CN110991471B

CN110991471B - Fault diagnosis method for high-speed train traction system

Info

Publication number: CN110991471B
Application number: CN201910708383.0A
Authority: CN
Inventors: 冒泽慧; 闫宇; 姜斌; 严星刚; 吕迅竑
Original assignee: Nanjing University of Aeronautics and Astronautics
Current assignee: Nanjing University of Aeronautics and Astronautics
Priority date: 2019-08-01
Filing date: 2019-08-01
Publication date: 2023-07-11
Anticipated expiration: 2039-08-01
Also published as: CN110991471A

Abstract

The invention discloses a fault diagnosis method for a traction system of a high-speed train, which relates to the field of fault diagnosis of the high-speed train and comprises the following steps: collecting sequence data from a semi-physical simulation platform and preprocessing the sequence data to obtain a data set, wherein the data set comprises a training set and a testing set; improving an LSTM self-encoder by utilizing state differential feedback control to obtain an LSTM self-encoder I, wherein the LSTM self-encoder I consists of L LSTM units; training the LSTM self-encoder I by utilizing the data set to obtain an LSTM self-encoder II; extracting original feature vectors from the second pair of test sets by using the LSTM self-encoder; performing feature dimension reduction on the original feature vector by using a t-SNE algorithm; performing fault diagnosis on the original feature vector after dimension reduction by a DBSCAN clustering method to obtain a diagnosis result. The invention can effectively solve the problems of difficult diagnosis and low diagnosis accuracy of the tiny gradual change fault of the high-speed train traction system.

Description

Fault diagnosis method for high-speed train traction system

Technical Field

The invention relates to the field of fault diagnosis of high-speed trains, in particular to a diagnosis method of tiny gradual change faults of a high-speed train traction system based on data learning.

Background

At present, high-speed trains play an increasingly important role in China passenger transport and freight transport, and a traction system is used as a power core system of the high-speed trains, so that accidents such as train stopping and delay are caused by faults of the traction system, and huge losses are caused. The micro gradual change fault occurs in the initial stage of the significant fault, and has the characteristics of unobvious fault characteristics and easy unknown disturbance and noise masking. Therefore, the detection and diagnosis of the minor gradual failure of the traction system are difficult. The small gradual change faults of the traction system are effectively detected and diagnosed, effective measures are timely taken, the safety of the system can be improved, and the maintenance cost and loss are reduced.

The effective fault diagnosis of the micro gradual change fault can improve the safety of the system, reduce the maintenance cost and loss, and has extremely important significance in the aspects of safety maintenance, health management of equipment and the like. Because the complexity of the traction system and the fault characteristics of the tiny gradual change faults are not obvious, the change characteristics are not obvious, and the detection and diagnosis of the tiny gradual change faults are difficult to effectively realize based on a data driving algorithm at present.

Disclosure of Invention

The invention aims to provide a fault diagnosis method for a high-speed train traction system, which aims to solve the technical problems of difficult diagnosis and low diagnosis accuracy of tiny gradual change faults in the prior art.

The invention provides a fault diagnosis method for a traction system of a high-speed train, which comprises the following steps:

(a) Collecting sequence data from a semi-physical simulation platform and preprocessing the sequence data to obtain a data set, wherein the data set comprises a training set and a testing set;

(b) Improving an LSTM self-encoder by utilizing state differential feedback control to obtain an LSTM self-encoder I, wherein the LSTM self-encoder I consists of L LSTM units;

(c) Training the LSTM self-encoder I by utilizing the data set to obtain an LSTM self-encoder II;

(d) Extracting original feature vectors from the second pair of test sets by using the LSTM self-encoder;

(e) Performing feature dimension reduction on the original feature vector by using a t-SNE algorithm;

(f) Performing fault diagnosis on the original feature vector after dimension reduction by a DBSCAN clustering method to obtain a diagnosis result.

