Received November 28, 2020, accepted December 20, 2020, date of publication December 22, 2020,

date of current version January 6, 2021.

Digital Object Identifier 10.1109/ACCESS.2020.3046681

Fault Diagnosis of Train Network Control

Management System Based on Dynamic Fault
Tree and Bayesian Network
CHONG WANG , LIDE WANG , HUANG CHEN , YUEYI YANG , AND YE LI
School of Electrical Engineering, Beijing Jiaotong University, Beijing 100044, China
Corresponding author: Chong Wang ([email protected])
This work was supported in part by the Fundamental Research Funds for the Central Universities under Grant 2018YSJ149, and in part by
the cooperation project of Beijing Jiaotong University and China Railway Rolling stock Corporation (CRRC) under Grant SRI2019070.

ABSTRACT Train network control management system (TCMS) is an important part of the High-speed
rail train. Because of the TCMS’s complex and redundant structure, long-term operation environment,
etc., breakdowns inevitably in the long-time running. Based on the historical fault data of the TCMS
accumulated during their online service, the working principles, failure modes, and effects analysis of TCMS
are researched and the dynamic fault tree (DFT) model of TCMS failure is built. Then, the dynamic fault
tree model is transformed into the Bayesian network (BN) model, which can model the reliability of such
types of systems. Finally, combining DFT with BN is used for fault probability estimation and reliability
assessment. The results present that increasing the reliability of key modules for the TCMS would be of
great help to High-speed rail train engineers in the fault diagnosis field.

INDEX TERMS Train network control management system, dynamic Fault tree analysis, Bayesian network,
fault diagnosis.

I. INTRODUCTION modes of TCMS are particularly complex [11]. For instance,

With the fast development of China’s Railway, the High- there are a lot of complex failure modes for the TCMS,
speed rail train, which is a large-scale intelligent system with which are related to training operation, mechanical structures,
sophisticated structures, has become a significant mode of electrical facilities, network control, and many other aspects.
transportation in China. However, the High-speed rail train The operation of the TCMS on a High-speed rail train is
breaks down inevitably in the long-time running. Fault diag- dominated by on-board equipment and microprocessors.
nosis will play a very significant role in the safety and effi- The TCMS involves mechanical structures and electrical
cient operations. One crucial technical system of high-speed equipment such as CCU, TCN Gateway, WTD, REP, RIOM,
rail train is the Train network Control Management System, and HMI. The complex failure modes of the TCMS consists
which is the core of ensuring safe and efficient train oper- of power supply anomaly, equipment anomaly, circuit board
ation. The failure of TCMS will seriously threaten the safe failure, communication anomaly, and system failure. The
running of the High-speed rail train. Therefore, the research relationship between different failures is complicated and
of TCMS fault diagnosis has a great significance to the further uncertain.
development of the High-speed railway. Hence, the TCMS of High-speed rail trains is regarded as
Complexity is defined in two forms: structural complexity a complex system, which is designed to achieve high levels
and fault complexity. The High-speed rail train system is a of reliability and redundancy. Failure of any such subsystems
complex system with redundancy, both in its structure and will have a heavy impact on the service itself, resulting in
in its failure modes. Although the structure of the TCMS obvious deterioration of performance, reduction of perceived
seems relatively simple, the working principles and failure quality, and increment of costs. A beneficial approach com-
bining dynamic fault tree analysis and Bayesian network
The associate editor coordinating the review of this manuscript and is applied to resolving the TCMS’s fault diagnosis with
approving it for publication was Baoping Cai . uncertainty reasoning, solving complex and polymorphism

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.

2618 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 9, 2021
C. Wang et al.: Fault Diagnosis of TCMS Based on DFT and BN

