Zero-shot online fault diagnosis method based on multi-resolution diffusion model

Abstract

This invention provides a zero-sample online fault diagnosis method based on a multi-resolution diffusion model. First, a source domain dataset U is constructed from data collected by a traction drive control simulation platform. S The target domain dataset U is constructed from data collected during actual train operation. T The fault category sets of the source and target domains do not overlap. A multi-resolution diffusion model is used to process the data, achieving outlier removal, missing data imputation, and standardization. A category-attribute description matrix A' is constructed, and fault features f from the source domain data are extracted using deep incremental supervised principal component analysis. i s Feature B s With attribute description a i The mapping relationship between them. Online acquisition of target domain data D. t Then, update the attribute descriptions of the fault categories and construct a new sparse matrix A. * The model parameters are then fine-tuned. Finally, Bayesian inference is used to perform real-time fault diagnosis on zero-sample data in the target domain, and the optimal fault category is inferred based on the maximum posterior probability, thus achieving accurate diagnosis and prediction of the target domain data.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for electric traction systems, and in particular to a zero-sample online fault diagnosis method based on a multi-resolution diffusion model. Background Technology

[0002] With the widespread application of electric traction systems in modern transportation systems, ensuring their operational stability and safety has become a significant technical challenge. As a core component of transportation vehicles such as railways and subways, the electric traction system primarily regulates train operation by controlling the speed and torque of the motor. However, electric traction systems can experience a wide variety of faults in actual operation, which not only affect train operating efficiency but may also lead to serious safety accidents.

[0003] Currently, existing fault diagnosis methods typically rely on training with large amounts of labeled data, which is often difficult to implement in practical applications, especially when new or rare fault types appear. These methods depend on numerous fault samples, making them ill-suited for fault detection in zero-sample scenarios during actual operation. Furthermore, existing fault diagnosis methods struggle to balance real-time performance, accuracy, and versatility when dealing with the complex and diverse operating conditions of electric traction systems. Therefore, a new method is urgently needed to overcome the shortcomings of existing technologies. Summary of the Invention

[0004] This invention provides a zero-sample online fault diagnosis method based on a multi-resolution diffusion model. By performing phased processing on source domain data and target domain data and combining it with deep incremental supervised learning, it achieves real-time detection and diagnosis of zero-sample faults in electric traction systems.

[0005] In a first aspect, embodiments of the present invention provide a zero-sample online fault diagnosis method based on a multi-resolution diffusion model, comprising the following steps:

[0006] S1) Define the data collected by the traction drive control simulation platform as source domain data, and define the data collected during actual train operation as target domain data; construct the source domain dataset U. S and target domain dataset U T Source domain dataset Where Y s E represents the source domain data sample space. s N represents the fault category space of the source domain data samples. s Indicates the number of data samples in the source domain; target domain data. Where Y t E represents the target domain data sample space. t N represents the fault category space of the target domain data samples. tIndicates the number of data samples in the target domain;

[0007] Construct fault category sets F for the source domain and the target domain respectively. s and F t The source domain's fault categories include common single faults and compound faults, while the target domain's fault categories include compound faults occurring under different operating conditions. Here, h is the number of fault categories in the source domain, q is the number of fault categories in the target domain, and the total number of fault categories is r = h + k. The set of fault categories in the source and target domains is F. s and F t They do not intersect, that is, F s ∩F t =φ;

[0008] S2) In establishing a multi-resolution diffusion model, the source domain U obtained from data acquisition... S and target domain U t The input data yields the reconstructed data. and It enables functions such as outlier removal, missing fault data filling, and standardization processing.

[0009] The multi-resolution diffusion model divides the time resolution into multiple stages from fine to coarse, and performs staged analysis and reconstruction of the operating data of the traction drive control system. Through decomposition and reconstruction at different time resolutions, fault signals are effectively recovered, and noise reduction performance and signal quality are improved.

