A method for detecting and diagnosing micro-faults of high-speed train traction system based on diffusion model
By learning the normal state distribution of the high-speed train traction system through a diffusion model, detecting minor faults using the difference of latent variables in multiple diffusion steps, and combining smoothing processing and continuous alarm decision, the sensitivity and type differentiation problems in the detection of minor faults in the high-speed train traction system are solved, and efficient fault diagnosis is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHANGCHUN UNIV OF TECH
- Filing Date
- 2026-04-16
- Publication Date
- 2026-07-21
AI Technical Summary
Existing methods for detecting minor faults in high-speed train traction systems lack sufficient sensitivity under complex disturbance conditions, making it difficult to effectively distinguish different fault types. Furthermore, they are insufficient in learning the distribution of normal states and rely heavily on precise mechanism models and expert experience.
The diffusion model is used to learn the normal state distribution using only fault-free operating samples. Fault detection statistics are constructed by the differences of latent variables in multiple diffusion steps. Combined with smoothing processing and continuous alarm decision mechanism, minor fault detection and fault type diagnosis are realized.
It improves the detection sensitivity and stability of minor faults, enhances the ability to distinguish different fault types, reduces the impact of complex disturbances and noise, and has high engineering applicability and diagnostic accuracy.
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent monitoring and fault diagnosis of complex systems, specifically involving a method for detecting and diagnosing minor faults in a high-speed train traction system based on a diffusion model. In particular, it relates to a method for detecting minor faults and diagnosing fault types by learning the normal state distribution using only fault-free operating samples and based on the differences in latent variables of multiple diffusion steps. Background Technology
[0002] The high-speed train traction system is a crucial component for power output and energy conversion, and its operational status directly affects the safety, reliability, and maintenance efficiency of train operation. With increasing service time, sensors, power devices, and actuators in the traction system may gradually age, experience parameter drift, and degrade in performance, further inducing faults such as bias, gain variations, and device malfunctions. Minor faults typically do not cause significant performance changes in the early stages, but rather develop gradually with small amplitudes, slow evolution, and unclear boundaries. If not identified in time, minor faults may further evolve into serious faults or even system failure.
[0003] Existing fault detection and diagnosis methods for high-speed train traction systems mainly include model-based methods and data-driven methods. Model-based methods typically rely on system dynamic equations, state observers, or residual generation mechanisms, achieving fault detection and diagnosis through the consistency deviation between measured outputs and model predicted outputs. While these methods possess certain physical meaning, for traction systems with complex structures, strong variable coupling, complex operating conditions, and significant external disturbances, accurate models are often difficult to establish, thus limiting their practical application. In contrast, data-driven methods can directly utilize operational monitoring data for feature extraction, state modeling, and fault identification, making them more suitable for complex equipment like high-speed train traction systems.
[0004] Patent CN104569819B discloses a fault detection method for asynchronous traction motors. This method constructs a discrete state-space model, combines output current, input voltage, and a pre-stored state range to achieve fault detection, and corrects and predicts the state range when there is no fault. This scheme provides a technical foundation for fault detection in traction systems, but its core still relies on the state-space model and the state range update mechanism. For minor faults in high-speed train traction systems, when abnormal deviations are weak and easily masked by normal fluctuations and random disturbances, this type of detection method based on model consistency analysis may still suffer from problems such as weak anomalies being difficult to highlight, insufficient detection sensitivity, and low stability under complex disturbances.
[0005] Patent CN110991472A discloses a method for diagnosing minor faults in high-speed train traction systems. This method extracts a subset of historical signal data, constructs a first-layer model using random forest, extreme random tree, XGBoost, and LightGBM, and then combines this with logistic regression to form a stacked generalization structure, thereby achieving minor fault diagnosis. While this approach provides a data-driven implementation path for minor fault diagnosis, its focus is primarily on using existing fault data and ensemble classification models to identify fault categories. It does not adequately address key issues such as learning normal state distributions, enhancing weak anomaly deviations, and stable detection under complex disturbances. Therefore, when early-stage anomaly information in minor faults is not yet significant, relying solely on a supervised classification framework for diagnosis makes it difficult to simultaneously consider early detection sensitivity, detection stability, and fault type differentiation capabilities.
