Intelligent Fault Migration Diagnosis Method Based on Generalized Network of Heterogeneous Federation Domain

Abstract

The present invention relates to an intelligent fault migration diagnosis method based on a heterogeneous federated domain generalized network, and belongs to the field of mechanical fault migration diagnosis. The method comprises: constructing a DDA model; constructing an HFDGN based on the MIWM mechanism; S4: inputting the divided training samples in all auxiliary domains of the source client and the source domain of the target client into the constructed DDA model, and using the optimization objective function in the DDA model to train the corresponding local client model; S5: uploading the parameters of the trained DDA model to a central server, and then using the constructed HFDGN network to perform federated migration faults; S6: after multiple iterations of training, the error curve tends to be stable, the HFDGN network training is completed, and the trained HFDGN network is used for heterogeneous multi-source federated migration diagnosis. The present invention can meet the application requirements of high data utilization and real-time diagnosis in actual engineering.

Description

Technical Field

[0001] The invention belongs to the field of mechanical fault migration diagnosis, and relates to an intelligent fault migration diagnosis method based on a heterogeneous federated domain generalized network. Background Art

[0002] Among the existing mechanical fault migration diagnosis technologies, there are some federated transfer learning intelligent diagnosis methods to solve the distribution differences and data privacy protection issues between the source domain and the target domain. However, their performance depends entirely on the source client and the target client corresponding to the mechanical equipment being isomorphic, that is, the mechanical data of the source client and the target client should come from the same mechanical parts. At the same time, these methods also require that the test target domain data samples in the target client are available during the training process. Considering that in actual engineering, the target domain data is usually invisible and heterogeneous with the source client data, the existing federated migration diagnosis methods are difficult to meet the application requirements of high data utilization and real-time diagnosis.

[0003] In order to solve the above problems, a novel heterogeneous federated domain generalization network (HFDGN) is urgently needed to fill the gap in heterogeneous multi-source federated diagnosis methods. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide an intelligent fault migration diagnosis method based on a heterogeneous federated domain generalized network. In the HFDGN network, a heterogeneous transfer learning framework (HFTL) is adopted, which can realize the generalized fault diagnosis of the target client through the public knowledge representation mapping of the heterogeneous source clients. In addition, a disentangled domain adaptation basic model (DDA) is used to remove the negative impact of noise. At the same time, the basic model can enhance the ability of domain confusion and extract features related to intrinsic faults.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] An intelligent fault migration diagnosis method based on a heterogeneous federated domain generalized network specifically includes the following steps:

[0007] S1: The original vibration signal on the mechanical equipment is collected through the acceleration sensor to facilitate the construction of subsequent migration diagnosis tasks; and the collected original vibration signal is sampled using the sliding window sampling technology;

[0008] S2: Construct a decoupled domain adaptation basic model, referred to as the DDA model;

[0009] S3: Build a heterogeneous federated transfer learning network based on the mutual information weight matching mechanism, where the mutual information weight matching mechanism is referred to as MIWM; the heterogeneous federated transfer learning network is referred to as HFDGN, which adopts a heterogeneous transfer learning framework;

[0010] S4: Input the divided training samples in all auxiliary domains of the source client and the source domain of the target client into the constructed DDA model, and use the optimization objective function in the DDA model to train the corresponding local client model;

[0011] S5: Upload the parameters of the trained DDA model to the central server, and then use the constructed HFDGN network to perform federated migration fault diagnosis;

[0012] S6: After multiple iterations of training, the error curve tends to be stable, and the HFDGN network training is completed. The trained HFDGN network is used for heterogeneous multi-source federated migration diagnosis.

[0013] Furthermore, in step S2, the backbone network of the DDA model consists of four parts: feature extractor G FE (θ FE ), decoupler G D (θ D ), Reconstructor G R (θ R ) and the fault classifier G FC (θ FC ), where θ FE ,θ D ,θ R and θ FC They represent the trainable weights of the corresponding network models respectively; the feature extractor is used to mine the knowledge of distribution differences and reduce the distribution differences, and then obtain the universal feature F G ; The fault classifier is dedicated to extracting label features F L To identify the fault type; the decoupler is used to decouple and separate the fault-related features F FR F FI ; The reconstructor reconstructs common features from fault-related features and fault-irrelevant features, so the decoupler and reconstructor can be regarded as the encoder and decoder structures in the autoencoder.

