Traction system fault detection method based on neighborhood restricted generalized autoencoder

Abstract

The application discloses a traction system fault detection method based on neighborhood restriction generalized autoencoder and belongs to the technical field of fault diagnosis. In view of defects such as great information loss, insufficient interpretability of potential variables and poor adaptability to nonlinear dynamic systems in existing high-speed train traction system fault detection methods, the method realizes high-precision fault detection through two stages of offline learning and online detection. In the offline learning stage, the Mahalanobis distance is used to determine a sample neighborhood set and weights, a neighborhood restriction generalized autoencoder loss function is constructed by fusing local linear reconstruction error and mutual information regularization terms, and optimal encoders and decoders are trained. Finally, normal state residuals are calculated, and a fault detection threshold is determined based on statistics. In the online detection stage, data are collected in real time and stacked data are constructed, real-time residuals are calculated by using the trained neighborhood restriction generalized autoencoder, and fault detection is realized by comparing statistics and the threshold.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology, and specifically discloses a fault detection method for traction systems based on a neighborhood-restricted generalized autoencoder. Background Technology

[0002] As high-speed trains develop towards higher speeds and greater intelligence, the operational stability of the traction system, as a core power component, directly affects train operation safety and the reliability of the entire rail transit system. Traction systems operate under high loads and complex conditions for extended periods, making them prone to sensor malfunctions and component damage. Failure to detect these issues promptly can lead to train delays.

[0003] Existing fault detection methods for high-speed train traction systems are mainly classified into three categories: data-driven, model-driven, and knowledge-driven. Among them, data-driven methods are widely used due to the abundant monitoring data provided by the numerous sensors equipped on high-speed trains. Within data-driven methods, machine learning-based fault detection methods have attracted considerable attention from scholars. Autoencoders, as an efficient machine learning tool, excel in feature extraction and nonlinear process modeling and have been widely used in the field of fault diagnosis. The first part of an autoencoder, called the encoder, compresses process data into latent variables. The second part, called the decoder, reconstructs the process data from the latent variables. However, traditional autoencoders have the following limitations:

[0004] The lack of interpretability of latent variables and the failure to effectively assess their information quality lead to significant information loss during data compression and reconstruction. Furthermore, the lack of sufficient consideration of the local linear structure and dynamic characteristics of the data results 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 achieve 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 significant information loss, poor adaptability to nonlinear dynamic systems, insufficient interpretability of latent variables, and low detection accuracy, this invention provides a traction system fault detection method based on a neighborhood-restricted generalized autoencoder. By introducing the concepts of mutual information and local linear embedding, the latent variables and data reconstruction process are optimized to improve the accuracy of fault detection.

[0007] A fault detection method for traction systems 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 phase.

[0010] The specific process of the offline learning phase is as follows:

[0011] Inputs from the high-speed train traction system during normal operation and output The data is standardized to eliminate the influence of units of measurement. The offline dataset, consisting of input and output, is as follows:

[0012] ,

[0013] in, This represents the total number of samples in the offline data.

[0014] Based on the dynamic characteristics of discrete time series, the sampled data is constructed as stacked data:

[0015] ,

[0016] in, The stacking window size is 5 sampling points; in the event of an additive fault, the input and output signals are:

[0017] ,

[0018] in, Indicates the magnitude of an additive fault;

[0019] By limiting the mutual information rate between the data and latent variables, information loss during data compression is reduced; latent variables are optimized based on the mutual information minimization criterion.

[0020] ,

[0021] in, For encoding mapping functions, Mutual information rate; the norm of latent variables and data is achieved by minimizing the norm of latent variables and data:

[0022] ,

[0023] in, Representing latent variables, Represents the L2 norm, Represents absolute value;

[0024] Calculating sample similarity based on Mahalanobis distance :

[0025] ,

[0026] in, The neighborhood window size is 10 sampling points. It is the covariance matrix of the dataset; it fully considers the relationships between the past, present, and future. :

[0027] ,

[0028] in, It is the square of the average Mahalanobis distance among all samples. Represented by natural constant An exponential function with base 0; design a neighborhood algorithm based on Mahalanobis distance:

[0029] ,

[0030] in, Represents an index set; therefore, The neighborhood weights can be described as:

[0031] ,

[0032] in, This represents the value of the independent variable that makes the function reach its minimum value;

[0033] Valid neighborhoods are selected using the local linear embedding criterion. ,and ,in, For the neighborhood set of latent variables;