[0006] Currently, existing methods for fault detection and diagnosis in high-speed train traction systems still lack sufficient ability to represent weak anomalies in minor fault scenarios, exhibit low detection stability under complex disturbance conditions, and require further improvement in distinguishing between different fault types. Therefore, there is an urgent need to establish a diffusion model-based method for the detection and diagnosis of minor faults in high-speed train traction systems. This method would learn the normal state distribution using only fault-free operating samples and combine multi-diffusion step latent variable differences, smoothing processing, and continuous alarm decision mechanisms to achieve sensitive detection and accurate diagnosis of minor faults in high-speed train traction systems. Summary of the Invention
[0007] This invention aims to address the shortcomings of existing methods for detecting and diagnosing minor faults in high-speed train traction systems, such as weak fault features, insufficient detection sensitivity under complex disturbances, limited ability to distinguish different fault types, and heavy reliance on precise mechanistic models and expert experience. It proposes a diffusion model-based method for detecting and diagnosing minor faults in high-speed train traction systems. This method learns the normal state distribution of the traction system using only fault-free operating samples and constructs fault detection statistics based on the differences in latent variables across multiple diffusion steps. After detecting an anomaly, it further combines fault diagnosis features with a classification model to diagnose the fault type.
[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 monitoring data of the high-speed train traction system under normal and fault conditions, and align and standardize the original timing signals; the monitoring data includes at least two sets of variables such as DC bus voltage, three-phase current, train speed, traction motor speed and torque;
[0011] Step 2: Reconstruct the monitoring data using a sliding time window to construct a multivariate time series sample representing the dynamic operating state of the traction system;
[0012] Step 3: Extract a set of fault-free samples from the time series samples, construct a diffusion model, and use the fault-free samples to perform forward diffusion and reverse denoising training to learn the data distribution and structural characteristics of the high-speed train traction system under normal operating conditions, and obtain the parameters of the trained diffusion model.
[0013] Step 4: Select multiple diffusion steps to form a diffusion step set, use the trained diffusion model to calculate the latent variables of the fault-free sample under the multiple diffusion steps and their back estimation results, further calculate the differences of latent variables under each diffusion step, and construct a multi-diffusion step latent variable difference vector.
[0014] Step 5: Establish a fault detection statistic based on the difference vector of latent variables in the multi-diffusion step of fault-free samples, determine the fault detection threshold according to the statistical distribution of fault-free samples, and establish a smoothing mechanism and continuous alarm decision rules.
[0015] Step 6: Collect labeled fault samples, extract the latent variable difference features corresponding to them under multiple diffusion steps, and construct a fault diagnosis training set;
[0016] Step 7: Establish a fault diagnosis model based on the fault diagnosis training set, and use KNN-5 as the final fault diagnosis model.
[0017] Online fault detection and diagnosis:
[0018] Step 1: Collect online monitoring data under the current operating status, and construct online test samples using the sliding time window method consistent with offline modeling;
[0019] Step 2: Input the online test samples into the trained diffusion model, obtain the corresponding latent variables and their back estimation results under the multiple diffusion steps, and calculate the difference vector of latent variables under multiple diffusion steps;
[0020] Step 3: Calculate the fault detection statistic based on the difference vector of the latent variables in the multi-diffusion step, smooth the fault detection statistic, and compare it with the preset fault detection threshold; when the smoothed detection statistic exceeds the detection threshold for multiple consecutive time periods, output the fault alarm result.
[0021] Step 4: When it is determined that an online sample has failed, extract the differential features of the multi-diffusion step latent variables corresponding to the sample, input them into the fault diagnosis model, and output the corresponding fault type diagnosis result.
[0022] The beneficial effects of this invention include: First, this invention trains the diffusion model using only fault-free operating samples and establishes a fault-free reference by learning the distribution of normal states, without relying on precise mechanistic models and complex expert knowledge, thus exhibiting high flexibility and engineering applicability. Second, this invention characterizes minute anomaly information through the difference of latent variables in multiple diffusion steps, which can more effectively depict the degree of deviation of test samples from normal states and improve the distinguishability of minute faults. Third, this invention, by introducing smoothing processing and continuous alarm decision mechanisms, can reduce the impact of random noise, instantaneous fluctuations, and complex disturbances on detection results, thereby improving the sensitivity and stability of fault detection. Fourth, this invention further constructs fault diagnosis features based on the difference of latent variables in multiple diffusion steps and combines them with KNN-5 to achieve fault type identification, which can improve the ability to distinguish between different minute faults and the accuracy of diagnosis, thus having good engineering application value. Attached Figure Description
[0023] Figure 1 This is a general framework diagram of the diffusion model-based method for detecting and diagnosing minor faults in high-speed train traction systems as described in this invention.
[0024] Figure 2 The diagram shows the fault detection results under fault type f1.
[0025] Figure 3 The diagram shows the fault detection results under fault type f2.