problems, which will be appropriate for the fault diagnosis showed that the estimation methodology could calculate the
and maintenance decision for High-speed Railways. RUL of structure systems with multiple influencing causes.
The traditional fault tree analysis is a graphical method that Yang et al. [18] have presented the Bayesian network-based
models how failures propagate through the system and how software reliability modeling method and task flow-oriented
component failures lead to system failure. It provides meth- software reliability simulation prediction method, which can
ods and tools to compute a wide range of properties and mea- utilize the prior information of software architecture, his-
sures [12]. FTA technique is widely used for both qualitative tory data, and software task flows to conduct the dynamic
and quantitative assessment. However, with the development reliability prediction and find the reliability weaknesses.
of modern engineering, the equipment is becoming more He et al. [19] have applied the dynamic fuzzy subset theory
and more complex. Systems designed to achieve high levels into the Bayesian network, in which the fault probability of
of reliability frequently employ high levels of redundancy, the leaf node fault state and fuzzy dynamic fault probability
dynamic redundancy management, and complex fault & error was developed and calculated. Halabi et al. [22] have applied
recovery techniques, Dugan et al. [6] presented dynamic the Bayesian network and multivariate logistic regression
fault tree modeling techniques for handling these difficulties. to model the relationship between sources of information.
This method solves the reliability analysis of three advanced Boudali and Dugan [23] have presented a novel reliability
fault-tolerant computer systems. modeling and analysis framework based on the Bayesian
Volk et al. [1] have presented a DFT modeling approach, network formalism.
which can serve to provide additional insights into the criti- However, there is another problem worth considering: it’s
cality of field elements, for infrastructure reliability analysis not easy to build the Bayesian network model for complex
of railway station areas. Kabir et al. [2] have proposed a engineering systems. The surprise is that transforming the
hybrid modularization scheme where independent sub-trees dynamic fault tree model into the Bayesian model will resolve
of a DFT are identified and quantified in a hierarchical the problem well.
order. This method could provide an approximate solution to Hamza and Hacene [13] have discussed the advantages of
DFTs without unacceptable loss of accuracy. Volk et al. [4] the Bayesian network over the fault tree in reliability and
have presented a new state-space generation approach for safety analysis. Zheng et al. [16] have presented the transfor-
dynamic fault trees (DFTs) that exploits several successful mation algorithm from the fault tree to the Bayesian network,
reduction techniques from the field of model checking. Ghad- aiming to develop a bridge crane spreader fault diagnosis sys-
hab et al. [5] have proposed the approach of constructing tem, which can effectively use historical fault data to support
dynamic fault trees, which can be used to evaluate various subsequent maintenance. Bobbio et al. [25] have introduced
quantitative measures by model checking, to model a variety the basic inference techniques of FT mapped into a BN, which
of safety concepts and E/E architectures for drive automation. can obtain some additional power both at the modeling and at
The DFT model can reveal clear failure mechanisms and the analysis level. Kabir et al. [26] have applied the fault tree
clear logical relationships between failures, however, as to analysis and the dynamic Bayesian network to assess the reli-
the complex and redundancy system, they cannot compute ability of flare systems. Khakzad et al. [27] have presented
the importance measure of basic event and have bidirectional the application of BNs in safety analysis and introduced the
inference. The Bayesian network takes advantage of proba- similarity between FTs and BNs. Rahman et al. [28] have
bility theory and graph theory reasoning uncertainty relation built up the fault tree model of marine logistics, then a
of events, which is more clearly on the logic. Moreover, Bayesian network approach is used to develop the risk model
the Bayesian network has bidirectional inference and update considering interdependencies and conditional relationships
related information according to the circumstance. From this among the contributing factors. Zhang et al. [29] have intro-
point of view, the Bayesian network is better than DFTA. duced a fuzzy AHP index system of respective expert judg-
Cai et al. [7] have proposed a novel real-time reliability ment ability and built up an event tree, then quantitative risk
evaluation methodology by combing the root cause diagnosis reasoning and sensitivity analysis of foundation pit collapse is
phase based on Bayesian networks, which can calculate achieved by using fuzzy Bayesian inference. Feng et al. [30]
the real-time reliability of the entire system by forwarding have established a fault tree model of gas pipelines and
inference of BNs. Cai et al. [8] have proposed a dynamic mapped this model to the Bayesian network. Qiao et al. [31]
Bayesian network-based fault diagnosis methodology, which have presented an event tree of the analytic hierarchical pro-
can identify the faulty components and distinguish the fault cess for the evaluation of expert capability, and then built up
types, in the presence of TF and IF for electronic systems. a Bayesian network model of maritime accident scenarios.
Cai et al. [9] have proposed using object-oriented Bayesian Xiao et al. [32] have transformed a seaplane risk evaluation
networks to reduce the overall complexities of BNs for indicator system into an event tree, and then proposed a novel
diagnosis and the reporting of faults that immediately occur. approach to modeling the risk of seaplane operation safety
Cai et al. [10] have presented a hybrid physics-model-based using a Bayesian network.
and data-driven remaining useful life estimation method- The rest of the paper is organized as follows: Section II
ology of structure systems considering the influence of describes the analysis of the Dynamic fault tree method.
multiple causes by using dynamic Bayesian. A case study An overview of Bayesian network analysis is also provided

VOLUME 9, 2021 2619

 C. Wang et al.: Fault Diagnosis of TCMS Based on DFT and BN

in this section. The structural analysis of TCMS and DFT

modeling of TCMS are described in section III. The descrip-
tion includes the structural analysis and working principle
of TCMS, failure mode, and effects analysis of TCMS,
and the dynamic fault tree modeling of TCMS. Section IV
shows the fault diagnosis progress of TCMS. The process
includes the method of mapping DFT into BN, the method
verification and case study, the Bayesian network model-
FIGURE 2. AND gate.
ing of TCMS, the results of experimental analysis, and the
comparison of other methods. Finally, Section V presents the
In the dynamic fault tree analysis method, Priority-AND
concluding remarks.
(PAND) gate, and Functional Dependency (FDEP) gate are
II. METHODOLOGIES USED used to describe the timing of part failure. Spare gate (SP)
The TCMS consists of many parts and complex structure, is mainly used for reliability modeling and evaluation when
moreover, the logical relationship, and the operating environ- some equipment has spare parts in some complex systems [6].
ment is not clear. The lack of effective data in the field data is
3) PAND GATE
also a problem to be considered in TCMS’s fault diagnosis.
Meanwhile, to deal with the limitations of redundancy anal- The input events of the PAND gate can be either the basic
ysis, the static fault tree is extended to a dynamic fault tree event or the output event of another logic gate. The failure
by introducing a dynamic logic gate [6]. Dynamic fault tree mechanism is as follows: when the input events of PAND
refers to the fault tree that contains at least one dynamic logic break down from left to right and the output events of the
gate so that it has can model and analyze the dynamic system. logic gate will fault occurs.
Bayesian network is suitable for handling multi-state Assume PAND has three input events, and its graphical
description of events. By setting different values of node notation is shown in Fig. 3. When B fails before A, C fails
variables to represent different states of the subsystem, and before A, or B and C fail simultaneously before a fails. Then
by setting the CPT of corresponding nodes to conveniently the output event of the logic gate will break down.
express the logical relationship between variables, the char-
acteristics of multi-state systems can be clearly described.
The combination of Dynamic fault tree analysis and the
Bayesian network method to achieve the reasonable expres-
sion of multi-source information will provide a new idea for
fault uncertainty processing of TCMS.

A. DYNAMIC FAULT TREE ANALYSIS

In the static fault tree analysis method [16], OR gate and AND
FIGURE 3. Priority-AND gate.
gate are the two most commonly used forms. The graphical
notations are shown in Fig. 1 and Fig. 2.
4) FDEP GATE
FDEP contains one triggering event and the related events
can be one or more. The failure mechanism is that related
events fail individually or triggering events failure leading to
all related events fail, then the output events of the logic gate
will fault occurs.
Assume FDEP has two related events A and B, and one
triggering event Tr. Then its graphical notation is shown
FIGURE 1. OR gate. in Fig. 4. When A and B fail, or Tr failure leading to A and B
fails, the output event of the logic gate will break down.
1) OR GATE
When at least one of the input events corresponding to the OR 5) SP GATE
gate is in a failure state, the output event of the logical gate The SP Gate can be divided into Cold spare (CSP), Warm
will fault occur. spare (WSP), and Hot spare (HSP). The graphical notations
are shown in Fig. 5. Besides, the ratio of failure rate during
2) AND GATE the backup period to failure rate during the operation period is
When all input events corresponding to AND gate are in a called Dormancy Factor β, and its graphical notation is shown
failure state, the output event of the logical gate will fault in Figure 4. Among them, the dormancy factors of CSP, WSP,
occur. and HSP are β = 0, 0 < β < 1, and β = 1 respectively.

