Finite data-oriented semi-supervised rolling bearing cross-domain fault diagnosis method

Abstract

The invention belongs to the technical field of fault diagnosis, and discloses a limited data-oriented semi-supervised rolling bearing cross-domain fault diagnosis method. Generating strong enhancement data from source domain labeled data through a strong enhancement module; the strong enhancement data and the source domain tagged data are processed by a decision boundary adaptive module to obtain a decision boundary; according to the decision boundary, determining a pseudo label for the unlabeled data of the source domain; weak enhancement is carried out on source domain labeled data, initial confidence is calculated according to weak enhancement data, and a confidence threshold value is generated; performing label screening on the pseudo labels of the unlabeled data in the source domain according to the confidence coefficient threshold value to obtain purified data; and inputting the purified data into a semantic alignment module and a health state classification module to a network model for training to obtain a model for diagnosis. The method provided by the invention not only solves the problem of potential huge cross-domain distribution difference, but also can solve the problem that only a small amount of source domain data has labels, and can further promote the application of a cross-domain fault diagnosis model in actual production.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a semi-supervised rolling bearing cross-domain fault diagnosis method oriented to limited data. Background Art

[0002] Rolling bearings are key components of rotating machinery and are widely used in many industrial fields such as electricity and chemical industry. Their operating conditions are of great significance to the stability, safety and efficiency of production systems. Due to long-term high-speed, complex force and variable working conditions, rolling bearings are prone to failure due to fatigue, wear and tear. If the fault is not discovered in time, it will cause mechanical vibration, increase noise, reduce efficiency, increase energy consumption and cost in minor cases; in serious cases, it may cause equipment shutdown, interrupt production, cause huge economic losses, and even cause safety accidents. Accurate fault diagnosis can monitor its status in real time and detect potential hidden dangers in advance, so as to take maintenance measures in time, such as planning maintenance and replacing parts. Therefore, it is very necessary to perform fault diagnosis on rolling bearings. Rotating parts will not always operate under fixed working conditions, and there will be different degrees of equipment wear and degradation. The emergence of this problem hinders the widespread application of fault diagnosis methods based on deep learning in actual production environments. The direct result of variable working conditions is that there are differences in data distribution, which may lead to the problem that the model may have good results in the source domain but poor results in the target domain. Therefore, it is a challenge to solve the problem of reduced generalization performance caused by the distribution shift of test data and training data. In addition, deep learning-based fault diagnosis methods require a large amount of labeled data in the training phase to ensure the learning effect of the model. However, in actual production environments, the collection and labeling of fault data is cumbersome and requires a lot of manual participation. Therefore, when the number of required samples and source domains is large, data labeling requires a lot of manpower. Therefore, how to train the model when only part of the data is labeled while ensuring the reliability of the model is a key challenge. Recently, the emergence of domain generalization methods has been used to solve the problem of distribution shift between training data and test data. The core idea is to eliminate cross-domain distribution differences by matching the feature distribution of samples from different domains in high-dimensional space and extracting domain-invariant features of each type of sample. These methods have been proven to achieve relatively good results in domain generalization fault diagnosis research. However, most of these studies assume that the samples of each domain in the fault data set used to train the model are labeled. In view of the labor cost, this assumption is not necessarily true, that is, the source domain will contain some unlabeled data. The existence of unlabeled data will cause the deep learning-based model to be unable to be trained normally. Summary of the invention

[0003] In view of the above problems, the present invention specifically proposes a semi-supervised rolling bearing cross-domain fault diagnosis method for limited data. The main idea is to not only label the unlabeled data with pseudo labels, but also ensure the reliability of the pseudo labels and learn the domain-invariant features of each type of fault.

[0004] The technical solution of the present invention is as follows: A semi-supervised rolling bearing cross-domain fault diagnosis method for limited data, the source domain has labeled data The data is enhanced by the strong enhancement module to generate strong enhanced data The strong enhancement data and the source domain has labeled data The decision boundary is obtained by the decision boundary adaptation module; according to the decision boundary, the source domain unlabeled data Determine its pseudo label For source domains with labeled data Perform weak enhancement to obtain weak enhancement data Calculate the initial confidence based on the weakly enhanced data and generate a confidence threshold; perform label screening on the pseudo labels of the unlabeled data in the source domain based on the confidence threshold to obtain purified data Design semantic alignment module and health status classification module to build network model, and convert the purified labeled data into The data is input into the network model for training to obtain a network model that can be used for diagnosis. When used later, the data to be tested is directly input into the trained network model.