[0014] The optimization objectives of the DDA model mainly include the following three parts: 1) Decoupled representation: learning a decoupled representation to separate the fault-irrelevant features caused by noise from the general features; 2) Separation and reconstruction: maximizing the distribution distance between fault-related features and fault-irrelevant features so as to ensure the independence of the extracted fault-related features; 3) Distribution adaptation: minimizing the distribution difference between any two fault-related features.

[0015] Furthermore, in step S2, in the optimization goal 1) of the DDA model, the decoupled representation is specifically: using the decoupler and the fault classifier to remove the irrelevant features of the fault caused by noise in an adversarial training manner; first, using the cross entropy loss L C Train feature extractors, decouplers, and fault classifiers to accurately identify fault types;

[0016]

[0017] in, represents the expectation of the sample domain; X i ,Y i Represent the i-th data sample and the corresponding label respectively; represents the sample domain, C represents the fault type; I() represents the indicator function, when argmax(Y i ) = c, I = 0;

[0018] Then, the weight parameters of the feature extractor and the fault classifier are fixed, and the decoupler is trained to deceive the fault classifier by maximizing the information entropy loss, thereby learning the fault-irrelevant features. The information entropy loss is a function that reflects the purity of the sample prediction label; the smaller the information entropy loss, the higher the purity of the predicted label probability. Information entropy (IE) loss L IE It is expressed as follows:

[0019]

[0020] in, It means to seek expectation on the sample domain; Represents sample X i The predicted label vector of represents the cth element in the predicted label vector, which is obtained by the following formula:

[0021]

[0022] Through adversarial training between formulas (1) and (2), the final fault-irrelevant features and fault-related features can be separated and obtained by the decoupler.

[0023] Furthermore, in step S2, the optimization goal 2) of the DDA model is to separate and reconstruct by maximizing the average difference loss L D To train the decoupler to expand the distribution difference between fault-related features and fault-irrelevant features:

[0024]

[0025] in, They represent the expectation of fault-irrelevant features and fault-related features respectively; φ(·) is the high-dimensional mapping function in the regenerated Hilbert space; represents the 2-norm of the regenerated Hilbert space;

[0026] Fault related features F FR F FI Obtained by the following formula:

[0027] {F FR ,F FI =G D (G FE (X)) (5)

[0028] At the same time, in order to avoid the over-training of formula (5) destroying the intrinsic properties of the extracted fault-related features, the reconstruction loss L R To train the reconstructor to reconstruct common features from fault-related features and fault-irrelevant features;

[0029]

[0030] in, It means to find the expectation of the sample domain.

[0031] Furthermore, in step S2, in the optimization goal 3) of the DDA model, distribution adaptation is specifically: using the minimization of the average difference index as the domain confusion loss L DC To train the feature extractor and decoupler to achieve distribution adaptation;

[0032]

[0033] in, represents the expectation of irrelevant fault features, φ(·) is the high-dimensional mapping function in the regenerated Hilbert space; represents the 2-norm of the regenerated Hilbert space;

[0034] Assume that the client has Q m There are K auxiliary domains, and there will be K domain confusion losses. in According to the transfer learning theory, we can know represents the distribution similarity of the corresponding two auxiliary domains, that is, The smaller the value, the more the feature extractor will focus on the corresponding Therefore, in order to enable the feature extractor to learn more general domain-invariant features, the purity loss L constructed using information entropy is used. P To train the feature extractor and decoupler to reduce the purity between all domain confusion losses, that is, to maximize the purity loss L P ;

[0035]

[0036] Furthermore, in step S2, the data of multiple auxiliary domains of the source client will fully participate in the three optimization objectives of the DDA model; because the target client assumes that the target domain data is inaccessible during the training process and can only be used for testing, the target client only has source domain data participating in the two optimization objectives of "decoupled representation" and "separation and reconstruction".

