Intelligent diagnosis method of mechanical faults based on generalized network of relation transfer domain

Abstract

The present invention relates to an intelligent diagnosis method for mechanical faults based on a relational migration domain generalization network, and belongs to the technical field of mechanical fault diagnosis. The method comprises: S1: collecting original vibration signals under different working conditions, and then dividing sample sets; S2: constructing a relational migration framework; S3: constructing a relational migration domain generalization network RTDGN; S4: training and optimizing RTDGN: in the TIDA stage, all task-irrelevant auxiliary data are input into the RTDGN, and the weight parameters of the constructed adversarial network are iteratively trained; in the TRDG stage: firstly inheriting the weight parameters of the feature extractor in the TIDA stage, and then inputting the source domain data in the task-related data into the RTDGN for fine-tuning training; S5: after multiple iterations, the trained RTDGN will be used for cross-working condition migration fault diagnosis under the relational migration task of rotating mechanical parts, and the diagnosis results with high generalization ability and strong robustness will be obtained.

Description

Technical Field

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

[0002] Key mechanical parts such as gears and bearings are widely used in important engineering fields such as aerospace, rail transportation, and petrochemicals. Once a failure occurs, it will cause serious economic losses and irreparable casualties. At present, a large number of migration diagnosis methods based on domain adaptation technology have been widely used in the field of mechanical diagnosis to address the inconsistency between the distribution of training data and test data; however, the success of these methods completely depends on the prior assumption that the target domain data can be obtained during model training. In most practical engineering applications, this prior assumption is often not true.

[0003] Therefore, a new mechanical intelligent diagnosis method is urgently needed to solve the above problems. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide an intelligent diagnosis method for mechanical faults based on a relational migration domain generalization network, so that the rotating mechanical parts have high generalization ability and strong robustness diagnosis results when the data distribution is inconsistent and the target domain data is invisible.

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

[0006] A mechanical fault intelligent diagnosis method based on a relational migration domain generalization network specifically comprises the following steps:

[0007] S1: Collect original vibration signals under different working conditions on mechanical equipment, and then divide the sample set according to the collected original vibration signals to facilitate the construction of subsequent relationship migration diagnosis tasks;

[0008] S2: Construct a relational transfer framework based on the relational transfer diagnosis task to be implemented;

[0009] S3: Based on the relationship transfer framework and convolutional neural network constructed in step S2, a relationship transfer domain generalization network (RTDGN) is constructed;

[0010] S4: The RTDGN training is optimized in two steps: ① In the task-irrelevant domain adaptation stage (TIDA): all task-irrelevant auxiliary data are input into the RTDGN, and then the weight parameters of the constructed adversarial network are iteratively trained, where the weight parameters of the adversarial network include the weight parameters θ of the feature extractor, fault classifier, and domain discriminator FE ,θ FC and ② In the task-relevant domain generalization stage (TRDG): first inherit the weight parameters of the feature extractor in the TIDA stage, and then input the source domain data in the task-relevant data into the RTDGN for fine-tuning training;

[0011] S5: After multiple iterations of training, observe whether the loss curve tends to converge smoothly. If the curve converges, the RTDGN training is completed. The trained RTDGN will be used for cross-condition migration fault diagnosis under the relational migration task of rotating mechanical parts.

[0012] Further, in step S2, building a relational migration framework specifically includes: first defining task-irrelevant auxiliary data It is composed of multiple auxiliary domains. Where Q represents the number of auxiliary domains, and denotes the i-th sample of the q-th auxiliary domain and its corresponding label vector, n q is the number of samples in the qth auxiliary domain; task-related data It consists of source domain data and target domain data, defined as in and denote the labeled source domain and the unseen unlabeled target domain, respectively. Respectively represent the i-th sample in the source domain and its corresponding label vector, n S is the number of samples in the source domain, represents the i-th sample in the target domain and its corresponding label vector, n T is the number of samples in the target domain;

[0013] Define the relationship between any two unrelated auxiliary domains of the task as where i,j∈[1,2,…,Q] and i≠j; φ I (·) represents the relationship mapping between task-irrelevant auxiliary data, ΔP ij represents the distribution difference between the ith auxiliary domain and the jth auxiliary domain, when ΔP ij = 0, Similarly, the relationship between the source and target domains of task-related data is defined as Among them, φ R (·) represents the relationship mapping between task-related data, ΔP ST represents the distribution difference between the source domain and the unlabeled target domain; according to the similarity of the distribution difference between the task-irrelevant auxiliary data and the task-relevant data, that is, φ R ≈φ I , and then the distribution difference relationship of task-irrelevant auxiliary data is migrated to task-relevant data through relationship transfer, thereby indirectly achieving the measurement and elimination of the distribution difference between the invisible target domain and the source domain.

