CN113640607A - Early fault diagnosis method for inverter circuit and motor of high-speed train - Google Patents

Abstract

The invention discloses an early fault diagnosis method for an inverter circuit and a motor of a high-speed train, which mainly aims at the inverter circuit and a three-phase squirrel cage asynchronous motor in a CRH (China railway high-speed) motor train unit train and comprises the following main steps: (1) six working modes are set up for the inverter circuit and the motor by utilizing a nuclear density estimation function; (2) in each mode, processing sensor data by using a new statistical analysis method, and judging whether a fault occurs according to a certain decision criterion; (3) and after the fault is found, extracting fault characteristics to form a fault characteristic matrix, and performing fault positioning. The method makes the statistic more sensitive to the fault and can find the early fault in time; and a new matrix distance criterion is adopted, so that the accuracy of the positioning result is improved.

Description

Early fault diagnosis method for inverter circuit and motor of high-speed train

Technical Field

The invention discloses an early fault diagnosis method for an inverter circuit and a motor of a high-speed train, and belongs to the technical field of fault diagnosis of electric traction systems of the high-speed train.

Background

At present, the operating mileage of the high-speed railway in China is the first to live in the world stably. The safe and stable operation of the high-speed train is closely related to the safety of passengers, and has important influence on the development of national economy. The losses caused by high-speed rail failures are very large when not discovered in time each year. Therefore, early fault finding and processing during train operation has become a research hotspot.

The electric traction system is one of core systems of a high-speed train, and the reliability of the electric traction system is an important guarantee for the safe and stable operation of the high-speed train. The electric traction converter is used as an important component for energy conversion in a traction system, bears main electric heating stress in the operation process, is mostly in electric, magnetic, thermal and mechanical multi-field coupling in application occasions, and has high failure rate. Meanwhile, the traction motor is an important component of a train traction system. When the traction motor operates, the electric energy is converted into mechanical energy, and when the traction motor brakes, the mechanical energy is converted into the electric energy, so that the traction motor is a component for converting the electric energy and the mechanical energy of the train. Traction inverter faults and motor faults account for about 50% of the total faults by statistics. Therefore, the method has important theoretical significance and great engineering value for improving the reliability of the electric traction system and reducing the fault loss by diagnosing the faults of the traction converter and the motor on line.

At present, fault detection applied to an electric traction system of a high-speed train is mainly based on a mathematical model. The method needs to know a specific system structure to establish a mathematical model, and the whole process is relatively complex; on the other hand, this method cannot detect an early failure although it can detect a failure, and has no way of determining the type of failure that has occurred.

Disclosure of Invention

The invention aims to provide an early fault diagnosis method for an inverter circuit and a motor of a high-speed train, aiming at the problems and defects in the prior art.

In order to achieve the purpose, the invention adopts the following technical scheme to realize the purpose:

an early fault diagnosis method for an inverter circuit and a motor of a high-speed train comprises the following steps:

firstly, according to six working states of an inverter circuit and a motor, constructing six working modes for the inverter circuit and the motor by using a kernel density estimation function to obtain a clustering center of each mode;

step two, dividing the sensor data on the train into different modes by using the clustering center of each mode obtained in the step one;

thirdly, processing the sensor data for multiple times by using deep PCA under each mode to highlight fault information;

step four, on the basis of the step three, calculating data statistics, and judging whether the statistic index exceeds a threshold value or not by combining a fault detection threshold value generated by the real-time working condition, the line condition and the load condition of the train, wherein when the statistic index exceeds the threshold value, the fault is indicated;

and step five, after the fault is found, extracting fault characteristics to form a fault characteristic matrix, and matching the fault characteristic matrix with the fault characteristic matrix in the database, thereby positioning the fault and finding out the fault reason.

