Rolling bearing unsupervised domain adaptive fault diagnosis method based on two-stage joint feature alignment maximum classifier difference network

Abstract

According to the rolling bearing unsupervised domain adaptive fault diagnosis method based on the two-stage joint feature alignment maximum classifier difference network, fault diagnosis of the rolling bearing under different working conditions is achieved. According to the method, a network structure comprising a feature extractor and double classifiers is constructed, the difference between output results of the double classifiers is measured by using a fusion classifier difference index, and the mobility and distinguishability of extracted features are improved by using a two-stage joint feature alignment strategy. In the first stage, the transferability of the features is improved by using the maximum mean value difference index, and in the second stage, on the basis of ensuring the transferability of the features, the potential category information of the target domain is fully utilized, and the distinguishability of the fault features is improved by using the local maximum mean value difference index. According to the method, the mobility and distinguishability of the extracted features are improved at the same time, and the certainty of a classifier output diagnosis result is ensured, so that the diagnosis precision of a rolling bearing variable working condition fault diagnosis task is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of cross-domain fault diagnosis of key components of rotating equipment such as rolling bearings, and more specifically, to an unsupervised domain adaptation fault diagnosis method for rolling bearings based on a two-stage joint feature alignment maximum classifier difference network. Background Art

[0002] Rolling bearings are an important part of the operation of modern rotating equipment. If rolling bearings are damaged but not discovered and repaired in time, it may trigger a series of chain reactions of failures, leading to equipment damage or serious safety accidents, causing irreparable losses. Therefore, it is of great significance to carry out fault diagnosis research on rolling bearings. However, due to the influence of factors such as speed and load, the operating conditions of rotating equipment are complex and changeable, which leads to a large difference in the data distribution between test samples and training samples, which in turn leads to a serious decline in the performance of data-driven diagnostic models. Therefore, complex and changeable working conditions are still a huge challenge for fault diagnosis in actual industrial scenarios.

[0003] Domain adaptation is an effective method for fault diagnosis under variable working conditions. The basic idea of ​​domain adaptation is to transfer the knowledge learned from the source domain task to the new related target domain task to achieve accurate fault diagnosis of the target domain samples. However, most domain adaptation methods need to rely on a large number of labeled target domain samples, which often faces challenges in practical applications.

[0004] Unsupervised domain adaptation effectively solves the problem of lack of labels in the target domain samples, that is, the problem of variable working condition diagnosis where the source task contains a large number of labeled samples and the target task contains only a small number of unlabeled samples. Unsupervised domain adaptation methods can be roughly divided into methods based on difference metrics and methods based on adversarial learning. Methods based on difference metrics achieve domain adaptation by comparing the differences between data with different distributions. An et al. designed an improved joint maximum mean difference to achieve synchronous alignment of marginal probability distribution and conditional probability distribution under unsupervised conditions. Zhou et al. used weight factors to dynamically adjust the marginal probability distribution and conditional probability distribution of data to complete the fault diagnosis task in the target domain. Adversarial learning-based methods use the idea of ​​adversarial learning to improve the feature extraction ability of the model, thereby obtaining domain-invariant feature representations. Yang et al. proposed a deep adversarial hybrid unsupervised domain adaptation network for fault diagnosis under different working conditions, which improved the diagnostic performance. Chen et al. proposed a general framework for fault diagnosis, namely adversarial domain-invariant generalization, which improved the diagnostic performance through feature normalization and customization strategies of adaptive weights. However, these unsupervised domain adaptation methods do not consider the relationship between target domain samples and classification boundaries when aligning distributions, resulting in a decrease in feature discriminability and some samples close to classification boundaries may be misclassified. Adversarial learning methods based on Maximum Classifier Discrepancy (MCD) effectively solve this problem. Saito et al. first proposed a new dual-classifier adversarial strategy to expand the decision boundary of the model through the game between feature extractors and classifiers in terms of output differences. Jiao et al. added an adaptive sample selection strategy and softmax cross entropy loss to the adversarial adaptation mechanism based on dual-classifier difference metrics to improve the accuracy of fault diagnosis. Xu et al. proposed an adversarial consistency learning method to enhance the discriminability of features by minimizing the triple loss function. The MCD-based method expands the decision boundary of the model and achieves feature alignment while maintaining feature discriminability.