[0005] The strong enhancement module includes a latent variable extraction module, a generative adversarial network (GAN) and a variational autoencoder (VAE); the latent variable extraction module, the generator and the discriminator of the GAN are all composed of residual convolution modules, and the encoder and decoder of the VAE are composed of several deep convolutional neural network modules; the first convolution layer of the latent variable extraction module uses a 16×1 convolution kernel to extract short-term features, assist the model in automatically learning fault-related features, and remove features that are not related to the fault; the remaining convolution layers all use a 3×1 convolution kernel; each convolution layer is followed by a batch normalization operation, an activation operation, and a maximum pooling operation; the output of the latent variable extraction module uses a Flatten operation to expand the extracted features to a one-dimensional feature representation as a latent variable x l , and used as the input of the generator; the generator of GAN generates a sample with the same size as the source domain data after several residual convolution operations on the extracted latent variables, recorded as The discriminator of GAN combines the original sample x and As input, the two are distinguished and the output is whether the current input sample is a real sample; the discriminator input layer of the GAN uses a large convolution kernel, and the subsequent layers use a 3×1 convolution kernel; the variational autoencoder VAE generates data Perform latent variable extraction and extract the latent variables And the latent variable x extracted from the original data l The difference is calculated and minimized to reduce the information loss in the data generation process.

[0006] The decision boundary adaptation module adopts a K-Means-based clustering method to learn sample decision boundaries and optimizes the decision boundaries by maximizing the sample average entropy.

[0007] The source domain has no labeled data Determine its pseudo label, specifically, the sample Map to high-dimensional space, calculate the Euclidean distance between the sample and the cluster center according to the optimal decision boundary, and assign the category label corresponding to the cluster center with the smallest distance to the sample Get the pseudo label of the sample

[0008] The weak enhancement module involves three enhancement methods:

[0009] 1) Flip: Flip a random segment of the input data to its front-to-back order;

[0010] 2) Splicing: divide the input data into several segments and splice them in random order;

[0011] 3) Mask, set the data at some random positions in the input data to 0;

[0012] The source domain has labeled data Perform weak enhancement to obtain weak enhancement data Weakly enhanced data deviates from the original source domain labeled data In this case, the domain offset between multiple source domains is used as the weight of the prediction confidence and is also involved in the threshold calculation of pseudo-label screening. The prediction confidence of each sample after weak enhancement is calculated by the Gaussian kernel function, which is K(x,y)=exp{-||xy|| 2 / 2σ 2}; For data point d i , and the sum of similarities with other data points in the same cluster is calculated as The sum of similarities relative to the entire dataset is Data point d i The prediction confidence for The domain offset introduces an exponential function to keep it monotonically increasing. The calculation formula is: Where a is a hyperparameter, D i For sample d iRelative to the distance between other source domains and the target domain, the adjusted domain offset is used as a weight and combined with the prediction confidence to generate a new confidence value to obtain the confidence threshold for sample purification in The samples whose pseudo-label category probability is less than the confidence threshold are screened out to obtain purified data.

[0013] The input of the semantic alignment module is labeled data in the source domain. Strong Enhanced Data and purification data after pseudo tag screening After being mapped to a high-dimensional space by the convolutional network, the cross-domain invariant features are learned to obtain the high-dimensional feature z, which is then input into the health status classification module to obtain the predicted output result of the model.

[0014] In the original data, one field has a label, and the other fields have no labels.

[0015] Beneficial effects of the present invention: The present invention proposes a semi-supervised rolling bearing cross-domain fault diagnosis method for limited data, which is mainly implemented based on a strong enhancement module, a decision boundary adaptive module, a weak enhancement module, a semantic alignment module and a health status classification module. First, a strong enhancement module is used to realize strong enhancement of samples; then, the sample obtained by strong enhancement is used to perform adaptive adjustment of the decision boundary to obtain a decision boundary; the pseudo-label of the source domain unlabeled data is determined according to the decision boundary; the source domain labeled data is weakly enhanced to obtain weakly enhanced data; the initial confidence is calculated according to the weakly enhanced data, and a confidence threshold is generated; the pseudo-label of the source domain unlabeled data is labeled based on the confidence threshold to obtain purified data; the semantic alignment module and the health status classification module are designed to construct a network model, and the purified labeled data is input into the network model for training to obtain a model that can be used for diagnosis. In general, the method proposed by the present invention not only solves the problem of model training caused by huge distribution differences in cross-domain fault diagnosis, but also solves the problem of how to train the model under limited labeled data. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a structural diagram of the first part of the overall framework of the present invention;

