Early fault diagnosis method for high-speed train suspension system

Abstract

The invention discloses an early fault diagnosis method for a high-speed train suspension system. The early fault diagnosis method comprises the following steps: step 1, respectively acquiring data sets under a normal working condition and a fault working condition of the high-speed train suspension system; 2, acquiring a load matrix of the data set processed in the step 1; 3, converting the load matrix into an overload matrix in a PRPCA residual error model, and constructing the PRPCA residual error model to obtain a residual error of the PRPCA residual error model of the data set; 4, setting a fault detection threshold value; and step 5, training fault identification and classification of the support vector machine through the overload matrix and the model residual error of the PRPCA, and using the trained support vector machine to diagnose the early fault of the high-speed train suspension system. According to the method, the measured data can be projected to a new space in a non-linear manner, so that the initial fault data and the normal data are effectively separated, the processed data contain more fault feature information, and more effective early fault detection is realized.

Description

Technical Field

[0001] The invention relates to the technical field of high-speed train suspension system fault diagnosis, and in particular to an early fault diagnosis method for a high-speed train suspension system based on PRPCA and SVM. Background Art

[0002] The suspension system is one of the important components of high-speed trains. It not only supports the bogie and the train body, but also isolates the force generated by the unevenness of the track, ensuring the riding experience of the passengers in the carriage. As the most important part of the bogie, the suspension system generally adopts a two-system suspension. The first-system suspension device mainly includes the axle box spring and the axle box vertical shock absorber, which are located between the axle box and the bogie frame; the second-system suspension device mainly consists of components such as traction device, shock absorber and air spring device, which are located between the bogie frame and the body. As the train runs on track for a longer time, some equipment in the suspension system will age and be damaged, such as air spring devices, shock absorbers, suspension couplings and suspension controllers, which will cause early failures such as slight cracking or deformation of the spring, damage to the seal inside the shock absorber, slight wear of the suspension coupling or partial failure of the suspension control electronic components, which may have an adverse effect on the performance and safety of the suspension system and bring potential dangers to the driving safety of the train.

[0003] At present, the existing early fault diagnosis methods for high-speed train suspension systems based on the combination of principal component analysis (PCA) and support vector machine (SVM) have some limitations in processing process data. The traditional method first uses PCA to reduce the dimension of the data, and then uses SVM for fault detection and classification. However, since PCA is a linear dimensionality reduction method, it cannot fully capture the relationship between complex nonlinear features in the data, resulting in the extracted feature information may be incomplete or insufficient to represent the actual system state. Therefore, the limitations of PCA on data processing affect the accuracy of fault detection and SVM-based fault classification. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a method for early fault diagnosis of a high-speed train suspension system to solve the technical problem that the prior art cannot fully capture the relationship between complex nonlinear features in the data, resulting in the extracted feature information being incomplete or insufficient to represent the actual system state.

[0005] The present invention provides a method for diagnosing early faults of a high-speed train suspension system, comprising the following steps:

[0006] Step 1: Obtain data sets under normal working conditions and fault conditions of the high-speed train suspension system respectively, and standardize the data sets;

[0007] Step 2: Obtain the load matrix of the data set processed in step 1;

[0008] Step 3: Convert the load matrix into the overload matrix in the PRPCA residual model, and construct the PRPCA residual model to obtain the residual of the PRPCA residual model of the data set;

[0009] Step 4: Set the fault detection threshold;

[0010] Step 5: The fault recognition and classification of the support vector machine are trained through the overload matrix and the model residual of PRPCA, and the trained support vector machine is used to diagnose the early faults of the high-speed train suspension system.

[0011] Furthermore, in step 2, the specific method for obtaining the load matrix is: performing principal component analysis on the data set processed in step 1, performing singular value decomposition on the covariance matrix corresponding to the analysis result, and obtaining the load matrix of the data set.

[0012] Furthermore, the specific method of setting the fault detection threshold in step 4 is as follows:

[0013] The Wasserstein distance index between the fault condition and the normal condition is obtained according to the model residual of the PRPCA residual model, and the fault detection threshold is set according to the Wasserstein distance.

[0014] Furthermore, the calculation formula of the Wasserstein distance is:

[0015]

[0016] Where R is the model residual under normal conditions; R if is the model residual under fault conditions; x and y are n-dimensional variables respectively.

[0017] Furthermore, in the Wasserstein distance calculation process, the expression of 2-Wasserstein distance between n-dimensional variables x and y is:

[0018]

[0019] Where Tr(·) is the trace of the matrix; ∑1=∑ 1 / 2 ∑ if ∑ 1 / 2 , ∑ is the covariance matrix of the n-dimensional variable x, ∑ if is the covariance matrix of the n-dimensional variable y; μ1=||μ-μif ||2, μ is the mean of the n-dimensional variable x, μ if is the mean of the n-dimensional variable y.

