An Open Set Domain Generalization Method for Bogie Fault Diagnosis

Abstract

This invention relates to the field of fault diagnosis technology, and in particular to an open-domain generalization method for bogie fault diagnosis. The method includes: acquiring the original vibration signal of the bogie's axle box bearing, performing data preprocessing and fast Fourier transform to obtain the vibration signal spectrum; dividing the dataset into a source domain and a target domain; constructing an open-domain generalization model based on an ellipsoidal prototype, including a feature extraction module, a domain confusion classification module, a fault classification module, and an ellipsoidal prototype metric module. The source domain data is input into the open-domain generalization model, and iterative training is performed using the backpropagation algorithm to obtain model parameters and the prototype, covariance matrix, and boundary threshold for each category; then, the fault type of the target domain sample is determined based on the known ellipsoidal boundary threshold. This invention constructs an ellipsoidal decision boundary for each category using Mahalanobis distance metric, providing a more flexible and universal method for identifying unknown faults under unknown operating conditions.

Description

Technical Field

[0001] This invention relates to the technical field of bogie fault diagnosis methods, and in particular to an open set domain generalization method for bogie fault diagnosis. Background Technology

[0002] The bogie is a critical component of high-speed railway vehicles, and its health directly affects the safety and stability of vehicle operation. However, due to the long-term operation of mechanical equipment, components such as axle box bearings, treads, and gearboxes in the bogie are prone to failure. Therefore, fault diagnosis research on bogies is particularly important.

[0003] Traditional deep learning-based fault diagnosis methods typically rely on sufficient labeled samples and assume that the training and testing data operate under identical conditions and have the same fault types. Therefore, traditional deep learning fault diagnosis models often exhibit poor generalization ability when faced with unknown fault types or different operating conditions, making it difficult to meet the diverse needs of actual production. Existing transfer learning-based fault diagnosis methods, using fine-tuning, domain feature alignment, and adversarial training, have addressed the issue of inconsistent operating conditions between training and testing data to some extent. However, they still require data from the target domain to participate in training and guide the transfer process. In real-world environments, target domain data cannot be obtained in advance; the training phase can only utilize multiple known source domains, and the target domain data may contain previously unseen fault types. This makes traditional transfer learning-based fault diagnosis methods equally ill-suited to real-world environments, resulting in poor generalization ability. Therefore, accurately identifying unknown faults in bogies under unknown operating conditions is a pressing problem that needs to be solved. Summary of the Invention

[0004] The purpose of this invention is to provide an open set domain generalization method for bogie fault diagnosis, which provides a new method for accurately identifying unknown faults under unknown operating conditions.

[0005] The open-domain generalization method for bogie fault diagnosis proposed in this invention includes the following steps:

[0006] S1: The original vibration signals of the axle box bearings in the bogie wheelset system under different operating conditions are collected using an accelerometer. The vibration signals are preprocessed and subjected to fast Fourier transform to obtain the spectrum of the vibration signals.

[0007] S2: Divide dataset D into source domain datasets according to different operating conditions. Target domain dataset Where, x i For source domain dataset The i-th sample, y i for The i-th fault label, z i for The i-th domain label, Let S represent the total number of samples in the source domain, M represent the number of operating conditions in the source domain, and x represent the total number of samples in the source domain. j For the target domain dataset D T The j-th sample, y j D T The j-th fault label, z j D T The j-th domain label, N T This represents the total number of samples in the target domain T;

[0008] S3: Construct an open-domain generalization model based on an ellipsoidal prototype, namely the EP-OSDG model, to identify unknown faults in unknown domains; the EP-OSDG model consists of a feature extraction module f φ (·), Domain Confusion Classification Module H θ (·), Fault Classification Module It consists of an ellipsoidal prototype measurement module. Among them, the fault classification module helps the feature extraction module extract fault category features, the domain confusion classification module helps the feature extraction module extract domain-invariant features from multiple source domain data through adversarial training, and the ellipsoidal measurement module constructs the ellipsoidal decision boundary for each category based on the prototype idea, providing a more flexible and general method for identifying unknown faults;

[0009] S4: During the training phase, the source domain dataset... enter

[0010] In the EP-OSDG model, the domain confusion loss L obtained from the domain confusion classification module is... domain The classification loss L obtained from the fault classification module class And the ellipsoidal metric loss L obtained from the ellipsoid prototype metric module ellipsoid Adding them together, we get the total loss function value L of the model. total Iterative training is performed using the backpropagation algorithm to obtain the parameters of the feature extraction module and the prototype vector p for each category. c , covariance matrix ∑ c and boundary threshold r c ;

