A high-speed train traction system fault detection and diagnosis method based on conditional variational autoencoder

Abstract

The present application aims at solving the problems of the existing high-speed train traction system fault detection and diagnosis method, such as the influence of data distribution drift when the working condition changes, the difficulty of new working condition migration, and the forgetting of historical knowledge when continuously updating, and proposes a high-speed train traction system fault detection and diagnosis method based on conditional variational autoencoder, belonging to the field of intelligent monitoring and fault diagnosis of complex systems. The present application trains normal state representation model and residual identification model using source working condition data, and realizes cross-working condition adaptation with the help of limited fault-free data of target working condition, thereby realizing cross-working condition fault detection and diagnosis of high-speed train traction system, and improving the detection reliability and diagnosis accuracy under multi-working condition and long-term operation conditions. The present application comprises: constructing a conditional variational autoencoder dynamic modeling network; introducing a maximum mean difference constraint to reduce the data distribution deviation between working conditions; combining transfer learning and continuous learning to update the model; and realizing online fault detection and diagnosis based on dynamic residual.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring and fault diagnosis of complex systems, specifically relating to a fault detection and diagnosis method for high-speed train traction systems based on conditional variational autoencoders, and particularly a fault detection and diagnosis method for high-speed train traction systems that combines maximum mean difference constraints, transfer learning, and continuous learning. Background Technology

[0002] With the rapid development and widespread application of high-speed rail technology, high-speed trains have become an indispensable and important part of modern transportation. Their operational safety and reliability are directly related to significant social benefits and the safety of life and property. As the core of power transmission and energy conversion in high-speed trains, the traction system operates under complex conditions of high speed, variable load, and strong coupling. Throughout the train's entire lifecycle, the traction system typically undergoes various operating states, including starting acceleration, constant speed operation, coasting, and regenerative braking, while simultaneously experiencing continuous changes in speed, load, and environmental conditions. This multi-condition, time-varying operational characteristic leads to continuous changes in the statistical properties and feature distribution of traction system monitoring data, i.e., distribution drift is prevalent between different operational stages. When a fault detection and diagnosis model is established based on data collected under a stable operating condition, directly applying it to new operating conditions can easily lead to increased false alarm rates, increased false negative rates, and confusion between different fault types due to inconsistencies in the distribution of training and test data, severely impacting train maintenance efficiency and operational safety.

[0003] However, due to the extremely complex internal structure of high-speed train traction systems and the high degree of electrical and mechanical coupling between subsystems, traditional physical model-based fault detection and diagnosis methods are insufficient to effectively address these challenges. Model-based methods typically rely on establishing accurate mathematical models or state observers; however, due to uncertainties in the modeling process, the nonlinearity of the system itself, and the influence of complex external disturbances, accurate models are often difficult to construct and struggle to adapt to the dynamic changes of the traction system during multi-condition switching. Therefore, data-driven methods that do not require accurate physical modeling have gradually become the mainstream research approach. Existing data-driven methods mainly include discriminative methods for fault classification and generative or reconstructive methods for health state modeling and anomaly detection. Among these, methods such as autoencoders and variational autoencoders can learn the normal state characteristics of the system from operational data and achieve fault detection and diagnosis through residual analysis, providing a new technical approach for intelligent operation and maintenance of traction systems.

[0004] Patent CN115358294A discloses a method for detecting minor faults in high-speed train traction systems. It constructs residuals through normalized variable analysis and uses Hellinger distance statistics for fault discrimination to improve the detection capability of minor faults in high-speed train traction systems. However, this approach focuses on extracting statistical features and detecting faults under existing operating conditions, without addressing the cross-condition distribution mismatch problem caused by traction system monitoring data under different speed, load, and environmental changes. It fails to establish a unified latent spatial representation and distribution alignment mechanism. When the model is generalized from the training conditions to new conditions, the features of healthy samples are prone to shift, leading to instability in the detection statistics and causing false alarms or missed alarms.