As a further preferred scheme of the present invention, in the first step, the specific contents and steps of constructing the six working modes are as follows:

step 1, obtaining data of current, train speed, motor rotating speed and electromagnetic torque from a sensor on a train; arbitrarily selecting a set of data as the cluster center of a class of data, i.e. x_i,j1,2,3.. 6; then take a group of data as x_new,j；

Step 2, calculating the value of the kernel density estimation function of the next group of data and the cluster center:

wherein, F_i(x_new) Is x_new,j(next set of data) and x_i,j(cluster center of i-th class data) kernel density estimation function value; m is the number of variables, and h is the width of the window; x is the number of_new,jRepresents the value of the jth data in the next set of data; x is the number of_i,jA value representing class i class center, jth data;

step 3, judging the nuclear density estimation calculated in the step 2Whether the value of the counting function exceeds a threshold value or not, and if not, taking the average value of the two groups of data as a new clustering center, namely x_i,j＝(x_new,j+x_i,j)/²Then take down a set of data as a new x_new,jThen, jumping back to the step 2; when the threshold value is exceeded, continuing to execute the step 4;

step 4. if the kernel density estimation function value calculated in step 2 exceeds the threshold value, it indicates that a cluster center is found, namely x_i,j(ii) a Then the searching for the clustering center is finished;

and 5, repeating the steps 1 to 4 until six clustering centers are found.

As a further preferable scheme of the present invention, in step three, the sensor data is processed by deep pca for multiple times in each mode to highlight the fault information, and the specific content and steps are as follows:

step 1) data in each modality is extracted, i.e.

Wherein z (N) represents data of the Nth measurement; n represents the total number of measurements; m represents the number of sensors;

step 2), calculating

Mean μ (j) and variance δ of each row²(j)；

Wherein j represents the jth row, z_j(i) Represents the ith number of the jth line;

step 3) of mixing

Data standardization of each row; obtaining a data set X with a mean value of 0 and a variance of 1;

wherein

Data obtained by normalizing the jth data are shown;

step 4), calculating a covariance matrix S of the X, and performing singular value decomposition on the S;

S＝P₁Λ_0,1P₁ ^T

wherein m represents the number of sensors; lambda_0,1∈R^m×m，Λ_0,1＝diag(λ_0,1,...λ_0,m)，λ_0,mAn mth singular value representing S; p₁∈R^m×mIs the singular value vector of S, takes P₁Is taken as P_1，2The rest part is taken as P_1，1Then P is₁＝[P_1,1,P_1,2]；

Step 5), decomposing X into two parts, X respectively_1,1,X_1,2I.e. by

X＝X_1,1+X_1,2

Wherein I is an identity matrix;

step 6), adding X_1,1,X_1,2Repeating the steps 4) and 5) to obtain X as new X_1,1,X_1,2Are each broken down into two parts, i.e.

X_1,1＝X_2,1+X_2,2

X_1,2＝X_2,3+X_2,4

X＝X_2,1+X_2,2+X_2,3+X_2,4

As a further preferable scheme of the present invention, in step four, on the basis of step three, the statistical quantity of the data is calculated, and in combination with the fault detection threshold generated by the real-time working condition, the line condition and the load condition of the train, whether the statistical quantity index exceeds the threshold is judged, and the exceeding of the threshold indicates that a fault occurs, and the specific content and the steps are as follows:

step (1) of calculating X separately_2,1,X_2,2,X_2,3,X_2,4Statistics for four data sets:

wherein the content of the first and second substances,

and SPE_j,kRepresentative data set X_j,kTwo statistics of (a);

is a reaction of with X_j,kCorresponding covariance matrix S_j,kA matrix of singular values of; p_(j+1),(2k-1)Is a reaction of with X_j,kCorresponding covariance matrix S_j,kPartial singular value vectors of (a); i is an identity matrix;

and (2) judging whether two statistics corresponding to each data exceed a set threshold value, wherein any one statistic exceeding the threshold value indicates that a fault occurs.

As a further preferred scheme of the present invention, in step five, after a fault is found, fault features are extracted to form a fault feature matrix, and the fault feature matrix is matched with a fault feature matrix in a database, so as to perform fault location, and the specific content and steps are as follows:

step I, extracting eight statistics of fault data and respective corresponding threshold values in real time to form a fault vector r each time a fault is found^c：

Wherein r is^cRepresents a type c fault;

and SPE_j,kRepresentative data set X_j,kTwo statistics of (a);

and J_SPE,j,kAre respectively

And SPE_j,kRespective threshold values; in the formula, j is 2;

step two, the fault vector r obtained in the step one is used^cProjection onto the interval [ 0.51) by the following equation, results in the vector p:

p＝[p₁,p₂,...p₈]^T

wherein the content of the first and second substances,

is represented by r^cThe kth value of (1);

during the c-th fault period, there will be N_cGrouping fault data, repeating the first step and the second step to obtain N_cArranging the vectors P in rows to form a fault characteristic matrix P^c；

