Bearing Fault Diagnosis Method Based on Heterogeneous Data Topological Association Embedding

Abstract

This invention discloses a bearing fault diagnosis method based on heterogeneous data topological association embedding, which solves the problem of insufficient modeling of the intrinsic representation relationship of heterogeneous data and effectively improves the fault prediction and health management of industrial bearings. The specific implementation process is as follows: (1) Collect heterogeneous data through sensors, construct an intrinsic association reconstruction module, and learn the intrinsic representation relationship between samples; (2) Construct a smooth association matrix and an intrinsic reconstruction weighted topology graph to maximize the correlation of heterogeneous data features, maintain topological association, and form a topologically constrained heterogeneous association embedding model; (3) Optimize the model to obtain the projection direction, and use spatial projection to directly obtain heterogeneous topological embedding features with good discriminative power, thereby obtaining the results of industrial bearing fault prediction and health management. Compared with the prior art, the method of this invention is more superior in industrial bearing fault prediction and health management.

Description

Technical Field

[0001] This invention relates to a bearing fault diagnosis method based on heterogeneous data topological association embedding, belonging to the field of machine learning and industrial fault prediction. Background Technology

[0002] Industrial bearings, as the core support and transmission components in rotating machinery systems, are widely used in wind power equipment, rail transportation, aerospace, CNC machine tools, and other industrial fields. Their fault prediction and health management directly determine the reliability and safety of the entire equipment. Traditional industrial bearing fault prediction methods are mainly based on signal processing technology, extracting fault features through time-domain statistical analysis, frequency-domain spectral analysis, and time-frequency domain analysis (such as wavelet transform and empirical mode decomposition), and then combining these with threshold judgments to achieve fault prediction and health management. With the development of artificial intelligence technology, machine learning-based industrial bearing fault prediction methods have gradually become mainstream. Algorithms such as support vector machines, backpropagation neural networks, and random forests are widely used in industrial bearing fault prediction and health management, achieving a leap from manual feature extraction to intelligent feature recognition. In recent years, deep learning technology, with its powerful automatic feature extraction capabilities, has further promoted the development of industrial bearing fault prediction technology. Models such as convolutional neural networks, long short-term memory networks, and generative adversarial networks have achieved good results in industrial bearing fault prediction and health management. However, most of the aforementioned methods model single-modal data (such as vibration signals). In actual industrial scenarios, the operating state of bearings is simultaneously reflected in multi-source heterogeneous data such as vibration, motor current, torque, and temperature. Single-modal data can only reflect local information about the state of industrial bearings and is easily affected by sensor noise and fluctuations in operating conditions, leading to insufficient reliability of prediction results. To address this, this invention obtains a set of heterogeneous industrial bearing fault data samples by collecting bearing signals through sensors. By learning the intrinsic representation relationships between samples and automatically adjusting the smoothing regularization coefficients based on data variance, an intrinsic correlation reconstruction module is constructed. Based on the reconstruction coefficients learned by the intrinsic correlation reconstruction module, a smoothing correlation matrix and an intrinsic reconstruction weighted topology graph are constructed to maximize the correlation of heterogeneous data features, further forming a topologically constrained heterogeneous correlation embedding model. The model is optimized and solved to obtain an analytical solution for the spatial projection direction of the heterogeneous data topological correlation embedding. Based on the projection direction, heterogeneous topological embedding features with good discriminative power are directly obtained, enhancing the topological correlation between heterogeneous data, reducing interference between different categories, and effectively improving the accuracy of fault prediction. Summary of the Invention

[0003] To capture the nonlinear manifold structure of data and address the problem of insufficient modeling of intrinsic representation relationships in heterogeneous data, this invention proposes an intrinsic correlation reconstruction module. This module learns the intrinsic representation relationships between samples and automatically adjusts the smoothing regularization coefficients based on data variance. Furthermore, a smoothing correlation matrix and an intrinsic reconstruction weighted topology graph are constructed to maximize the correlation of heterogeneous data features, forming a topologically constrained heterogeneous correlation embedding model. The model is then optimized to obtain an analytical solution for the spatial projection direction of the heterogeneous data topological correlation embedding. Based on the projection direction, heterogeneous topological embedding features with good discriminative power are directly obtained, leading to industrial bearing fault prediction and health management results. The specific implementation steps of this invention are as follows: 1. Bearing signals are acquired through sensors, and 18 time-domain features, 4 frequency-domain features, and 5 time-frequency-domain features are extracted to construct a heterogeneous data sample set of industrial bearing faults. and ,in To represent a mode, Indicates another mode, express The sample dimension express The sample dimension This indicates the number of samples. The samples are divided into training and test sets according to a certain ratio, and the test set data samples are randomized for each experiment.

[0004] 2. Construct an internally related refactoring module.

