A high-speed train traction system fault detection method based on domain adaptive variational autoencoder

Abstract

The present application aims to solve the problems of the existing high-speed train traction system fault detection method under variable speed working conditions, such as the decline of detection performance caused by data distribution drift, and the difficulty of effective cross-condition adaptation under the condition of lack of sufficient labeled samples in target working conditions, and proposes a high-speed train traction system fault detection method based on domain adaptive variational autoencoder, belonging to the field of intelligent monitoring and fault diagnosis of complex systems. The present application can use source working condition fault-free data and target working condition fault-free data to perform cross-condition fault detection on the high-speed train traction system, and improve the detection reliability and accuracy under variable speed working conditions and complex operating conditions. The content of the present application includes the following steps: constructing a domain adaptive variational autoencoder dynamic modeling network; introducing a maximum mean difference constraint to reduce the data distribution deviation between working conditions; combining source working condition pre-training and target working condition adaptation to complete model updating; and finally realizing online fault detection based on dynamic residual.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring and fault diagnosis of complex systems, specifically involving a fault detection method for high-speed train traction systems based on a domain adaptive variational autoencoder, and particularly a fault detection method for high-speed train traction systems that combines dynamic feature construction, potential spatial distribution alignment and residual statistical discrimination. Background Technology

[0002] With the rapid development of high-speed rail technology, high-speed trains have become an important transportation equipment in the modern rail transit system. Their operational safety and reliability are directly related to transportation efficiency, social benefits, and the safety of passengers' lives and property. As the core device for high-speed trains to achieve power conversion and traction drive, the traction system operates under high-speed, variable-load, strong coupling, and complex disturbance environments for extended periods. It is easily affected by speed fluctuations, load changes, and external environmental factors, leading to various faults such as sensor bias, drift, and abnormal gain. If these faults are not detected in a timely and accurate manner, it will not only affect the stable operation of the traction system but may also further induce performance degradation or even safety accidents. Therefore, it is necessary to conduct research on fault detection methods for high-speed train traction systems.

[0003] Fault detection in high-speed train traction systems mainly includes mechanistic model-based methods and data-driven methods. Mechanism-based methods typically rely on precise mathematical models of the controlled object or state observers, constructing residuals to achieve fault detection. However, high-speed train traction systems exhibit significant nonlinearity, time-varying characteristics, and multi-subsystem coupling, making it difficult to establish precise mechanistic models. Furthermore, they are susceptible to parameter perturbations, external disturbances, and changes in operating conditions during actual operation, limiting the adaptability of such methods under complex conditions. In contrast, data-driven methods do not rely on precise physical models and can directly utilize system operating data to mine state characteristics, thus gradually becoming an important research direction in the field of traction system fault detection.

[0004] However, in actual operation, high-speed train traction systems are often under dynamic conditions with constantly changing speeds, such as acceleration, deceleration, and switching between different speed modes. At this time, because the operating conditions corresponding to the training and testing phases are different, the monitoring data often exhibits significant differences in statistical characteristics and feature distribution, i.e., there is a cross-condition distribution drift problem. Traditional data-driven fault detection methods are usually based on the assumption that the training and testing data satisfy the same distribution. When this assumption is not true, the model's generalization ability will significantly decrease, leading to problems such as increased false alarm rate, increased false negative rate, and decreased detection stability. Especially when there is a lack of sufficient labeled samples for the target operating condition, how to achieve reliable fault detection under variable speed conditions has become a key issue in the intelligent monitoring of high-speed train traction systems.

[0005] Patent CN114065613B discloses a multi-condition fault detection and diagnosis method for process industries based on deep transfer learning. This method utilizes a variational autoencoder to construct a fault detection model and combines maximum mean difference to align the distribution of multi-condition data. It then determines whether the system is abnormal by comparing the loss function of the detection model with a threshold. However, this solution is primarily geared towards process industries and organizes data in a two-dimensional matrix, without specifically addressing the construction of multi-sensor time-series dynamic characteristics and residual statistical detection under variable-speed operation conditions in high-speed train traction systems.