[0026] Figure 4 The diagram shows the fault detection results under fault type f3.
[0027] Figure 5 The confusion matrix diagram represents the optimal classifier and the overall diagnostic process. Detailed Implementation
[0028] 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.
[0029] This embodiment provides a method for detecting and diagnosing minor faults in a high-speed train traction system based on a diffusion model. For example... Figure 1 The diagram shown is an overall framework diagram of the method of the present invention; Figures 2 to 4 The figures show the fault detection results under three types of minor fault conditions. Figure 5 The confusion matrix diagram represents the optimal classifier and the overall diagnostic process. Based on the above framework, the specific implementation steps are as follows:
[0030] Step 1: Collect monitoring data of the high-speed train traction system under normal and fault conditions, and align and standardize the original time-series signals.
[0031] At discrete sampling time The monitoring vector of the high-speed train traction system is defined as:
[0032] (1)
[0033] in, and These represent the two sets of DC bus voltages, , and Indicates three-phase current. Indicates train speed. Indicates the traction motor speed. This represents torque. The original time-domain signals from each acquired channel are time-aligned, and standardization is used to eliminate dimensional differences between different variables.
[0034] Step 2: Reconstruct the monitoring data using a sliding time window to construct a multivariate time series sample representing the dynamic operating state of the traction system.
[0035] Because the traction system operating data exhibits significant temporal correlation and variable coupling characteristics, a sliding time window is used to construct dynamic samples. Let the window length be... Then the first Each sample is represented as:
[0036] (2)
[0037] Therefore, each sample consists of continuous The monitoring vectors at each sampling time are stacked in chronological order to characterize the dynamic operating status of the high-speed train traction system over a period of time.
[0038] Step 3: Extract a set of fault-free samples from the time series samples, construct a diffusion model, and use the fault-free samples to perform forward diffusion and reverse denoising training to learn the data distribution and structural characteristics of the high-speed train traction system under normal operating conditions, and obtain the parameters of the trained diffusion model.
[0039] Extract fault-free samples from the dynamic samples constructed in step 2 to form a fault-free training set:
[0040] (3)
[0041] in, This represents the number of fault-free training samples. For any fault-free sample... Establish the forward diffusion process:
[0042] (4)
[0043] in, , Indicates the first Latent variables under each diffusion step Represents the noise scheduling coefficient. This represents the total number of diffusion steps. Further, it can be written as:
[0044] (5)
[0045] in,
[0046] (6)
[0047] In the reverse process, the diffusion model learns to gradually recover the representation of fault-free samples from noisy latent variables, and its reverse mapping is expressed as:
[0048] (7)
[0049] in, These are the model parameters. A noise prediction loss function is used during training:
[0050] (8)
[0051] in, Indicates Gaussian noise. The noise is represented by the model prediction. The parameters of the trained diffusion model are obtained by minimizing equation (8).
[0052] Step 4: Select multiple diffusion steps to form a diffusion step set, use the trained diffusion model to calculate the latent variables of the fault-free sample under the multiple diffusion steps and their back estimation results, further calculate the differences of latent variables under each diffusion step, and construct a multi-diffusion step latent variable difference vector.
[0053] Multiple diffusion steps are selected to form a diffusion step set:
[0054] (9)
[0055] For any sample The latent variables at each diffusion step are obtained through the forward diffusion process. And obtain the estimate of its previous diffusion step through the back-mapping after training:
[0056] (10)
[0057] in, This represents the inverse denoising mapping after training. In the... At each diffusion step, the difference in latent variables is defined as follows:
[0058] (11)
[0059] And the differences in latent variables across multiple diffusion steps are organized into a difference vector:
[0060] (12)
[0061] For fault-free samples, the evolution of their latent variables is consistent with the learned fault-free distribution, so the difference is small; for samples containing minor faults, the consistency of their latent variable evolution weakens, so the difference increases.
[0062] Step 5: Establish a fault detection statistic based on the difference vector of latent variables in the multi-diffusion step of fault-free samples, determine the fault detection threshold according to the statistical distribution of fault-free samples, and establish a smoothing mechanism and continuous alarm decision rules.
[0063] Covariance matrix estimation based on the difference vector of fault-free samples Further construct fault detection statistics:
[0064] (13)
[0065] in, This indicates the overall deviation of the sample from a fault-free reference state. To reduce the impact of measurement noise, instantaneous fluctuations, and complex disturbances, the detection statistic sequence is smoothed.
[0066] (14)
[0067] in, This represents the length of the smoothing window. The detection threshold is determined based on the statistical distribution of fault-free samples at a given significance level. To improve detection reliability, a continuous alarm decision rule is introduced: when the smoothed detection statistic satisfies...
