Bearing fault diagnosis method based on unsupervised depth discrimination transfer learning

Abstract

The invention discloses a bearing fault diagnosis method based on unsupervised depth discrimination transfer learning, and the method comprises the steps: collecting an initial bearing vibration signal under an actual working condition, carrying out the preprocessing, and constructing preprocessing data; constructing a feature extractor based on a convolutional neural network, and performing preliminary feature extraction on the source domain data set and the target domain data set; performing pseudo label generation and dynamic updating based on the source domain feature data and the target domain feature data; carrying out feature alignment based on joint distribution adaptation loss; and training and optimizing the unsupervised deep discrimination transfer learning network, and carrying out bearing state diagnosis on the target domain feature data. According to the method, marginal distribution and conditional distribution are considered, and two loss metrics of MMD and CORAL are combined, so that adaptive alignment of source domain and target domain features is successfully realized, and the accuracy and robustness of cross-domain bearing fault diagnosis are remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of mechanical engineering, and particularly relates to a bearing fault diagnosis method based on unsupervised deep discriminative transfer learning. BACKGROUND

[0002] In the technical field of mechanical engineering, bearings as core rotating components are prone to failure under long-term high load, high temperature and harsh environment. Once a failure occurs, it may cause serious mechanical failure, and even endanger equipment and personnel safety. Therefore, bearing fault diagnosis is crucial to ensure the safe and stable operation of mechanical equipment. Bearing failures usually occur in the inner ring, outer ring and rolling elements. Among them, vibration signals as the most common monitoring method can directly reflect the running state and potential failure of bearings.

[0003] With the rapid development of intelligent technology, data-driven fault diagnosis methods have become a research hotspot. In particular, the introduction of deep learning and transfer learning technology has greatly improved the diagnostic accuracy and adaptability. However, existing bearing fault diagnosis technologies still face challenges such as insufficient data, improper feature selection and poor adaptability to complex working conditions. Traditional signal processing methods require a large amount of manual feature extraction, which is often difficult to fully reflect the fault state of bearings. At the same time, the fault mode of bearings is complex, and environmental factors such as noise, temperature and load have a greater impact on the signal, increasing the difficulty of fault diagnosis.

[0004] Transfer learning technology can effectively solve the problems of insufficient data and poor model generalization by transferring knowledge from the source domain to the target domain. In particular, in bearing fault diagnosis, source domain data usually comes from experimental benches, containing rich fault modes and labels, while target domain data may come from actual train operation, with large differences in working conditions. In order to achieve effective fault diagnosis in the target domain, the working condition differences between the source domain and the target domain must be fully considered, and appropriate transfer learning methods must be applied for knowledge transfer. SUMMARY

[0005] The application provides a bearing fault diagnosis method based on unsupervised deep discriminative transfer learning to solve the above technical problems.

[0006] To solve the above technical problems, one technical solution adopted by the application is: a bearing fault diagnosis method based on unsupervised deep discriminative transfer learning, comprising:

[0007] S1. Collect initial bearing vibration signals under actual working conditions and pre-process them to construct pre-processed data;

[0008] S2. Constructing a feature extractor based on a convolutional neural network based on preprocessed data, performing preliminary feature extraction on the source domain and target domain data sets, and obtaining source domain feature data and target domain feature data;

[0009] S3. Generating and dynamically updating pseudo labels based on source domain feature data and target domain feature data;

[0010] S4. Constructing an unsupervised deep discriminative transfer learning network based on joint distribution adaptation loss;

[0011] S5. Training and optimizing the unsupervised deep discriminative transfer learning network, and performing bearing state diagnosis on the target domain feature data.

[0012] Further, the method of step S1 comprises:

[0013] S11. Collecting bearing vibration signals with real labels from an experimental platform and unlabeled bearing vibration signals from a high-speed train in a real scenario to obtain initial vibration signals;

[0014] S12. Based on the initial vibration signals, resampling is performed to unify the sampling frequency to 32 kHz to obtain first vibration signals;

[0015] S13. Based on the first vibration signals, band-pass filtering is used for denoising, and then data augmentation is performed by introducing Gaussian noise to obtain second vibration signals;

[0016] S14. Based on the second vibration signals, preprocessed data of source domain data with real labels and target domain data without labels are constructed.