Step four, calculating a fault characteristic matrix P^cMean vector mu of^cWith the covariance matrix ∑^c：

μ^c＝[μ^c ₁,μ^c ₂,...μ^c ₈]^T

Wherein the content of the first and second substances,

is P^cRow j, th value; n is a radical of_CIs P^cThe number of columns; mu.s^c _jRepresentation matrix P^cMean of j row; i is an identity matrix;

fifthly, calculating a fault characteristic matrix P through the following formula^cAnd fault special matrix in database

KLD distance of (d);

wherein the content of the first and second substances,

representing a feature matrix P^cAnd fault special matrix in database

KLD distance of (d); i is an identity matrix;

finding out a database model corresponding to the minimum KLD distance, and if the minimum KLD distance is less than or equal to a set threshold, distinguishing the distance threshold between classes to show that the same fault as the database model occurs; otherwise, a new fault type is generated, and the database needs to be updated.

Compared with the prior art, the invention adopting the technical scheme has the following technical effects: the method of the invention does not need specific circuit knowledge and mathematical models, and is simpler to implement; meanwhile, the method is more sensitive to early faults of the inverter circuit and the motor; finally, after the fault is detected, the method can also carry out fault positioning to find out the fault reason.

Drawings

Fig. 1 is a general flowchart of an early failure diagnosis method.

Fig. 2 is a schematic diagram of a high-speed train electric traction system configuration.

FIG. 3 is a flowchart of steps.

FIG. 4 is a step bipartite flow chart.

FIG. 5 is a step three-section flow chart.

Fig. 6 is a step quartering flowchart.

FIG. 7 is a flow chart of steps five.

FIG. 8 is a data set statistic T²Graph is shown.

FIG. 9 is a plot of the data set statistic SPE.

Fig. 10 is a graph of calculation results of KLD.

Detailed Description

The technical scheme of the invention is further explained in detail by combining the attached drawings:

the invention designs an early fault diagnosis method for an inverter circuit and a motor of a high-speed train, which comprises the following steps as shown in figure 1:

step one, according to six working states of an inverter circuit and a motor, six working modes are built for the inverter circuit and the motor by using a kernel density estimation function, and a clustering center of each mode is obtained, as shown in fig. 3, the steps are as follows:

step 1. As shown in FIG. 2, in the variationThe output port of the current device and the sensor are arranged in the motor to respectively obtain current signals i_a，i_b，i_cTrain speed Sp, motor speed Rs, and electromagnetic torque data Te. Arbitrarily selecting a set of data x ═ i_a,i_b,i_c,Sp,Rs,Te]^TAs a cluster center for a class of data, i.e. x_i,j1,2,3.. 6. Then take a group of data as x_new,j。

Step 2, calculating the value of the kernel density estimation function of the next group of data and the cluster center:

wherein, F_i(x_new) Is x_new,j(next set of data) and x_i,j(cluster center of i-th class data) kernel density estimation function value; m is the number of variables, and h is the width of the window; x is the number of_new,jRepresents the value of the jth data in the next set of data; x is the number of_i,jA value representing class i class center, jth data;

step 3, judging whether the kernel density estimation function value calculated in the step 2 exceeds a threshold value, and if not, taking the average value of the two groups of data as a new clustering center, namely x_i,j＝(x_new,j+x_i,j) 2, then take down a set of data as a new x_new,jThen, jumping back to the step 2; when the threshold value is exceeded, continuing to execute the step 4;

step 4. if the kernel density estimation function value calculated in step 2 exceeds the threshold value, it indicates that a cluster center is found, namely x_i,j(ii) a Then the searching for the clustering center is finished;

and 5, repeating the steps 1 to 4 until six clustering centers are found.

Step two: dividing the sensor data on the train into different modes by using the clustering center of each mode obtained in the step one, as shown in fig. 4:

after six working modes of the inverter circuit and the motor are established, the train is operatedIn running, real-time data x of the train are obtained through a sensor_t＝[i_a,i_b,i_c,Sp,Rs,Te]^TAnd calculating the distances between the real-time data and the six working modes through a kernel density estimation function, and dividing the real-time data into the nearest working modes.

