Rotating machine fault diagnosis method based on conditional decoupling representation enhanced network

Abstract

The invention discloses a rotating machine fault diagnosis method based on a conditional decoupling representation enhanced network, and the method comprises the steps: capturing fault features from a known source domain and an unknown target domain for an acquired rotating machine vibration signal through a deep convolution feature extractor for a scene of rotating machine data category imbalance and cross-domain fault diagnosis; high-quality class semantic information is generated through a decoupling representation enhancement module to enhance minority class features in a high-dimensional feature space, so that inter-class data distribution is balanced, the influence of number difference of different classes of samples is reduced, then, the attention degree on minority classes is improved through condition discrimination comparison loss, and the accuracy of classification is improved. The intra-class feature compactness of the same class cross-domain is enhanced; and finally, a required unbalanced domain generalization mechanical fault diagnosis model is formed. According to the method, the influence of class imbalance can be overcome, the generalization ability and robustness of an unknown target domain can be improved, and the accuracy of rotating machine fault diagnosis is effectively improved.

Description

Technical Field

[0001] The invention belongs to the field of unbalanced cross-operating condition fault diagnosis, and specifically is a rotating machinery fault diagnosis method based on a conditional decoupling representation enhanced network. Background Art

[0002] With the continuous improvement of productivity, precision rotating machinery has been developed and integrated on a large scale. However, the harsh production environment and high-intensity operation can easily lead to mechanical failure. Therefore, in order to ensure the stability and safety of the machinery itself, the monitoring and diagnosis of rotating machinery has become increasingly important. As the industry moves towards higher digitalization and automation, intelligent fault diagnosis models have received more and more attention. In recent years, data-driven deep learning methods have become a common method for fault diagnosis. However, traditional deep learning techniques assume that the training data set and the test data set are independent and identically distributed, which is unreliable in fault diagnosis scenarios. Various factors lead to significant differences in the distribution of the obtained data, which is called domain shift. Assuming that data from multiple working environments or different machinery are treated as a whole to train traditional deep learning models, the diagnostic performance will decline due to the significant distribution differences of the data.

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

[0004] In order to make full use of labeled data to perform cross-domain diagnosis on unknown domains, domain generalization (DG) is developed as an extension of DA. DG uses data from known source domains to train deep models, which are then directly applied to diagnosis tasks in previously unknown target domains. It is worth noting that the DG method eliminates the dependence on target domain data during training and facilitates end-to-end processing of new target domains. The main goal of DG is to enhance the generalization and robustness of the trained model and apply it to any target domain with the same fault type.

[0005] In fault diagnosis scenarios, rotating machinery data often exhibit class imbalance, that is, the data is unevenly distributed among different classes. Rotating machinery equipment currently in use mainly operates in a healthy state and stops operating once a fault is detected. This results in that in actual industrial environments, healthy state samples are easier to obtain than fault state samples. Therefore, the number of healthy class samples is significantly greater than that of fault class samples. In addition, the performance of domain generalization methods is also affected by the significant difference in the number of samples in different classes, resulting in a bias towards learning the majority class and ignoring the minority class. Therefore, in the domain generalization methods of fault diagnosis based on deep learning, the class imbalance of rotating machinery data is a major challenge that needs urgent attention.

[0006] In summary, how to provide a new rotating machinery fault diagnosis method that can overcome the problems existing in the above-mentioned prior art and thus ensure the accuracy of rotating machinery fault diagnosis is the research direction required by the present invention. Summary of the invention

[0007] In view of the problems existing in the above-mentioned prior art, the present invention provides a rotating machinery fault diagnosis method based on a conditional decoupled representation enhanced network. According to the characteristics of domain offset and class imbalance of rotating machinery data under different working conditions, a conditional decoupled representation enhanced network is adopted to reduce the impact of class imbalance distribution on the model and improve the generalization ability of the model, thereby effectively improving the accuracy of rotating machinery fault diagnosis.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is: a rotating machinery fault diagnosis method based on a conditional decoupling representation enhanced network, the specific steps are:

[0009] Step 1: For the same type of rotating mechanical equipment, collect vibration monitoring data of each device in normal state and fault state under different working conditions, and label the data according to the fault type to obtain mechanical data sets for different working conditions;

