AU2021101433A4

AU2021101433A4 - LSTM Method For Fault Detection Of The High-Speed Train Steering System Based On Generative Adversarial Network

Info

Publication number: AU2021101433A4
Application number: AU2021101433A
Authority: AU
Inventors: Qian Shi; Hui Zhang
Original assignee: Beihang University
Current assignee: Beihang University
Priority date: 2020-10-28
Filing date: 2021-03-19
Publication date: 2021-05-13
Anticipated expiration: 2029-03-19
Also published as: CN112326276B; CN112326276A

Abstract

The present disclosure discloses an LSTM method for fault detection of the high-speed train steering system based on GAN (generative adversarial network), which pertains to the technical field of fault diagnosis. First of all, a signal sensor is loaded on the train; vibration signals are collected in multiple times of normal cases and in multiple times of fault conditions, respectively, and are synthesized into vectors and labeled; then, all the labeled vectors are incorporated into a data set which is divided into a training set and a test set; wherein in the training set, n fault vectors and m normal vectors are selected, and the fault vectors are oversampled by using the GAN to obtain m-n new fault status vectors. The LSTM network fault detection model is trained by using new fault state vectors, n fault vectors, m normal vectors, and 2m pieces of data, and fault detection is performed on the test set; with the test results, the index G-mean is calculated for verification. According to the present disclosure, the GAN is introduced for oversampling, so that detection errors caused by data imbalance are reduced. 1/2 Vibration signal sensors are loaded at different places on a high-speed train Vibration signals are measured at various positions by using all vibration signal sensors in the case that the high-speed train steering system works normally Vibration signals are measured at various positions by using all vibration signal sensors in the case that the high-speed train steering system fails For each measurement, the group of data in normal condition is combined into one vector that is labeled as 1; and the group of data in fault condition is combined into one vector that is labeled as -1 Measurements are performed respectively in the normal state and the fault state for several times, respective vectors are labeled correspondingly, and all labeled vectors are combined into a data set which consists of a training set and a test set In the training set, n fault state vectors Fl, F2, ... , Fn and m normal state vectors NI, N2, ... , Nm are selected; GAN is used to over-sample n fault state vectors so as to obtain m-n new fault state vectors A total of 2m pieces of data, including m-n new fault state vectors, n fault vectors F1, F2, ... , Fn, and m normal vectors N, N2, ... , Nm, are used to train the LSTM network fault detection model, and the LSTM fault detection model that has been trained is used for fault detection of a test set By using the LSTM fault detection result, the size of the index G-mean is calculated and verified FIG.1

Description

1/2

Vibration signal sensors are loaded at different places on a high-speed train

Vibration signals are measured at various positions by using all vibration signal sensors in the case that the high-speed train steering system works normally

Vibration signals are measured at various positions by using all vibration signal sensors in the case that the high-speed train steering system fails

For each measurement, the group of data in normal condition is combined into one vector that is labeled as 1; and the group of data in fault condition is combined into one vector that is labeled as -1

Measurements are performed respectively in the normal state and the fault state for several times, respective vectors are labeled correspondingly, and all labeled vectors are combined into a data set which consists of a training set and a test set

In the training set, n fault state vectors Fl, F2, ... , Fn and m normal state vectors NI, N2, ... , Nm are selected; GAN is used to over-sample n fault state vectors so as to obtain m-n new fault state vectors

A total of 2m pieces of data, including m-n new fault state vectors, n fault vectors F1, F2, ... , Fn, and m normal vectors N, N2, ... , Nm, are used to train the LSTM network fault detection model, and the LSTM fault detection model that has been trained is used for fault detection of a test set

By using the LSTM fault detection result, the size of the index G-mean is calculated and verified

FIG.1

LSTM METHOD FOR FAULT DETECTION OF THE HIGH-SPEED TRAIN STEERING SYSTEM BASED ON GENERATIVE ADVERSARIAL NETWORK

TECHNICAL FIELD The present disclosure relates to the fault diagnosis of high-speed train steering system, which pertains to the technical field of fault diagnosis, in particularly to an LSTM method for fault detection of the high-speed train steering system based on GAN (generative adversarial network).