Wherein, F_i(x_t) Is x_t(real time data) and x_i,j(cluster center of i-th class data) kernel density estimate function value. m is the number of variables and h is the width of the window. x is the number of_t,jIndicating the value of the jth data in the real-time data. x is the number of_i,jThe value of the ith class center, jth data is represented.

Step three: the sensor data is processed by deep pca multiple times in each modality, and fault information is highlighted, as shown in fig. 5, the steps are as follows:

step 1) data in each modality is extracted, i.e.

Wherein z (N) represents data of the Nth measurement; n represents the total number of measurements; m represents the number of sensors;

step 2), calculating

Mean μ (j) and variance δ of each row²(j)；

Wherein j represents the jth row, z_j(i) Represents the ith number of the jth line;

step 3) of mixing

Data standardization of each row; obtaining a data set X with a mean value of 0 and a variance of 1;

wherein

Data obtained by normalizing the jth data are shown;

step 4), calculating a covariance matrix S of the X, and performing singular value decomposition on the S;

S＝P₁Λ_0,1P₁ ^T

wherein m represents the number of sensors; lambda_0,1∈R^m×m，Λ_0,1＝diag(λ_0,1,...λ_0,m)，λ_0,mAn mth singular value representing S; p₁∈R^m×mIs the singular value vector of S, takes P₁Is taken as P_1，2The rest part is taken as P_1，1Then P is₁＝[P_1,1,P_1,2]；

Step 5), decomposing X into two parts, X respectively_1,1,X_1,2I.e. by

X＝X_1,1+X_1,2

Wherein I is an identity matrix;

step 6), adding X_1,1,X_1,2Repeating the steps 4) and 5) to obtain X as new X_1,1,X_1,2Are each broken down into two parts, i.e.

X_1,1＝X_2,1+X_2,2

X_1,2＝X_2,3+X_2,4

X＝X_2,1+X_2,2+X_2,3+X_2,4

Step four: on the basis of the third step, calculating data statistics, and determining whether the statistic index exceeds a threshold value by combining a fault detection threshold value generated by the real-time working condition, the line condition and the load condition of the train, wherein when the statistic index exceeds the threshold value, a fault is indicated, and as shown in fig. 6, the steps are as follows:

step (1) of calculating X separately_2,1,X_2,2,X_2,3,X_2,4Statistics for four data sets:

wherein the content of the first and second substances,

and SPE_j,kRepresentative data set X_j,kTwo statistics of (a);

is a reaction of with X_j,kCorresponding covariance matrix S_j,kA matrix of singular values of; p_(j+1),(2k-1)Is a reaction of with X_j,kCorresponding covariance matrix S_j,kPartial singular value vectors of (a); i is an identity matrix;

and (2) judging whether two statistics corresponding to each data exceed a set threshold value, wherein any one statistic exceeding the threshold value indicates that a fault occurs.

Step five: after a fault is found, extracting fault features to form a fault feature matrix, matching the fault feature matrix with a fault feature matrix in a database, and then positioning the fault to find out the fault reason, as shown in fig. 7, the steps are as follows:

step I, extracting eight statistics of fault data and respective corresponding threshold values in real time to form a fault vector r each time a fault is found^c：

Wherein r is^cRepresents a type c fault;

and SPE_j,kRepresentative data set X_j,kTwo statistics of (a);

and J_SPE,j,kAre respectively

And SPE_j,kRespective threshold values; in the formula, j is 2;

step two, the fault vector r obtained in the step one is used^cProjection onto the interval [ 0.51) by the following equation, results in the vector p:

p＝[p₁,p₂,...p₈]^T

wherein the content of the first and second substances,

is represented by r^cThe kth value of (1);

during the c-th fault period, there will be N_cGrouping fault data, repeating the first step and the second step to obtain N_cArranging the vectors P in rows to form a fault characteristic matrix P^c；

Step four, calculating a fault characteristic matrix P^cMean vector mu of^cWith the covariance matrix ∑^c：

μ^c＝[μ^c ₁,μ^c ₂,...μ^c ₈]^T

Wherein the content of the first and second substances,

is P^cRow j, th value; n is a radical of_CIs P^cThe number of columns; mu.s^c _jRepresentation matrix P^cMean of j row; i is an identity matrix;

fifthly, calculating a fault characteristic matrix P through the following formula^cAnd fault special matrix in database

KLD distance of (d);

wherein the content of the first and second substances,

representing a feature matrix P^cAnd fault special matrix in database

KLD distance of (d); i is a unitA matrix;

finding out a database model corresponding to the minimum KLD distance, and if the minimum KLD distance is less than or equal to a set threshold, distinguishing the distance threshold between classes to show that the same fault as the database model occurs; otherwise, a new fault type is generated, and the database needs to be updated.