[0005] The specific construction steps of the intrinsic association reconstruction module are as follows: (2a) For any sample We assume that each sample can be reconstructed from a linear combination of other samples, that is: in For the sample For the sample The reconstruction coefficients are used to reconstruct the problem of all samples in matrix form: in Let represent the reconstructed coefficient matrix, and satisfy . .

[0006] (2b) In order to learn smooth and stable reconstruction coefficients, we construct the following regularized optimization problem: in for Norm, for Norm, To smooth out the regularization coefficients, and to address the poor generalization problem of fixed coefficients, we propose an intrinsic coefficient design: in For heterogeneous data sample sets The variance of all elements; (2c) Solving the regularization optimization problem, the gradient is: The proximal gradient descent method is used to solve the regularized optimization problem, with the learning rate based on... Constant determined: in Let be the spectral norm of the matrix; the iterative update formula is: in For the number of iterations, For soft thresholding function: Additionally, set the diagonal elements to 0 after each iteration: The iteration termination condition is: The intrinsic association reconstruction module learns the intrinsic representational relationships between samples and automatically adjusts the sparsity regularization coefficient based on the data variance.

[0007] 3. Construct a smooth correlation matrix and an intrinsically reconstructed weighted topology graph to form a topologically constrained heterogeneous correlation embedding model.

[0008] The specific construction steps of the topologically constrained heterogeneous association embedding model are as follows: Calculations based on heterogeneous data samples : Further, the distance matrix between samples is obtained. : In addition, define samples Minimum distance The sample set is Construct a smooth correlation matrix : Based on this, calculate : Constructing an intrinsically reconstructed weighted topology graph : Define hyperparameters To balance maximizing relevance and preserving topological association, a topologically constrained heterogeneous association embedding model is constructed: Topologically constrained heterogeneous association embedding models maximize the correlation of heterogeneous data features while maintaining the topological association of heterogeneous data, thus adapting to complex heterogeneous data distributions.

[0009] 4. The model is optimized and solved to obtain the analytical solution of the spatial projection direction of the heterogeneous data topology association embedding. Based on the projection direction, the heterogeneous topology embedding features with good discriminative power are directly obtained, and the results of industrial bearing fault prediction and health management are obtained.

[0010] The method of the present invention has the following advantages: (1) This invention addresses the machine learning modeling needs of industrial bearing fault diagnosis by constructing an intrinsic correlation reconstruction module. By learning the intrinsic representation relationship between samples and automatically adjusting the sparse regularization coefficient according to the data variance, it solves the problem of poor generalization of fixed parameters and improves the stability and accuracy of the reconstruction coefficient. (2) Based on the engineering application goal of fault prediction and health management, this invention constructs a topologically constrained heterogeneous association embedding model. Based on the reconstruction coefficients learned by the intrinsic association reconstruction module, it combines the Gaussian kernel function to perform weight smoothing and nearest neighbor selection, constructs an intrinsically reconstructed weighted topological graph that can adapt to complex data distribution, and derives the complete objective function and optimization solution process. While maximizing the cross-modal feature correlation, it provides effective unsupervised constraints through the topological graph regularization term, suppresses the learning of false correlations, and significantly improves the quality of the common subspace. (3) This invention optimizes the model through theoretical derivation to obtain an analytical solution for the spatial projection direction of heterogeneous data topological association embedding. By randomly sampling data, heterogeneous topological embedding features with good discriminative power are directly obtained based on the projection direction, which enhances the topological association between heterogeneous data, reduces interference between different categories, and effectively improves the accuracy of industrial bearing fault prediction. Attached Figure Description

[0011] The present invention will be further described below with reference to the accompanying drawings and examples.

[0012] Figure 1 This is a flowchart of the present invention, wherein... For the number of sensors, This represents the number of bearing failure categories.

[0013] Figure 2 It is the average accuracy under increasing sample size. Detailed Implementation

[0014] The specific implementation steps of this invention are as follows: 1. Bearing signals are acquired through sensors, and 18 time-domain features, 4 frequency-domain features, and 5 time-frequency-domain features are extracted to construct a heterogeneous data sample set of industrial bearing faults. and ,in To represent a mode, Indicates another mode, express The sample dimension express The sample dimension This indicates the number of samples. The samples are divided into training and test sets according to a certain ratio, and the test set data samples are randomized for each experiment.

[0015] 2. Based on the topologically constrained heterogeneous association embedding model, the model is as follows: The model is optimized and solved to obtain an analytical solution for the spatial projection direction of the heterogeneous data topological association embedding.

[0016] 3. By randomly sampling data, heterogeneous topological embedding features with good discriminative power are directly obtained based on the projection direction, thus yielding industrial bearing fault prediction and health management results.

