Equipment fault diagnosis method and system based on federated domain generalization

Abstract

The present application discloses a method and system for equipment fault diagnosis based on federated domain generalization, which relates to the field of fault diagnosis technology, wherein the method comprises: obtaining domain data and data to be diagnosed, inputting the domain data into a domain-invariant feature learning module and a domain feature learning module respectively, obtaining domain-invariant features, predicted domains, and predicted faults; adjusting the parameters of the local model to obtain a trained local model; aggregating the parameters of each trained local model to establish a global model; inputting the data to be diagnosed into the global model to obtain the final diagnosis result. The method of the present application introduces a federated hybrid domain generalization mechanism, wherein the data distribution of each source client is assumed to be a mixture of several predefined domains, and the global model is trained by learning domain-invariant features and features of specific domains, and adopting different aggregation strategies in each module, thereby improving the generalization and robustness of the global model.

Description

Technical Field

[0001] The present application relates to the technical field of fault diagnosis, and in particular to a method and system for equipment fault diagnosis based on federal domain generalization. Background Art

[0002] With the rapid development of information technology and industrial operation technology in recent years, the exponential growth of equipment sensor monitoring data has provided a large amount of equipment data, which has promoted the real-time and comprehensive acquisition of equipment operation status. At the same time, machine learning technology represented by deep learning has developed rapidly and has been successfully applied to equipment fault diagnosis (EFD). These research works are usually based on three basic assumptions: (1) there are a large number of labeled samples with rich typical fault information; (2) training and test samples meet the assumption of independent and identical distribution; (3) the possibility of sharing and integrating equipment operation and maintenance data across different factories.

[0003] However, in actual situations, EFD exhibits obvious characteristics: (1) Equipment usually operates for a long time under normal conditions, so it is challenging to obtain sufficient labeled data for specific, particularly rare fault categories. This causes operation and maintenance data to be sparse, unbalanced, and exhibit a long-tail distribution. (2) Industrial equipment usually operates under harsh, variable, and complex conditions, resulting in domain shift, that is, there is a significant distribution difference between training data and test data, which poses a challenge to the assumption of independent and identical distribution. (3) Equipment maintenance data often contains sensitive information such as enterprise production process details and enterprise core process parameters. Due to potential conflicts of interest, local equipment data movement may lead to serious privacy violations, causing industrial users to refuse to share their data.

[0004] Federal Domain Adaptation (FDA) provides a feasible solution to the above problems of EFD by integrating deep learning in feature representation, domain adaptation in knowledge transfer, and the advantages of federated learning in privacy protection. This approach has received widespread attention and made initial progress in EFD. However, these techniques usually assume that samples from the target domain can be used for model training. However, in actual application scenarios, it is extremely challenging to obtain complete data from the target domain, and even the working conditions of the target domain equipment may be unknown.

[0005] Federal Domain Generalization (FDG) further develops a highly practical and general federated learning framework in which data from all source clients are prohibited from communicating with any other client, and data from the target client is not available during the model training phase. This approach is critical to EFD because it relaxes the strict assumptions about the availability of samples from the target client and ensures the security of private data across different clients. Although a few researchers have proposed FDG methods for EFD, existing studies usually assume that each source client has only one domain and that the domains across clients do not overlap, which is inconsistent with many practical situations and limits the application of the technology. Summary of the invention

[0006] The embodiments of the present application provide a method and system for equipment fault diagnosis based on federal domain generalization, so as to solve the problem that the situations targeted by the fault diagnosis methods proposed in the prior art are inconsistent with many actual situations.

[0007] On the one hand, an embodiment of the present application provides an equipment fault diagnosis method based on federation domain generalization, including:

[0008] Obtain domain data of multiple source clients, each source client including multiple domain data;

[0009] The domain data of each source client is input into the domain invariant feature learning module. The domain invariant feature learning module includes a fault feature extractor, a generator and a discriminator. The fault feature extractor extracts the actual fault feature of each domain data. The generator generates the fault feature based on random noise. The discriminator obtains the domain invariant feature according to the actual fault feature and the generated fault feature.

