Methods of Track Circuit Fault Diagnosis Based on Hmm

International Journal of Research in Engineering and Science (IJRES)
ISSN (Online): 2320-9364, ISSN (Print): 2320-9356
www.ijres.org Volume 5 Issue 3 ǁMar. 2017 ǁPP.45-51
www.ijres.org 45 | Page
Methods of Track Circuit Fault Diagnosis Based on Hmm
Gao Tongfei1
, Zhang Wenyan2
, Deng Renzhen3
College Of Urban Railway Transportation, Shanghai University Of Engineering Science,
Shanghai, China
ABSTRACT: A fault diagnosis method of track circuit based on HMM (Hidden Markov Model) was
proposed. On the basis of division of failure mechanism of the track circuit, a training mechanism of multi -
sample HMM model was established, and a track circuit fault diagnosis system was composed by multiple fault
classifiers. Because of the universality of this system, taking the hump section of railway and a section of ZPW-
2000A non-insulated track circuit as examples, the correctness and effectiveness of the system were verified.
The result shows that the fault diagnosis method of track circuit which is based on HMM can effectively achieve
six kinds of track circuit fault diagnosis. And compared with BP Neural Network fault diagnosis, it has a higher
accuracy rate and has a faster computing speed, which can be used as a new solution for fault diagnosis of track
circuits.
Keyword: Track circuits, Fault diagnosis, Hidden Markov Model, Multi-sample training
I. INTRODUCTION
Track circuit is the basic equipment of the rail transit signal automatic control. The track circuit fault
affects the normal driving, shunting operations, and even security risks. In the actual work, the track circuit fault
diagnosis almost uses the manual regular maintenance. And the maintenance personnel professional standards,
experience accumulation and accuracy of the measurement equipment may have different degrees of impact to
track circuit fault diagnosis. Therefore, it has a great practical significance through scientific means to diagnose
the track circuit fault.
There are also a few applications on the various types of algorithms in the track circuit fault diagnosis.
In recent years, fault diagnosis algorithms based on expert system, neural network, genetic algorithm, support
vector machine algorithm and other algorithms have been widely studied, and they also made a lot of
achievements. In addition, the algorithm based on state space model is more and more applied in the field of
fault detection and prediction. Among them, Hidden Markov Model (HMM) has become a new research
direction in the field of fault diagnosis as it reveals the transition process of hidden state and it is easy to achieve
by using programming software. There are some literatures discuss the possibility of HMM applied to the
rolling bearing fault diagnosis. In addition there are many cases of HMM model optimization or complex for the
fault diagnosis study. Combined with the application of HMM in the field of related machinery, it is introduced
into the track circuit fault diagnosis research, putting forward more possibilities for the track circuit fault
diagnosis method. In this paper, HMM-based track circuit fault diagnosis method is universal. The following is
the railway hump section data as an example to introduce.
Fault Analysis of Hump Section Track Circuit
Principles of Hump Section Track Circuit
JWXC-2.3 AC closed-circuit hump track circuit is one of the main equipment of various hump
marshaling yards in China. From end to end, it is composed of interactive rheostat, two groups of diodes
composed of bridge rectifier, transformer BG1, breaker and track relay DG. In addition, the hump track circuit
set the two-segment track circuit at the shunt turnout, respectively, to protect the section DG1 and turnout
section DG. When the vehicles enter the protection zone, the track relays of the two sections immediately
demagnetize and drop. When the vehicles in the protection section lost shunt in a short time, DG section of the
track relay slowly pick up; and when the vehicles clean from the DG section, DG section of the track relay
immediately pick up, which prevent the track circuit’s wrong action caused by the shunt faults. JWXC-2.3 AC
closed-circuit hump rail circuit figure is as follows.