[0010] Step 2: Select data sets under some working conditions from the mechanical data set in step 1 as multi-source domain data sets, and set the category distribution imbalance for each source domain data set in the multi-source domain data set, that is, the number of samples in the healthy class is much larger than the number of samples in the fault 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 working conditions in step 1 as the target domain dataset;

[0012] Step 4: construct a generalized mechanical fault diagnosis model in the imbalance domain and train the model using the training set in step 3;

[0013] Step 5: Use the validation set in step 3 to optimize the trained imbalanced domain generalized mechanical fault diagnosis model;

[0014] Step six, input the target domain data set of 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 in the target domain data set, and compares it with the actual situation to determine its fault diagnosis accuracy (since the equipment status corresponding to each vibration data in the target domain data set is known, but unknown to the model, because the data set is not used for model training and optimization, the unknown vibration data is simulated by the target domain data set so that the model performs fault diagnosis based on the data, thereby determining the model judgment accuracy). If the required accuracy is achieved, it is used to perform fault diagnosis on the same type of rotating mechanical equipment; 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 data set, each source domain data set corresponds to a working condition, and the health status labeling space in each domain is the same, and the number of samples of each source domain category is set to be unbalanced.

[0016] Furthermore, the step 4 trains to form an unbalanced domain generalized mechanical fault diagnosis model, specifically:

[0017] Establish a conditional decoupled representation enhancement network, which includes a deep convolutional feature extractor, a linear classifier, and a multi-source domain cross entropy loss L CE , enhanced cross entropy loss L` CE , decoupled representation enhancement module and conditional discriminant contrast loss L CD The conditional decoupling representation enhancement network training phase is divided into two parts. The first is the warm-up phase: the training set enters the deep convolutional feature extractor to obtain high-dimensional latent features; then the high-dimensional latent features are input into the linear classifier to obtain the prediction score, and the multi-source domain cross entropy loss L is calculated. CE , and at the same time, conditional discriminant contrast loss L is performed on high-dimensional latent features CD Calculation; 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 computational graph; and perform gradient calculations on each loss to train a generalized mechanical fault diagnosis model for imbalanced domains.

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

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

[0020]

[0021] Where K is the number of source domains, N i is the number of samples in each source domain, l is the cross entropy loss, f(h) is the prediction score, y a is the true label of the sample;

[0022] The enhanced cross entropy loss is specifically:

[0023]

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

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

[0026] One-dimensional instance normalization (1D-InstanceNorm) is used to decompose the one-dimensional mechanical vibration signal into semantic information and domain-related nuisance. Specifically, considering a sample (x, y, d), the latent space feature is defined as h = G(x)∈R L , where L is the characteristic length of the latent space, and the one-dimensional instance is normalized 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 instance respectively:

[0029]

[0030] The normalized instance z(h) is regarded as semantic information, while μ(h) and σ(h) are regarded as domain-related nuisances;

[0031] Enhance the feature representation of minority classes by recombining semantic information and domain-related nuisance:

[0032]

[0033] where h i and h j From different instances, due to The semantic representation comes from h i , thus enhancing the instance label With h iThe same; however, the reorganization of only two feature instances cannot capture class generalization information, so the following complex interpolation is performed on z(h), μ(h), σ(h) to enhance them;

[0034] In order to improve the invariance of the same type of representation in different domains, the normalized instances of the same type of features in different domains are averaged; define r c is the average value of semantic information of class c across domains. Note that class c belongs to the minority fault type:

[0035]

[0036] Where n c is the number of samples in category c. Taking into account the diversity of category representation, a standardized individual instance z(h c ), by analyzing the individual c-type semantic information z(h c ) and the semantic mean value r of the same category c Interpolate to obtain enhanced semantics 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 nuisance calculated for all features in a certain field is averaged to provide specific information u for that field d ,v d :

[0040]

[0041] Where n d For domain d j The number of samples, the nuisance (μ(h d ),σ(h d )) interpolates the nuisance mean value corresponding to the source domain to which the instance belongs, simulating domain-dependent nuisance;

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

[0043] Finally, combined with the above formula, the final enhanced feature is expressed as:

[0044]

[0045] According to 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 contrast loss L CD Specifically:

