Traction motor fault sample generation method and system

Abstract

The invention relates to the technical field of fault sample generation, and discloses a traction motor fault sample generation method and system, and the method achieves the correlation between a potential vector and the fault semantics of a traction motor through the learning of the semantics of the traction motor. Through decoupling of attribute potential vectors, establishment of a one-to-one correspondence relationship between the attribute potential vectors and fault semantics, effective learning of real semantics of fault samples, realization of controllable sample generation through editing of the attribute potential vectors, generation of data of preset fault types, severity and operation conditions, and realization of zero sample generation. The method is easy to implement, does not need additional hardware equipment, and can improve the running state health monitoring level of the motor and the rail transit vehicle.

Description

Technical Field

[0001] This invention belongs to the field of fault sample generation technology, and particularly relates to a method and system for generating fault samples of traction motors. Background Technology As modern industrial production equipment continues to evolve towards structured, automated, and intelligent systems, electric motors, as the most crucial power and drive devices, are widely used in high-speed trains, subways, and intercity rail transit. Traction motors are a major source of frequent train malfunctions, and their proper functioning directly impacts train safety. Because rail trains employ a fail-safe mechanism—meaning they are quickly stopped upon fault detection—and because faults themselves are low-probability events, most data is collected under healthy operating conditions. Furthermore, physical experiments can only obtain incomplete data on a small number of fault types and single severity levels under specific operating conditions. Training diagnostic models with incomplete fault datasets significantly impacts generalization performance and fault diagnosis accuracy. However, fault data generation can obtain unseen fault data with the same distribution as real-world data, improving the completeness of the fault dataset and providing a feasible solution for deep learning to handle incomplete fault datasets.

[0002] Currently, existing fault sample generation methods are black-box models, which make it difficult to establish the correspondence between the true semantics of faults and potential vectors, and cannot achieve zero-sample generation by editing potential vectors to improve the completeness of fault datasets. Summary of the Invention

[0003] This invention proposes a method and system for generating traction motor fault samples to address the problem that existing methods struggle to establish the correspondence between the true semantics of faults and potential vectors, and are unable to achieve zero-sample generation by editing potential vectors to improve the completeness of fault datasets.

[0004] To achieve the above objectives, the present invention employs the following technical solution: In a first aspect, the present invention provides a method for generating traction motor fault samples, comprising: S1: collecting traction motor current signals under different operating conditions, obtaining corresponding attribute labels, constructing the collected samples into a total sample set, and dividing the total sample set into a training sample set and a test sample set; S2: Construct an encoder model, which includes an encoder and a decoder; S3: Based on the encoder, the training sample set is encoded into a low-dimensional latent vector. Based on the decoder and the low-dimensional latent vector, the samples are reconstructed to obtain reconstructed samples, and a reconstruction loss function is constructed. S4: Find the unique overlapping attribute sample pair and all overlapping attribute sample pairs in the training sample set; divide the low-dimensional latent vector into N types of attribute latent vectors, exchange the attribute latent vectors of the unique overlapping attribute sample pair and all overlapping attribute sample pairs in a set manner, and construct attribute supervision constraint terms. S5: Estimate the upper bound of mutual information between pairwise attribute latent vectors, introduce total correlation, and decouple attribute latent vectors to minimize total correlation to obtain total correlation constraints. S6: Construct convex path loss functions for degree, rotational speed, load, and other attribute latent vectors in sequence, and construct a convex path loss function for the total attribute latent vector; construct the total loss function based on the reconstruction loss function, attribute supervision constraints, total correlation constraints, and convex path loss function. S7: Using the total loss function as the objective, train the optimal model parameters, use the optimal model parameters as the target model, and test the target model through the test sample set. After the test is passed, input the real-time collected traction motor current signal into the encoder model of the target model to obtain the corresponding latent vector. Perform attribute segmentation on the latent vector and reorganize it to obtain a new latent vector, which is then input into the decoder model of the target model to achieve zero-sample generation.

[0005] Secondly, this application also provides a traction motor fault sample generation system, including a memory, a processor, and a computer program and a large language model stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described in the first aspect above.

[0006] Beneficial effects: This invention provides a method for generating traction motor fault samples. By constructing a signal encoder model, a text encoder model, and a signal-text embedding alignment comprehensive loss function, semantic association between sensor signals and natural language can be achieved. Combined with a knowledge graph in the motor domain and a large language model, zero-sample generation is achieved by editing latent vectors. This method can make effective and interpretable operation and maintenance decisions for motors under complex operating conditions. The method is easy to implement, requires no additional hardware equipment, and can improve the health monitoring level of motors and rail transit vehicles.