[0005] Patent CN115165366A discloses a fault diagnosis method and system for rotating machinery under varying operating conditions. It addresses fault diagnosis scenarios where the source domain is labeled and the target domain is unlabeled. By introducing multi-core maximum mean difference and improved joint distribution adaptation, it achieves transfer learning to improve fault diagnosis performance under the target operating condition. However, this solution focuses on solving the transfer diagnosis problem under varying operating conditions and does not adequately consider the sequential updating and knowledge retention issues when multiple new operating conditions occur consecutively during the long-term operation of high-speed trains. When the model is continuously updated based on subsequent new operating condition data, it easily overwrites the effective knowledge learned under existing operating conditions, leading to a decline in diagnostic performance under historical operating conditions. This makes it difficult to meet the application requirements of high-speed train traction systems under long-term multi-operating-condition scenarios.

[0006] Currently, existing data-driven fault detection and diagnosis methods for high-speed train traction systems are insufficient in suppressing data distribution drift caused by changes in multiple operating conditions. Furthermore, they do not adequately consider target condition adaptation, cross-condition knowledge transfer, and knowledge retention during sequential condition updates, resulting in low detection stability and diagnostic reliability under long-term operation and complex operating conditions. Therefore, there is an urgent need to establish a fault detection and diagnosis method for high-speed train traction systems based on conditional variational autoencoders to improve detection reliability, diagnostic accuracy, and engineering applicability under multiple operating conditions. Summary of the Invention

[0007] This invention aims to address the problems of existing high-speed train traction system fault detection and diagnosis methods, such as performance degradation due to data distribution drift under changing operating conditions, difficulty in effective transfer due to insufficient labeled samples under new operating conditions, and easy forgetting of historical operating condition knowledge during continuous updates in long-term operation. It proposes a high-speed train traction system fault detection and diagnosis method based on conditional variational autoencoder.

[0008] The technical solution adopted in this invention is divided into two parts: offline modeling and online fault detection and diagnosis, specifically including the following steps:

[0009] Offline modeling:

[0010] Step 1: Collect fault-free input and output data under source and target operating conditions, and construct dynamic samples using a sliding time window;

[0011] Step 2: Construct a conditional variational autoencoder model and introduce a maximum mean difference constraint in the latent space to align fault-free samples under different operating conditions to a uniform distribution;

[0012] Step 3: Pre-train the model using source operating condition fault-free data to obtain basic model parameters;

[0013] Step 4: Use a limited number of fault-free samples under the target operating condition to perform transfer learning updates on the model, thereby adapting the model from the source operating condition to the target operating condition;

[0014] Step 5: For cases where multiple target working conditions occur sequentially, continuous learning and updates are performed by combining memory playback and parameter regularization to reduce the model's forgetting of historical working condition knowledge;

[0015] Step 6: Establish dynamic residuals, residual statistics, and fault detection thresholds based on the fault-free samples of the target operating condition;

[0016] Step 7: Construct a fault residual feature library based on the labeled fault samples.

[0017] Online fault detection and diagnosis:

[0018] Step 1: Collect online monitoring data under the current operating conditions and construct dynamic samples;

[0019] Step 2: Input the online samples into the updated model, obtain the predicted output, and calculate the dynamic residuals;

[0020] Step 3: Calculate the fault detection statistics based on the dynamic residuals, compare them with the detection thresholds, and output the fault detection results;

[0021] Step 4: When the detection statistic exceeds the threshold, match the current residual with the fault residual feature library and output the fault diagnosis result.

[0022] The beneficial effects of this invention include: First, by introducing a maximum mean difference constraint into the latent space of the conditional variational autoencoder, the distribution alignment of fault-free samples under different operating conditions is achieved, reducing the impact of data distribution drift caused by changes in operating conditions on detection performance; Second, by combining source operating condition pre-training with target operating condition transfer learning, the dependence on labeled samples of the target operating condition is reduced, and the cross-operating condition adaptability is improved; Third, by introducing a continuous learning mechanism of memory playback and parameter regularization, the knowledge forgetting problem in the continuous model update process is reduced; Fourth, online fault detection and diagnosis are achieved through dynamic residual statistics and fault mode matching, which has good engineering application value. Attached Figure Description

[0023] Figure 1 This is a general framework diagram of a fault detection and diagnosis method for high-speed train traction systems based on conditional variational autoencoders.

