting model materials, all bearing components are uniformly set as elastic materials; in the ANSYS-workbench interface, GCr15 steel is used as the material for the inner ring, outer ring, ball, and cage of the bearing. The density is set to 7850kg / m 3 , the elastic modulus is 2.00×10^11Pa, and the Poisson's ratio is 0.3.

[0016] S1-4: Setting contact

[0017] During normal operation of rolling bearings, the contact modes are mainly divided into the contact between the inner ring and the rolling element, the contact between the outer ring and the rolling element, and the contact between the rolling element and the cage. The friction coefficient between the inner and outer ring raceways and the rolling element is set to 0.1, and the friction coefficient between the cage and the rolling element is set to 0.05.

[0018] S1-5: Applying loads and constraints

[0019] In LS-prepost, the inner and outer ring meshes are respectively connected to the inner and outer ring shell meshes. The loads and constraints applied to the inner and outer shells are transferred to the inner and outer ring meshes. The outer ring of the bearing is fixed to limit its movement and rotational freedom in the x, y, and z directions. Due to the loads and constraints applied to the inner ring, the rotational freedom in the y and z directions and the movement freedom in the x direction are limited. Apply a force of 2000N in the y direction and set the speed required for the current working condition with the x axis as the rotation axis direction.

[0020] S1-6: Data Simulation

[0021] The twin model is calculated in ANSYS to obtain the model twin data, and the software is continuously adjusted to obtain the twin data.

[0022] Furthermore, step S2 is specifically as follows:

[0023] S2-1: Generate samples

[0024] The twin data is input into the generator, which continuously learns the distribution of real sample data and generates generated data that the discriminator cannot distinguish.

[0025] S2-2: Using the U-Net network architecture, the size and number of channels of the feature map are reduced through the depth layer and the pooling layer, and Wassertein GAN with gradient penalty is added. The formula is as follows:

[0026]

[0027] Where α is the penalty coefficient, is the actual data distribution P r The purpose of uniform sampling along the straight line between the points of the generated distribution is to make the convergence of the loss function more stable during training.

[0028] S2-3: Use the acceleration signal data of real samples and the acceleration signal data of generated samples for training to prevent overfitting

[0029] S2-4: Sample differentiation

[0030] The discriminator is used to distinguish between real samples and generated samples, so that the discriminator and the generator can be trained adversarially. The objective function is set as follows:

[0031] m G inm D axE x~Pr(x) [logD(x)]+E z ~P z (z)[log(1-D(G(z)))]

[0032] where x is from the real data distribution P r (x) is the real data sampled, and z is from a uniform distribution P z (z) is a noise vector sampled from the generator G, which generates synthetic data G(z). This equation shows that treating the discriminator as a classifier can minimize the difference between the true data distribution and the data distribution assumed by the generator.

[0033] Furthermore, step S3 is specifically as follows:

[0034] S3-1: Classified Data

[0035] The generated data is resampled with a signal length of 1024 and a sliding window overlap of 0.5. The resampled data is then placed in another table with one row corresponding to one signal, and a classification label is added at the end.

[0036] S3-2: Create datasets and labels

[0037] Convert data and labels from Pandas data structures into tensors that PyTorch can handle for training and prediction in neural networks.

[0038] Furthermore, step S4 is specifically as follows:

[0039] S4-1: Define the network architecture

[0040] CNN-BIGRU consists of input, feature extraction, distance measurement and output stages. The main structural parameters are shown in Table 1. Two layers of convolution are used. Each convolution layer is followed by a pooling layer. The features output by the second pooling layer are input to the BIGRU layer. Finally, a fully connected layer is connected to map the features into feature vectors F(x). The two feature vectors F(x) output i ) and F(x j ) is:

[0041] D i,j =||F(x i )-F(x j )||

[0042] The calculated distance is input into the fully connected layer FC with a sigmoid activation function to obtain a probability value P i,j Represents the similarity between the input sample pairs. The larger the probability value, the more similar the two samples are, and the greater the possibility that they belong to the same category.

[0043] P i,j =sigmoid{FC[||F(x i )-F(y i )||]}

[0044] Two samples are randomly selected from the training set to form a sample input into the model, and the output is the similarity P obtained by the following formula i,j , to determine whether the input samples belong to the same category. Then calculate P i,j With sample pair label T i,j The cross entropy loss function is used and the Adam optimizer is used to optimize the network.

[0045] L(x i ,x i )=T i,j log e (P i,j )+(1-T i,j )log e (1-P i,j )

[0046] Furthermore, step S5 is specifically as follows:

[0047] First, randomly select a sample x from each of the 10 categories of training samples. i , with different labels y i , forming the support set.