[0005] However, the above schemes still have the following shortcomings: 1) In the process of obtaining domain-invariant features, most unsupervised domain adaptation fault diagnosis methods, including MCD-based methods, only align the global feature distribution of source and target domain samples, but ignore the difference in feature distribution between samples of the same category in the source and target domains, resulting in misclassification of target domain samples close to the classification boundary. 2) MCD-based methods use the difference between dual classifiers to learn transferable feature representations, which often ignores the certainty of the output results of the target domain sample classifier, resulting in reduced feature discriminability. Summary of the invention

[0006] The object of the present invention is to provide an unsupervised domain adaptation fault diagnosis method for rolling bearings based on a two-stage joint feature alignment MCD network, which is used to effectively perform cross-domain fault diagnosis on rolling bearings under variable working conditions. In order to reduce the difference between the source domain and the target domain distribution, a two-stage joint feature alignment strategy is designed to align the global feature distribution of samples collected under different working conditions and the feature distribution between samples of the same category of rolling bearings. In the first stage, the maximum mean difference index is used to reduce the global feature distribution difference between the source domain and the target domain samples, thereby improving the transferability of fault features. On the basis of ensuring the transferability of fault features in the first stage, the second stage makes full use of the potential category information of the target domain samples, obtains the target domain sample output probability vector as the target domain sample pseudo label through the network model trained in the first stage, and uses the local maximum mean difference index to reduce the feature distribution difference between the source domain and the target domain samples of the same category, so as to improve the discriminability of fault features. In order to ensure that the dual classifier outputs a certain diagnosis result, a fusion classifier difference index is designed to measure the difference between the output results of the dual classifiers. The index includes a joint difference index and an individual difference index, which are respectively used to ensure the consistency and certainty of the final output results of the dual classifiers. The present invention is based on a two-stage joint feature alignment MCD network, which makes full use of the potential fault category information of the target domain samples, improves both the transferability and the discriminability of the extracted fault features, and ensures the certainty of the diagnostic results output by the dual classifiers, thereby significantly improving the diagnostic accuracy of the rolling bearing variable operating condition fault diagnosis task.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] An unsupervised domain adaptation fault diagnosis method for rolling bearings based on a two-stage joint feature alignment MCD network is characterized by comprising the following steps:

[0009] A1 obtains historical operating data of rolling bearings under different working conditions (different working conditions represent different load conditions), uses the sliding window method to divide the original vibration signal into data of the same length and normalizes it, constructs a labeled source domain sample set and an unlabeled target domain sample set (the source domain sample set and the target domain sample set belong to different working conditions respectively), and then divides the unlabeled target domain sample set into a training set and a test set.

[0010] A2 constructs a network structure including a feature extractor and a dual classifier and initializes the network parameters. It uses the fusion classifier difference index to measure the difference between the output results of the dual classifiers and is used to construct the network loss function.

[0011] A3 adopts a two-stage joint feature alignment strategy to reduce the difference between the source domain and the target domain distribution. Each stage contains three steps, each step corresponds to a total loss function, and the overall network, classifier, and feature extractor are trained in turn. In the first stage, the labeled source domain samples and the unlabeled target domain training samples are input into the network model, and the maximum mean difference is introduced to reduce the global feature distribution difference between the source domain and the target domain samples. The three steps of this stage will be trained repeatedly until the preset maximum number of iterations is reached. In the second stage, the labeled source domain samples and the unlabeled target domain training samples are input into the network model. On the basis of the first stage, the potential category information of the target domain samples is fully utilized. The target domain sample output probability vector is obtained through the network model trained in the first stage as the target domain sample pseudo label. The local maximum mean difference is introduced to reduce the feature distribution difference between the same category samples in the source domain and the target domain. The three steps are trained repeatedly until the maximum number of iterations in the second stage is reached, and the training is stopped to generate an MCD network based on two-stage joint feature alignment with good generalization performance.

