Rotating machinery fault diagnosis method and system based on joint domain adaptive network

Abstract

The application provides a rotating machinery fault diagnosis method and system based on a joint domain adaptive network, comprising the following steps: an initial pre-training model is trained by combining consistency regularization loss and pseudo labels, so as to improve the prediction accuracy of target domain data in the initial stage of the domain adaptation process; and the utilization rate of fault information in unlabeled data is improved by generating pseudo labels for the target domain unlabeled data.

Description

Technical Field

[0001] This invention relates to the field of rotating machinery fault diagnosis technology, and in particular to a rotating machinery fault diagnosis method and system based on joint domain adaptive networks. Background Technology

[0002] Traditional intelligent fault diagnosis methods rely on sufficient labeled data to train diagnostic models. However, in real-world industrial scenarios, most equipment operates normally, making fault data difficult to obtain. Based on the idea that diagnostic knowledge can be reused across multiple related machines—for example, diagnostic knowledge of laboratory bearings may help identify the health status of bearings in industrial settings—domain adaptation combined with semi-supervised learning can transfer diagnostic knowledge learned from labeled source domain data to unlabeled target domain data, thus effectively addressing the problem of unlabeled data in some real-world industrial scenarios. However, current domain adaptation-based fault diagnosis mainly suffers from the following problems:

[0003] Some methods, such as distance metrics and adversarial learning, learn domain-invariant features between the source and target domains. However, these approaches only align the data from the perspective of the feature space, ignoring the influence of label information on the alignment. This results in only globally aligning the two data distributions, neglecting the alignment between class distributions.

[0004] Using a single domain-adaptive alignment method will greatly affect the effectiveness of this alignment method when there are large differences in data distribution, and it will be unable to capture the multimodal information contained in the data.

[0005] The method of directly using a model trained on source domain data to predict target domain data and then performing domain adaptation alignment is problematic in the early stages of training. Due to the significant differences in the distribution between source and target domain data, the prediction performance of the target domain data is poor. Furthermore, domain adaptation depends on the prediction results of the target domain data. If the prediction performance of the target domain data is poor, the model will be difficult to train and will not converge. Summary of the Invention

[0006] The purpose of this invention is to provide a rotating machinery fault diagnosis method and system based on a joint domain adaptation network, so as to solve the problem of poor fault diagnosis effect of existing domain adaptation-based fault diagnosis methods.

[0007] To address the aforementioned technical problems, this invention provides a method for diagnosing rotating machinery faults based on joint domain adaptive networks, comprising:

[0008] By combining consistency regularization loss and pseudo-labels, an initial pre-trained model is trained to improve the prediction accuracy of target domain data in the initial stage of domain adaptation; and

[0009] By generating pseudo-labels for unlabeled data in the target domain, the utilization rate of fault information in unlabeled data can be improved.

[0010] Optionally, the rotating machinery fault diagnosis method based on joint domain adaptive networks further includes:

[0011] By combining joint maximum mean difference domain adaptation and conditional adversarial domain adaptation, and aligning the joint distribution of features between source and target domain data in multiple ways, as well as the joint distribution of labels between source and target domain data in multiple ways, this approach can capture the multimodal information contained in the data; and

[0012] The method involves globally aligning two joint distributions and aligning between different categories to improve the alignment accuracy under different data distribution differences and the alignment effect of similar features between the source and target domains.

[0013] Optionally, the rotating machinery fault diagnosis method based on joint domain adaptive networks further includes:

[0014] By introducing a semi-supervised pre-trained model, higher-confidence target domain data is introduced during the semi-supervised training process through threshold screening. This replaces the model trained directly with source domain data in the pre-adaptation stage to predict target domain data, thereby improving prediction accuracy.

[0015] Optionally, the rotating machinery fault diagnosis method based on joint domain adaptive networks further includes:

[0016] Two modules are constructed to perform domain adaptation for the source and target domains;

[0017] By leveraging the joint distribution difference between input features and output labels, domain adaptation is performed in the feature extraction and classification layers through joint maximum mean difference domain adaptation.