BACKGROUND In recent years, the high-speed railway construction scale and innovation achievements in high-speed railway in China have made great progress. The rapid growth of high-speed railway operating mileage and the rapid improvement of run timing have put great challenges in front of the high-speed train safety guarantee technology. It has become one of important directions to research in the field of high-speed railway that how the safety and passenger comfort of the high-speed railway may be improved. Bogie, an important component of the train body, has its fault state reflected in abnormal vibrations of the bogie and the train body, wherein the vibration signals are all timing signals. Traditional fault diagnosis methods are all based on static signals, while most of the methods cannot be directly applied in the fault diagnosis based on vibration signals. Therefore, an LSTM-based (long-term and short-term memory network) method for fault diagnosis of the high-speed train bogie came into being. In view of that during the training of the LSTM fault diagnosis model, the bogie fault data is far less than normal data, which leads to imbalance of data and affects the accuracy and fault detection effect of the LSTM fault diagnosis model. Therefore, a method using the GAN is necessary to generate fault data, reduce the imbalance of training data, and thus improve the accuracy of fault diagnosis model.

SUMMARY In view of the above problems, the present disclosure provides an LSTM method for fault detection of the high-speed train steering system based on a generative adversarial network (GAN), which uses the GAN to generate fault data, establishes an LSTM fault detection model after sampling, and then performs fault diagnosis by using the fault detection model. The LSTM method for fault detection of the high-speed train steering system based on

GAN includes the specific steps as follows: In Step 1, vibration signal sensors are loaded at different places on a high-speed train; In Step 2, under the condition that the train steering system works properly, each vibration signal sensor is used for measuring the vibration signal at each position; wherein the vibration signals measured at all normal positions every time are grouped as a set of data; In Step 3, under the condition that the train steering system fails, each vibration signal sensor is used for measuring the vibration signal at each position; wherein the vibration signals measured at all fault positions every time are grouped as into a group of data; In Step 4, for each measurement, the group of data in normal condition is combined into one vector that is labeled as 1; and the group of data in fault condition is combined into one vector that is labeled as -1; In Step 5, measurements are performed respectively in the normal state and the fault state for several times, respective vectors are labeled correspondingly, and all labeled vectors are combined into a data set which consists of a training set and a test set; In Step 6, in the training set, n di-dimensional vectors F1, F2, ... , Fn in fault state and m di-dimensional vectors N1, N2, ... , Nm in normal state are selected, wherein n << m; the GAN is used to over-sample n fault state vectors so as to obtain m-n new fault state vectors. The specific steps are as follows: In Step 601, a three-layer perceptron neural generative network G is established, and the weight parameters are randomly initialized for even distribution within (0,1): The formula of the neural generative network G is as follows:

f(x) 1- exp(-2x) 1+exp(-2x) (2) k,j Wherein, represents a weight of neurons from the J th neuron at the i- 1 th layer to

the k th neuron at the i th layer in the neural generative network G. i=1,2,3 is a layer label

of the three-layer perceptron network, k represents a label of the i th layer, k =1,2,...,K j

represents a label of a neuron at the i1th layer, j=1,2,...,J Ji- represents a total

number of neurons at the I- 1 th layer,g9j 1 represents an output of the Ith neuron at the

- 1 th layer of the generative network, and g represents an output of the k th neuron at the th layer of the generative network.

After done with all i,k ,we may obtain final output vectors 3 ''''' g of the neural generative network G, with a number of cells in the output layer being di which equals to the dimension of the fault state vector. Here is the actual operation process: Firstly, the relation between m-n and n is discriminated. When m-n is less than or equal to J. n, 0 is initialized and set as a first fault vector F1; formula (1) is used for calculating three

layers i=1,2,3 in sequence to obtain an output 93,1 corresponding to the fault vector. Similarly, the first m-n fault vectors are selected in sequence from fault vectors F2, ... , Fn and substituted into formula (1) respectively to obtain their respectively corresponding output

vectors: 9 3,1 ,' 3 ,2 , --- 3,.-n so as to form m-n fault vectors.