[0037] Further, in step S3, the MIWM mechanism is used to evaluate the contribution of each source client to the target client during the weight parameter allocation process;

[0038] The contribution of the i-th source client to the target client is MI i The expression is:

[0039]

[0040] in, and represents the fault classifier and decoupler in the DDA model trained for the i-th source client; F G and F L They represent the common features and label features obtained by training the DDA model with source domain data in the target client; I Θ Represents two data sample domains The mutual information between them can be expressed as:

[0041]

[0042] Among them, T(θ) represents the function set formed by the network parameters; and Respectively represent and The marginal probability distribution of is the joint probability distribution between the two domains; represents the product of the marginal probability distribution between two domains; Θ represents the weight parameter set, θ represents the weight parameter,

[0043] It means to find the expectation of the joint probability of two domains. It represents the expectation of the product of the marginal probabilities of two domains.

[0044] Further, in step S5, the central server aggregates the weight parameters of the feature extractors from all source clients, and the calculation formula is:

[0045]

[0046] Where M represents the number of source clients; represents the weight parameter of the feature extractor from the i-th source client; Represents the weight parameter after aggregation; w i The weight parameters for matching the corresponding source client and target client are obtained through the MIWM mechanism;

[0047]

[0048] Where τ represents the temperature coefficient;

[0049] The central server then distributes the aggregated weight parameters to each source and destination client.

[0050] Furthermore, in step S5, the DDA model parameters uploaded to the central server include: for all source clients, the weight parameters of the feature extractor, decoupler and fault classifier in the DDA model need to be uploaded; for the target client, the general feature F extracted from the source domain data in the DDA model needs to be uploaded. G and label feature F L .

[0051] The beneficial effect of the present invention is that the present invention can solve the application requirements of high data utilization and real-time diagnosis in actual engineering. The present invention adopts a heterogeneous transfer learning framework (HFTL) in a heterogeneous federated domain generalization network (HFDGN), which can realize generalized fault diagnosis of the target client through the public knowledge representation mapping of the heterogeneous source clients. In addition, the present invention adopts a decoupled domain adaptive basic model (DDA) to remove the negative impact of noise. At the same time, the basic model can enhance the ability of domain confusion and extract features related to intrinsic faults.

[0052] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0054] Figure 1 It is the structure diagram of the decoupled domain adaptation basic model (DDA);

[0055] Figure 2 This is a schematic diagram of the heterogeneous federated transfer learning framework (HFTL);

[0056] Figure 3 It is a DDS test bench;

[0057] Figure 4 For the CWRU test bed;

[0058] Figure 5 It is a RDS test bench;

[0059] Figure 6 It is a test bench for SWJTU;

[0060] Figure 7 To compare the experimental results. DETAILED DESCRIPTION

[0061] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0062] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0063] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0064] See also Figure 1 to Figure 7 The embodiment of the present invention provides a cross-bearing migration diagnosis method based on a heterogeneous federated domain generalized network, and the specific process is as follows:

[0065] Step 1: The original vibration signal on the mechanical equipment is collected by the acceleration sensor to facilitate the construction of the subsequent migration diagnosis task. Then, the collected original vibration signal is expanded using the sliding window sampling technology.

[0066] Step 2: Build a decoupled domain adaptation basic model (DDA) based on a one-dimensional convolutional neural network.

[0067] like Figure 1 As shown in Figure 2, the backbone network of the decoupled domain adaptation basic model (DDA) consists of four parts: feature extractor G FE (θ FE ), decoupler G D (θ D ), Reconstructor G R (θ R ) and the fault classifier G FC (θ FC ), where θ FE ,θ D ,θ R and θ FC They represent the trainable weights of the corresponding network models. The feature extractor is used to mine the knowledge of distribution differences and reduce the distribution differences, and then obtain the universal feature F G ; The fault classifier is dedicated to extracting label features F L To identify the fault type; the decoupler is used to decouple and separate the fault-related features F FR F FI ; The reconstructor reconstructs common features from fault-related features and fault-irrelevant features, so the decoupler and reconstructor can be regarded as the encoder and decoder structures in the autoencoder.

