A method for fault diagnosis of rotating machinery based on conditional decoupling representation enhancement network

Abstract

This invention discloses a rotating machinery fault diagnosis method based on a conditional decoupled representation enhancement network. Addressing scenarios involving class imbalance and cross-domain fault diagnosis in rotating machinery, the method utilizes a deep convolutional feature extractor to capture fault features from both the known source domain and the unknown target domain in the acquired vibration signals. A decoupled representation enhancement module generates high-quality class semantic information to enhance minority class features in the high-dimensional feature space, thereby balancing the data distribution between classes and mitigating the impact of differences in the number of samples from different classes. Subsequently, a conditional discriminant contrastive loss is used to increase attention to the minority class and strengthen the compactness of intra-class features across domains. Finally, the desired imbalanced domain generalized mechanical fault diagnosis model is formed. This method not only overcomes the influence of class imbalance but also improves the generalization ability and robustness to unknown target domains, effectively enhancing the accuracy of rotating machinery fault diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of unbalanced cross-condition fault diagnosis, specifically a rotating machinery fault diagnosis method based on conditional decoupling representation enhancement network. Background Technology

[0002] With the continuous improvement of productivity, precision rotating machinery has been developed and integrated on a large scale. However, harsh production environments and high-intensity operations can easily lead to mechanical failures. Therefore, to ensure the stability and safety of the machinery itself, the monitoring and diagnosis of rotating machinery has become increasingly important. As industry moves towards higher levels of digitalization and automation, intelligent fault diagnosis models are receiving more and more attention. In recent years, data-driven deep learning methods have become a common approach for fault diagnosis. However, traditional deep learning techniques assume that the training and test datasets are independent and identically distributed, which is unreliable in fault diagnosis scenarios. Various factors cause significant differences in the distribution of the acquired data, a phenomenon known as domain shift. If data from multiple working environments or different machines are treated as a single entity to train a traditional deep learning model, the diagnostic performance will degrade due to the significant differences in data distribution.

[0003] In the context of transfer learning, domain adaptation (DA) is an effective method for solving the domain transfer problem. By reducing distributional differences between domains and capturing transferable representations, DA can learn the ability to distinguish fault types. However, DA assumes that the target domain data can be obtained in advance, which limits its applicability in certain situations, such as when data cannot be collected in advance or when real-time diagnostics are required.

[0004] To fully leverage labeled data for cross-domain diagnostics in unknown domains, Domain Generalization (DG) was developed as an extension of Data Acquisition (DA). DG uses data from a known source domain to train a deep model, which is then directly applied to diagnostic tasks in a previously unknown target domain. Notably, the DG method eliminates the dependency on target domain data during training, facilitating end-to-end processing of new target domains. The primary goal of DG is to enhance the generalization ability and robustness of the trained model, enabling its application to any target domain with the same fault type.

[0005] In fault diagnosis scenarios, rotating machinery data often exhibits class imbalance, meaning the data is unevenly distributed across different classes. Currently, most rotating machinery in use operates in a healthy state, ceasing operation upon detecting a fault. This leads to a situation where healthy state samples are easier to obtain than faulty state samples in real-world industrial environments. Consequently, the number of healthy class samples significantly exceeds that of faulty class samples. Furthermore, the performance of domain generalization methods is also significantly affected by the substantial difference in the number of samples in different classes, resulting in a bias towards learning the majority class while neglecting the minority class. Therefore, class imbalance in rotating machinery data is a significant challenge that urgently needs attention in deep learning-based fault diagnosis domain generalization methods.

[0006] In summary, the research direction required by this invention is to provide a new method for diagnosing rotating machinery faults that can overcome the problems existing in the prior art and thus ensure the accuracy of rotating machinery fault diagnosis. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides a rotating machinery fault diagnosis method based on a conditionally decoupled representation enhancement network. Taking into account the domain offset and class imbalance characteristics of rotating machinery data under different operating conditions, the method employs a conditionally decoupled representation enhancement network to reduce the impact of class imbalance distribution on the model, improve the model's generalization ability, and thus effectively enhance the accuracy of rotating machinery fault diagnosis.