[0010] The domain data and actual fault features of each source client are input into the domain feature learning module. The domain feature learning module includes a domain feature extractor, a domain classifier, and multiple fault classifiers. The domain feature extractor extracts the domain features of each domain data. The domain classifier classifies the domain features to obtain the predicted domain of the domain features. The fault classifier corresponding to the actual fault features among the multiple fault classifiers classifies the actual fault features to obtain the predicted faults.

[0011] Adjust the parameters of the local model based on the actual fault features, generated fault features, domain invariant features, predicted domains, and the differences between the predicted faults and the corresponding labels in the domain data to obtain a trained local model;

[0012] Input the parameters of each trained local model into a weighted aggregation module to obtain global parameters, and obtain a global model based on the global parameters. The global model includes a global domain feature extractor, a global domain classifier, a global fault feature extractor, and a global fault classifier.

[0013] Obtain the target client's data to be diagnosed;

[0014] The data to be diagnosed is input into the global model, the global domain feature extractor extracts the domain features to be diagnosed of the data to be diagnosed, the global domain classifier determines the similarity between the domain features to be diagnosed and the domain features of the source client based on the domain features to be diagnosed, the global fault feature extractor extracts the fault features to be diagnosed of the data to be diagnosed, the global fault classifier first performs a preliminary classification of the fault features to be diagnosed to obtain a preliminary diagnosis result, and then combines the preliminary diagnosis result with the similarity to obtain the final diagnosis result.

[0015] On the other hand, an embodiment of the present application further provides an equipment fault diagnosis system, including:

[0016] A domain data acquisition module is used to acquire domain data of multiple source clients, each source client contains multiple domain data;

[0017] A domain-invariant feature learning module includes a fault feature extractor, a generator, and a discriminator. The fault feature extractor is used to extract actual fault features of each domain data, the generator is used to generate fault features based on random noise, and the discriminator is used to obtain domain-invariant features based on actual fault features and generated fault features.

[0018] A domain feature learning module includes a domain feature extractor, a domain classifier, and multiple fault classifiers. The domain feature extractor is used to extract domain features of each domain data. The domain classifier is used to classify the domain features to obtain the predicted domain of the domain features. The fault classifier corresponding to the actual fault feature among the multiple fault classifiers is used to classify the actual fault feature to obtain the predicted fault.

[0019] A parameter adjustment module, used to adjust the parameters of the local model based on the actual fault features, the generated fault features, the domain invariant features, the predicted domain, and the difference between the predicted fault and the corresponding labels in the domain data, to obtain a trained local model;

[0020] A weighted aggregation module is used to aggregate the parameters of each trained local model to obtain global parameters, and obtain a global model based on the global parameters. The global model includes a global domain feature extractor, a global domain classifier, a global fault feature extractor, and a global fault classifier;

[0021] A module for obtaining data to be diagnosed, used to obtain the data to be diagnosed of the target client;

[0022] The fault diagnosis module is used to input the data to be diagnosed into the global model. The global domain feature extractor extracts the domain features to be diagnosed of the data to be diagnosed. The global domain classifier determines the similarity between the domain features to be diagnosed and the domain features of the source client based on the domain features to be diagnosed. The global fault feature extractor extracts the fault features to be diagnosed of the data to be diagnosed. The global fault classifier first performs a preliminary classification on the fault features to be diagnosed to obtain a preliminary diagnosis result, and then combines the preliminary diagnosis result with the similarity to obtain the final diagnosis result.

[0023] The equipment fault diagnosis method and system based on federated domain generalization in this application have the following advantages:

[0024] A federated mixed-domain generalization mechanism is introduced, in which the data distribution of each source client is assumed to be a mixture of several predefined domains. The global model is trained by learning domain-invariant features as well as domain-specific features, and different aggregation strategies are adopted in each module to improve the generalization and robustness of the global model. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0026] Figure 1 An architectural diagram of an equipment fault diagnosis method based on federal domain generalization provided in an embodiment of the present application.