[0047] In the anchor sample x a ∈B={b1,b2,···,b k}, where B represents the set of batches from K source domains, and the set of the same type 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 latent space; the conditional discriminant contrast loss is expressed as follows:

[0048]

[0049] Among them, N i is the number of samples in the source domain K, |pos(g)| is the cardinality of the positive sample set, σ(a,b)=(a T ·b / |a|·|b|) is the similarity score of vectors a and b, τ is the temperature factor, is the class discriminant factor; specifically, is the y corresponding to the domain of the two latent space features h(a) and h(p) a The sum of the number of classes. When the class corresponding to h is the majority class, the smaller 1 / n is, the smaller the loss ratio of the majority class in the total loss is, which reduces the dominant position of the majority class in the conditional contrast loss. Similarly, when h belongs to the minority class, the larger 1 / n is, the larger the loss ratio of the minority class in the total loss is, the model increases the attention to the minority class, and the semantic representation of the minority class is easier to gather together, so that the model can learn the semantic representation of the minority class. In addition, by minimizing L CD , the model can minimize the similarity between different classes and maximize the similarity between the same class.

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

[0051] L w-total =λL CE +(1-λ)L CD

[0052] The total loss function in the formal stage is:

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

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

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

[0056] Compared with the prior art, the present invention aims at the characteristics of domain shift and class imbalance of rotating machinery data under different working conditions. For the collected rotating machinery vibration signal, a parameter-sharing one-dimensional deep convolutional feature extractor is used to capture high-level abstract fault features from the known source domain and the unknown target domain. Then, a decoupled representation enhancement module is designed to generate high-quality class semantic information to enhance the minority class features in the high-dimensional feature space, thereby balancing the data distribution between classes; since the decoupled representation enhancement module enhances the minority class features, the influence of the difference in the number of samples of different categories is reduced. Subsequently, the conditional discriminant contrast loss can maximize the similarity between similar features in the hidden space, and pay more attention to the minority class, reduce the influence of domain shift, thereby increasing the attention to the minority class, and strengthening the compactness of intra-class features of the same cross-domain; finally, the desired unbalanced domain generalized mechanical fault diagnosis model is formed. This method can not only overcome the influence of class imbalance, but also improve the generalization ability and robustness of the unknown target domain. Experimental verification shows that the invention can effectively improve the accuracy of rotating machinery fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

[0059] Figure 3 It is a comparison chart of the diagnostic results of the present invention and other mainstream methods in experimental proof. DETAILED DESCRIPTION

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

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

[0062] Step 1: For the same type of rotating mechanical equipment, collect vibration monitoring data of each device in normal state and fault state under different working conditions, and label the data according to the fault type to obtain mechanical data sets for different working conditions;

[0063] Step 2: Select data sets under some working conditions from the mechanical data set in step 1 as multi-source domain data sets, and set the category distribution imbalance for each source domain data set in the multi-source domain data sets, that is, the number of samples of the healthy class is much larger than the number of samples of the fault class; in the multi-source domain data sets, each source domain data set corresponds to a working condition, and the health status labeling space in each domain is the same, and the distribution of the number of samples of each source domain category is set to be imbalanced.

[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 working conditions in step 1 as the target domain dataset;

[0065] Step 4: Build a generalized mechanical fault diagnosis model in the imbalance domain and train the model using the training set in step 3. The specific process is as follows:

[0066] Establish a conditional decoupled representation enhancement network, which includes a deep convolutional feature extractor, a linear classifier, and a multi-source domain cross entropy loss L CE , enhanced cross entropy loss L` CE , decoupled representation enhancement module and conditional discriminant contrast loss L CD ; The deep convolutional feature extractor includes six layers of one-dimensional convolutional neural network and one fully connected layer, and the linear classifier includes three fully connected layers, such as Figure 2 shown.