[0012] A4 uses the trained MCD network based on two-stage joint feature alignment to identify the bearing fault types of target domain test samples under variable working conditions and complete the final fault diagnosis task.

[0013] A rolling bearing unsupervised domain adaptation fault diagnosis method based on a two-stage joint feature alignment MCD network, characterized in that the specific steps of A1 are as follows:

[0014] A11) Obtain the actual operation history data of the rolling bearing under different working conditions (different working conditions represent different load conditions), use the sliding window method to divide the original vibration signal into data of the same length and perform normalization processing, the length of each data is 1024, and the normalization processing formula is as follows:

[0015]

[0016] Among them, X is the original data, X norm is the normalized data, X min and X max are the minimum and maximum values ​​in the data set, respectively.

[0017] A12) Construct a labeled source domain sample set n s represents the number of source domain samples, represents the i-th source domain sample, for Fault type labels; Construct unlabeled target domain sample set represents the jth target domain sample, n TRepresents the number of samples in the target domain. The source domain sample set and the target domain sample set belong to different working conditions. 80% of the target domain sample set is randomly divided into the training set, and the remaining 20% ​​is divided into the test set.

[0018] A rolling bearing unsupervised domain adaptation fault diagnosis method based on a two-stage joint feature alignment MCD network, characterized in that the specific steps of A2 are as follows:

[0019] A21) Construct a network structure including a feature extractor and a dual classifier and initialize the network parameters: the feature extractor consists of five convolution blocks, each of which consists of a convolution layer, a batch normalization processing layer, a ReLU activation function layer and a maximum pooling layer. The convolution kernel size of the convolution layer of the first convolution block is 1×64, and the convolution kernel size of the convolution layers of the remaining four convolution blocks is 1×3. The classifier consists of two fully connected layers. The first fully connected layer uses ReLU as the activation function, and the second fully connected layer uses Softmax classifier as the activation function. The structures of the two classifiers are the same.

[0020] A22) The fusion classifier difference index is used to measure the difference between the output results of the two classifiers. The fusion classifier difference index L f By formula (2), we can obtain:

[0021] L f =μL u +ωL e(2)

[0022] Among them, μ and ω are equilibrium parameters, μ is 1.6, and ω is 1. u is the joint difference index, obtained by formula (3); L e is the individual difference index, obtained by formula (4).

[0023]

[0024] Where G represents the feature extractor, C 1 and C 2 Represent two classifiers, δ c Represents the c-th element of the softmax output.

[0025]

[0026] An unsupervised domain adaptation fault diagnosis method for rolling bearings based on a two-stage joint feature alignment MCD network is characterized in that: in step A3, a two-stage joint feature alignment strategy is adopted to reduce the difference between the source domain and the target domain distribution. Each stage includes three steps, each step corresponds to a total loss function, and the overall network, classifier and feature extractor are trained in sequence. The specific steps are as follows:

[0027] A31) Input source domain samples, source domain sample labels and target domain training samples, and calculate the first step total loss function. This step total loss function is calculated by calculating the cross entropy loss L S and the maximum mean difference index L mmd The sum is obtained as shown in formula (5) and formula (6), and the overall network is trained.

[0028]

[0029] Among them, p(y|x s ) represents the softmax output probability of the source domain sample after the classifier, k∈{1,...,K}, K represents the number of fault categories and K=10, I [k =y s ] represents the indicator vector of the source domain sample label, and E represents the mathematical expectation.

[0030]

[0031] Wherein, H represents the reproducing kernel Hilbert space, and Φ represents the function mapped to the reproducing kernel Hilbert space.

[0032] A32) Input source domain samples, source domain sample labels and target domain training samples, and calculate the second step total loss function. This step total loss function is calculated by calculating the cross entropy loss L s (x s ,y s ) and the fusion classifier difference index L f The sum is obtained and two classifiers are trained.

[0033] A33) Input the target domain training sample and calculate the total loss function of the third step. The total loss function of this step is calculated by calculating the difference index of the fusion classifier - L f Obtain and train a feature extractor.