[0018] Domain adversarial training is performed between features and predicted labels to reduce domain drift;

[0019] Based on the uncertainty of the domain classifier's predictions, the samples in the domain classifier are reweighted using entropy-calculated weights ω to reduce the impact of target instances with uncertain predictions; and

[0020] The system utilizes two modules to maximize the differentiation between categories and to adapt to different domains across multiple modes.

[0021] Optionally, in the aforementioned rotating machinery fault diagnosis method based on joint domain adaptation networks, an initial pre-trained model is trained by combining consistency regularization loss and pseudo-labels, including:

[0022] Weak augmentation is performed on labeled data;

[0023] For unlabeled data, weak enhancement and strong enhancement are performed respectively, with the proportion of noise added for weak enhancement being different from that for strong enhancement.

[0024] For weakly augmented unlabeled data, predict pseudo-labels. When the model produces predictions above a threshold, retain them as pseudo-labels.

[0025] For strongly augmented unlabeled data, predict classification probabilities and use cross-entropy loss to measure the consistency between the strong and weak predictions; and

[0026] By combining the supervised loss of labeled data and the consistency regularization loss of unlabeled data, a pre-trained model is obtained, which reduces the error in predicting the target domain data during the domain adaptation stage.

[0027] Optionally, in the aforementioned rotating machinery fault diagnosis method based on joint domain adaptation networks, improving the alignment effect of similar features between the source and target domains includes:

[0028] Weak augmentation is performed on labeled source domain data and unlabeled target domain data respectively, and then fed into the pre-trained model obtained from the weak augmentation in the pre-training stage.

[0029] Perform multi-level linear transformations on the final features and labels to represent the joint distribution of features and labels;

[0030] A joint maximum mean difference domain adaptation method and a conditional domain adversarial domain adaptation method are employed to align the joint distribution of features and labels; and

[0031] By combining the loss from the label classification joint maximum mean difference domain adaptation method and the loss from the conditional domain adversarial domain adaptation method, domain-invariant features of the source and target domains are trained to improve the accuracy of fault diagnosis through transfer learning.

[0032] Optionally, in the aforementioned rotating machinery fault diagnosis method based on joint domain adaptation networks, the pre-training further includes:

[0033] Pre-training introduces consistency regularization loss;

[0034] Consistency regularization loss includes: after injecting noise into unlabeled data, the classifier outputs the same probability distribution, forcing an unlabeled sample to be classified into the same category as its augmented sample; and

[0035] By forcing an unlabeled sample to be classified into the same category as its augmented sample, information from unlabeled data is introduced into the pre-trained model, making the accuracy of the pre-trained model higher than that of a model trained directly using source domain data when predicting target domain data, thus avoiding the model's inability to converge in the early stages.

[0036] Optionally, in the rotating machinery fault diagnosis method based on joint domain adaptation networks, the joint distribution of alignment features and labels of the two domain adaptation modules further includes:

[0037] Combining two different domain adaptation methods to complementaryly align the joint distribution of data from two domains avoids situations where the difference in data distribution between the source and target domains exceeds a data distribution threshold: aligning the distribution between features while ignoring category information, and aligning the distributions of the two features holistically while ignoring the relationship between categories; and

[0038] By combining two different domain adaptation methods to complementarily align the joint distribution of data from two domains, we can avoid using a single domain adaptation module to align the joint distribution: because the complexity of the difference information between the distributions exceeds the complexity threshold, a single domain adaptation module cannot capture the multimodal information.