[0068] The optimization objectives of the decoupled domain adaptation basic model mainly include the following three parts: ① Decoupled representation: learning a decoupled representation to separate the fault-irrelevant features caused by noise from the general features; ② Separation and reconstruction: maximizing the distribution distance between fault-related features and fault-irrelevant features, so as to ensure the independence of the extracted fault-related features; ③ Distribution adaptation: minimizing the distribution difference between any two fault-related features. The following will introduce these three parts in detail:

[0069] ① Decoupled representation: In order to achieve feature decoupling, the decoupler and fault classifier are used to remove the irrelevant features of the fault caused by noise in an adversarial training manner. First, the cross entropy loss L is used C Feature extractors, decouplers, and fault classifiers are trained to accurately identify fault types.

[0070]

[0071] Among them, X i ,Y i Represent the i-th data sample and the corresponding label respectively; represents the sample domain, C represents the fault type; I() represents the indicator function, when argmax(Y i )=c, I=0.

[0072] Then, the weight parameters of the feature extractor and the fault classifier are fixed, and the decoupler is trained to deceive the fault classifier by maximizing the information entropy loss, thereby learning the fault-irrelevant features. The information entropy loss is a value that can reflect the purity of the sample prediction label. The smaller the information entropy loss, the higher the purity of the predicted label probability. Information entropy (IE) loss L IE It is expressed as follows:

[0073]

[0074] in, Represents sample X i The predicted label vector of represents the cth element in the predicted label vector, which is obtained by the following formula:

[0075]

[0076] Through adversarial training between formulas (1) and (2), the final fault-irrelevant features and fault-related features can be separated and obtained by the decoupler.

[0077] ② Separation and reconstruction: In order to enhance the independence of the extracted fault-related features, the average difference loss L is maximized. DTo train the decoupler to expand the distribution difference between fault-related features and fault-irrelevant features:

[0078]

[0079] where φ(·) is the high-dimensional mapping function in the regenerated Hilbert space; where the fault-related feature F FR F FI Obtained by the following formula:

[0080] {F FR ,F FI =G D (G FE (X)) (5)

[0081] At the same time, in order to avoid the over-training of formula (5) destroying the intrinsic properties of the extracted fault-related features, the reconstruction loss L R To train the reconstructor to reconstruct common features from fault-related features and fault-irrelevant features;

[0082]

[0083] ③ Distribution Adaptation: After the above two processes, in order to align the distribution of fault-related features extracted between different domains, the domain confusion loss L is also minimized using the average difference indicator DC To train the feature extractor and disentangler, thus achieving distribution adaptation.

[0084]

[0085] Assume that the client has Q m There are K auxiliary domains, and there will be K domain confusion losses. in According to the transfer learning theory, we can know represents the distribution similarity of the corresponding two auxiliary domains, that is, The smaller the value, the more the feature extractor will focus on the corresponding Therefore, in order to enable the feature extractor to learn more general domain-invariant features, the purity loss L constructed using information entropy is used. P To train the feature extractor and decoupler to reduce the purity between all domain confusion losses, that is, to maximize the purity loss L P .

[0086]

[0087] The data of multiple auxiliary domains of the source client will fully participate in the above three optimization processes. Because the target client assumes that the target domain data is inaccessible during the training process and can only be used for testing, the target client only uses the source domain data to participate in the two optimization processes of "decoupled representation" and "separation and reconstruction".

[0088] Step 3: Build a heterogeneous federated transfer learning network (HFDGN) based on mutual information weight matching mechanism (MIWM), such as Figure 2 As shown, a heterogeneous transfer learning framework (HFTL) is adopted.

[0089] The heterogeneous transfer learning framework proposed in this invention is elaborated in detail from three aspects:

[0090] ① Heterogeneous transfer learning framework structure

[0091] like Figure 2 As shown in the figure, the heterogeneous federated transfer learning framework consists of multiple source clients, a target domain client, a central server and a mutual information weight matching (MIWM) mechanism. The source clients and the target clients can be heterogeneous; the central server is dedicated to the allocation of weight parameters of the decoupled domain adaptation basic model (DDA) trained by all source clients; the mutual information weight matching mechanism (MIWM) is used to evaluate the contribution of each source client to the target client during the weight parameter allocation process.

[0092] ②Federal Communication Paradigm

[0093] pass Figure 2 From the federated communication paradigm demonstrated in , we can know that all source domain clients need to upload the weight parameters of feature extractors, decouplers, and fault classifiers in the decoupled domain adaptation basic model (DDA) to the central server; for the target client, the universal feature F extracted from the source domain data in the decoupled domain adaptation basic model (DDA) G and label feature F L Need to be uploaded to the central server. Because the high-dimensional features extracted by the feature extractor are more general and transferable, the central server only aggregates the weight parameters of the feature extractors from all source clients.