2620 VOLUME 9, 2021

C. Wang et al.: Fault Diagnosis of TCMS Based on DFT and BN

M taking advantage of Bayes theorem:

P(X , M ) P(X , M )
P(X /M ) = =P (2)
P(M ) X P(X , M )
A basic Bayesian network is shown in Fig. 6.

FIGURE 4. Functional Dependency gate.

FIGURE 6. Basic Bayesian network.

FIGURE 5. Spare gate.

III. STRUCTURAL ANALYSIS AND DYNAMIC FAULT TREE
MODELING OF TCMS
Since all the devices in the TCMS are in the state of hot
In this section, we are going to introduce the TCMS system
spare redundancy, we will mainly use the HSP as the dynamic
that belongs to the 200km/h intercity EMU train KDZ15,
fault tree logic gate.
which is independently developed by CRRC Corp.
In the HSP, both the host and back-up parts enter the
working state at the same time. When the host part fails,
A. STRUCTURAL ANALYSIS AND WORKING PRINCIPLE
any back-up part can be regarded as the host part to ensure
OF TCMS
the normal working state of the whole system. Therefore,
the output event breaks down only when both the host and The 200km/h intercity EMU is a fixed grouping of 8 vehicles,
back-up components fail. in which a traction unit is defined for every 4 vehicles. And
the names of train marshaling are Mc1-Tp2-M3-T4-Tb5-M6-
B. BAYESIAN NETWORK
Tp7-Mc8, Where M represents the bullet train, Mc represents
the bullet train with a cab, which is located at the head of
BN is an organic set of Bayesian methods and graph the-
the train. T represents the trailer, and Tp represents the trailer
ory. It is a directed acyclic graph (DAG), which contains a
with a pantograph.
probability table and is composed of network structure and
Among them, TCMS is responsible for the transmission
parameters. The network structure of BN includes two parts:
of instructions and braking control of EMUs and monitoring
node-set and directed edge set. Each node can be regarded as a
the main equipment’s status on the train. What’s more, TCMS
variable with discrete or continuous values, and each directed
can be capable of fault diagnosis and fault recording function.
edge represents the dependency between nodes. BN parame-
The network topology of the KDZ15 train is shown in Fig. 7.
ter refers to a set of conditional probability distributions of the
By utilizing the train network, TCMS collects all kinds of
Bayesian network model, which represents the conditional
information related to the running condition of vehicles, such
probability distribution (CPD) of a node under a certain
as traction, braking, auxiliary power supply, air conditioning,
value state of its parent node. In the case of discrete values
and other subsystems, makes a comprehensive logical judg-
of network nodes, CPD can be expressed as a conditional
ment on these data, and then sends the results back to each
probability table (CPT).
subsystem after processing to realize management of each
Consider a two-state system with N components, each of
subsystem.
which is represented by Xi . Because the Bayesian network has
The composition of TCMS consists of network topology
a strong conditional independent relationship, when given a
lines and network devices, which are attached to High-speed
parent node of one node, it is conditional independent from
rail trains via the WTB line and the MVB line. What’s
all other nodes except its descendants.
more, the main components of TCMS include Central Control
According to the conditional independence and the chain
Unit (CCU), Input / Output Module in Cab (IOM Cab),
rule, BN represents the joint probability distribution P(X) of
Remote Input / Output Module (RIOM), TCN Gateway
variables of any BNs as [13]:
(GW), Repeater (REP), Human Machine Interface (HMI)
n
\ and Wireless Transmission Device (WTD). Besides, the main
P(X1 , X2 , · · · , Xn ) = P(Xi /parent(Xi )) (1) functions of each component are shown in Table 1.
i=1 At the beginning of TCMS’s design, the reliability factor
BN can update the prior probability of any events given has been taken into account. For instance, the WTB line and
new information (posterior probability), called evidence the MVB line are designed with redundancy. Furthermore,

VOLUME 9, 2021 2621

 C. Wang et al.: Fault Diagnosis of TCMS Based on DFT and BN

FIGURE 7. Train network topology.

hot standby redundant design has also been applied to the 1) FAILURE MODE 1:
important network equipment, such as CCU, TCN GW, IOM CCU failure, resulting in the CCU communication anomaly.
CAB, RIOM, HMI, WTD, and REP. When they work prop- Probable causes:
erly, one is defined as the master device and the other as the (a) MVB equipment board card failure.
slave device. The master and slave equipment will monitor (b) Equipment power supply anomaly.
the working state of each other in real-time. If the master (c) CPU card fault.
equipment fails, the slave equipment will take over its work (d) CCU system failure.
automatically. After the switch is completed, the system can
continue to work. 2) FAILURE MODE 2
B. FAILURE MODE AND EFFECTS ANALYSIS (FMEA) TCN GW failure, resulting in the TCN Gateway communica-
OF TCMS tion anomaly.
If the TCMS is in normal working condition, each component Probable causes:
must be in normal condition. While the technology of train (a) MVB equipment board card failure.
control is so advanced, the TCMS’s failure is inevitable. (b) Equipment power supply anomaly.
Hence, it is necessary to make research on fault diagnosis and (c) CPU card fault.
reliability analysis of TCMS. (d) CCU system failure.
The TCMS of EMU train is a typical and complex system,
its normal operation is done by the cooperation of hundreds 3) FAILURE MODE 3
of board cards. When the onboard equipment and its board IOM CAB failure, resulting in the IOM CAB communication
cards in the system break down, it will lead to a fault alarm anomaly.
of TCMS. Probable causes:
Through the way of classifying and researching fault track- (a) MVB equipment board card failure.
ing records, the most frequent TCMS failure modes can be (b) Equipment power supply anomaly.
summarized as follows. (c) CPU card fault.

2622 VOLUME 9, 2021

C. Wang et al.: Fault Diagnosis of TCMS Based on DFT and BN

TABLE 1. Main functions of the TCMS component. Probable causes:

(a) Equipment power supply anomaly.
(b) Hardware failure of Repeater.

