Traction system fault detection method based on neighborhood limitation generalized auto-encoder

Abstract

The invention discloses a traction system fault detection method based on a neighborhood limitation generalized auto-encoder, and belongs to the technical field of fault diagnosis. Aiming at the defects of large information loss, insufficient potential variable interpretation, poor adaptability to a nonlinear dynamic system and the like in the existing high-speed train traction system fault detection method, the method provided by the invention realizes high-precision fault detection through two stages of offline learning and online detection. An off-line learning stage: determining a sample neighborhood set and a weight by using a mahalanobis distance, fusing a local linear reconstruction error and a mutual information regular term to construct a neighborhood limitation generalized auto-encoder loss function, and training to obtain an optimal encoder and an optimal decoder; and finally, calculating a normal state residual error, and determining a fault detection threshold value based on the statistical magnitude. And an online detection stage: collecting data in real time and constructing stacked data, calculating a real-time residual error by using the trained neighborhood limitation generalized auto-encoder, and realizing fault detection by comparing a statistical magnitude with a threshold value.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault diagnosis, and specifically discloses a traction system fault detection method based on a neighborhood-restricted generalized autoencoder. Background Art

[0002] As high-speed trains evolve toward higher speeds and greater intelligence, the operational stability of the traction system, a core power component, is directly linked to train safety and the reliability of the entire rail transit system. Traction systems are subject to long-term high loads and complex operating conditions, making them prone to sensor failures and component damage. Failure to detect these problems in a timely manner can lead to train delays.

[0003] Existing fault detection methods for high-speed train traction systems are mainly divided into three categories: data-driven, model-driven, and knowledge-driven. Among them, data-driven methods are widely used because high-speed trains are equipped with a large number of sensors that can provide rich monitoring data. Among data-driven methods, fault detection methods based on machine learning have attracted the attention of many scholars. As an efficient machine learning tool, autoencoders perform well in feature extraction and nonlinear process modeling, and have been widely used in the field of fault diagnosis. The first half of the autoencoder is called the encoder, which can compress process data into latent variables. The second half of the autoencoder is called the decoder, which can reconstruct process data through latent variables. However, traditional autoencoders have the following limitations:

[0004] The latent variables lack interpretability and their information quality has not been effectively evaluated, resulting in significant information loss during data compression and reconstruction. The local linear structure and dynamic characteristics of the data are not fully considered, resulting in poor adaptability to nonlinear dynamic systems such as high-speed train traction systems.

[0005] Therefore, developing a fault detection method that can effectively reduce information loss, adapt to nonlinear dynamic characteristics and have high detection accuracy has become an urgent need for the safe operation and maintenance of high-speed train traction systems. Summary of the Invention

[0006] To address the shortcomings of existing high-speed train traction system fault detection methods, such as high information loss, poor adaptability to nonlinear dynamic systems, insufficient interpretability of latent variables, and low detection accuracy, this paper provides a traction system fault detection method based on a neighborhood-restricted generalized autoencoder. By introducing the concepts of mutual information and locally linear embedding, this method optimizes latent variables and the data reconstruction process to improve fault detection accuracy.

[0007] A traction system fault detection method based on a neighborhood-restricted generalized autoencoder includes two stages: offline learning and online detection. The specific steps are as follows:

[0008] Offline learning phase.

[0009] Online fault detection stage.

[0010] The specific process of the offline learning stage is as follows: Collect input from high-speed train traction systems during normal operation and output , the data is standardized to eliminate the dimension effect, where the offline data set consisting of input and output is: , in, Indicates the total number of samples of offline data; According to the dynamic characteristics of discrete time series, the sampled data is constructed as stacked data: , in, The stacking window size is 5 sampling points; when an additive fault occurs, the input and output signals are: , in, represents the fault amplitude of additive fault; By limiting the mutual information rate between data and latent variables, the information loss during data compression is reduced; the latent variables are optimized based on the mutual information minimization criterion: , in, is the encoding mapping function, is the mutual information rate; the norm of the latent variable and the data is achieved by minimizing the norm of the latent variable and the data: , in, represents the latent variable, represents the two-norm, Indicates absolute value; Calculate sample similarity based on Mahalanobis distance : , in, The neighborhood window size is 10 sampling points, is the covariance matrix of the data set; fully considering the relationship between the past, present and future : , in, is the square of the average Mahalanobis distance between all samples, Expressed as a natural constant An exponential function with base ; design a neighborhood algorithm based on Mahalanobis distance: , in, represents an indexed collection; therefore, The neighborhood weight of can be described as: , in, Indicates the value of the independent variable that makes the function achieve the minimum value; Screening effective neighborhoods by locally linear embedding criterion ,and ,in, is the neighborhood set of latent variables; Fusion of local linear embedding criterion and mutual information to construct the loss function of neighborhood-restricted generalized autoencoder : , in, is the decoder mapping function, is the regularization coefficient, Represents the function's compliance operator; the optimal encoder can be obtained by optimizing the loss function and decoder ; Designing a residual generator based on a neighborhood-restricted generalized autoencoder : ; Based on Taylor's formula, the residual signal can be analyzed as follows: , in, is a Hessian matrix, Indicates the partial derivative. When it is known, the last term is zero; therefore, ; design Test statistic: , in, is the covariance matrix of the fault-free residual signal data set; Design a threshold for the test statistic : ;