[0017] Further, the method of step S2 comprises:

[0018] S21. Constructing a feature extractor based on a convolutional neural network;

[0019] S22. Based on the feature extractor, input the source domain data to obtain the source domain features;

[0020] S23. Based on the feature extractor, input the target domain data to obtain the target domain features.

[0021] Further, the method of step S3 comprises:

[0022] S31. Based on the feature extractor and Softmax classifier trained by the source domain data set, input the samples in the target domain data to obtain pseudo labels;

[0023] S32. Based on the prediction confidence, calculate the posterior probability of the samples in the target domain data belonging to the pseudo labels;

[0024] S33. If the posterior probability is greater than or equal to the prediction confidence, the pseudo label is kept unchanged;

[0025] S34. If the posterior probability is less than the prediction confidence, a random other fault type is selected as the new pseudo label.

[0026] Further, the method of step S4 comprises:

[0027] S41. Obtain the marginal distribution alignment loss between the source domain and the target domain based on formula (1) ; wherein formula (1) is:

[0028] (1) ;

[0029] wherein, is the source domain feature; is the target domain feature;

[0030] S42. Obtain the conditional distribution alignment loss between the source domain and the target domain based on formula (2) ; wherein formula (2) is:

[0031] (2) ;

[0032] wherein, is the total number of categories, and are the sample distributions of the source domain and the target domain under the category

[0033] S43. Based on formula (3), the joint distribution adaptation loss for aligning the source domain and the target domain feature distributions is formed by combining the marginal distribution alignment loss and the conditional distribution alignment loss ; wherein formula (3) is:

[0034] (3) ;

[0035] S44. Based on formula (4), the marginal distribution difference and the conditional distribution difference in the loss function are evaluated; wherein formula (4) is:

[0036] (4) ;

[0037] wherein, indicates the maximum mean difference, indicates the correlation alignment; specifically as shown in the following formulas (5)-(6):

[0038] (5) ;

[0039] (6) ;​

[0040] wherein, and are the number of samples in the source domain and the target domain, respectively, is a Gaussian kernel mapping function, denotes the norm in the reproducing kernel Hilbert space, Cov and Cov denote the covariance matrix of the source domain and the target domain, respectively, is the dimension of the eigenvector, denotes the Frobenius norm.

[0041] S45. Based on formula (7), the final loss is formed, which is specifically represented as follows:

[0042]

[0043] (7).

[0044] S46. Based on the final loss and the back propagation algorithm, the network parameters of the unsupervised deep discriminative transfer learning network are updated.

[0045] Further, the method of step S5 comprises:

[0046] S51. Based on formula (8), the classification loss of the source domain is calculated; wherein, formula (8) is:

[0047] (8).

[0048] wherein, is the feature score of the source domain sample on the class , is the score of the th sample on the th class;

[0049] S52. Based on formula (9), the classification loss of the target domain is calculated; wherein, formula (9) is:

[0050] (9).

[0051] wherein, is the feature score of the target domain sample on the pseudo-label class , is the score of the th target domain sample on the th class;

[0052] S53. Calculate the total classification loss based on formula (10) ; wherein formula (10) is:

[0053] (10);

[0054] S54. Obtain the final target loss of the unsupervised deep discriminative transfer learning network based on formula (11) ; wherein formula (11) is:

[0055] (11);

[0056] wherein,

[0057] S55. Train and optimize the unsupervised deep discriminative transfer learning network based on the final target loss;

[0058] S56. Input the test samples of the target domain into the trained unsupervised deep discriminative transfer learning network, and finally output the classification result.