Compared with the prior art, the diagnosis method for the micro gradual change fault of the high-speed train traction system based on data learning has the following beneficial effects:

the method can better extract the original characteristics of the micro gradual change fault by carrying out characteristic extraction on the basis of the sequence data of the micro gradual change fault by the improved LSTM unit self-encoder;

the original characteristics are subjected to dimension reduction processing through a t-SNE algorithm, redundancy of the original characteristic information is reduced, diagnosis performance is improved, and calculation complexity is reduced;

the unsupervised diagnosis of the micro gradual change faults is realized by the DBSCAN clustering method, the marking in advance is not needed, the diagnosis accuracy is very high, the unknown faults have certain diagnosis capability, and the micro gradual change fault diagnosis requirement is met.

Aiming at a traction motor system of a high-speed train, the method is used for diagnosing tiny gradual faults such as a middle capacitor, a middle resistor, a speed sensor degradation fault and the like of the traction system, has a wider application range and has engineering application value.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings that are needed in the description of the embodiments or the prior art will be briefly described, and it is obvious that the drawings in the description below are some embodiments of the present invention, and other drawings can be obtained according to the drawings without inventive effort for a person skilled in the art.

FIG. 1 shows an improved LSTM cell structure provided by an embodiment of the present invention;

FIG. 2 (a) is a block diagram of an overall framework and training process for a codec model according to an embodiment of the present invention;

FIG. 2 (b) is a general framework and flow diagram providing a fault diagnosis model for an embodiment of the present invention;

FIG. 3 (a) is an effect diagram of different dimension reduction methods for providing known fault types according to an embodiment of the present invention;

FIG. 3 (b) is an effect diagram of a method for dimension reduction of t-SNE for unknown fault types according to an embodiment of the present invention;

FIG. 4 (a) provides a graph of known fault class clustering effects for an embodiment of the present invention;

FIG. 4 (b) provides a graph of unknown fault class clustering effects for an embodiment of the present invention;

FIG. 5 is a schematic diagram of an architecture of a high speed train traction system according to an embodiment of the present invention;

Detailed Description

The following description of the embodiments of the present invention will be made apparent and fully in view of the accompanying drawings, in which some, but not all embodiments of the invention are shown. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

In the description of the present invention, it should be noted that the directions or positional relationships indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. are based on the directions or positional relationships shown in the drawings, are merely for convenience of describing the present invention and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a specific orientation, be configured and operated in a specific orientation, and thus should not be construed as limiting the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance.

In the description of the present invention, it should be noted that, unless explicitly specified and limited otherwise, the terms "mounted," "connected," and "connected" are to be construed broadly, and may be either fixedly connected, detachably connected, or integrally connected, for example; can be mechanically or electrically connected; can be directly connected or indirectly connected through an intermediate medium, and can be communication between two elements. The specific meaning of the above terms in the present invention will be understood in specific cases by those of ordinary skill in the art.

The following describes the embodiments of the present invention in further detail with reference to the drawings.

As shown in fig. 2 (b), the diagnosis method of the micro gradual change fault of the high-speed train traction system based on data learning mainly comprises the following steps:

1. collecting sequence data from a semi-physical simulation platform and preprocessing to obtain a data set, wherein the data set comprises a training set and a testing set

Preprocessing the acquired data: supplementing missing values, modifying abnormal values, smoothing filtering, reducing noise and normalizing.

Specifically, filling missing values in data by adopting a 2-nearest neighbor method for missing values of the training set and the testing set; analyzing abnormal values of the sequence data through a box graph, and modifying the numerical value which is judged to be the abnormal value in the data by adopting a 2-nearest neighbor method; the sequence data is subjected to smooth filtering in a wavelet filtering mode, so that noise in the sequence data is reduced; and normalizing the sequence data by adopting a Z-normalization method.

2. Improving an LSTM self-encoder by using state differential feedback control to obtain an LSTM self-encoder I, wherein the LSTM self-encoder I consists of L LSTM units

The basic LSTM unit is improved, and the state differential feedback control is applied to the LSTM unit: state differential feedback control is applied in the modified LSTM cell.

Specifically, the state differential signal p of the state vector of the dynamic system at the time t-1 and the time t-2 _t-1 Updated at time t-1 and fed back to the system at time t to participate in controlling LSTM hidden unit state vector c _t Forget and update of (a). The modified LSTM cell description is shown in the following formula:

wherein f _t Representing forget gate layer output, r _t Representing updated gate layer output, c _t The state of the hidden unit is indicated,

representing the new state quantity to be selected, p _t Differential vector, o, representing a unit hidden state vector _t Represents the output gate layer output, h _t Representing the LSTM cell output.