[0034] A34) Repeat steps A31) to A33) until the maximum number of iterations of the first stage is reached, and the number of iterations of the first stage is 15,000.

[0035] A35) Input source domain samples, source domain sample labels and target domain training samples, and calculate the first step total loss function. This step total loss function is calculated by calculating the cross entropy loss L s (x s ,y s ) and the local maximum mean difference index L lmmd (x s ,x T ) is obtained by summing up, as shown in formula (5) and formula (7), and training the feature overall network.

[0036]

[0037] in, and represent source domain samples and target domain samples respectively, and They represent the weights of the k-th fault category in the source domain and the target domain respectively. The calculation is shown in formula (8):

[0038]

[0039] in, represents the probability that the sample belongs to fault category k, the source domain sample The kth item of the corresponding one-hot vector label is taken as The target domain samples are collected from unlabeled conditions, so the kth item of the probability vector output by the network model is used as the target domain sample. of

[0040] A36) Input source domain samples, source domain sample labels and target domain training samples, and calculate the second step total loss function. This step total loss function is calculated by calculating the cross entropy loss L s (x s ,y s ) and the fusion classifier difference index L f The sum is obtained and two classifiers are trained.

[0041] A37) Input the target domain training sample and calculate the total loss function of the third step. The total loss function of this step is calculated by calculating the difference index of the fusion classifier - L f Obtain and train a feature extractor.

[0042] A38) Repeat steps A35) to A37 until the maximum number of iterations of the second stage is reached, which is 5000. Stop training and generate an MCD network based on two-stage joint feature alignment with good generalization performance.

[0043] A rolling bearing unsupervised domain adaptation fault diagnosis method based on a two-stage joint feature alignment MCD network, characterized in that the specific steps of A4 are as follows:

[0044] A41) Input the target domain test sample into the trained MCD network based on two-stage joint feature alignment to obtain the output results of the two classifiers, and then average the two results to obtain the final target domain test sample diagnosis result.

[0045] Beneficial Effects

[0046] The present invention discloses an unsupervised domain adaptation fault diagnosis method for rolling bearings based on a two-stage joint feature alignment MCD network. A two-stage joint feature alignment strategy is designed to align the global feature distribution of samples collected under different working conditions of rolling bearings and the feature distribution between samples of the same category, while improving the transferability and discriminability of the extracted fault features. In the first stage, the maximum mean difference index is used to reduce the global feature distribution difference between source domain and target domain samples, thereby improving the transferability of the fault features. On the basis of ensuring the transferability of the fault features in the first stage, the second stage makes full use of the potential category information of the target domain samples, uses the network model trained in the first stage to obtain the target domain sample output probability vector as the target domain sample pseudo label, and uses the local maximum mean difference index to reduce the feature distribution difference between the source domain and target domain samples of the same category, so as to improve the discriminability of the fault features. A fusion classifier difference index is designed to measure the difference between the classifier output results. The index includes a joint difference index and an individual difference index, which are respectively used to ensure the consistency and certainty of the final output results of the dual classifiers. The present invention makes full use of the latent category information of the target domain samples, which not only improves the transferability of the extracted fault features, but also improves the discriminability of the extracted fault features, while ensuring the certainty of the diagnostic results output by the dual classifiers. It has the advantages of strong feature extraction capability, strong generalization capability, and high diagnostic accuracy, and is of great significance for the unsupervised domain adaptive fault diagnosis of rolling bearings under variable operating conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Shown is a flow chart of the method of the present invention;

[0048] Figure 2 The figure shows the rolling bearing unsupervised domain adaptation fault diagnosis method based on the two-stage joint feature alignment MCD network of the present invention;

[0049] Figure 3 The following is a two-stage joint feature alignment strategy. (a) The first stage; (b) The second stage;

[0050] Figure 4 Shown is a schematic diagram of the bearing test rig at Case Western Reserve University (CWRU);

[0051] Figure 5 Precision evaluation indicators of different models in the CWRU dataset transfer learning task 3-0;

[0052] Figure 6 Recall evaluation indicators of different models in the CWRU dataset transfer learning task 3-0;

[0053] Figure 7 F1 score evaluation indicator of different models in the CWRU dataset transfer learning task 3-0. DETAILED DESCRIPTION

[0054] In view of the shortcomings of the prior art, the present invention provides an unsupervised domain adaptation fault diagnosis method for rolling bearings based on a two-stage joint feature alignment MCD network. This method can effectively align the feature distribution of samples collected under different working conditions of rolling bearings, thereby improving the accuracy of rolling bearing fault diagnosis under variable working conditions.