C. DYNAMIC FAULT TREE MODELING OF TCMS

The core function of TCMS is to be able to run normally with-
out failure. Moreover, TCMS adopts a modular structure, and
the component units are designed with a redundant structure.
For instance, the CCU, TCN GW, IOM CAB, RIOM, HMI,
WTD, and REP have applied for the hot spare redundancy.
It is assumed that the failures of the components other than
those with dynamic failure mechanisms are independent of
each other, and the life distribution of the comment units
conforms to an exponential law.
Based on the classification of the High-speed emu system
and the cause of failure of all modules, a dynamic fault
tree was established by taking the TCMS system fault of
KDZ15 emu, as shown in Fig. 8.
The dynamic fault tree model is composed of two logic
gates, one is the ‘‘OR’’ gate, and the other is the ‘‘HSP’’ gate.
The symbol ‘‘T’’ is on behalf of the top event, which means
‘‘TCMS fails to work’’.
The symbol ‘‘M∗ ’’ is on behalf of the intermediate events,
which consist of M1: CCU communication anomaly, M2:
TCN GW communication anomaly, M3: IOM CAB commu-
nication anomaly, M4: RIOM communication anomaly, M5:
HMI communication anomaly, M6: WTD communication
anomaly, M7: Repeater communication anomaly.
4) FAILURE MODE 4 The symbol ‘‘M∗∗ ’’ represents the redundant events, which
RIOM failure, resulting in the RIOM communication consist of M11: CCU host failure, M12: CCU back-up failure,
anomaly. M21: TCN GW host failure, M22: TCN GW back-up failure,
Probable causes: M31: IOM CAB host failure, M32: IOM CAB back-up fail-
(a) MVB equipment board card failure. ure, M41: RIOM host failure, M42: RIOM back-up failure,
(b) Equipment power supply anomaly. M51: HMI host failure, M52: HMI back-up failure, M61:
(c) CPU card fault. WTD host failure, M62: WTD back-up failure.
The symbol ‘‘X∗ ’’ is on behalf of the basic events, which
5) FAILURE MODE 5 are made up of X1: Power supply board card no output,
HMI failure, resulting in the HMI communication anomaly. X2: MVB board card communication anomaly, X3: CPU
Probable causes: board card failure, X4: DI acquisition boards anomaly, X5:
(a) MVB equipment board card failure. DO acquisition boards anomaly, X6: Backboard failure, X7:
(b) Equipment power supply anomaly. Power supply board card no output, X8: MVB board card
(c) CPU card fault. communication anomaly, X9: CPU board card failure, X10:
(d) CCU system failure. WTB boards hardware failure, X11: Backboard failure, X12:
Power supply board card failure, X13: CPU board card fail-
6) FAILURE MODE 6 ure, X14: IOM Cab Backboard failure, X15: DI acquisition
WTD failure, resulting in the wireless transmission devices card failure, X16: DO output card anomaly, X17: Power
communication anomaly. supply board card failure, X18: CPU board card failure, X19:
Probable causes: RIOM Backboard failure, X20: DI acquisition card failure,
(a) MVB equipment board card failure. X21: DO output card anomaly, X22: Power supply anomaly,
(b) Equipment power supply anomaly. X23: MVB board card communication anomaly, X24: CPU
(c) CPU card fault. board card failure, X25: LCD failure, X26: Power supply
(d) CCU system failure. anomaly, X27: MVB board card communication failure, X28:
CPU board card failure, X29: Communication interface card
7) FAILURE MODE 7 failure, X30: Backboard failure,X31: Communication func-
REP failure, resulting in the Repeater communication tion board card failure, X32: REP host failure, X33: REP
anomaly. back-up failure.

VOLUME 9, 2021 2623

 C. Wang et al.: Fault Diagnosis of TCMS Based on DFT and BN

FIGURE 8. DFTA modeling of the TCMS.

By analyzing the structure-function and failure process TABLE 2. Reliability parameters of basic events.
of the object in detail, the dynamic fault tree model of the
system is established according to the basic flow of fault tree
modeling. This step is the foundation of the entire modeling
and analysis process. It is necessary to accurately and rea-
sonably select the top event and determine the failure logic
relationship between various parts of the system, to ensure
the correctness of the final analysis results. According to the
established dynamic fault tree model, the events in the model
are mapped to the nodes of the Bayesian network model layer
by layer.
The components’ hidden trouble incidence calculated by
field historical data and the reliability parameters obtained by
expert evaluation. Furthermore, we have a deep cooperative
relationship with CRRC Corp thus their experts provide us
with the practical FMECA data, which can be applied in the
field of fault diagnosis and reliability analysis. As is shown
in Table 2.

IV. FAULT DIAGNOSIS OF TCMS

A. THE METHOD OF MAPPING DFT INTO BN
Before modeling, not only the scope and boundary of the
model need to be determined, but also the important variables
and states need to be defined. Dynamic fault tree makes
use of the logic gate to express the relationship between the
event and the Bayesian network is used to connect the edge
nodes. A Bayesian network’s parents’ nodes represent the
input events of the fault tree and the child nodes represent
the output.
Furthermore, the transformation of dynamic fault tree and corresponding relationships between them and the simplified
Bayesian network model consists of two parts, one is the processes are as shown in Fig. 9.
transformation of structure between DFT and BN, the other Note that the dynamic fault tree shown in Figure 8 has
is the establishment of CPT for non-root nodes. Concrete applied static logic gate and HSP logic gate, hence the

2624 VOLUME 9, 2021

C. Wang et al.: Fault Diagnosis of TCMS Based on DFT and BN

FIGURE 11. Transformation of HSP logic gate.

FIGURE 9. Simplified process of Bayesian network based on dynamic

fault tree [26].

FIGURE 12. Case model of dynamic fault tree.