[0017] The effectiveness of this invention was further verified through the following experiments: Experiments were conducted on the bearing datasets of the artificial damage test bench and accelerated life test bench at the University of Paderborn. The artificial damage test bench consisted of the following modules: (1) a motor; (2) a torque measurement shaft; (3) a rolling bearing test module; (4) a flywheel; and (5) a loading motor. The accelerated life test bench contained a bearing housing and a motor, which provided power to the shafts of four test bearings in the bearing housing. The bearings operated under the radial load applied by the spring screw mechanism. Three types of heterogeneous data from the dataset were selected for experimental verification: motor current signal, torque signal, and vibration signal. The sampling frequency of the motor current signal and vibration signal was 64 kHz, and the sampling frequency of the torque signal was 4 kHz. All test bearings were 6203 type rolling bearings. Two sets of artificial damage fault data, one type of accelerated life damage data, and one type of fault-free data were selected. The artificial damage was inner ring fault and outer ring fault, denoted as RF1 and RF2; the accelerated life damage was a mixed inner and outer ring fault, denoted as RF3; and the fault-free data was denoted as RF4. Both current and vibration signals were divided into 250 samples, with a single sample sampling length of 1024; torque signals were also divided into 250 samples, with a single sample sampling length of 64. To avoid experimental randomness, the same random experiment was repeated ten times, and the average of the ten experimental results was taken as the final average recognition rate. Figure 2 This visually demonstrates the average accuracy of bearing fault diagnosis under 10 randomized experiments with each class of training samples increasing in number. From Figure 2 As can be seen, the method of this invention has high accuracy, and the accuracy increases with the number of training samples, demonstrating good stability. Experimental results show that the method disclosed in this invention is an accurate and effective method for predicting industrial bearing failures and managing their health.

Claims

1. A bearing fault diagnosis method based on heterogeneous data topological association embedding, characterized in that, The method includes the following steps: (1) Bearing signals were collected by sensors, and 18 time-domain features, 4 frequency-domain features and 5 time-frequency-domain features were obtained through feature extraction to construct a heterogeneous data sample set of industrial bearing faults. and ,in To represent a mode, Indicates another mode, express The sample dimension express The sample dimension This indicates the number of samples, which are divided into training and test sets according to a certain ratio, and the test set data samples are randomized for each experiment. (2) Construct an internally related reconstruction module; (3) Construct a topologically constrained heterogeneous association embedding model; (4) The model is optimized and solved to obtain the analytical solution of the spatial projection direction of the heterogeneous data topology association embedding. Based on the projection direction, the heterogeneous topology embedding features with good discriminativeness are directly obtained, and the results of industrial bearing fault prediction and health management are obtained.

2. The bearing fault diagnosis method based on heterogeneous data topology association embedding according to claim 1, characterized in that, Step (2) involves constructing the intrinsic association reconstruction module, and the steps are as follows: (2a) For any sample We assume that each sample can be reconstructed from a linear combination of other samples, that is: in For the sample For the sample The reconstruction coefficients are used to reconstruct the problem of all samples in matrix form: in Let represent the reconstructed coefficient matrix, and satisfy . ; (2b) In order to learn smooth and stable reconstruction coefficients, we construct the following regularized optimization problem: in for Norm, for Norm, To smooth out the regularization coefficients, and to address the poor generalization problem of fixed coefficients, we propose an intrinsic coefficient design: in For heterogeneous data sample sets The variance of all elements; (2c) Solving the regularization optimization problem, the gradient is: The proximal gradient descent method is used to solve the regularized optimization problem, with the learning rate based on... Constant determined: in Let be the spectral norm of the matrix; the iterative update formula is: in For the number of iterations, For soft thresholding function: Additionally, set the diagonal elements to 0 after each iteration: The iteration termination condition is: The intrinsic association reconstruction module learns the intrinsic representational relationships between samples and automatically adjusts the smoothing regularization coefficients based on the data variance.

3. The bearing fault diagnosis method based on heterogeneous data topology association embedding according to claim 1, characterized in that, Step (3) involves constructing a topologically constrained heterogeneous association embedding model, which is performed as follows: To maximize the correlation of heterogeneous data features while maintaining the topological association of heterogeneous data, a topologically constrained heterogeneous association embedding model is constructed by combining the intrinsic association reconstruction module: in is a hyperparameter used to balance maximizing correlation and preserving topological association; Weighted topology for internal reconstruction: in Defined as: in The smoothed incidence matrix is ​​defined as follows: in For the sample Minimum distance A sample set, The distance matrix between samples: in Calculations based on heterogeneous data samples: Based on the reconstruction coefficients learned by the intrinsic association reconstruction module, the topologically constrained heterogeneous association embedding model can adapt to complex heterogeneous data distributions.

Images