[0020] Beneficial effects of the present invention:

[0021] The present invention uses the PRPCA method to process process data, which can nonlinearly project the measured data into a new space, thereby effectively separating the initial fault data from the normal data, solving the problem that traditional PCA cannot fully capture the nonlinear fault characteristics in the process data, so that the processed data contains more fault characteristic information, thereby achieving more effective early fault detection.

[0022] The present invention uses a nonlinear PRPCA algorithm to reduce the dimension of the original fault data, groups the large sample data after dimensionality reduction, then uses variance to measure the degree of deviation of each group of data, and then uses a support vector machine to classify the fault type, thereby further improving the accuracy of SVM classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The features and advantages of the present invention will be more clearly understood by referring to the accompanying drawings, which are schematic and should not be construed as limiting the present invention in any way. In the accompanying drawings:

[0024] Figure 1 It is a schematic diagram of a flow chart of a specific embodiment of the present invention;

[0025] Figure 2 This is a comparison diagram of the effects of the present invention and PCA in IDV3 fault detection;

[0026] Figure 3 This is a comparison diagram of the effects of the present invention and PCA in IDV0 fault detection;

[0027] Figure 4 This is a comparison diagram of the effects of the present invention and PCA in IDV1 fault detection;

[0028] Figure 5 This is a comparison diagram of the effects of the present invention and PCA in IDV2 fault detection;

[0029] Figure 6 This is a comparison chart of the classification performance of the present invention and PCA-SVM under two types of fault conditions: IDV0 and IDV3;

[0030] Figure 7 This is a comparison chart of the classification performance of the present invention and PCA-SVM under the three types of faults IDV0, IDV2, and IDV3;

[0031] Figure 8It is a comparison chart of classification performance between the present invention and PCA-SVM under four types of fault conditions: IDV0, IDV1, IDV2, and IDV3. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0033] The present invention is further illustrated below in conjunction with specific embodiments. Those skilled in the art should understand that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention, and modifications to various equivalent forms of the present invention fall within the scope defined by the appended claims of this application.

[0034] like Figure 1 As shown, the present invention provides a method for early fault diagnosis of a high-speed train suspension system, comprising the following steps:

[0035] Step 1: Obtain data sets under normal working conditions and fault conditions of the high-speed train suspension system respectively, and standardize the data sets;

[0036] The formula for standardizing the original data is as follows:

[0037]

[0038] In the formula, xi j is the sample corresponding to the i-th row and j-th column; s j is the standard deviation of the samples in the jth column; x s·j is the j-th column sample after standardization and centering.

[0039] The definition of the processed raw data is:

[0040] X=[Z C ,Z B1 ,Z B2 ,P C ,P B1 ,P B2 ,T C ,T B1 ,T B2 ]

[0041] Where X∈R N×m is N samples measured by m sensors, m ≥ 1; Z C , Z B1 , ZB2 Represent the vertical displacement of the car body and bogie respectively; P C , P B1 , P B2 The copy represents the pitch angle of the car body and bogie; T C , T B1 , T B2 Represent the inclination angles of the car body and bogie respectively.

[0042] Step 2: Perform principal component analysis on the data set processed in step 1, perform singular value decomposition on the covariance matrix corresponding to the analysis result, and obtain the load matrix P of the data set to achieve preliminary dimensionality reduction and feature extraction of the data set;

[0043] Step 3: Convert the load matrix into the overload matrix in the PRPCA residual model, and construct the PRPCA residual model to obtain the residual of the PRPCA residual model of the data set;

[0044] In the process of converting the load matrix into the overload matrix in the PRPCA residual model, the probability density function is first obtained by the kernel density estimation (KDE) method:

[0045]

[0046] Where, t j represents the j-th column element of the score vector T; N is the number of samples; h is the bandwidth; K is the kernel function;

[0047] Then determine the probability correlation matrix W by the probability density function:

[0048] W=diag(w1,…,w m )

[0049] Among them, the boundary parameter β is given, and P is determined according to the boundary parameter and probability density function a and P b , and then determine the probability correlation coefficient w based on the Logistic function:

[0050]

[0051] Finally, the overload matrix P in the PRPCA algorithm is obtained by the following formula New :

[0052] P New =W -1 / 2 P

[0053] The process of constructing the PRPCA residual model is:

[0054] Determine the number of pivots r:

[0055]

[0056] CVR(r)≥v

[0057] Where CVR represents the cumulative variance contribution rate, v is the confidence ratio of the variance, 0≤v≤1.

[0058] After the dimension is confirmed, the PRPCA model under the r dimension can be expressed as:

[0059]

[0060] Where P rNew is the matrix P New The overload matrix after retaining the first r-dimensional eigenvectors; X fc It can be used as the original data for subsequent fault classification.