[0006] Patent CN113281680B discloses a method for diagnosing open-circuit faults in single-phase three-level rectifiers of high-speed railway traction systems. This method achieves fault diagnosis and location by establishing a hybrid logic dynamic model, designing an adaptive sliding mode observer, and combining the relationship between residual features and adaptive thresholds. However, this scheme mainly targets specific open-circuit faults in traction rectifiers, focusing on observer design and fault location, without considering the data distribution drift problem under variable speed conditions, and without establishing a cross-condition unified feature learning mechanism based on domain adaptation.

[0007] Current fault detection methods for high-speed train traction systems are insufficient in suppressing data distribution drift under variable speed conditions. They fail to comprehensively consider the synergistic relationship between multi-sensor dynamic temporal feature representation, cross-condition distribution alignment, and online residual detection, resulting in low stability and accuracy of fault detection under varying operating conditions. Therefore, there is an urgent need to establish a fault detection method for high-speed train traction systems based on a domain adaptive variational autoencoder. This method aims to improve the reliability and engineering applicability of fault detection under variable speed conditions by constructing dynamic features and aligning them with the potential spatial distribution. Summary of the Invention

[0008] This invention aims to address the problems of existing high-speed train traction system fault detection methods, such as the susceptibility of detection performance to data distribution drift under variable speed conditions and the difficulty in achieving effective cross-condition adaptation when there are insufficient labeled samples in the target condition. The invention proposes a high-speed train traction system fault detection method based on a domain adaptive variational autoencoder.

[0009] The technical solution adopted in this invention is divided into two parts: offline modeling and online fault detection. Steps 1 to 6 are offline modeling steps, and steps 7 to 8 are online fault detection steps, as detailed below:

[0010] Step 1: Establish a dynamic representation model of the high-speed train traction system, and represent the system input and output monitoring data as a dynamic time sequence process;

[0011] Step 2: Acquire fault-free input / output monitoring data under source and target operating conditions, and reassemble the timing signals through a sliding time window to form dynamic samples;

[0012] Step 3: Construct a domain-adaptive variational autoencoder model and introduce a maximum mean difference constraint in the latent space to reduce the distribution deviation between samples under different working conditions;

[0013] Step 4: Pre-train the domain adaptive variational autoencoder model using source operating condition fault-free data to obtain basic model parameters;

[0014] Step 5: While retaining the pre-trained encoder parameters, the source working condition samples and the target working condition fault-free samples are used together for model optimization. Through domain adaptive correction and joint constraints of multiple losses, the model is transferred and adapted to the target working condition.

[0015] Step 6: Calculate the residuals of the model using the fault-free samples from the validation phase, and establish fault detection statistics and corresponding thresholds based on the statistical characteristics of the residuals;

[0016] Step 7: Collect online monitoring data under the current operating conditions, construct dynamic samples in the same way as the offline stage, input them into the trained domain adaptive variational autoencoder model, obtain the corresponding output estimation results, and generate dynamic residuals based on the deviation between the model output and the actual output.

[0017] Step 8: Calculate the fault detection statistic based on the dynamic residual and compare it with a predetermined detection threshold; if the fault detection statistic is greater than or equal to the detection threshold, it is determined that the high-speed train traction system has malfunctioned; otherwise, it is determined that the system is in normal operation.

[0018] The beneficial effects of this invention include: First, by constructing dynamic features from the multi-sensor input and output data of the high-speed train traction system, the temporal dynamic information during system operation can be characterized. Second, by introducing a maximum mean difference constraint in the latent space of the domain adaptive variational autoencoder and combining it with a domain adaptive correction network, the feature distributions of the source condition and the target condition are aligned, thereby reducing the impact of data distribution drift caused by variable speed conditions on fault detection performance. Third, by combining source condition pre-training with target condition joint training, the model's dependence on the labeled samples of the target condition is reduced, and the cross-condition adaptability is improved. Fourth, by constructing fault detection statistics based on the output reconstruction residual and performing threshold discrimination, online fault detection of the high-speed train traction system under variable speed conditions is realized, which has good engineering application value. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the fault detection method for the domain adaptive variational autoencoder described in this invention;

[0020] Figure 2This is a schematic diagram of the domain adaptive variational autoencoder structure described in this invention;

[0021] Figure 3 The fault detection results are shown in the diagram for the three types of faults under Mode A.