[0059] The beneficial effects of the present application are: the present application considers the marginal distribution and the conditional distribution, combines the MMD and CORAL two loss metrics, successfully realizes the adaptive alignment of the source domain and the target domain features, and significantly improves the accuracy and robustness of the cross-domain bearing fault diagnosis. The present application uses a convolutional neural network to extract deep features of the source domain and the target domain, and through the combination of feature mapping and domain adaptation mechanism, not only improves the accuracy of feature extraction, but also enhances the adaptability of the model under different working conditions. The present application solves the problem of unlabeled data in the target domain through pseudo-label generation and dynamic updating technology, and provides a new solution for the cross-machine fault diagnosis task. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 is a flowchart of an embodiment of the bearing fault diagnosis method based on unsupervised deep discriminative transfer learning of the present application;

[0061] Figure 2 is a flowchart of an embodiment of step S1 in Figure 1 ;

[0062] Figure 3 is a flowchart of an embodiment of step S2 in Figure 1 ;

[0063] Figure 4 is a flowchart of an embodiment of step S3 in Figure 1 ;

[0064] Figure 5 is a UMAP visualization diagram before transfer of the present application;

[0065] Figure 6 UMAP visualization diagram after migration of the present application;

[0066] Figure 7 MDA loss curve diagram of marginal distribution of an embodiment of the bearing fault diagnosis method based on unsupervised deep discriminative transfer learning of the present application;

[0067] Figure 8 CDA loss curve diagram of conditional distribution of an embodiment of the bearing fault diagnosis method based on unsupervised deep discriminative transfer learning of the present application;

[0068] Figure 9 Loss comparison diagram of source domain and target domain of an embodiment of the bearing fault diagnosis method based on unsupervised deep discriminative transfer learning of the present application;

[0069] Figure 10 Training loss curve diagram of the unsupervised deep discriminative transfer learning network of the present application;

[0070] Figure 11 Training and testing accuracy curve diagram of the unsupervised deep discriminative transfer learning network of the present application. DETAILED DESCRIPTION

[0071] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with specific embodiments.

[0072] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, however, the present application can also be implemented in other ways different from those described herein, therefore, the present application is not limited to the specific embodiments disclosed in the following description.

[0073] Reference Figure 1 , Figure 1 is a flowchart of an embodiment of the bearing fault diagnosis method based on unsupervised deep discriminative transfer learning of the present application. The method comprises:

[0074] S1. Collect initial bearing vibration signals under actual working conditions and pre-process to construct pre-processed data.

[0075] Specifically, referring to Figure 2 , the method of step S1 comprises:

[0076] S11. Collect bearing vibration signals with real labels from an experimental platform and unlabeled bearing vibration signals from a high-speed train under real conditions to obtain initial vibration signals.

[0077] S12. Based on the initial vibration signals, resample to a uniform sampling frequency of 32 kHz to obtain first vibration signals.

[0078] S13. Based on the first vibration signal, band-pass filtering is used for denoising, and then data enhancement is performed by introducing Gaussian noise to obtain a second vibration signal;

[0079] S14. Based on the second vibration signal, pre-processing data of source domain data with real labels and target domain data without labels are constructed, wherein a single sample in the source domain represents , and the corresponding label is , and a single sample in the target domain represents . .

[0080] S2. Based on the pre-processing data, a feature extractor based on a convolutional neural network is constructed to perform preliminary feature extraction on the source domain and target domain data sets to obtain source domain feature data and target domain feature data.

[0081] Specifically, referring to Figure 3 , the method of step S2 includes:

[0082] S21. Constructing a feature extractor based on a convolutional neural network ; the network is composed of multiple one-dimensional convolutional blocks, each of which contains a convolutional layer, a batch normalization layer, and a max-pooling layer;

[0083] S22. Based on the feature extractor , input the source domain data to obtain the source domain features ;

[0084] S23. Based on the feature extractor , input the target domain data to obtain the target domain features .

[0085] S3. Based on the source domain feature data and the target domain feature data, pseudo-label generation and dynamic updating are performed.

[0086] Specifically, referring to Figure 4 , the method of step S3 includes:

[0087] S31. Based on the feature extractor and Softmax classifier trained based on the source domain data set, input the sample in the target domain data to obtain a pseudo-label.

[0088] Specifically, the sample in the target domain data set is input to the feature extractor trained based on the source domain data set and the subsequent Softmax classifier to predict the sample​​ , obtain pseudo label .

[0089] S32. Calculate the posterior probability of the sample in the target domain data belonging to the pseudo label based on the prediction confidence.

