A method and device for detecting the main circuit grounding fault of a traction system

Abstract

The present invention discloses a main circuit grounding fault detection method and device for a traction system, including: obtaining the current detection voltage sequences, including the secondary side voltage sequence of the rectifier, the DC bus voltage sequence, and the half bus voltage sequence; inputting the data online into a classification model and outputting a fault classification result; the classification model is obtained by offline learning and training with a feature matrix; the feature matrix is constructed with multiple feature vectors as inputs and the corresponding fault types as outputs; the feature vectors correspond to historical fault points and are obtained by fusing the feature indicators corresponding to the historical fault points; the feature indicators are obtained by performing short-time Fourier transform on the feature variables and statistically calculating the feature variables; the feature variables corresponding to the historical fault points are obtained by fusing multi-dimensional voltages associated with the fault types corresponding to the historical fault points; according to the fault classification result, determining the fault location and fault type corresponding to the current detection voltage sequences. Thus, the fault location and fault type can be accurately found out.

Description

Technical Field

[0001] The present invention relates to the field of main circuit grounding faults of high-speed train traction converters, and particularly to a detection method and a detection device for main circuit grounding faults of a traction system. Background Art

[0002] The only power source of a high-speed train is the traction drive system, which is a power electronics system composed of a traction transformer, a traction converter, and traction motors. Due to the complexity of the train operation environment, the traction drive system is easily affected by factors such as ambient temperature, ambient humidity, and system power supply surges, resulting in faults during the train operation.

[0003] Main circuit grounding fault is a common abnormal phenomenon during train operation. When a single-point fault occurs, the signal of the grounding detection circuit will be abnormal. At this time, the harm to the train can be ignored and will not affect the normal operation of the system. However, when two or more points are grounded, a large short-circuit current may be generated, causing damage to the components of the electric drive system. In severe cases, it may even lead to locomotive breakdown. Currently, only regular maintenance can be used to avoid faults during train operation, and the fault location and fault type cannot be accurately found. Summary of the Invention

[0004] The present invention provides a detection method and a detection device for main circuit grounding faults of a traction system, which can accurately find the fault location and fault type.

[0005] To solve the above technical problems, the present invention provides a detection method for main circuit grounding faults of a traction system, including the following steps:

[0006] Obtain the current detection voltage sequence, where the current detection voltage sequence includes the rectifier secondary side voltage sequence, the DC bus voltage sequence, and the half bus voltage sequence;

[0007] Input the current detection voltage sequence online into a preset classification model to enable the classification model to output a fault classification result; wherein, the classification model is obtained by offline learning and training of a feature matrix; the feature matrix is constructed with multiple feature vectors as inputs and the corresponding fault types of each feature vector as outputs; each feature vector corresponds to a historical fault point and is obtained by fusing four feature indicators corresponding to the historical fault point; the feature indicators are obtained by performing short-time Fourier transform on the feature variables and statistically calculating the feature variables; the feature variables corresponding to the historical fault point are obtained by fusing multi-dimensional voltages associated with the fault type corresponding to the historical fault point; the feature indicators include: the average value of the sequence values after short-time Fourier transform of the feature variables, the variance of the sequence values after short-time Fourier transform of the feature variables, the average value of the fused feature variables, and the variance of the fused feature variables.

[0008] Based on the fault classification result, determine the fault location and fault type corresponding to the current detected voltage sequence.

[0009] The present invention proposes a method for detecting the grounding fault of the main circuit of a traction system. By obtaining the current detected voltage sequence and inputting it online into a classification model to obtain a fault classification result, and then determining the fault location and fault type corresponding to the detected voltage sequence according to the classification result. The classification model is obtained by training an initial extreme learning machine with multiple feature vectors as inputs and the fault types corresponding to the feature vectors as outputs. Each feature vector corresponds to a historical fault point and is obtained by fusing four feature indicators obtained by performing a short-time Fourier transform on the feature variables corresponding to the historical fault point. The above method reduces the feature variables, enabling the feature variables to reach a minimum of one, and at the same time establishes a classification model of four-dimensional feature indicators. Through the above method, the fault location and fault model corresponding to the current detection can be obtained quickly.

[0010] As a preferred example, the feature matrix:

[0011]

[0012] where FV is a matrix, C n is the fault type, and I n (k) is the feature vector corresponding to the fault type.

[0013] In this preferred example, a classification model is established with the feature vector as the input quantity and the corresponding fault type as the output quantity. According to the feature vector, the fault type can be accurately found.

[0014] As a preferred example, the online input of the current detected voltage sequence into a preset classification model to enable the classification model to output a fault classification result is specifically:

[0015] Calculate the feature vector corresponding to the current detected voltage sequence, and input the feature vector into the classification model to enable the classification model to output a fault analysis result.

[0016] In this preferred example, by calculating the feature vector of the detected voltage sequence and inputting it into the classification model, the fault location is output through the classification model, and further a fault analysis result is obtained.

[0017] As a preferred example, the feature variables corresponding to the historical fault points are obtained by fusing multi-dimensional voltages associated with the fault types corresponding to the historical fault points, specifically:

[0018] Fuse the multi-dimensional voltages associated with the fault types corresponding to each historical fault point through the feature variable formula, and construct the corresponding feature variables; calculate the feature vector corresponding to the current detected voltage sequence, and input the feature vector into the classification model;

[0019] The formula for the characteristic variable is as follows

[0020]

[0021] where U2 is the secondary side voltage of the rectifier, U d1 is the DC bus voltage, and U d2 is half of the bus voltage, and F x1 is the characteristic variable.

[0022] In this preferred example, the characteristic variables are obtained by fusing the voltages detected in the current detection, reducing the characteristic variables so that the characteristic variables can be as low as one at least.