[0034] By integrating the local linear embedding criterion and mutual information, a loss function for a neighborhood-restricted generalized autoencoder is constructed. :

[0035] ,

[0036] in, It is the decoder mapping function. It is the regularization coefficient. The operator represents the convention of a function; the optimal encoder can be obtained by optimizing the loss function. and decoder ;

[0037] Design a residual generator based on a neighborhood-restricted generalized autoencoder. :

[0038] ;

[0039] Based on Taylor's formula, analysis of the residual signal yields the following:

[0040] ,

[0041] in, It is a Hessian matrix. This indicates taking the partial derivative; given the given information, the last term is zero. Therefore... ;

[0042] design Test statistic:

[0043] ,

[0044] in, It is the covariance matrix of the fault-free residual signal dataset;

[0045] Design a threshold for the test statistic. :

[0046] ;

[0047] The specific process of the online fault detection phase is as follows:

[0048] Collect input from online operation and output Build online stacked data ;

[0049] Reconstructing data using a trained neighborhood-restricted generalized autoencoder and calculating real-time residuals. :

[0050] ;

[0051] Calculate the test statistic for real-time residuals And perform fault detection tasks:

[0052] .

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] By introducing the mutual information minimization criterion to design potential variables, we can ensure the preservation of key information during data compression and improve the identifiability of fault characteristics.

[0055] By combining local linear embedding and neighborhood constraints, and making full use of the temporal correlation and local linearity of data, the reconstruction accuracy of nonlinear dynamic systems can be improved.

[0056] It can achieve real-time fault detection of high-speed train traction systems without relying on precise system mathematical models, and is suitable for complex industrial scenarios. Attached Figure Description

[0057] Figure 1This is a framework diagram of the neighborhood-restricted generalized autoencoder neural network described in this invention;

[0058] Figure 2 This is a framework diagram of the residual generator based on the neighborhood-restricted generalized autoencoder described in this invention;

[0059] Figure 3-6 The traction system sensor described in this invention includes data graphs showing no faults and faults;

[0060] Figure 7-10 The images shown are fault detection results based on the neighborhood-restricted generalized autoencoder described in this invention. Detailed Implementation

[0061] The present invention will be further described below with reference to the embodiments and the accompanying drawings:

[0062] Taking the high-speed train traction system 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.

[0063] Table 1 Main parameters of the traction system

[0064]

[0065] Seven signals were used, including current. , , , ,Voltage , and speed .in, , , These are variables within the controller of the traction system. , , It is the input signal of the traction system. It is the output signal of the traction system.

[0066] Figure 3 The fluctuations of seven signals during stable operation are shown. To verify the fault detection performance of the neighborhood-restricted generalized autoencoder, three different types of faults were injected into the traction system for fault detection experiments. Figure 4-6 It describes the signal fluctuations under the influence of faults.

[0067] The fault types are as follows:

[0068] (1) Bias fault The paranoid fault will occur at 40ms. The fault amplitude was initially 0.05A when injected into the sensor. The fault amplitude increased by 0.05A at 70ms. Figure 4 The signal fluctuated. Due to the closed-loop control structure of the high-speed train traction system, the addition of a bias fault... Afterwards, the signals from the seven sets of sensors did not change significantly;

[0069] (2) Constant Fault A constant fault was injected into the sensor at 40ms. The fault amplitude is 0.2A. Figure 5 This demonstrates the current after adding a constant fault. The signal changes significantly. Because the high-speed train traction system is highly robust, after a period of time following a constant fault, the traction system's input and output gradually return to normal.

[0070] (3) Intermittent failure Inject intermittent faults between 40ms and 70ms. The fault range is 30 km / h. Figure 6 This shows the change in the vehicle speed sensor after an intermittent fault occurred. When an intermittent fault... When this occurs, the system inputs and outputs cannot return to normal; when the fault is intermittent... When it disappears, the system returns to normal.

[0071] The false alarm rate (FAR) and the missed detection rate (MDR) are defined as follows:

[0072] ,

[0073] in, This represents the total number of fault-free items. Indicates the total number of faults. Indicates the total number of false alarms. This indicates the total number of cases that were missed.

[0074] Figure 7-10 The detection results for three different types of faults are presented using a neighborhood-restricted generalized autoencoder (GAO) and a traditional autoencoder method. Dashed lines represent thresholds, and solid lines represent test statistics. Since the neighborhood-restricted GAO is an improved method of autoencoders, the detection results show similarities but also significant differences.