[0061] The calculation formula of the PRPCA model residual is as follows:

[0062] R=XX fc

[0063] Step 4: Obtain the Wasserstein distance index between the fault condition and the normal condition based on the model residual of the PRPCA residual model, and set the fault detection threshold based on the Wasserstein distance. The specific process is as follows:

[0064] Select two sets of offline data under normal working conditions and process them in the previous steps to obtain the new model residuals R and R if , R is used as the reference signal and is set to X, R if Let Y be the moving window size and m, and then find the mean and variance of X and Y. The Wasserstein distance is as follows:

[0065]

[0066] Where R is the model residual under normal conditions; R if is the model residual under fault conditions; x and y are n-dimensional variables respectively.

[0067] Assume that the distributions of n-dimensional variables x and y are both Gaussian distributions with means μ and μ respectively. if ; The covariance matrices are ∑ and ∑ if Therefore, the expression for the 2-Wasserstein distance between x and y can be expressed as follows:

[0068]

[0069] Where Tr(·) is the trace of the matrix; ∑1=∑ 1 / 2 ∑ if ∑1 / 2 , ∑ is the covariance matrix of the n-dimensional variable x, ∑ if is the covariance matrix of the n-dimensional variable y; μ1=||μ-μ if ||2, μ is the mean of the n-dimensional variable x, μ if is the mean of the n-dimensional variable y.

[0070] The moving window can be moved forward one sample at a time to obtain the mean μ of W2(P,Q) WD and variance σ WD , until the last sample of R.

[0071] The threshold h is obtained using the threshold calculation formula, which is as follows:

[0072] h=μ WD +4σ WD

[0073] In the formula, μ WD and σ WD are the mean and standard deviation of WD under normal working conditions.

[0074] For the early failure of the high-speed train suspension system actuator and secondary suspension, the fault detection mechanism is set as follows:

[0075]

[0076] Where W2 is the Wasserstein distance indicator.

[0077] Step 5: The fault identification and classification of the support vector machine (SVM) are trained through the overload matrix and the model residual of PRPCA, and the trained support vector machine is used to diagnose the early faults of the high-speed train suspension system. The SVM construction process is as follows:

[0078] For a given training data x i ∈R m ,y i ∈{1,2,3,4},(i=1,2,3,...,N), where m is the dimension of the input data and N is the number of training data for SVM. SVM classification can be expressed as the following quadratic programming (QP) problem:

[0079]

[0080]

[0081] i=1,2,3,4,...,N

[0082] Where w is the normalized weight vector of the same dimension as x; b is the normalized deviation of the hyperplane; and g i The inequality is transformed into the following equation:

[0083]

[0084] Where w is the normalized weight vector of the same dimension as x; b is the normalized deviation of the hyperplane.

[0085] The KKT condition is obtained by using the Lagrange multiplier method:

[0086]

[0087] The KKT condition problem can be transformed into an equivalent Lagrangian dual problem:

[0088]

[0089] st i ∈0,i=1,2,3,4,...,N

[0090] In the formula, λ i Represents the Lagrange multiplier of the i-th sample. Solving the dual problem has the following two advantages: first, it can make the expression and constraint conditions more concise; second, it is given by It can be seen that the optimal solution is determined only by the dot product result of the support vector, that is, it is only determined by the spatial similarity of the support vector.

[0091] SVM constructs a soft interval classifier by introducing slack variables, and then adds a penalty factor before the slack variables to indicate the tolerance for the loss caused by outliers. When C approaches infinity, the decision hyperplane is forced to correctly divide all samples, causing the classifier to degenerate into a hard interval classifier. The specific formula is as follows:

[0092]

[0093] i=1,2,3,4,...,N

[0094] Where ξ is the slack variable and C is the penalty factor.

[0095] The kernel function is used to nonlinearly map the nonlinear inseparable model in the low-dimensional space to the high-dimensional space, transforming it into a linearly separable model, where the linear decision function in the dual form is:

[0096]

[0097] Where sign(·) is a sign function; k′ is a kernel function that returns the dot product of the feature space mapping of a data point in the low-dimensional space.

[0098] The present invention adopts Gaussian function as the kernel function of SVM, and the formula is as follows:

[0099]

[0100] Where γ is the bandwidth of the Gaussian kernel function.

[0101] In the specific embodiment of the present invention, a comprehensive evaluation index is also constructed to evaluate the method of the present invention, which is as follows:

[0102] Set the detection progress: Detection accuracy is used to describe the percentage of correctly detected samples, that is, the ratio of correctly detected original samples to faulty samples. The formula for detection accuracy is as follows:

[0103]

[0104] In the formula, TP (true positive) represents the number of faulty samples that are correctly detected; TN (true negative) represents the number of normal samples that are correctly detected; FP (false positive) represents the number of normal samples that are incorrectly detected as faulty and FN (false negative) represents the number of faulty samples that are incorrectly detected as normal.