[0007] In a further technical solution, by estimating the upper bound of the mutual information between pairwise attribute latent vectors and minimizing the mutual information, the attribute latent vectors are decoupled, establishing a one-to-one correspondence between the attribute latent vectors and the fault semantics. This effectively learns the true semantics of fault samples. Controllable sample generation can be achieved by editing the attribute latent vectors, generating data with preset fault types, severity, and operating conditions, achieving zero-sample generation, and improving the completeness of fault samples. Therefore, the generalization ability and diagnostic accuracy of the fault diagnosis model are improved. Attached Figure Description

[0008] Figure 1 This is one of the flowcharts of a traction motor fault sample generation method in an embodiment of the present invention; Figure 2 This is the second flowchart of a method for generating traction motor fault samples in an embodiment of the present invention; Figure 3 This is the result of fault data generation using a traction motor fault sample generation method in an embodiment of the present invention. Detailed Implementation

[0009] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0010] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms "an" or "a," and similar terms, do not indicate a quantity limitation, but rather indicate the presence of at least one.

[0011] Please see Figure 1 The present application provides a method for generating traction motor fault samples, comprising: S1: collecting traction motor current signals under different operating conditions, obtaining corresponding attribute labels, constructing the collected samples into a total sample set, and dividing the total sample set into a training sample set and a test sample set; S2: Construct an encoder model, which includes an encoder and a decoder; S3: Based on the encoder, the training sample set is encoded into a low-dimensional latent vector. Based on the decoder and the low-dimensional latent vector, the samples are reconstructed to obtain reconstructed samples, and a reconstruction loss function is constructed. S4: Find the unique overlapping attribute sample pair and all overlapping attribute sample pairs in the training sample set; divide the low-dimensional latent vector into N types of attribute latent vectors, exchange the attribute latent vectors of the unique overlapping attribute sample pair and all overlapping attribute sample pairs in a set manner, and construct attribute supervision constraint terms. S5: Estimate the upper bound of mutual information between pairwise attribute latent vectors, introduce total correlation, and decouple attribute latent vectors to minimize total correlation to obtain total correlation constraints. S6: Construct convex path loss functions for the potential vectors of degree, rotation speed, and load attributes in sequence, and construct the convex path loss function for the total potential vector of attributes; S7: Construct constant-speed loss functions for the potential vectors of degree, rotation speed, and load attributes in sequence, and construct a constant-speed loss function for the total potential vector of attributes; construct the total loss function based on the reconstruction loss function, attribute supervision constraints, total correlation constraints, convex path loss, and constant-speed loss function. S8: Using the total loss function as the objective, train the optimal model parameters, use the optimal model parameters as the target model, and test the target model through the test sample set. After the test is passed, input the real-time collected traction motor current signal into the encoder model of the target model to obtain the corresponding latent vector. Perform attribute segmentation on the latent vector and reorganize the new latent vector to input it into the decoder model of the target model to achieve zero-sample generation.

[0012] The aforementioned traction motor fault sample generation method, by constructing a signal encoder model, a text encoder model, and a signal-text embedding alignment comprehensive loss function, can realize the semantic association between sensor signals and natural language. Combined with a knowledge graph and a large language model in the motor domain, it achieves zero-sample generation by editing latent vectors. This method can make effective and interpretable operation and maintenance decisions for motors under complex operating conditions. The method is easy to implement, requires no additional hardware equipment, and can improve the health monitoring level of motors and rail transit vehicles.

[0013] Optionally, S1 includes: S11: Collect data under different categories, degrees, speeds, and loads. Traction motor current signal ,in, m Set it to a constant and obtain its corresponding label. The types of attribute tags include category, degree, speed, load, and others. , Indicates sample Length, Represent the sample space; These are the real samples and their corresponding labels. , , , These are the attribute label values ​​for category, degree, speed, and load, respectively; these values ​​are discrete. The attribute label values ​​for "other" are... This is set to the default value because the noise level and initial phase of each sample are random and arbitrarily determined, making it impossible to obtain the exact attribute label value.

[0014] S12: Construct the overall sample set based on the collected samples. The sample set was then randomly divided in a 3:2 ratio. Divided into training set and test set .in,( This represents the traction motor current signal and its corresponding attribute label in the training sample set. This represents the traction motor current signal and its corresponding attribute label in the test sample set.