[0055] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. 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.

[0056] Reference Figure 2 The present invention provides an unsupervised domain adaptation fault diagnosis method for rolling bearings based on a two-stage joint feature alignment MCD network, which overcomes the problem of difficulty in achieving accurate rolling bearing fault diagnosis under variable working conditions. The proposed two-stage joint feature alignment strategy can better reduce the difference between the source domain and the target domain distribution, such as Figure 3 As shown. At the same time, a classifier difference index, called the fusion classifier difference index, is designed to measure the difference between classifiers and is used to construct a loss function based on the two-stage joint feature alignment MCD network. Ultimately, the extraction of highly transferable and highly discriminative features of source and target domain samples is achieved to maintain the accuracy of rolling bearing variable condition fault diagnosis.

[0057] The Case Western Reserve University (CWRU) public bearing dataset is widely used to verify the performance of fault diagnosis. Figure 4 The CWRU bearing test bench is shown, which consists of two motors, torque sensors, power test meters and other control devices. There are four health conditions in this dataset: normal state (N), roller element fault (RF), inner raceway fault (IF) and outer raceway fault (OF). At the same time, there are three different damage levels of 0.007 inches, 0.014 inches and 0.021 inches on the bearings of each fault type. In this experiment, a drive end bearing dataset with a sampling frequency of 12kHz was selected. The experimental platform was operated under four different load conditions of 0HP, 1HP, 2HP and 3HP, and ten fault types were selected. The four different working conditions are shown in Table 1. Each experiment selects two from the four conditions as the source domain and the target domain. For example, Task 1-2 means that the source domain samples and the target domain samples are collected under working conditions 1 and working conditions 2, respectively. To reduce the influence of experimental randomness, each task is repeated 10 times to obtain the average diagnosis accuracy.

[0058] In terms of hyperparameter setting, the SGD optimizer was used for model optimization. The batch size of the training model was set to 1024, the initial learning rate was set to 0.01, and the dynamic learning rate update technology was used. The number of iterations in the first stage of the model was 15,000, and the number of iterations in the second stage was 5,000. In terms of model structure design, a network structure containing a feature extractor and a dual classifier was constructed. The feature extractor consists of five convolution blocks, each of which consists of a convolution layer, a batch normalization processing layer, a ReLU activation function layer, and a maximum pooling layer. The classifier consists of two fully connected layers. The first fully connected layer uses ReLU as the activation function, and the second fully connected layer uses the Softmax classifier as the activation function. The structures of the two classifiers are the same.

[0059] Based on the above description and according to the content of the invention, the specific implementation process is as follows:

[0060] 1) Obtain the actual operation history data of rolling bearings under different working conditions (different working conditions represent different load conditions), use the sliding window method to split the original vibration signal into data of the same length and perform normalization processing to construct a labeled source domain sample set n s represents the number of source domain samples, represents the i-th source data sample, for Fault labels; Construct unlabeled target domain sample set represents the jth target domain sample, n T Represents the number of samples in the target domain. The source domain sample set and the target domain sample set belong to different working conditions. 80% of the target domain sample set is randomly divided into the training set, and the remaining 20% ​​is divided into the test set.

[0061] 2) Input source domain samples, source domain sample labels and target domain training samples, and calculate the total loss function L of the network at this time using formula (5) and formula (6): s (x s ,y s )+L mmd (x s ,x T ), train the overall network.

[0062] 3) Input source domain samples, source domain sample labels and target domain training samples, and calculate the total loss function L of the network at this time by using formula (5) and formula (2): s (x s ,y s )+L f , train two classifiers.

[0063] 4) Input the target domain training sample and calculate the total loss function of the network at this time using formula (2): f , train the feature extractor.