[0068] (15)
[0069] The system has been determined to have malfunctioned. Indicates the length of continuous alarms.
[0070] Step 6: Collect labeled fault samples, extract the latent variable difference features corresponding to them under multiple diffusion steps, and construct a fault diagnosis training set.
[0071] For labeled fault samples, calculate the corresponding latent variable differences under multiple diffusion steps identical to step 4, and construct the fault diagnosis feature vector:
[0072] (16)
[0073] Therefore, a fault diagnosis training set is established:
[0074] (17)
[0075] in, Indicates the first Diagnostic feature vector of each fault sample, Indicates its fault category label, This represents the number of training samples for fault diagnosis.
[0076] Step 7: Establish a fault diagnosis model based on the fault diagnosis training set, and use KNN-5 as the final fault diagnosis model.
[0077] Based on the fault diagnosis training set constructed in step 6, multiple classification models, including KNN-5, LDA, Linear-SVM, RBF-SVM, and Bagged Trees, were compared, and KNN-5 was ultimately selected as the fault diagnosis model. For any sample to be diagnosed... The fault category prediction results are expressed as follows:
[0078] (18)
[0079] in, Indicates sample The set of 5 nearest neighbor samples, Indicates an indicator function.
[0080] Online fault detection and diagnosis
[0081] Step 1: Collect online monitoring data under the current operating status, and construct online test samples using the sliding time window method consistent with offline modeling.
[0082] The monitoring data collected from the currently operating traction system is aligned and standardized, and then an online test sample is constructed in accordance with the method described in formula (2).
[0083] Step 2: Input the online test samples into the trained diffusion model, obtain the corresponding latent variables and their inverse estimation results under the multiple diffusion steps, and calculate the difference vector of latent variables under multiple diffusion steps.
[0084] For each online test sample, the trained diffusion model and the pre-selected set of diffusion steps are used. Calculate the differences in latent variables at each diffusion step, and then calculate the difference vector corresponding to Equation (12). .
[0085] Step 3: Calculate the fault detection statistic based on the difference vector of the latent variables in the multi-diffusion step, smooth the fault detection statistic, and compare it with the preset fault detection threshold; when the smoothed detection statistic exceeds the detection threshold for multiple consecutive times, output the fault alarm result.
[0086] The fault detection statistic is calculated according to equation (13). The smoothed detection statistic is calculated according to equation (14). The fault alarm decision is made in conjunction with equation (15). If the continuous threshold crossing condition is not met, the system is determined to be operating normally; if the continuous threshold crossing condition is met, the fault alarm result is output.
[0087] Step 4: When it is determined that an online sample has failed, extract the differential features of the multi-diffusion step latent variables corresponding to the sample, input them into the fault diagnosis model, and output the corresponding fault type diagnosis result.
[0088] For online samples that have already triggered an alarm, their corresponding feature vectors are... Input the KNN-5 fault diagnosis model established in step 7 and output the corresponding fault category.
[0089] The above embodiments of the present invention are merely illustrative of the technical solutions 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 detecting and diagnosing minor faults in a high-speed train traction system based on a diffusion model, characterized in that, The method comprises two parts: offline modeling and online fault detection and diagnosis, specifically: Offline modeling: Step 1: Collect monitoring data of the high-speed train traction system under normal and fault conditions, and align and standardize the original timing signals; the monitoring data includes at least two sets of DC bus voltage, three-phase current, train speed, traction motor speed and torque; Step 2: Reconstruct the monitoring data using a sliding time window to construct a multivariate time series sample characterizing the dynamic operating state of the traction system; Step 3: Extract a set of fault-free samples from the multivariate time series samples, construct a diffusion model, and use the fault-free samples to perform forward diffusion and reverse denoising training to learn the data distribution and structural characteristics of the high-speed train traction system under normal operating conditions, and obtain the parameters of the trained diffusion model. Step 4: Select multiple diffusion steps to form a diffusion step set, use the trained diffusion model to calculate the latent variables of the fault-free sample under the multiple diffusion steps and their back estimation results, further calculate the differences of latent variables under each diffusion step, and construct a multi-diffusion step latent variable difference vector. Step 5: Establish a fault detection statistic based on the difference vector of latent variables in the multi-diffusion step of fault-free samples, determine the fault detection threshold according to the statistical distribution of fault-free samples, and establish a smoothing mechanism and continuous alarm decision rules. Step 6: Collect labeled fault samples, extract the latent variable difference features corresponding to them under multiple diffusion steps, and construct a fault diagnosis training set; Step 7: Establish a fault diagnosis model based on the fault diagnosis training set. Online fault detection and diagnosis: Step 1: Collect online monitoring data under the current operating status, and construct online test samples using the sliding time window method consistent with offline modeling; Step 2: Input the online test samples into the trained diffusion model, obtain the corresponding latent variables and their back estimation results under the multiple diffusion steps, and calculate the difference vector of latent variables under multiple diffusion steps; Step 3: Calculate the fault detection statistic based on the difference vector of the latent variables in the multi-diffusion step, smooth the fault detection statistic, and compare it with the preset fault detection threshold; when the smoothed detection statistic exceeds the detection threshold for multiple consecutive time periods, output the fault alarm result. Step 4: When it is determined that an online sample has failed, extract the differential features of the multi-diffusion step latent variables corresponding to the sample, input them into the fault diagnosis model, and output the corresponding fault type diagnosis result.