When m-n is larger than n, 0 is initialized and respectively set as the fault vectors F1,

F2, ... , Fn; formula (1) is used for calculating three layers i=1,2,3 in sequence to obtain

output fault vectors 9,1 3 ' , , --- 32 3 ,n corresponding to each fault vector; then fault vectors are

selected in cycle as Fl, F2, ... , Fn, and formula (1) is used to obtain outputs 93,"1'K3,n+2, --- K3,2n

corresponding to each fault vector; in this manner, the cycle goes on until a fault amount reaches

m-n, resulting in final m-n fault vectors: In Step 602, a three-layer perceptron neural discrimination network D is established, which is used to discriminate whether the vector generated by the neural generative network G is true k,j or not, and the weight parameter ' is randomly initialized for even distribution within (0,1): dk = f(ZJ4 '* d/i)

Wherein, represents a weight of neurons from the i th neuron at the - 1 th layer to the k th neuron at the i th layer of the discrimination network D. In the discrimination

network D, d represents an output of the i th neuron at the -1th layer, and di

represents an output of the k th neuron at the i th layer of the discrimination network D.

After done with all i,k we may obtain an output of the neural discrimination network D, with a number of cells in the output layer of the neural discrimination network D being 1.

dj During the actual operation, 0 is initialized and set in sequence as the output vectors

93,1,K3,2,..93,.-n of the generative network G to be substituted into formula (3) in cycle, so as to

generate respectively corresponding m-n output vectors: d 3 ,d 3 2 ,. d 3 ,m-n, thereby forming m-n output probabilities. In Step 603, by setting a current number of training iterative steps as t=1, output sets

93,1,K3,2,..93,.-n of the neural generative network G act as the sets Fl, F2, ... , Fn containing

false fault state vectors and true fault state vectors which act as training data together; and the neural discrimination network D is trained by using the stochastic gradient descent (SGD)

method, and the weight parameter 'i of the discrimination network D is updated for

maximizing the objective function V(D) The specific process is as follows:

The false fault state vectors ,13 g , 3 2 ,..93,-- are sequentially input into the discrimination network D, and the fault vectors Fl, F2, ... , Fn are also input into the discrimination network D

in sequence and in cycle for m-n times in total before the objective function V(D) of the discrimination network is calculated. When the number of fault vectors F1, F2, ... , Fn is less than m-n, the fault vectors are input in a sequence from Fl to Fn, then once again from Fl to Fn, repeating this sequence in cycle until the number reaches m-n. When the number of fault vectors F1, F2, ... , Fn is larger or equal to m-n, first m-n vectors are selected in sequence starting from Fl.

The objective function V(D) is as below:

V(D)= log(D(Fl))+log(1-D(g,))

Wherein, D(F) represents an output result of the discrimination network D with Fl

being the input, and D(g 3 ,) represents an output result of the discrimination network D with a

false fault state vector g 3 ' being the input.

In Step 604, while the updated weight parameter of the neural discrimination network D is kept unchanged, the neural generative network G is trained by using the stochastic gradient k,j descent (SGD) method, and the weight parameter wi is updated to minimize the objective parameter ;c(G). The specific process is as follows:

The false fault state vectors 9 3,1 ,' , ,---K3,.-- 3 2 are input into the neural discrimination

network D in sequence, and the objective function 7(G) of the generative network is

calculated, wherein the objective function 7(G) is as below:

rc(G)= log(1-D(g,))

In Step 605, by setting a number of iterative steps as t=t+1, Steps 603-604 are repeated until the number of iterative steps t becomes larger than the maximum iterative steps tmax to kj k,j obtain optimal weight parameters 1I and 1 at this point; In Step 606, the fault vectors Fl, F2, ... , Fn are input into the neural generative network G k,j corresponding to the optimal weight parameter 'j to obtain a final output of m-n new fault

state vectors 1g3,1'g3,2,---93,.-- which are taken as a false fault state to be put into the

training set and labeled as -1. In Step 7, a total of 2m pieces of data, including the new fault state vectors