[0039] Optionally, the rotating machinery fault diagnosis method based on joint domain adaptive networks further includes:

[0040] Transferring data under different operating conditions using the same dataset; based on the differences in bearing or gear vibration signals collected under different operating conditions, and the fact that rotating machinery operates under varying operating conditions and the amount of data collected under at least one operating condition is less than the data amount threshold, the operating conditions with a data amount greater than the data amount threshold are used as source domain data, and the diagnostic knowledge trained from the source domain data is transferred to other operating conditions; the differences include load parameters and speed parameters.

[0041] Migration between different datasets; use bearing data collected through laboratory simulations; based on diagnostic knowledge of laboratory bearings, identify the health status of bearings in engineering scenarios for use in bearing fault diagnosis in engineering scenarios; and

[0042] Migration under different operating conditions with mixed fault types; based on the case where bearings and gears fail simultaneously, and the data distribution difference collected under different operating conditions is greater than the data distribution threshold, multiple domain adaptation methods and / or information from unlabeled data are adopted, and the impact of data distribution differences is reduced by introducing a pre-trained model combined with multiple domain adaptation modules.

[0043] The present invention also provides a rotating machinery fault diagnosis system based on a joint domain adaptive network, comprising:

[0044] The pre-trained model module is configured to train an initial pre-trained model by combining consistency regularization loss and pseudo-labels to improve the prediction accuracy of the target domain data in the initial stage of the domain adaptation process; and

[0045] The semi-supervised training module is configured to improve the utilization rate of fault information in unlabeled data by generating pseudo-labels for unlabeled data in the target domain.

[0046] In the rotating machinery fault diagnosis method and system based on joint domain adaptation network provided by the present invention, an initial pre-trained model is trained by combining consistency regularization loss and pseudo-labels, so that the prediction effect of the target domain data in the initial stage of the domain adaptation process is not too bad, avoiding the difficulty of model training and non-convergence, and solving the problem that the target domain data is not fully utilized in the existing domain adaptation fault diagnosis algorithm.

[0047] Furthermore, this invention combines two domain adaptation alignment methods: joint maximum mean difference and conditional adversarial domain adaptation. By aligning the joint distribution of source domain data and target domain data features and labels in multiple ways, it can capture the multimodal information contained in the data. It not only aligns the two distributions globally, but also considers the alignment between different categories, thus achieving a good alignment effect even when the data distributions are significantly different. This solves the problem of poor alignment of similar features between the source and target domains.

[0048] This invention proposes a semi-supervised rotating machinery fault diagnosis method based on a joint domain adaptation network, aiming to achieve good diagnostic results even in real-world industrial scenarios where labeled data is unavailable. This approach combines the ideas of semi-supervised learning and domain adaptation, designing a semi-supervised rotating machinery fault diagnosis method based on a joint domain adaptation network. It requires no prior knowledge, automatically learning domain-invariant features through a neural network. The target domain does not require a label; this method is a semi-supervised algorithm independent of target domain labels, thus addressing the problem of insufficient data labeling.

[0049] This invention provides a fault diagnosis algorithm with good diagnostic performance for scenarios with insufficient or no labeled fault data. Compared to traditional fault diagnosis algorithms that rely on a single domain adaptation alignment, this invention combines two domain adaptation alignment methods, considers the category information of the data, and aligns the joint distribution of the two data points rather than their marginal distributions, significantly reducing the probability of negative transfer and thus improving the transfer effect. Unlike some transfer fault diagnosis algorithms that directly use a model trained on source domain data to predict target domain data and then perform domain adaptation alignment, this invention first employs a pre-training method to introduce information from unlabeled data into the pre-trained model, improving the prediction effect on the target domain data to a certain extent and avoiding the problems of difficult model training and non-convergence. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the pre-trained topology of a rotating machinery fault diagnosis method based on a joint domain adaptive network according to an embodiment of the present invention;

[0051] Figure 2This is a schematic diagram of the domain adaptation topology of a rotating machinery fault diagnosis method based on a joint domain adaptation network according to an embodiment of the present invention. Detailed Implementation

[0052] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0053] It should be noted that the components in the various figures may be shown exaggeratedly for illustrative purposes and are not necessarily to scale. In each figure, the same reference numerals are used for components that are identical or have the same function.