[0027] Figure 2 This is the effect of deleting the domain feature learning module and the weighted aggregation module in the ablation experiment. DETAILED DESCRIPTION

[0028] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0029] Figure 1 The present invention provides an equipment fault diagnosis method based on federated domain generalization, which includes a collaborative training phase and a model generalization phase, wherein the collaborative training phase includes:

[0030] S100, obtaining domain data of multiple source clients, each source client including multiple domain data.

[0031] Exemplarily, the source client of the present application has N Each source client has M domain data, each domain data comes from a domain.

[0032] The above-mentioned field data may be data on certain aspects of the equipment, such as vibration, temperature, etc.

[0033] S110, input the domain data of each source client into the domain invariant feature learning module. The domain invariant feature learning module includes a fault feature extractor, a generator and a discriminator. The fault feature extractor extracts the actual fault features of each domain data, the generator generates the fault features based on random noise, and the discriminator obtains the domain invariant features according to the actual fault features and the generated fault features.

[0034] Exemplarily, the purpose of the domain invariant feature learning module is to perform adversarial training between the true distribution and the reference distribution of domain data of different source clients, thereby eliminating the distribution differences between different source clients and learning generalized features between different source clients.

[0035] S120, the domain data and actual fault features of each source client are input into the domain feature learning module. The domain feature learning module includes a domain feature extractor, a domain classifier and multiple fault classifiers. The domain feature extractor extracts the domain features of each domain data. The domain classifier classifies the domain features to obtain the predicted domain of the domain features. The fault classifier corresponding to the actual fault features among the multiple fault classifiers classifies the actual fault features to obtain the predicted faults.

[0036] Exemplarily, the purpose of the domain feature learning module is to learn domain-specific features of domain data in each domain to facilitate the evaluation of the similarity between unseen domain samples and known domain samples in the model generalization stage.

[0037] S130, adjusting the parameters of the local model based on the actual fault features, the generated fault features, the domain invariant features, the predicted domain, and the difference between the predicted fault and the corresponding labels in the domain data to obtain a trained local model.

[0038] Exemplarily, the difference between the actual fault feature and the corresponding label in the domain data is the actual fault loss, the difference between the generated fault feature and the corresponding label in the domain data is the generated fault loss, the difference between the domain invariant feature and the corresponding label in the domain data is the domain invariant feature loss, the difference between the predicted domain and the corresponding label in the domain data is the domain loss, and the difference between the predicted fault and the corresponding label in the domain data is the fault loss. Based on the actual fault loss, the generated fault loss, the domain invariant feature loss, the domain loss and the fault loss, the final loss corresponding to the source client is determined, and the parameters of the corresponding local model are adjusted according to the final loss.

[0039] Specifically, the actual fault loss, generated fault loss, and domain invariant feature loss are collectively referred to as adversarial loss. In the adversarial learning process, the actual fault features are considered as negative samples, while the generated fault features are considered as positive samples. The domain invariant feature loss can be expressed as:

[0040]

[0041] The actual failure loss can be expressed as:

[0042]

[0043] The generation fault loss can be expressed as:

[0044]

[0045] in, , and Respectively represent n domain-invariant feature loss, actual fault loss, and generated fault loss for each source client. represents the mathematical expectation, Indicates n Domain data of the source client, express The characteristic distribution of express The distribution probability of express Follow the distribution probability , represents the operation of the discriminator to extract domain invariant features from the actual fault features and the generated fault features, Expressed as n The random noise samples assigned by the source client are used to generate fault features derived from the corresponding random noise samples. represents the feature distribution of generated fault features, express The distribution probability of express Follow the distribution probability , c Indicates related, and express and with the same actual fault label y Related.