[0022] Figure 4 The fault detection results are shown in the diagram for the three types of faults under Mode B.

[0023] Figure 5 The diagram shows the fault detection results for three types of faults under Mode C.

[0024] Figure 6 The diagram shows the fault detection results for the three types of faults under Mode D. Detailed Implementation

[0025] To make the objectives, technical solutions, and beneficial effects 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 following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of this invention.

[0026] This embodiment provides a fault detection method for high-speed train traction systems based on a domain adaptive variational autoencoder. For example... Figure 1 The diagram shown is a schematic flowchart of the domain adaptive variational autoencoder fault detection method described in this invention; Figure 2 The diagram shown is a schematic representation of the domain adaptive variational autoencoder structure described in this invention; as follows: Figures 3 to 6 The image shows the detection results for three types of faults under different speed modes. The specific implementation steps are as follows:

[0027] Step 1: Establish a dynamic representation model of the high-speed train traction system

[0028] Let discrete time Below, the input vector of the high-speed train traction system is The output vector is The state vector is Then the system can be represented as:

[0029] (1)

[0030] (2)

[0031] in, For process noise, To measure noise, and These represent the state transition mapping and the output mapping, respectively.

[0032] Near a given operating point, the system can be approximated as a discrete linear time-invariant system:

[0033] (3)

[0034] (4)

[0035] in, , , and These are the state matrix, input matrix, output matrix, and direct transfer matrix, respectively.

[0036] Step 2: Construct dynamic samples of source and target operating conditions

[0037] Fault-free data collected under stable operating conditions is defined as source operating condition data, and fault-free data collected under variable speed operating conditions is defined as target operating condition data. The source and target operating condition datasets are represented as follows:

[0038] (5)

[0039] (6)

[0040] in, , These represent the source operating condition input and output data, respectively. , These represent the target operating condition input and output data, respectively.

[0041] To characterize the dynamic timing features of the high-speed train traction system, a sliding time window is used to reassemble the input and output signals. Let the window length be... Then the dynamic vectors of the input and output are represented as follows:

[0042] (7)

[0043] (8)

[0044] Further construct dynamic samples:

[0045] (9)

[0046] in, This is the dynamic sample corresponding to the current moment.

[0047] Step 3: Construct a domain-adaptive variational autoencoder model

[0048] Establish a domain-adaptive variational autoencoder model, the model including a feature encoder. Domain Adaptive Correction Network and output decoder .

[0049] The feature encoder maps dynamic samples to the latent space, whose posterior distribution is represented as:

[0050] (10)

[0051] in, For feature encoder parameters, and These are the mean and variance of the latent variable distribution, respectively.

[0052] Latent variables are obtained using reparameterization:

[0053] (11)

[0054] in, , This indicates element-wise multiplication.

[0055] Domain adaptive correction networks correct latent variables, and their mapping relationship is as follows:

[0056] (12)

[0057] in, To adaptively correct network parameters for the domain.

[0058] The output decoder generates the output predicted value based on the corrected latent variables, and its expression is:

[0059] (13)

[0060] in, To output decoder parameters, Output the prediction results for the model.

[0061] Step 4: Construct the source condition pre-training objective function

[0062] The feature encoder and output decoder are pre-trained using fault-free samples from the source operating condition. In this stage, the reconstruction loss is defined as:

[0063] (14)

[0064] Simultaneously, a maximum mean difference constraint between the latent variable distribution of the source operating conditions and the standard normal prior distribution is introduced into the latent space to construct the pre-training objective function for the source operating conditions:

[0065] (15)

[0066] in, It follows a standard normal prior distribution.