The method of the invention is verified by simulation as follows:

step 1, simulating train operation by using a TDCS-FIB platform developed by the university of China and south. The simulation time is 3 seconds, and the motor bearing fault is injected in 1 second.

And 2, writing a program by using Matlab to realize the content of the invention.

And 3, adopting Matlab/TDCS-FIB combined simulation, importing the data generated in the step 1 into a Matlab program, and performing fault diagnosis.

FIG. 8.a is a data set X_2,1Statistic T of² _2,1A scaled plot; the broken line in the figure is the set threshold. As shown in fig. 8.a, after the failure occurs, the statistics in the graph rapidly exceed the threshold, and the change trend is very obvious. Although the value of the ordinate in FIG. 8.a is not large, in practice T is² _2,1The calculated value is very large, about 10²³This indicates the statistic T² _2,1Is very sensitive to faults.

FIG. 8.b is data set X_2,2Statistic T of² _2,2A scaled plot; the broken line in the figure is the set threshold. As shown in fig. 8.b, after the failure occurs, the statistics in the graph rapidly exceed the threshold, with a similar trend to fig. 8. d. Although the value of the ordinate in FIG. 8.b is not large, in practice T is² _2,2The calculated value is very large, about 10²³This indicates the statistic T² _2,2Is very sensitive to faults.

FIG. 8.c is data set X_2,3Statistic T of² _2,3A scaled plot; the broken line in the figure is the set threshold. As shown in fig. 8.c, after the failure occurs, the statistics in the graph quickly exceed the threshold and the maximum value varies significantly.Although the value of the ordinate in FIG. 8.c is not large, in practice T is² _2,3The calculated value is very large, about 10²³This indicates the statistic T² _2,3Is very sensitive to faults.

FIG. 8.d is data set X_2,4Statistic T of² _2,4A scaled plot; the broken line in the figure is the set threshold. As shown in fig. 8.d, after the failure occurs, the statistics in the graph rapidly exceed the threshold, with a similar trend to fig. 8. b. Although the value of the ordinate in FIG. 8.d is not large, in practice T is² _2,4The calculated value is very large, about 10²³This indicates the statistic T² _2,4Is very sensitive to faults.

FIG. 9.a is data set X_2,1Statistic SPE_2,1A scaled plot; the broken line in the figure is the set threshold. As shown in fig. 9.a, after the failure occurs, the statistic in the graph rapidly exceeds the threshold value, and the trend of the change is obvious. Theoretical statistics SPE_2,1Is not sensitive to faults, and when faults occur, the statistic SPE_2,1And not necessarily detected. However, the method of the present invention improves SPE_2,1Sensitivity to faults, so that when a fault occurs, it is passed through SPE_2,1A fault can be easily detected.

FIG. 9.b is data set X_2,2Statistic SPE_2,2A scaled plot; the broken line in the figure is the set threshold. As shown in fig. 9.b, after the failure occurs, the statistics in the graph quickly exceed the threshold and the maximum value is large. Theoretical statistics SPE_2,2Is not sensitive to faults, and when faults occur, the statistic SPE_2,2And not necessarily detected. However, the method of the present invention improves SPE_2,2Sensitivity to faults, so that when a fault occurs, it is passed through SPE_2,2A fault can be easily detected.

FIG. 9.c is data set X_2，3Statistic SPE_2,3A scaled plot; the broken line in the figure is the set threshold. As shown in fig. 9.c, after the failure occurs, the statistics in the graph quickly exceed the threshold, with a similar trend to fig. 9. a. Theoretical statistic SPE_2,3Is not sensitive to faults, and when faults occur, the statistic SPE_2,3And not necessarily detected. However, the method of the present invention improves SPE_2,3Sensitivity to faults, so that when a fault occurs, it is passed through SPE_2,3A fault can be easily detected.