, n fault vectors Fl, F2, ... , Fn, and m normal vectors NI, N2, ... , Nm, are

used to train the LSTM network fault detection model, and the LSTM fault detection model that has been trained is used for fault detection of a test set. In Step 8, by using the LSTM fault detection result, the size of the index G-mean is calculated and verified; For a certain test set, the LSTM fault detection model is used to group the result into fault data and normal data; and it is used to determine as follows by combining true normal labels and fault labels of the test set: If the detected result is grouped as fault data and the true label is a fault label, a counting unit TP is output; If the detected result is grouped as fault data and the true label is a normal label, a counting unit FP is output; If the detected result is grouped as normal data and the true label is a fault label, a counting

unit FN is output; If the detected result is grouped as normal data and the true label is a normal label, a counting unit TN is output; Finally, statistics is performed for an amount of each counting unit, and the following formula is used to calculate a G-mean of the resultant fault detection:

TP TN Gmean = * W TP+FN TN+FP In comparison to the prior art, the present disclosure provides the following advantages and positive effects: 1) The present disclosure provides an LSTM method for fault detection of the high-speed train steering system based on a generative adversarial network, which performs the fault detection on the basis of LSTM and capable of processing a vibration signal which is a timing signal; 2) The present disclosure provides an LSTM method for fault detection of the high-speed train steering system based on a generative adversarial network, which introduces the generative adversarial network for over-sampling, thereby reducing detection errors due to data imbalance.

BRIEF DESCRIPTION OF THE DRAWINGS Fig. 1 is a flowchart of the LSTM method for fault detection of the high-speed train steering system based on generative adversarial network according to the present disclosure; Fig. 2 is a flowchart of over-sampling by using the generative adversarial network to obtain a new fault state vector according to the present disclosure.

DETAILED DESCRIPTION So that those of ordinary skills in the art may understand and implement the present disclosure more easily, the present disclosure will be further described in detail and in depth with reference to accompanying drawings. The present disclosure provides an LSTM method for fault detection of the high-speed train steering system based on a generative adversarial network, which firstly generates fault data by using the generative adversarial network, and performs over-sampling of the fault data, and, based on the above processes, trains and tests an LSTM model. Specifically, sensors are loaded on the high-speed train steering system to measure vibration signals at various positions of the high-speed train steering system under normal and fault conditions. Each group of signals are combined into a vector xi, and the normal and fault states corresponding to the vector are labeled respectively. The state is labeled as yi=1 if normal, and labeled as yi=-1 if fault. All collected vectors xi and yi are combined into a data set which is equally divided into a training set and a test set. In the training set, an amount of fault data n is obtained by counting the fault state vectors, and an amount of normal data m is obtained by counting the normal state vectors. In the case of the amount of fault data n«m, the fault data vectors are oversampled by using the generative adversarial network. All the oversampled training data are combined into X, the training starts for the LSTM fault detection model, and the trained LSTM fault detection model is used to detect faults. As shown in Fig. 1, the specific steps are as follows: In Step 1, vibration signal sensors are loaded at different places on a high-speed train; In Step 2, under the condition that the train steering system works properly, each vibration signal sensor is used for measuring the vibration signal at each position; wherein the vibration signals measured at all normal positions every time are grouped as a set of data; In Step 3, under the condition that the train steering system fails, each vibration signal sensor is used for measuring the vibration signal at each position; wherein the vibration signals measured at all fault positions every time are grouped as into a group of data; In Step 4, for each measurement, the group of data in normal condition is combined into one vector that is labeled as 1; and the group of data in fault condition is combined into one vector that is labeled as -1; In Step 5, measurements are performed respectively in the normal state and the fault state for several times, respective vectors are labeled correspondingly, and all labeled vectors are combined into a data set which consists of a training set and a test set; In Step 6, in the training set, n di-dimensional vectors F1, F2, ... , Fn in fault state and m di-dimensional vectors N1, N2, ... , Nm in normal state are selected, wherein n « m; the GAN is used to over-sample n fault state vectors so as to obtain m-n new fault state vectors. The specific steps are as follows: In Step 601, a three-layer perceptron neural generative network G is established, and the weight parameters are randomly initialized for even distribution within (0,1): The formula of the neural generative network G is as follows: f(x) 1- exp(-2x) 1+exp(-2x) (2) k,j Wherein, represents a weight of neurons from the i th neuron at the i- 1 th layer to the k th neuron at the i th layer in the neural generative network G. i=1,2,3 is a layer label of the three-layer perceptron network, k represents a label of the i th layer, k =1,2,...,K j represents a label of a neuron at the i1th layer, j=1,2,...,J, Ji- represents a total number of neurons at the i- 1 th layer, g i 1 represents an output of the th neuron at the