[0011] S5: During the testing phase, the target domain dataset will be... Input the trained feature extraction module f φ (·), and then based on the known prototype vector p c , covariance matrix ∑ c and boundary threshold r c Determine the fault type of the target domain sample. When sample x j With prototype vector pc The square of the Mahalanobis distance is less than or equal to the ellipsoid boundary threshold r. c When, sample x j The category is c, and conversely, the category of sample x is c. j This refers to a fault that has not been previously observed.

[0012] The beneficial effects of adopting the above technical solution are as follows: This invention constructs an open-domain generalization method for bogie fault diagnosis, which can identify unknown faults under unknown working conditions, effectively improving the intelligence and practicality of the diagnostic method. Based on the idea of ​​an ellipsoidal prototype, this invention uses Mahalanobis distance to construct an asymmetric ellipsoidal decision boundary, considering the correlation and variance of various dimensions of sample features, and can adapt to different scales and correlations of features in the data. Compared with the symmetric boundary constructed based on Euclidean distance using a hypersphere prototype, the ellipsoidal prototype has higher flexibility and applicability. The domain confusion classification module, through adversarial training, enables the model to learn domain-invariant features from multiple source domains, thereby enhancing the model's generalization ability and enabling it to better adapt to new and unseen domains. The fault classification module, as an auxiliary task, helps the feature extraction module extract features conducive to fault classification, which is beneficial for generating a more accurate ellipsoidal decision boundary. Attached Figure Description

[0013] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0014] Figure 1 This is a flowchart of the fault diagnosis process of the present invention;

[0015] Figure 2 This is a diagram of the open set-domain generalized fault diagnosis model of the present invention; Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0017] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0018] like Figure 1As shown, this invention discloses an open-domain generalization method for bogie fault diagnosis, which specifically includes the following steps:

[0019] S1: The original vibration signals of the axle box bearings in the bogie wheelset system under different operating conditions are collected using an accelerometer. The vibration signals are preprocessed and subjected to fast Fourier transform to obtain the spectrum of the vibration signals.

[0020] In this embodiment, the raw vibration signals of the bogie axle box bearings under different operating conditions are first collected using an accelerometer. Then, the vibration signals undergo data preprocessing, including Z-score normalization and sliding window slicing, with the sliding window size set to 2048 and the step size to 512. Finally, the spectrum of the vibration signal is obtained through a fast Fourier transform, with a spectrum length of 1024.

[0021] S2: Divide the dataset D into source domains according to different operating conditions. target domain The source domain contains data under multiple operating conditions, i.e., multiple source domains;

[0022] This embodiment uses the Rolling Bearing Test Dataset (CWRU) from Case Western Reserve University, USA, for experiments. The data focuses on the drive-end bearing, model 6205-2RS JEM SKF, with a sampling frequency of 12kHz. Bearing fault types are categorized into inner ring fault (IR), rolling element fault (B), outer ring six o'clock direction fault (OR), and normal data (N). Each load type includes three fault types and one normal data type, for a total of four data categories. The experimental dataset contains four load types: 0hp, 1hp, 2hp, and 3hp. The CWRU dataset operating condition information is shown in Table 1, and the CWRU dataset category information is shown in Table 2.

[0023] Table 1 - CWRU Dataset Operating Condition Information

[0024] Load (HP) 0 1 2 3 Rotational speed (r / min) 1797 1772 1750 1730

[0025] Table 2 - CWRU Dataset Category Information

[0026] Label 0 1 2 3 Fault type N B IR OR

[0027] To verify the effectiveness of the proposed EP-OSDG model in identifying unknown faults under unknown operating conditions, an open-domain generalized fault diagnosis experiment was designed. The source domain operating conditions were loads of 0HP, 1HP, and 2HP, and the target domain operating condition was load 3HP. Each source domain operating condition contained only samples labeled {0, 1, 2}, while the target domain operating condition (3HP) contained samples labeled {0, 1, 2, 3}. Furthermore, 100 samples were selected for training or testing for each fault type under each operating condition.