Methods Of Track Circuit Fault Diagnosis Based On HMM
Figure 1 JWXC-2.3 AC closed-circuit hump rail circuit
The Fault Analysis of Hump Section Track Circuit Through the fault status analysis of JWXC-2.3 type
AC closed-circuit hump track circuit, the fault is divided into indoor faults and outdoor faults. Indoor faults
include indoor open circuit, indoor short circuit, while outdoor failure including outdoor open circuit (open
circuit point before the test point, open circuit point after the test point), outdoor short circuit (short circuit point
before the test point, short circuit point after the test point). Compared the measured voltage of the fault track
circuit in a certain section of the hump section, these five data about the current limiter step-down of sending
side, rail voltage of sending side, rail voltage of receiving side, the adjustment state voltage of relay and the
adjustment state voltage of relay after throwing the line are selected to describe the fault.
The Fundamental of HMM
The principle cognition of HMM
The Hidden Markov Model (HMM) is based on the Markov chain. The Markov chain is a stochastic
process that describes a state sequence. In this sequence, the state of the future (which is the future of now) is
independent of the state of the past (which is the past of now), and each state data depends only on one of its
previous states, namely, assuming the value of
means the state at time and the value of depends only on the value of . At the same time, the
probability distribution of the system state at time depends only on the state at time, namely,.
The Markov chain
model can be described by three elements. The first one is the non-empty state set composed by all possible
states of the system. The second one is the state transition probability matrix of the system, and the last one is
the initial probability distribution of the system. In the state transition probability, since each state transition
probability is positive, it is available that, and each state transition process is transferred from one state to
another (or the same) state. So there is, in which means the probability of state turns into state.
In the Markov chain, since all states are directly visible, the state transition probability is the only
parameter. However, for some practical reasons, the observer cannot directly observe the real state of the
system, but they can observe the output depends on its state. At this time, there is a possible probability
distribution for each state to each output. This model is called the hidden Markov model, which describes a
Markov process with implicit unknown parameters. It can be seen that the hidden Markov model is a double
stochastic process. One is a hidden Markov chain with a certain number of states, and the other is a set of
random functions.

Figure 2 The evolution process of HMM model running state
At this point, the hidden Markov model can be expressed by five corresponding elements:
Implied state : These states usually can’t be obtained by direct observation, and it can use state collection to be
expressed as , in which means the number of imply states.
Observable state : they are associated with the implied state and they can be obtained by direct observation, and
it can use state collection to be showed as , in which means the number of observable states.
Initial state probability matrix : it represents the probability matrix of the implied state at the initial moment,
and it can be represented as , in which means the probability of imply state at the initial moment, namely
Implicit state transition probability matrix : it describes the transition probability between each hidden state.
So we can get , which shows the possibility of state turn into state .
Observation state transition probability matrix : it describes the probability of transition between the
observed states, in which . And this indicates that at some point, the probability of implied state is while the
observing state is .
In summary, the hidden Markov model has a successful application as a statistical model in the field of fault
diagnosis.
The basic problem of HMM
Assessment problem: For a given model, find the probability of an observation sequence. Give an
observation sequence and model parameters. The question is how to calculate the probability of generating this
observation sequence, and then make the relevant assessment of the HMM. The forward algorithm is a powerful
way to solve this problem. It can compare each probability of generating sequence produced by hidden Markov
model after calculating and select the optimal HMM model among all the results.
Decoding problem: For a given model and observation sequence, find the state sequence with the
highest probability. Give an observation sequence and model parameters . The question is how to find the
optimal hidden state sequence. In such problems, the Viterbi algorithm is usually used to find.
Learning problem: For a given sequence of observations, adjust the parameters and make the
probability of observations is greatest. That is, when the HMM model parameters are unknown, how to make
the probability of the observation sequence as large as possible by adjusting these parameters. The Baum-
Welch algorithm and the Reversed Viterbi algorithm can be used to solve this problem properly.
Fault Diagnosis Method of Rail Circuit Based on HMM
Sample normalization
In the 1.2, the study objects of fault are indoor open circuit, indoor short circuit, outdoor open circuit
(open circuit point before the test point, open circuit point after the test point), outdoor short circuit (short circuit
point before the test point, short circuit point after the test point). These five feature data about the current
limiter step-down of sending side, rail voltage of sending side, rail voltage of receiving side, the adjustment state
voltage of relay and the adjustment state voltage of relay after throwing the line are selected to describe the
fault. And 240 sets of sample data (including training data and verification data) are selected as normalized
pretreatment in order to prevent the sick matrix during the training process and easy to discriminate the
program. The parameters are normalized using the following formula:
Among them, is raw data, while, mean the maximum and minimum of the 240 groups of raw data
respectively. is the data after the normalization process. Partially processed data are shown as follows:

Table 1 Partially processed data
Numbe
r
sending side receiving
side
relay Fault type
current limiter
step-down
rail
voltage
rail voltage adjustment
state voltage
adjustment
state voltage of
relay after
throwing the
line
1 0.0285 0.9915 0.9801 0.9972 0.9943 indoor open circuit
2 0.0456 0.9972 0.9685 0.9915 0.9986 indoor open circuit
3 0.9744 0.0399 0.0285 0.0684 0.9801 indoor short circuit
4 0.9715 0.0427 0.0313 0.1254 0.9772 indoor short circuit
5 0.9858 0.7293 0.7179 0.1880 0.1795 outdoor short
circuit(before)
6 0.9915 0.7236 0.7123 0.1083 0.1026 outdoor short
circuit(before)
7 0.9886 0.0342 0.0228 0.1339 0.1282 outdoor short
circuit(after)
8 0.9858 0.0370 0.0256 0.1852 0.1766 outdoor short
circuit(after)
9 0.0256 0.7236 0.7123 0.0598 0.0285 outdoor open
circuit(before)
10 0.0370 0.7208 0.7094 0.0712 0.0342 outdoor open
circuit(before)
11 0.0256 0.9915 0.9801 0.0570 0.0285 outdoor open
circuit(after)
12 0.0342 0.9801 0.9687 0.0570 0.0256 outdoor open
circuit(after)
The Realization and Application of Fault Diagnosis Method for Rail Circuit Based on HMM In this
paper, the fault diagnosis method of track circuit is based on evaluation and learning problems of HMM. Six
kinds of HMM models are designed to perform multi-sample HMM model training for six different types of
faults. The fault classification is written in the Matlab environment. First, the number of hidden states is 6, and
the Markov chain in this case is shown in Figure 3. Then, the observation sequences of HMM are composed of
training data in five types of data after normalization, and as input variables, they will be input into six track
circuit fault classifications. The "Forward-Backward algorithm" is used to train the fault recognition and the
state transition probability matrix, output probability matrix and initial probability distribution of each model are
obtained. Finally, the fault classification model is completed. After that, the pre-processed verification data is
used as the observation sequence in order to input classification. The probability of the fault diagnosis model
generating a specific observation sequence is calculated by using the "Forward algorithm". The fault state of the
fault classification with the maximum probability is the fault state of the track circuit.
Figure 3 The Markov chain with six hidden states