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

[0068]

[0069] Where K is the number of source domains, N i is the number of samples in each source domain, l is the cross entropy loss, f(h) is the prediction score, y a is the true label of the sample;

[0070] The enhanced cross entropy loss is specifically:

[0071]

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

[0073] The training phase of the conditional decoupling representation enhancement network is divided into two parts. The first is the warm-up phase: the training set enters the deep convolutional feature extractor to obtain high-dimensional latent features; then the high-dimensional latent features are input into the linear classifier to obtain the prediction score, and the multi-source domain cross entropy loss L is calculated. CE , and at the same time, conditional discriminant contrast loss L is performed on high-dimensional latent features CD Calculation; 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 computational graph; after training, a deep convolutional feature extractor and linear classifier with generalization ability are obtained;

[0074] The enhancement process of the decoupled representation enhancement module is as follows:

[0075] One-dimensional instance normalization (1D-InstanceNorm) is used to decompose the one-dimensional mechanical vibration signal into semantic information and domain-related nuisance. Specifically, considering a sample (x, y, d), the latent space feature is defined as h = G(x)∈R L , where L is the characteristic length of the latent space, and the one-dimensional instance is normalized 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 instance respectively:

[0078]

[0079] The normalized instance z(h) is regarded as semantic information, while μ(h) and σ(h) are regarded as domain-related nuisances;

[0080] Enhance the feature representation of minority classes by recombining semantic information and domain-related nuisance:

[0081]

[0082] where h i and h j From different instances, due to The semantic representation comes from h i , thus enhancing the instance label With h i The same; however, the reorganization of only two feature instances cannot capture class generalization information, so the following complex interpolation is performed on z(h), μ(h), σ(h) to enhance them;

[0083] In order to improve the invariance of the same type of representation in different domains, the normalized instances of the same type of features in different domains are averaged; define r c is the average value of semantic information of class c across domains. Note that class c belongs to the minority fault type:

[0084]

[0085] Where n c is the number of samples in category c. Taking into account the diversity of category representation, a standardized individual instance z(h c ), by analyzing the individual c-type semantic information z(h c ) and the semantic mean value r of the same category c Interpolate to obtain enhanced semantics 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 nuisance calculated for all features in a certain field is averaged to provide specific information u for that field d ,v d :

[0089]

[0090] Where n d For domain d j The number of samples, the nuisance (μ(h d ),σ(h d )) interpolates the nuisance mean value corresponding to the source domain to which the instance belongs, simulating domain-dependent nuisance;

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

[0092] Finally, combined with the above formula, the final enhanced feature is expressed as:

[0093]

[0094] According to 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 contrast loss L CD Specifically:

[0096] In the anchor sample x a ∈B={b1,b2,…,b k}, where B represents the set of batches from K source domains, and the set of the same type 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 latent space; the conditional discriminant contrast loss is expressed as follows:

[0097]

[0098] Among them, N i is the number of samples in the source domain K, |pos(g)| is the cardinality of the positive sample set, σ(a,b)=(a T ·b / |a|·|b|) is the similarity score of vectors a and b, τ is the temperature factor, is the class discriminant factor; specifically, is the y corresponding to the domain of the two latent space features h(a) and h(p) a The sum of the number of classes. When the class corresponding to h is the majority class, the smaller 1 / n is, the smaller the loss ratio of the majority class in the total loss is, which reduces the dominant position of the majority class in the conditional contrast loss. Similarly, when h belongs to the minority class, the larger 1 / n is, the larger the loss ratio of the minority class in the total loss is, the model increases the attention to the minority class, and the semantic representation of the minority class is easier to gather together, so that the model can learn the semantic representation of the minority class. In addition, by minimizing L CD, the model can minimize the similarity between different classes and maximize the similarity between the same class.

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

[0100] See also Figure 1 , Figure 1 The conditional decoupling represents the enhanced network structure diagram. When the number of iterations is less than T0, the model is in the warm-up stage, and its training process includes the following steps:

[0101] Obtain multi-source domain hidden layer fault features h through feature extractor;

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

[0103]

[0104] Input the hidden feature h into the classifier to obtain the prediction score;

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

[0106]

[0107] Calculate the target optimization function during the warm-up phase:

[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 is greater than T0, the model is in the formal stage, and its training process includes the following steps:

[0111] Obtain multi-source domain hidden layer features h through feature extractor;

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

[0113]

[0114] Enhanced features are obtained by enhancing the minority classes through the decoupled representation enhancement module

[0115] The hidden layer features h and the enhanced features Input to the classifier to obtain the prediction score;

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

[0117]

[0118] Calculate the objective optimization function of 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 also Figure 1 The decoupling in represents the enhancement module, and according to the above enhancement process, the category y is obtained i The enhanced feature representation of is:

[0122]

[0123] The gradient of each loss is calculated. At the same time, the back propagation and Adam optimization algorithms are used to optimize and adjust the imbalanced domain generalized fault diagnosis model during the training process, and finally the total loss function of the model is minimized to complete the training process of the imbalanced domain generalized fault diagnosis model.