[0011] The specific process of the online fault detection stage is as follows: Collect input from online operations and output , build online stacking data ; Reconstruct data using trained neighborhood-restricted generalized autoencoders and calculate real-time residuals : ; Compute test statistics for real-time residuals , and perform fault detection tasks: .

[0012] Compared with the prior art, the present invention has the following beneficial effects:

[0013] By introducing the mutual information minimization criterion to design latent variables, key information is retained during data compression, improving the identifiability of fault characteristics.

[0014] By integrating local linear embedding and neighborhood constraints, the temporal correlation and local linearity of data are fully utilized to improve the reconstruction accuracy of nonlinear dynamic systems.

[0015] Without relying on precise system mathematical models, real-time fault detection of high-speed train traction systems can be achieved through data-driven means, making it suitable for complex industrial scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a framework diagram of the neighborhood-restricted generalized autoencoder neural network according to the present invention;

[0017] Figure 2 This is a framework diagram of the residual generator based on the neighborhood-restricted generalized autoencoder according to the present invention;

[0018] Figure 3-6 The traction system sensor of the present invention includes a data map of normal and faulty data;

[0019] Figure 7-10 They are respectively diagrams of the fault detection results based on the neighborhood-restricted generalized autoencoder described in the present invention. DETAILED DESCRIPTION

[0020] The present invention will be further described below in conjunction with the embodiments and accompanying drawings:

[0021] The high-speed train traction system is used as the research object. The experimental platform includes a permanent magnet synchronous motor (PMSM), a high-voltage control panel, a controller, and a data acquisition board. The main parameters are shown in Table 1 below: Table 1 Main parameters of traction system

[0022] Seven signals are used, including current 、 、 、 ,Voltage 、 and speed .in, 、 、 is a variable inside the controller of the traction system. 、 、 is the input signal of the traction system, is the output signal of the traction system.

[0023] Figure 3 The fluctuations of seven signals during stable operation are shown. In order to test the fault detection performance of the neighborhood-restricted generalized autoencoder, three different types of faults are injected into the traction system to conduct fault detection experiments. Figure 4-6 Describes signal fluctuations under the influence of faults.

[0024] The fault types are as follows: (1) Bias failure . At 40ms, the paranoid fault Injected into the sensor, the initial fault amplitude is 0.05 A. At 70ms, the fault amplitude increases by 0.05A. Figure 4 Due to the closed-loop control structure of the high-speed train traction system, when a bias fault is added Afterwards, there was no significant change in the seven sensor signals; (2) Constant fault At 40ms the sensor injects a constant fault , the fault amplitude is 0.2A. Figure 5 shows that after adding a constant fault, the current The signal changes significantly. Since the high-speed train traction system has strong robustness, after a period of constant fault, the input and output of the traction system gradually return to normal; (3) Intermittent failure Inject intermittent faults between 40ms and 70ms , the fault amplitude is 30km / h. Figure 6 Shows the changes in the vehicle speed sensor after an intermittent fault occurs. When an intermittent fault occurs, the system input and output cannot return to normal state; When it disappears, the system returns to normal state.

[0025] The false alarm rate (FAR) and missed detection rate (MDR) are defined as: , in, Indicates the total number of fault-free Indicates the total number of failures, represents the total number of false positives, Indicates the total number of missed reports.

[0026] Figure 7-10 The detection results for three different types of faults using a neighborhood-restricted generalized autoencoder and a traditional autoencoder are shown. The dashed line represents the threshold, and the solid line represents the test statistic. Because the neighborhood-restricted generalized autoencoder is an improved autoencoder, the detection results are similar, but also significantly different.

