See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/362174484

Train-Centric Communication Based Autonomous Train Control System

Article in IEEE Transactions on Intelligent Vehicles · July 2022

DOI: 10.1109/TIV.2022.3192476

CITATIONS READS
10 554

5 authors, including:

Haifeng Song Shigen Gao

Technische Universität Braunschweig Beijing Jiaotong University
50 PUBLICATIONS 383 CITATIONS 71 PUBLICATIONS 1,176 CITATIONS

SEE PROFILE SEE PROFILE

Yidong Li
Beijing Jiaotong University
197 PUBLICATIONS 1,868 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Development, evaluation and verification of a metro train collision avoidance system using Colored Petri net View project

High-speed train tight tracking protection perception and cooperative control View project

All content following this page was uploaded by Haifeng Song on 27 January 2023.

The user has requested enhancement of the downloaded file.

 1

Train-centric Communication based Autonomous

Train Control System
Haifeng Song∗ † , Shigen Gao† , Yidong Li† , Ling Liu‡ , Hairong Dong†
∗ Schoolof Electronic and Information Engineering, Beihang University, Beijing 100191, China
† State Key Laboratory of Rail Traffic Control and Safety, Beijing Jiaotong University, Beijing 100044, China
‡ Beijing National Railway Research &Design Institute of Signal &Communication, Beijing 100070, China

Abstract—Driverless train control systems are becoming pop- The reason is mainly due to the unbalanced growth of high
ular in the world. Notably, this technology has already been operating mileage and passenger demands. In the last decades,
applied for several years. However, these kinds of Automatic the operating mileage has increased by five times. However,
Train Operation systems are still under the framework that
network elements and movement authority are generated from the demand for passenger transportation has increased by 15
the ground control center. Hence, the aforementioned rail systems times. The urgent need for sustainable development. However,
are “automated” but not “autonomous”, and the architecture with the rapid growth and changing customer expectations,
utilization and train tracking interval are limited. With the the current train operation strategy can not meet the practical
development of advanced sensors and control algorithms, the demands [1]. Integrating new technologies into the railway can
train intends to be equipped with a decision-making capability.
In order to further improve train movement coordination and provide possible solutions to address these challenges.
efficiency, an Autonomous Train Control System (ATCS) is Hence, the next-generation train control system’s develop-
proposed. This paper provides three main contributions: Firstly, ment direction is to make the train run more efficient, more
the system structure is described, and the information following accessible, more reliable, and much safer [2]. ATCS is a
in ATCS is discussed with general operation scenarios. Secondly,
data isolation is broken under ATCS, and data prediction and
possible solution, which can efficiently and autonomously use
edge-based information fusion are applied to process the real resources to achieve higher utilization of the ground equipment
dynamic data and estimate the train’s speed and position. [3]. In order to put forward the implementation of ATCSs, the
Thirdly, with the assistant of data fusion, model predictive and following aspects are intended to carry out:
data-based control methodologies are discussed.
• running trains closer;
Index Terms—Autonomous train control, train-centric commu- • dispatching trains faster;
nication, multiple scenarios, train control system.
• more valuable data;
• optimum energy use;
I. I NTRODUCTION • low-cost railway solutions;
High-speed rail has become the main body of inter-regional • higher safety level.
rapid transportation in the comprehensive transportation sys- As releases by Japan Transportation and Tourism Research
tem. The development of the train control signaling system is Institute, there are currently more than 64 fully automated train
put further by the serious safety accident, operation require- lines around the world. More than half of them are located
ment, and technology improvement. Take China, which is the in Asia [4]. Japan proposed a next-generation railway system,
only country in the world where high-speed railways are net- and modifications are intended to make in the system structure
worked, as a case study, in 2019, the operating mileage reached and wayside equipment [5]. Electric cables are replaced by
35,000 kilometers. The proportion of railway passenger traffic wireless or optical fiber. As a result, Train-to-train communi-
has reached nearly 60%, which is three times more than air cation is available, and more essential data can be transferred
passenger traffic. to ground for optimal operation decision making.
Even though high-speed rail operation and under con-
Similarly, the on-board equipment is taken more localization
struction mileage are increasing globally, some inter-regional
functions in Positive Train Control (PTC) system, which is
routes still can not meet the passenger demand. Taking the
released by America [6]. The on-board equipment transmits its
Beijing-Shanghai, China, high-speed line as a case study, it
own status and location information to the PTC server through
is becoming more saturated. The inter-city Beijing-Tianjin,
the wireless communication network. The server calculates the
China, is a further example, and the tickets are hard to order
movement authority based on the running train information
in rush hours. The Shinjuku station, Japan transferred more
data and the vehicle control information obtained by the
than 775 thousands passengers by JR East Japan in one day
dispatching system, and sends it to the train through wireless
in 2019.
communication. To reduce development cost, GE Tempo is
This work is supported by National Natural Science Foundation of China introduced by GE Transportation. A unified safety platform
(No. 61925302, 61790573, 61903021, 62073027). Beijing Natural Science is used for the ground and on-board equipment. It makes
Foundation L211021. Prof. Hairong Dong is the corresponding author
([email protected]), and she is with the State Key Laboratory of Rail Traffic possible to operate train both in main railways and urban lines.
Control and Safety, Beijing Jiaotong University. Configuration and optimized structure are flexible, and life
 2