2) TRANSFORMATION OF HSP LOGIC GATE

The transformation of structure for the HSP logic gate is
shown in the Fig. 11.
Make M = [A1 , A2 , . . . , Ag ], g is the number of input
events for HSP logic gate, Ai (i = 1, 2, . . . , g) is the state
variable of the input event. Make N is the set of state variables
for the output event. There is the conditional probability
FIGURE 10. Transformation of static gate. distribution of N:
(
1 k=j
P(N = k|M ) = (5)
transformation of structure and the establishment of CPT on 0 k 6= j
these two logic gates are presented respectively below.
where j = max (A1 , A2 , . . . , Ag ).
1) TRANSFORMATION OF STATIC GATE
B. METHOD VERIFICATION AND CASE ANALYSIS
The transformation of structure for OR gate and AND gate is
shown in the Fig. 10. Taking the dynamic fault tree shown in Fig. 12 as an example.
Make X1 = [A1 , A2 , . . . , Am ], m is the number of input To verify the accuracy of the above method, which is com-
events for OR gate, Ai (i = 1, 2, . . . , m) is the state variable pared with the computation results of the Binary Decision
of the input event. Make Y1 is the set of state variables for the Diagrams (BDD) method [33].
output event. There is the conditional probability distribution The dynamic fault tree model is composed of a static logic
of Y1: gate and dynamic logic gate, which are the OR gate and the
( HSP gate respectively.
1 k=j Firstly, the topology of the dynamic fault tree model is
P(Y 1 = k|X 1) = (3)
0 k 6= j transformed into the structure of the Bayesian network model.
Then, according to the construction method of CPT according
where j = min (A1 , A2 , . . . , Am ). to Formula (1) – (3). The input of CPT and the construction
Make X2 = [A1 , A2 , . . . , An ], n is the number of input of the BN model were carried out in GeNIe software. Fur-
events for AND gate, Ai (i = 1, 2, . . . , n) is the state variable thermore, the fault diagnosis of the BN model and reliability
of the input event. Make Y2 is the set of state variables for the analysis was accomplished by GeNIe software.
output event. There is the conditional probability distribution The transformed Bayesian network model is shown
of Y2: in Fig. 13. Moreover, the corresponding relationship between
(
1 k=j the nodes of each layer and the events in Fig. 12 is described
P(Y 2 = k|X 2) = (4) as follows:
0 k 6= j
i) The First layer: Node 1, 2, 3, 4, and 5 corresponds to the
where j = max (A1 , A2 , . . . , An ). basic event X2, X3, X4, X5, and X1.

VOLUME 9, 2021 2625

 C. Wang et al.: Fault Diagnosis of TCMS Based on DFT and BN

TABLE 4. Node definition.

FIGURE 13. Case model of Bayesian network structure.

TABLE 3. Comparison of BN and BDD simulation results.

TABLE 5. The CPT of node M11.

ii) The Second layer: Node 6 and 7 corresponds to the

intermediate events M1 and M2.
iii) The Third layer: Node 8 corresponds to the top event T.
Assuming that the failure rate of each basic event in the
dynamic fault tree respectively are P1(X1 = 1, X2 = 1,
X3 = 1, X4 = 1, X5 = 1) = (0, 0, 0, 0, 0), P2(X1 = 1,
X2 = 1, X3 = 1, X4 = 1, X5 = 1) = (0.1, 0.2, 0.3, 0.4,
0.5), P3(X1 = 1, X2 = 1, X3 = 1, X4 = 1, X5 = 1) = (0.9,
0.8, 0.7, 0.6, 0.5), P4(X1 = 1, X2 = 1, X3 = 1, X4 = 1,
X5 = 1)= (0.1, 0.7, 0.5, 0.8, 0.2), P5(X1 = 1, X2 = 1, X3 = computation results based on the Binary Decision Diagrams
1, X4 = 1, X5 = 1) = (0.6, 0.3, 0.7, 0.2, 0.9), P6(X1 = 1, method [34], as shown in Table 3.
X2 = 1, X3 = 1, X4 = 1, X5 = 1) = (0.5, 0.5, 0.5, 0.5, 0.5), It can be seen from the Table 3 that the results calcu-
P7(X1 = 1, X2 = 1, X3 = 1, X4 = 1, X5 = 1) = (1, 1, 1, lated by the two methods are extremely close, which verifies
1, 1), among them the number ‘‘1’’ in parenthesis mark after the effectiveness of the Bayesian network for the reliability
the letter ‘‘P∗ ’’ represents the corresponding event has failed. assessment of the system.
And the numbers in parenthesis mark after the equal sign
represents the failure rate of each basic event. According to C. THE BAYESIAN NETWORK MODEL OF TCMS
the hypothesis, the fault diagnosis simulation of the Bayesian The Bayesian network model has the function of bidirectional
network model is carried out, which is compared to the inference, which can express the association relation between

2626 VOLUME 9, 2021

C. Wang et al.: Fault Diagnosis of TCMS Based on DFT and BN

TABLE 6. The CPT of node M1. TABLE 8. Posterior probability.

FIGURE 14. Bayesian network model of TCMS.

TABLE 7. The CPT of node T.

The meanings of the nodes represented are shown

in Fig. 14.
Due to the limited space, the CPT of leaf node T and inter-
mediate node M1 and M11 were given in Table 5, 6, and 7.
Moreover, the number ‘‘0’’ represents the normal state and
the number ‘‘1’’ represents the failure state.

D. RESULTS OF THE EXPERIMENTAL ANALYSIS

We utilize the simulation software named GeNIe version
2.0 for the inference usage of the developed BN models.
Cheng et al. [17] modeled the dynamic Bayesian network,
which is based on the GeNIe software, to locate the fault
of the train control system and deduce the system main-
tenance strategies. Yang et al. [18] modeled the integrated
and sub-Bayesian network of TCMS, which is based on the
GeNIe software, to utilize the prior information and find the
reliability weaknesses. Halabi et al. [22] presented the BN of
reliability and risk model, which is modeled in the GeNIe
version 2.0. Kabir et al. [26] built the DBN of the FT model
for the flare system, which is based on GeNIe software.
elements more clearly and intuitively, providing a solid foun- Rahman et al. [28] built the modified BN model for marine
dation for security modeling. offshore logistics operation, which is based on GeNIe soft-
According to the corresponding relationship and trans- ware, to address the interdependencies and conditional rela-
formation rules, the dynamic fault tree model in Fig. 8 is tionships among the critical factors. Feng et al. [30] modeled
transformed into the Bayesian network model in Fig. 14. the Bayesian network of pipe failure and risk diagnosis-based

VOLUME 9, 2021 2627

 C. Wang et al.: Fault Diagnosis of TCMS Based on DFT and BN

FIGURE 15. Simulation diagram of Bayesian network.

FIGURE 16. Bayesian network based on the prior failure rate.