[0027] exist Figure 7 In the example, when a bias fault with a fault amplitude of 0.05A occurs, the traditional autoencoder method has poor detection effect. When a bias fault with a fault amplitude of 0.1A occurs, the traditional autoencoder method can accurately detect it. Finally, no matter the bias fault has an amplitude of 0.05A or 0.1A, the fault detection method based on the neighborhood restricted generalized autoencoder can respond quickly and effectively detect the occurrence of the fault.

[0028] exist Figure 8 In this study, both traditional autoencoders and neighborhood-restricted generalized autoencoders performed poorly in fault detection. This is due to the inherent robustness of the high-speed train traction system. The input and output signals of the high-speed train traction system return to normal after 50 milliseconds, making them ineffective for traditional autoencoders and neighborhood-restricted generalized autoencoders.

[0029] exist Figure 9 In the event of an intermittent fault Finally, the fault detection method based on neighborhood restricted generalized autoencoder has good detection performance. However, the fault detection method based on traditional autoencoder has a high MDR.

[0030] When joining As input signal, add 、 As the output signal, Figure 10 In the fault detection method based on neighborhood restricted generalized autoencoder, At this time, the fault detection method based on the neighborhood-restricted generalized autoencoder has an MDRs of 3.00% and a FARs of 1.00%.

[0031] The fault detection method based on neighborhood-restricted generalized autoencoders is compared with four methods: principal component analysis, independent component analysis, autoencoders, and local linear generalized autoencoders. The evaluation metrics are summarized in Table 2. Table 2 Comparison of fault detection performance

[0032] Table 3 shows the ablation study results of the fault detection method based on neighborhood-restricted generalized autoencoders in P1 (only applying the minimum statistic) and P2 (only applying the local linear embedding), verifying the effectiveness of the proposed method. Table 3 Ablation study of fault detection method based on neighborhood restricted generalized autoencoder

Claims

1. A traction system fault detection method based on a neighborhood-restricted generalized autoencoder, characterized in that: It includes two stages: offline learning and online detection. The specific steps are as follows: Collect input from high-speed train traction systems during normal operation and output , the data is standardized to eliminate the dimension effect, where the offline data set consisting of input and output is: ， in, Indicates the total number of samples of offline data; According to the dynamic characteristics of discrete time series, the sampled data is constructed as stacked data: ， in, The stacking window size is 5 sampling points; when an additive fault occurs, the input and output signals are: ， in, represents the fault amplitude of additive fault; By limiting the mutual information rate between data and latent variables, the information loss during data compression is reduced; the latent variables are optimized based on the mutual information minimization criterion: ， in, is the encoding mapping function, is the mutual information rate; the norm of the latent variable and the data is achieved by minimizing the norm of the latent variable and the data: ， in, represents the latent variable, represents the two-norm, Indicates absolute value; Calculate sample similarity based on Mahalanobis distance : ， in, The neighborhood window size is 10 sampling points, is the covariance matrix of the data set; fully considering the relationship between the past, present and future : ， in, is the square of the average Mahalanobis distance between all samples, Expressed as a natural constant An exponential function with base ; design a neighborhood algorithm based on Mahalanobis distance: ， in, represents an indexed collection; therefore, The neighborhood weight of can be described as: ， in, Indicates the value of the independent variable that makes the function achieve the minimum value; Screening effective neighborhoods by locally linear embedding criterion ,and ,in, is the neighborhood set of latent variables; Fusion of local linear embedding criterion and mutual information to construct the loss function of neighborhood-restricted generalized autoencoder : ， in, is the decoder mapping function, is the regularization coefficient, Represents the function's compliance operator; the optimal encoder can be obtained by optimizing the loss function and decoder ; Designing a residual generator based on a neighborhood-restricted generalized autoencoder : ； Based on Taylor's formula, the residual signal can be analyzed as follows: ， in, is a Hessian matrix, Indicates the partial derivative. When it is known, the last term is zero; therefore, ; design Test statistic: ， in, is the covariance matrix of the fault-free residual signal data set; Design a threshold for the test statistic : ； Collect input from online operations and output , build online stacking data ; Reconstruct data using trained neighborhood-restricted generalized autoencoders and calculate real-time residuals : ； Compute test statistics for real-time residuals , and perform fault detection tasks: 。

Images