[0023] As a preferred example, the characteristic index is obtained by performing a short-time Fourier transform on the characteristic variable and statistically calculating the characteristic variable. Specifically:[[]]

[0024] Perform a short-time Fourier transform on each of the characteristic variables, and extract the average value of the DC component of the sequence after the short-time Fourier transform of the characteristic variable and the variance of the corresponding sequence value after the short-time Fourier transform of the characteristic variable; perform statistical calculation of the characteristic variables on each of the characteristic variables, and extract the average value of the fused characteristic variable and the variance of the fused characteristic variable;

[0025] Fuse the multiple corresponding characteristic indexes to construct a corresponding characteristic vector.

[0026] In this preferred example, by performing a short-time Fourier transform on the characteristic variable, the one-dimensional characteristic variable is converted into a four-dimensional characteristic index, and the four-dimensional characteristic indexes are fused into a characteristic vector, making the fault detection and analysis more accurate.

[0027] As a preferred example,

[0028] Perform a short-time Fourier transform on each of the characteristic variables, and extract the average value of the DC component of the sequence after the short-time Fourier transform of the characteristic variable and the variance of the corresponding sequence value after the short-time Fourier transform of the characteristic variable; perform statistical calculation of the characteristic variables on each of the characteristic variables, and extract the average value of the fused characteristic variable and the variance of the fused characteristic variable. Specifically:[[]]

[0029] Perform a short-time Fourier transform on each of the characteristic variables;

[0030] The formula for the short-time Fourier transform is

[0031]

[0032] [S, f, t] = STFT(F x1 , win, hop, nfft, f s )

[0033] Among them, x(τ) is the signal, h(τ - t) is the signal segmentation analysis window function, S is a two-dimensional matrix [m, n], win is the window function, hop is the translation step size, nff is the number of points in the sliding window, and f s is the sampling frequency;

[0034] Extract the average value of the sequence values after the short-time Fourier transform of the corresponding feature variables, the variance of the sequence values after the short-time Fourier transform of the feature variables, the average value of the fused feature variables, and the variance of the fused feature variables;

[0035] The average value of the DC component of the sequence after the short-time Fourier transform of the feature variable is

[0036]

[0037] The variance of the corresponding sequence values after the short-time Fourier transform of the feature variable is

[0038] I2(k) = pk_pk{S(t, f), S(t + 1, f),..., S(t + N - 1, f)}

[0039] The average value of the fused feature variable is

[0040]

[0041] The variance of the fused feature variable is

[0042]

[0043] Among them, N is the number of sampling points, k is the point of the feature index, and pk_pk is the peak-to-peak value of the sequence.

[0044] In this preferred example, by performing a short-time Fourier transform on the feature variables, the one-dimensional feature variables are converted into four-dimensional feature indices, making the fault analysis results more accurate.

[0045] As a preferred example, the fusion of the multiple corresponding feature indices to construct the corresponding feature vector is specifically as follows:

[0046] Construct the corresponding feature vector from the multiple corresponding feature indices through a vector matrix;

[0047] The corresponding feature vector matrix is

[0048] FV(k) = [I1(k) I2(k) I3(k) I4(k)]

[0049] Among them, FV(k) is the vector matrix.

[0050] In this preferred example, by fusing multiple corresponding characteristic indicators into one characteristic vector, the original one-dimensional characteristic variables are converted into four-dimensional characteristic indicators, and a corresponding four-dimensional characteristic vector is constructed, making the fault analysis result more accurate.

[0051] As a preferred example, inputting the characteristic vector into the classification model so that the classification model outputs a fault analysis result specifically includes:

[0052] Comparing the variance of the sequence values after the short-time Fourier transform of the characteristic variable with a preset value. When the variance of the sequence values after the short-time Fourier transform of the characteristic variable is greater than the preset value, the classification model outputs a fault classification result.

[0053] In this preferred example, by comparing the variance of the sequence values after the short-time Fourier transform of the characteristic variable with the preset value, it is judged whether the classification model will output a fault classification result. When the variance of the sequence values after the short-time Fourier transform of the characteristic variable is greater than the preset value, it proves that a fault has occurred at the detection point, and a fault classification result is output; when the variance of the sequence values after the short-time Fourier transform of the characteristic variable is less than the preset value, it proves that no fault has occurred at the detection point, and there is no need to output a fault classification result.

[0054] The present invention also provides a traction system main circuit grounding fault detection device, including an acquisition module, a classification enabling module, and a fault tracing module;

[0055] The acquisition module is used to acquire the current detection voltage sequence, and the current detection voltage sequence includes the rectifier secondary side voltage sequence, the DC bus voltage sequence, and the half-bus voltage sequence;

[0056] The classification enabling module is used to online input the current detection voltage sequence into a preset classification model so that the classification model outputs a fault classification result; wherein, the classification model is obtained by offline learning and training with a characteristic matrix; the characteristic matrix is constructed with multiple characteristic vectors as inputs and the corresponding fault types of each characteristic vector as outputs; each characteristic vector corresponds to a historical fault point and is obtained by fusing four characteristic indicators corresponding to the historical fault point; the characteristic indicators are obtained by performing short-time Fourier transform on the characteristic variable and statistically calculating the characteristic variable; the characteristic variable corresponding to the historical fault point is obtained by fusing multi-dimensional voltages associated with the fault type corresponding to the historical fault point; the characteristic indicators include: the average value of the sequence values after the short-time Fourier transform of the characteristic variable, the variance of the sequence values after the short-time Fourier transform of the characteristic variable, the average value of the fused characteristic variable, and the variance of the fused characteristic variable;

[0057] The fault tracing module is used to determine the fault location and fault type corresponding to the current detection voltage sequence according to the fault classification result.