[0010] S3) Construct fault category sets F for the source domain and the target domain respectively. s and F t Summarize source domain fault samples Y s Attribute description a i (i) Construct a category-attribute description matrix A' and transform it into a sparse category-attribute description matrix A using a one-hot encoder;

[0011] S4) will Fault category space E s Associate the attribute labels with the category attribute description matrix A to obtain the attribute labels L. s And deep incremental supervised principal component analysis (DI-SPCA) is used to extract fault features f from source domain data samples. i s Construct the feature matrix B of the source domain samples. s ; By training the learner model α i (i = 1, 2, 3, ..., C) Perform attribute learning to obtain the features B of the source domain samples. s With attribute description a i The mapping relationship between them;

[0012] S5), Online acquired runtime data D t (Target domain data), update the attribute descriptions of fault categories, and construct a new sparse category-attribute description matrix A. * ; and based on the updated matrix, the attribute learner α i Perform parameter fine-tuning;

[0013] S6) Input target domain samples According to the output a of the attribute learner t Using Bayesian inference to infer samples from attribute values The fault categories are defined, and the posterior probability of each fault category is given. The category with the highest posterior probability is selected as the optimal matching fault type, enabling real-time inference and diagnosis of fault types for zero-sample data in the target domain.

[0014] In some embodiments, the operating parameters of the source domain dataset and the target domain dataset mentioned in step S1) include: inverter three-phase current {I sa I sb I sc Train speed SP, motor speed Rs, motor electromagnetic torque Te, and voltage of the upper and lower supporting capacitors in the intermediate DC circuit {U}. cd1 U cd2}, AC voltage U on the secondary side of the transformer n and current data U n .

[0015] In some embodiments, the multi-resolution diffusion model described in step S2) performs multi-stage processing on source domain data and target domain data through stages of progressively increasing time resolution. Each stage involves staged analysis and reconstruction of data at different resolutions to gradually restore fault characteristics.

[0016] S21) Divide the preprocessed data into training set S train and test set S test The initial data is segmented into time series using a historical window (LW) and a current window (CW). The historical window (LW) covers data from 1 to 5 seconds, and the current window (CW) covers data from 1.5 to 5.5 seconds. The segmented data is then further decomposed into coarse-to-fine time resolutions of 10 μs, 100 μs, 1 ms, and 10 ms to generate multi-resolution data.

[0017] S22) The forward diffusion process adds noise to the source data, generating noisy data. Adding noise using the diffusion coefficient and Gaussian noise: Where Z is Gaussian noise sampled from the standard normal distribution N(0,1), and β Sγ is the variance coefficient of Z. S =1-β S It is the forward diffusion coefficient. It is the cumulative diffusion coefficient;

[0018] S23) The reverse reconstruction process adopts a Markov process, where the state at each time step t depends only on the reconstructed state of the previous step. Reverse reconstruction includes a conditional network and a reconstruction network. In phase S, the signal to be reconstructed is used... Signal X with the same time resolution ls As a condition for reconstruction, obtain the reconstructed fault data;

[0019] S24) The model uses the piecewise error SmoothL1 as the loss function, and the loss function is:

[0020]

[0021] Where, θ S These are all the parameters in the conditional network and reconstruction network during the S-stage. yes Under reconstruction condition c S The output of the denoising network is used below; this loss measures the difference between the noise generated by the network and the real noise, and θ is adjusted and optimized through iterative training. S Minimize L SmoothL1 (θ S This is to improve the performance of the reconstructed network.

[0022] In some embodiments, in step S3), the source domain fault samples Y are summarized. s Attribute description a i (i) consists of C attributes, and the description of h-class fault data and categories is represented as a category-attribute description matrix A'. A one-hot encoder is used to transform the source domain fault description matrix A' into a sparse category-attribute description matrix A: A∈R h×c =one-hot(A'), where all elements in matrix A are either 1 or 0, representing whether the attribute exists or not.

[0023] In some embodiments, in steps S3) and S4), DI-SPCA extracts the main features from high-dimensional data through layer-by-layer dimensionality reduction, mapping the original data to a low-dimensional space. The feature extraction mapping function ζ of DI-SPCA is used to map the source domain data sample Y. s and corresponding attribute tags As input, extract the fault feature vector f i s Represented as:

[0024]

[0025] In some embodiments, during step S5), the updated fault description matrix A is generated using a unique thermal encoder during the online update process. *′ ∈R r*c ; Transform into a new sparse category-attribute description matrix A * This is to reflect new types of faults and changes in operating conditions.

[0026] In some embodiments, in steps S5) and S6), the posterior probability of the fault category is represented as:

[0027]

[0028] in, Because there is a category-attribute description matrix A * Provides association information between target fault categories and attributes, such as p(a t |t)=1, comprehensive sample Fault characteristics and predict attribute label a t The posterior probability is expressed as:

[0029]

[0030] in, It is in feature f i t Under the condition, attribute The probability of belonging to fault category t; It is an attribute The prior probability across all fault categories; p(t) is the prior probability for each category.