[0048] S={(x1,x1)…(x 10 ,x 10 )}

[0049] Secondly, for the test samples to be classified The 10 sample pairs are input into the CNN-BIGRU twin network model in sequence to obtain the probability that the test sample belongs to the same class as the 10 training samples. The bearing state with the largest probability value is selected as the bearing state of the test sample. The calculation formula is:

[0050]

[0051] The final support set consists of 10 different classes, each with 5 samples (S1, ..., S N ), and then similar to the previous test, determine the category of the test sample from the 10 subsets

[0052]

[0053] Determine the probability that the test sample and a certain fault in the support set belong to the same category. The fault type with the highest probability is this bearing.

[0054] Compared with the prior art, the present invention has the following effects:

[0055] The purpose of the present invention is to provide a rolling bearing fault diagnosis method driven by a digital twin-driven CNN-BIGRU learning model. First, the bearing is three-dimensionally modeled using the three-dimensional modeling software UG, meshing is performed in Hypermesh, contact settings are performed in workbench to output a K file, and the K file is imported into Ls-prepost for keyword settings to output the corresponding twin data. Then the output corresponding twin data is used to generate samples using the data in the generative adversarial network, and the generated real samples are distinguished from the generated samples by the discriminator. Finally, the real samples and the generated samples are input into the CNN-BIGRU learning model to divide the labels and divide the training set, test set and verification set according to the ratio of 7:2:1. A sample is randomly selected from each of the 10 categories of training samples, and different labels are provided to form a support set. The 10 sample pairs are sequentially input into the CNN-BiGRU twin network model to obtain the probability that the test sample belongs to the same category as the 10 training samples, and the bearing state with the largest probability value is selected as the bearing state of the test sample. Twin data is generated by constructing a digital twin model of rolling bearings, and the difference between real fault data and virtual twin data is eliminated through the generative adversarial network field. The real fault data and the data generated by the adversarial network are then applied to the fault diagnosis model, which solves the problem of difficulty in collecting fault data and data missing in actual working conditions of rolling bearings. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Schematic diagram of rolling bearing mesh division;

[0057] Figure 2 It is the contact diagram between the outer ring and the rolling element;

[0058] Figure 3 It is the contact diagram between the inner ring and the rolling element;

[0059] Figure 4 It is the contact diagram between rolling element and cage;

[0060] Figure 5 Provide a flow chart for digital twin modeling and finite element analysis;

[0061] Figure 6 This is a schematic diagram of the Gan network structure;

[0062] Figure 7 It is a CNN-BiGRU network structure model;

[0063] Figure 8 Schematic diagram of rolling bearing fault diagnosis based on digital twin. DETAILED DESCRIPTION

[0064] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0065] A rolling bearing fault diagnosis method driven by a digital twin-driven CNN-BIGRU learning model comprises the following steps:

[0066] S1: Output twin data:

[0067] S1-1: Constructing a 3D bearing model

[0068] In the modeling software UG, the three-dimensional bearing model of the rolling bearing in a healthy state is drawn according to the physical parameters of the SKF6205 bearing.

[0069] S1-2: Grid division

[0070] Import the 3D model of the bearing into Hypermesh for meshing. Use SOLID MAP to clean and segment each part, obtain a series of SOLIDs that meet the MAP conditions, and then perform 3D meshing separately. After meshing, the number of units in the 3D model of the bearing is 450816, and the number of nodes is 501136. Hexahedral unit mesh is used for each part of the bearing. For the key spheres in the bearing, the size is adjusted and the mesh is refined to more accurately reflect the actual situation. In the software, shell164 shell meshes are established for the inner and outer rings respectively, and set as rigid body materials to facilitate the subsequent application of loads and constraints.

[0071] S1-3: Setting materials

[0072] In view of the dynamic characteristics of rolling bearings, when selecting model materials, all bearing components are uniformly set as elastic materials; in the ANSYS-workbench interface, GCr15 steel is used as the material for the inner ring, outer ring, ball, and cage of the bearing. The density is set to 7850kg / m 3 , the elastic modulus is 2.00×10^11Pa, and the Poisson's ratio is 0.3.

[0073] S1-4: Setting contact

[0074] During normal operation of rolling bearings, the contact modes are mainly divided into the contact between the inner ring and the rolling element, the contact between the outer ring and the rolling element, and the contact between the rolling element and the cage. The friction coefficient between the inner and outer ring raceways and the rolling element is set to 0.1, and the friction coefficient between the cage and the rolling element is set to 0.05.

[0075] S1-5: Applying loads and constraints

[0076] In LS-prepost, the inner and outer ring meshes are respectively connected to the inner and outer ring shell meshes. The loads and constraints applied to the inner and outer shells are transferred to the inner and outer ring meshes. The outer ring of the bearing is fixed to limit its movement and rotational freedom in the x, y, and z directions. Due to the loads and constraints applied to the inner ring, the rotational freedom in the y and z directions and the movement freedom in the x direction are limited. Apply a force of 2000N in the y direction and set the speed required for the current working condition with the x axis as the rotation axis direction.