[0017] Figure 2 This is a structural diagram of the second part of the overall framework of the present invention;

[0018] Figure 3 The following is a description of the flipping operation of the weak enhancement module in the present invention;

[0019] Figure 4 This is a description of the mask operation of the weak enhancement module in the present invention;

[0020] Figure 5This is a description of the splicing operation of the weak enhancement module in the present invention;

[0021] Figure 6 The comparative experimental results of the present invention on the CWRU dataset;

[0022] Figure 7 This is a purification confidence threshold result diagram in Task 2 of the present invention;

[0023] Figure 8 This is the ROC curve diagram of the pseudo-labels screened with different thresholds in Task 2 of the present invention. DETAILED DESCRIPTION

[0024] The current fault diagnosis training model requires a lot of labeled raw data. It is cumbersome to obtain and label the raw data, and the existing domain generalization methods have low diagnostic accuracy and poor generalization ability. This method is based on deep learning and domain generalization. It proposes a theory that only one labeled source domain data and several unlabeled source domain data are used for model training. A semi-supervised cross-domain fault diagnosis method for limited data is designed. A strong enhancement module, a decision boundary adaptive adjustment module, a weak enhancement module, a pseudo-label marking and screening, a semantic alignment module and a health status classification module are designed. The overall framework of the method is as follows Figure 1 and Figure 2 As shown in the figure. First, a strong enhancement module is designed, which combines the deep generative adversarial network GAN and the variational autoencoder VAE to achieve strong enhancement of samples and generate auxiliary samples. Secondly, after obtaining the strongly enhanced data, the clustering method is used to learn the decision boundary, and the output entropy is controlled to ensure that the decision boundary can cross the low-density area in the data. After determining the decision boundary, the Gaussian kernel function is used to transform the classification prediction confidence of each category in the weakly enhanced labeled data; and the prediction confidence is used to filter the pseudo-label data to obtain high-quality pseudo-label data. In the pseudo-label marking and screening, the unlabeled data is assigned to the category with the maximum probability to obtain the pseudo-label, and then the source domain labeled data is weakly enhanced, and the domain offset generated by the enhanced data is used as the weight of the prediction confidence and as the threshold for purifying the pseudo-label samples. Finally, the semantic alignment module and the health status classification module are designed, and the purified labeled data is input into the deep network for training to obtain the diagnosis model.

[0025] The following describes the implementation of the present invention in detail, and provides a detailed implementation method and operation process.

[0026] The environment and dependent software packages of this implementation are shown in the following table.

[0027] Table 1 Environment and dependent versions

[0028]

[0029] The overall structure is as follows Figure 1 and Figure 2 As shown in the figure, the framework mainly includes a strong enhancement module, a decision boundary adaptive adjustment module, a weak enhancement module, a pseudo-label marking and screening part, a semantic alignment module, and a health status classification module. The input of the framework is a one-dimensional original vibration signal from multiple source domains, of which only one source domain has a corresponding label, and the rest of the source domains have no labels. The output is the predicted health status of the bearing.

[0030] First, the present invention analyzes some problems existing in the existing fault diagnosis research, that is, the current mainstream fault diagnosis research methods based on domain generalization all assume that the fault data and all source domains used for model training have corresponding labels, which requires a lot of manpower to label the data in the actual production process and is difficult to be widely carried out. On the one hand, most of the existing methods for pseudo-labeling raw data lack sufficient reliable samples and a good sample screening mechanism, resulting in low quality of pseudo-labels for samples, which in turn leads to poor model training effect. On the other hand, during the training process, since the training data cannot contain all possibilities, the decision boundary of the model may not be able to pass through the low-density area of ​​the feature space during model training, which in turn leads to the model being unable to correctly identify and classify samples in the low-density area. The above problems will cause the existing methods to be unable to learn domain-invariant features well, thereby reducing the generalization performance of the model for unseen target domains. Under the premise that only one domain has a label and the other domains have no labels, the present invention improves the quality of sample pseudo-labels as much as possible, improves the diagnostic accuracy and generalization performance of the cross-domain fault diagnosis method, and realizes more accurate and reliable rolling bearing fault diagnosis.