[0058] The present invention proposes a main circuit grounding fault detection device for a traction system. An acquisition module acquires the historical detected voltage sequence for the current detection and inputs it into a classification model of a classification enabling module to judge the fault result, and a fault tracing module determines the fault location and fault type corresponding to the detected voltage sequence. The classification enabling module constructs historical fault data to obtain corresponding feature vectors, and constructs a classification model, so that a judgment module inputs the current monitoring data into the classification model to judge whether the model triggers a diagnosis. When a diagnosis is triggered, a diagnosis module obtains the diagnosis result and determines the current fault location and fault type.

[0059] As a preferred example, the feature matrix is:

[0060]

[0061] where FV is a matrix, C n is the fault type, and I n (k) is the feature vector corresponding to the fault type.

[0062] In this preferred example, a classification model is established with the feature vector as the input quantity and the corresponding fault type as the output quantity, and the fault type can be accurately found according to the feature vector.

[0063] As a preferred example, the classification enabling module includes a first unit;

[0064] The first unit is used to calculate the feature vector corresponding to the current detected voltage sequence, and input the feature vector into the classification model, so that the classification model outputs a fault analysis result.

[0065] In this preferred example, the feature vector of the detected voltage sequence is calculated and input into the classification model, and the fault location is output through the classification model to further obtain a fault analysis result.

[0066] As a preferred example, the classification enabling module further includes a second unit;

[0067] The second unit is used to fuse the multi-dimensional voltages associated with the fault types corresponding to each historical fault point through a feature variable formula and construct corresponding feature variables; calculate the feature vector corresponding to the current detected voltage sequence, and input the feature vector into the classification model;

[0068] The feature variable formula is as follows

[0069]

[0070] where U2 is the voltage on the secondary side of the rectifier, U d1 is the DC bus voltage, U d2 is half of the bus voltage, and F x1 is the feature variable.

[0071] In this preferred example, by fusing the voltages detected in the current detection to obtain characteristic variables, the number of characteristic variables is reduced, and the number of characteristic variables can be minimized to one.

[0072] As a preferred example, the classification enabling module further includes a third unit and a fourth unit;

[0073] The third unit is used to perform short-time Fourier transform on each of the characteristic variables, and extract the average value of the DC component of the sequence after the short-time Fourier transform of the characteristic variable and the variance of the corresponding sequence value after the short-time Fourier transform of the characteristic variable; perform statistical calculation on each of the characteristic variables to calculate the characteristic variables, and extract the average value of the fused characteristic variables and the variance of the fused characteristic variables;

[0074] The fourth unit is used to fuse the multiple corresponding characteristic indicators to construct a corresponding characteristic vector.

[0075] In this preferred example, by performing short-time Fourier transform on the characteristic variables, the one-dimensional characteristic variables are converted into four-dimensional characteristic indicators, and the four-dimensional characteristic indicators are fused into a characteristic vector, making the fault detection and analysis more accurate.

[0076] As a preferred example, when performing short-time Fourier transform on each of the characteristic variables to extract four corresponding characteristic indicators, specifically:

[0077] Perform short-time Fourier transform on each of the characteristic variables;

[0078] The short-time Fourier transform formula is

[0079]

[0080] [S,f,t]=STFT(F x1 ,win,hop,nfft,f s )

[0081] where x(τ) is the signal, h(τ - t) is the signal segmentation analysis window function, S is a two-dimensional matrix [m, n], win is the window function, hop is the translation step, nff is the number of points of the sliding window, and f s is the sampling frequency;

[0082] Extract the average value of the sequence values after the short-time Fourier transform of the corresponding characteristic variables, the variance of the sequence values after the short-time Fourier transform of the characteristic variables, the average value of the fused characteristic variables, and the variance of the fused characteristic variables;

[0083] The average value of the DC component of the sequence after the short-time Fourier transform of the characteristic variable is

[0084]

[0085] The variance of the corresponding sequence values after short-time Fourier transform of the characteristic variable is

[0086] I2(k) = pk_pk{S(t,f), S(t + 1,f),..., S(t + N - 1,f)}

[0087] The average value of the fused characteristic variable is

[0088]

[0089] The variance of the fused characteristic variable is

[0090]

[0091] Where N is the number of sampling points, k is the point of the characteristic index, and pk_pk is the peak-to-peak value of the sequence.

[0092] In this preferred example, by performing short-time Fourier transform on the characteristic variable, the one-dimensional characteristic variable is converted into four-dimensional characteristic indexes, making the fault analysis result more accurate.

[0093] As a preferred example, the step of fusing the multiple corresponding characteristic indexes to construct a corresponding characteristic vector is specifically as follows:

[0094] Construct a corresponding characteristic vector from the multiple corresponding characteristic indexes through a vector matrix;

[0095] The corresponding characteristic vector matrix is

[0096] FV(k) = [I1(k) I2(k) I3(k) I4(k)]

[0097] Where FV(k) is the vector matrix.

[0098] In this preferred example, by fusing multiple corresponding characteristic indexes into one characteristic vector, the original one-dimensional characteristic variable is converted into four-dimensional characteristic indexes, and a corresponding four-dimensional characteristic vector is constructed, making the fault analysis result more accurate.

[0099] As a preferred example, the classification enabling module further includes a fifth unit;

[0100] The fifth unit is used to compare the variance of the sequence values after short-time Fourier transform of the characteristic variable with a preset value. When the variance of the sequence values after short-time Fourier transform of the characteristic variable is greater than the preset value, the classification model outputs a fault analysis result.