[0124] Step 5: Use the validation set in step 3 to optimize the trained imbalanced domain generalized mechanical fault diagnosis model;

[0125] The target domain data set in 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 in the target domain data set, and compares it with the actual situation to determine its fault diagnosis accuracy (since the equipment status corresponding to each vibration data in the target domain data set is known, but unknown to the model, because the data set is not used for model training and optimization, the unknown vibration data is simulated by the target domain data set so that the model performs fault diagnosis based on the data, thereby determining the model judgment accuracy). If the required accuracy is achieved, it is used to perform fault diagnosis on the same type of rotating mechanical equipment; 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] Experimental proof:

[0127] In order to verify that the present invention has better performance than existing methods in generalized fault diagnosis of rotating machinery imbalance domain, 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 every 4096 sampling points into one sample. The number of data samples used in different fields is described as follows.

[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, data set division: The data set has four operating speed conditions: 1) speed up, 2) deceleration, 3) speed up first and then decelerate, 4) decelerate first and then speed up; the above four different speed conditions are set as four domains from Al to A4, the source domain is set as class imbalance for model training, and the target domain is set as 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 the present 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, the diagnostic accuracy is introduced; the 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: In order to evaluate the superiority of the present invention, we compared the method proposed by the present invention with other current mainstream fault diagnosis methods. The comparison results are as follows: Figure 3 As shown. The method proposed in the present invention is represented by Proposed. Figure 3 It can be seen that the highest accuracy was obtained in tasks 1 to 4. Therefore, the method proposed in the present invention has better robustness and generalization ability for cross-domain fault diagnosis of rotating machinery under the condition of class imbalance. It can learn diagnostic knowledge from multiple source domain data and effectively improve the fault diagnosis accuracy of unknown target domains.

[0137] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for fault diagnosis of rotating machinery based on conditional decoupling representation enhanced network, characterized in that: The specific steps are: Step 1: For the same type of rotating mechanical equipment, collect vibration monitoring data of each device in normal state and fault state under different working conditions, and label the data according to the fault type to obtain mechanical data sets for different working conditions; Step 2: Select data sets under some working conditions from the mechanical data set in step 1 as multi-source domain data sets, and set the category distribution imbalance for each source domain data set in the multi-source domain data set, that is, the number of samples in the healthy class is much larger than the number of samples in the fault 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 working conditions in step 1 as the target domain dataset; Step 4: construct a generalized mechanical fault diagnosis model in the imbalance domain and train the model using the training set in step 3; Step 5: Use the validation set in step 3 to optimize the trained imbalanced domain generalized mechanical fault diagnosis model; Step six, input the target domain data set of 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 in the target domain data set, and compares it with the actual situation to determine its fault diagnosis accuracy. If the required accuracy is achieved, it is used to perform fault diagnosis on the same type of rotating mechanical equipment; 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 enhanced network according to claim 1 is characterized in that: In the multi-source domain data set, each source domain data set corresponds to a working condition, and the health status labeling space in each domain is the same, and the number of samples of each source domain category is set to be unbalanced.

3. The rotating machinery fault diagnosis method based on conditional decoupling representation enhanced network according to claim 1 is characterized in that: The fourth step is to train and form a generalized mechanical fault diagnosis model in the imbalance domain, specifically: Establish a conditional decoupled representation enhancement network, which includes a deep convolutional feature extractor, a linear classifier, and a multi-source domain cross entropy loss L CE , enhanced cross entropy loss L` CE , decoupled representation enhancement module and conditional discriminant contrast loss L CD The conditional decoupling representation enhancement network training phase is divided into two parts. The first is the warm-up phase: the training set enters the deep convolutional feature extractor to obtain high-dimensional latent features; then the high-dimensional latent features are input into the linear classifier to obtain the prediction score, and the multi-source domain cross entropy loss L is calculated. CE , and at the same time, conditional discriminant contrast loss L is performed on high-dimensional latent features CD Calculation; 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 computational graph; and perform gradient calculations on each loss to train a generalized mechanical fault diagnosis model for imbalanced domains.