[0075] exist Figure 7 In the context of fault detection, the traditional self-encoder method performs poorly when a bias fault with an amplitude of 0.05A occurs. However, it can accurately detect a bias fault with an amplitude of 0.1A. Subsequently, regardless of whether the bias fault amplitude is 0.05A or 0.1A, the fault detection method based on the neighborhood-restricted generalized autoencoder can react quickly and effectively detect the occurrence of the fault.

[0076] exist Figure 8 In this study, both traditional autoencoders and neighborhood-restricted generalized autoencoders yielded poor fault detection results. This is because the high-speed train traction system itself possesses strong robustness. The input and output signals of the high-speed train traction system tend to normalize after 50ms, making traditional autoencoders and neighborhood-restricted generalized autoencoders ineffective in detecting faults.

[0077] exist Figure 9 In the event of intermittent failure Subsequently, the fault detection method based on neighborhood-restricted generalized autoencoders exhibits good detection performance. However, traditional autoencoder fault detection methods suffer from high mean deviation rate (MDR).

[0078] When joining As input signal, added , After being used as an output signal, Figure 10 In the context of fault detection methods based on neighborhood-restricted generalized autoencoders, the following applies: It then responds immediately. At this point, its fault detection method based on a neighborhood-restricted generalized autoencoder has an MDR of 3.00% and a FAR of 1.00%.

[0079] The fault detection method based on neighborhood-restricted generalized autoencoders is compared with four other methods: principal component analysis, independent component analysis, autoencoders, and locally linear generalized autoencoders. Evaluation metrics are summarized in Table 2.

[0080] Table 2 Comparison of Fault Detection Performance

[0081]

[0082] Table 3 shows the ablation study results of the fault detection method based on neighborhood-restricted generalized autoencoders in P1 (applying only minimum statistics) and P2 (applying only local linear embedding), which verifies the effectiveness of the proposed method.

[0083] Table 3 Ablation Study of Fault Detection Method Based on Neighborhood-Restricted Generalized Autoencoder

[0084]

Claims

1. A method for traction system fault detection based on neighborhood-constrained generalized autoencoder, characterized in that, The method includes two stages: offline learning and online detection. The specific steps are as follows: Inputs from the high-speed train traction system during normal operation and output The data is standardized to eliminate the influence of units of measurement. The offline dataset, consisting of input and output, is as follows: ， wherein, represents the total number of samples of offline data; According to the dynamic characteristics of the discrete time series, the sampling data is constructed into stacked data: ， wherein is the stack window size of 5 samples; in the presence of an additive fault, the input and output signals are: ， wherein denotes the fault magnitude of an additive fault; By limiting the mutual information rate of data and latent variables, the information loss in data compression process is reduced; and the latent variables are optimized based on the mutual information minimization criterion: ， wherein, is an encoding mapping function, is a mutual information rate; the norm minimization of the latent variable and the data is achieved by the norm minimization of the latent variable and the data: ， wherein, denotes a latent variable, denotes a two-norm, denotes an absolute value; Calculating sample similarity based on mahalanobis distance : ， wherein is the neighborhood window size of 10 sample points, is the covariance matrix of the dataset; taking into account the relationship between past, present and future : ， where is the square of the average Mahalanobis distance between all samples, denotes the exponential function with base of the natural constant ; design neighborhood algorithm based on Mahalanobis distance: ， wherein, denotes a set of indices; thus, The neighborhood weight of can be described as: ， wherein denotes the value of the argument of the function that yields the minimum value; Filtering effective neighborhoods by local linear embedding criteria , and wherein, is a set of potential variable neighborhoods; The loss function of the neighborhood-restricted generalized autoencoder is constructed by fusing the local linear embedding criterion and mutual information : ， wherein, is a decoder mapping function, is a regularization coefficient, denotes a composition operator of functions; the optimal encoder and decoder are found by optimizing a loss function Designing residual generator based on neighborhood restricted generalized autoencoder : ； Based on Taylor formula, the residual signal is analyzed to obtain ， where is a Hessian matrix, denotes partial differentiation, and the last term is zero in the known case; thus, ; Design Test statistic: ， wherein, is the covariance matrix of the fault-free residual signal dataset; Designing threshold values for inspection statistics : ； Collecting input on-line running and output , building on-line stack data ; reconstructing the data using the trained neighborhood-constrained generalized autoencoder, calculating real-time residuals : ； Computing a test statistic for real-time residuals and perform a fault detection task: 。

Images