If LSTM control vector

And if so, the LSTM unit hidden unit state c is stabilized under the action of stabilizing the vector to be updated, so that the LSTM network is stabilized. Control vector->

Related to the state differential vector p of LSTM cell state c. When the sequence sample of the tiny gradual change fault sequence data is short, the differential vector of the fault information variable changes linearly when the Taylor expansion is first-order approximation. When the LSTM unit learns that the hidden unit state c has higher correlation with the tiny gradual change fault information variable, the differential vector p of the LSTM unit hidden unit state c is stable along with the iterative input of the sequence data. But->

Has a strong correlation with the differential vector p, thus generating a stable control vector +.>

At this time, the LSTM unit hidden unit state variable is not easy to forget and update. When the correlation between the hidden unit state c learned by the LSTM unit and the tiny gradual change fault information variable is weak, the differential vector p is unstable, and the hidden unit state c of the LSTM unit is controlled by the unstable control vector +>

Can be forgotten and updated frequently. Secondly, the LSTM self-encoder after training can acquire differential trend signals representing fault information variables, and can diagnose micro gradual change faults earlier.

3. Training the LSTM self-encoder I with the data set to obtain the LSTM self-encoder II

And acquiring the LSTM self-encoder by using the feature vector of the training set to unsupervised train the codec model. Improved LSTM based self-encoderAfter the system sequence data is processed, fault information variables are embedded into the original characteristic vectors with fixed lengths. During the training of the codec model, the LSTM reads from the encoder the sequence samples X of the traction system, each sample

Is vector type sequence data, wherein X=x ₁ ,x ₂ ...,x _L-1 ,x _L ，N _v Representing the vector dimension, L represents the sequence data sample length. After the reading process of the sequence data sample X is completed, LSTM is embedded in the encoder from the last layer hidden unit state +.>

Will be retained and serve as the original feature vector. The GRU decoder initializes the hidden unit state of the first layer to the original feature vector +.>

And outputs the target sequence +.>

Wherein->

Training is targeted to codec output sequences

Remains similar under the metric to the input sequence X, but in reverse order. In the present invention, by minimizing +.>

And square error J between X _ed Training:

wherein D is _N Representing the entire training set, in the present invention, the plaitedThe decoder model trained optimizer is Adam optimizer.

4. Extracting original feature vectors from the test set by using the LSTM self-encoder and using a trained LSTM self-encoder model, inputting a sequence sample X of the test set at an input part, each sample

Will be retained and used as the original feature vector to complete the task of the LSTM self-encoder to extract the original feature vector of the test set data.

The trained LSTM self-encoder extracts the original feature vector from the sequence data of the small gradual change fault of the traction system.

5. Performing feature dimension reduction on the original feature vector by using t-SNE algorithm

The original feature vector is subjected to t-SNE algorithm

The dimension reduction is carried out, and the problems of high dimension of the original feature vector and information redundancy are solved. In the t-SNE algorithm, the loss function J _SNE The joint probability P of data distribution in the original feature vector space and the target dimension reduction space _D And Q is equal to _D Is a difference in (a) between the two. The probability distribution difference uses a Kullback-Leibler divergence. Loss function J of T-SNE algorithm _SNE The expression is as follows:

wherein p is _ij And q _ij Is defined as:

subscripts i, j, k, l denote samples numbered i, j, k, l, respectively, of the samples,/->

Representing the original feature vector in the high-dimensional space, z representing the feature vector after dimension reduction, σ representing the sum of the sum and the original feature vector +.>

The bandwidth of the associated gaussian kernel function. In the optimization process, the KL divergence is minimized by a random gradient descent method. Loss function J _SNE The gradient information of (2) is given by:

in the t-SNE algorithm, the similarity of feature vector distribution before and after dimension reduction is kept by minimizing the KL distance from the high-dimension original feature vector space to the low-dimension feature vector space.