[0028] S3: Construct an open set domain generalization model based on the ellipsoid prototype, namely the EP-OSDG model, to identify unknown faults in unknown domains;

[0029] like Figure 2 As shown, the open-set domain generalized fault diagnosis model consists of a feature extraction module, a domain confusion classification module, a fault classification module, and an ellipsoid prototype metric module. The fault classification module helps the feature extraction module extract fault category features. The domain confusion classification module, through adversarial training, helps the feature extraction module extract domain-invariant features from multiple source domain data. The ellipsoid metric module, based on the prototype concept, constructs the ellipsoidal decision boundary for each category, providing a more flexible and general method for identifying unknown faults.

[0030] The feature extraction module has 10 layers, including 5 convolutional layers and 5 pooling layers. Layers 1, 3, 5, 7, and 9 are convolutional layers, layers 2, 4, 6, and 8 are average pooling layers, and layer 10 is an adaptive average pooling layer.

[0031] The domain confusion classification module consists of 6 layers, including 2 convolutional layers, 2 pooling layers, and 2 fully connected layers. Layers 1 and 3 are convolutional layers, layers 2 and 4 are average pooling layers, and layers 5 and 6 are fully connected layers.

[0032] The fault classification module consists of 6 layers, including 2 convolutional layers, 2 pooling layers, and 2 fully connected layers. Layers 1 and 3 are convolutional layers, layers 2 and 4 are average pooling layers, and layers 5 and 6 are fully connected layers.

[0033] The ellipsoid measurement module is defined as follows:

[0034] First, the formula for an ellipsoid expressed by Mahalanobis distance is as follows:

[0035]

[0036] Among them, X i It is sample x i ∑ is the eigenvector of , where μ is the mean vector. -1 It is the inverse of the covariance matrix, defining the shape and orientation of the ellipsoid. r0 is a constant that controls the size of the ellipsoid.

[0037] Then, based on multiple source domains Category C Sample feature set The mean vector μ of category c is calculated. c The sum of the covariance matrix ∑ c , where the mean μ c That is, the prototype vector p c , The number of samples of class c in the source domain S;

[0038] prototype vector p c Calculation formula:

[0039]

[0040] covariance matrix ∑ c Calculation formula:

[0041]

[0042] Secondly, based on the obtained prototype vector p c With covariance matrix ∑ c Calculate the Mahalanobis distance M(p) between the features of this category of samples and the prototype vector. c ,X i ):

[0043]

[0044] The maximum Mahalanobis distance between all sample features of this category and the prototype vector is used as the boundary threshold r. c An ellipsoidal decision boundary is formed, such that points whose Mahalanobis distance is less than the threshold represent samples of that class.

[0045] Finally, we obtain the prototype vector p representing this category. c With the covariance matrix ∑ c With boundary threshold r c .

[0046] In some embodiments, the network parameter settings for the feature extraction module, domain confusion classification module, and fault classification module in the open set domain generalized fault diagnosis model are shown in Tables 3, 4, and 5, respectively.

[0047] Table 3 - Feature Extraction Module Parameter Settings

[0048]

[0049] Table 4. Parameter Settings for Domain Confusion Classification Module

[0050]

[0051] Table 5 Fault Classification Module Parameter Settings

[0052]

[0053] S4: During the training phase, the source domain dataset... enter

[0054] In the EP-OSDG model, the domain confusion loss L obtained from the domain confusion classification module is... domainThe classification loss L obtained from the fault classification module class And the ellipsoidal metric loss L obtained from the ellipsoid prototype metric module ellipsoid Adding them together, we get the total loss function value L of the model. total Iterative training is performed using the backpropagation algorithm to obtain the parameters of the feature extraction module and the prototype vector p for each category. c , covariance matrix ∑ c and boundary threshold r c ;

[0055] The training process of the EP-OSDG model is as follows:

[0056] (1) Source domain data Input to feature extraction module f φ (·), to obtain the feature vector X of the source domain sample. i =f φ (x i );

[0057] (2) All source domain sample feature vectors X i Input the fault classification module to calculate the fault classification loss L. class ;

[0058] (3) All source domain sample feature vectors X i The input domain confusion classification module, feature extraction module, and domain confusion classification module are trained adversarially. The domain classifier attempts to accurately distinguish samples from different source domains, while the feature extractor attempts to obfuscate the differences between these source domains to extract domain-invariant features. Finally, the domain confusion loss L is calculated. domain ;

[0059] (4) All source domain sample feature vectors X i Input the ellipsoid metric module to calculate the ellipsoid metric loss L. ellipsoid And the prototype vector p for each category c , covariance matrix ∑ c and boundary threshold r c ;

[0060] (5) The fault classification loss L of the source domain sample features class Domain confusion loss L domain and ellipsoidal metric loss L ellipsoid The weighted sum is used as the total loss function value L. total Iterative training is performed using the backpropagation algorithm (BP) until the model converges, yielding the model parameters and the prototype vector p for each class. c , covariance matrix ∑ c and boundary threshold r c .