After done with all i,k ,we may obtain final output vectors 3 ''"'5g of the neural generative network G, with a number of cells in the output layer being di which equals to the dimension of the fault state vector. Here is the actual operation process: Firstly, the relation between m-n and n is discriminated. When m-n is less than or equal to J. n, 0 is initialized and set as a first fault vector F1; formula (1) is used for calculating three

vectors: , 9 3,1 1' 3 2 , .3,.-n so as to form m-n fault vectors.

output fault vectors 9,1 , 3 1' 3 2 .. 3 ,n corresponding to each fault vector; then fault vectors are

selected in cycle as Fl, F2, ... , Fn, and formula (1) is used to obtain outputs 93,"+1'K3,n+2,--K3,2n

m-n, resulting in final m-n fault vectors: In Step 602, a three-layer perceptron neural discrimination network D is established, which is used to discriminate whether the vector generated by the neural generative network G is true k,j or not, and the weight parameter'4 is randomly initialized for even distribution within (0,1): dk = fJ(Z * di)

) k,j Wherein, , represents a weight of neurons from the ith neuron at the - 1 th layer to the k th neuron at the i th layer of the discrimination network D. In the discrimination network

1 th layer, and di represents an D, dil represents an output of the Ith neuron at the i- output of the k th neuron at the i th layer of the discrimination network D.

After done with all i,k , we may obtain an output scalar d' of the neural discrimination network D, with a number of cells in the output layer of the neural discrimination network D

being 1, wherein the output d represents a probability that the input comes from the true fault data.

During the actual operation, di0 is initialized and set in sequence as the output vector

93,1,93,2..93,.-n of the generative network G to be substituted into formula (3) in cycle, so as to

generate respectively corresponding m-n output vectors: d3,1 3 ,2, .d3,--n, thereby forming m-n output probabilities. When a probability value is larger than 0.5, it indicates that the output vector corresponding to the neural generative network G is true, otherwise false. In Step 603, by setting a current number of training iterative steps as t=1, output sets

93,1,93,2...93,.-n of the neural generative network G act as the set Fl, F2, ... , Fn containing false

fault state vectors and true fault state vectors which act as training data together; and the neural discrimination network D is trained by using the stochastic gradient descent (SGD) method, and

the weight parameter' of the discrimination network D is updated for maximizing the

objective function V(D) .

The specific process is as follows:

The false fault state vectors ,3 1' , 3 2 .. 3,.-- are sequentially input into the discrimination network D, and the fault vectors F1, F2, ... , Fn are also input into the discrimination network D

in sequence and in cycle for m-n times in total before the objective function V(D) of the discrimination network is calculated. When the number of fault vectors F1, F2, ... , Fn is less than m-n, the fault vectors are input in a sequence from Fl to Fn, then once again from Fl to Fn, repeating this sequence in cycle until the number reaches m-n. When the number of fault vectors F, F2, ... , Fn is larger or equal to m-n, first m-n vectors are selected in sequence starting from F1.

The objective function V(D) is as below:

V(D)= Ylog(D(Fl)) +log(1 -D(g3,1)) 1=1 (4)

being the input, and D(g 3 ,1 ) represents an output result of the discrimination network D with a

false fault state vector 93 ' being the input.