[0014] Further, in step S3, the construction of RTDGN includes two stages, TIDA and TRDG. Both stages use convolutional neural networks as the backbone network to extract features. In the TIDA stage, a general feature extractor with domain confusion and separability is constructed based on task-irrelevant auxiliary data. Then, the distribution relationship knowledge learned by the feature extractor in the TIDA stage is transferred to the TRDG stage through the relational transfer framework to achieve domain confusion.

[0015] In the backbone network, batch normalization (BN) is used to accelerate the convergence of the neural network and avoid variance drift; in addition, Dropout layer and global average pooling (GAP) are used to prevent network overfitting; the hidden layer uses ReLU as the activation function, and the output layer uses Softmax and Sigmoid activation functions.

[0016] Furthermore, in step S3, the TIDA stage, an adversarial network with multiple domain discriminators is constructed, including a feature extractor G FE (θ FE ), a fault classifier G FC (θ FC ) and K binary classification domain discriminators where k∈[1,2…,K] and represents the number of combinations of Q elements, θ FE ,θ FC and denote the weight parameters of the feature extractor, fault classifier and domain discriminator respectively.

[0017] Furthermore, in step S3, the TRDG stage, a domain generalization diagnosis network is constructed, including a feature extractor G FE (θ FE ) and residual fusion classifier G RC (θ RC); The structure and weight parameters of the feature extractor in the TRDG stage are the same as those in the TIDA stage; the residual fusion classifier is a fault classifier that fuses multiple features through residual connections.

[0018] Further, in step S4, in the TIDA stage, the weight parameters of the constructed adversarial network are iteratively trained, specifically including: using the Adam optimizer and the domain loss L D , classification loss L C , inverse entropy loss L IE , the weight parameter θ of the adversarial network in the TIDA stage FE ,θ FC and Iteratively updated:

[0019]

[0020]

[0021]

[0022] Among them, ε0 represents the learning rate in the TIDA phase, μ and λ are trade-off parameters, and θ D represents the weight parameters of all domain classifiers, and In order to directly perform gradient parameter updates, the adversarial network introduces a gradient reversal layer (GRL).

[0023] Furthermore, the domain loss L D Defined as:

[0024]

[0025] Where d represents the one-hot encoded vector of the domain label, represents the k-th binary cross entropy loss;

[0026] Classification loss L C Defined as:

[0027]

[0028] Among them, Y represents the label vector corresponding to the auxiliary domain sample, represents the multi-classification cross entropy loss of the qth auxiliary domain;

[0029] Inverse entropy loss L IE Defined as:

[0030]

[0031] in,

[0032] Furthermore, in step S4, the weight parameter θ of the TRDG stage FE ,θ RC The update method is as follows:

[0033]

[0034]

[0035] Among them, ε1 and ε2 represent the learning rate, satisfying ε1<ε2; Represents the multi-class cross entropy loss of the source domain.

[0036] Furthermore, the multi-classification cross entropy loss of the source domain Defined as: where Y S Represents the label vector corresponding to the source domain sample.

[0037] The beneficial effects of the present invention are:

[0038] (1) The present invention constructs a relational migration framework; through distribution difference similarity, the framework can migrate the distribution difference information obtained from a rotating machinery component with a complete label data set (task-irrelevant auxiliary data) to another rotating machinery component with incomplete data (task-relevant data). It can not only maximize the use of data resources, but also help rotating machinery components with incomplete data to indirectly measure and reduce the distribution difference between the source domain and the target domain when the target domain data cannot be obtained. Ultimately, the constructed migration diagnosis model obtains diagnostic results with high generalization ability and strong robustness in the case of inconsistent data distribution and invisible target domain data.

[0039] (2) Based on the constructed relational transfer framework, the present invention designs a mechanical intelligent diagnosis method based on the relational transfer domain generalization network (RTDGN). In the RTDGN, two stages are divided: 1) a task-irrelevant domain adaptation stage (TIDA) for task-irrelevant auxiliary data; 2) a task-relevant domain generalization stage (TRDG) for task-relevant data. In the TIDA stage, an adversarial network with multiple domain discriminators is constructed to enhance the domain confusion capability; in addition, the present invention designs an inverse entropy loss to help the network obtain a more general domain confusion capability. In the TRDG stage, the present invention constructs a residual fusion classifier to effectively suppress the fault classification from falling into overfitting, while improving the generalization ability of the fault classifier.