[0061] The fault classification loss L obtained by the fault classification module class for:

[0062]

[0063] In the formula: Let S be the total number of samples in the source domain S, M be the number of operating conditions in the source domain, and y be the total number of samples in the source domain S. ic Indicates sample x i One-hot encoding of the real category, when x i When the category is c, y ic =1, otherwise 0. Indicates sample x i The predicted probability of belonging to category c;

[0064] The domain confusion loss L obtained by the domain confusion classification module domain for:

[0065]

[0066] In the formula: z im Indicates sample x i One-hot encoding of the real source domain, when x i When the domain label is m, z im =1, otherwise 0. Indicates sample x i The predicted probability of belonging to domain label m;

[0067] The ellipsoid measurement loss L obtained by the ellipsoid measurement module ellipsoid for:

[0068]

[0069] In the formula, C represents the number of categories, and λ is a hyperparameter. It is a regularization term, r c L represents the boundary threshold for category c. intra It is the intra-class loss function, L inter It is the inter-class loss function.

[0070] Intra-class loss function L intra The calculation formula is as follows:

[0071]

[0072] In the formula: express These are samples belonging to the same category as class c; Indicates sample The distance to the prototype of category c, α controls the sensitivity to the distance gap;

[0073] It is a penalty term, when the sample distance is greater than r. c At that time, increase the punishment. It is a reward item when the sample distance is less than r. c When the time comes, a reward will be given. intra It is a hyperparameter that controls the intensity of the reward;

[0074] It is an indicator function, when When the value is 1;

[0075] Inter-class loss function L inter The calculation formula is as follows:

[0076]

[0077] In the formula: express Samples of class c are different from those of class c; Indicates sample The distance to the prototype of category c, -β is used to make it more sensitive to differences in distance to negative samples.

[0078] The formula for calculating the total loss function is as follows:

[0079] L total =ωL class +ξL domain +L ellipsoid

[0080] In the formula: ω,ξ are the hyperparameters of fault classification loss and domain confusion loss.

[0081] S5: During the testing phase, the target domain dataset will be... Input the trained feature extraction module f φ (·), and then based on the known prototype vector p c , covariance matrix ∑ c and boundary threshold r c Determine the fault type of the target domain samples. When the feature vector X j With prototype vector p c The square of the Mahalanobis distance is less than or equal to the ellipsoid boundary threshold r. c When, the eigenvector X j The category is c, and conversely, the feature vector X j This refers to a fault that has not been previously observed.

[0082] The specific process of using the trained EP-OSDG model for open set domain generalized fault diagnosis is as follows:

[0083] (1) Target domain data Input to feature extraction module f φ(·), to obtain the feature representation X of the target domain sample. j =f φ (x j );

[0084] (2) The target domain sample feature vector X j Input the ellipsoidal prototype metric module to compute the prototype vector p for each sample feature. c Mahalanobis distance M(p) c ,X i ):

[0085]

[0086] Then, a judgment is made based on the discrimination criteria:

[0087] If M(p) c ,X j )≤r c Then the eigenvector X is considered to be j Inside the ellipsoid, it belongs to category c.

[0088] If M(p) c ,X i )>r c Then the eigenvector X is considered to be j Outside the ellipsoid, i.e., not belonging to category c.

[0089] To demonstrate the effectiveness of the EP-OSDG model, the HP-OSDG method, an open set-domain generalization method based on the hypersphere prototype, was selected and compared with the EP-OSDG method proposed in this invention. The fault diagnosis accuracy of the HP-OSDG method was 84.95%, while that of the EP-OSDG method was 95%. Experimental results show that the EP-OSDG method proposed in this invention has better diagnostic performance.