[0090] Specifically, input the sample in the target domain data set to the feature extractor trained based on the source domain data set and the subsequent Softmax classifier, predict the sample , and obtain the pseudo label .

[0091] S33. If the posterior probability is greater than or equal to the prediction confidence, the pseudo label is kept unchanged.

[0092] Specifically, if , the pseudo label is kept unchanged.

[0093] S34. If the posterior probability is less than the prediction confidence, randomly select other fault types as new pseudo labels.

[0094] Specifically, if , randomly select other fault types as new pseudo labels.

[0095] S4. Construct an unsupervised deep discriminative transfer learning network based on feature alignment of joint distribution adaptation loss.

[0096] Specifically, the method of step S4 comprises:

[0097] S41. Calculate the marginal distribution alignment loss between the source domain and the target domain using the maximum mean difference , which is specifically expressed as follows:

[0098] (1);

[0099] wherein, is the source domain feature; is the target domain feature;

[0100] S42. Calculate the conditional distribution alignment loss between the source domain and the target domain using correlation alignment , which is specifically expressed as follows:

[0101] (2);

[0102] wherein, is the total number of categories, and are the category ​​sample distribution of the source domain and the target domain under the following formula (1) and formula (2) respectively:

[0103] S43. The boundary distribution loss in step S41 and the conditional distribution loss in step S42 are combined to form a joint distribution adaptation loss for aligning the feature distribution of the source domain and the target domain, which is specifically represented as follows:

[0104] (3).

[0105] S44. A new metric is defined to evaluate the marginal distribution difference and the conditional distribution difference in the loss function, which is specifically represented as follows:

[0106] (4).

[0107] wherein, represents the maximum mean difference, represents the correlation alignment, which is specifically represented as shown in the following formula (5) and formula (6):

[0108] (5).

[0109] (6).

[0110] wherein, and are the number of samples in the source domain and the target domain respectively, is a Gaussian kernel mapping function, represents the norm in the reproducing kernel Hilbert space, Cov and Cov are the covariance matrices of the source domain and the target domain respectively, is the dimension of the feature vector, represents the Frobenius norm.

[0111] S45. Based on the in step S43 and the metric in step S44, a final loss function is formed, which is specifically represented as follows:

[0112]

[0113] (7).

[0114] S46. Based on the final loss function and the back propagation algorithm, the network parameters of the unsupervised deep discriminative transfer learning network are updated.

[0115] ​S5. Train and optimize the unsupervised deep discriminative transfer learning network, and perform bearing state diagnosis on the target domain feature data.

[0116] Specifically, the method of step S5 comprises:

[0117] S51. Calculate the classification loss of the source domain using the Softmax function , which is specifically represented as follows:

[0118] (8) ;

[0119] wherein, is the feature score of the source domain sample on the class , and is the score of the th sample on the th class;

[0120] S52. Calculate the classification loss of the target domain using the Softmax function , which is specifically represented as follows:

[0121] (9) ;

[0122] wherein, is the feature score of the target domain sample on the pseudo-label class , and is the score of the th target domain sample on the th class;

[0123] S53. Calculate the total classification loss based on formula (10) ;

[0124] (10) ;

[0125] S54. Obtain the final target loss of the unsupervised deep discriminative transfer learning network based on the total classification loss of step S53 and the joint distribution loss of step S45 , which is specifically represented as follows:

[0126] (11) ;

[0127] wherein,

[0128] S55. Train and optimize the unsupervised deep discriminative transfer learning network based on the final target loss .

[0129] S56. Input the test sample of the target domain into the trained unsupervised deep discriminative transfer learning network, and finally output the classification result.

[0130] Embodiment 1

[0131] In this embodiment, the bearing vibration signals collected from the experimental platform and from the actual working conditions of high-speed trains are resampled to a sampling frequency of 32 kHz, the original data is denoised by band-pass filtering, and data enhancement is performed by introducing Gaussian noise to construct a source domain dataset with real labels and a target domain dataset containing 16 samples without labels. A feature extractor based on a convolutional neural network is constructed to extract source domain and target domain feature representations, respectively, and the network is pre-trained on the source domain dataset. Based on the pre-trained classification network, pseudo-labels for target domain samples are generated by using target domain features, and the prediction confidence is set to 0.85 to achieve dynamic updating of the pseudo-labels. Considering the marginal distribution and conditional distribution, a joint adaptation loss is constructed to complete the feature alignment between the source domain and the target domain.