[0031] In some embodiments, in step S6), during the calculation of the posterior probability of the fault category, it is assumed that the prior probabilities of all categories are equal, i.e., p(t) = constant, and the next step of maximum a posteriori calculation has no impact on the result; because Select the class with the highest posterior probability as the sample. Optimal matching fault type:

[0032]

[0033] in, This represents the attribute description vector for the e-th fault type in the target domain sample. yes The i-th element; thus, the category-attribute description matrix A *It plays a role in knowledge transfer. By using the attribute information of known fault categories, the model applies this knowledge to fault samples of unknown categories to achieve real-time reasoning and diagnosis of fault types in zero-sample data of the target domain. After the reasoning and diagnosis are completed, the fault category prediction results generated by the model will be used to guide the maintenance and fault handling of the system.

[0034] In a second aspect, embodiments of the present invention also provide an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the computer program, it implements the zero-sample online fault diagnosis method based on a multi-resolution diffusion model as described in the first aspect.

[0035] Thirdly, embodiments of the present invention also provide a computer-readable storage medium storing computer-executable instructions for performing the zero-sample online fault diagnosis method based on a multi-resolution diffusion model as described in the first aspect.

[0036] According to the zero-sample online fault diagnosis method based on a multi-resolution diffusion model provided in this embodiment of the invention, firstly, a source domain dataset U is constructed from the data collected by the traction drive control simulation platform. S The target domain dataset U is constructed from data collected during actual train operation. T The fault category sets of the source and target domains do not overlap. A multi-resolution diffusion model is used to process the data, achieving outlier removal, missing data imputation, and standardization. Next, a category-attribute description matrix A' is constructed, and fault features f from the source domain data are extracted using deep incremental supervised principal component analysis (DI-SPCA). i s Feature B s With attribute description a i The mapping relationship between them. Online acquisition of target domain data D. t Then, update the attribute descriptions of the fault categories and construct a new sparse matrix A. * The model parameters are then fine-tuned. Finally, Bayesian inference is used to perform real-time fault diagnosis on zero-sample data in the target domain. The optimal matching fault category is inferred based on the maximum a posteriori probability, achieving accurate diagnosis and prediction of the target domain data, thereby guiding system maintenance and fault handling. Based on this, this embodiment of the invention utilizes the attributes of known fault categories to diagnose new fault types and compound faults. By using fault categories and attributes as learning objects, attribute transfer from source domain fault categories to target domain fault categories can be achieved even without target domain fault category data samples, thus realizing zero-sample fault diagnosis. Attached Figure Description

[0037] Figure 1This is a flowchart of a zero-sample online fault diagnosis method based on a multi-resolution diffusion model provided in an embodiment of the present invention;

[0038] Figure 2 This is a waveform of periodic interference fault of inverter A-phase current sensor at a resolution of [10e-6] provided by an embodiment of the present invention;

[0039] Figure 3 This is a waveform diagram of periodic interference fault of inverter A-phase current sensor at a resolution of [5e-3] provided by an embodiment of the present invention;

[0040] Figure 4 This is a waveform diagram of periodic interference fault of inverter A-phase current sensor at a resolution of [20e-3] provided by an embodiment of the present invention;

[0041] Figure 5 This is a flowchart of a denoising network for a multi-resolution diffusion model provided in an embodiment of the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0043] In this embodiment of the invention, the terms "furthermore," "exemplarily," or "preferredly" are used as examples, illustrations, or descriptions and should not be construed as being more preferred or advantageous than other embodiments or designs. The use of terms such as "furthermore," "exemplarily," or "preferredly" is intended to present the relevant concepts in a specific manner.

[0044] Currently, existing fault diagnosis methods typically rely on training with large amounts of labeled data, which is often difficult to implement in practical applications, especially when new or rare fault types appear. These methods depend on numerous fault samples, making them ill-suited for fault detection in zero-sample scenarios during actual operation. Furthermore, existing fault diagnosis methods struggle to balance real-time performance, accuracy, and versatility when dealing with the complex and diverse operating conditions of electric traction systems. Therefore, a new method is urgently needed to overcome the shortcomings of existing technologies.