[0046] The failure loss can be expressed as:

[0047]

[0048] in, Indicates n The failure loss of domain data of a source client, Indicates n The source client m The fault loss of domain data is expressed by the cross entropy loss function as follows:

[0049]

[0050] in, Indicates n Source Client m The number of samples of domain data, represents the cross entropy loss function, Indicates n Source Client No. m The actual fault labels of the domain data, Indicates n The source client m Field data j samples, Indicates that the fault feature extractor is The actual fault features extracted from express and The difference between.

[0051] The domain loss is expressed by the cross entropy loss function as:

[0052]

[0053] in, m is the label of the domain data, The representation domain feature extractor is The domain features extracted from Represents the domain classifier pair The predicted field obtained after classification.

[0054] Furthermore, the final loss is a linear combination of the actual fault loss, generated fault loss, domain invariant feature loss, domain loss and fault loss.

[0055] Specifically, no. n The final loss of domain data of the source client It is expressed as:

[0056]

[0057] in, l 1 and l 2 is an adjustable weight hyperparameter used to balance the contributions of the domain-invariant feature learning module and the domain feature learning module.

[0058] S140, inputting the parameters of each trained local model into a weighted aggregation module to obtain global parameters, and obtaining a global model based on the global parameters. The global model includes a global domain feature extractor, a global domain classifier, a global fault feature extractor and a global fault classifier.

[0059] Exemplarily, the weighted aggregation module aggregates the parameters of the fault feature extractors in all local models based on a variance minimization strategy, and the weighted aggregation module also aggregates the parameters of the domain classifiers and fault classifiers in all local models based on a sample size strategy.

[0060] Specifically, in order to accelerate model training and enhance robustness, inspired by the FDG method, this application adopts an aggregation mechanism based on minimizing variance differences. However, the FDG method requires two rounds of communication between the server and the source client to update the global model once, which greatly increases the communication cost. Therefore, this application proposes a new update mechanism based on minimizing the variance difference, which only requires one communication for each update. The basic idea of ​​this mechanism is to determine the maximum average difference between the actual fault features and the generated fault features corresponding to each source client, and adjust the weights of the parameters in the corresponding local model according to the maximum average difference. The larger the maximum average difference, the poorer the generalization ability of the global model on the client, so the greater the weight of the parameter in the local model. By increasing the weight of the local model of the source client with large differences, the generalization ability of the global model is enhanced and the variance between multiple source clients is reduced. Although this adjustment may cause the difference of other source clients to increase, for local models with good performance, the reduction in weight has little effect on the difference. In the collaborative training process, adjusting the weight based on the maximum average difference not only better maintains data privacy, but also prevents the global model from being biased towards source clients with a large sample size.

[0061] Furthermore, in order to prevent the parameter aggregation of the domain invariant feature learning module and the domain feature learning module, which will lead to excessive differences in weights belonging to the same source client, parameter aggregation is performed in sequence according to the order of domain data in the same source client, and the parameter weights in each aggregation are obtained by referring to the weight parameters obtained in the previous aggregation.

[0062] Specifically, when determining the weight of the parameters in the local model, the parameters are aggregated in turn according to each domain data in the source client. In each parameter aggregation process, the maximum average difference of the current domain data is first determined, and then the weight of the parameter after this aggregation is determined based on the weight obtained in the previous aggregation and the current maximum average difference. The weight of the aggregated parameter is expressed as:

[0063]

[0064]

[0065] in, Indicates n Field data of the source client r The parameter weight of the sub-aggregation, and Yes The result after normalization is It is for n Field data of the source client r -1 aggregation parameter weight The result after normalization is Indicates n The maximum average difference between the actual fault signature and the generated fault signature in the source client, max means taking the maximum value, m express N The average of the maximum average differences among the source clients, Indicates r The parameter that controls the weight change during the sub-aggregation is expressed as:

[0066]

[0067] in, r is the number of aggregation times, Indicates r The number of source clients in the subaggregation, d It is used to control the size of the weight change at each aggregation, and its value range is (0,1).