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

5. The rotating machinery fault diagnosis method based on conditional decoupling representation enhanced network according to claim 3 is characterized in that: The enhancement process of the decoupling representation enhancement module is as follows: One-dimensional instance normalization (1D-InstanceNorm) is used to decompose the one-dimensional mechanical vibration signal into semantic information and domain-related nuisance. Specifically, considering a sample (x, y, d), the latent space feature is defined as h = G(x)∈R L , where L is the characteristic length of the latent space, and the one-dimensional instance is normalized as follows: where z(h),μ(h),σ(h)∈R 1 , μ(h) and σ(h) represent the mean and standard deviation of the instance respectively: The normalized instance z(h) is regarded as semantic information, while μ(h) and σ(h) are regarded as domain-related nuisances; Enhance the feature representation of minority classes by recombining semantic information and domain-related nuisance: where h i and h j From different instances, due to The semantic representation comes from h i , thus enhancing the instance label With h i The same; however, the reorganization of only two feature instances cannot capture class generalization information, so the following complex interpolation is performed on z(h), μ(h), σ(h) to enhance them; In order to improve the invariance of the same type of representation in different domains, the normalized instances of the same type of features in different domains are averaged; define r c is the average value of semantic information of class c across domains. Note that class c belongs to the minority fault type: where n c is the number of samples in category c. Taking into account the diversity of category representation, a standardized individual instance z(h c ), by analyzing the individual c-type semantic information z(h c ) and the semantic mean value r of the same category c Interpolate to obtain enhanced semantics z'(h c ), to balance diversity and cross-domain invariance; z'(h c )=λ c z(h c )+(1-λ c )r c where λ c ~Beta(α c ,α c ) is the category interpolation factor; The nuisance calculated for all features in a certain field is averaged to provide specific information u for that field d ,v d : where n d For domain d j The number of samples, the nuisance (μ(h d ),σ(h d )) interpolates the nuisance mean value corresponding to the source domain to which the instance belongs, simulating domain-dependent nuisance; μ'(h)=λ d μ(h d )+(1-λ d )u d ,σ'(h)=λ d σ(h d )+(1-λ d )v d Finally, combined with the above formula, the final enhanced feature is expressed as: According to 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.

6. The rotating machinery fault diagnosis method based on conditional decoupling representation enhanced network according to claim 3 is characterized in that: The conditional discriminant contrast loss L CD Specifically: In the anchor sample x a ∈B={b1,b2,···,b k }, the set of the same type 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 latent space; the conditional discriminant contrast loss is expressed as follows: Among them, N i is the number of samples in the source domain K, |pos(g)| is the cardinality of the positive sample set, σ(a,b)=(a T ·b / |a|·|b) is the similarity score of vectors a and b, τ is the temperature factor, is the class discriminant factor; specifically, is the y corresponding to the domain of the two latent space features h(a) and h(p) a The sum of the number of classes. When the class corresponding to h is the majority class, the smaller 1 / n is, the smaller the loss ratio of the majority class in the total loss is, which reduces the dominant position of the majority class in the conditional contrast loss. Similarly, when h belongs to the minority class, the larger 1 / n is, the larger the loss ratio of the minority class in the total loss is, the model increases its attention to the minority class, and the semantic representation of the minority class is easier to cluster together, enabling the model to learn the semantic representation of the minority class.

7. The rotating machinery fault diagnosis method based on conditional decoupling representation enhanced network according to claim 3 is characterized in that: The loss function of the warm-up phase is: THE w-total =λL CE +(1-λ)L CD The total loss function in the formal stage is: L total =λ(L CE +L' CE )+(1-λ)L CD Where λ is a hyperparameter that is used to balance the cross entropy loss and contrastive loss.

8. The rotating machinery fault diagnosis method based on conditional decoupling representation enhanced network according to claim 3 is characterized in that: During the training process, back propagation and Adam optimization algorithms are used to optimize and adjust the imbalanced domain generalized fault diagnosis model, ultimately minimizing the total loss function of the model and completing the imbalanced domain generalized fault diagnosis model training process.

Images