[0067] Step 5: Conduct target operating condition adaptation training

[0068] After completing the source condition pre-training, the feature encoder parameters are fixed, and both the source condition samples and the target condition fault-free samples are input into the model. The latent features are corrected through a domain adaptive correction network, and the output decoder is jointly optimized. The corresponding joint training objective function is:

[0069] (16)

[0070] in, To jointly reconstruct losses, To correct the distribution alignment terms between the source and target domains, For intra-domain constraints, This is a regularization term. Through this step, the distribution of potential features under the target operating condition is aligned with the healthy state of the source operating condition, thereby achieving model migration and adaptation from the source operating condition to the target operating condition.

[0071] Step 6: Establish fault detection statistics and detection thresholds

[0072] After offline training is completed, the fault-free samples from the validation phase are input into the trained model to obtain the output prediction values ​​and construct the residuals:

[0073] (17)

[0074] Based on the residuals, establish fault detection statistics:

[0075] (18)

[0076] in, This represents the matrix trace operation. Furthermore, the maximum value of the statistic corresponding to the fault-free sample is used as the fault detection threshold:

[0077] (19)

[0078] Step 7: Acquire online monitoring data and generate dynamic residuals

[0079] During the actual operation of the high-speed train traction system, online monitoring data under the current operating conditions is collected, and dynamic samples are constructed according to the method in step 2. These dynamic samples are then input into the trained domain adaptive variational autoencoder model to obtain the corresponding output prediction values. And generate dynamic residuals according to equation (17). .

[0080] Step 8: Complete fault diagnosis

[0081] Fault detection statistics are calculated based on the dynamic residual. and compare it with a predetermined detection threshold. The comparison is performed. The criterion for judgment is:

[0082] Fault (20)

[0083] No faults (21)

[0084] This completes the online fault identification process for high-speed train traction systems based on domain adaptive variational autoencoders.

[0085] The above embodiments of the present invention are merely illustrative of the implementation 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 fault detection method for high-speed train traction systems based on a domain adaptive variational autoencoder, characterized in that, The method comprises two parts: offline modeling and online fault detection, specifically: Offline modeling: Step 1: Acquire fault-free input / output monitoring data under source and target operating conditions, and reassemble the timing signals through a sliding time window to form dynamic samples; Step 2: Construct a domain-adaptive variational autoencoder model and introduce a maximum mean difference constraint in the latent space to reduce the distribution deviation between samples under different working conditions; Step 3: Pre-train the model using source operating condition fault-free data to obtain basic model parameters; Step 4: While retaining the pre-trained encoder parameters, the source working condition samples and the target working condition fault-free samples are used together for model optimization. Through domain adaptive correction and joint constraints of multiple losses, the model is transferred and adapted to the target working condition. Step 5: Calculate the residuals of the model using the fault-free samples from the verification phase, and determine the fault detection statistics and corresponding thresholds based on the statistical characteristics of the residuals. Online fault detection: Step 1: Collect online monitoring data under the current operating conditions and construct dynamic samples in the same way as the offline stage; Step 2: Input the dynamic samples into the trained domain adaptive variational autoencoder model to obtain the corresponding output estimation results; Step 3: Generate dynamic residuals based on the deviation between the model output and the actual output; Step 4: Calculate the fault detection statistic from the dynamic residual and compare it with a predetermined detection threshold; Step 5: If the fault detection statistic is greater than or equal to the detection threshold, the high-speed train traction system is determined to have malfunctioned; otherwise, the system is determined to be in normal operation.