[0015] Optionally, S3 includes: S31: Samples from the training dataset Input encoder This yields the low-dimensional latent vector projected onto the latent space. : (1) It is a low-dimensional latent vector with dimension O(n). ,and , Represent the latent vector space; the decoder receives the latent vectors. To reconstruct the original input sample, the output obtained is the reconstructed sample. Reconstructing samples and input samples In the same sample space, that is ; S22: Construct the reconstruction loss function, which can be expressed as: (2) In the formula, For reconstruction loss; Characterization decoder; All samples in the training sample set are input into the autoencoder to obtain the corresponding latent vectors and reconstructed samples.

[0016] Optionally, S4 includes: S41: Find a pair of samples that have exactly one attribute label value that are the same. This pair of samples is a unique pair of overlapping attribute samples. It satisfies the following relationship: (3) Traversing the training set Find all unique overlapping attribute sample pairs; Two samples are considered as samples with overlapping attributes if they have the same attribute label values ​​(except for "other"). They satisfy: (4) Traversing the training set Find all overlapping attribute sample pairs; S42: Segment the semantic space of the latent vectors and encode the low-dimensional latent vectors obtained by the autoencoder. Latent attribute vectors segmented into categories The latent vector of degree attributes The latent vector of rotational speed attributes The latent vector of load properties Other attribute latent vectors The latent vector of the category attribute is used as the first latent vector, the latent vector of the degree attribute is used as the second latent vector, the latent vector of the rotation speed attribute is used as the third latent vector, the latent vector of the load attribute is used as the fourth latent vector, and the other latent vectors are used as the first latent vector. The dimensions of each sub-vector satisfy the following relationship: (5) In the formula, These are the category attribute vector dimension, degree attribute vector dimension, rotation speed attribute vector dimension, load attribute vector dimension, and other attribute vector dimension, respectively. S43: Swap the latent vectors of attributes corresponding to the same attribute label values ​​in unique overlapping attribute sample pairs, keeping the remaining latent vectors unchanged. For the first... n Unique overlapping attribute sample pairs with the same attribute value For n=1, 2, 3, 4, the attribute swapping operation between them is represented as follows: (6) For unique overlapping attribute sample pairs Using the attribute value corresponding to the severity level as the second attribute value, assuming their second attribute values ​​are the same, swapping their second attribute values ​​is represented as follows: (7) In the formula, Indicates the first training sample set i The sample and the first j The latent attribute vectors of each sample are swapped after the second attribute value is swapped. All unique overlapping attribute sample pairs are traversed, and the attribute latent vectors corresponding to the same attribute label values ​​between them are swapped, while keeping the low-dimensional latent vectors of other attributes unchanged. Swap the latent vector of the 5th attribute of all overlapping attribute sample pairs, while keeping the latent vectors of the other attributes unchanged; S44: Establish a constraint function based on the consistency requirement of sample reconstruction before and after attribute swapping for unique attribute sample pairs, satisfying the following relationship: (8) In the formula, Represents the probability density function. Expressing expectations, Indicates the first training sample set i The latent vector obtained by attribute swapping of samples Indicates the first training sample set i The sample and the first j The low-dimensional latent attribute vectors of the samples are interchanged. n The latent attribute vector following the attribute value; The attribute constraint items are represented as follows: (9) In the formula, This indicates a sample swapping operation; The constraint terms under attribute supervision can be represented as a loss function, satisfying the following relationship: (10) Optionally, S5 includes: S51: Constructing a mutual information upper bound estimation function. Taking the estimation of the mutual information upper bound between the first attribute latent vector and the second attribute latent vector of the i-th sample as an example, it is expressed as follows: (11) In the formula, Indicates the upper bound of mutual information. Representation of attribute latent vector and The true distribution of conditions between them The parameter is The variational conditional distribution, using parameters Variational network learns attribute latent vectors and The true distribution of conditions between them This constraint term represents the conditional distribution and is expressed as follows: (12) Using parameters respectively , , Variational network learns attribute latent vectors and , and , and The true distribution of conditions between them; S52: Introduce total correlation to minimize the total correlation and the upper bound of mutual information of all latent attribute vectors, thereby achieving pairwise decoupling between latent attribute vectors. The total correlation constraint satisfies the following relationship: (13) In the formula, Variables representing estimates A and B The upper bound of mutual information between them, where A and B are respectively , , , ; Optionally, S6 includes: S61: Construct convex path loss functions for the latent vectors of degree, rotational speed, load, and other attributes in sequence. The convex path loss function for the latent vector of degree attribute satisfies the following relationship: (15) In the formula, Indicates the first i The latent vector of the degree attribute of each sample With the j The latent vector of the degree attribute of each sample The geodesic distance, which can be calculated using the local tangent space permutation algorithm, and the convex path loss function of the latent vector of rotational speed attribute satisfy the following relationship: (16) In the formula, Indicates the first i Latent vector of rotational speed attribute of each sample With the j Latent vector of rotational speed attribute of each sample The geodesic distance and the convex path loss function of the latent vector of load attributes satisfy the following relationship: (17) In the formula, Indicates the first i The latent vector of the load attributes of each sample With the j The latent vector of the load attributes of each sample The geodesic distance; S62: Constructing the convex path loss function for the total latent attribute vector: (19) Optionally, S7 includes: S71: A polynomial function is used as the interpolation network to fit the latent vector of the degree attribute. If a cubic function is used, the polynomial can be expressed as: (20) In the formula, , , , For about k The coefficients of the polynomial from the highest to the lowest terms; S72: Use this interpolation network as the input to the decoder, and apply the following constant-rate constraint, which can be expressed as: (twenty one) In the formula, Indicates the first training sample set i The vector obtained after inputting the degree attribute vector of each sample into the decoder. express about k The derivative of . It can be considered as along . exist k The speed at that location, and This indicates the magnitude of the velocity, corresponding to the length of the tangent vector. The expression is: (twenty two) Through constant rate constraint terms To obtain a constant-speed interpolation of a certain degree; S73: Similarly, construct a constant speed constraint term. : (twenty three) In the formula, Indicates the first training sample set i The vector obtained after inputting the rotational speed attribute vector of each sample into the decoder. express about k The derivative of . It can be considered as along . exist k The speed at that location, and This indicates the magnitude of the velocity, corresponding to the length of the tangent vector. The expression is: (twenty four) S74: Similarly, construct the constant-rate load constraint term. : (25) In the formula, Indicates the first training sample set i The vector obtained after the load attribute vector of each sample is input into the decoder. express about k The derivative of . It can be considered as along . exist k The speed at that location, and This indicates the magnitude of the velocity, corresponding to the length of the tangent vector. The expression is: (26) S75: Constructing a constant-rate loss function for the total latent attribute vector: (27) S76: Construct the overall loss function for the attribute-supervised decoupled representation data generation method, satisfying the following relationship: (28) In the formula, To reconstruct the loss, For attribute supervision constraints, For total relevance constraints, For convex path loss constraints, For constant rate loss constraints, , , and All are scalar coefficients, used to adjust the loss weights of the attribute supervision constraint, mutual information constraint, convex path constraint, and constant velocity constraint, respectively.