[0054] In this invention, unless otherwise specified, "arranged on," "arranged above," and "arranged on" do not exclude the possibility of an intermediate element between them. Furthermore, "arranged on or above" merely indicates the relative positional relationship between two components, and in certain cases, such as when the product orientation is reversed, it can also be converted to "arranged below or under," and vice versa.

[0055] In this invention, the various embodiments are merely intended to illustrate the solutions of the invention and should not be construed as limiting.

[0056] In this invention, unless otherwise specified, the quantifiers “a” and “one” do not exclude scenarios involving multiple elements.

[0057] It should also be noted that, in the embodiments of the present invention, only a portion of the components or parts may be shown for clarity and simplicity. However, those skilled in the art will understand that, under the teachings of the present invention, necessary components or parts can be added as needed for specific scenarios. Furthermore, unless otherwise stated, features in different embodiments of the present invention can be combined with each other. For example, a feature in the second embodiment can replace a corresponding or functionally identical or similar feature in the first embodiment, and the resulting embodiment will also fall within the scope of disclosure or description of this application.

[0058] It should also be noted that, within the scope of this invention, the terms "same," "equal," and "equal to" do not imply that the two values ​​are absolutely equal, but rather allow for a certain reasonable margin of error. In other words, the terms also encompass "substantially the same," "substantially equal," and "substantially equal to." Similarly, in this invention, the directional terms "perpendicular to," "parallel to," etc., also encompass the meanings of "substantially perpendicular to" and "substantially parallel to."

[0059] Furthermore, the numbering of the steps in the methods of the present invention does not limit the execution order of the method steps. Unless otherwise specified, the method steps may be executed in different orders.

[0060] The following detailed description, in conjunction with the accompanying drawings and specific embodiments, provides a more comprehensive overview of the rotating machinery fault diagnosis method and system based on joint domain adaptive networks proposed in this invention. The advantages and features of this invention will become clearer from the following description. It should be noted that the accompanying drawings are in a very simplified form and use non-precise proportions, and are only used to facilitate and clarify the illustration of the embodiments of this invention.

[0061] The purpose of this invention is to provide a rotating machinery fault diagnosis method and system based on a joint domain adaptation network, so as to solve the problem of poor fault diagnosis effect of existing domain adaptation-based fault diagnosis methods.

[0062] To achieve the above objectives, the present invention provides a method and system for diagnosing rotating machinery faults based on a joint domain adaptation network, comprising: training an initial pre-trained model by combining consistency regularization loss and pseudo-labels to improve the prediction accuracy of target domain data in the initial stage of the domain adaptation process; and making full use of fault information in unlabeled data by generating pseudo-labels for unlabeled data in the target domain.

[0063] Figures 1-2 The first embodiment of the present invention is provided, which shows the pre-training and domain adaptation topology schematic diagrams of the rotating machinery fault diagnosis method based on joint domain adaptation networks.

[0064] This invention proposes a semi-supervised fault diagnosis method for rotating machinery based on a joint domain adaptation network, aiming to achieve good diagnostic results even in real-world industrial scenarios where labeled data is unavailable. This approach primarily combines the ideas of semi-supervised learning and domain adaptation to design a semi-supervised fault diagnosis method for rotating machinery based on a joint domain adaptation network; it offers the following advantages:

[0065] No prior knowledge is required; domain-invariant features are learned automatically entirely through neural networks.

[0066] The target domain does not require labels. This method is a semi-supervised algorithm that does not rely on target domain labels, which can solve the problem of insufficient data labeling.