[0064] 5) Repeat steps 2) to 4) until the maximum number of iterations of the first stage is reached. The number of iterations of the first stage is 15,000.

[0065] 6) Input source domain samples, source domain sample labels and target domain training samples, and calculate the total loss function L of the network at this time using formula (5) and formula (7): s (x s ,y s )+L lmmd (x s ,x T ), train the overall network.

[0066] 7) Input source domain samples, source domain sample labels and target domain training samples, and calculate the total loss function L of the network at this time using formula (5) and formula (2): s (x s ,y s )+L f , train two classifiers.

[0067] 8) Input the target domain training sample and calculate the total loss function of the network at this time using formula (2): f , train the feature extractor.

[0068] 9) Repeat steps 6) to 8) until the maximum number of iterations of the second stage is reached, which is 5000. Stop training and generate an MCD network based on two-stage joint feature alignment with good generalization performance.

[0069] 10) Input the target domain test sample into the MCD network based on two-stage joint feature alignment to obtain the output results of the two classifiers, and average the two results to obtain the final target domain test sample diagnosis result.

[0070] In order to fully demonstrate the effectiveness of the model, five domain adaptation methods were selected for comparison. For the difference metric-based methods, the deep domain adaptation network (DAN) based on the maximum mean difference metric and the CORAL method based on correlation matching were selected. For the domain adversarial method, the deep adversarial network (DANN) was selected. The MCD method was selected for the dual classifier adversarial method. In addition, the domain adaptation method CMUTLN based on local alignment was also selected.

[0071] In order to comprehensively evaluate the performance of the proposed fault diagnosis method, accuracy, precision, recall and F1-score are used as indicators to evaluate the fault diagnosis performance. These four indicators are defined as:

[0072]

[0073]

[0074]

[0075]

[0076] Among them, TP, FP, FN and TN represent the positive samples predicted by the model as positive, the negative samples predicted by the model as positive, the positive samples predicted by the model as negative, and the negative samples predicted by the model as negative, respectively.

[0077] A total of 12 groups of transfer learning tasks were set up in this experiment, and the diagnosis results are shown in Table 2. The present invention has achieved relatively high diagnostic accuracy in almost all transfer learning tasks. Given 12 groups of transfer learning tasks, the average diagnostic accuracy is 98.78%, and fault diagnosis under variable working conditions has a high diagnostic accuracy. This is mainly due to the reduction of the feature distribution differences between samples of the same category in the source domain and the target domain, making the extracted features more discriminative. However, for methods such as CORAL, DAN, DANN and MCD that only perform global feature distribution alignment, the diagnostic results are not ideal because the feature distribution differences between samples of the same category in the source domain and the target domain are not considered. In addition, the diagnostic result of the worst performing method on the most difficult task is 88.54%, while that of the present invention is 95.76%.

[0078] In order to further demonstrate the superiority of the present invention, the performance of the proposed model and the comparison model was evaluated using standard indicators such as precision, recall and F1 score. The precision, recall and F1 evaluation indicators of different models in task 3-0 on the CWRU dataset are shown in the figure below. Figure 5 , Figure 6 and Figure 7 shown.

[0079] Among different fault types, the F1 values ​​of CORAL and DAN are generally low. This is because the unsupervised domain adaptation fault diagnosis method based on difference metric only focuses on the global feature distribution difference between the source domain and the target domain samples, resulting in poor discriminability of the extracted domain invariant features. For adversarial unsupervised domain adaptation fault diagnosis methods such as DANN and MCD, although DANN adds a domain classifier and MCD uses a dual classifier through a special network structure to better extract domain invariant features, they also only focus on the global feature distribution difference between the source domain and the target domain samples, ignoring the potential fault category information of the target domain samples, resulting in poor discriminability of the extracted fault features, and then leading to incorrect fault diagnosis results. CMUTLN uses a local alignment method for domain adaptation, but because the classifier difference index cannot ensure the certainty of the final output result of the classifier, some target domain samples close to the classification boundary will be misclassified. In contrast, the method proposed in the present invention not only reduces the feature distribution difference between the same category samples in the source domain and the target domain, but also ensures the certainty of the final output result, showing good performance under the F1 index.