[0008] To achieve the above objectives, the technical solution adopted by this invention is: a method for diagnosing rotating machinery faults based on conditional decoupling representation enhancement networks, the specific steps of which are as follows:

[0009] Step 1: For the same type of rotating machinery, collect vibration monitoring data of each piece of equipment under different working conditions, including normal and fault states. Label the data according to the fault type to obtain a mechanical dataset for different working conditions.

[0010] Step 2: Select a dataset under some working conditions from the mechanical dataset in Step 1 as a multi-source domain dataset. Set the class distribution to be unbalanced for each source domain dataset in the multi-source domain dataset, that is, the number of samples in the healthy class is much greater than the number of samples in the faulty class.

[0011] Step 3: Divide the multi-source domain dataset processed in Step 2 into a training set and a validation set, and use the dataset under the remaining conditions in Step 1 as the target domain dataset.

[0012] Step 4: Construct an unbalanced domain generalized mechanical fault diagnosis model and train the model using the training set from Step 3;

[0013] Step 5: Optimize the trained unbalanced domain generalized mechanical fault diagnosis model using the validation set from Step 3.

[0014] Step Six: Input the target domain dataset from Step Three into the unbalanced domain generalized mechanical fault diagnosis model optimized in Step Five for testing. The unbalanced domain generalized mechanical fault diagnosis model performs fault diagnosis on each data point in the target domain dataset and compares it with the actual situation to determine its fault diagnosis accuracy (since the equipment state corresponding to each vibration data point in the target domain dataset is known, but unknown to the model, as this dataset was not used for model training and optimization, simulating unknown vibration data using the target domain dataset allows the model to perform fault diagnosis based on the data, thereby determining the model's judgment accuracy). If the required accuracy is achieved, it can be used to diagnose faults on the same type of rotating machinery; otherwise, repeat Steps Four and Five to continue training and optimizing the model until the trained and optimized model reaches the required accuracy.

[0015] Furthermore, in the multi-source domain dataset, each source domain dataset corresponds to a working condition, and the health status label space in each domain is the same, and the distribution of the number of samples in each source domain category is set to be unbalanced.

[0016] Furthermore, step four involves training to form an unbalanced domain generalized mechanical fault diagnosis model, specifically as follows:

[0017] Establish a conditional decoupling representation enhancement network, which includes a deep convolutional feature extractor, a linear classifier, and a multi-source cross-entropy loss Li. CE Enhanced cross-entropy loss L` CE Decoupling representation enhancement module and conditional discrimination contrast loss L CD Conditional decoupling means that the training phase of the augmented network is divided into two parts. The first is the warm-up phase: the training set is fed into a deep convolutional feature extractor to obtain high-dimensional latent features; then the high-dimensional latent features are input into a linear classifier to obtain prediction scores, and the multi-source domain cross-entropy loss L is calculated. CE Simultaneously, conditional discriminant comparison loss L is applied to high-dimensional latent features. CD Computation; Formal Stage: Input the high-dimensional latent features into the decoupled representation enhancement module, generate a certain amount of minority class features based on the difference in the number of majority and minority classes, input the enhanced minority class features into the linear classifier, and calculate the enhanced cross-entropy loss L'. CE The computational graph is enhanced by preserving features; and gradient calculation is performed for each loss, thereby training a generalized mechanical fault diagnosis model in the unbalanced domain.

[0018] Furthermore, the deep convolutional feature extractor comprises six layers of one-dimensional convolutional neural networks and one fully connected layer, and the linear classifier comprises three fully connected layers.

[0019] The multi-source domain cross-entropy loss is specifically as follows:

[0020]

[0021] Where K is the number of source domains, N i Let be the number of samples for each source domain, l be the cross-entropy loss, f(h) be the prediction score, and y be the number of samples for each source domain. a The true labels for the samples;

[0022] The enhanced cross-entropy loss is specifically as follows:

[0023]

[0024] Where N' i To increase the number of samples, l represents the cross-entropy loss. To enhance feature prediction scores, To enhance the label.