[0101] In this preferred example, by comparing the variance of the sequence values after the short-time Fourier transform of the characteristic variables with a preset value, it is determined whether the classification model will output a fault classification result. When the variance of the sequence values after the short-time Fourier transform of the characteristic variables is greater than the preset value, it proves that a fault has occurred at the detection point, and the fault classification result is output; when the variance of the sequence values after the short-time Fourier transform of the characteristic variables is less than the preset value, it proves that no fault has occurred at the detection point, and there is no need to output the fault classification result. BRIEF DESCRIPTION OF THE DRAWINGS

[0102] Figure 1 is a flowchart of a method for detecting a ground fault in the main circuit of a traction system according to an embodiment provided by the present invention;

[0103] Figure 2 is a learning diagram of a method for detecting a ground fault in the main circuit of a traction system according to an embodiment provided by the present invention;

[0104] Figure 3 is a structural diagram of a device for detecting a ground fault in the main circuit of a traction system according to an embodiment provided by the present invention;

[0105] Figure 4 is a flowchart of an algorithm for a method for detecting a ground fault in the main circuit of a traction system according to an embodiment provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0106] Next, the technical solutions in the present invention will be clearly and completely described in conjunction with the accompanying drawings in the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0107] A method and device for detecting a ground fault in the main circuit of a traction system provided by an embodiment of the present invention are applicable to.

[0108] Please refer to Figure 1 , in an embodiment of the present invention, there is provided Figure 1 a schematic flowchart of a method for detecting a ground fault in the main circuit of a traction system as shown, and the method includes steps S1 to S3. The specific steps are as follows:

[0109] S1. Obtain the current detection voltage sequence, where the current detection voltage sequence includes the rectifier secondary side voltage sequence, the DC bus voltage sequence, and the half-bus voltage sequence;

[0110] S2. Input the current detected voltage sequence online into a preset classification model so that the classification model outputs a fault classification result. The classification model is obtained through offline learning and training of a feature matrix, which is constructed with multiple feature vectors as inputs and the corresponding fault types of each feature vector as outputs. Each feature vector corresponds to a historical fault point and is obtained by fusing four feature indicators corresponding to the historical fault point. The feature indicators are obtained by performing short-time Fourier transform on the feature variables and statistically calculating the feature variables. The feature variables corresponding to the historical fault points are obtained by fusing multi-dimensional voltages associated with the fault types corresponding to the historical fault points. The feature indicators include: the average value of the sequence values after short-time Fourier transform of the feature variables, the variance of the sequence values after short-time Fourier transform of the feature variables, the average value of the fused feature variables, and the variance of the fused feature variables.

[0111] S3. Determine the fault location and fault type corresponding to the current detected voltage sequence according to the fault classification result.

[0112] In a traction system main circuit grounding fault detection method provided by the present invention, by obtaining the current detected voltage sequence and inputting it into a classification model to obtain a fault classification result, and then determining the fault location and fault type corresponding to the detected voltage sequence according to the classification result. The classification model is obtained by training an initial extreme learning machine with multiple feature vectors as inputs and the corresponding fault types of the feature vectors as outputs. Each feature vector corresponds to a historical fault point and is obtained by fusing four feature indicators obtained by performing short-time Fourier transform on the feature variables corresponding to the historical fault point. The above method reduces the feature variables, enabling the feature variables to reach a minimum of one, and at the same time establishes a classification model with four-dimensional feature indicators. Through the above method, the fault location and fault model corresponding to the current detection can be quickly obtained.

[0113] According to the fault mechanism, the corresponding relationship between the grounding point and the fault type is shown in Table 1, and from the system signals collected by the traction drive system, select the traction drive system signals related to the main circuit grounding fault: the rectifier secondary side voltage sequence U2, the DC bus voltage sequence U measured by VH2 d1 and the one-half bus voltage sequence U measured by the VH1 voltage sensor d2 .

[0114]

[0115]

[0116] Please refer to Table 1 for the common main circuit grounding fault points of the traction system Figure 2 and Figure 4 , in an embodiment of the present invention, the feature matrix is:

[0117]

[0118] Among them, FV is a matrix, and C n is the fault type, and I n (k) is the eigenvector corresponding to the fault type.

[0119] In the embodiment of the present invention, a classification model is established with the eigenvector as the input quantity and the corresponding fault type as the output quantity. According to the eigenvector, the fault type can be accurately found.

[0120] Please refer to Figure 2 and Figure 4 , in this embodiment, based on a machine learning algorithm such as a neural network, a grounding fault location and identification model is constructed for the established classification model. The model parameters of the grounding fault location and identification model include: machine learning input layer parameters, hidden layer parameters, and the input weights and node thresholds of the hidden layer are optimized.

[0121] Please refer to Figure 2 and Figure 4 , in another embodiment of the present invention, a machine learning algorithm is used to construct a grounding fault location and identification model. Based on the extreme learning machine (ELM), a grounding fault location and identification model is constructed. The grounding fault location and identification model includes an input layer, a hidden layer, and an output layer. The characteristic index is used as the input, the input layer parameters of the machine learning model are set as a 4-dimensional vector, the number of hidden layer nodes is 10, the input weights and node thresholds of the hidden layer, and the output layer is 6-dimensional, and the activation function is sig; the characteristic data set extracted from the historical samples is used as the input X, and C1 and C4 are the outputs. By randomly generating ELM parameters multiple times, a set of parameters with good classification effects is selected.

[0122] In a certain embodiment of the present invention, the online input of the current detection voltage sequence into a preset classification model to enable the classification model to output a fault classification result is specifically:

[0123] Calculate the eigenvector corresponding to the current detection voltage sequence, and input the eigenvector into the classification model to enable the classification model to output a fault analysis result.

[0124] In the embodiment of the present invention, a detection voltage sequence is obtained, a characteristic matrix is calculated, a trained classification model is called, and the fault location is output through the classification model, and further a fault analysis result is obtained.

[0125] In a certain embodiment of the present invention, the characteristic variable corresponding to the historical fault point is obtained by fusing multi-dimensional voltages associated with the fault type corresponding to the historical fault point, specifically:

[0126] Fuse the multi-dimensional voltages associated with the corresponding fault types of each of the historical fault points through a characteristic variable formula, and construct corresponding characteristic variables; calculate the characteristic vector corresponding to the current detected voltage sequence, and input the characteristic vector into a classification model;

[0127] The characteristic variable formula is as follows

[0128]

[0129] where U2 is the secondary side voltage of the rectifier, U d1 is the DC bus voltage, U d2 is half of the bus voltage, F x1 is the characteristic variable.