[0068] After adjusting the weights, the parameters of the corresponding global model can be expressed as:

[0069]

[0070]

[0071] in, represents the parameters of the fault feature extractor in the global model, Indicates n The parameters of the fault feature extractor in the local model, represents the parameters of the generator in the global model, Indicates n The parameters of the generator in each local model.

[0072] The weights of the parameters in the domain feature learning module are as follows:

[0073]

[0074]

[0075] in, Indicates n Source Client m The parameter weights of domain data, Indicates n The parameter weight of each source client, Indicates n The number of samples in a source client.

[0076] Based on the above parameter weights, the parameters of the corresponding global model can be expressed as:

[0077]

[0078]

[0079]

[0080] in, Indicates the global model m The parameters of the fault classifier, Indicates n In the local model m The parameters of the fault classifier, represents the parameters of the domain feature extractor in the global model, Indicates n The parameters of the domain feature extractor in the local model, represents the parameters of the domain classifier in the global model, Indicates n Parameters of the local model domain classifier.

[0081] The model generalization phase includes:

[0082] S150, obtaining the to-be-diagnosed data of the target client.

[0083] Exemplarily, the target client may be a new factory. After a global model is established with the help of domain data of equipment in an existing factory, the global model may be used to diagnose faults of equipment in the new factory.

[0084] S160, input the data to be diagnosed into the global model, the global domain feature extractor extracts the domain features to be diagnosed of the data to be diagnosed, the global domain classifier determines the similarity between the domain features to be diagnosed and the domain features of the source client according to the domain features to be diagnosed, the global fault feature extractor extracts the fault features to be diagnosed of the data to be diagnosed, the global fault classifier first performs a preliminary classification of the fault features to be diagnosed to obtain a preliminary diagnosis result, and then combines the preliminary diagnosis result with the similarity to obtain a final diagnosis result.

[0085] For example, the similarity can be expressed as:

[0086]

[0087] in, s Represents similarity, which is multiple similarities A collection of Represents the characteristics of the domain to be diagnosed and m The similarity of the domains, x Indicates data to be diagnosed. Represents the global domain feature extractor from x The operation of extracting the features of the domain to be diagnosed, Represents the global domain classifier pair Perform field classification operations.

[0088] After classifying the faults using the global fault classifier, the preliminary classification is expressed as:

[0089]

[0090] in, Indicates m The preliminary classification obtained by the global fault classifier is Represents the global fault feature extractor from x Extract the fault features to be diagnosed from Indicates m A global fault classifier pair Perform fault classification operations.

[0091] The final diagnosis result can be expressed as:

[0092]

[0093] in, lIndicates the final diagnosis result.

[0094] The present application also provides an equipment fault diagnosis system, which includes:

[0095] A domain data acquisition module is used to acquire domain data of multiple source clients, each source client contains multiple domain data;

[0096] A domain-invariant feature learning module includes a fault feature extractor, a generator, and a discriminator. The fault feature extractor is used to extract actual fault features of each domain data, the generator is used to generate fault features based on random noise, and the discriminator is used to obtain domain-invariant features based on actual fault features and generated fault features.

[0097] A domain feature learning module includes a domain feature extractor, a domain classifier, and multiple fault classifiers. The domain feature extractor is used to extract domain features of each domain data. The domain classifier is used to classify the domain features to obtain the predicted domain of the domain features. The fault classifier corresponding to the actual fault feature among the multiple fault classifiers is used to classify the actual fault feature to obtain the predicted fault.