[0025] Furthermore, the enhancement process of the decoupling representation enhancement module is as follows:

[0026] One-dimensional mechanical vibration signals are decomposed into semantic information and domain-correlated perturbations using 1D-InstanceNorm. Specifically, considering a sample (x, y, d), the hidden space features are defined as h = G(x) ∈ R. L Where L is the feature length of the hidden space, and the one-dimensional instance normalization is as follows:

[0027]

[0028] Where z(h), μ(h), σ(h) ∈ R 1 μ(h) and σ(h) represent the mean and standard deviation of the instances, respectively.

[0029]

[0030] Normalized instances z(h) are considered as semantic information, while μ(h) and σ(h) are considered as domain-related disturbances.

[0031] Enhance the feature representation of a few classes by reorganizing semantic information and domain-related perturbations:

[0032]

[0033] Where h i and h j From different instances, due to The semantic representation comes from h i Therefore, the tags of enhanced instances are... With h iThe same applies; however, class generalization information cannot be captured simply by recombination of two feature instances. Therefore, complex interpolation is performed on z(h), μ(h), and σ(h) respectively to enhance them.

[0034] To improve the invariance of similar representations across different domains, normalized instances of the same type of feature from different domains are averaged; define r c This represents the average of the cross-domain semantic information for class c. Note that class c belongs to a minority of fault types.

[0035]

[0036] Where n c Given the number of samples in class c, and considering the diversity of class representations, a normalized individual instance z(h) is selected from a minority of classes c. c ), through the semantic information z(h) of individual class c c ) and the semantic average r of the same category c Perform interpolation to obtain the enhanced semantic z'(h) c To balance diversity and cross-domain invariance;

[0037] z'(h c )=λ c z(h c )+(1-λ c )r c

[0038] Where λ c ~Beta(α) c ,α c ) is the category interpolation factor;

[0039] The average of the disturbances calculated from all features in a certain domain provides domain-specific information. d ,v d :

[0040]

[0041] Where n d For domain d j The number of samples, representing the disturbance of a randomly sampled instance (μ(h)). d ),σ(h d Interpolate the average disturbance value corresponding to the source domain to which the instance belongs to simulate domain-related disturbance;

[0042] μ'(h)=λ d μ(h d )+(1-λ d )u d ,σ'(h)=λ d σ(hd )+(1-λ d )v d

[0043] Finally, combining the above formulas, the final enhanced feature is represented as:

[0044]

[0045] Based on the difference between the number of majority and minority health status categories, a certain number of minority category enhanced feature representations are generated to balance the category distribution of features in the latent space.

[0046] Furthermore, the conditional discrimination comparison loss L CD Specifically:

[0047] In anchor sample x a ∈B={b1,b2,···,b k In the case of}, where B represents the set of batches from K source domains, and the set of the same kind is defined as pos(a) = {x p ∈B:y p =y a The set of negative classes is defined as neg(a) = {x} q ∈B:y q ≠y a The latent space feature is defined as h = G(x) ∈ R. L Where L is the feature length of the hidden space; the conditional discriminative contrastive loss is expressed as follows:

[0048]

[0049] Where, N i Let K be the number of samples in the source domain, |pos(g)| be the cardinality of the positive sample set, and σ(a,b) = (a T ·b / |a|·|b|) represents the similarity score between vectors a and b, and τ is the temperature factor. As a class discriminant factor; specifically, Let y be the domain corresponding to the two hidden space features h(a) and h(p). a The sum of the number of classes: When h corresponds to the majority class, the smaller 1 / n is, the smaller the proportion of loss of the majority class in the total loss, which reduces the dominant position of the majority class in conditional contrastive loss. Similarly, when h belongs to the minority class, the larger 1 / n is, the larger the proportion of loss of the minority class in the total loss, the more attention the model pays to the minority class, and the more easily the semantic representations of the minority class cluster together, enabling the model to learn the semantic representations of the minority class. Furthermore, by minimizing L... CD The model can minimize the similarity between different classes and maximize the similarity between classes.

[0050] Furthermore, the loss function for the preheating stage is:

[0051] L w-total =λL CE +(1-λ)L CD

[0052] The total loss function for the formal phase is:

[0053] L total =λ(L CE +L' CE )+(1-λ)L CD

[0054] Where λ is a hyperparameter used to balance cross-entropy loss and contrastive loss.