[0080] Finally, it can be concluded that the present invention makes full use of the latent category information of the target domain samples, which not only improves the transferability and discriminability of the extracted fault features, but also ensures the certainty of the diagnostic results output by the classifier, thereby significantly improving the diagnostic performance indicators of the rolling bearing variable operating condition fault diagnosis task, and has a great beneficial effect on rolling bearing fault diagnosis.

[0081] Table 1 Four different working conditions on the CWRU bearing dataset

[0082]

[0083] Table 2 Diagnostic accuracy of each model on the CWRU bearing dataset

[0084]

Claims

1. An unsupervised domain adaptation fault diagnosis method for rolling bearings based on a two-stage joint feature alignment maximum classifier difference network, characterized in that: The following steps are involved: A1 obtains historical operating data of rolling bearings under different working conditions. Different working conditions represent different load conditions. The original vibration signal is divided into data of the same length using the sliding window method and normalized. A labeled source domain sample set and an unlabeled target domain sample set are constructed. The source domain sample set and the target domain sample set belong to different working conditions. Then the unlabeled target domain sample set is divided into a training set and a test set. A2 builds a network structure including a feature extractor and a dual classifier and initializes the network parameters. It uses the fusion classifier difference index to measure the difference between the dual classifier output results and is used to construct the network loss function. A3 adopts a two-stage joint feature alignment strategy to reduce the difference between the source domain and the target domain distribution. Each stage contains three steps, each of which corresponds to a total loss function. The overall network, classifier, and feature extractor are trained in sequence. In the first stage, labeled source domain samples and unlabeled target domain training samples are input into the network model, and the maximum mean difference is introduced to reduce the global feature distribution difference between the source domain and the target domain samples. The three steps of this stage will be trained repeatedly until the preset maximum number of iterations is reached; in the second stage, the labeled source domain samples and the unlabeled target domain training samples are input into the network model. On the basis of the first stage, the potential category information of the target domain samples is fully utilized. The target domain sample output probability vector is obtained through the network model trained in the first stage as the target domain sample pseudo-label. The local maximum mean difference is introduced to reduce the feature distribution difference between the same category samples in the source domain and the target domain. The three steps are trained repeatedly until the maximum number of iterations of the second stage is reached, and the training is stopped to generate a maximum classifier difference network based on the two-stage joint feature alignment; A4 uses a trained maximum classifier difference network based on two-stage joint feature alignment to identify the bearing fault types of target domain test samples under variable working conditions and complete the final fault diagnosis task.

2. The rolling bearing unsupervised domain adaptation fault diagnosis method based on two-stage joint feature alignment maximum classifier difference network according to claim 1 is characterized by: The specific steps of A1 are as follows: A11) Obtain the actual operation history data of the rolling bearing under different working conditions (different working conditions represent different load conditions), use the sliding window method to divide the original vibration signal into data of the same length and perform normalization processing, the length of each data is 1024, and the normalization processing formula is as follows: Among them, X is the original data, X norm is the normalized data, X min and X max are the minimum and maximum values ​​in the data set, respectively; A12) Construct a labeled source domain sample set n s represents the number of source domain samples, represents the i-th source domain sample, for Fault type labels; Construct unlabeled target domain sample set represents the jth target domain sample, n T represents the number of samples in the target domain; the source domain sample set and the target domain sample set belong to different working conditions respectively; the target domain sample set is randomly divided into a training set and a test set.