In Step 604, while the updated weight parameter of the neural discrimination network D is kept unchanged, the neural generative network G is trained by using the stochastic gradient k,j descent (SGD) method, and the weight parameter u' is updated to minimize the objective

parameter rc(G). The specific process is as follows:

The false fault state vectors 93 ,1'g 3 ,2 ,---K3,.-- are input into the neural discrimination

network D in sequence, and the objective function r(G) of the generative network is

calculated, wherein the objective function rc(G) is as below:

rc (G)=Y log(1 -D(g3,1)) 1=1

In Step 605, by setting a number of iterative steps as t=t+1, Steps 603-604 are repeated until the number of iterative steps t becomes larger than the maximum iterative steps tmax to kj k,j obtain optimal weight parameters UI and 1 at this point; In Step 606, the fault vectors Fl, F2, ... , Fn are input into the neural generative network G k,j corresponding to the optimal weight parameter uj to obtain a final output of m-n new fault

state vectors 1g3,1'g3,2,-.'3,.-- which are taken as a false fault state to be put into the

, n fault vectors Fl, F2, ... , Fn, and m normal vectors NI, N2, ... , Nm, are

used to train the LSTM network fault detection model, and the LSTM fault detection model that has been trained is used for fault detection of a test set. The specific process is as follows:

Firstly, the training data is used to take the cross entropy as the training target, and the Adam optimization algorithm is used to train the long-term and short-term memory network

(LSTM) f, and then obtains the parameter , b of LSTM after training. The cross entropy is maximized as shown in formula (5). z= Ls(x, P, b)

a = i

k

L -ZyIna, (5)

Wherein, X represents the training data, , bft represents the LSTM network

parameter, zi represents an element of the output z vector of the LSTM network, i=1,2;

a represents the fault and normal probabilities predicted by the LSTM network for the

detection signal input, wherein a 1 is the fault probability, a 2 is the normal probability; Yi

represents an element of the true label Y corresponding to the detection signal input, wherein

if the detection signal is labeled as fault, Y=1, Y2 = 0; otherwise Y1 = 0 y2 = 1

Then the test data is tested according to the following steps:

(C1) The resultant LSTM network parameter , b/f is input into all the network layers;

Wherein P includes

(C2) The collected data V to be tested is input into a fully connected layer for calculation by using formula (6):

x = g(P1 J"±+bf (6) )

e-e g(m)= tanh(m) = e e Wherein e" +e". (C3) After calculation of the fully connected layer, the calculation result is transmitted to the LSTM network layer for carrying out the following operation:

Wherein the subscript L represents an input gate, f represents a forget gate, and 0 represents an output gate; Input gate:

aL = PAX' + PLb'- + PLS

bt = f(a) (7)

1 Wherein, 1+ex Forget gate:

a' =Px' + Pyfb'- + Pf S'

b'= f(a') (8)

Cell state:

a' =Px' + Pb' S' = b' S'- + b'g(a' )

Output gate:

a' = P,,x' + Pob' +

b't = f (a') t(10 (10) Final output:

b' = b'g(S')

(C4) The LSTM network layer transmits the calculation result to the fully connected layer for calculating as shown in formula (12);

z = g(P,'b,+bj') (12) (C5) The output result of the fully connected layer is X, which is transmitted to the softmax layer, and calculated by using formula (13) to obtain the fault and normal probabilities from the detection.

eZ a = ezk

(13) Finally, the higher one of the resultant output probabilities is taken as the detection result. In Step 8, by using the LSTM fault detection result, the size of the index G-mean is calculated and verified; For a certain test set, the LSTM fault detection model is used to group the result into fault data and normal data; and it is used to determine as follows by combining true normal labels and fault labels of the test set: If the detected result is grouped as fault data and the true label is a fault label, a counting unit TP is output; If the detected result is grouped as fault data and the true label is a normal label, a counting unit FP is output;

If the detected result is grouped as normal data and the true label is a fault label, a counting unit FN is output; If the detected result is grouped as normal data and the true label is a normal label, a counting unit TN is output; Finally, statistics is performed for an amount of each counting unit, and the following formula is used to calculate a G-mean of the resultant fault detection:

TP TN Gmean = * W TP+FN TN+FP The G-mean size is discriminated for the resultant fault detection. Since the fault detection model obtained through the training data over-sampled by the generative adversarial network put more consideration on the fault data, a higher G-mean may be obtained which means that the classification result will be more sensitive to the fault data. Throughout the specification and claims, unless the context requires otherwise, the word ''comprise" or variations such as "comprises" or "comprising", will be understood to imply the inclusion of a stated integer or group of integers but not the exclusion of any other integer or group of integers.