[0067] By combining consistency regularization loss and pseudo-labels, an initial pre-trained model is obtained, ensuring that the prediction performance of the target domain data in the initial stage of domain adaptation is not too poor, avoiding difficulties in model training and convergence, and solving the problem of underutilization of target domain data in existing domain adaptation fault diagnosis algorithms; and

[0068] By combining two domain adaptation alignment methods—joint maximum mean difference and conditional adversarial domain adaptation—the joint distribution of features and labels of source and target domain data can be aligned in multiple ways. This captures the multimodal information contained in the data, aligning the two distributions globally and considering the alignment between different categories. As a result, it can achieve good alignment even when the data distributions are significantly different, thus solving the problem of poor alignment of similar features between the source and target domains.

[0069] like Figure 1 As shown, the pre-training process includes: weak augmentation for labeled data; weak and strong augmentation for unlabeled data, differing only in the proportion of noise added; predicting pseudo-labels for weakly augmented unlabeled data, retaining pseudo-labels only when the model produces predictions above a threshold; predicting classification probabilities for strongly augmented unlabeled data, using cross-entropy loss to measure the consistency between the strong and weak predictions; and combining the supervision loss for labeled data and the consistency regularization loss for unlabeled data to obtain a pre-trained model, thereby reducing the error in predicting target domain data during the domain adaptation stage.

[0070] like Figure 2 As shown, domain adaptation includes: performing weak augmentation on labeled source domain data and unlabeled target domain data respectively, and feeding them into the pre-trained model obtained from the weak augmentation in the pre-training stage; performing multi-level linear transformation on the final features f and labels l to represent the joint distribution of features and labels; aligning the joint distribution of features and labels using two domain adaptation methods, Joint Maximum Mean Difference (JMMD) and Conditional Domain Adversarial Analysis (CDA); and combining label classification loss, JMMD loss, and CDA loss to train and obtain domain-invariant features of the source and target domains, achieving a good transfer learning fault diagnosis effect.

[0071] This invention introduces a semi-supervised pre-trained model. Compared with a model trained directly with source domain data in the pre-adaptation stage to predict target domain data, the semi-supervised training process introduces relatively reliable target domain data through threshold screening, which can improve the accuracy of prediction.

[0072] This invention constructs two modules to achieve domain adaptation between the source and target domains. On one hand, it utilizes the joint distribution difference between input features and output labels to perform domain adaptation at the feature extraction and classification layers using JMMD. On the other hand, it performs domain adversarial training between features and predicted labels to reduce domain drift. Simultaneously, considering the uncertainty of the domain classifier's predictions, it reweights the samples in the domain classifier using entropy-calculated weights ω to reduce the influence of target instances with uncertain predictions. These two modules not only achieve maximum distinction between categories but also realize multi-modal domain adaptation.

[0073] Specifically, pre-training is necessary because it introduces consistency regularization loss. The idea behind consistency regularization loss is that even after unlabeled data is injected with noise, the classifier should still output the same probability distribution, forcing an unlabeled sample to be classified into the same category as its augmented counterpart. In this way, information from the unlabeled data is incorporated into the pre-trained model, resulting in higher accuracy for the pre-trained model compared to a model trained directly on source domain data. This avoids the problems of difficulty in early model training and convergence.

[0074] Furthermore, combining the alignment of the joint distribution of features and labels using two domain adaptation modules is necessary: ​​when the data distributions of the source and target domains differ significantly, simply aligning the distributions of features while ignoring category information only aligns the two feature distributions holistically, neglecting the relationships between categories, thus greatly affecting the alignment effect. Similarly, using only a single domain adaptation module to align the joint distribution is insufficient due to the complex differences between the distributions; a single module struggles to capture the inherent multimodal information and achieve a good alignment result. Therefore, combining two different domain adaptation methods is essential for complementary alignment of the joint distribution of data from the two domains.