[0098] A parameter adjustment module, used to adjust the parameters of the local model based on the actual fault features, the generated fault features, the domain invariant features, the predicted domain, and the difference between the predicted fault and the corresponding labels in the domain data, to obtain a trained local model;

[0099] A weighted aggregation module is used to aggregate the parameters of each trained local model to obtain global parameters, and obtain a global model based on the global parameters. The global model includes a global domain feature extractor, a global domain classifier, a global fault feature extractor, and a global fault classifier;

[0100] A module for obtaining data to be diagnosed, used to obtain the data to be diagnosed of the target client;

[0101] The fault diagnosis module is used to input the data to be diagnosed into the global model. The global domain feature extractor extracts the domain features to be diagnosed of the data to be diagnosed. The global domain classifier determines the similarity between the domain features to be diagnosed and the domain features of the source client based on the domain features to be diagnosed. The global fault feature extractor extracts the fault features to be diagnosed of the data to be diagnosed. The global fault classifier first performs a preliminary classification on the fault features to be diagnosed to obtain a preliminary diagnosis result, and then combines the preliminary diagnosis result with the similarity to obtain the final diagnosis result.

[0102] Experimental proof

[0103] The four task categories of the hybrid domain included in the source client and the target client are illustrated in Table 1.

[0104] Table 1 Task categories of the mixed domains included in the source client and the target client

[0105]

[0106] Table 2 shows the domain distribution α And label distribution β Each source client has a domain distribution, which is drawn from the Dirichlet distribution. α >0 as a concentration parameter, governing the consistency of domain distribution among source clients. α →0, then each source client has data from only a single randomly selected domain. In contrast, if α →∞, all source clients show the same domain distribution consistent with the prior distribution. Similarly, when β →0, the heterogeneity of label distribution reaches its maximum value. β →∞, the label distribution tends to be uniform.

[0107] Table 2 Domain-based distribution α And label distribution β Specific tasks in each task category

[0108]

[0109] The diagnostic results of Task T1 on the CWRU (Case Western Reserve University) and KAT (Konstruktions and Antriebstechnik) datasets can be seen in Tables 3 and 4. The total sample size of each source client is set to 2000. The initial sample lengths of the vibration signals of CWRU and KAT are 512 and 1024, respectively. Subsequently, 256 and 512 frequency coefficients derived by fast Fourier transform are used as inputs to the model. The proposed method consistently outperforms all the compared methods.

[0110] Table 3 Diagnosis results of task T1 on the CWRU dataset

[0111]

[0112] Table 4 Diagnosis results of task T1 on the KAT dataset

[0113]

[0114] At CWRU, this application approach has three distinct advantages:

[0115] (1) The average precision is improved from 10.43% to 47.73%. This enhancement is attributed to the effectiveness of the domain feature learning module and domain invariant feature learning module in this application, which learn domain-specific knowledge and domain-invariant knowledge in a mixed-domain setting, respectively.

[0116] (2) Compared with other methods, the standard deviation of the method proposed in this application is smaller, which further demonstrates that the method is stable and robust.

[0117] (3) Taking S1 as an example, as β As it decreases from ∞ to 1, the accuracy drops significantly due to the transition from uniform to non-uniform label distribution. β When it drops from 1 to 0.5 and then to 0.1, the trend reverses and shows an increase. This is because, in a highly heterogeneous setting, each domain contains only one label that is consistent across all domains. In addition, the domain on each source client randomly selects samples with this label, resulting in instances of the same sample. β is fixed, and α When decreasing from ∞ to 0.1, the accuracy of the comparison methods such as FedAvg tends to decrease. The proposed method does not show a clear pattern, which can be attributed to the domain-aware multi-head classifier, making the proposed method less sensitive to the changes in the mixed domain distribution.

[0118] In the quantitative diagnosis results of multiple methods in 16 tasks of task T1 on the KAT dataset, the results obtained are basically consistent with those obtained on the CWRU dataset. First, the average accuracy is 7.62% higher than the best performing comparison algorithm. Second, the standard deviation of the algorithm is low, indicating high stability. Finally, the best comparison algorithm results are achieved under homogeneous label distribution. In all other heterogeneous scenarios, the performance of the algorithm is better than the comparison algorithm.