3. The rolling bearing unsupervised domain adaptation fault diagnosis method based on two-stage joint feature alignment maximum classifier difference network according to claim 1 is characterized by: The specific steps of A2 are as follows: A21) Construct a network structure including a feature extractor and a dual classifier and initialize the network parameters: the feature extractor consists of five convolution blocks, each of which consists of a convolution layer, a batch normalization layer, a ReLU activation function layer and a maximum pooling layer. The convolution kernel size of the convolution layer of the first convolution block is 1×64, and the convolution kernel size of the convolution layers of the remaining four convolution blocks is 1×3. The classifier consists of two fully connected layers. The first fully connected layer uses ReLU as the activation function, and the second fully connected layer uses Softmax classifier as the activation function. The structures of the two classifiers are the same; A22) The fusion classifier difference index is used to measure the difference between the output results of the two classifiers. The fusion classifier difference index L f By formula (2), we can obtain: L f =μL u +ωL e (2) Among them, μ and ω are equilibrium parameters, μ is 1.6, ω is 1; L u is the joint difference index, obtained by formula (3); L e is the individual difference index, obtained by formula (4); Where G represents the feature extractor, C1 and C2 represent two classifiers, c Represents the cth element of the softmax output; 4. The rolling bearing unsupervised domain adaptation fault diagnosis method based on two-stage joint feature alignment maximum classifier difference network according to claim 1 is characterized by: In step A3, a two-stage joint feature alignment strategy is adopted to reduce the difference between the source domain and the target domain distribution. Each stage contains three steps, each step corresponds to a total loss function, and the overall network, classifier and feature extractor are trained in turn. The specific steps are as follows: A31) Input source domain samples, source domain sample labels and target domain training samples, and calculate the first step total loss function. This step total loss function is calculated by calculating the cross entropy loss L S and the maximum mean difference index L mmd The sum is obtained and the overall network is trained; A32) Input source domain samples, source domain sample labels and target domain training samples, and calculate the second step total loss function. This step total loss function is calculated by calculating the cross entropy loss L s (x s ,y s ) and the fusion classifier difference index L f The sum is obtained and two classifiers are trained; A33) Input the target domain training sample and calculate the total loss function of the third step. The total loss function of this step is calculated by calculating the difference index of the fusion classifier - L f Obtain and train a feature extractor; A34) Repeat steps A31) to A33) until the maximum number of iterations of the first stage is reached, and the number of iterations of the first stage is 15000; A35) Input source domain samples, source domain sample labels and target domain training samples, and calculate the first step total loss function. This step total loss function is calculated by calculating the cross entropy loss L s (x s ,y s ) and the local maximum mean difference index L lmmd (x s ,x T ) and train the overall network; A36) Input source domain samples, source domain sample labels and target domain training samples, and calculate the second step total loss function. This step total loss function is calculated by calculating the cross entropy loss L s (x s ,y s ) and the fusion classifier difference index L f The sum is obtained and two classifiers are trained; A37) Input the target domain training sample and calculate the total loss function of the third step. The total loss function of this step is calculated by calculating the difference index of the fusion classifier - L f Obtain and train a feature extractor; A38) Repeat steps A35) to A37 until the maximum number of iterations of the second stage is reached, which is 5000. Stop training and generate a maximum classifier difference network based on two-stage joint feature alignment.

5. The rolling bearing unsupervised domain adaptation fault diagnosis method based on two-stage joint feature alignment maximum classifier difference network according to claim 4 is characterized by: In step A31, the total loss function is calculated by L S Cross Entropy Loss and L mmd The sum of the maximum mean difference index is obtained, as shown in formula (5) and formula (6): Among them, p(y|x s ) represents the softmax output probability of the source domain sample after the classifier, k∈{1,...,K}, K represents the number of fault categories and K=10, I [k =y s ] represents the indicator vector of the source domain sample label, and E represents the mathematical expectation; Wherein, H represents the reproducing kernel Hilbert space, and Φ represents the function mapped to the reproducing kernel Hilbert space.

6. The rolling bearing unsupervised domain adaptation fault diagnosis method based on two-stage joint feature alignment maximum classifier difference network according to claim 4 is characterized by: In step A35, the local maximum mean difference index L in the overall loss function lmmd By formula (7), we can obtain: in, and represent source domain samples and target domain samples respectively, and Represent the weights of the kth fault category in the source domain and the target domain respectively; The calculation formula is as follows: in, represents the probability that the sample belongs to fault category k, the source domain sample The kth item of the corresponding one-hot vector label is taken as The target domain samples are collected from unlabeled conditions, so the kth item of the probability vector output by the network model is used as the target domain sample. of

Images