[0031] The semi-supervised rolling bearing cross-domain fault diagnosis method for limited data of the present invention mainly includes the following parts:

[0032] Strong enhancement module: It consists of a latent variable extraction module, a generative adversarial network (GAN) and a variational auto-encoder (VAE). The latent variable extraction module, the generator and discriminator of GAN are all composed of residual convolution modules, and the encoder and decoder of VAE are composed of several deep convolutional neural network modules. First, the original data is passed through the latent variable extraction module to obtain the original data latent variable x l , and then perform a Flatten operation on it to obtain a one-dimensional feature vector, which is then input into the GAN generator for sample generation to obtain a sample Then the sample and the original sample x are input into the GAN discriminator to determine whether the current sample is a real sample; at the same time, the sample It will also be input into VAE and obtain the hidden variables of VAE and will With x l Comparison is made to reduce the information loss in the data generation process.

[0033] Decision boundary adaptation module: The decision boundary adaptation module uses a K-Means-based clustering method to learn the sample decision boundary, and optimizes the output entropy to ensure that the decision boundary learned by the model can pass through the low-density area of ​​the sample, so as to improve the model's feature learning ability in the entire data space. The source domain unlabeled data is obtained by using the Gaussian kernel function to transform the decision boundary. 's pseudo-labels.

[0034] Weak enhancement module: used to flip, mask and splice the original data to achieve weak enhancement operations on the data.

[0035] Pseudo label marking and screening part: Get the decision boundary that can cross the low-density area and convert the unlabeled data points u i Assign to the category with the highest probability, that is is the assigned pseudo label.

[0036] Then use flipping, masking and splicing methods to reconstruct the labeled data in the source domain. Weak enhancement is performed. In this case, the enhanced data will deviate from the original sample. In this case, the domain offset between multiple source domains is used as the weight of the prediction confidence. The prediction confidence of each sample after weak enhancement is calculated using the Gaussian kernel function, which is K(x,y)=exp{-||xy|| 2 / 2σ 2}, for data point d i , and the sum of similarities with other data points in the same cluster is calculated as The sum of similarities relative to the entire dataset is Data point d i The prediction confidence for The domain offset introduces an exponential function to keep it monotonically increasing. The calculation formula is: The adjusted domain offset is used as a weight and combined with the prediction confidence to generate a new confidence value. This gives the confidence threshold for sample purification: in In this way, the pseudo-label assignment and sample purification of the source domain data are completed.

[0037] The semantic alignment module is mainly used to learn domain-invariant features. Contrastive learning does not require manual design of complex feature extraction rules, and can automatically learn representative and discriminative feature representations from raw fault data. By comparing the relationship between different samples, contrastive learning can enable the fault diagnosis model to better understand the semantic connotation of fault data. It not only focuses on the surface features of the data, but also deeply explores the essence of the fault mode behind the data, and can more accurately classify and identify faults at the semantic level. Therefore, this module is mainly implemented using contrastive learning to promote the model to better learn domain-invariant features.

[0038] The health status classification module is used to output the prediction results of the input data. By introducing the gated recurrent unit, the attention paid to temporal information in the model feature extraction process is enhanced.

[0039] Furthermore, in order to fully capture and extract fault-related features from the original signal, the present invention designs and uses convolution kernels of different sizes in the strong enhancement module. Among them, for the latent variable extraction module, it mainly includes four residual convolution layers and four maximum pooling layers. The first convolution layer uses a larger convolution kernel to extract short-term features, assist the model to automatically learn fault-related features, and remove features that are not related to the fault; the remaining convolution layers all use a smaller 3×1 convolution kernel to help deepen the network structure while reducing the risk of model overfitting. Each layer of convolution is followed by batch normalization, activation, and maximum pooling operations. Among them, BN reduces internal covariate transfer by converting the input distribution into a Gaussian distribution, speeds up the calculation speed, and stabilizes the training process; the activation function uses a rectified linear unit (RectifiedLinear Unit, ReLU) to achieve nonlinear transformation, improve the nonlinear representation ability of the neural network, and increase the generalization ability of the network to a certain extent; the pooling layer is used for downsampling, and in order to retain as many fault peak information features as possible in the original data, maximum pooling is selected. For the output of the latent variable extraction module, the Flatten operation is used to expand the extracted features into a one-dimensional feature representation as a latent variable and as the input of the generator. The parameters of the strong enhancement module are detailed in Table 2.