[0119] Ablation experiment

[0120] The ablation experiment analyzes the contributions of the Domain Specific Feature Learning (DSFL) module and the Weighted Aggregation (WA) module. Figure 2Results of the S4 task on the CWRU and KAT datasets after removing the domain feature learning module and the weighted aggregation module are shown in Figure 2. From the results, we observe that the domain feature learning module is essential for both datasets as removing it significantly reduces the accuracy. The weighted aggregation module proves to be important for the CWRU dataset, but its impact on the KAT dataset is limited. This is mainly due to the higher inter-domain similarity between multi-source clients in KAT compared to CWRU, which makes specialized aggregation methods less necessary.

[0121] Hyperparameter Analysis

[0122] Hyperparameters are discussed here l 1. l 2. l 0 and d The impact of different settings on the performance of the model in this application. Taking tasks S1-1 and S1-0 on the CWRU dataset as an example, l 1The parameter values ​​considered are {0.95, 0.9, 0.85, 0.8, 0.75, 0.7, 0.65}; l 0, the value is {1×10 −4 , 1×10 −3 , 1×10 −2 , 1×10 −1};for d , these values ​​are {0.2, 0.1, 0.05, 0.01, 0}. l The optimal value of 1 is 0.85, l The optimal value corresponding to 2 is 0.15. l The optimal value of 0 is 1×10 −4 In different settings, d Considering the overall performance of various tasks, this application will d Set to 0.1.

[0123] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0124] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for equipment fault diagnosis based on federated domain generalization, characterized in that: include: Acquire domain data of multiple source clients, each of the source clients comprising multiple domain data; Input the domain data of each source client into a domain invariant feature learning module, the domain invariant feature learning module includes a fault feature extractor, a generator and a discriminator, the fault feature extractor extracts the actual fault feature of each domain data, the generator generates a fault feature based on random noise, and the discriminator obtains a domain invariant feature according to the actual fault feature and the generated fault feature; Input the domain data and the actual fault features of each source client into a domain feature learning module, wherein the domain feature learning module includes a domain feature extractor, a domain classifier, and multiple fault classifiers, wherein the domain feature extractor extracts domain features of each domain data, the domain classifier classifies the domain features to obtain a predicted domain of the domain features, and the fault classifier corresponding to the actual fault features among the multiple fault classifiers classifies the actual fault features to obtain a predicted fault; The parameters of the local model are adjusted based on the actual fault feature, the generated fault feature, the domain invariant feature, the predicted domain, and the difference between the predicted fault and the corresponding label in the domain data to obtain a trained local model; wherein the difference between the actual fault feature and the corresponding label in the domain data is the actual fault loss, the difference between the generated fault feature and the corresponding label in the domain data is the generated fault loss, the difference between the domain invariant feature and the corresponding label in the domain data is the domain invariant feature loss, the difference between the predicted domain and the corresponding label in the domain data is the domain loss, and the difference between the predicted fault and the corresponding label in the domain data is the fault loss; based on the actual fault loss, the generated fault loss, the domain invariant feature loss, the domain loss, and the fault loss, a final loss corresponding to the source client is determined, and the parameters of the corresponding local model are adjusted according to the final loss; Inputting the parameters of each trained local model into a weighted aggregation module to obtain global parameters, and obtaining a global model based on the global parameters, wherein the global model includes a global domain feature extractor, a global domain classifier, a global fault feature extractor, and a global fault classifier; Obtain the target client's data to be diagnosed; The data to be diagnosed is input into the global model, the global domain feature extractor extracts the domain features to be diagnosed of the data to be diagnosed, the global domain classifier determines the similarity between the domain features to be diagnosed and the domain features of the source client according to the domain features to be diagnosed, the global fault feature extractor extracts the fault features to be diagnosed of the data to be diagnosed, the global fault classifier first performs a preliminary classification on the fault features to be diagnosed to obtain a preliminary diagnosis result, and then combines the preliminary diagnosis result with the similarity to obtain a final diagnosis result.

2. According to the equipment fault diagnosis method based on federated domain generalization according to claim 1, it is characterized in that: The final loss is a linear combination of the actual fault loss, the generated fault loss, the domain invariant feature loss, the domain loss, and the fault loss.