[0017] In one example, fault data is learned and generated from data of a CRH2 traction motor (three-phase asynchronous motor). In this example, under four operating conditions—2800 r / min, 3200 r / min, and 3600 r / min—normal, inter-turn short circuit, and air gap eccentricity fault data were collected. The inter-turn short circuit and air gap eccentricity fault data each included A and B phase current sensor data of five different severity levels: 0.005, 0.05, 0.05, 0.08, and 0.1. The data sampling frequency was 2500 Hz, and the data sampling duration for each operating condition was 30 seconds. After preprocessing such as standardization and sliding window (window length 1024), the collected samples were randomly divided into two groups at a 3:2 ratio. Divided into training set and test set 5m real training samples Encoded into a latent vector by an encoder. Next, based on the fact that the fault samples have four attributes, the latent vectors are randomly and uniformly divided into four parts, resulting in four attribute latent vectors. It's worth noting that at this point, there is no direct correspondence between these four attribute latent vectors and the fault attributes. The latent attribute vectors of the training dataset can be represented as... .

[0018] In this embodiment, the parameter settings for training the traction motor fault sample generation model based on attribute-supervised decoupled representation are as follows: parameters λp=5, λTC=5, latent vector The batch size is B=64, the number of iterations is N=10, the learning rate of the encoder and decoder is 1e-4, and the learning rate of the variational approximation network is 1e-3. The training process uses the Adam optimizer to minimize the loss term.