[0040] Table 2 Parameters of latent variable extraction module

[0041]

[0042] For the generator of GAN, multiple layers of residual convolution modules with the same size are used here to realize data generation. The discriminator of GAN is used to distinguish whether the current sample is a real sample. It needs to be able to extract the fault information in the original data well. Therefore, the discriminator is similar to the latent variable extraction module and contains five residual convolution modules. The first residual convolution module uses a larger convolution kernel to extract short-term features. Subsequent convolution modules all use a 3×1 convolution kernel, and a fully connected layer is added after that to map the high-dimensional features to the output space. The output size of the fully connected layer is 2. The parameters of the GAN module are detailed in Table 3.

[0043] Table 3 GAN module parameters

[0044]

[0045]

[0046] For the VAE part, its main function is to extract latent variables from the generated data and compare them with the latent variables extracted from the original data to reduce the difference between the two and reduce information loss in the data generation process. Therefore, the VAE module also uses convolution kernels of different sizes for data processing. The first convolution layer uses a larger convolution kernel, and the rest use 3×1 convolution kernels. In the latent variable output part, the output is controlled to be the same size as the aforementioned latent variable. Given that VAE has good data fitting ability, this part is constructed using ordinary convolution modules. See Table 4 for VAE module parameters.

[0047] Table 4 VAE module parameters

[0048]

[0049]

[0050] The loss function of the strong enhancement module as a whole is shown in formula (1), where L KL is the KL divergence loss, is the latent variable reconstruction loss, L D represents the discriminator loss, L AC represents the autocorrelation loss, L rec represents the reconstruction loss.

[0051]

[0052] Formula (2) is the KL divergence loss, where σ j is the standard deviation of the input samples, μ j is the mean of the input sample, σ0 is the standard deviation of the prior distribution, and μ0 is the mean of the prior distribution.

[0053]

[0054] Formula (3) is the latent variable reconstruction loss, where z j is the latent variable extracted from the original data, are latent variables extracted from the generated data.

[0055]

[0056] Formula (4) is the discriminator loss, where G(·) is the generator generation process and D(·) is the discriminator discrimination process.

[0057]

[0058] Formula (5) is the autocorrelation loss, where [m] represents the signal sequence, represents the autocorrelation function.

[0059]

[0060] Formula (6) is the reconstruction loss of the generator.

[0061]

[0062] Furthermore, the decision boundary adaptive module is mainly implemented using the K-Means clustering algorithm, where the parameter k is the same as the number of fault types in the data set. The decision boundary can cross the low-density area by optimizing the entropy, where the formulas for entropy and average entropy are shown in equations (7) and (8). Entropy is calculated from the probability of the data points, where the expression of probability is shown in equation (9).

[0063]

[0064] where x i represents the sample, c represents the category, represents the enhanced labeled source domain data, center j represents the center of each cluster.

[0065] Furthermore, the weak enhancement module mainly performs flipping, masking and splicing operations on the original data to obtain weakly enhanced samples. Figure 3 , Figure 4 , Figure 5 illustrate.

[0066] Furthermore, the pseudo-label marking and screening part converts the unlabeled data points u i Assign to the category with the highest probability, that is That is the assigned pseudo label. Then the labeled data in the source domain is weakly enhanced by flipping, masking and splicing. At this time, the enhanced data will deviate from the original sample. At this time, the domain offset between multiple source domains is used as the weight of the prediction confidence and is also involved in the threshold calculation of pseudo label screening. The prediction confidence of each sample after weak enhancement is calculated by the Gaussian kernel function, which is K(x,y)=exp{-||xy|| 2 / 2σ 2}, for data point d i , and the sum of similarities with other data points in the same cluster is calculated as The sum of similarities relative to the entire dataset is Data point d i The prediction confidence for The domain offset introduces an exponential function to keep it monotonically increasing. The calculation formula is: The adjusted domain offset is used as a weight and combined with the prediction confidence to generate a new confidence value. This gives the confidence threshold for sample purification: in In this way, the pseudo-label assignment and sample purification of the source domain data are completed.