[0132] The high-dimensional features learned by the model are reduced to two-dimensional space for visualization using UMAP technology, as shown in Figure 5-6 By comparing the UMAP graphs at different stages (before and after transfer), it can be observed that before feature alignment, the source domain samples (circles) and the unlabeled target domain samples (triangles) usually show a clear separation. However, with the action of the JDA mechanism, the target domain samples gradually approach the samples of the corresponding classes in the source domain, and the samples of the same classes in the source domain and the target domain gradually merge. This indicates that by constructing a joint distribution adaptation loss, knowledge transfer between the source domain and the target domain is completed.

[0133] The joint adaptation loss and the classification loss of the source domain and the target domain are weighted and fused to form the final target loss function of the unsupervised deep discriminative transfer learning network, and the network is trained and optimized based on this loss function. Finally, the test samples of the target domain are input into the trained network, and the classification results are shown in Table 1. Among the 16 samples of the target domain, there are 3 rolling element failures, 6 inner ring failures, 4 outer ring failures, and 3 normal.

[0134] Table 1 shows the classification results

[0135] Target sample number Status Target sample number Status Target sample number Status Target sample number Status 1 Inner race failure 5 Rolling element failure 9 Outer race failure 13 Rolling element failure 2 Normal 6 Inner race failure 10 Inner race failure 14 Inner race failure 3 Outer race failure 7 Inner race failure 11 Outer race failure 15 Rolling element failure 4 Normal 8 Normal 12 Outer race failure 16 Inner race failure

[0136] Figure 7 The MDA loss curve is shown in Figure 6. The initial loss is high, then it rapidly decreases, and finally it tends to be stable. This indicates that the model can effectively reduce the distribution difference between the source domain and the target domain when performing marginal distribution alignment, and successfully realizes domain adaptation.

[0137] Figure 8 ​The CDA loss curve focuses on the alignment of conditional distribution. At the early stage of training, the CDA loss has a sharp fluctuation, which may indicate that the model needs to adapt to the conditional distribution of the source domain and the target domain at the early stage. As the training proceeds, the CDA loss gradually stabilizes and maintains at a lower level, indicating that the alignment of conditional distribution gradually stabilizes and the domain adaptation process achieves good results.

[0138] Figure 9 The target domain classification loss and the source domain classification loss are compared. Both the target domain loss and the source domain loss show a relatively rapid decline, especially at the early stage of training. The early decline is relatively steep, because the model can effectively learn some commonalities between the source domain and the target domain at the beginning. During the training process, the target class loss and the source class loss are almost stable at a lower level, indicating that the model has achieved a good balance in aligning the features of the source domain and the target domain.

[0139] Figure 10 The total loss curve is the sum of the classification loss and the domain adaptation loss, and the curve shows a similar trend to the MDA and CDA loss, but the convergence is relatively flat.

[0140] Figure 11 The accuracy of the model training and testing is shown. The accuracy of the model on the training set is close to 1, indicating that it has learned the source domain well. The test accuracy fluctuates greatly, especially at the early stage of training, which may be due to the difference between the distribution of the target domain data and the source domain data. After domain adaptation, the accuracy improves rapidly, but there is a continuous shock due to the lack of labels.

[0141] The above is only an embodiment of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation using the content of the specification and drawings, or direct or indirect application in other related technical fields, is also included in the patent protection scope of the present application.

Claims

1. A bearing fault diagnosis method based on unsupervised deep discriminative transfer learning, characterized in that, The method comprises the following steps: S1. Collecting initial bearing vibration signals under actual working conditions and preprocessing, constructing preprocessed data; S2. Based on the preprocessed data, a feature extractor based on a convolutional neural network is constructed to preliminarily extract features of source domain and target domain data sets, and source domain feature data and target domain feature data are obtained; S3. Based on the source domain feature data and the target domain feature data, pseudo-label generation and dynamic updating are performed; S4. Based on the feature alignment of the joint distribution adaptation loss, an unsupervised deep discriminative transfer learning network is constructed; S5. The unsupervised deep discriminative transfer learning network is trained and optimized, and the target domain feature data is subjected to bearing state diagnosis.