[0040] 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

[0041] 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:

[0042] Figure 1 A schematic diagram of a relationship migration task in an embodiment of the present invention;

[0043] Figure 2 This is the structural diagram of the Relation Transfer Domain Generalization Network (RTDGN);

[0044] Figure 3 This is a schematic diagram of the DDS test bench;

[0045] Figure 4 Schematic diagram of the RDS test bench. DETAILED DESCRIPTION

[0046] 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.

[0047] 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.

[0048] 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.

[0049] See also Figure 1 to Figure 4 This embodiment provides a mechanical fault intelligent diagnosis method based on a relational transfer domain generalization network (RTDGN) of bearings and gears, which specifically includes the following steps:

[0050] Step 1: Install an accelerometer on the mechanical equipment to collect the original vibration signals under different working conditions. Then, divide the sample set according to the collected original vibration signals to facilitate the construction of subsequent relationship migration diagnosis tasks.

[0051] Step 2: Construct a relational migration framework based on the relational migration diagnostic task to be implemented, such as Figure 1 shown.

[0052] Building a relational transfer framework: First, define task-irrelevant auxiliary data It is composed of multiple auxiliary domains. Where Q represents the number of auxiliary domains, and denotes the i-th sample of the q-th auxiliary domain and its corresponding label vector, n q Defined as the number of samples in the qth auxiliary domain. Task-related data It consists of source domain data and target domain data, defined as in and They are defined as the labeled source domain and the unseen unlabeled target domain, The i-th sample of the source domain and its corresponding label vector, n S is the number of samples in the source domain, represents the i-th sample of the unlabeled target domain and its corresponding label vector, n T is the number of samples in the unlabeled target domain.

[0053] In the relation transfer framework, the relationship between any two task-unrelated auxiliary domains is defined as where i,j∈[1,2,…,Q] and i≠j; φ I(·) represents the relationship mapping, ΔP ij represents the distribution difference between the ith auxiliary domain and the jth auxiliary domain, when ΔP ij = 0, Similarly, the relationship between the source and target domains of task-related data is defined as However, since the target domain samples are unknown, the relationship between the source domain and the target domain cannot be measured and eliminated. However, the distribution difference similarity (φ R ≈φ I ), and then migrate the distribution difference relationship of task-irrelevant auxiliary data to task-relevant data through relation transfer, thereby indirectly achieving the measurement and elimination of the distribution difference between the invisible target domain and the source domain.

[0054] like Figure 1 As shown, the gear→bearing relationship migration task is taken as an example. Assume that the bearing knows the vibration signal and its label prior knowledge under one of the three working conditions (source domain), and needs to diagnose the fault category under other working conditions (target domain). In the domain generalization task, the target domain data is not available during training, so the distribution differences caused by different loads will greatly reduce the robustness and generalization ability of the diagnostic model trained from the source domain. Through the relationship migration framework proposed in the present invention, the distribution differences caused by load changes can be indirectly obtained from complete and labeled task-irrelevant auxiliary data, thereby achieving domain confusion of task-related data without visibility in the target domain.

[0055] Step 3: Based on the relation transfer framework and convolutional neural network constructed in step 2, a relation transfer domain generalization network (RTDGN) is constructed, as shown in Figure 2 shown.

[0056] RTDGN includes two stages: TIDA and TRDG. Figure 2 As shown. Both stages use convolutional neural networks as the backbone network to extract features. In the TIDA stage, a general feature extractor with domain confusion and separability is constructed based on task-irrelevant auxiliary data. This feature extractor is then migrated to the TRDG stage through the relational migration framework proposed in this invention. We used a batch normalization layer (BN) in the backbone network to accelerate the convergence of the neural network and avoid variance drift; in addition, the Dropout layer and global average pooling (GAP) were employed to prevent the network from overfitting. The hidden layer uses ReLU as the activation function, and the output layer uses Softmax and Sigmoid activation functions. The TIDA and TRDG stages will be introduced in detail below:

[0057] 1) TIDA stage

[0058] In the TIDA stage, in order to reduce the distribution differences between the Q auxiliary domains and ensure the robustness of relation transfer, an adversarial network with multiple domain discriminators is constructed. Figure 2 As shown, this adversarial network consists of a feature extractor G FE (θ FE ), a fault classifier G FC (θ FC ) and K binary classification domain discriminators Composition, where k∈[1,2…,K] and θ FE ,θ FC and Represent the weight parameters of the feature extractor, fault classifier and domain discriminator respectively. Through adversarial training between K domain discriminators and feature extractors, domain confusion between any two auxiliary domains can be achieved. Through the following three loss functions, domain-invariant and separable high-dimensional universal features can be obtained.