[0067] Furthermore, the semantic alignment module and the health status classification module are mainly used to realize the extraction of domain invariant features and the output of prediction results. The contrastive learning framework is used to build this part of the module. By comparing the relationship between different samples, the health status classification module can better understand the semantic information of the fault data. In addition, the gated recurrent unit is introduced to increase the model's attention to the temporal relationship of the data. The contrastive learning loss function is shown in formula (11), where sim(x,x + ) is the similarity metric, and τ is the temperature parameter. The loss function of the health status classification module is shown in formula (12), where y c is the label of category c, is the predicted probability.

[0068]

[0069] Prepare experimental dataset.

[0070] The purpose of the data set is to verify the diagnostic performance of the algorithm. As public bearing fault data sets, the JNU data set and the CWRU data set are widely used in the field of fault diagnosis. Therefore, the present invention selects the JNU data set and the CWRU data set as evaluation data sets. In order to ensure the smooth progress of the experiment, the present invention performs data preprocessing on the data set. Data preprocessing generates a data set for training and evaluating the model from the collected original signal, which mainly includes three steps: sample segmentation, normalization, and data set construction. Among them, sample segmentation intercepts samples from the collected original signal. Under different working conditions, the amplitude of the signal varies greatly. For deep neural networks, it is difficult to process samples with such a large amplitude variation range. Therefore, each sample is normalized with its own mean and variance to limit the amplitude of the sample, which helps to improve the generalization ability of the deep neural network. In order to simulate the actual problem of incomplete labels, the normalized samples are used to construct a training data set with only one domain having a label and multiple domains without labels.

[0071] Train the model.

[0072] Step 1: Use the deep learning framework PyTorch to prepare the network, initialize and configure the training model.

[0073] Step 2: Load the training data and preprocess the original data to ensure that one domain in the original data has a label and the other domains have no labels, and set the parameters of the algorithm.

[0074] Step 3: Train the network. The entire training process is described as follows: Input the original data, first perform strong enhancement, generate strong enhancement samples, and use the decision boundary adaptive adjustment module to learn the optimal decision boundary. Then perform weak enhancement on the original samples, use the domain shift of the weakly enhanced samples as the weight of the prediction confidence, and perform sample purification. Finally, after obtaining the labeled samples, the semantic alignment module and classification module are used to give the final output of the model.

[0075] Step 4: Adjust parameters to achieve the best model effect and obtain the optimal training model.

[0076] Step 4: Test the model. This step is mainly to obtain the predicted value on the test data, which is the basis for effect analysis. After preprocessing the test data, load the model weights into the network, and input the test data into the network to obtain the predicted result output for the final performance evaluation.

[0077] In order to evaluate the performance of the proposed method, experiments were conducted on two public bearing datasets (CWRU dataset and JNU dataset). Assuming there are four experimental tasks in total, the difference between each task is that the target domain data is different: for example, there are four domains in total, and the four tasks are to learn models from the first three domains and apply them to the fourth domain, Task 1: 123->4, Task 2: 124->3, Task 2: 134->2, Task 3: 234->1. First, a comparative test was conducted with several popular semi-supervised methods to verify the effectiveness and superiority of the proposed method. Some experimental results are shown in Figure 2. Figure 6-8 As shown in the figure, secondly, ablation experiment verification and fault feature visualization were carried out to verify the effectiveness and necessity of each module in the proposed method; finally, comparative experiments with different parameters were carried out to explore the performance of the proposed method under different parameter settings. The experimental results show that the proposed method effectively solves the problem of cross-domain fault diagnosis under limited data sets, has important application value, and promotes the application of cross-domain fault diagnosis models in actual industry.

Claims

1. A semi-supervised rolling bearing cross-domain fault diagnosis method for limited data, characterized in that: The source domain has labeled data The data is enhanced by the strong enhancement module to generate strong enhanced data The strong enhancement data and the source domain has labeled data The decision boundary is obtained by the decision boundary adaptation module; according to the decision boundary, the source domain unlabeled data Determine its pseudo label For source domains with labeled data Perform weak enhancement to obtain weak enhancement data Calculate the initial confidence based on the weakly enhanced data and generate a confidence threshold; perform label screening on the pseudo labels of the unlabeled data in the source domain based on the confidence threshold to obtain purified data Design semantic alignment module and health status classification module to build network model, and convert the purified labeled data into The data is input into the network model for training to obtain a network model that can be used for diagnosis.