[0075] This invention provides a fault diagnosis algorithm with good diagnostic performance for scenarios with insufficient or no labeled fault data. Compared to traditional fault diagnosis algorithms that rely on a single domain adaptation alignment, this invention combines two domain adaptation alignment methods, considers the category information of the data, and aligns the joint distribution of the two data points rather than their marginal distributions, significantly reducing the probability of negative transfer and thus improving the transfer effect. Compared to some transfer fault diagnosis algorithms that directly use a model trained on source domain data to predict target domain data and then perform domain adaptation alignment, this invention first employs a pre-training method to introduce information from unlabeled data into the pre-trained model, thereby improving the prediction effect on the target domain data to a certain extent and avoiding the problems of difficult model training and non-convergence.

[0076] Furthermore, this invention enables the transfer of the same dataset under different operating conditions. The vibration signals of bearings or gears collected under different operating conditions are different, such as different loads and different speeds. Considering that rotating machinery often operates under variable operating conditions, and the amount of data collected under some operating conditions may be insufficient, if this invention has sufficient labeled data under one of the operating conditions, this invention can use it as source domain data. Through this method, the diagnostic knowledge learned from the source domain data can be transferred to the operating conditions with insufficient data, and a good diagnostic effect can also be achieved.

[0077] Furthermore, this invention enables data transfer between different datasets. In real-world industrial scenarios, faults are rare, and the collected data is insufficient to train a high-performing model. In such cases, this invention can utilize bearing data collected through laboratory simulations. Diagnostic knowledge from laboratory bearings may help identify the health status of bearings in engineering scenarios, and this solution can be used for bearing fault diagnosis in engineering settings.

[0078] Furthermore, this invention enables migration under different operating conditions with mixed fault types. In real industrial scenarios, bearings and gears may fail simultaneously, resulting in significant differences in data distribution collected under different operating conditions. The mutual influence of different fault types further exacerbates these differences. In such cases, using a single domain adaptation method or without incorporating unlabeled data is unlikely to achieve effective alignment. By introducing a pre-trained model combined with multiple domain adaptation modules, the impact of these differences can be significantly reduced.

[0079] In summary, the above embodiments have provided detailed descriptions of different configurations of the rotating machinery fault diagnosis method and system based on joint domain adaptive networks. Of course, this invention includes, but is not limited to, the configurations listed in the above embodiments. Any modifications made based on the configurations provided in the above embodiments are within the scope of protection of this invention. Those skilled in the art can apply the knowledge gained from the above embodiments to other situations.

[0080] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0081] The above description is merely a description of preferred embodiments of the present invention and is not intended to limit the scope of the present invention in any way. Any changes or modifications made by those skilled in the art based on the above disclosure shall fall within the protection scope of the claims.

Claims

1. A rotating machinery fault diagnosis method based on a joint domain adaptive network, characterized in that, The method comprises the following steps: An initial pre-training model is trained based on consistency regularization loss and pseudo labels to improve the prediction accuracy of target domain data in the initial stage of domain adaptation; And The utilization rate of fault information in unlabeled data is improved by generating pseudo labels for target domain unlabeled data. The method further comprises: Migrating in different working conditions of the same data set; collecting bearing or gear vibration signals under different working conditions; and migrating diagnostic knowledge trained from source domain data to other working conditions when the rotating machinery is running under variable working conditions and the amount of data collected under at least one working condition is less than a data amount threshold; the difference includes load parameters and rotational speed parameters; Migrating between different data sets; collecting bearing data collected in a laboratory; identifying the health status of bearings in an engineering scene based on the diagnostic knowledge of laboratory bearings for bearing fault diagnosis in the engineering scene; and Migrating in mixed fault type and different working conditions; based on the condition that the bearing and gear faults occur simultaneously and the data distribution difference collected under different working conditions is greater than a data distribution threshold, a plurality of domain adaptation methods and / or information of unlabeled data are used to reduce the influence of the data distribution difference by introducing a pre-training model combined with a plurality of domain adaptation modules.