BN of Noisy-OR gates, which is conducted on GeNIe exponential distribution. The failure efficiency of basic events
software. is shown in Table 2. Furthermore, mapping each level of
It is assumed that the states of parts and systems are only events and failure logic relationships of the dynamic fault tree
normal and failure and the failure of parts is subject to the into each level of nodes of the Bayesian network, as can be

2628 VOLUME 9, 2021

C. Wang et al.: Fault Diagnosis of TCMS Based on DFT and BN

FIGURE 17. The posterior probability of T node failure.

seen in Fig. 15, the BN model has 1 leaf node, 19 intermediate

nodes, and 33 root nodes.
After the BN model of TCMS is built, the GeNIe is used to
conduct simulation analysis on it, which is shown in Fig. 16.
When we insert the inputs of each root node into the BN
model, we can get the occurrence probability of intermediate
nodes and leaf node T. The result shows that the probability
of the top event is 2.2214327e−11 .
When the TCMS has fault occurs, the state of T is set
to ‘‘1’’ in GeNIe. Therefore, the failure probability of each
basic event is computed by using the reverse reasoning of the
Bayesian network, as is shown in Table 8 and Fig. 17.
The variation trend of the posterior probability of each
basic and intermediate event was illustrated in Fig. 18 and
Fig. 19. The Fig. 20 shows the comparison of prior proba- FIGURE 18. The posterior probability of the basic event.
bility and posterior probability changes, which indicates that
although the X23 has the maximum original failure rate when The higher the module reliability is, the higher the TCMS
the top event is in the failure state, the X27 and X30 have the reliability is, which makes sense. We can conclude that to
maximum occurrence probability instead. ensure the reliability of TCMS, all events’ reliability should
Through data analysis, we can find that X16 (Status of be improved and the reliability of key modules such as WTD
IOM CAB’s DO card) has the minimum probability of fail- has a greater impact on the system reliability should be
ure. Then, X27 (Status of WTD’s MVB board card), X30 ensured at a high level.
(Status of WTD’s CPU board card), and M6 (Status of WTD
communication) have the maximum probability of failure, E. COMPARISON OF METHODS
the corresponding parts and components are the weak links The experimental data in this paper, which has the features
of the TCMS reliability service. Therefore, the stability of of long duration and incompleteness, is based on the field
WTD equipment is more sensitive to failure and system operation data and historical statistics of the High-speed rail
reliability. train equipment. Moreover, the fault diagnosis and reliability

VOLUME 9, 2021 2629

 C. Wang et al.: Fault Diagnosis of TCMS Based on DFT and BN

TABLE 9. Comparison between DFT-BN and other methods [16].

FIGURE 19. The posterior probability of the mediate event.

FIGURE 20. The comparison of prior probability and posterior probability
changes.
analysis of the TCMS demands a precise inference approach,
which can make full use of the prior information.
The comparison among diverse methods is shown fit for dealing with the uncertainty problem according to the
in Table 9. And the presented method has combined the prior information, but also diagnosing faults and analyzing
advantages of DFT and BN. The DFT model can be easily reliabilities of the complex systems. In general, this method
built from the FMEA analysis and previous data, and the DAG has the best capability in this real instance compared to other
and CPT of BN can be acquired. Furthermore, BN is not only methods.