[0130] In the embodiment of the present invention, fusing the voltages detected each time to obtain characteristic variables reduces the number of characteristic variables, and the characteristic variables can be reduced to as low as one.

[0131] In a certain embodiment of the present invention, the characteristic indicators are obtained by performing short-time Fourier transform on the characteristic variables and statistically calculating the characteristic variables. Specifically:

[0132] Perform short-time Fourier transform on each of the characteristic variables, and extract the average value of the DC component of the sequence after the short-time Fourier transform of the characteristic variable and the variance of the corresponding sequence values after the short-time Fourier transform of the characteristic variable; perform statistical calculation of the characteristic variables on each of the characteristic variables, and extract the average value of the fused characteristic variables and the variance of the fused characteristic variables;

[0133] Fuse the multiple corresponding characteristic indicators to construct a corresponding characteristic vector.

[0134] In the embodiment of the present invention, performing short-time Fourier transform on the characteristic variables converts the one-dimensional characteristic variables into four-dimensional characteristic indicators, and fuses the four-dimensional characteristic indicators into a characteristic vector, making the fault detection and analysis more accurate. Only use characteristic variables to reduce the number of characteristic variables; combine statistical and short-time Fourier transform techniques to extract time-frequency domain characteristic indicators, so that the number of characteristic variables can be reduced by 1, and at the same time only 4 characteristic indicators are needed, reducing the dimension.

[0135] In a certain embodiment of the present invention, perform short-time Fourier transform on each of the characteristic variables, and extract the average value of the DC component of the sequence after the short-time Fourier transform of the characteristic variable and the variance of the corresponding sequence values after the short-time Fourier transform of the characteristic variable; perform statistical calculation of the characteristic variables on each of the characteristic variables, and extract the average value of the fused characteristic variables and the variance of the fused characteristic variables. Specifically:

[0136] Perform short-time Fourier transform on each of the characteristic variables;

[0137] The short-time Fourier transform formula is

[0138]

[0139] [S,f,t] = STFT(F x1 , win, hop, nfft, f s )

[0140] where x(τ) is the signal, h(τ - t) is the signal segment analysis window function, S is a two-dimensional matrix [m, n], win is the window function, hop is the translation step, nff is the number of points in the sliding window, and f s is the sampling frequency;

[0141] Extract the average value of the sequence values after the short-time Fourier transform of the corresponding feature variables, the variance of the sequence values after the short-time Fourier transform of the feature variables, the average value of the fused feature variables, and the variance of the fused feature variables;

[0142] The average value of the DC component of the sequence after the short-time Fourier transform of the feature variable is

[0143]

[0144] The variance of the corresponding sequence values after the short-time Fourier transform of the feature variable is

[0145] I2(k) = pk_pk{S(t,f), S(t + 1,f),..., S(t + N - 1,f)}

[0146] The average value of the fused feature variable is

[0147]

[0148] The variance of the fused feature variable is

[0149]

[0150] where N is the number of sampling points, k is the point of the feature index, and pk_pk is the peak-to-peak value of the sequence.

[0151] In the embodiment of the present invention, the short-time Fourier transform is performed on the feature variables, converting the one-dimensional feature variables into four-dimensional feature indices and fusing the four-dimensional feature indices into a feature vector, making the fault detection and analysis more accurate. Only use the feature variables, reducing the number of feature variables; combining statistical and short-time Fourier transform technologies to extract time-frequency domain feature indices, so that the number of feature variables can be reduced by 1, and at the same time only 4 feature indices are required, reducing the dimension.

[0152] In a certain embodiment of the present invention, the fusion of the multiple corresponding feature indices to construct the corresponding feature vector is specifically:

[0153] Construct a corresponding feature vector from the multiple corresponding feature indicators through a vector matrix;

[0154] The corresponding feature vector matrix is

[0155] FV(k) = [I1(k) I2(k) I3(k) I4(k)]

[0156] where FV(k) is the vector matrix.

[0157] In an embodiment of the present invention, multiple corresponding feature indicators are fused into one feature vector, converting the original one-dimensional feature variable into a four-dimensional feature indicator, and constructing a corresponding four-dimensional feature vector, making the fault analysis result more accurate. The constructed one-dimensional feature variable is extracted into a 4D feature subscript sequence and its matrix is constructed, that is, a four-dimensional feature vector, so as to diagnose faults by exploring the characteristics of the four-dimensional feature vector.

[0158] In a certain embodiment of the present invention, inputting the feature vector into a classification model so that the classification model outputs a fault analysis result specifically includes:

[0159] Compare the variance of the sequence values after the short-time Fourier transform of the feature variable with a preset value. When the variance of the sequence values after the short-time Fourier transform of the feature variable is greater than the preset value, the classification model outputs a fault classification result.

[0160] In an embodiment of the present invention, compare the variance of the sequence values after the short-time Fourier transform of the feature variable with a preset value to determine whether the classification model will output a fault classification result. When the variance of the sequence values after the short-time Fourier transform of the feature variable is greater than the preset value, it proves that a fault has occurred at the detection point and a fault classification result is output; when the variance of the sequence values after the short-time Fourier transform of the feature variable is less than the preset value, it proves that no fault has occurred at the detection point and there is no need to output a fault classification result.

[0161] In another embodiment of the present invention, the following method is used to detect whether the classification model outputs a fault classification result:

[0162]

[0163] I th is the preset value. When flag = 1, the classification model outputs a fault classification result.