6. Performing fault diagnosis on the original feature vector subjected to dimension reduction by using a DBSCAN clustering method to obtain a diagnosis result

And performing unsupervised training on the coder-decoder model by using the training set sequence data to obtain the LSTM self-encoder model. Raw eigenvector of sequence data to be diagnosed

The LSTM is extracted from the encoder by training. Original feature vector +.>

The dimension is high, information redundancy exists, in order to reduce the information redundancy and the complexity of algorithm operation, the original feature vector c is subjected to the t-SNE algorithm _tL And obtaining a final feature vector z after dimension reduction. And finally, clustering the feature vector z by a DBSCAN clustering method to realize the diagnosis of the micro gradual change faults of the traction system.

The following simulation verification is carried out on the diagnosis method of the tiny gradual change fault of the high-speed train traction system based on data learning:

and 1, firstly, confirming the types of micro gradual change faults, wherein the micro gradual change fault types comprise an intermediate capacitance degradation fault (fault I), an intermediate resistance degradation fault (fault II) and a speed sensor degradation fault (fault III) in the verification. Also included in this experimental verification is a speed sensor bias fault (fault IV), which is a minor abrupt fault. The fault I and the fault II are used for unsupervised fault diagnosis model training and final fault diagnosis model testing processes. The untrained fault class (fault III and fault IV in the experimental verification) is only used for testing and verifying the trained fault diagnosis model, and the fault III and the fault IV are used for testing the generalization capability of the model to unknown faults. The detailed descriptions of the four faults in the experimental verification are shown in table 1.

TABLE 1 description of fault types

The data of the faults I and II are divided into a training set and a testing set, and the data form of each sample is a vector of sequence data. There are a total of about 14000 samples in the form of sequence data vectors in the overall data. 60% of the data in the whole dataset was randomly selected to make up the training set (8500 samples), with the remaining samples making up the test set (5455 samples). Wherein, the healthy sample number is 1180, the fault I sample number is 990, and the fault II sample number is 935. Likewise, the number of samples for failure III is 1175, the number of samples for failure IV is 1175, and the length of each sequence data sample ranges from 20 to 40.

And 2, extracting original feature vectors of the fault I and fault II test sets by the improved LSTM self-encoder which is completed through training. And applying t-SNE to perform feature vector dimension reduction. Fig. 3 (a) shows the dimensionality reduction effect of different dimensionality reduction methods (including t-SNE, kernel PCA and Isomap) on feature vectors of known fault types. Where, + represents a healthy sample point, + represents a sample point with fault I, x represents a sample point with fault II. As can be seen from the PCA dimension reduction result and the Isomap dimension reduction result, the feature vectors of the health samples represent stable annular distribution in the low-dimensional space, and the two types of micro-gradual-change faults of the faults I and II represent conical-like distribution starting from the health annular in the low-dimensional space, which is consistent with the slow time-varying characteristic of the fault information variable in the micro-gradual-change fault, so that the LSTM self-encoder is proved to successfully extract the feature vectors related to the micro-gradual-change fault information variable. In the present invention, an experimental verification test is performed on the type of failure (unknown failure class) that is not used for codec training. The test sets of faults III and IV are used to test the generalization ability of the fault diagnosis algorithm to unknown fault types. Fault III is a minor gradual fault in the traction system and fault IV is a minor abrupt fault occurring in the traction system. Fig. 3 (b) shows the dimension reduction effect of the feature vector for an unknown fault using the t-SNE algorithm, with good inter-class and intra-class distances for the unknown fault class.

And step 3, clustering the feature vector samples subjected to dimension reduction by using a DBSCAN clustering algorithm to finish fault diagnosis of the test set. The DBSCAN clustering algorithm is a density-based clustering algorithm that generally assumes that the class can be determined by how tightly the sample is distributed. Samples of the same class are closely connected, that is, there must be samples of the same class around any sample of that class. By grouping closely connected samples into one class, a cluster class is thus obtained. By grouping all closely connected sets of samples into different categories we get the final all clusters category result. The method describes how tight the sample set is based on a set of neighbors. The DBSCAN method is simple, it selects a core object without category as seed, then finds out the sample set with reachable density of the core object, which is a cluster. Then, another core object without category is continuously selected to search a sample set with reachable density, and thus, another cluster is obtained. Run until all core objects have a class.