[0019] The optimal model is obtained through training with the above parameters. Then, zero-sample generation is achieved through attribute latent vector segmentation and recombination. For example, inter-turn short circuit data with a rotation speed of 2800 r / min and a fault severity of 0.005 and inter-turn short circuit data with a rotation speed of 3200 r / min and a fault severity of 0.05 can be used to generate inter-turn short circuit data with a rotation speed of 3200 r / min and a fault severity of 0.005 through attribute segmentation and recombination. This generates new attribute data and improves the completeness of the fault dataset.

[0020] The example in this study underwent eight tests, generating the following results using the learned semantics: Figure 3 The table shows eight different types of fault data, with generated data on the left and real data on the right. The meanings of the generated data codes are shown in Table 1. Figure 3 The values ​​of T = 1, D = 0.05, and S = 3200 shown represent inter-turn short-circuit fault data with a severity of 0.05 at a speed of 3200 r / min. Figure 3 As shown, the generated samples maintain a high degree of similarity with the real samples, indicating that the proposed method has truly learned the real semantics of traction motor faults. This can be used to improve the completeness of fault data, thereby improving the generalization ability and diagnostic accuracy of the fault diagnosis model.

[0021] Table 1. Meaning of Generated Data Codes

[0022] This application also provides a traction motor fault sample generation system, including a memory, a processor, and a predictive machine program stored in the memory and run on the processor. When the processor executes the computer program, it implements the steps corresponding to the above method.

[0023] The motor fault data generation system described above can implement all embodiments of the motor fault data generation method described above and achieve the same beneficial effects. Further details are omitted here.

[0024] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for generating traction motor fault samples, characterized in that, include: S1: Collect traction motor current signals under different operating conditions, obtain corresponding attribute labels, construct a total sample set from the collected samples, and divide the total sample set into a training sample set and a test sample set; S2: Construct an encoder model, which includes an encoder and a decoder; S3: Based on the encoder, the training sample set is encoded into a low-dimensional latent vector. Based on the decoder and the low-dimensional latent vector, the samples are reconstructed to obtain reconstructed samples, and a reconstruction loss function is constructed. S4: Find the unique overlapping attribute sample pair and all overlapping attribute sample pairs in the training sample set; divide the low-dimensional latent vector into N types of attribute latent vectors, exchange the attribute latent vectors of the unique overlapping attribute sample pair and all overlapping attribute sample pairs in a set manner, and construct attribute supervision constraint terms. S5: Estimate the upper bound of mutual information between pairwise attribute latent vectors, introduce total correlation, and decouple attribute latent vectors to minimize total correlation to obtain total correlation constraints. S6: Construct convex path loss functions for the potential vectors of degree, rotation speed, and load attributes in sequence, and construct the convex path loss function for the total potential vector of attributes; S7: Construct constant-speed loss functions for the potential vectors of degree, rotation speed, and load attributes in sequence, and construct a constant-speed loss function for the total potential vector of attributes; construct the total loss function based on the reconstruction loss function, attribute supervision constraints, total correlation constraints, convex path loss, and constant-speed loss function. S8: Using the total loss function as the objective, train the optimal model parameters, use the optimal model parameters as the target model, and test the target model through the test sample set. After the test is passed, input the real-time collected traction motor current signal into the encoder model of the target model to obtain the corresponding latent vector. Perform attribute segmentation on the latent vector and reorganize the new latent vector to input it into the decoder model of the target model to achieve zero-sample generation.

2. The traction motor fault sample generation method according to claim 1, characterized in that, S1 includes: S11: Collect data under different categories, degrees, speeds, and loads. m Traction motor current signal ,in, m It is a constant, and the corresponding attribute tag is obtained. The types of attribute tags include category, degree, speed, load, and others. , Represents the real number field. Indicates sample Length, Represent the sample space; These are the real samples and their corresponding labels. , , , These are the attribute label values ​​for category, degree, speed, and load; other attribute label values... Set as default; S12: Construct the overall sample set based on the collected samples. The overall sample set was randomly divided at a ratio of 3:

2. Divided into training sample set and test sample set ,in,( This represents the traction motor current signal and its corresponding attribute label in the training sample set. This represents the traction motor current signal and its corresponding attribute label in the test sample set.

3. The method for generating traction motor fault samples according to claim 1, characterized in that, The encoder in S2 includes an N-layer convolutional neural network, and the decoder includes a deconvolutional layer that is structurally symmetrical with the encoder.