Claims

1. An open-domain generalization method for bogie fault diagnosis, characterized in that... Includes the following steps: S1: The original vibration signals of the axle box bearings in the bogie wheelset system under different operating conditions are collected using an accelerometer. The vibration signals are preprocessed and subjected to fast Fourier transform to obtain the spectrum of the vibration signals. S2: Divide dataset D into source domain datasets according to different operating conditions. Target domain dataset Where, x i For source domain dataset The i-th sample, y i for The i-th fault label, z i for The i-th domain label, Let S represent the total number of samples in the source domain, M represent the number of operating conditions in the source domain, and x represent the total number of samples in the source domain. j For the target domain dataset D T The j-th sample, y j D T The j-th fault label, z j D T The j-th domain label, N T This represents the total number of samples in the target domain T; S3: Construct an open-domain generalization model based on an ellipsoidal prototype, namely the EP-OSDG model, to identify unknown faults in unknown domains; the EP-OSDG model consists of a feature extraction module f φ (·), Domain Confusion Classification Module H θ (·), Fault Classification Module and the ellipsoid prototype measurement module E δ (·) consists of: the fault classification module helps the feature extraction module extract fault category features; the domain confusion classification module, through adversarial training, helps the feature extraction module extract domain-invariant features from multiple source domain data; and the ellipsoid metric module, based on the prototype idea, constructs the ellipsoidal decision boundary for each category, providing a more flexible and universal method for identifying unknown faults. S4: During the training phase, the source domain dataset... Inputting the domain confusion loss L obtained from the domain confusion classification module into the EP-OSDG model domain The classification loss L obtained from the fault classification module class And the ellipsoidal metric loss L obtained from the ellipsoid prototype metric module ellipsoid Adding them together, we get the total loss function value L of the model. total Iterative training using the backpropagation algorithm is performed to obtain the feature extraction module f. φ The parameters of (·) and the prototype vector p for each category c , covariance matrix ∑ c and boundary threshold r c ; S5: During the testing phase, the target domain dataset will be... Input the trained feature extraction module f φ (·), and then based on the known prototype vector p c , covariance matrix ∑ c and boundary threshold r c Determine the fault type of the target domain sample; when sample x j With prototype vector p c The square of the Mahalanobis distance is less than or equal to the ellipsoid boundary threshold r. c When, sample x j The category is c, and conversely, the category of sample x is c. j This refers to a fault that has not been previously observed.

2. The open-domain generalization method for bogie fault diagnosis as described in claim 1, characterized in that, Step S1 specifically includes the following steps: S1-1: Use an accelerometer to collect the original vibration signals of the axle box bearing under different operating conditions; S1-2: Z-score normalization of the vibration signal reduces the distribution differences between different operating conditions. The mathematical formula is shown below: in, σ is the mean of the original data, and σ is the standard deviation of the original data. S1-3: The normalized vibration signal is sliced ​​into sliding window segments. The sliding window size is set to 2L and the step size is I. Each segment of the vibration signal is a sample. S1-4: The spectrum of the vibration signal is obtained by fast Fourier transform, and the spectrum length is L.

3. The open-domain generalization method for bogie fault diagnosis as described in claim 1, characterized in that: The feature extraction module in the EP-OSDG model has a total of 10 layers, including 5 convolutional layers and 5 pooling layers. Layers 1, 3, 5, 7, and 9 are convolutional layers, layers 2, 4, 6, and 8 are average pooling layers, and layer 10 is an adaptive average pooling layer. The domain confusion classification module has a total of 6 layers, including 2 convolutional layers, 2 pooling layers and 2 fully connected layers. The first and third layers are convolutional layers, the second and fourth layers are average pooling layers, and the fifth and sixth layers are fully connected layers. The fault classification module has a total of 6 layers, including 2 convolutional layers, 2 pooling layers and 2 fully connected layers. The first and third layers are convolutional layers, the second and fourth layers are average pooling layers, and the fifth and sixth layers are fully connected layers. The ellipsoid measurement module is defined as follows: First, the formula for an ellipsoid expressed by Mahalanobis distance is as follows: Among them, X i It is sample x i The eigenvectors of , μ is the mean vector, and ∑ -1 It is the inverse of the covariance matrix, which defines the shape and orientation of the ellipsoid, and r0 is a constant that controls the size of the ellipsoid; Then, based on the source domain dataset Category C Sample feature set The mean vector μ of category c is calculated. c The sum of the covariance matrix ∑ c , where the mean μ c That is, the prototype vector p c , The number of samples of class c in the source domain S; prototype vector p c Calculation formula: covariance matrix ∑ c Calculation formula: Secondly, based on the obtained prototype vector p c With the covariance matrix ∑ c Calculate the Mahalanobis distance M(p) between the features of this category of samples and the prototype vector. c ,X i ): The maximum Mahalanobis distance between all sample features of this category and the prototype vector is used as the boundary threshold r. c This forms an ellipsoidal decision boundary, such that points with a Mahalanobis distance less than the threshold represent samples of that class. Finally, we obtain the prototype vector p representing this category. c With the covariance matrix ∑ c With boundary threshold r c .