[0055] Furthermore, backpropagation and the Adam optimization algorithm are used to optimize and adjust the unbalanced domain generalized fault diagnosis model during the training process, ultimately minimizing the total loss function of the model and completing the training process of the unbalanced domain generalized fault diagnosis model.

[0056] Compared with existing technologies, this invention addresses the domain offset and class imbalance characteristics of rotating machinery data under different operating conditions. It utilizes a parameter-shared one-dimensional deep convolutional feature extractor to capture high-level abstract fault features from both the known source domain and the unknown target domain in the acquired rotating machinery vibration signals. Then, a decoupled representation enhancement module generates high-quality class semantic information to enhance minority class features in the high-dimensional feature space, thereby balancing the data distribution between classes. Because the decoupled representation enhancement module enhances minority class features, it mitigates the impact of differences in the number of samples from different classes. Subsequently, conditional discriminant contrastive loss maximizes the similarity between similar features in the hidden space, focusing more attention on the minority class, mitigating the impact of domain offset, and thus increasing attention to the minority class and strengthening the compactness of intra-class features across domains. Finally, the desired imbalanced domain generalized mechanical fault diagnosis model is formed. This method not only overcomes the influence of class imbalance but also improves the generalization ability and robustness to unknown target domains. Experimental verification shows that this invention effectively improves the accuracy of rotating machinery fault diagnosis. Attached Figure Description

[0057] Figure 1 This is a diagram of the conditional decoupling representation enhancement network structure of the present invention.

[0058] Figure 2 This is a diagram of the backbone structure of the deep convolutional feature extractor and the linear classifier in this invention.

[0059] Figure 3 This is a comparison chart of the diagnostic results of this invention and other mainstream methods in experimental verification. Detailed Implementation

[0060] The present invention will be further described below.

[0061] like Figure 1 As shown, the specific steps of this invention are as follows:

[0062] Step 1: For the same type of rotating machinery, collect vibration monitoring data of each piece of equipment under different working conditions, including normal and fault states. Label the data according to the fault type to obtain a mechanical dataset for different working conditions.

[0063] Step 2: Select a subset of working conditions from the mechanical dataset in Step 1 as a multi-source domain dataset. Set an imbalanced class distribution for each source domain dataset, i.e., the number of samples in the healthy class is much greater than the number of samples in the faulty class. In the multi-source domain dataset, each source domain dataset corresponds to a working condition, and the health status label space in each domain is the same. Set an imbalanced distribution of the number of samples in each source domain category.

[0064] Step 3: Divide the multi-source domain dataset processed in Step 2 into a training set and a validation set, and use the dataset under the remaining conditions in Step 1 as the target domain dataset.

[0065] Step 4: Construct an unbalanced domain generalized mechanical fault diagnosis model and train the model using the training set from Step 3; the specific process is as follows:

[0066] Establish a conditional decoupling representation enhancement network, which includes a deep convolutional feature extractor, a linear classifier, and a multi-source cross-entropy loss Li. CE Enhanced cross-entropy loss L` CE Decoupling representation enhancement module and conditional discrimination contrast loss L CD The deep convolutional feature extractor comprises six one-dimensional convolutional neural networks and one fully connected layer, while the linear classifier comprises three fully connected layers, as shown below. Figure 2 As shown.

[0067] The multi-source domain cross-entropy loss is specifically as follows:

[0068]

[0069] Where K is the number of source domains, N i Let be the number of samples for each source domain, l be the cross-entropy loss, f(h) be the prediction score, and y be the number of samples for each source domain. a The true labels for the samples;

[0070] The enhanced cross-entropy loss is specifically as follows:

[0071]

[0072] Where N' i To increase the number of samples, l represents the cross-entropy loss. To enhance feature prediction scores, To enhance the label.