FIG. 9.d is data set X_2,4Statistic SPE_2,4A scaled plot; the broken line in the figure is the set threshold. As shown in fig. 9.d, after the failure occurs, the statistics in the graph not only quickly exceed the upper limit (threshold), but also quickly exceed the lower limit. Theoretical statistics SPE_2,4Is not sensitive to faults, and when faults occur, the statistic SPE_2,4And not necessarily detected. However, the method of the present invention improves SPE_2,4Sensitivity to faults, so that when a fault occurs, it is passed through SPE_2,4A fault can be easily detected.

Fig. 10 is a calculation result of KLD. The calculated KLD values for the different fault types are very large, about 10²⁰(ii) a KLD values calculated for the same type are very small, about 10¹. The fault type can be more accurately judged by using the KLD.

The attached figures 8-10 show that the method can effectively realize the detection and the positioning of the early faults of the inverter circuit and the motor of the high-speed train, effectively solve the problems of the diagnosis of the early faults of the inverter circuit and the motor and the engineering application thereof, and has important significance for the safe operation of the high-speed train.

The embodiments of the present invention have been described in detail with reference to the drawings, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the gist of the present invention.

Claims (5)

1. An early fault diagnosis method for an inverter circuit and a motor of a high-speed train is characterized by comprising the following steps:

firstly, according to six working states of an inverter circuit and a motor, constructing six working modes for the inverter circuit and the motor by using a kernel density estimation function to obtain a clustering center of each mode;

step two, dividing the sensor data on the train into different modes by using the clustering center of each mode obtained in the step one;

thirdly, processing the sensor data for multiple times by using deep PCA under each mode to highlight fault information;

step four, on the basis of the step three, calculating data statistics, and judging whether the statistic index exceeds a threshold value or not by combining a fault detection threshold value generated by the real-time working condition, the line condition and the load condition of the train, wherein when the statistic index exceeds the threshold value, the fault is indicated;

and step five, after the fault is found, extracting fault characteristics to form a fault characteristic matrix, and matching the fault characteristic matrix with the fault characteristic matrix in the database, thereby positioning the fault and finding out the fault reason.

2. The method for diagnosing the early faults of the inverter circuit and the motor of the high-speed train according to claim 1, wherein in the step one, specific contents and steps of constructing six working modes are as follows:

step 1, obtaining data of current, train speed, motor rotating speed and electromagnetic torque from a sensor on a train; arbitrarily selecting a set of data as the cluster center of a class of data, i.e. x_i,j1,2,3.. 6; then take a group of data as x_new,j；

Step 2, calculating the value of the kernel density estimation function of the next group of data and the cluster center:

wherein, F_i(x_new) Is x_new,j(next set of data) and x_i,j(cluster center of i-th class data) kernel density estimation function value; m is the number of variables, and h is the width of the window; x is the number of_new,jRepresents the value of the jth data in the next set of data; x is the number of_i,jValues representing class i center, j data；

Step 3, judging whether the kernel density estimation function value calculated in the step 2 exceeds a threshold value, and if not, taking the average value of the two groups of data as a new clustering center, namely x_i,j＝(x_new,j+x_i,j) 2, then take down a set of data as a new x_new,jThen, jumping back to the step 2; when the threshold value is exceeded, continuing to execute the step 4;

step 4. if the kernel density estimation function value calculated in step 2 exceeds the threshold value, it indicates that a cluster center is found, namely x_i,j(ii) a Then the searching for the clustering center is finished;

and 5, repeating the steps 1 to 4 until six clustering centers are found.

3. The method for diagnosing the early failure of the inverter circuit and the motor of the high-speed train as claimed in claim 1, wherein in the third step, the deep pca is used for processing the sensor data for a plurality of times under each mode to highlight the failure information, and the specific content and the steps are as follows:

step 1) data in each modality is extracted, i.e.