2. The rotating machinery fault diagnosis method based on the joint domain adaptive network according to claim 1, wherein, Further comprising: Combining the alignment methods of joint maximum mean difference domain adaptation and conditional adversarial domain adaptation, the joint distribution of features between source domain data and target domain data is aligned in multiple ways, and the joint distribution of labels between source domain data and target domain data is aligned in multiple ways, so as to capture the multi-modal information contained in the data; And Aligning the two joint distributions globally and aligning different categories to improve the alignment accuracy under different data distribution differences and the alignment effect of similar features between source domain and target domain.

3. The rotating machinery fault diagnosis method based on the joint domain adaptive network according to claim 2, characterized in that, Further comprising: Introducing a semi-supervised pre-training model, introducing target domain data with higher reliability through threshold screening in the semi-supervised training process, and replacing the model trained by source domain data in the pre-adaptation stage to predict target domain data, so as to improve the prediction accuracy.

4. The rotating machinery fault diagnosis method based on the joint domain adaptive network according to claim 3, characterized in that, Further comprising: Constructing two modules for domain adaptation of source domain and target domain; Using the joint distribution difference of input features and output labels, the domain adaptation is performed in the feature extraction and classification layer through joint maximum mean difference domain adaptation; Performing field adversarial training between features and predicted labels to reduce field drift; Based on the uncertainty of the domain classifier prediction, the samples in the domain classifier are reweighted by entropy calculation weight ω to reduce the influence of target instances with uncertain prediction; And Through the two modules, the maximum distinction between categories and the field adaptation under multiple modes are performed.

5. The rotating machinery fault diagnosis method based on the joint domain adaptive network according to claim 4, characterized in that, Training an initial pre-training model by combining consistency regularization loss and pseudo labels comprises: Weak enhancement is performed on labeled data; Weak enhancement and strong enhancement are performed on unlabeled data respectively, and the proportion of noise added by weak enhancement is different from the proportion of noise added by strong enhancement; Pseudo labels are predicted for weak enhanced unlabeled data, and when the model produces a prediction higher than a threshold, the prediction is retained as a pseudo label. The consistency of the strong and weak predictions is measured by cross-entropy loss on the strongly augmented unlabeled data prediction classification probability; and The supervised loss of the labeled data and the consistency regularization loss of the unlabeled data are combined to obtain a pre-training model, so that the prediction error of the target domain data in the domain adaptation stage is reduced.

6. The rotating machinery fault diagnosis method based on the joint domain adaptive network according to claim 5, wherein, The source domain target domain similar feature alignment effect includes: The weakly augmented source domain data and the unlabeled target domain data are respectively sent to the pre-training model obtained by the weakly augmented pre-training stage; The final feature and label are subjected to multi-layer linear transformation to represent the joint distribution of the feature and the label; The joint maximum mean difference domain adaptation method and the conditional domain contrast domain adaptation method are used to align the joint distribution of the feature and the label; and The domain-invariant features of the source domain and the target domain are trained by combining the loss of the label classification joint maximum mean difference domain adaptation method and the loss of the conditional domain contrast domain adaptation method, so as to improve the accuracy of the transfer learning fault diagnosis.

7. The rotating machinery fault diagnosis method based on the joint domain adaptive network according to claim 6, wherein, The pre-training further includes: The pre-training introduces a consistency regularization loss; The consistency regularization loss includes: after the unlabeled data is injected with noise, the classifier outputs the same probability distribution for it, and forces a unlabeled sample to be classified into the same classification as its augmented sample; and By forcing a unlabeled sample to be classified into the same classification as its augmented sample, the information of the unlabeled data is introduced into the pre-training model, so that the accuracy of the pre-training model is higher than that of the model trained directly by the source domain data to predict the target domain data, thereby avoiding the model from not converging in the early stage.