[0073] Conditional decoupling represents the reinforcement network training phase, which is divided into two parts. The first is a warm-up phase: the training set is fed into a deep convolutional feature extractor to obtain high-dimensional latent features; then, the high-dimensional latent features are input into a linear classifier to obtain prediction scores, and the multi-source domain cross-entropy loss L is calculated. CE Simultaneously, conditional discriminant comparison loss L is applied to high-dimensional latent features. CD Computation; Formal Stage: Input the high-dimensional latent features into the decoupled representation enhancement module, generate a certain amount of minority class features based on the difference in the number of majority and minority classes, input the enhanced minority class features into the linear classifier, and calculate the enhanced cross-entropy loss L'. CE Enhance the feature-preserving computation graph; after training, obtain a deep convolutional feature extractor and a linear classifier with generalization ability;

[0074] The decoupling representation enhances the module's enhancement process as follows:

[0075] One-dimensional mechanical vibration signals are decomposed into semantic information and domain-correlated perturbations using 1D-InstanceNorm. Specifically, considering a sample (x, y, d), the hidden space features are defined as h = G(x) ∈ R. L Where L is the feature length of the hidden space, and the one-dimensional instance normalization is as follows:

[0076]

[0077] Where z(h), μ(h), σ(h) ∈ R 1 μ(h) and σ(h) represent the mean and standard deviation of the instances, respectively.

[0078]

[0079] Normalized instances z(h) are considered as semantic information, while μ(h) and σ(h) are considered as domain-related disturbances.

[0080] Enhance the feature representation of a few classes by reorganizing semantic information and domain-related perturbations:

[0081]

[0082] Where h i and h j From different instances, due to The semantic representation comes from h i Therefore, the tags of enhanced instances are... With h i The same applies; however, class generalization information cannot be captured simply by recombination of two feature instances. Therefore, complex interpolation is performed on z(h), μ(h), and σ(h) respectively to enhance them.

[0083] To improve the invariance of similar representations across different domains, normalized instances of the same type of feature from different domains are averaged; define r c This represents the average of the cross-domain semantic information for class c. Note that class c belongs to a minority of fault types.

[0084]

[0085] Where n c Given the number of samples in class c, and considering the diversity of class representations, a normalized individual instance z(h) is selected from a minority of classes c. c ), through the semantic information z(h) of individual class c c ) and the semantic average r of the same category c Perform interpolation to obtain the enhanced semantic z'(h) c To balance diversity and cross-domain invariance;

[0086] z'(h c )=λ c z(h c )+(1-λ c )r c

[0087] Where λ c ~Beta(α) c ,α c ) is the category interpolation factor;

[0088] The average of the disturbances calculated from all features in a certain domain provides domain-specific information. d ,v d :

[0089]

[0090] Where n d For domain d j The number of samples, representing the disturbance of a randomly sampled instance (μ(h)). d ),σ(h d Interpolate the average disturbance value corresponding to the source domain to which the instance belongs to simulate domain-related disturbance;

[0091] μ'(h)=λ d μ(h d )+(1-λ d )u d,σ'(h)=λ d σ(h d )+(1-λ d )v d

[0092] Finally, combining the above formulas, the final enhanced feature is represented as:

[0093]

[0094] Based on the difference between the number of majority and minority health status categories, a certain number of minority category enhanced feature representations are generated to balance the category distribution of features in the latent space.

[0095] The conditional discriminant comparison loss L CD Specifically:

[0096] In anchor sample x a ∈B={b1,b2,…,b k In the case of}, where B represents the set of batches from K source domains, and the set of the same kind is defined as pos(a) = {x p ∈B:y p =y a The set of negative classes is defined as neg(a) = {x} q ∈B:y q ≠y a The latent space feature is defined as h = G(x) ∈ R. L Where L is the feature length of the hidden space; the conditional discriminative contrastive loss is expressed as follows:

[0097]

[0098] Where, N i Let K be the number of samples in the source domain, |pos(g)| be the cardinality of the positive sample set, and σ(a,b) = (a T ·b / |a|·|b|) represents the similarity score between vectors a and b, and τ is the temperature factor. As a class discriminant factor; specifically, Let y be the domain corresponding to the two hidden space features h(a) and h(p). a The sum of the number of classes: When h corresponds to the majority class, the smaller 1 / n is, the smaller the proportion of loss of the majority class in the total loss, which reduces the dominant position of the majority class in conditional contrastive loss. Similarly, when h belongs to the minority class, the larger 1 / n is, the larger the proportion of loss of the minority class in the total loss, the more attention the model pays to the minority class, and the more easily the semantic representations of the minority class cluster together, enabling the model to learn the semantic representations of the minority class. Furthermore, by minimizing L... CDThe model can minimize the similarity between different classes and maximize the similarity between classes.