[0094]

[0095] Where M represents the number of source clients; represents the weight parameter of the feature extractor from the i-th source client; w iThe weight parameters of the corresponding source client and target client are obtained through the mutual information weight matching mechanism (MIWM). Finally, the central server distributes the aggregated weight parameters to each source client and target client.

[0096] ③ Mutual Information Weight Matching Mechanism (MIWM)

[0097] In the mutual information weight matching mechanism, mutual information is used to evaluate the contribution of each source client to the target client. Mutual information can reflect the similarity between two data sample domains. The greater the mutual information, the higher the similarity. The mutual information between is defined as follows:

[0098]

[0099] in, and Respectively represent and The marginal probability distribution of is the joint probability distribution between the two domains; Represents the product of the marginal probability distributions between two domains. The relationship between the joint probability distribution and the marginal probability distribution is shown as follows:

[0100]

[0101] Because it is difficult to calculate the mutual information between unknown continuous variables in equation (10), we reformulate equation (10) through Donsker-Varadhan variational theory as follows:

[0102]

[0103] Where T(θ) represents the function set formed by the network parameters. Through the mutual information evaluation of formula (12), the contribution of the i-th source client to the target client MI i It can be obtained by the following formula:

[0104]

[0105] in, and represents the fault classifier and decoupler in the decoupled domain adaptation base model trained for the i-th source client; F G and F L They represent the general features and label features obtained by training the decoupled domain adaptation basic model with source domain data in the target client. Finally, the weight w in formula (9) i can be expressed as:

[0106]

[0107] Among them, τ represents the temperature coefficient. The smaller the temperature coefficient, the larger {MI1,MI2,…,MI M}.

[0108] Combining the decoupled domain adaptation base model and the heterogeneous federated transfer learning framework, the final heterogeneous federated domain generalization network (HFDGN) is constructed to achieve data federation under privacy protection.

[0109] Step 4: Input the divided training samples in all auxiliary domains of the source client server and the source domain of the target client into the constructed decoupled domain adaptation basic model (DDA), and use the optimization objective function in the decoupled domain adaptation basic model (DDA) to train the corresponding local client model.

[0110] Step 5: Upload the parameters of the trained decoupled domain adaptation basic model of the target client and the source client to the central server, and then use the built heterogeneous federated transfer learning network (HFDGN) to perform federated migration fault diagnosis.

[0111] Step 6: After multiple iterations of training, the error curve tends to be stable, the model training is completed, and the trained heterogeneous federated domain generalization network will be used for heterogeneous multi-source federated migration diagnosis.

[0112] The following experimental results illustrate the effectiveness of the above intelligent diagnosis method.

[0113] Verification experiment: This experiment collected vibration signals of three bearing fault simulation test benches (CWRU, RDS, SWJTU) and a vibration signal of a gear fault simulation test bench (DDS).

[0114] ① Such as Figure 3 As shown in the figure, the DDS test bench is mainly composed of five parts: motor, planetary gearbox, parallel gearbox and magnetic powder brake. Different working condition signals can be obtained through magnetic powder brake loading simulation. The planetary gearbox has five kinds of health information: normal, surface wear, tooth root crack, gear tooth defect, and gear tooth breakage. In addition, three working conditions are simulated: 0N·m (G1), 1.4N·m (G2), and 2.8N·m (G3).

[0115] ② The CWRU bearing dataset is the CWRU standard bearing dataset published by Case Western Reserve University in the United States. The schematic diagram of the CWRU bearing dataset is as follows: Figure 4 As shown in the figure, it consists of a motor, bearings at both ends of the motor, a torque sensor and a power meter. Four faults are simulated, including normal, inner race fault, rolling element fault and outer race fault. It contains four working condition information: hp(C1), 1hp(C2), 2hp(C3), and 3hp(C4).