[0059] ① Domain loss L D

[0060] The domain loss is employed to remove the distribution difference between two arbitrary auxiliary domains, which will help the feature extractor extract common domain-invariant features, defined as follows:

[0061]

[0062] Where d represents the one-hot encoded vector of the domain label, represents the k-th binary cross entropy loss, θ D represents the weight parameters of all domain classifiers, and

[0063] ②Classification loss L C

[0064] Through domain loss, the feature extractor can only achieve domain confusion, but loses the classification ability. We hope that the feature extractor will eventually obtain domain-invariant and separable features. Therefore, through supervised training, classification loss is used to minimize the classification risk.

[0065]

[0066] in, represents the multi-class cross entropy loss of the q-th auxiliary domain.

[0067] ③ Anti-entropy loss L IE

[0068] Assuming that all domain discriminators have obtained excellent discrimination ability, the loss output of each domain discriminator can be used to measure the distribution similarity between two auxiliary domains. For example, the larger the k-th binary cross entropy loss, the more the feature extractor pays attention to the distribution alignment between the two auxiliary domains corresponding to the k-th domain discriminator. Usually we hope that the feature extractor can evenly achieve domain confusion between all auxiliary domains, rather than focusing too much on the distribution alignment between two domains, which will cause the domain invariant features learned by the feature extractor to overfit. Therefore, in order to effectively implement the relationship transfer framework proposed in the present invention, an inverse entropy loss is designed to help the feature extractor obtain more universal high-dimensional general features. In order to facilitate the derivation of the inverse entropy loss, the domain loss of the k-th domain discriminator is first defined:

[0069]

[0070] Then corresponding to p k The weights can be obtained:

[0071]

[0072] in, The final inverse entropy loss is expressed as:

[0073]

[0074] like Figure 2 As shown, the inverse entropy loss only acts on the weight parameter update of the feature extractor.

[0075] Through the Adam optimizer and the above three losses, the weight parameters (θ FE ,θ FC and ) can be updated iteratively:

[0076]

[0077]

[0078]

[0079] Among them, ε0 represents the learning rate in the TIDA stage, and μ and λ are trade-off parameters. In order to directly perform the update of gradient parameters, the adversarial network introduces a gradient reversal layer (GRL).

[0080] 2) TRDG stage

[0081] like Figure 2 As shown, in the TRDG stage, a feature extractor G FE (θFE ) and residual fusion classifier G RC (θ RC ) domain generalization diagnosis network is constructed. The structure and weight parameters of the feature extractor in the TRDG stage are the same as those in the TIDA stage. Because the features obtained by the feature extractor are more general and transferable than those of the residual fusion classifier. It also reflects that the residual fusion classifier is prone to overfitting in the source domain. Therefore, in order to effectively alleviate the classifier from falling into overfitting and obtain a more generalized diagnostic effect, a fault classifier that fuses multiple features through residual connections is constructed.

[0082] For the TRDG stage, since the target domain is invisible, the distribution difference between the source domain and the target domain is immeasurable, reflecting that the domain generalization network trained only by source domain samples is a network without transfer diagnosis. However, through the relationship transfer framework proposed in the present invention, the distribution relationship knowledge learned by the transfer feature extractor in the TIDA stage is transferred to the TRDG stage to achieve domain confusion.

[0083] The fusion classification loss of the residual fusion classifier is defined as:

[0084]

[0085] in, Represents the multi-class cross entropy loss of the source domain.

[0086] In the gradient back propagation of the fusion classification loss, we use two optimizers with different learning rates to perform fine-tuning operations. The optimizer with a smaller learning rate is used for the update learning of the feature extractor, and the optimizer with a larger learning rate is used for the fusion classifier. The weight parameter update (θ FE ,θ RC )as follows:

[0087]

[0088]

[0089] Among them, the learning rate ε1 and the learning rate ε2 should satisfy ε1<ε2.