[0164] Please refer to Figure 3 , in an embodiment of the present invention, a traction system main circuit grounding fault detection device is further provided, including an acquisition module, a classification enabling module, and a fault tracing module;

[0165] The obtaining module is used to obtain the current detected voltage sequence, and the current detected voltage sequence includes the secondary side voltage sequence of the rectifier, the DC bus voltage sequence, and the half bus voltage sequence;

[0166] The classification enabling module is used to input the current detected voltage sequence online into a preset classification model, so that the classification model outputs a fault classification result; wherein, the classification model is obtained by offline learning and training of a feature matrix; the feature matrix is constructed with multiple feature vectors as inputs and the corresponding fault types of each feature vector as outputs; each feature vector corresponds to a historical fault point and is obtained by fusing four feature indicators corresponding to the historical fault point; the feature indicators are obtained by performing short-time Fourier transform and statistical calculation of the feature variables on the feature variables; the feature variables corresponding to the historical fault point are obtained by fusing multi-dimensional voltages associated with the fault type corresponding to the historical fault point; the feature indicators include: the average value of the sequence values after short-time Fourier transform of the feature variables, the variance of the sequence values after short-time Fourier transform of the feature variables, the average value of the fused feature variables, and the variance of the fused feature variables;

[0167] The fault tracing module is used to determine the fault location and fault type corresponding to the current detected voltage sequence according to the fault classification result.

[0168] In a traction system main circuit grounding fault detection device provided by the present invention, the obtaining module obtains the historical current detected voltage sequence and inputs it into the classification model of the classification enabling module to judge the fault result, and the fault tracing module determines the fault location and fault type corresponding to the detected voltage sequence. The classification enabling module constructs historical fault data to obtain corresponding feature vectors, constructs a classification model, so that the judgment module inputs the current monitoring data into the classification model to judge whether the model triggers a diagnosis. When the diagnosis is triggered, the diagnosis module obtains the diagnosis result and determines the current fault location and fault type.

[0169] Please refer to Figures 2 to 4 , in an embodiment of the present invention, the feature matrix is:

[0170]

[0171] wherein, FV is a matrix, C n is the fault type, and I n (k) is the feature vector corresponding to the fault type.

[0172] In an embodiment of the present invention, a classification model is established with feature vectors as input quantities and corresponding fault types as output quantities, and the fault type can be accurately found according to the feature vectors.

[0173] Please refer to Figures 2 to 4, in this embodiment, the established classification model is based on a machine learning algorithm, such as a neural network, to construct a grounding fault location recognition model. The model parameters of the grounding fault location recognition model include: machine learning input layer parameters, hidden layer parameters, optimized hidden layer input weights, and node thresholds.

[0174] Please refer to Figures 2 to 4 , in another embodiment of the present invention, a grounding fault location recognition model is constructed using a machine learning algorithm. The grounding fault location recognition model is constructed based on the Extreme Learning Machine (ELM). The grounding fault location recognition model includes an input layer, a hidden layer, and an output layer. The characteristic indicators are used as inputs. The input layer parameters of the machine learning model are set as a 4-dimensional vector, the number of hidden layer nodes is 10, the hidden layer input weights and node thresholds, the output layer is 6-dimensional, and the activation function is sig. The feature data set extracted from the historical samples is used as the input X, and C1 and C4 are the outputs. By randomly generating ELM parameters multiple times, a set of parameters with good classification effects is selected.

[0175] Please refer to Figure 3 , in a certain embodiment of the present invention, the classification enabling module includes a first unit;

[0176] The first unit is used to calculate the feature vector corresponding to the current detection voltage sequence and input the feature vector into the classification model so that the classification model outputs a fault analysis result.

[0177] In the embodiment of the present invention, a detection voltage sequence is obtained, a feature matrix is calculated, a trained classification model is called, and the fault location is output through the classification model, and further a fault analysis result is obtained.

[0178] Please refer to Figure 3 , in a certain embodiment of the present invention, the classification enabling module further includes a second unit;

[0179] The second unit is used to fuse the multi-dimensional voltages associated with the corresponding fault types of each historical fault point through a feature variable formula and construct a corresponding feature variable; calculate the feature vector corresponding to the current detection voltage sequence and input the feature vector into the classification model;

[0180] The feature variable formula is as follows

[0181]

[0182] Among them, U2 is the secondary side voltage of the rectifier, U d1 is the DC bus voltage, U d2 is half of the bus voltage, F x1 is the feature variable.

[0183] In an embodiment of the present invention, the voltages detected in the current detection are fused to obtain characteristic variables, reducing the number of characteristic variables so that the number of characteristic variables can be as low as one.

[0184] Please refer to Figure 3 , in an embodiment of the present invention, the classification enabling module further includes a third unit and a fourth unit;

[0185] The third unit is used to perform short-time Fourier transform on each of the characteristic variables, extract the average value of the DC component of the sequence after the short-time Fourier transform of the characteristic variable and the variance of the corresponding sequence values after the short-time Fourier transform of the characteristic variable; perform statistical calculation on each of the characteristic variables to calculate the characteristic variables, and extract the average value of the fused characteristic variables and the variance of the fused characteristic variables;

[0186] The fourth unit is used to fuse the multiple corresponding characteristic indicators to construct a corresponding characteristic vector.

[0187] In an embodiment of the present invention, short-time Fourier transform is performed on the characteristic variables, converting one-dimensional characteristic variables into four-dimensional characteristic indicators, and fusing the four-dimensional characteristic indicators into a characteristic vector, making the fault detection and analysis more accurate. Only characteristic variables are used, reducing the number of characteristic variables; combining statistical and short-time Fourier transform technologies to extract time-frequency domain characteristic indicators, enabling the number of characteristic variables to be reduced by one, and at the same time only four characteristic indicators are required, reducing the dimension.