[0116] ③ Schematic diagram of the test bench for collecting RDS bearing data sets Figure 5 As shown in the figure, it consists of a servo motor, a coupling, a rotor, and the bearings and bearing seats at both ends. Like the CWRU bearing dataset, four health states (normal, inner race fault, rolling element fault, and outer race fault) are also simulated. By applying different weights to the rotor, four load conditions are simulated: 0N (R1), 14N (R2), 28N (R3), and 44N (R4).

[0117] ④SWJTU bearing data set comes from Southwest Jiaotong University, and its test bench diagram is as follows Figure 6 As shown in Figure 1, it consists of a motor, a support bearing, a faulty bearing used for testing, and a loading system. Four health states (normal, inner ring fault, rolling element fault, and outer ring fault) and four load conditions (0 kN (B1), 1 kN (B2), 2 kN (B3), and 3 kN (B4)) are simulated.

[0118] Through the above four datasets, a CWRU+RDS+SWJTU→DDS heterogeneous multi-source federated diagnosis instance is constructed to evaluate the diagnostic performance of HFDGN.

[0119] Comparative experiment:

[0120] In order to demonstrate the superiority of the migration diagnosis method based on heterogeneous federated domain generalization network (HFDGN) in this paper, compared with the current typical domain generalization migration diagnosis model (WDCNN, Whitening-net) and the classic domain adaptive diagnosis model (DANN, DDC), the experimental results of six generalization migration diagnosis tasks are as follows: Figure 7 As shown. Figure 7 It can be concluded that the heterogeneous federated domain generalized network proposed in the present invention has higher migration diagnosis accuracy and stronger generalization ability. In addition, the average diagnostic accuracy of HFDGN on six generalized migration diagnosis tasks reached more than 89.85%, and compared with the domain adaptive diagnosis model (DANN, DDC), it has an approximately the same diagnostic effect. In particular, for the G1→G3, G2→G3 and G3→G2 migration diagnosis tasks, HFDGN has a more significant improvement in diagnostic accuracy compared to all other comparison models.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. An intelligent fault migration diagnosis method based on a heterogeneous federated domain generalized network, characterized in that: The method specifically comprises the following steps: S1: Collect the original vibration signal on the mechanical equipment through the sensor, and expand the sample of the collected original vibration signal; S2: Construct a decoupled domain adaptation basic model, referred to as the DDA model; The backbone network of the DDA model consists of four parts: feature extractor G FE (θ FE ), decoupler G D (θ D ), Reconstructor G R (θ R ) and the fault classifier G FC (θ FC ), where θ FE ,θ D ,θ R and θ FC They represent the trainable weights of the corresponding network models respectively; the feature extractor is used to mine the knowledge of distribution differences and reduce the distribution differences, and then obtain the universal feature F G ; The fault classifier is dedicated to extracting label features F L To identify the fault type; the decoupler is used to decouple and separate the fault-related features F FR F FI ;The reconstructor reconstructs common features from fault-related features and fault-irrelevant features; The optimization objectives of the DDA model include the following three parts: 1) Decoupled representation: learning a decoupled representation to separate the fault-irrelevant features caused by noise from the general features; 2) Separation and reconstruction: maximizing the distribution distance between fault-related features and fault-irrelevant features; 3) Distribution adaptation: minimizing the distribution difference between any two fault-related features; S3: Build a heterogeneous federated transfer learning network based on the mutual information weight matching mechanism, where the mutual information weight matching mechanism is referred to as MIWM; the heterogeneous federated transfer learning network is referred to as HFDGN, which adopts a heterogeneous transfer learning framework; S4: Input the divided training samples in all auxiliary domains of the source client and the source domain of the target client into the constructed DDA model, and use the optimization objective function in the DDA model to train the corresponding local client model; S5: Upload the parameters of the trained DDA model to the central server, and then use the built HFDGN network to perform federated migration fault diagnosis; S6: After multiple iterations of training, the error curve tends to be stable, and the HFDGN network training is completed. The trained HFDGN network is used for heterogeneous multi-source federated migration diagnosis.