[0024] Figure 2 This is a schematic diagram of the electric drive experimental platform used to verify the method of the present invention;

[0025] Figure 3 A confusion matrix diagram for identifying fault types under different operating conditions;

[0026] Figure 4 This is a graph showing the changes in fault detection statistics and threshold decision under different operating conditions. Detailed Implementation

[0027] 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 specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0028] This embodiment provides a method for fault detection and diagnosis of high-speed train traction systems based on conditional variational autoencoders. For example... Figure 1 The diagram shown is an overall framework diagram of the method of the present invention; as follows: Figure 2 The diagram shows the structure of an electric drive experimental platform used to verify the method of the present invention. Based on the above framework and platform, the specific implementation steps are as follows:

[0029] Step 1: Establish a discrete dynamic model of the traction system

[0030] Let discrete time The input vector of the high-speed train traction system is as follows:

[0031] (1)

[0032] The output vector is:

[0033] (2)

[0034] in, and These represent the upper and lower DC bus voltages, respectively. , , These represent the three-phase traction currents, respectively. Indicates train speed.

[0035] The high-speed train traction system is represented as a discrete nonlinear dynamic system:

[0036] (3)

[0037] (4)

[0038] in, Let be the system state vector. and These are process noise and measurement noise, respectively. and These represent actuator fault items and sensor fault items, respectively.

[0039] Step 2: Describe the differences in data distribution under different operating conditions

[0040] Because high-speed trains operate under different speeds, loads, and environmental conditions, the input and output data under the source and target operating conditions typically do not satisfy the same distribution condition, which can be expressed as:

[0041] (5)

[0042] in, Indicates the distribution of source operating condition data. This indicates the distribution of target operating condition data.

[0043] Step 3: Construct dynamic samples

[0044] A historical information vector is constructed by using a sliding time window for the input and output time series data:

[0045] (6)

[0046] Therefore, dynamic samples are constructed:

[0047] (7)

[0048] in, This represents the length of the sliding time window.

[0049] Step 4: Construct source and target operating condition sample sets

[0050] The fault-free sample set under the source operating condition is represented as:

[0051] (8)

[0052] No. The fault-free sample set under each target operating condition is represented as follows:

[0053] (9)

[0054] in, and These represent the source operating condition label and the first... Target operating condition label.

[0055] Step 5: Construct the conditional variational autoencoder model

[0056] For dynamic samples The conditional variational autoencoder satisfies:

[0057] (10)

[0058] in, As a latent variable, For operating condition labels, These are decoder parameters.

[0059] The posterior distribution of the encoder output latent variable is:

[0060] (11)

[0061] in, For encoder parameters, and Let represent the mean and standard deviation of the latent variable distribution, respectively.

[0062] Step 6: Perform latent variable sampling and output prediction

[0063] Latent variables are obtained using reparameterization:

[0064] (12)

[0065] The decoder obtains the current output prediction value:

[0066] (13)

[0067] in, Indicates the condition label The prediction result of the system output at the current time by the conditional variational autoencoder under the given conditions.

[0068] Step 7: Establish source operating condition pre-training targets

[0069] Define the sample reconstruction loss as:

[0070] (14)

[0071] Under the original operating condition, a maximum mean difference constraint is introduced to construct the pre-trained objective function:

[0072] (15)

[0073] in, The reconstruction loss of the source operating condition sample, This represents the maximum mean difference between the potential distribution of the source operating condition sample and the reference distribution. These are the weighting coefficients.

[0074] Step 8: Establish target working condition transfer learning objectives

[0075] By using fault-free samples under the target operating condition to perform transfer updates on the pre-trained model under the source operating condition, a learning objective function under the target operating condition is constructed:

[0076] (16)

[0077] in, For the first Reconstruction loss under a target operating condition This is the corresponding distribution alignment item.

[0078] Step 9: Establish continuous learning goals

[0079] When multiple target conditions arrive sequentially, representative samples are selected from the learned conditions to form a memory set. And form a joint training set with the current target working condition samples:

[0080] (17)

[0081] Based on this, construct the continuous learning objective function:

[0082] (18)

[0083] in, For memory playback items, For parameter regularization terms, These are the weighting coefficients.