2. The fault detection method for a high-speed train traction system based on a domain adaptive variational autoencoder according to claim 1, characterized in that, The specific process of offline modeling step 1 is as follows: Let discrete time Below, the input vector of the high-speed train traction system is The output vector is The state vector is Then the high-speed train traction system can be represented as a discrete nonlinear dynamic system: (1) (2) in, For process noise, To measure noise, and These represent the state transition mapping and the output mapping, respectively. Near a given operating point, the system is approximated as a discrete linear time-invariant system: (3) (4) in, , , and These are the state matrix, input matrix, output matrix, and direct transfer matrix, respectively. The source and target load condition datasets are represented as follows: (5) (6) in, , These represent the source operating condition input and output data, respectively. , These represent the target operating condition input and output data, respectively. The input and output dynamic vectors are constructed using a sliding time window: (7) (8) Further construct dynamic samples: (9) in, This represents the length of the sliding time window.

3. The fault detection method for a high-speed train traction system based on a domain adaptive variational autoencoder according to claim 2, characterized in that, The specific process of offline modeling step 2 is as follows: Construct a domain-adaptive variational autoencoder model, the model including a feature encoder. Domain Adaptive Correction Network and output decoder . For dynamic samples The feature encoder maps dynamic samples to the latent space, whose posterior distribution is represented as: (10) in, For feature encoder parameters, and These are the mean and variance of the latent variable distribution, respectively. Latent variables are obtained using reparameterization: (11) in, , ⊙ indicates element-wise multiplication. Latent variables are corrected using a domain adaptive correction network: (12) in, To adaptively correct network parameters for the domain. The output decoder generates the output predicted value based on the corrected latent variables: (13) in, To output decoder parameters, The model outputs prediction results. A maximum mean difference constraint is introduced into the latent space to align the feature distributions of the source and target operating conditions.

4. The fault detection method for a high-speed train traction system based on a domain adaptive variational autoencoder according to claim 3, characterized in that, The specific process of offline modeling step 3 is as follows: Define the refactoring loss under the source condition as: (14) Under the original operating condition, a constraint of the maximum mean difference between the latent variable distribution and the standard normal prior distribution is introduced to construct the pre-training objective function for the original operating condition: (15) in, The prior distribution is a standard normal distribution. The feature encoder and output decoder are trained by minimizing the pre-training objective function of the source condition to obtain the initial model parameters.

5. The fault detection method for a high-speed train traction system based on a domain adaptive variational autoencoder according to claim 4, characterized in that, The specific process of offline modeling step 4 is as follows: After completing the source condition pre-training, the feature encoder parameters are fixed, and both the source condition samples and the target condition fault-free samples are input into the model. The latent features are corrected through a domain adaptive correction network, and a joint training objective function is established. (16) in, To jointly reconstruct losses, To correct the distribution alignment terms between the source and target domains, For intra-domain constraints, This is a regularization term. By minimizing the joint training objective function, the domain adaptive correction network and the output decoder are jointly optimized to achieve model transfer and adaptation from the source operating condition to the target operating condition.

6. The fault detection method for a high-speed train traction system based on a domain adaptive variational autoencoder according to claim 5, characterized in that, The specific process of offline modeling step 5 is as follows: The fault-free samples are input into the trained domain adaptive variational autoencoder model to obtain the output prediction value, and a dynamic residual is constructed: (17) Based on the residuals, establish fault detection statistics: (18) in, This represents the matrix trace operation. The maximum value of the statistic corresponding to the fault-free sample is used as the fault detection threshold: (19) This completes the establishment of fault detection statistics and detection thresholds.

7. The fault detection method for a high-speed train traction system based on a domain adaptive variational autoencoder according to claim 6, characterized in that, The specific process of the online fault detection is as follows: The online monitoring data under the current working conditions are used to construct dynamic samples in the same way as the offline modeling stage, and then input into the trained domain adaptive variational autoencoder model to obtain the output prediction value. A dynamic residual is constructed based on the actual output and the predicted output value, and a fault detection statistic is calculated. ; The fault detection statistics With detection threshold In comparison, when At that time, it was determined that the high-speed train's traction system had malfunctioned. At that time, it was determined that the high-speed train traction system was in normal operating condition.

Images