2630 VOLUME 9, 2021

C. Wang et al.: Fault Diagnosis of TCMS Based on DFT and BN

V. CONCLUSION [11] L.-L. Song, T.-Y. Wang, X.-W. Song, L. Xu, and D.-G. Song, ‘‘Research
In this paper, we have addressed the limitations of DFT and application of FTA and Petri nets in fault diagnosis in the pantograph-
type current collector on CRH EMU trains,’’ Math. Problems Eng.,
and BN among their applications in Introduction section. vol. 2015, pp. 1–12, Oct. 2015, doi: 10.1155/2015/169731.
Then we analyzed and summarized the critical basic events [12] E. Ruijters and M. Stoelinga, ‘‘Fault tree analysis: A survey of the state-
which contributed to the TCMS failure condition. Moreover, of-the-art in modeling, analysis and tools,’’ Comput. Sci. Rev., vols. 15–16,
pp. 29–62, Feb. 2015, doi: 10.1016/j.cosrev.2015.03.001.
we systematically discussed the working principles and fail- [13] Z. Hamza and S. Hacene, ‘‘Reliability and safety analysis using fault tree
ure modes of TCMS. The dynamic fault tree was established and Bayesian networks,’’ Int. J. Comput. Aided Eng. Technol., vol. 11,
by analyzing the logical relationship between the redundancy no. 1, pp. 73–86, 2019.
[14] R. S. Chanda and P. K. Bhattacharjee, ‘‘A reliability approach to transmis-
systems. Then, the dynamic fault tree model was transformed sion expansion planning using fuzzy fault-tree model,’’ Electr. Power Syst.
into the Bayesian network model. Hence, we explained the Res., vol. 45, no. 2, pp. 101–108, May 1998.
method of mapping DFT into BN and introduced the features [15] S. M. Lavasani, A. Zendegani, and M. Celik, ‘‘An extension to fuzzy
of the two methods that help to contrast them. BN is one fault tree analysis (FFTA) application in petrochemical process indus-
try,’’ Process Saf. Environ. Protection, vol. 93, pp. 1–48, Jan. 2014, doi:
of the significant ways to resolve the problem for uncertain 10.1016/j.pSep. 2014.05.001.
inference of the period. The experimental simulation and [16] Y. Zheng, F. Zhao, and Z. Wang, ‘‘Fault diagnosis system of bridge crane
fault diagnosis of the BN model were conducted on GeNIe equipment based on fault tree and Bayesian network,’’ Int. J. Adv. Manuf.
Technol., vol. 105, no. 9, pp. 3605–3618, Dec. 2019, doi: 10.1007/s00170-
software. 019-03793-0.
This method can significantly increase fault analysis effi- [17] Y. Cheng, T. L. Xu, and S. Yang, ‘‘Bayesian network based fault diagnosis
ciency in engineering practice, and it can make full use of and maintenance for high-speed train control systems,’’ in Proc. Int. Conf.
Qual., Rel., Risk, Maintenance, Saf. Eng., Jul. 2013, pp. 1753–1757.
prior knowledge and historical fault data to improve the effec- [18] S. Yang, M. Lu, and L. Ge, ‘‘Bayesian network based software reliability
tiveness of pursuant fault analysis. The major contribution prediction by dynamic simulation,’’ in Proc. 7th Int. Conf. Softw. Secur.
of this article is to combine two approaches, which include Rel., Jun. 2013, pp. 13–20, doi: 10.1109/SERE.2013.24.
[19] Q. He, R. Zhang, T. Liu, Y. Zha, and J. Liu, ‘‘Multi-state system reliability
dynamic fault tree and Bayesian network, to resolve a signif- analysis methods based on Bayesian networks merging dynamic and fuzzy
icant and actual case engineering problem and has a certain fault information,’’ Int. J. Rel. Saf., vol. 13, nos. 1–2, pp. 44–59, 2019.
promotional value. [20] B. Uzunoğlu, ‘‘Bayesian approach with subjective opinion fusions for
wind turbine maintenance,’’ J. Phys., Conf. Ser., vol. 1037, Jun. 2018,
Art. no. 062021, doi: 10.1088/1742-6596/1037/6/062021.
REFERENCES [21] H. Wang, F. Duan, and J. Ma, ‘‘Reliability analysis of complex uncertainty
multi-system based on Bayesian network,’’ Maintenance Rel., vol. 21,
[1] M. Volk, N. Weik, J. Katoen, and N. Nießen, ‘‘A DFT modeling approach
no. 3, pp. 410–429, 2019, doi: 10.17531/ein.2019.3.8.
for infrastructure reliability analysis of railway station areas,’’ in Proc. Int.
[22] A. Halabi, R. S. Kenett, and L. Sacerdote, ‘‘Modeling the relationship
Workshop Formal Methods Ind. Crit. Syst., in Lecture Notes in Computer
between reliability assessment and risk predictors using Bayesian networks
Science, vol. 11687, 2019, pp. 40–58, doi: 10.1007/978-3-030-27008-7_3.
and a multiple logistic regression model,’’ Qual. Eng., vol. 30, no. 4,
[2] S. Kabir, K. Aslansefat, I. Sorokos, Y. Papadopoulos, and
pp. 663–675, Oct. 2017, doi: 10.1080/08982112.2017.1368556.
S. Konur, ‘‘A hybrid modular approach for dynamic fault tree
[23] H. Boudali and J. B. Dugan, ‘‘A discrete-time Bayesian network reliability
analysis,’’ IEEE Access, vol. 8, pp. 97175–97188, 2020, doi:
modeling and analysis framework,’’ Rel. Eng. Syst. Saf., vol. 87, no. 3,
10.1109/ACCESS.2020.2996643.
pp. 337–349, Mar. 2005, doi: 10.1016/j.ress.2004.06.004.
[3] S. Kabir, T. K. Geok, M. Kumar, M. Yazdi, and F. Hossain, ‘‘A method [24] M. Li, D. Wang, and H. Shan, ‘‘Risk assessment of mine ignition sources
for temporal fault tree analysis using intuitionistic fuzzy set and using fuzzy Bayesian network,’’ Process Saf. Environ. Protection, vol. 125,
expert elicitation,’’ IEEE Access, vol. 8, pp. 980–996, 2020, doi: pp. 297–306, May 2019, doi: 10.1016/j.psep.2019.03.029.
10.1109/ACCESS.2019.2961953.
[25] A. Bobbio, L. Portinale, M. Minichino, and E. Ciancamerla, ‘‘Improving
[4] M. Volk, S. Junges, and J.-P. Katoen, ‘‘Fast dynamic fault tree analysis by the analysis of dependable systems by mapping fault trees into Bayesian
model checking techniques,’’ IEEE Trans. Ind. Informat., vol. 14, no. 1, networks,’’ Rel. Eng. Syst. Saf., vol. 71, no. 3, pp. 249–260, Mar. 2001.
pp. 370–379, Jan. 2018, doi: 10.1109/TII.2017.2710316. [26] S. Kabir, M. Taleb-Berrouane, and Y. Papadopoulos, ‘‘Dynamic reliability
[5] M. Ghadhab, S. Junges, J.-P. Katoen, M. Kuntz, and M. Volk, ‘‘Safety assessment of flare systems by combining fault tree analysis and Bayesian
analysis for vehicle guidance systems with dynamic fault trees,’’ Rel. Eng. networks,’’ Energy Sour., A, Recovery, Utilization, Environ. Effects, vol.
Syst. Saf., vol. 186, pp. 37–50, Jun. 2019, doi: 10.1016/j.ress.2019.02.005. 2019, pp. 1–18, Sep. 2019, doi: 10.1080/15567036.2019.1670287.
[6] J. B. Dugan, S. J. Bavuso, and M. A. Boyd, ‘‘Dynamic fault-tree models [27] N. Khakzad, F. Khan, and P. Amyotte, ‘‘Safety analysis in process
for fault-tolerant computer systems,’’ IEEE Trans. Rel., vol. 41, no. 3, facilities: Comparison of fault tree and Bayesian network approaches,’’
pp. 363–377, Sep. 1992, doi: 10.1109/24.159800. Rel. Eng. Syst. Saf., vol. 96, no. 8, pp. 925–932, Aug. 2011, doi:
[7] B. Cai, Y. Liu, Y. Ma, Z. Liu, Y. Zhou, and J. Sun, ‘‘Real-time reliability 10.1016/j.ress.2011.03.012.
evaluation methodology based on dynamic Bayesian networks: A case [28] M. S. Rahman, F. Khan, A. Shaikh, S. Ahmed, and S. Imtiaz, ‘‘A condi-
study of a subsea pipe ram BOP system,’’ ISA Trans., vol. 58, pp. 595–604, tional dependence-based marine logistics support risk model,’’ Rel. Eng.
Sep. 2015, doi: 10.1016/j.isatra.2015.06.011. Syst. Saf., vol. 93, pp. 1–15, Jan. 2020, doi: 10.1016/j.ress.2019.106623.
[8] B. Cai, Y. Liu, and M. Xie, ‘‘A dynamic-Bayesian-network-based fault [29] G. Zhang, C. Wang, Y. Jiao, H. Wang, W. Qin, W. Chen, and G. Zhong,
diagnosis methodology considering transient and intermittent faults,’’ ‘‘Collapse risk analysis of deep foundation pits in metro stations using
IEEE Trans. Autom. Sci. Eng., vol. 14, no. 1, pp. 276–285, Jan. 2017, doi: a fuzzy Bayesian network and a fuzzy AHP,’’ Math. Problems Eng.,
10.1109/TASE.2016.2574875. vol. 2020, pp. 1–18, Apr. 2020, doi: 10.1155/2020/4214379.
[9] B. Cai, H. Liu, and M. Xie, ‘‘A real-time fault diagnosis method- [30] X. Feng, J. Jiang, and W. Wang, ‘‘Gas pipeline failure evaluation method
ology of complex systems using object-oriented Bayesian networks,’’ based on a Noisy-OR gate Bayesian network,’’ J. Loss Prevention Process
Mech. Syst. Signal Process., vol. 80, pp. 31–44, Dec. 2016, doi: Industries, vol. 66, pp. 1–10, Jul. 2020, doi: 10.1016/j.jlp.2020.104175.
10.1016/j.ymssp.2016.04.019. [31] W. Qiao, Y. Liu, X. Ma, and Y. Liu, ‘‘Human factors analysis for mar-
[10] B. Cai, X. Shao, Y. Liu, X. Kong, H. Wang, H. Xu, and W. Ge, ‘‘Remain- itime accidents based on a dynamic fuzzy Bayesian network,’’ Risk Anal.,
ing useful life estimation of structure systems under the influence of vol. 40, no. 5, pp. 1–24, 2020, doi: 10.1111/risa.13444.
multiple causes: Subsea pipelines as a case study,’’ IEEE Trans. Ind. [32] Q. Xiao, F. Luo, and Y. Li, ‘‘Risk assessment of seaplane operation safety
Electron., vol. 67, no. 7, pp. 5737–5747, Jul. 2020, doi: 10.1109/TIE.2019. using Bayesian network,’’ Symmetry, vol. 12, no. 6, p. 888, 2020, doi:
2931491. 10.3390/sym12060888.