[0105] The specific embodiment of the present invention comprehensively considers the fault detection and fault identification accuracy, and adopts the F1 evaluation index for comprehensive evaluation. The specific formula of the F1 evaluation index is as follows:

[0106]

[0107] In the formula, is the fault detection accuracy, is the fault classification accuracy.

[0108] The simulation results of the method of the present invention and the existing method are compared through the following simulation experiment process:

[0109] Step A1: Set the characteristic information IDV0 of the actuator fault as 10% failure of the front bogie actuator, the fault injection time as the 20th second, the fault process time as 20 seconds to 50 seconds, and perform software injection of the fault through the fault injection module to establish a fault model;

[0110] Step A2: Set the characteristic information IDV1 of the mechanical component failure to be 10% failure of the secondary suspension spring, IDV2 to be 10% failure of the secondary suspension damper, the fault injection time to be the 20th second, the fault process time to be 20 seconds to 50 seconds, and perform software injection of the fault through the fault injection module to establish a fault model;

[0111] Step A3: Set the characteristic information IDV3 of the sensor fault as 10% sensor failure, the fault start time as the 20th second, and the fault process time as 20 seconds to 50 seconds. The fault is injected into the software through the fault injection module to establish a fault model.

[0112] Step A4: Import the track-train coupling model established in SIMPACK into Matlab / Simulink, and establish a train vertical suspension system control simulation model in Simulink. Set the vehicle running speed to 250 km / h and the simulation time to 50 seconds, generating a total of 5,000 sets of sample data.

[0113] The corresponding relationship between the above fault types and fault descriptions is shown in Table 1 below:

[0114]

[0115] Table 1

[0116] Using the above simulation process, the simulation effect of the present invention (PRPCA) and the simulation effect of the prior art (PCA) are shown in Tables 2 and 3 and Figure 2-8 As shown:

[0117]

[0118] Table 2

[0119]

[0120] Table 3

[0121] It can be seen from the figures and tables that the method of the present invention can more effectively realize the detection of early faults of the high-speed train suspension system and more effectively solve the classification of early multiple faults of the high-speed train suspension system compared with the traditional PCA-SVM method. At the same time, it effectively solves the detection of early faults under the closed-loop control structure and its engineering practical problems, which is of great significance for the early warning and real-time monitoring of high-speed train suspension faults.

[0122] Although the embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention, and such modifications and variations are all within the scope defined by the appended claims.

Claims

1. A method for early fault diagnosis of a high-speed train suspension system, characterized in that: The steps include: Step 1: Obtain data sets under normal working conditions and fault conditions of the high-speed train suspension system respectively, and standardize the data sets; Step 2: Obtain the load matrix of the data set processed in step 1; Step 3: Convert the load matrix into the overload matrix in the PRPCA residual model, and construct the PRPCA residual model to obtain the residual of the PRPCA residual model of the data set; Step 4: Set the fault detection threshold; Step 5: The fault recognition and classification of the support vector machine are trained through the overload matrix and the model residual of PRPCA, and the trained support vector machine is used to diagnose the early faults of the high-speed train suspension system.

2. The high-speed train suspension system early fault diagnosis method according to claim 1, characterized in that: In step 2, the specific method for obtaining the load matrix is: perform principal component analysis on the data set processed in step 1, perform singular value decomposition on the covariance matrix corresponding to the analysis result, and obtain the load matrix of the data set.

3. The method for early fault diagnosis of a high-speed train suspension system according to claim 1, characterized in that: The specific method of setting the fault detection threshold in step 4 is as follows: The Wasserstein distance index between the fault condition and the normal condition is obtained according to the model residual of the PRPCA residual model, and the fault detection threshold is set according to the Wasserstein distance.

4. The method for early fault diagnosis of a high-speed train suspension system according to claim 3, characterized in that: The calculation formula of the Wasserstein distance is: Where R is the model residual under normal conditions; R if is the model residual under fault conditions; x and y are n-dimensional variables respectively.

5. The method for early fault diagnosis of a high-speed train suspension system according to claim 3 or 4, characterized in that: In the Wasserstein distance calculation process, the expression of the 2-Wasserstein distance between n-dimensional variables x and y is: Where Tr(·) is the trace of the matrix; ∑1=∑ 1 / 2 ∑ if ∑ 1 / 2 , ∑ is the covariance matrix of the n-dimensional variable x, ∑ if is the covariance matrix of the n-dimensional variable y; μ1=||μ-μ if ||2, μ is the mean of the n-dimensional variable x, μ if is the mean of the n-dimensional variable y.

Images