Fig. 4 (a) clusters data of known fault types, and the clustering result shows that the fault diagnosis algorithm based on the improved LSTM self-encoder can successfully realize degradation fault diagnosis of the intermediate capacitance and the intermediate resistance of the traction system, namely realize micro gradual change fault diagnosis of the traction system. Fig. 4 (b) clusters data of unknown fault types, and the clustering result shows that the algorithm provided by the invention has good diagnosis effect on some unknown faults.

As can be seen from FIG. 4, the method of the invention can effectively diagnose the tiny gradual change faults of the traction motor of the high-speed train. For some unlabeled faults, the method also has good diagnosis effect and relatively strong generalization capability.

As shown in fig. 5, the embodiment of the present invention further provides a schematic architecture of a traction system for a high-speed train, where the traction system 120 includes: at least one processor 121, such as a CPU, at least one network interface 124 or other user interface 123, memory 125, and at least one communication bus 122. The communication bus 122 is used to enable connected communication between these components. Optionally, a user interface 123 is also included, including a display, keyboard or pointing device (e.g., mouse, trackball, touch pad or touch sensitive display). The memory 125 may comprise a high-speed RAM memory or may further comprise a non-volatile memory (non-volatile memory), such as at least one disk memory. Memory 125 may optionally include at least one storage device located remotely from the aforementioned processor 121.

In some implementations, the memory 125 stores the following elements, executable modules or data structures, or a subset thereof, or an extended set thereof:

an operating system 1251 containing various system programs for implementing various basic services and handling hardware-based tasks;

the application 1252 includes various applications for implementing various application services.

Specifically, the processor 121 is configured to execute the processes corresponding to steps (1) - (4); in this process, the processor 121 needs to accept the signal processing sent by the sensor set 126 through the network interface 124.

Meanwhile, the processor 121 is further configured to: the fault diagnosis result is output to the visual interface through the user interface 123.

Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention, and not for limiting the same; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some or all of the technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit of the invention.

Claims

1. A fault diagnosis method for a traction system of a high-speed train is characterized by comprising the following steps of:

(f) Performing fault diagnosis on the original feature vector subjected to dimension reduction by a DBSCAN clustering method to obtain a diagnosis result;

the (c) includes:

training the LSTM self-encoder I by utilizing the data set to obtain an LSTM self-encoder II;

the LSTM reads the data set from the encoder, the data set

For vector sequence data, x=x ₁ ,x ₂ ...,x _L-1 ,x _L ，N _v Representing vector dimensions, L representing sequence data sample length;

the LSTM self-encoder has a last layer hidden unit state

As a feature vector, the hidden unit state of the first layer of the GRU decoder is +.>

The GRU decoder output sequence +.>

Wherein->

For the said

Minimizing the square error J with X _ed The training is carried out by the device,

wherein D is _N Representing the entire training set, the training remains similar under the metrics, in reverse order.

2. The high-speed train traction system fault diagnosis method as claimed in claim 1, wherein: the pretreatment in (a) comprises:

filling the missing value of the sequence data by using a k-nearest neighbor method;

analyzing abnormal values of the sequence data through a box graph, and modifying the sequence data judged to be the abnormal values by using a k-nearest neighbor method;

smoothing and filtering the sequence data by utilizing wavelet filtering, and normalizing the sequence data by utilizing a Z-score normalization method.

3. The high-speed train traction system fault diagnosis method as claimed in claim 1, wherein: in the (d), further comprising:

the LSTM extracts an original feature vector from encoder two, the original feature vector comprising fault information, and embeds the fault information variable into the original feature vector of fixed length.

4. The high-speed train traction system fault diagnosis method as claimed in claim 1, wherein: the (e) includes:

definition of the loss function J _SNE The loss function J _SNE The probability P is combined for the original feature vector and the data distribution in the target dimension reduction _D And Q is equal to _D The probability P _D And Q is equal to _D The difference probability distribution difference of (1) adopts a Kullback-Leibler divergence;

Kullback-Leibler divergence is minimized by a random gradient descent method.

5. The high-speed train traction system fault diagnosis method as claimed in claim 1, wherein: in the above-mentioned (f),

and performing fault diagnosis on the feature vector subjected to dimension reduction by a DBSCAN clustering method to obtain a diagnosis result.