[0188] In an embodiment of the present invention, performing short-time Fourier transform on each of the characteristic variables and extracting four corresponding characteristic indicators specifically includes:

[0189] Performing short-time Fourier transform on each of the characteristic variables;

[0190] The short-time Fourier transform formula is

[0191]

[0192] [S,f,t]=STFT(F x1 ,win,hop,nfft,f s )

[0193] where x(τ) is the signal, h(τ - t) is the signal segmentation analysis window function, S is a two-dimensional matrix [m, n], win is the window function, hop is the translation step, nff is the number of points of the sliding window, and f s is the sampling frequency;

[0194] Extracting the average value of the sequence values after the short-time Fourier transform of the corresponding characteristic variables, the variance of the sequence values after the short-time Fourier transform of the characteristic variables, the average value of the fused characteristic variables, and the variance of the fused characteristic variables;

[0195] The average value of the DC component of the sequence after the short-time Fourier transform of the feature variable is

[0196]

[0197] The variance of the corresponding sequence values after the short-time Fourier transform of the feature variable is

[0198] I2(k) = pk_pk{S(t,f), S(t + 1,f),..., S(t + N - 1,f)}

[0199] The average value of the fused feature variable is

[0200]

[0201] The variance of the fused feature variable is

[0202]

[0203] Where N is the number of sampling points, k is the point of the feature index, and pk_pk is the peak-to-peak value of the sequence.

[0204] In the embodiment of the present invention, the short-time Fourier transform is performed on the feature variable, converting the one-dimensional feature variable into a four-dimensional feature index, and fusing the four-dimensional feature indexes into a feature vector, making the fault detection and analysis more accurate. Only using the feature variable reduces the number of feature variables; combining statistical and short-time Fourier transform technologies to extract time-frequency domain feature indexes, so that the number of feature variables can be reduced by 1, and only 4 feature indexes are required, reducing the dimension.

[0205] In a certain embodiment of the present invention, the fusion of the multiple corresponding feature indexes to construct the corresponding feature vector is specifically as follows:

[0206] Construct the corresponding feature vector by the vector matrix for the multiple corresponding feature indexes;

[0207] The corresponding feature vector matrix is

[0208] FV(k) = [I1(k) I2(k) I3(k) I4(k)]

[0209] Where FV(k) is the vector matrix.

[0210] In the embodiment of the present invention, multiple corresponding feature indexes are fused into one feature vector, converting the original one-dimensional feature variable into a four-dimensional feature index, and constructing the corresponding four-dimensional feature vector, making the fault analysis result more accurate. The constructed one-dimensional feature variable is extracted into a 4D feature sub-index sequence and its matrix is constructed, that is, the four-dimensional feature vector, so as to diagnose faults by exploring the characteristics of the four-dimensional feature vector.

[0211] Please refer to Figure 3 , in an embodiment of the present invention, the classification enabling module further includes a fifth unit;

[0212] The fifth unit is used to compare the variance of the sequence values after the short-time Fourier transform of the feature variable with a preset value. When the variance of the sequence values after the short-time Fourier transform of the feature variable is greater than the preset value, the classification model outputs a fault analysis result.

[0213] In an embodiment of the present invention, the variance of the sequence values after the short-time Fourier transform of the feature variable is compared with a preset value to determine whether the classification model will output a fault classification result. When the variance of the sequence values after the short-time Fourier transform of the feature variable is greater than the preset value, it proves that a fault has occurred at the detection point, and a fault classification result is output; when the variance of the sequence values after the short-time Fourier transform of the feature variable is less than the preset value, it proves that no fault has occurred at the detection point, and there is no need to output a fault classification result.

[0214] In another embodiment of the present invention, the following method is used to detect whether the classification model outputs a fault classification result:

[0215]

[0216] I th is the preset value. When flag = 1, the classification model outputs a fault classification result.

[0217] The above is the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements are also regarded as the protection scope of the present invention.

Claims

1. A method for detecting the ground fault of the main circuit of a traction system, characterized in that, Including the following steps: Obtain the current detected voltage sequence, where the current detected voltage sequence includes the rectifier secondary side voltage sequence, the DC bus voltage sequence, and the half bus voltage sequence; Online input the current detected voltage sequence into a preset classification model so that the classification model outputs a fault classification result; wherein, the classification model is obtained through offline learning and training of a feature matrix; the feature matrix is constructed with multiple feature vectors as inputs and the corresponding fault types of each feature vector as outputs; each feature vector corresponds to a historical fault point and is obtained by fusing four feature indicators corresponding to the historical fault point; the feature indicators are obtained by performing short-time Fourier transform on the feature variables and statistically calculating the feature variables; the feature variables corresponding to the historical fault point are obtained by fusing the multi-dimensional voltage sequences associated with the historical fault point corresponding fault type; the feature indicators include: the average value of the sequence values after short-time Fourier transform of the feature variable, the variance of the sequence values after short-time Fourier transform of the feature variable, the average value of the fused feature variable, and the variance of the fused feature variable; According to the fault classification result, determine the fault location and fault type corresponding to the current detected voltage sequence; Wherein, the feature variables corresponding to the historical fault point are obtained by fusing the multi-dimensional voltages associated with the historical fault point corresponding fault type, specifically: Fuse the multi-dimensional voltages associated with the corresponding fault type of each historical fault point through the feature variable formula and construct the corresponding feature variables; calculate the feature vector corresponding to the current detected voltage sequence and input the feature vector into the classification model; The feature variable formula is as follows Among them, U2 is the secondary side voltage of the rectifier, and U d1 is the DC bus voltage, and U d2 is half of the bus voltage, and F x1 is the characteristic variable; The feature indicators are obtained by performing short-time Fourier transform on the feature variables and statistically calculating the feature variables, specifically: Perform short-time Fourier transform on each feature variable, extract the average value of the DC component of the sequence after short-time Fourier transform of the feature variable and the variance of the corresponding sequence values after short-time Fourier transform of the feature variable; perform statistical calculation of the feature variables on each feature variable and extract the average value of the fused feature variable and the variance of the fused feature variable; Fuse the multiple corresponding feature indicators to construct the corresponding feature vector.