2. The intelligent fault migration diagnosis method according to claim 1, characterized in that: In step S2, in the optimization goal 1) of the DDA model, the decoupled representation is specifically: using the decoupler and the fault classifier to remove the irrelevant features of the fault caused by noise in an adversarial training manner; first, using the cross entropy loss L C Train feature extractors, decouplers, and fault classifiers to accurately identify fault types; in, represents the expectation of the sample domain; X i ,Y i Represent the i-th data sample and the corresponding label respectively; represents the sample domain, C represents the fault type; I(·) represents the indicator function, when argmax(Y i )=c, I=0; Then, the weight parameters of the feature extractor and the fault classifier are fixed, and the decoupler is trained to deceive the fault classifier by maximizing the information entropy loss, thereby learning the fault-irrelevant features; the information entropy loss L IE It is expressed as follows: in, It means to seek expectation on the sample domain; Represents sample X i The predicted label vector of represents the cth element in the predicted label vector, which is obtained by the following formula: Through adversarial training between formulas (1) and (2), the final fault-irrelevant features and fault-related features are separated by the decoupler.

3. The intelligent fault migration diagnosis method according to claim 1, characterized in that: In step S2, the optimization goal 2) of the DDA model is to separate and reconstruct by maximizing the average difference loss L D To train the decoupler to expand the distribution difference between fault-related features and fault-irrelevant features: in, They represent the expectation of fault-irrelevant features and fault-related features respectively; φ(·) is the high-dimensional mapping function in the regenerated Hilbert space; represents the 2-norm of the regenerated Hilbert space; Fault related features F FR F FI Obtained by the following formula: {F FR ,F FI }=G D (G FE (X)) (5) At the same time, through the reconstruction loss L R To train the reconstructor to reconstruct common features from fault-related features and fault-irrelevant features; in, It means to find the expectation of the sample domain.

4. The intelligent fault migration diagnosis method according to claim 1, characterized in that: In step S2, in the optimization goal 3) of the DDA model, distribution adaptation is specifically: using the minimization of the average difference index as the domain confusion loss L DC To train the feature extractor and decoupler to achieve distribution adaptation; in, represents the expectation of irrelevant fault features, φ(·) is the high-dimensional mapping function in the regenerated Hilbert space; represents the 2-norm of the regenerated Hilbert space; Assume that the client has Q m auxiliary domains, there will be K domain confusion losses in Purity loss L constructed using information entropy P To train the feature extractor and decoupler to reduce the purity between all domain confusion losses, that is, to maximize the purity loss L P ; 5. The intelligent fault migration diagnosis method according to claim 1, characterized in that: In step S2, the data of multiple auxiliary domains of the source client will fully participate in the three optimization objectives of the DDA model; the target client only has the source domain data participating in the two optimization objectives of "decoupled representation" and "separation and reconstruction".

6. The intelligent fault migration diagnosis method according to claim 1, characterized in that: In step S3, the MIWM mechanism is used to evaluate the contribution of each source client to the target client during the weight parameter allocation process; The contribution of the i-th source client to the target client is MI i The expression is: in, and represents the fault classifier and decoupler in the DDA model trained for the i-th source client; F G and F L They represent the common features and label features obtained by training the DDA model with source domain data in the target client; I Θ Represents two data sample domains The mutual information between them can be expressed as: Among them, T(θ) represents the function set formed by the network parameters; and Respectively represent and The marginal probability distribution of is the joint probability distribution between the two domains; represents the product of the marginal probability distribution between two domains; Θ represents the weight parameter set, θ represents the weight parameter, represents the expectation of the joint probability of two domains, It represents the expectation of the product of the marginal probabilities of two domains.

7. The intelligent fault migration diagnosis method according to claim 6, characterized in that: In step S5, the central server aggregates the weight parameters of the feature extractors from all source clients, and the calculation formula is: Where M represents the number of source clients; represents the weight parameter of the feature extractor from the i-th source client; Represents the weight parameter after aggregation; w i The weight parameters for matching the corresponding source client and target client are obtained through the MIWM mechanism; Where τ represents the temperature coefficient; The central server then distributes the aggregated weight parameters to each source and destination client.

8. The intelligent fault migration diagnosis method according to claim 1 or 7, characterized in that: In step S5, the DDA model parameters uploaded to the central server include: for all source clients, the weight parameters of the feature extractor, decoupler and fault classifier in the DDA model need to be uploaded; for the target client, the general feature F extracted from the source domain data in the DDA model needs to be uploaded G and label feature F L .

Images