[0045] Based on this, the present invention provides a zero-sample online fault diagnosis method based on a multi-resolution diffusion model. First, a source domain dataset U is constructed from the data collected by the traction drive control simulation platform. S The target domain dataset U is constructed from data collected during actual train operation. TThe fault category sets of the source and target domains do not overlap. A multi-resolution diffusion model is used to process the data, achieving outlier removal, missing data imputation, and standardization. Next, a category-attribute description matrix A' is constructed, and fault features f from the source domain data are extracted using deep incremental supervised principal component analysis (DI-SPCA). i s Feature B s With attribute description a i The mapping relationship between them. Online acquisition of target domain data D. t Then, update the attribute descriptions of the fault categories and construct a new sparse matrix A. * The model parameters are then fine-tuned. Finally, Bayesian inference is used to perform real-time fault diagnosis on zero-sample data in the target domain. The optimal matching fault category is inferred based on the maximum a posteriori probability, achieving accurate diagnosis and prediction of the target domain data, thereby guiding system maintenance and fault handling. Based on this, this embodiment of the invention utilizes the attributes of known fault categories to diagnose new fault types and compound faults. By using fault categories and attributes as learning objects, attribute transfer from source domain fault categories to target domain fault categories can be achieved even without target domain fault category data samples, thus realizing zero-sample fault diagnosis.

[0046] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0047] like Figure 1 As shown, Figure 1 This is a flowchart of a zero-shot online fault diagnosis method based on a multi-resolution diffusion model provided in an embodiment of the present invention. The zero-shot online fault diagnosis method based on a multi-resolution diffusion model may include, but is not limited to, the following steps.

[0048] S1) Define the data collected by the traction drive control simulation platform as source domain data, and define the data collected during actual train operation as target domain data; construct the source domain dataset U. S and target domain dataset U T Source domain dataset Where Y s E represents the source domain data sample space. s N represents the fault category space of the source domain data samples. s Indicates the number of data samples in the source domain; target domain data. Where Y t E represents the target domain data sample space. t N represents the fault category space of the target domain data samples. t Indicates the number of data samples in the target domain;

[0049] Construct fault category sets F for the source domain and the target domain respectively. s and Ft The source domain's fault categories include common single faults and compound faults, while the target domain's fault categories include compound faults occurring under different operating conditions. Here, h is the number of fault categories in the source domain, q is the number of fault categories in the target domain, and the total number of fault categories is r = h + k. The set of fault categories in the source and target domains is F. s and F t They do not intersect, that is, F s ∩F t =φ;

[0050] S2) In establishing a multi-resolution diffusion model, the source domain U obtained from data acquisition... S and target domain U t The input data yields the reconstructed data. and Y t * It enables functions such as outlier removal, missing fault data filling, and standardization processing.

[0051] S3) Construct fault category sets F for the source domain and the target domain respectively. s and F t Summarize source domain fault samples Y s Attribute description a i (i) Construct a category-attribute description matrix A' and transform it into a sparse category-attribute description matrix A using a one-hot encoder;

[0052] S4) will Fault category space E s Associate the attribute labels with the category attribute description matrix A to obtain the attribute labels L. s And deep incremental supervised principal component analysis (DI-SPCA) is used to extract fault features f from source domain data samples. i s Construct the feature matrix B of the source domain samples. s ; By training the learner model α i (i = 1, 2, 3, ..., C) Perform attribute learning to obtain the features B of the source domain samples. s With attribute description a i The mapping relationship between them;

[0053] S5), Online acquired runtime data D t (Target domain data), update the attribute descriptions of fault categories, and construct a new sparse category-attribute description matrix A. * ; and based on the updated matrix, the attribute learner α i Perform parameter fine-tuning;

[0054] S6) Input target domain samples Based on the output a of the attribute learner t Using Bayesian inference to infer samples from attribute values The fault categories are defined, and the posterior probability of each fault category is given. The category with the highest posterior probability is selected as the optimal matching fault type, enabling real-time inference and diagnosis of fault types for zero-sample data in the target domain.

[0055] Preferably, in step S2), the operating parameters of the source and target domain datasets include: inverter three-phase current {I sa I sb I sc Train speed SP, motor speed Rs, motor electromagnetic torque Te, and voltage of the upper and lower supporting capacitors in the intermediate DC circuit {U}. cd1 U cd2}, AC voltage U on the secondary side of the transformer n and current data U n .

[0056] Preferably, in step S2), the multi-resolution diffusion model performs multi-stage processing on the source and target domain data by gradually increasing the time resolution. In each stage, the fault characteristics are gradually restored by performing staged analysis and reconstruction on the data at different resolutions.