2. The method of claim 1, wherein, The method of step S1 comprises: S11. Collecting bearing vibration signals with real labels from an experimental platform and unlabeled bearing vibration signals from a high-speed train in a real scene to obtain the initial vibration signals; S12. Based on the initial vibration signals, resampling is performed to unify the sampling frequency to 32 kHz to obtain first vibration signals; S13. Based on the first vibration signals, band-pass filtering is used for denoising, and then data enhancement is performed by introducing Gaussian noise to obtain second vibration signals; S14. Based on the second vibration signals, the preprocessed data of the source domain data with real labels and the target domain data without labels are constructed.

3. The method of claim 2, wherein, The method of step S2 comprises: S21. Constructing a feature extractor based on a convolutional neural network; S22. Based on the feature extractor, input the source domain data to obtain source domain features; S23. Based on the feature extractor, input the target domain data to obtain target domain features.

4. The method of claim 3, wherein, The method of step S3 comprises: S31. Based on the feature extractor and Softmax classifier trained by the source domain data set, input the samples in the target domain data to obtain pseudo-labels; S32. Based on the prediction confidence, calculate the posterior probability that the samples in the target domain data belong to the pseudo-labels; S33. If the posterior probability is greater than or equal to the prediction confidence, the pseudo-labels remain unchanged; S34. If the posterior probability is less than the prediction confidence, randomly select other fault types as new pseudo-labels.

5. The method of claim 1, wherein, The method of step S4 comprises: In S41, a marginal distribution alignment loss between the source domain and the target domain is obtained based on formula (1). ; wherein the formula (1) is: （1）； wherein, is the source domain feature; is the target domain feature; In S42, the conditional distribution alignment loss between the source domain and the target domain is obtained based on formula (2). ; wherein the formula (2) is: （2）； wherein, is the total number of classes, and is the sample distribution of the source domain and the target domain under the class . S43. Based on formula (3), the marginal distribution alignment loss and the conditional distribution alignment loss are combined to form a joint distribution adaptation loss for aligning the source domain and the target domain feature distributions ; wherein the formula (3) is: （3）； S44. Based on formula (4), evaluate the marginal distribution difference and conditional distribution difference in the loss function; wherein the formula (4) is: （4）； wherein, denotes the maximum mean difference, denotes the correlation alignment; specifically, as shown in the following equations (5) - (6): （5）； （6）； where, and are the number of samples in the source and target domains, respectively, is a Gaussian kernel mapping function, denotes the norm in the reproducing kernel Hilbert space, Cov and Cov denote the covariance matrices of the source and target domains, respectively, is the dimension of the feature vector, denotes the Frobenius norm. S45. Forming the final loss based on equation (7) which is expressed as follows: （7）； S46. Based on the final loss and the back propagation algorithm, update the network parameters of the unsupervised deep discriminative transfer learning network.

6. The method of claim 1, wherein, The method of step S5 comprises: S51. Based on formula (8), calculate the classification loss of the source domain; wherein the formula (8) is: （8）； in, Source domain sample In category Feature scores on It is the first The sample at the th Scores on the class; S52. Based on formula (9), calculate the classification loss of the target domain; wherein the formula (9) is: （9）； wherein, is the target domain sample The feature score on the pseudo-label class , is the score of the th target domain sample on the th class; S53. Calculate the total classification loss based on formula (10) ; wherein the formula (10) is: （10）； S54. Based on formula (11), the final target loss of the unsupervised deep discriminative transfer learning network is obtained ; wherein the formula (11) is: （11）； wherein, S55. Based on the final target loss, train and optimize the unsupervised deep discriminative transfer learning network; S56. Input the test samples of the target domain into the trained unsupervised deep discriminative transfer learning network to finally output the classification results.

Images