4. The method for generating traction motor fault samples according to claim 1, characterized in that, S3 includes: S31: Transfer samples from the training sample set Input encoder This yields a low-dimensional latent vector projected onto the latent space. It satisfies the following relationship: (1) In the formula, It is a low-dimensional latent vector with dimension O(n). ,and , Indicates sample Length, Represent the latent vector space; the decoder receives low-dimensional latent vectors. To reconstruct the original input sample, the output obtained is the reconstructed sample. Reconstructing samples and input samples In the same sample space, that is ; S32: Construct a reconstruction loss function that satisfies the following relationship: (2) In the formula, To reconstruct the loss, Characterization decoder; All samples in the training sample set are input into the autoencoder to obtain the corresponding low-dimensional latent vectors and reconstructed samples.

5. The method for generating traction motor fault samples according to claim 1, characterized in that, S4 includes: S41: Find sample pairs that have exactly one identical attribute label value and consider them as unique overlapping attribute sample pairs. It satisfies the following relationship: (3) Traversing the training sample set Find all unique overlapping attribute sample pairs; Find sample pairs where all attribute label values ​​are the same except for other types of attribute labels, and consider them as all overlapping attribute sample pairs. It satisfies the following relationship: (4) Traversing the training sample set Find all overlapping attribute sample pairs; S42: The low-dimensional latent vector obtained by the autoencoder. Latent attribute vectors segmented into categories The latent vector of degree attributes The latent vector of rotational speed attributes The latent vector of load properties Other attribute latent vectors The latent vector of the category attribute is used as the first latent vector, the latent vector of the degree attribute is used as the second latent vector, the latent vector of the rotation speed attribute is used as the third latent vector, the latent vector of the load attribute is used as the fourth latent vector, and the other latent vectors are used as the first latent vector. The dimensions of each sub-vector satisfy the following relationship: (5) In the formula, These are the category attribute vector dimension, degree attribute vector dimension, rotation speed attribute vector dimension, load attribute vector dimension, and other attribute vector dimension, respectively. S43: Swap the latent vectors of attributes corresponding to the same attribute label values ​​in unique overlapping attribute sample pairs, keeping the remaining latent vectors unchanged. For the first... n Unique overlapping attribute sample pairs with the same attribute value For n=1, 2, 3, 4, the attribute swapping operation between them is represented as follows: (6) For unique overlapping attribute sample pairs Using the attribute value corresponding to the severity level as the second attribute value, assuming their second attribute values ​​are the same, swapping their second attribute values ​​is represented as follows: (7) In the formula, Indicates the first training sample set i The sample and the first j The latent attribute vectors of each sample are swapped after the second attribute value is swapped. All unique overlapping attribute sample pairs are traversed, and the attribute latent vectors corresponding to the same attribute label values ​​between them are swapped, while keeping the low-dimensional latent vectors of other attributes unchanged. Swap the latent vector of the 5th attribute of all overlapping attribute sample pairs, while keeping the latent vectors of the other attributes unchanged; S44: Establish a constraint function based on the consistency requirement of sample reconstruction before and after attribute swapping for unique attribute sample pairs, satisfying the following relationship: (8) In the formula, Represents the probability density function. Expressing expectations, Indicates the first training sample set i The latent vector obtained by attribute swapping of samples Indicates the first training sample set i The sample and the first j The low-dimensional latent attribute vectors of the samples are interchanged. n The latent attribute vector following the attribute value; The attribute constraint items are represented as follows: (9) In the formula, This indicates a sample swapping operation; The constraint terms under attribute supervision can be represented as a loss function, satisfying the following relationship: (10)。 6. The method for generating traction motor fault samples according to claim 1, characterized in that, S5 includes: S51: Constructing a mutual information upper bound estimation function. Taking the estimation of the mutual information upper bound between the first attribute latent vector and the second attribute latent vector of the i-th sample as an example, it is expressed as follows: (11) In the formula, Indicates the upper bound of mutual information. Representation of attribute latent vector and The true distribution of conditions between them The parameter is The variational conditional distribution, using parameters Variational network learns attribute latent vectors and The true distribution of conditions between them This constraint term represents the conditional distribution and is expressed as follows: (12) Using parameters respectively , , Variational network learns attribute latent vectors and , and , and The true distribution of conditions between them; S52: Introduce total correlation to minimize the total correlation and the upper bound of mutual information of all latent attribute vectors, thereby achieving pairwise decoupling between latent attribute vectors. The total correlation constraint satisfies the following relationship: (13) In the formula, Variables representing estimates A and B The upper bound of mutual information between them, where A and B are respectively , , , .