[0099] Set global parameters: learning rate lr, number of warm-up iterations T0, maximum number of iterations T; initialize parameter θ. G The feature extractor and parameters are θ F The classifier.

[0100] See Figure 1 , Figure 1 To enhance the network structure diagram for conditional decoupling representation, when the number of iterations is less than T0, the model is in the warm-up phase, and its training process includes the following steps:

[0101] The multi-source domain hidden layer fault feature h is obtained through a feature extractor;

[0102] Calculate the conditional discriminative loss for the hidden layer feature h:

[0103]

[0104] The hidden layer feature h is input into the classifier to obtain the prediction score;

[0105] Calculate the multi-source domain cross-entropy loss between the predicted score and the true label:

[0106]

[0107] Calculate the objective optimization function for the preheating stage:

[0108] L w-total =λL CE +(1-λ)L CD ;

[0109] Optimize model parameters: θ G →θ G ',θ F →θ F '

[0110] When the number of iterations exceeds T0, the model is in the formal stage, and its training process includes the following steps:

[0111] The hidden layer features h of the multi-source domain are obtained through a feature extractor;

[0112] Calculate the conditional discriminative loss for the hidden layer feature h:

[0113]

[0114] Enhanced features are obtained by decoupling the representation enhancement module to enhance the minority class.

[0115] Combine hidden layer features h with enhanced features Input the data into the classifier to obtain a predicted score;

[0116] Calculate the multi-source domain cross-entropy loss L CE and enhanced cross-entropy loss L' CE

[0117]

[0118] Calculate the objective function for the formal phase:

[0119] L total =λ(L CE +L' CE )+(1-λ)L CD ;

[0120] Optimize model parameters: θ G '→θ G * ,θ F '→θ F * ;

[0121] See Figure 1 The decoupling representation enhancement module in the above process is used to obtain the category y. i Enhanced feature representation:

[0122]

[0123] Gradient calculation is performed for each loss, and backpropagation and Adam optimization algorithms are used to optimize and adjust the unbalanced domain generalized fault diagnosis model during training, so as to minimize the total loss function of the model and complete the training process of the unbalanced domain generalized fault diagnosis model.

[0124] Step 5: Optimize the trained unbalanced domain generalized mechanical fault diagnosis model using the validation set from Step 3.

[0125] The target domain dataset from step three is input into the unbalanced domain generalized mechanical fault diagnosis model optimized in step five for testing. The unbalanced domain generalized mechanical fault diagnosis model performs fault diagnosis on each data point in the target domain dataset and compares it with the actual situation to determine its fault diagnosis accuracy (since the equipment state corresponding to each vibration data point in the target domain dataset is known, but unknown to the model, as this dataset was not used for model training and optimization, the model is made to perform fault diagnosis based on the data by simulating unknown vibration data using the target domain dataset, thus determining the model's judgment accuracy). If the required accuracy is achieved, it is used to diagnose faults on the same type of rotating machinery; otherwise, steps four and five are repeated to continue training and optimizing the model until the trained and optimized model reaches the required accuracy.

[0126] Experiments have shown that:

[0127] To verify that the present invention has better performance in generalizing fault diagnosis of unbalanced domain of rotating machinery compared with existing methods, the University of Ottawa bearing fault dataset is used as an example to verify the effectiveness of the present invention.

[0128] Step 1, Data Processing: Divide each set of 4096 sampling points into one sample. The number of data samples used in different fields is described below.

[0129] Table 1: Sample size of different health types in the source and target domains of the University of Ottawa bearing dataset.

[0130]

[0131] Step 2, Dataset Partitioning: This dataset has four operating speed conditions: 1) acceleration, 2) deceleration, 3) acceleration followed by deceleration, and 4) deceleration followed by acceleration. These four different speed conditions are assigned to four domains, A1 to A4. The source domain is set to class imbalance for model training, and the target domain is set to class balance for evaluation. Four cross-condition fault diagnosis tasks are set, as shown in Table 2.