3. According to claim 1, a method for equipment fault diagnosis based on federated domain generalization is characterized in that: The weighted aggregation module aggregates the parameters of the fault feature extractors in all the local models based on a variance minimization strategy, and the weighted aggregation module also aggregates the parameters of the domain classifiers and the fault classifiers in all the local models based on a sample size strategy.

4. According to the equipment fault diagnosis method based on federated domain generalization according to claim 3, it is characterized in that: When the weighted aggregation module aggregates the parameters of the fault feature extractors in all the local models based on the variance minimization strategy, it determines the maximum average difference between the actual fault feature and the generated fault feature corresponding to each source client, and adjusts the weights of the corresponding parameters in the local model according to the maximum average difference.

5. According to claim 4, a method for equipment fault diagnosis based on federated domain generalization is characterized in that: The larger the maximum average difference is, the larger the weight of the corresponding parameter in the local model is.

6. The equipment fault diagnosis method based on federated domain generalization according to claim 4 is characterized in that: When determining the weights of the parameters in the local model, parameter aggregation is performed in sequence according to each of the domain data in the source client. In each parameter aggregation process, the maximum average difference of the current domain data is first determined, and then the weights of the parameters after this aggregation are determined based on the weights obtained in the previous aggregation and the current maximum average difference.

7. A system for applying the equipment fault diagnosis method based on federated domain generalization as described in any one of claims 1 to 6, characterized in that: include: A domain data acquisition module, used to acquire domain data of multiple source clients, each of which contains multiple domain data; A domain invariant feature learning module, comprising a fault feature extractor, a generator and a discriminator, wherein the fault feature extractor is used to extract actual fault features of each of the domain data, the generator is used to generate fault features based on random noise, and the discriminator is used to obtain domain invariant features according to the actual fault features and the generated fault features; A domain feature learning module, comprising a domain feature extractor, a domain classifier, and a plurality of fault classifiers, wherein the domain feature extractor is used to extract domain features of each of the domain data, the domain classifier is used to classify the domain features to obtain a predicted domain of the domain features, and the fault classifier corresponding to the actual fault feature among the plurality of fault classifiers is used to classify the actual fault feature to obtain a predicted fault; A parameter adjustment module, configured to adjust the parameters of the local model based on the actual fault feature, the generated fault feature, the domain invariant feature, the predicted domain, and the difference between the predicted fault and the corresponding label in the domain data to obtain a trained local model; wherein the difference between the actual fault feature and the corresponding label in the domain data is the actual fault loss, the difference between the generated fault feature and the corresponding label in the domain data is the generated fault loss, the difference between the domain invariant feature and the corresponding label in the domain data is the domain invariant feature loss, the difference between the predicted domain and the corresponding label in the domain data is the domain loss, and the difference between the predicted fault and the corresponding label in the domain data is the fault loss; a final loss corresponding to the source client is determined based on the actual fault loss, the generated fault loss, the domain invariant feature loss, the domain loss, and the fault loss, and the parameters of the corresponding local model are adjusted according to the final loss; A weighted aggregation module, used to aggregate the parameters of each trained local model to obtain global parameters, and obtain a global model based on the global parameters, wherein the global model includes a global domain feature extractor, a global domain classifier, a global fault feature extractor, and a global fault classifier; A module for obtaining data to be diagnosed, used to obtain the data to be diagnosed of the target client; A fault diagnosis module is used to input the data to be diagnosed into the global model, the global domain feature extractor extracts the domain features to be diagnosed of the data to be diagnosed, the global domain classifier determines the similarity between the domain features to be diagnosed and the domain features of the source client according to the domain features to be diagnosed, the global fault feature extractor extracts the fault features to be diagnosed of the data to be diagnosed, the global fault classifier first performs a preliminary classification on the fault features to be diagnosed to obtain a preliminary diagnosis result, and then combines the preliminary diagnosis result with the similarity to obtain a final diagnosis result.

Images