[0084] Step 10: Online Fault Detection

[0085] After completing offline training, fix the updated model parameters. For online time... Input the current dynamic sample into the model to obtain the predicted output and construct the residual:

[0086] (19)

[0087] Constructing fault detection statistics based on the statistical characteristics of fault-free samples:

[0088] (20)

[0089] in, The mean of the fault-free residuals. This is the fault-free residual covariance matrix. The statistics... With a predetermined detection threshold Compare; when When the system is in a normal state, it is determined that the system is in a normal state; when When this happens, the system is determined to have malfunctioned and enters the fault diagnosis phase.

[0090] Step 11: Perform online fault diagnosis and output the results.

[0091] A fault residual feature library is established based on labeled fault samples. For the first... For each fault type, calculate the matching distance between the current residual and the residual pattern for that fault type:

[0092] (twenty one)

[0093] in, and They represent the first The mean vector and covariance matrix of the residuals for each type of fault.

[0094] Compare the matching distances corresponding to various faults, and output the category with the smallest distance as the current fault diagnosis result:

[0095] (twenty two)

[0096] Therefore, when the fault detection statistics of the online sample exceed the detection threshold, a fault alarm signal is output, and the fault type diagnosis result is given in combination with Equation (22); when the fault detection statistics of the online sample do not exceed the detection threshold, the normal operation result of the system is output.

[0097] The above embodiments of the present invention are merely illustrative of the calculation process of the present invention and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the protection scope of the present invention.

Claims

1. A method for fault detection and diagnosis of high-speed train traction system based on conditional variational autoencoder, characterized in that, The method comprises two parts: offline modeling and online fault detection and diagnosis, specifically: Offline modeling: Step 1: Collect fault-free input and output data of the high-speed train traction system under source and target operating conditions, and construct dynamic samples using a sliding time window; Step 2: Construct a conditional variational autoencoder model and introduce a maximum mean difference constraint in the latent space to align the fault-free samples under different operating conditions to a unified reference distribution. Step 3: Pre-train the conditional variational autoencoder model using source operating condition fault-free data to obtain basic model parameters; Step 4: Use a limited number of fault-free samples under the target working condition to perform transfer learning to update the pre-trained model, so as to adapt the model from the source working condition to the target working condition. Step 5: When multiple target working conditions appear in sequence, combine memory playback and parameter regularization for continuous learning and updating to reduce the model's forgetting of historical working condition knowledge; Step 6: Establish dynamic residuals, residual statistics, and fault detection thresholds based on the fault-free samples of the target operating condition; Step 7: Construct a fault residual feature library based on the labeled fault samples. Online fault detection and diagnosis: Step 1: Collect online monitoring data under the current operating conditions and construct dynamic samples; Step 2: Input the online samples into the updated conditional variational autoencoder model to obtain the predicted output and calculate the dynamic residuals; Step 3: Calculate the fault detection statistics based on the dynamic residuals, compare them with the detection threshold, and output the fault detection results; Step 4: When the fault detection statistics exceed the detection threshold, match the current residual with the fault residual feature library and output the fault diagnosis result.

2. The method for fault detection and diagnosis of high-speed train traction system based on conditional variational autoencoder according to claim 1, characterized in that, The specific process of offline modeling step 1 is as follows: The high-speed train traction system at discrete moments The input vector is represented as The output vector is represented as The input vector includes at least the upper and lower DC bus voltages, and the output vector includes at least the three-phase traction current and the train speed. The high-speed train traction system is represented as a discrete nonlinear dynamic system: (1) (2) in, Let be the system state vector. and These are process noise and measurement noise, respectively. and These represent actuator fault items and sensor fault items, respectively. For changes in operating conditions under different speeds, loads, and environmental conditions, the data distribution of the source and target operating conditions is represented as follows: (3) in, Indicates the distribution of source operating condition data. This indicates the distribution of target operating condition data. Constructing historical information vectors using a sliding time window: (4) And construct dynamic samples: (5) in, The sliding time window length is used. Further, a fault-free sample set under the source operating condition is constructed. and the Fault-free sample set under target operating conditions .