Wherein z (N) represents data of the Nth measurement; n represents the total number of measurements; m represents the number of sensors;

step 2), calculating

Mean μ (j) and variance δ of each row²(j)；

Wherein j represents the jth row, z_j(i) Represents the ith number of the jth line;

step 3) of mixing

Data standardization of each row; obtaining a data set X with a mean value of 0 and a variance of 1;

wherein

Data obtained by normalizing the jth data are shown;

step 4), calculating a covariance matrix S of the X, and performing singular value decomposition on the S;

S＝P₁Λ_0,1P₁ ^T

wherein m represents the number of sensors; lambda_0,1∈R^m×m，Λ_0,1＝diag(λ_0,1,...λ_0,m)，λ_0,mAn mth singular value representing S; p₁∈R^m×mIs the singular value vector of S, takes P₁Is taken as P_1，2The rest part is taken as P_1，1Then P is₁＝[P_1,1,P_1,2]；

Step 5), decomposing X into two parts, X respectively_1,1,X_1,2I.e. by

X＝X_1,1+X_1,2

Wherein I is an identity matrix;

step 6), adding X_1,1,X_1,2Repeating the steps 4) and 5) to obtain X as new X_1,1,X_1,2Are each broken down into two parts, i.e.

X_1,1＝X_2,1+X_2,2

X_1,2＝X_2,3+X_2,4

X＝X_2,1+X_2,2+X_2,3+X_2,4。

4. The method for diagnosing the early faults of the inverter circuit and the motor of the high-speed train as claimed in claim 1, wherein in the fourth step, on the basis of the third step, the statistic of the data is calculated, and the fault detection threshold generated by combining the real-time working condition, the line condition and the load condition of the train is combined to judge whether the statistic index exceeds the threshold, wherein the exceeding of the threshold indicates the occurrence of the fault, and the specific content and the steps are as follows:

step (1) of calculating X separately_2,1,X_2,2,X_2,3,X_2,4Statistics for four data sets:

wherein the content of the first and second substances,

and SPE_j,kRepresentative data set X_j,kTwo statistics of (a);

is a reaction of with X_j,kCorresponding covariance matrix S_j,kA matrix of singular values of; p_(j+1),(2k-1)Is a reaction of with X_j,kCorresponding covariance matrix S_j,kPartial singular value vectors of (a); i is an identity matrix;

and (2) judging whether two statistics corresponding to each data exceed a set threshold value, wherein any one statistic exceeding the threshold value indicates that a fault occurs.

5. The method for diagnosing the early faults of the inverter circuit and the motor of the high-speed train as claimed in claim 1, wherein in the fifth step, after the fault is found, the fault characteristics are extracted to form a fault characteristic matrix which is matched with the fault characteristic matrix in the database so as to carry out fault location, and the specific content and the steps are as follows:

step I, extracting eight statistics of fault data and respective corresponding threshold values in real time to form a fault vector r each time a fault is found^c：

Wherein r is^cRepresents a type c fault;

and SPE_j,kRepresentative data set X_j,kTwo statistics of (a);

and J_SPE,j,kAre respectively

And SPE_j,kRespective threshold values; in the formula, j is 2;

step two, the fault vector r obtained in the step one is used^cBy the following formulaOn interval [ 0.51), we get vector p:

p＝[p₁,p₂,...p₈]^T

wherein the content of the first and second substances,

is represented by r^cThe kth value of (1);

during the c-th fault period, there will be N_cGrouping fault data, repeating the first step and the second step to obtain N_cArranging the vectors P in rows to form a fault characteristic matrix P^c；

Step four, calculating a fault characteristic matrix P^cMean vector mu of^cWith the covariance matrix ∑^c：

μ^c＝[μ^c ₁,μ^c ₂,...μ^c ₈]^T

Wherein the content of the first and second substances,

is P^cRow j, th value; n is a radical of_CIs P^cThe number of columns; mu.s^c _jRepresentation matrix P^cMean of j row; i is an identity matrix;

fifthly, calculating a fault characteristic matrix P through the following formula^cAnd fault special matrix in database

KLD distance of (d);

wherein the content of the first and second substances,

representing a feature matrix P^cAnd fault special matrix in database

KLD distance of (d); i is an identity matrix;

finding out a database model corresponding to the minimum KLD distance, and if the minimum KLD distance is less than or equal to a set threshold, distinguishing the distance threshold between classes to show that the same fault as the database model occurs; otherwise, a new fault type is generated, and the database needs to be updated.

Images