4. The open-domain generalization method for bogie fault diagnosis as described in claim 1, characterized in that: The fault classification loss L obtained by the fault classification module class for: In the formula: Let S be the total number of samples in the source domain S, M be the number of operating conditions in the source domain, and y be the total number of samples in the source domain S. ic Indicates sample x i One-hot encoding of the real category, when x i When the category is c, y ic =1, otherwise 0. Indicates sample x i The predicted probability of belonging to category c; The domain confusion loss L obtained by the domain confusion classification module domain for: In the formula: z im Indicates sample x i One-hot encoding of the real source domain, when x i When the domain label is m, z im =1, otherwise 0. Indicates sample x i The predicted probability of belonging to domain label m; The ellipsoid measurement loss L obtained by the ellipsoid measurement module ellipsoid for: In the formula, C represents the number of categories, and λ is a hyperparameter. It is a regularization term, r c L is the boundary threshold for category c. intra It is the intra-class loss function, L inter It is the inter-class loss function; Intra-class loss function L intra The calculation formula is as follows: In the formula: express These are samples belonging to the same category as class c; Indicates sample The distance to the prototype of category c, α controls the sensitivity to the distance difference. It is a penalty term, when the sample distance is greater than r. c At that time, increase the punishment. It is a reward item when the sample distance is less than r. c When the time comes, a reward will be given. intra It is a hyperparameter that controls the intensity of the reward. It is an indicator function, when When the value is 1; Inter-class loss function L inter The calculation formula is as follows: In the formula: express Samples of a different class than class c Indicates sample The distance to the prototype of category c, -β is used to make it more sensitive to differences in distance to negative samples; The formula for calculating the total loss function is as follows: L total =ωL class +ξL domain +L ellipsoid In the formula: ω,ξ are the hyperparameters of fault classification loss and domain confusion loss.

5. The open-domain generalization method for bogie fault diagnosis as described in claim 1, characterized in that, The training process of the EP-OSDG model includes the following steps: S4-1: Transfer the source domain dataset Input to feature extraction module f φ (·), to obtain the feature vector X of the source domain sample. i =f φ (x i ); S4-2: Combine the feature vectors X of all source domain samples i Input the fault classification module to calculate the fault classification loss L. class ; S4-3: Combine the feature vectors X of all source domain samples i The input domain confusion classification module, feature extraction module, and domain confusion classification module are trained adversarially. The domain classifier attempts to accurately distinguish samples from different source domains, while the feature extractor attempts to obfuscate the differences between these source domains to extract domain-invariant features. Finally, the domain confusion loss L is calculated. domain ; S4-4: Combine the feature vectors X of all source domain samples i Input the ellipsoid metric module to calculate the ellipsoid metric loss L. ellipsoid And the prototype vector p for each category c , covariance matrix ∑ c and boundary threshold r c ; S4-5: Fault classification loss L of source domain sample features class Domain confusion loss L domain and ellipsoidal metric loss L ellipsoid The weighted sum is used as the total loss function value L. total Iterative training is performed using the backpropagation algorithm (BP) until the model converges, yielding the model parameters and the prototype vector p for each class. c , covariance matrix ∑ c and boundary threshold r c .

6. The open-domain generalization method for bogie fault diagnosis as described in claim 1, characterized in that, The testing process for the EP-OSDG model includes the following steps: S5-1: Target domain dataset Input to feature extraction module f φ (·), to obtain the feature vector X of the target domain sample. j =f φ (x j ); S5-2: The target domain sample feature vector X j Input the ellipsoidal prototype metric module to compute the prototype vector p for each sample feature. c Mahalanobis distance M(p) c ,X j ): Then, a judgment is made based on the discrimination criteria: If M(p) c ,X j )≤r c Then the eigenvector X is considered to be j Inside the ellipsoid, it belongs to category c; If M(p) c ,X i )>r c Then the eigenvector X is considered to be j Outside the ellipsoid, i.e., not belonging to category c.

Images