The results and analysis of examples
There are 240 sets of data and six fault types in this paper. Because of its same implementation and
application, only take the outdoor short circuit fault (short-circuit point before the test point) as an example of a
detailed demonstration.
Get data: The collected data samples are pre-processed in 3.1 and divided into 30 groups of training samples
and 10 groups of test samples.
Train HMM fault diagnosis model: The 30 sets of training data are input into HMM model as observation
sequences, which constitute the outdoor short circuit detection classification (short-circuit point before the test
point). Set the maximum iteration step to 30 steps and convergence error to. From the following diagram of
HMM model training curve we can see that the observed values of HMM model gradually converge after
training with the increase of the number of iterations. When the number of iterations reaches 14, the model
achieves the convergence requirement. In the actual training, the six models satisfy the convergence condition in
the 20th iteration, and it can also be seen that the HMM model has the advantage of rapid training.
Figure 4 The HMM training curve of outdoor short circuit
The implicit state transition probability matrix of outdoor short circuit (short-circuit point before the test point)
is
Thus, the fault classification of outdoor short circuit (short-circuit point before the test point) is completed.
Test fault classification: The 10 sets of test data of outdoor short circuit (short-circuit point before the test point)
are input into six fault classification respectively in order to test the effect of the fault classifier. The figure
below shows the output probability. It is clear that when verify the data into their own fault classification
(outdoor short circuit, short-circuit point before the test point), the maximum output probability is 50%, the
minimum output probability is 12.5%. However, when verify these ten data into other five fault classifications;
the output probability is close to 0%. The results of the other five faults are basically the same as those of the
outdoor short circuit (short-circuit point at the test point).. The statistical results are shown in the figure below.
Figure 5 The output probability of the fault classification
Table 2 Results of fault state classification based on HMM
Types of
fault
Indoor
open
circuit
Indoor
short
circuit
Outdoor
short circuit
(before)
Outdoor
short
circuit
(after)
Outdoor
open
circuit
(before)
Outdoor
open circuit
(after)
recogniti
on rates
Indoor
open
circuit
10 100%
Indoor
short
circuit
10 100%
Outdoor
short
circuit
(before)
10 100%
Outdoor
short
circuit
(after)
10 100%
Outdoor
open
circuit
(before)
10 100%
Outdoor
open
circuit
(after)
10 100%
Thus, methods of track circuit fault diagnosis based on HMM are completed.
Model comparison and expansion
First, compare HMM fault diagnosis model with BP neural network fault diagnosis model. When 120
sets of sample data are brought into the BP neural network, it can be seen that although the successful
recognition rate is consistent with the HMM model, the BP model repeats 3 iterations (360 calculations) to meet
the convergence conditions, far more than 20 calculation of HMM model. As a result, there are obvious
advantages for the model of track circuit fault diagnosis based on HMM in computing speed.

Second, it is observed that all fault recognition rates are 100% in the matter of practical application in
hump section. The reason for this result may be due to the quite different voltage and other related data. So there
is no cross-overlapping area between fault data. Then, adjust the sample to a section of ZPW-2000A non-
insulated track circuit to the actual data. The faults are divided into the main track fault, small track fault,
common transmission channel fault, indoor fault and attenuation box fault. The input voltage of the main rail,
the input voltage of the small rail, the rail voltage, the measured voltage of the attenuation box and the voltage
of the simulation disk are used as fault characterization data, and then normalize these data and bring them into
the HMM fault diagnosis system. Take small track fault as an example and the result is shown below. It is
obvious that HMM fault detection classification output still has a clear degree of recognition of output fault
probability.
Figure 6 Output fault probability of small track fault classification
However, when the sample data is brought into the BP neural network model for diagnosis, the obvious
diagnostic results are appeared. First, bring 110 sample data into the model. It can be seen that the model repeats
4 iterations to meet the convergence conditions, which is, a total of 440 calculations. Nevertheless, HMM
model’s iterations still kept within 20 times. It is obvious that HMM model has distinct advantages in terms of
computing speed. What’s more, there are some problems of the fault diagnosis by BP neural network. Some of
the errors are shown in the following table.
Table 3 Verification sample classification T/F of BP neural network
Verificat
ion
sample
number
The verification sample classification result of BP neural network Verification
sample
classificatio
n results
Verific
ation
sample
Source
Clas
sific
ation
T/F
Main
track
fault
Small track
fault
Common
transmissi
on channel
fault
Indoor
fault
Attenuation
box
fault
1 0.86843 0.44578 0.76396 0.87501 0.51145 F1 F1 √
2 0.91152 0.46935 0.7845 0.86147 0.52032 F1 F1 √
3 0.62587 0.86775 0.70534 0.84636 0.53875 F2 F2 √
4 0.68947 0.82456 0.67742 0.87803 0.5977 F3 F2 ×
5 0.86726 0.88378 0.78853 0.78732 0.58523 F3 F3 √
6 0.86575 0.88524 0.79249 0.78651 0.58479 F3 F3 √
7 0.69322 0.46722 0.63455 0.84035 0.53244 F3 F4 ×
8 0.57778 0.41799 0.64325 0.72455 0.69454 F4 F4 √
9 0.59462 0.41234 0.1349 0.84532 0.52378 F5 F5 √
10 0.59342 0.40154 0.2589 0.8923 0.58923 F5 F5 √
Compare the result of HMM fault diagnosis method with the result of BP fault diagnosis. And the result is
shown in the table below.
Table 4 The correct rate comparison of two fault diagnosis methods
Method
Main track
fault
Small track
fault
Common transmission channel fault
indoor
fault
Attenuation box
fault
BP 95% 80% 90% 85% 80%
HMM 100% 100% 100% 100% 100%
In summary, high accuracy and faster computing speed can be provided by methods of track circuit fault
diagnosis based on HMM. It’s a good choice to introduce HMM algorithm and HMM optimization algorithm
into track circuit fault diagnosis field. And it’s clear that this fault diagnosis will become a new direction for the
development of track circuit fault diagnosis method.
ACKNOWLEDGEMENTS
This project is sponsored by Shanghai University of Engineering Science Innovation (NO.E3-0800-16-02131).
REFERENCES
[1]. Bilmes,J.A. (2006).What HMMs can do.IEICE TRANSACTIONS on Information and Systems,
89(3):1－24.
[2]. GUI Lin, WU Xiaoyue. (2008).Discrete hidden semi-Markov model based on prognosis mode.
Application Research of Computer, 25(11):3320-3327.
[3]. HUANG Zanwu, WEI Xueye, LIU Ze.(2012). Fault Diagnosis of Railway Track Circuits Using Fuzzy
Neural Network. Journal Of The China Railway Society,34(11):52-75.