[0077] S1-6: Data Simulation

[0078] The twin model is calculated in ANSYS to obtain the model twin data, and the software is continuously adjusted to obtain the twin data.

[0079] S2: Enhanced data samples

[0080] S2-1: Generate samples

[0081] As a separate network structure in the network model, the generator can generate samples in random noise and latent space. The twin data is input into the generator, which continuously learns the distribution of real sample data and generates generated data that the discriminator cannot distinguish.

[0082] S2-2: Using the U-Net network architecture, the size and number of channels of the feature map are reduced through the depth layer and the pooling layer, and Wassertein GAN with gradient penalty is added. The formula is as follows:

[0083]

[0084] Where α is the penalty coefficient, is the actual data distribution P r The purpose of uniform sampling along the straight line between the points of the generated distribution is to make the convergence of the loss function more stable during training.

[0085] S2-3: Use the acceleration signal data of real samples and the acceleration signal data of generated samples for training to prevent overfitting

[0086] S2-4: Sample differentiation

[0087] The discriminator is used to distinguish between real samples and generated samples, so that the discriminator and the generator can be trained adversarially. The objective function is set as follows:

[0088]

[0089] where x is from the real data distribution Pr (x) is the real data sampled, and z is from a uniform distribution P z (z) is a noise vector sampled from the generator G, which generates synthetic data G(z). This equation shows that treating the discriminator as a classifier can minimize the difference between the true data distribution and the data distribution assumed by the generator.

[0090] S3: Data classification and partitioning labels

[0091] S3-1: Classified Data

[0092] The generated data is resampled with a signal length of 1024 and a sliding window overlap of 0.5. The resampled data is then placed in another table with one row corresponding to one signal, and a classification label is added at the end.

[0093] S3-2: Create datasets and labels

[0094] Convert data and labels from Pandas data structures into tensors that PyTorch can handle for training and prediction in neural networks.

[0095] S4: Define the CNN-BIGRU network model

[0096] S4-1: Define the network architecture

[0097] CNN-BIGRU consists of input, feature extraction, distance measurement and output stages. The main structural parameters are shown in Table 1. Two layers of convolution are used. Each convolution layer is followed by a pooling layer. The features output by the second pooling layer are input to the BIGRU layer. Finally, a fully connected layer is connected to map the features into feature vectors F(x). The two feature vectors F(x) output i ) and F(x j ) is:

[0098] D i,j =||F(x i )-F(x j )||

[0099] The calculated distance is input into the fully connected layer FC with a sigmoid activation function to obtain a probability value P i,j Represents the similarity between the input sample pairs. The larger the probability value, the more similar the two samples are, and the greater the possibility that they belong to the same category.

[0100] P i,j =sigmoid{FC[||F(x i )-F(y i )||]}

[0101] Two samples are randomly selected from the training set to form a sample input into the model, and the output is the similarity P obtained by the following formula i,j , to determine whether the input samples belong to the same category. Then calculate P i,j With sample pair label T i,j The cross entropy loss function is used and the Adam optimizer is used to optimize the network.

[0102] L(x i ,x i )=T i,j log e (P i,j )+(1-T i,j )log e (1-P i,j )

[0103] S5: Diagnosing faults

[0104] First, randomly select a sample x from each of the 10 categories of training samples. i , with different labels y i , forming the support set.

[0105] S={(x1,x1)…(x 10 ,x 10 )}

[0106] Secondly, for the test samples to be classified The 10 sample pairs are input into the CNN-BIGRU twin network model in sequence to obtain the probability that the test sample belongs to the same class as the 10 training samples. The bearing state with the largest probability value is selected as the bearing state of the test sample. The calculation formula is:

[0107]

[0108] The final support set consists of 10 different classes, each with 5 samples (S1, ..., S N ), and then similar to the previous test, determine the category of the test sample from the 10 subsets

[0109]

[0110] Determine the probability that the test sample and a certain fault in the support set belong to the same category. The fault type with the highest probability is this bearing.

[0111] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and improved concepts of the present invention within the technical scope disclosed by the present invention, and they should be covered by the protection scope of the present invention.

Claims

1. A rolling bearing fault diagnosis method based on a digital twin-driven CNN-BIGRU learning model, characterized in that: The following steps are involved: S1: Output twin data; S2: enhanced data sample; S3: data classification and partitioning labels; S4: Define the CNN-BIGRU network model; S5: Diagnose fault.

Citation Information

Patent Citations

  • Rolling bearing fault diagnosis method of digital twin drive deep transfer learning model

    CN116952583A

  • Bearing fatigue period prediction method based on maximum entropy and generative adversarial network

    CN117332315A

  • Full-process automatic calculation method and device for impact resistance of fiber reinforced composite material

    CN118484981A