[0057] Preferably, in step S2), where:

[0058] S21) Divide the preprocessed data into training set S train and test set S test ;

[0059] The initial data is divided into time series segments using a historical window (LW) and a current window (CW), with the LW covering data from 1 to 5 seconds and the CW covering data from 1.5 to 5.5 seconds.

[0060] like Figure 2 , Figure 3 and Figure 4 As shown, the segmented data is decomposed in stages from fine to coarse according to three different time resolutions: 10e-6, 5e-3, and 20e-3, generating multi-resolution data X. c1 X c2 X c3 X s1 X s2 X s3 .

[0061] S22) The forward diffusion process adds noise to the source data, generating noisy data.

[0062] In the forward diffusion network, noise is added using the diffusion coefficient and Gaussian noise:

[0063]

[0064] Where Z is Gaussian noise sampled from the standard normal distribution N(0,1), and β S γ is the variance coefficient of Z. S =1-β S It is the forward diffusion coefficient. It is the cumulative diffusion coefficient;

[0065] The goal is to enable the model to learn the data distribution under different noise levels by perturbing the noise.

[0066] S23) The reverse reconstruction process adopts a Markov process, and the state at each time step t depends only on the reconstructed state of the previous step, making the reconstruction process simple, stable and easy to calculate.

[0067] Reverse reconstruction includes conditional networks and reconstruction networks;

[0068] In the S-phase, the conditional network uses a signal X with the same time resolution as the signal to be reconstructed. ls Generate refactoring conditions, the refactoring conditions are: c s =concat(z lookback ,X c(s+1) ),s=K,…,1;

[0069] Among them, z lookback It is for X ls Tensors obtained by performing linear mapping;

[0070] The reconstruction network uses the piecewise error (SmoothL1) as the loss function, which is:

[0071]

[0072] Where, θ S These are all the parameters in the conditional network and reconstruction network during the S-stage. Under reconstruction condition c S The output of the denoising network is as follows: Figure 5 As shown:

[0073]

[0074] This loss measures the difference between the noise generated by the network and the real noise, and θ is adjusted and optimized through iterative training. S Minimize L SmoothL1 (θ STo improve the performance of the reconstructed network, data reconstruction can be continuously improved, reconstruction quality can be ensured, and fault data after reconstruction can be obtained.

[0075] Preferably, in step S3), in order to identify and diagnose the target fault with zero samples, the source domain fault samples Y are summarized. s Attribute description a i (i).

[0076] Composed of C attributes, the description of fault data and categories of type h can be represented as a category-attribute description matrix A'.

[0077] One-hot encoders are used to transform the source domain fault description matrix A' into a sparse category-attribute description matrix A:

[0078] A∈R h×c = one-hot(A')

[0079] In matrix A, all elements are either 1 or 0, indicating whether an attribute exists or not.

[0080] Preferably, in steps S3) and S4), DI-SPCA extracts the main features from the high-dimensional data by performing layer-by-layer dimensionality reduction, mapping the original data to a low-dimensional space and reducing computational complexity.

[0081] The feature extraction mapping function ζ of DI-SPCA is used to map the source domain data samples Y. s and corresponding attribute tags As input.

[0082] Extracting the fault feature vector f i s Represented as:

[0083]

[0084] Preferably, in step S5), during the online update process, the updated fault description matrix is ​​represented by a single-thermal encoder as follows:

[0085] A *′ ∈R r*c

[0086] Transformed into a new sparse category-attribute description matrix A * This is to reflect new types of faults and changes in operating conditions.

[0087] Preferably, in steps S3) and S4), the posterior probability of the fault category is expressed as:

[0088]

[0089] in,

[0090] There is a category-attribute description matrix A * Provides association information between target fault categories and attributes, such as p(a t |t)=1;

[0091] Comprehensive sample Fault characteristics and predict attribute label a t The posterior probability is expressed as:

[0092]

[0093] in, It is in feature f i t Under the condition, attribute The probability of belonging to fault category t;

[0094] It is an attribute The prior probability across all fault categories; p(t) is the prior probability for each category;

[0095] Preferably, in step S5), the calculation of the posterior probability of the fault category assumes that the prior probabilities of all categories are equal, i.e.:

[0096] p(t) = constant

[0097] The next step of maximum a posteriori calculation has no impact on the results;

[0098] because Select the class with the highest posterior probability as the sample. Optimal matching fault type:

[0099]

[0100] in, This represents the attribute description vector for the e-th fault type in the target domain sample. yes The i-th element.