3. The method for fault detection and diagnosis of high-speed train traction system based on conditional variational autoencoder according to claim 2, characterized in that, The specific process of offline modeling step 2 is as follows: For dynamic samples and working condition labels Construct a conditional variational autoencoder whose decoder satisfies: (6) in, As a latent variable, These are the decoder parameters. Set the potential prior distribution to a standard normal distribution: (7) The posterior distribution of the encoder output latent variable is: (8) Latent variables are obtained using reparameterization: (9) The decoder obtains the current output prediction value: (10) in, For encoder parameters, For decoder parameters, and Let represent the mean and standard deviation of the latent variable distribution, respectively.

4. The method for fault detection and diagnosis of high-speed train traction system based on conditional variational autoencoder according to claim 3, characterized in that, The specific processes of offline modeling steps 3 and 4 are as follows: Define the sample reconstruction loss as: (11) Under the source load condition, construct the aggregated posterior distribution of the source load condition. And based on the maximum mean difference between the aggregated posterior distribution and the standard normal reference distribution, a pre-training objective function for the source operating condition is established: (12) in, The reconstruction loss term for the source operating condition sample. This represents the maximum mean difference between the potential distribution of the source operating condition sample and the reference distribution. These are the weighting coefficients. The basic model parameters are obtained by minimizing the pre-trained objective function based on the source operating condition. . Under the target operating condition, using the pre-trained parameters of the source operating condition as initialization parameters, and utilizing the first... Fault-free sample set under target operating conditions Establish the target working condition transfer learning objective function: (13) in, For the first The reconstruction loss term under the target operating condition For the corresponding distribution alignment item, These are the weighting coefficients.

5. The method for fault detection and diagnosis of high-speed train traction system based on conditional variational autoencoder according to claim 4, characterized in that, The transfer learning update in step 4 of the offline modeling process specifically refers to: The network parameters of the conditional variational autoencoder are decomposed into shared parameters and adaptive parameters: (14) in, and Shared layer parameters inherited from the source condition remain fixed during migration to the target condition. and For the first The adaptive layer parameters corresponding to each target working condition are updated during the target working condition migration process. By fixing the parameters of the shared layer and updating the parameters of the adaptive layer, the model can be adapted to the distribution of target working condition data.

6. The method for fault detection and diagnosis of high-speed train traction system based on conditional variational autoencoder according to claim 5, characterized in that, The continuous learning update in step 5 of the offline modeling process specifically refers to: When multiple target operating conditions arrive sequentially according to the task sequence, representative fault-free samples are selected from the learned operating conditions to form a memory set. And it is combined with the current target working condition sample set to form a joint training set: (15) And a memory replay loss term is established based on the joint training set. . Record the model parameters after the previous learning stage. And construct parameter regularization terms based on the importance weights of each parameter: (16) Further construct the overall objective function for continuous learning: (17) in, represents the weighting coefficients. By minimizing the overall objective function of continuous learning, the model achieves sequential updates under new operating conditions while mitigating the forgetting of historical operating condition knowledge.

7. The method for fault detection and diagnosis of high-speed train traction system based on conditional variational autoencoder according to claim 6, characterized in that, The specific process of the online fault detection is as follows: After completing offline training and updates, fix the updated model parameters; for online time... Input the current dynamic sample into the model to obtain the predicted output. And construct dynamic residuals: (18) Calculate the mean of the fault-free residual based on the fault-free sample under the target operating condition. and fault-free residual covariance matrix Therefore, fault detection statistics are constructed: (19) And based on the empirical distribution of the fault-free phase statistic and the given significance level. Determine the detection threshold . The fault detection statistics are compared with the detection threshold. When... The system is determined to be in a normal state when... The system will determine when a fault has occurred and enter the fault diagnosis phase.

8. The method for fault detection and diagnosis of high-speed train traction system based on conditional variational autoencoder according to claim 7, characterized in that, The specific process of online fault diagnosis is as follows: Establish a fault residual feature library based on labeled fault samples; for the first For each type of fault, calculate the mean vector of the residual samples corresponding to that type of fault. Covariance Matrix . When the fault detection statistic of the online sample exceeds the detection threshold, calculate the current residual and the result of the first step. Matching distance between fault-like residual patterns: (20) The matching distances for various types of faults are compared, and the category with the smallest matching distance is output as the current fault diagnosis result.

Images