[4]. Leveson,S.E. Rabiner, L .(1983).An introduction to the application of the theory of probabilistic
functions of Markov process to automatic speech recognition.Bell System Journal , (4):1035 -1047.
[5]. LI Quan. (2010).Research on Fault Diagnosis of Rolling Bearing Based on HMM.(Master dissertation).
Kunming University of Science and Technology.
[6]. LI Zhihua, ZHU Hui, PAN Yue. (2014).The Application of Genetic Algorithm Particle Swarm
Optimization Algorithm Optimize the BPNN in Analog Circuit Fault Diagnosis. Control Engineering
of China, 21(3):450-454.
[7]. LUO Xiliang, YANG Shixi, CAI Qinghua. (2006).Design and Realization of HMM-based Power
Transformer Fault Diagnosis System Software. Industrial Control Computer, 19(14),36-39.
[8]. Rabiner,L.(1989).A tutorial on hidden Markov models and selected applications in speech
recognition.Proc.IEEE (2):257-286.
[9]. TAN Xiaodong. (2008).Fault Prognostic Technology of Rolling Element Bearings Based on
HSMM.(Master dissertation). National University of Defense Technology.
[10]. TANG Lin. (2014).Multi-sensor fault detection and diagnosis mechanism based on Markov
prediction.Transducer and Microsystem Technologies, 33(11):52-75.
[11]. WANG Tong.(2014). Track Circuit Fault Diagnosis Method Based on Least Squares Support Vector
Machine.Railway Standard Design, 58(2):89-93.
[12]. WANG Xinmin.(2004). A Modified Training Algorithm of Hidden Markov Model. Journal Of
XiaoGan University, 24(3):74-77.
[13]. XIAO Wen-bin.(2011).Study on Coupled Hidden Markov Model Based Rolling Element Bearing Fault
Diagnosis and Performance Degradation Assessment. (Master dissertation). Shanghai Jiao Tong
University.
[14]. YUE Xia.(2012). HMM-Based Fault Diagnosis Technology under Complex Conditions. (Master
dissertation). South China University of Technology.
[15]. ZHAO Linhai,RAN Yikui,MU Jiancheng.(2010).A Comprehensive Fault Algorithm.China Railway
Science, 31(3):107-113.)
[16]. ZHU Ming, GUO Chunsheng. (2010).Hidden Markov Model and Its latest Application and Progress.
Computer Systems & Applications, 19(7):255-216.

Methods of Track Circuit Fault Diagnosis Based on Hmm

Recommended

More Related Content

Similar to Methods of Track Circuit Fault Diagnosis Based on Hmm (20)

More from IJRESJOURNAL (20)

Recently uploaded (20)

Methods of Track Circuit Fault Diagnosis Based on Hmm