8. The rotating machinery fault diagnosis method based on the joint domain adaptive network according to claim 7, wherein, The two domain adaptation modules combined to align the joint distribution of the feature and the label further include: The two different domain adaptation methods are combined to complementarily align the joint distribution of the two domain data, so as to avoid the case where the distribution difference between the source domain and the target domain data is greater than the data distribution threshold: aligning the distribution between the features, ignoring the class information, and aligning the two feature distributions as a whole, ignoring the relationship between the classes; and The two different domain adaptation methods are combined to complementarily align the joint distribution of the two domain data, so as to avoid the case where a single domain adaptation module is used to align the joint distribution: based on the difference between the distribution information, the complexity is greater than the complexity threshold, and the single domain adaptation module cannot capture the multi-modal information.

9. A rotating machinery fault diagnosis system based on a joint domain adaptive network, characterized in that, It includes: The pre-training model module is configured to train an initial pre-training model by combining the consistency regularization loss and the pseudo label, so as to improve the prediction accuracy of the target domain data in the initial stage of the domain adaptation process; And The semi-supervised training module is configured to improve the utilization rate of fault information in the unlabeled data by generating pseudo labels for the target domain unlabeled data; Wherein: The same data set is migrated under different working conditions; the bearing or gear vibration signals collected under different working conditions have differences, and the rotating machinery is running under variable working conditions, and the amount of data collected under at least one working condition is less than the data amount threshold, the working condition with the data amount greater than the data amount threshold is taken as the source domain data, and the diagnostic knowledge trained from the source domain data is migrated to other working conditions; the differences include load parameters and rotational speed parameters; The pre-training further includes: The pre-training introduces a consistency regularization loss; The consistency regularization loss includes: after the unlabeled data is injected with noise, the classifier outputs the same probability distribution for it, and forces a unlabeled sample to be classified into the same classification as its augmented sample; and By forcing a unlabeled sample to be classified into the same classification as its augmented sample, the information of the unlabeled data is introduced into the pre-training model, so that the accuracy of the pre-training model is higher than that of the model trained directly by the source domain data to predict the target domain data, thereby avoiding the model from not converging in the early stage. The two domain adaptation modules combined to align the joint distribution of the feature and the label further include: The two different domain adaptation methods are combined to complementarily align the joint distribution of the two domain data, so as to avoid the case where the distribution difference between the source domain and the target domain data is greater than the data distribution threshold: aligning the distribution between the features, ignoring the class information, and aligning the two feature distributions as a whole, ignoring the relationship between the classes; and The two different domain adaptation methods are combined to complementarily align the joint distribution of the two domain data, so as to avoid the case where a single domain adaptation module is used to align the joint distribution: based on the difference between the distribution information, the complexity is greater than the complexity threshold, and the single domain adaptation module cannot capture the multi-modal information. It includes: The pre-training model module is configured to train an initial pre-training model by combining the consistency regularization loss and the pseudo label, so as to improve the prediction accuracy of the target domain data in the initial stage of the domain adaptation process; And The semi-supervised training module is configured to improve the utilization rate of fault information in the unlabeled data by generating pseudo labels for the target domain unlabeled data; Wherein: In the same data set, the migration is carried out under different working conditions; the bearing or gear vibration signals collected under different working conditions have differences, and the rotating machinery is running under variable working conditions, and the amount of data collected under at least one working condition is less than the data amount threshold, the working condition with the data amount greater than the data amount threshold is taken as the source domain data, and the diagnostic knowledge trained from the source domain data is migrated to other working conditions; the differences include load parameters and rotational speed parameters; Migrate between different datasets; use bearing data collected by laboratory simulation; identify the health status of bearings in engineering scenarios based on diagnostic knowledge of laboratory bearings for bearing fault diagnosis in engineering scenarios; and Migrate under different working conditions of mixed fault types; under the condition that the simultaneous failure of bearings and gears occurs and the data distribution difference collected under different working conditions is greater than the data distribution threshold, use multiple domain adaptation methods and / or introduce unannotated data information, and reduce the influence of data distribution difference by introducing a pre-trained model combined with multiple domain adaptation modules.

Images