[0090] Step 4: Training optimization is performed in two steps: ① In the TIDA stage, all task-irrelevant auxiliary data are input into the relational transfer domain generalization network, and then the weight parameters of the constructed network are iteratively trained using equations (6) to (8). ② In the TRDG stage, the weight parameters of the feature extractor in the TIDA stage are first inherited, and then the source domain data in the task-relevant data is input into the relational transfer domain generalization network for fine-tuning training using equations (10) to (11).

[0091] Step 5: After multiple iterations of training, observe whether the loss curve tends to converge smoothly. If the curve converges, the network model training is completed. The trained relationship transfer domain generalization network will be used for cross-condition migration fault diagnosis under the relationship migration task of bearings and gears.

[0092] The above are the process steps of using the proposed relational migration domain generalization network for cross-operating condition migration fault diagnosis under the relational migration task of bearings and gears. The following experiments are used to illustrate the effectiveness of the method of the present invention.

[0093] Verification experiment:

[0094] The original gear vibration signal collected in this experiment comes from the DDS test bench, such as Figure 3 As shown. 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, which are 0N.m (G1), 1.4Nm (G2) and 2.8Nm (G3). The planetary gearbox has five types of health information: normal (NC), surface wear (SW), tooth root crack (RC), gear tooth defect (CT), and gear tooth breakage (MT). The acceleration sensor is installed on the planetary gearbox to collect the original vibration signal.

[0095] The original bearing vibration signals collected in this experiment come from the RDS test bench, such as Figure 4 As shown. The RDS test bench consists of a servo motor, a coupling, a rotor, and bearings and bearing seats at both ends. Different operating condition signals can be obtained with rotors of different masses, which are 0N (B1), 28N (B2) and 44N (B3). The bearing has four types of health information: normal (NC), inner ring fault (IF), outer ring fault (OF) and rolling element fault (BF). The acceleration sensor is installed on the planetary bearing seat to collect the original vibration signal.

[0096] Through the bearing and gear data sets collected above, a total of two relationship migration tasks can be constructed, and each migration task contains six cross-condition migration diagnosis tasks, as shown in Tables 1 and 2.

[0097] Comparative experiment:

[0098] In order to prove the superiority of the relational migration framework and the constructed relational migration domain generalization network proposed in the present invention, the current typical fault diagnosis network model is compared, and the experimental results are shown in Table 1 and Table 2. According to the data results in Table 1 and Table 2, it can be concluded that the relational migration framework and the constructed relational migration domain generalization network (RTDGN) proposed in the present invention have higher migration diagnosis accuracy and stronger robustness and generalization ability.

[0099] Table 1 Experimental results of gear→bearing relationship transfer task

[0100]

[0101]

[0102] Table 2 Experimental results of bearing→gear relationship transfer task

[0103]

[0104] In Table 1 and Table 2, RTDGN (w / o) represents the diagnostic model without the inverse entropy loss on RTDGN, which belongs to the ablation experiment and is used to verify the effectiveness of the inverse entropy loss constructed by the present invention.