7. The method for generating traction motor fault samples according to claim 1, characterized in that, S6 includes: S61: Construct convex path loss functions for the latent vectors of degree, rotational speed, load, and other attributes in sequence. The convex path loss function for the latent vector of degree attribute satisfies the following relationship: (15) In the formula, Indicates the first i The latent vector of the degree attribute of each sample With the j The latent vector of the degree attribute of each sample The geodesic distance, which can be calculated using the local tangent space permutation algorithm, and the convex path loss function of the latent vector of rotational speed attribute satisfy the following relationship: (16) In the formula, Indicates the first i Latent vector of rotational speed attribute of each sample With the j Latent vector of rotational speed attribute of each sample The geodesic distance and the convex path loss function of the latent vector of load attributes satisfy the following relationship: (17) In the formula, Indicates the first i The latent vector of the load attributes of each sample With the j The latent vector of the load attributes of each sample The geodesic distance; S62: Constructing the convex path loss function for the total latent attribute vector: (19)。 8. The method for generating traction motor fault samples according to claim 1, characterized in that, S7 includes: S71: A polynomial function is used as the interpolation network to fit the latent vector of the degree attribute. This polynomial is expressed as: (20) In the formula, , , , For about k The coefficients of the polynomial from the highest to the lowest terms; S72: Use the interpolation network as the input to the decoder, and apply the following constant-rate constraint, which is expressed as: (21) In the formula, Indicates the first training sample set i The vector obtained after inputting the degree attribute vector of each sample into the decoder. express about k The derivative is considered to be along exist k The speed at that point, This indicates the magnitude of the velocity, corresponding to the length of the tangent vector. This indicates calculating the mean. The expression is: (22) Through constant rate constraint terms To obtain a constant-speed interpolation of a certain degree; S73: Construct constant rotation speed constraint terms : (23) In the formula, Indicates the first training sample set i The vector obtained after inputting the rotational speed attribute vector of each sample into the decoder. express about k The derivative is considered to be along exist k The speed at that location, and This indicates the magnitude of the velocity, corresponding to the length of the tangent vector. The expression is: (24) S74: Construct constant-rate load constraints : (25) In the formula, Indicates the first training sample set i The vector obtained after the load attribute vector of each sample is input into the decoder. express about k The derivative is considered to be along exist k The speed at that point, This indicates the magnitude of the velocity, corresponding to the length of the tangent vector. The expression is: (26) S75: Constructing a constant-rate loss function for the total latent attribute vector: (27) S76: Construct the overall loss function for the attribute-supervised decoupled representation data generation method, satisfying the following relationship: (28) In the formula, To reconstruct the loss, For attribute supervision constraints, For total relevance constraints, For convex path loss constraints, For constant rate loss constraints, , , and All are scalar coefficients, used to adjust the loss weights of the attribute supervision constraint, mutual information constraint, convex path constraint, and constant velocity constraint, respectively.

9. The method for generating traction motor fault samples according to claim 1, characterized in that, S8 includes: S81: Set the number of iterations to... N Batch size is m The learning rate is lr , training set All unique overlapping attribute sample pairs and all overlapping attribute samples The input is used in the constructed traction motor fault sample generation model based on attribute-supervised decoupled representation to minimize the total loss function. With the goal of training, the optimal generative model parameters are obtained, which include the optimal encoder parameters. Optimal decoder parameters ; S82: Test sample set Input the trained optimal encoder To obtain the low-dimensional latent vector Based on step S42, segmentation is performed to obtain the latent attribute vectors of the categories. The latent vector of degree attributes The latent vector of rotational speed attributes The latent vector of load properties Other attribute latent vectors ; S83: Reorganize the latent attribute vectors. Based on the attribute requirements of the target fault samples, select and reorganize the latent attribute vectors of different test samples to generate new latent vector combinations. ; S84: The recombined latent vector Input Decoder The generated fault samples are obtained. To achieve fault data generation in zero-sample scenarios, the generated samples meet the following requirements. And it is consistent with the spatial distribution of the real samples. This represents the sample length.

10. A traction motor fault sample generation system according to claim 1, comprising a memory, a processor, and a computer program and a large language model stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method described in any one of claims 1 to 9.

Images