VOLUME 9, 2021 2631

 C. Wang et al.: Fault Diagnosis of TCMS Based on DFT and BN

[33] F. P. G. Marquez, I. S. Ramirez, B. M. Ivatloo, and A. P. Marugan, LIDE WANG received the M.Eng. degree from
‘‘Reliability dynamic analysis by fault trees and binary decision diagrams,’’ Southwest Jiaotong University, China, in 1986.
Information, vol. 11, no. 6, p. 324, 2020, doi: 10.3390/info11060324. He is currently a Professor with the School of
[34] S. Nozhati, B. R. Ellingwood, and H. Mahmoud, ‘‘Understanding com- Electrical Engineering, Beijing Jiaotong Univer-
munity resilience from a PRA perspective using binary decision dia- sity. His research interests include control of elec-
grams,’’ Risk Anal., vol. 39, no. 10, pp. 2127–2142, Oct. 2019, doi: tric traction systems, computer control network
10.1111/risa.13321. technique, and embedded systems and applica-
[35] S. Zhou, J. Xiang, and W. Wong, ‘‘Reliability analysis of dynamic
tions. He is a Senior Member of the China Railway
fault trees with spare gates using conditional binary decision dia-
Society.
grams,’’ J. Syst. Softw., vol. 170, Dec. 2020, Art. no. 110766, doi:
10.1016/j.jss.2020.110766.
[36] C. Liu, X. Li, and P. Yang, ‘‘Train control management system safety
assessment,’’ in Proc. Int. Conf. Electr. Inf. Technol. Rail Transp., vol. 2,
2014, pp. 583–591, doi: 10.1007/978-3-642-53751-6_62. HUANG CHEN received the B.Eng. degree
[37] H. Zhao, Z. Huang, and Y. Mei, ‘‘High-speed EMU TCMS design and LCC from Beijing Jiaotong University, Beijing, China,
technology research,’’ Engineering, vol. 3, no. 1, pp. 122–129, Feb. 2017, in 2014, where he is currently pursuing the Ph.D.
doi: 10.1016/J.ENG.2017.01.004. degree in electrical engineering with the School
[38] H. Yin, K. Wang, Y. Qin, Q. Hua, and Q. Jiang, ‘‘Reliability analysis of Electrical Engineering. His research interests
of subway vehicles based on the data of operational failures,’’ EURASIP include train communication networks and heuris-
J. Wireless Commun. Netw., vol. 2017, no. 1, p. 212, Dec. 2017, doi: tic algorithm.
10.1186/s13638-017-0996-y.
[39] S. Kabir, M. Walker, and Y. Papadopoulos, ‘‘Dynamic system safety
analysis in HiP-HOPS with Petri nets and Bayesian networks,’’ Saf. Sci.,
vol. 105, pp. 55–70, Jun. 2018, doi: 10.1016/j.ssci.2018.02.001.
[40] S. Kabir and Y. Papadopoulos, ‘‘Applications of Bayesian networks and
Petri nets in safety, reliability, and risk assessments: A review,’’ Saf. Sci., YUEYI YANG received the B.Eng. degree in elec-
vol. 115, pp. 154–175, Jun. 2019, doi: 10.1016/j.ssci.2019.02.009. trical engineering from the Zhongyuan University
[41] Y. F. Li, J. Mi, H. Huang, N. Xiao, and S. Zhu, ‘‘System reliability of Technology, in 2011, and the M.Sc. degree
modeling and assessment for solar array drive assembly based on Bayesian in electrical engineering from Southwest Jiaotong
networks,’’ Eksploatacja i Niezawodnosc-Maintenance Rel., vol. 15, no. 2, University, in 2014. He is currently pursuing the
pp. 117–122, 2013. Ph.D. degree in electrical engineering with Beijing
Jiaotong University, Beijing, China. His research
interests include networked and distributed sys-
tems, and industrial cyber-physical systems.

CHONG WANG received the B.Eng. degree from YE LI received the B.Eng. degree in electrical engi-
Shandong Agricultural University, China, in 2016. neering and automation from Shanxi University,
He is currently pursuing the Ph.D. degree in elec- Shanxi, China, in 2019. He is currently pursuing
trical engineering with the School of Electrical the master’s degree in electrical engineering with
Engineering, Beijing Jiaotong University, Beijing, Beijing Jiaotong University, Beijing, China. His
China. His research interests include fault diagno- research interests include data communication and
sis and reliability analysis of train communication fault diagnosis.
networks and Bayesian network theory.

2632 VOLUME 9, 2021