[0101] Therefore, the category-attribute description matrix A * It plays a role in knowledge transfer. By using the attribute information of known fault categories, the model can apply this knowledge to fault samples of unknown categories.

[0102] Real-time inference and diagnosis of fault types based on zero-sample data in the target domain;

[0103] This process relies not only on the current input data but also on category attribute information from historical data in the source domain, thereby ensuring the accuracy of the diagnosis.

[0104] In addition, an embodiment of the present invention discloses an electronic device, comprising: at least one processor; at least one memory for storing at least one program; and, when the at least one program is executed by the at least one processor, implementing the zero-sample online fault diagnosis method based on the multi-resolution diffusion model as described in any of the preceding embodiments.

[0105] In addition, one embodiment of the present invention discloses a computer-readable storage medium storing computer-executable instructions for performing the zero-sample online fault diagnosis method based on a multi-resolution diffusion model as described in any of the preceding embodiments.

[0106] The system architecture and application scenarios described in the embodiments of this invention are for the purpose of more clearly illustrating the technical solutions of the embodiments of this invention, and do not constitute a limitation on the technical solutions provided by the embodiments of this invention. As those skilled in the art will know, with the evolution of system architecture and the emergence of new application scenarios, the technical solutions provided by the embodiments of this invention are also applicable to similar technical problems.

[0107] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0108] In hardware implementations, the division between functional modules / units mentioned in the above description does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software may be distributed on a computer-readable medium, which may include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, as is known to those skilled in the art, communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0109] The terms “component,” “module,” “system,” etc., used in this specification are used to refer to computer-related entities, hardware, firmware, combinations of hardware and software, software, or software in execution. For example, a component can be, but is not limited to, a process running on a processor, a processor, an object, an executable file, an execution thread, a program, or a computer. As illustrated, applications running on computing devices and computing devices can both be components. One or more components may reside in a process or execution thread, and components may be located on a single computer or distributed among two or more computers. Furthermore, these components can be executed from various computer-readable media on which various data structures are stored. Components can communicate, for example, via local or remote processes based on signals having one or more data packets (e.g., data from two components interacting with another component between a local system, a distributed system, or a network, such as the Internet interacting with other systems via signals).

Claims

1. A zero-shot online fault diagnosis method based on a multi-resolution diffusion model, characterized in that, Includes the following steps: S1) Define the data collected by the traction drive control simulation platform as source domain data, and define the data collected during actual train operation as target domain data; construct the source domain dataset. and target domain dataset Source domain dataset ,in, Represents the source domain data sample space. Represents the fault category space of the source domain data samples. Indicates the number of data samples in the source domain; target domain dataset. ,in, Represents the target domain data sample space. This represents the fault category space of the target domain data samples. Indicates the number of data samples in the target domain; Construct fault category sets for the source domain and the target domain respectively. and The source domain's fault categories include single faults and compound faults, while the target domain's fault categories include compound faults occurring under different operating conditions. It is the number of source domain fault categories. This represents the number of fault categories in the target domain; the total number of fault categories is... Source domain and target domain fault category sets and They do not interact with each other, that is ; S2) Establish a multi-resolution diffusion model to analyze the source domain obtained from data acquisition. and target domain The input data yields the reconstructed data. and ; S3) Fault category set based on source and target domains and Summarize source domain fault samples Attribute description Create a category-attribute description matrix And through a one-hot encoder, the matrix Transform into a sparse category-attribute description matrix ; S4) will Fault category space With category attribute description matrix Perform attribute tag association and retrieve attribute tags. Furthermore, deep incremental supervised principal component analysis (DI-SPCA) was used to extract fault characteristics from source domain data samples. Construct the feature matrix of source domain samples ; By training the learner model Attribute learning is performed, among which, To obtain the features of the source domain samples With attribute description The mapping relationship between them; S5) Online acquisition of operational data Update the attribute descriptions of the fault categories and construct a new sparse category-attribute description matrix. And based on the updated matrix, the attribute learner... Perform parameter fine-tuning; S6) Input target domain samples According to the output of the attribute learner Using Bayesian inference to infer samples from attribute values The fault categories are identified, the posterior probability of each fault category is calculated, and the category with the highest posterior probability is selected as the optimal matching fault type.