Claims

WHAT IS CLAIMED IS:

1. An LSTM method for fault detection of the high-speed train steering system based on Generative adversarial network, comprising:

in Step 1, vibration signal sensors are loaded at different places on a high-speed train;

in Step 2, under the condition that the train steering system works properly, each vibration signal sensor is used for measuring the vibration signal at each position;

wherein the vibration signals measured at all normal positions every time are grouped as a set of data;

in Step 3, under the condition that the train steering system fails, each vibration signal sensor is used for measuring the vibration signal at each position;

wherein the vibration signals measured at all fault positions every time are grouped as into a group of data;

in Step 4, for each measurement, the group of data in normal condition is combined into one vector that is labeled as 1; and the group of data in fault condition is combined into one vector that is labeled as -1;

in Step 5, measurements are performed respectively in the normal state and the fault state for several times, respective vectors are labeled correspondingly, and all labeled vectors are combined into a data set which consists of a training set and a test set;

in Step 6, in the training set, n di-dimensional vectors Fl, F2, ... , Fn in fault state and m di-dimensional vectors N1, N2, . . , Nm in normal state are selected, wherein n « m; the GAN is used to over-sample n fault state vectors so as to obtain m-n new fault state vectors;

the specific steps are as follows:

in Step 601, a three-layer perceptron neural generative network G is established, and the weight parameters are randomly initialized for even distribution within (0,1):

the formula of the neural generative network G is as follows:

gk = f(' w * g. 1)

1 x)I- exp(-2x) 1+exp(-2x) (2)

wherein, wki represents a weight of neurons from the th neuron at the -th - layer to the k th neuron at the ith layer in the neural generative network G; i=1,2,3 is a layer label of the three-layer perceptron network, k represents a label of the ith layer, k =1,2,...,K j represents a label of a neuron at the i1th layer, j=1,2,..., Ji- J- represents a total number of neurons at the i- 1 th layer, g' represents an output of the Ith neuron at the k i 1 th layer of the generative network, and gi represents an output of the k th neuron at the i th layer of the generative network; after done with all i,k we may obtain final output vectors 3 =[g 3 ,--#3 3 ]of the neural generative network G, with a number of cells in the output layer being di which equals to the dimension of the fault state vector; in Step 602, a three-layer perceptron neural discrimination network D is established, which is used to discriminate whether the vector generated by the neural generative network G is true or not, and the weight parameter rk''j is randomly initialized for even distribution within (0,1): d= f( ikj* di) 1 j=1 (3) ) wherein, 5 is a weight of neurons from the i th neuron at the i- 1 th layer to the k th neuron at the ith layer in the discrimination network D; di ,represents an output of the ith neuron at the i - 1 th layer of the discrimination network D, and di represents an output of the k th neuron at the i th layer of the discrimination network D; after done with allk , we may obtain an output d of the neural discrimination network D, with a number of cells in the output layer of the neural discrimination network D being 1; and during the actual operation, dj is initialized and set in sequence as the output vectors

93,193,2 --- 93,n-n of the generative network G to be substituted into formula (3) in cycle, so as to

generate respectively corresponding m-n output vectors: d3,d3,...d3"-", thereby forming m-n output probabilities;

in Step 603, by setting a current number of training iterative steps as t=1, output sets

9 3 ,g 3, 2 ,---.g 3 ,.n of the neural generative network G act as the sets F1, F2, ..., Fn containing

false fault state vectors and true fault state vectors which act as training data together; and the neural discrimination network D is trained by using the stochastic gradient descent (SGD) method, and the weight parameter r' of the discrimination network D is updated for maximizing the objective function V(D) the objective function V(D) is as below:

V(D)= log(D(Fl))+log(1- D(g 3 1, ))

being the input, and D(g,,) represents an output result of the discrimination network D with a

false fault state vector 3 ',' being the input;

in Step 604, while the updated weight parameter r' of the neural discrimination network