2. The method for detecting and diagnosing minor faults in a high-speed train traction system based on a diffusion model according to claim 1, characterized in that, The offline modeling steps 1 and 2 are specifically as follows: The high-speed train traction system at discrete sampling times The monitoring vector is represented as: (1) in, and These represent the two sets of DC bus voltages, , and Represents three-phase current. Indicates train speed. Indicates the traction motor speed. This indicates torque. Dynamic samples are constructed using a sliding time window: (2) in, The length of the sliding time window. For the first A dynamic sample.
3. The method for detecting and diagnosing minor faults in a high-speed train traction system based on a diffusion model according to claim 2, characterized in that, The offline modeling step 3 specifically involves: Extract fault-free samples from the dynamic samples constructed in step 2 to form a fault-free training set: (3) A forward diffusion process is established based on the aforementioned fault-free training set: (4) And a noise prediction loss function is used: (5) The diffusion model is trained to obtain the parameters of the trained diffusion model.
4. The method for detecting and diagnosing minor faults in a high-speed train traction system based on a diffusion model according to claim 3, characterized in that, The offline modeling step 4 specifically involves: Multiple diffusion steps are selected to form a diffusion step set: (6) The latent variable differences of samples at the multiple diffusion steps are calculated based on the trained diffusion model, and a multi-diffusion-step latent variable difference vector is constructed: (7) The multi-diffusion step latent variable difference vector is used to characterize the degree of deviation of a sample from the fault-free latent variable evolution mode.
5. The method for detecting and diagnosing minor faults in a high-speed train traction system based on a diffusion model according to claim 4, characterized in that, The offline modeling step 5 specifically involves: Covariance matrix estimation based on the difference vector of fault-free samples And construct fault detection statistics: (8) Smooth the detection statistic sequence: (9) When the smoothed detection statistic satisfies: (10) When this happens, the system is deemed to have malfunctioned; among them, This is the fault detection threshold. To smooth the window length, This is the continuous alarm length.
6. The method for detecting and diagnosing minor faults in a high-speed train traction system based on a diffusion model according to claim 5, characterized in that, The offline modeling step 6 specifically involves: For labeled fault samples, the differences in corresponding latent variables are calculated under multiple diffusion steps, and a fault diagnosis feature vector is constructed: (11) Therefore, a fault diagnosis training set is established: (12) in, Indicates the first Diagnostic feature vector of each fault sample, Indicates its fault category label, This represents the number of training samples for fault diagnosis.
7. The method for detecting and diagnosing minor faults in a high-speed train traction system based on a diffusion model according to claim 6, characterized in that, The fault diagnosis model is a K-nearest neighbor fault diagnosis model, where the number of nearest neighbors K is 5; for any sample to be diagnosed, its fault category prediction result is expressed as: (13) in, Indicates sample The set of 5 nearest neighbor samples, Indicates an indicator function.
8. A method for detecting and diagnosing minor faults in a high-speed train traction system based on a diffusion model, as described in any one of claims 1 to 6, characterized in that, The online fault detection and diagnosis specifically includes: Online monitoring data is collected from the currently operating traction system, and after alignment and standardization, online test samples are constructed in accordance with the method described in claim 2. Input the online test samples into the trained diffusion model and calculate the multi-diffusion step latent variable difference vector in accordance with the method described in claim 4; The fault detection statistic is calculated and smoothed according to the method described in claim 5. When the smoothed detection statistic exceeds the detection threshold for multiple consecutive times, the fault alarm result is output. For online samples that have triggered alarms, extract the corresponding multi-diffusion step latent variable difference features, input them into the fault diagnosis model, and output the fault category.