2. The method for semi-supervised rolling bearing cross-domain fault diagnosis based on limited data according to claim 1 is characterized in that: The strong enhancement module includes a latent variable extraction module, a generative adversarial network (GAN) and a variational autoencoder (VAE); the latent variable extraction module, the generator and the discriminator of the GAN are all composed of residual convolution modules, and the encoder and decoder of the VAE are composed of several deep convolutional neural network modules; the first convolution layer of the latent variable extraction module uses a 16×1 convolution kernel to extract short-term features, assist the model in automatically learning fault-related features, and remove features that are not related to the fault; the remaining convolution layers all use a 3×1 convolution kernel; each convolution layer is followed by a batch normalization operation, an activation operation, and a maximum pooling operation; the output of the latent variable extraction module uses a Flatten operation to expand the extracted features to a one-dimensional feature representation as a latent variable x l , and used as the input of the generator; the generator of GAN generates a sample with the same size as the source domain data after several residual convolution operations on the extracted latent variables, recorded as The discriminator of GAN combines the original sample x and As input, the two are distinguished and the output is whether the current input sample is a real sample; the discriminator input layer of the GAN uses a large convolution kernel, and the subsequent layers use a 3×1 convolution kernel; the variational autoencoder VAE generates data Perform latent variable extraction and extract the latent variables And the latent variable x extracted from the original data l The difference is calculated and minimized to reduce the information loss in the data generation process.

3. The method for semi-supervised rolling bearing cross-domain fault diagnosis based on limited data according to claim 1 is characterized in that: The decision boundary adaptation module adopts a K-Means-based clustering method to learn sample decision boundaries and optimizes the decision boundaries by maximizing the sample average entropy.

4. The method for semi-supervised rolling bearing cross-domain fault diagnosis based on limited data according to claim 1 is characterized in that: The source domain has no labeled data Determine its pseudo label, specifically, the sample Map to high-dimensional space, calculate the Euclidean distance between the sample and the cluster center according to the optimal decision boundary, and assign the category label corresponding to the cluster center with the smallest distance to the sample Get the pseudo label of the sample 5. The method for semi-supervised rolling bearing cross-domain fault diagnosis based on limited data according to claim 1 is characterized in that: The weak enhancement module involves three enhancement methods: 1) Flip: Flip a random segment of the input data to its front-to-back order; 2) Splicing: divide the input data into several segments and splice them in random order; 3) Mask, set the data at some random positions in the input data to 0; The source domain has labeled data Perform weak enhancement to obtain weak enhancement data Weakly enhanced data deviates from the original source domain labeled data In this case, the domain offset between multiple source domains is used as the weight of the prediction confidence; the prediction confidence of each sample after weak enhancement is calculated by the Gaussian kernel function, and the Gaussian kernel function is K(x,y)=exp{-||xy|| 2 / 2σ 2 }; For data point d i , and the sum of similarities with other data points in the same cluster is calculated as The sum of similarities relative to the entire dataset is Data point d i The prediction confidence for The domain offset introduces an exponential function to keep it monotonically increasing. The calculation formula is: Where a is a hyperparameter, D i For sample d i Relative to the distance between other source domains and the target domain, the adjusted domain offset is used as a weight and combined with the prediction confidence to generate a new confidence value to obtain the confidence threshold for sample purification. in The samples whose pseudo-label category probability is less than the confidence threshold are screened out to obtain purified data.

6. The method for semi-supervised rolling bearing cross-domain fault diagnosis based on limited data according to claim 1 is characterized in that: The input of the semantic alignment module is the labeled data in the source domain. Strong Enhanced Data and purification data after pseudo tag screening After being mapped to a high-dimensional space by the convolutional network, the cross-domain invariant features are learned to obtain the high-dimensional feature z, which is then input into the health status classification module to obtain the predicted output result of the model.

7. The method for semi-supervised rolling bearing cross-domain fault diagnosis based on limited data according to claim 1 is characterized in that: In the original data, one field has a label, and the other fields have no labels.

Images