[0132] Table 2: Source and Target Domain Settings for Cross-Condition Tasks

[0133]

[0134] Step 3: Input the data into the model of this invention for training and testing. Set the learning rate lr to 0.001, the number of warm-up iterations T0 to 500, and the maximum number of iterations T to 1500.

[0135] Step 4: In order to evaluate the performance of fault diagnosis, a measure of diagnostic accuracy is introduced; diagnostic accuracy is defined as the ratio of the number of correctly classified test samples to the total number of test samples.

[0136] Step 5: To evaluate the superiority of this invention, we compared the method proposed in this invention with other currently mainstream fault diagnosis methods. The comparison results are as follows: Figure 3 As shown. The method proposed in this invention is denoted as Proposed, and is implemented through... Figure 3 As can be seen, the highest accuracy was achieved in tasks 1 through 4. Therefore, the method proposed in this invention has better robustness and generalization ability for cross-domain fault diagnosis of rotating machinery under class imbalance conditions. It can learn diagnostic knowledge from multiple source domain data, effectively improving the fault diagnosis accuracy for unknown target domains.

[0137] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for fault diagnosis of rotating machinery based on conditional decoupling representation enhancement networks, characterized in that, The specific steps are as follows: Step 1: For the same type of rotating machinery, collect vibration monitoring data of each piece of equipment under different working conditions, including normal and fault states. Label the data according to the fault type to obtain a mechanical dataset for different working conditions. Step 2: Select a dataset under some working conditions from the mechanical dataset in Step 1 as a multi-source domain dataset. Set the class distribution to be unbalanced for each source domain dataset in the multi-source domain dataset, that is, the number of samples in the healthy class is much greater than the number of samples in the faulty class. Step 3: Divide the multi-source domain dataset processed in Step 2 into a training set and a validation set, and use the dataset under the remaining conditions in Step 1 as the target domain dataset. Step 4: Construct an imbalanced domain generalized mechanical fault diagnosis model and train the model using the training set from Step 3. Specifically, this involves establishing a conditional decoupling representation enhancement network, which includes a deep convolutional feature extractor, a linear classifier, and a multi-source domain cross-entropy loss L0. CE Enhanced cross-entropy loss L` CE Decoupling representation enhancement module and conditional discrimination contrast loss L CD Conditional decoupling means that the training phase of the augmented network is divided into two parts. The first is the warm-up phase: the training set is fed into a deep convolutional feature extractor to obtain high-dimensional latent features; then the high-dimensional latent features are input into a linear classifier to obtain prediction scores, and the multi-source domain cross-entropy loss L is calculated. CE Simultaneously, conditional discriminant comparison loss L is applied to high-dimensional latent features. CD Computation; Formal Stage: Input the high-dimensional latent features into the decoupled representation enhancement module, generate a certain amount of minority class features based on the difference in the number of majority and minority classes, input the enhanced minority class features into the linear classifier, and calculate the enhanced cross-entropy loss L'. CE Enhance the feature-preserving computation graph; and perform gradient calculation for each loss, thereby training to form an unbalanced domain generalized mechanical fault diagnosis model; Step 5: Optimize the trained unbalanced domain generalized mechanical fault diagnosis model using the validation set from Step 3. Step Six: Input the target domain dataset from Step Three into the unbalanced domain generalized mechanical fault diagnosis model optimized in Step Five for testing. The unbalanced domain generalized mechanical fault diagnosis model performs fault diagnosis on each data point in the target domain dataset and compares it with the actual situation to determine its fault diagnosis accuracy. If the required accuracy is achieved, it can be used to diagnose faults in the same type of rotating machinery. Otherwise, repeat Steps Four and Five to continue training and optimizing the model until the trained and optimized model reaches the required accuracy.

2. The rotating machinery fault diagnosis method based on conditional decoupling representation enhancement network according to claim 1, characterized in that, In the multi-source domain dataset, each source domain dataset corresponds to a working condition, and the health status label space in each domain is the same, and the distribution of the number of samples in each source domain category is set to be unbalanced.