2. The method for detecting the ground fault of the main circuit of a traction system according to claim 1, characterized in that, The feature matrix is: Among them, FV is a matrix, C n is the fault type, I n (k) is the eigenvector corresponding to the fault type.

3. The method for detecting the ground fault of the main circuit of a traction system according to claim 1, characterized in that, The online input of the current detected voltage sequence into the preset classification model so that the classification model outputs a fault classification result is specifically: Calculate the feature vector corresponding to the current detected voltage sequence and online input the feature vector into the classification model so that the classification model outputs a fault analysis result.

4. The method for detecting the ground fault of the main circuit of a traction system according to claim 1, characterized in that, Perform short-time Fourier transform on each feature variable, extract the average value of the DC component of the sequence after short-time Fourier transform of the feature variable and the variance of the corresponding sequence values after short-time Fourier transform of the feature variable; perform statistical calculation of the feature variables on each feature variable and extract the average value of the fused feature variable and the variance of the fused feature variable, specifically: Perform short-time Fourier transform on each feature variable; The short-time Fourier transform formula is [S,f,t] = STFT(F x1 , win, hop, nfft, f s ) Among them, \(x(\tau)\) is the signal, \(h(\tau - t)\) is the signal segmentation analysis window function, \(S\) is a two-dimensional matrix \([m, n]\), win is the window function, hop is the translation step size, nff is the number of points of the sliding window, and \(f\) s is the sampling frequency; Extract the corresponding average value of the sequence values after short-time Fourier transform of the feature variable, the variance of the sequence values after short-time Fourier transform of the feature variable, the average value of the fused feature variable, and the variance of the fused feature variable; The average value of the DC component of the sequence after the short-time Fourier transform of the characteristic variable is The variance of the corresponding sequence value after the short-time Fourier transform of the characteristic variable is I2(k) = pk_pk{S(t,f), S(t + 1,f),..., S(t + N - 1,f)} The average value of the fused characteristic variable is The variance of the fused characteristic variable is Where N is the number of sampling points, k is the point of the characteristic index, and pk_pk is the peak-to-peak value of the sequence.

5. A method for detecting a ground fault in the main circuit of a traction system according to claim 1, characterized in that, The fusion of the multiple corresponding characteristic indices to construct a corresponding characteristic vector is specifically as follows: Construct a corresponding characteristic vector from the multiple corresponding characteristic indices through a vector matrix; The corresponding characteristic vector matrix is FV(k) = [I1(k) I2(k) I3(k) I4(k)] Where FV(k) is the vector matrix.

6. A method for detecting a ground fault in the main circuit of a traction system according to claim 3, characterized in that, The input of the characteristic vector into the classification model to enable the classification model to output a fault analysis result is specifically as follows: Compare the variance of the sequence value after the short-time Fourier transform of the characteristic variable with a preset value. When the variance of the sequence value after the short-time Fourier transform of the characteristic variable is greater than the preset value, the classification model outputs a fault classification result.

7. A device for detecting a ground fault in the main circuit of a traction system, characterized in that, It includes an acquisition module, a classification enabling module, and a fault tracing module; The acquisition module is used to acquire the current detection voltage sequence, and the current detection voltage sequence includes the rectifier secondary side voltage sequence, the DC bus voltage sequence, and the half-bus voltage sequence; The classification enabling module is used to input the current detection voltage sequence online into a preset classification model to enable the classification model to output a fault classification result; wherein, the classification model is obtained through offline learning and training of a characteristic matrix; the characteristic matrix is constructed with multiple characteristic vectors as inputs and the corresponding fault types of each characteristic vector as outputs; each characteristic vector corresponds to a historical fault point and is obtained by fusing four characteristic indices corresponding to the historical fault point; the characteristic indices are obtained by performing short-time Fourier transform on the characteristic variable and statistically calculating the characteristic variable; the characteristic variable corresponding to the historical fault point is obtained by fusing multi-dimensional voltages associated with the fault type corresponding to the historical fault point; the characteristic indices include: the average value of the sequence value after the short-time Fourier transform of the characteristic variable, the variance of the sequence value after the short-time Fourier transform of the characteristic variable, the average value of the fused characteristic variable, and the variance of the fused characteristic variable; The fault tracing module is used to determine the fault location and fault type corresponding to the current detection voltage sequence according to the fault classification result; Wherein, the characteristic variable corresponding to the historical fault point is obtained by fusing multi-dimensional voltages associated with the fault type corresponding to the historical fault point, specifically as follows: Fuse the multi-dimensional voltages associated with the fault type corresponding to each historical fault point through the characteristic variable formula and construct a corresponding characteristic variable; calculate the characteristic vector corresponding to the current detection voltage sequence and input the characteristic vector into the classification model; The characteristic variable formula is as follows Among them, U2 is the secondary side voltage of the rectifier, and U d1 is the DC bus voltage, and U d2 is half of the bus voltage, and F x1 is the characteristic variable; The characteristic indices are obtained by performing short-time Fourier transform on the characteristic variable and statistically calculating the characteristic variable, specifically as follows: Perform short-time Fourier transform on each of the said characteristic variables, and extract the average value of the DC component of the sequence after the short-time Fourier transform of the characteristic variable and the variance of the corresponding sequence value after the short-time Fourier transform of the characteristic variable; perform statistical calculation on each of the said characteristic variables to calculate the characteristic variables, and extract the average value of the fused characteristic variables and the variance of the fused characteristic variables; Fuse the said multiple corresponding characteristic indicators to construct a corresponding characteristic vector.

8. A device for detecting a ground fault in the main circuit of a traction system according to claim 7, characterized in that, The said characteristic matrix is: Among them, FV is a matrix, C n is the fault type, I n (k) is the eigenvector corresponding to the fault type.

Images