D is kept unchanged, the neural generative network G is trained by using the stochastic gradient descent (SGD) method, and the weight parameter W<'j is updated to minimize the objective

parameter rc(G);

the specific process is as follows:

the false fault state vectors 93,323'."9",-" are input into the neural discrimination

network D in sequence, and the objective function ;T(G) of the generative network is

calculated, wherein the objective function ;T(G) is as below: rn-n

rc(G)= log(1- D(g 3 1, ))

in Step 605, by setting a number of iterative steps as t=t+1, Steps 603-604 are repeated until the number of iterative steps t becomes larger than the maximum iterative steps tmax to obtain k, k, optimal weight parameters W and at this point;

in Step 606, the fault vectors Fl, F2, ... , Fn are input into the neural generative network G k,j corresponding to the optimal weight parameter wi to obtain a final output of m-n new fault

state vectors '3,1'3,2' '-"I which are taken as a false fault state to be put into the training set and labeled as -1;

in Step 7, a total of 2m pieces of data, including the new fault state vectors

, n fault vectors F1, F2, ... , Fn, and m normal vectors NI, N2, ... , Nm, are used to train the LSTM network fault detection model, and the LSTM fault detection model that has been trained is used for fault detection of a test set; in Step 8, by using the LSTM fault detection result, the size of the index G-mean is calculated and verified.

2. The LSTM method for fault detection of the high-speed train steering system based on the generative adversarial network according to claim 1, wherein Step 601 comprises actual operation processes as follows:

firstly, the relation between m-n and n is discriminated; when m-n is less than or equal to n,

g is initialized and set as a first fault vector F1; formula (1) is used for calculating three layers i=1,2,3 in sequence to obtain an output c o rresponding to the fault vector, the first m-n

fault vectors are selected in sequence from fault vectors F2, ... , Fn and substituted into formula (1) respectively to obtain their respective and corresponding output vectors: g 3 1,g 3 ,2 ,-- 3,-

so as to form m-n fault vectors;

when m-n is larger than n, 0 is initialized and respectively set as the fault vectors Fl, F2,

Fn; formula (1) is used for calculating three layers i=1,2,3 in sequence to obtain output fault vectors g 31 ,g 32 ,-. . 3 ,n corresponding to each fault vector; then fault vectors are selected

in cycle as Fl, F2, ... , Fn, and formula (1) is used to obtain outputs 9 , 3 ,n.2 1--K3,2.

corresponding to each fault vector; in this manner, the cycle goes on until a fault amount reaches m-n, resulting in the final m-n fault vectors: g3,g3,2,-. 3,--n

3. The LSTM method for fault detection of the high-speed train steering system based on the generative adversarial network according to claim 1, wherein Step 603 comprises following specific processes:

the false fault state vectors 9 , 3 1 , 3 2 ..- 3,.-" are sequentially input into the discrimination network D, and the fault vectors F1, F2, ... , Fn are also input into the discrimination network D

in sequence and in cycle for m-n times in total before the objective function V(D) of the discrimination network is calculated;

when the number of fault vectors Fl, F2, ... , Fn is less than m-n, the fault vectors are input in a sequence from Fl to Fn, then once again from Fl to Fn, repeating this sequence in cycle until the number reaches m-n; when the number of fault vectors Fl, F2, ... , Fn is larger or equal to m-n, first m-n vectors are selected in sequence starting from Fl.

4. The LSTM method for fault detection of the high-speed train steering system based on the generative adversarial network according to claim 1, wherein Step 8 comprises following specific processes:

for a certain test set, the LSTM fault detection model is used to group the result into fault data and normal data; and it is used to determine as follows by combining true normal labels and fault labels of the test set:

if the detected result is grouped as fault data and the true label is a fault label, a counting unit TP is output;

if the detected result is grouped as fault data and the true label is a normal label, a counting unit FP is output;

if the detected result is grouped as normal data and the true label is a fault label, a counting unit FN is output;

if the detected result is grouped as normal data and the true label is a normal label, a counting unit TN is output;

finally, statistics is performed for an amount of each counting unit, and the following formula is used to calculate a G-mean of the resultant fault detection:

Gmean TP= TP+FN TN+FP