3. The rotating machinery fault diagnosis method based on conditional decoupling representation enhancement network according to claim 1, characterized in that, The deep convolutional feature extractor comprises six one-dimensional convolutional neural networks and one fully connected layer, and the linear classifier comprises three fully connected layers. The multi-source domain cross-entropy loss is specifically as follows: ； Where K is the number of source domains, N i Let be the number of samples for each source domain, l be the cross-entropy loss, f(h) be the prediction score, and y be the number of samples for each source domain. a The true labels for the samples; The enhanced cross-entropy loss is specifically as follows: ； Where N' i To increase the number of samples, l is the cross-entropy loss, f( To enhance feature prediction scores, To enhance the label.

4. The rotating machinery fault diagnosis method based on conditional decoupling representation enhancement network according to claim 1, characterized in that, The decoupling representation enhancement module enhancement process is as follows: One-dimensional instance normalization is used to decompose a one-dimensional mechanical vibration signal into semantic information and domain-correlated perturbation. Specifically, consider a sample The hidden space features are defined as follows: Where L is the feature length of the hidden space, and the one-dimensional instance normalization is as follows: ； Where z(h), μ(h), σ(h) ∈ R 1 , and These represent the mean and standard deviation of the instances, respectively. ； ； Normalized instances It is regarded as semantic information, and and Considered a nuisance related to the domain; Enhance the feature representation of a few classes by reorganizing semantic information and domain-related perturbations: ； in and From different instances, due to The semantic representation comes from Therefore, the tags of enhanced instances are... and The same applies; however, class generalization information cannot be captured simply by recombination of two feature instances. Therefore, complex interpolation is performed on z(h), μ(h), and σ(h) respectively to enhance them. To improve the invariance of similar representations across different domains, normalized instances of the same type of feature across different domains are averaged; a definition is defined. This represents the average of the cross-domain semantic information for class c. Note that class c belongs to a minority of fault types. ； in Given the number of samples in class c, and considering the diversity of class representations, a normalized individual instance is selected from a minority of classes c. By analyzing the semantic information of individual class c and the semantic average of the same category Perform interpolation to obtain enhanced semantics To balance diversity and cross-domain invariance; ； in It is a category interpolation factor; The average of the disturbances calculated from all features in a certain domain provides domain-specific information. : ； in For domain d j The number of samples, representing the perturbation of a randomly sampled instance (μ(h)). d ),σ(h d Interpolate the average disturbance value corresponding to the source domain to which the instance belongs to simulate domain-related disturbance; ； Finally, combining the above formulas, the final enhanced feature is represented as: ； Based on the difference between the number of majority and minority health status categories, a certain number of minority category enhanced feature representations are generated to balance the category distribution of features in the latent space.

5. The rotating machinery fault diagnosis method based on conditional decoupling representation enhancement network according to claim 1, characterized in that, The conditional discriminant comparison loss L CD Specifically: In anchor sample In this case, a set of the same kind is defined as The set of negative classes is defined as The hidden space features are defined as follows: Where L is the feature length of the hidden space; the conditional discriminative contrastive loss is expressed as follows: ； Where, N i The number of samples in the source domain K. The cardinality of the positive sample set. Let τ be the similarity score between vectors a and b, and τ be the temperature factor. As a class discriminant factor; specifically, For two hidden space features The domain corresponding to y a The sum of the number of classes: when h corresponds to the majority class, the smaller 1 / n is, the smaller the proportion of loss of the majority class in the total loss, which reduces the dominant position of the majority class in conditional contrastive loss; similarly, when h belongs to the minority class, the larger 1 / n is, the larger the proportion of loss of the minority class in the total loss, the more attention the model pays to the minority class, and the semantic representations of the minority class are more likely to cluster together, enabling the model to learn the semantic representations of the minority class.

6. The rotating machinery fault diagnosis method based on conditional decoupling representation enhancement network according to claim 1, characterized in that, The loss function for the preheating stage is: ； The total loss function for the formal phase is: ； in It is a hyperparameter used to balance cross-entropy loss and contrastive loss.

7. The rotating machinery fault diagnosis method based on conditional decoupling representation enhancement network according to claim 1, characterized in that, During the training process, backpropagation and the Adam optimization algorithm are used to optimize and adjust the unbalanced domain generalized fault diagnosis model, ultimately minimizing the model's total loss function and completing the training process of the unbalanced domain generalized fault diagnosis model.

Images