[0105] 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. A mechanical fault intelligent diagnosis method based on a relational migration domain generalization network, characterized in that: The method specifically comprises the following steps: S1: Collect original vibration signals under different working conditions on mechanical equipment, and then divide the sample set according to the collected original vibration signals; S2: Construct a relational transfer framework based on the relational transfer diagnosis task to be implemented; S3: Based on the relation transfer framework and convolutional neural network constructed in step S2, a relation transfer domain generalization network, namely RTDGN, is constructed; S4: The RTDGN training is optimized in two steps: ① In the task-irrelevant domain adaptation stage, i.e., the TIDA stage: all task-irrelevant auxiliary data are input into the RTDGN, and then the weight parameters of the constructed adversarial network are iteratively trained, where the weight parameters of the adversarial network include the weight parameters θ of the feature extractor, fault classifier, and domain discriminator. FE ,θ FC and θ Dk ;②In the task-related domain generalization stage, namely the TRDG stage: first inherit the weight parameters of the feature extractor in the TIDA stage, and then input the source domain data in the task-related data into the RTDGN for fine-tuning training; S5: After multiple iterations of training, observe whether the loss curve tends to converge steadily. If the curve converges, the RTDGN training is completed. The trained RTDGN will be used for cross-condition migration fault diagnosis under the relational migration task of rotating mechanical parts; In step S2, building a relational migration framework specifically includes: first defining task-irrelevant auxiliary data It is composed of multiple auxiliary domains. Where Q represents the number of auxiliary domains, and denotes the i-th sample of the q-th auxiliary domain and its corresponding label vector, n q is the number of samples in the qth auxiliary domain; task-related data It consists of source domain data and target domain data, defined as in and denote the labeled source domain and the unseen unlabeled target domain, respectively. Respectively represent the i-th sample in the source domain and its corresponding label vector, n S is the number of samples in the source domain, represents the i-th sample of the unlabeled target domain and its corresponding label vector, n T is the number of samples in the target domain; Define the relationship between any two unrelated auxiliary domains of the task as where i,j∈[1,2,…,Q] and i≠j; φ I (·) represents the relationship mapping between task-irrelevant auxiliary data, ΔP ij represents the distribution difference between the ith auxiliary domain and the jth auxiliary domain, when ΔP ij = 0, Similarly, the relationship between the source and target domains of task-related data is defined as Among them, φ R (·) represents the relationship mapping between task-related data, ΔP ST represents the distribution difference between the source domain and the unlabeled target domain; according to the similarity of the distribution difference between the task-irrelevant auxiliary data and the task-relevant data, that is, φ R ≈φ I , and then migrate the distribution difference relationship of task-irrelevant auxiliary data to task-relevant data through relationship migration, thereby indirectly achieving the measurement and elimination of the distribution difference between the invisible target domain and the source domain; In step S3, the construction of RTDGN includes two stages, TIDA and TRDG. Both stages use convolutional neural networks as the backbone network to extract features. In the TIDA stage, a general feature extractor with domain confusion and separability is constructed based on task-irrelevant auxiliary data. Then, the distribution relationship knowledge learned by the feature extractor in the TIDA stage is transferred to the TRDG stage through the relational transfer framework to achieve domain confusion. In the backbone network, a batch normalization layer is used to accelerate the convergence of the neural network and avoid variance drift. In addition, a Dropout layer and global average pooling are used to prevent network overfitting. The hidden layer uses ReLU as the activation function, and the output layer uses Softmax and Sigmoid activation functions. In the TIDA stage, an adversarial network is constructed, including a feature extractor G FE (θ FE ), a fault classifier G FC (θ FC ) and K binary classification domain discriminators where k∈[1,2…,K] and represents the number of combinations of Q elements, θ FE ,θ FC and denote the weight parameters of the feature extractor, fault classifier, and domain discriminator respectively; In the TRDG stage, a domain generalization diagnosis network is constructed, including the feature extractor G FE (θ FE ) and residual fusion classifier G RC (θ RC ); The structure and weight parameters of the feature extractor in the TRDG stage are the same as those in the TIDA stage; the residual fusion classifier is a fault classifier that fuses multiple features through residual connections.

2. The intelligent diagnosis method for mechanical failure according to claim 1, characterized in that: In step S4, in the TIDA stage, the weight parameters of the constructed adversarial network are iteratively trained, specifically including: using the Adam optimizer and the domain loss L D , classification loss L C , inverse entropy loss L IE , the weight parameter θ of the adversarial network in the TIDA stage FE ,θ FC and Iteratively updated: Among them, ε0 represents the learning rate in the TIDA phase, μ and λ are trade-off parameters, and θ D represents the weight parameters of all domain classifiers, and In order to directly perform gradient parameter updates, the adversarial network introduces a gradient reversal layer.

3. The intelligent diagnosis method for mechanical failure according to claim 2 is characterized in that: Domain loss L D Defined as: Where d represents the one-hot encoded vector of the domain label, represents the k-th binary cross entropy loss; Classification loss L C Defined as: Among them, Y represents the label vector corresponding to the auxiliary domain sample, represents the multi-classification cross entropy loss of the qth auxiliary domain; Inverse entropy loss L IE Defined as: in, 4. The intelligent diagnosis method for mechanical failure according to claim 1, characterized in that: In step S4, the weight parameter θ of the TRDG stage FE ,θ RC The update method is as follows: Among them, ε1 and ε2 represent the learning rate, satisfying ε1<ε2; Represents the multi-class cross entropy loss of the source domain.

5. The intelligent diagnosis method for mechanical failure according to claim 4 is characterized in that: Multi-class cross entropy loss in source domain Defined as: where Y S Represents the label vector corresponding to the source domain sample.

Images