2. The method according to claim 1, characterized in that: The operating parameters of the source domain dataset and the target domain dataset mentioned in step S1) include: inverter three-phase current. Train speed Motor speed Motor electromagnetic torque The voltage of the upper and lower supporting capacitors in the intermediate DC circuit AC voltage on the secondary side of the transformer and current data .

3. The method according to claim 1, characterized in that: The multi-resolution diffusion model described in step S2) performs multi-stage processing on source domain data and target domain data through a progressively thickening phase of time resolution. Each stage involves phased analysis and reconstruction of data at different resolutions to gradually restore fault characteristics. S21) Divide the preprocessed data into training sets. and test set The initial data was segmented into time series using a historical window (LW) and a current window (CW). The historical window (LW) covered data from 1 to 5 seconds, and the current window (CW) covered data from 1.5 to 5.5 seconds. The segmented data was then further decomposed into coarse-to-fine time resolutions of 10 μs, 100 μs, 1 ms, and 10 ms to generate multi-resolution data. and ; S22), the forward diffusion process, in each stage To multi-resolution data Noise is added to the data using the diffusion coefficient and Gaussian noise to generate noisy data. : ,in, From the standard normal distribution Gaussian noise in the mid-sample, yes The variance coefficient, It is the forward diffusion coefficient. It is the cumulative diffusion coefficient; S23) The reverse reconstruction process adopts a Markov process, where the state at each time step t depends only on the reconstructed state of the previous step. Reverse reconstruction includes a conditional network and a reconstruction network. Phase, using the signal to be reconstructed Signals with the same time resolution As a condition for reconstruction, obtain the reconstructed fault data; S24) The model uses the piecewise error SmoothL1 as the loss function, and the loss function is: （1） in, Is All parameters in the stage's conditional network and reconstruction network, yes Under reconstruction conditions The output of the denoising network is used below; this loss measures the difference between the noise generated by the network and the real noise, and is adjusted and optimized through iterative training. , minimize To improve the performance of the reconstructed network.

4. The method according to claim 1, characterized in that: In step S3), source domain fault samples are summarized. Attribute description It consists of C attributes, and the description of fault category h is represented as a category-attribute description matrix. A unique thermal encoder is used to describe source domain faults. Transform the matrix into a sparse category-attribute description matrix : ,matrix All elements in the array are either 1 or 0, representing whether the attribute exists or not, respectively.

5. The method according to claim 1, characterized in that: In steps S3) and S4), DI-SPCA extracts the main features from high-dimensional data through layer-by-layer dimensionality reduction, mapping the original data to a low-dimensional space. The feature extraction mapping function ζ of DI-SPCA is used to transform the source domain data space samples. and corresponding attribute tags As input, extract the fault feature vector. Represented as: （2）。 6. The method according to claim 1, characterized in that: In step S5), during the online update process, the updated fault description matrix is ​​transmitted via a one-hot encoder. Transform into a new sparse category-attribute description matrix This is to reflect new types of faults and changes in operating conditions.

7. The method according to claim 1, characterized in that: In steps S5) and S6), the posterior probability of the fault category is expressed as: （3） in, Because of the category-attribute description matrix Provides association information between target fault categories and attributes. Comprehensive sample Fault characteristics and predicted attribute labels The posterior probability is expressed as: （4） in, It is an attribute Prior probabilities across all fault categories; It represents the prior probability of each category.

8. The method according to claim 7, characterized in that: In step S6), the calculation of the posterior probability of the fault category assumes that the prior probabilities of all categories are equal, i.e. The next step, maximum a posteriori calculation, has no impact on the result; because Select the class with the highest posterior probability as the sample. Optimal matching fault type: （5） in, This represents the attribute description vector for the e-th fault type in the target domain sample. yes The i-th element; thus, the new sparse category-attribute description matrix. It plays a role in knowledge transfer. By using the attribute information of known fault categories, the model applies this knowledge to fault samples of unknown categories to achieve real-time reasoning and diagnosis of fault types in zero-sample data of the target domain. After the reasoning and diagnosis are completed, the fault category prediction results generated by the model will be used to guide the maintenance and fault handling of the system.

9. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the computer program, it implements the zero-sample online fault diagnosis method based on a multi-resolution diffusion model as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing computer-executable instructions for performing the zero-sample online fault diagnosis method based on a multi-resolution diffusion model as described in any one of claims 1 to 8.

Images