cycle costs are efficiently reduced. control systems. Otherwise, train contract system requirements
ERTMS-Regional has removed the track-side equipment, specification1 are highly recommended for starters.
canceled the track circuit’s occupancy inspection of the train,
and carried out the train positioning and integrity inspection
A. Train operation and control procedure
autonomously by the vehicle, and used the object controller
to control the turnout. In order to ensure the train can The high-speed railway is composed of infrastructure, mo-
have an efficient communication and feedback with various bile equipment, and operation control system. The operation
sensors and scenarios, Artificial Intelligent (AI) technologies control and dispatch command system in the operation control
are involved [7]. Project RAILS, which is funded by the system is the brain and central nervous system of the high-
Shift2Rail Joint Undertaking, intends to apply AI in the areas speed railway system. The improvement of the sequential
of rail safety, automation, traffic planning and management [8]. operation’s intelligence level is also the key to improving
Sydney Metro Northwest implemented a Grade of Automation the capacity. Two main concepts in the train control system
4 (GOA 4) metro, which can fully automated carry the function are Automatic Train Protection (ATP) and Automatic Train
of departing, arriving, door operation, emergency handling. Operation (ATO).
Alstom in France put a simplified CBTC system, in which (1) Based on the timetable, the Centralized Traffic Control
the ground equipment has been streamlined and the ground (CTC) system sends train operation plans and dispatching
interlocking and area controllers have been eliminated, named commands to each Radio Block Center (RBC) system within
Urbalis Fluence into operation on the Lille Line 1. Fur- its jurisdiction. According to the received dispatching com-
thermore, Alstom also deployed an incremental train control mand, RBC combines the position information of the pre-
system (ITCS) in low density railway networks [9]. ITCS ceding train, the line condition, and temporary speed limit
requires very little infrastructure on the ground since track-side information to generate the movement authority (MA) required
devices are eliminated by using Global Positioning System by the on-board control subsystem. MA and line conditions
(GPS) to determine the train location. are sent to the train through the train to ground communication
Ultimately, ATCS has to handle all these unexpected events network.
on the rail network, which is similar to autonomous vehicles. (2) The train control subsystem calculates and generates a
Event though the equipment and line route in railway are less speed profile based on the received MA and real-time train op-
variable than the road traffic, trains are potentially have more erating parameters. Simultaneously, it uses train speed/distance
catastrophic consequences [10]. The latest autonomous train sensors to collect real-time train operating status, and pe-
systems are being designed to adapt to complications such as riodically sending it to RBC through the train to ground
train delays and modifications to the schedule. communication network.
Comprehensive and efficient use of these resources and (3) The automatic train protection (ATP) system monitors
technologies, perception of high-speed rail-mobile equipment, the train running speed in real-time. When the actual speed
fixed infrastructure, and internal and external environment in- exceeds the safe allowable speed, ATP automatically controls
formation can be carried out. As a new generation of intelligent the train to decelerate.
high-speed railway system, ATCS is unique in interconnection, (4) According to the speed profile, the on-board ATO
fusion processing, active learning, and scientific decision- subsystem, which is activated in the automatic driving mode,
making [11]. Hence, in this paper, we propose the definition or the driver, who is in charge in the manual driving mode,
of an ATCS and discuss the main modifications different from drives the train in a safe, punctual and smooth operation. As
the current one, which is introduced in Section II. a result, a safe tracking interval with the preceding train is
The remainder of the paper is organized as follows: Based maintaining.
on our previous work, the system structure and main com-
ponents of the ATCS are proposed in Section III. To further B. Automatic Train Protection and Automatic Train Operation
analyze the system logic of ATCS, data information flow and
fusion prediction methodology are discussed in Section III-C ATP is the safety system that ensures a safe distance
and Section V, respectively. With the broken of data isolation, between the trains and gives sufficient warnings when to stop.
which is an obvious limitation in current train control system, However, ATO is the non-safe part of train operations when
ATCS operation scenarios and traffic securing principle are it comes to stopping and starting stations and shows the stop
illustrated in Section IV. Finally, Section VII concludes this position for the train. As soon as the train stops, ATP has
paper and presents future works. confirmed that the line is clear.
ATP provides the driver a limited speed, which is calculated
based on the information from the ground, training practical
II. PRELIMINARIES parameters, and line condition; it supervises the train opera-
tion and executes a normal or emergency brake if the train
In order to make the paper more readable, some basic pre- exceeded the speed limit. Usually, the braking actions can be
liminary notions, which can be mentioned for the discussion divided into three levels: normal (Vn (t)), maximum (Vm (t)),
that will be conducted in the following sections, are illustrated and emergency (Ve (t)) braking.
in this section. What is more, we assume that the reader has
a basic knowledge of the current and next-generation train 1 ERTMS/ETCS System Requirements Specification (SUBSET-026)
 3

D
CTC
T2G Timetable

RMU

Transmitter Train to infrastructure communication

Balise
control unit Track
Infrastructure Switch
Radio DMI Controller
Interlocking unit Door Light
T2T
Data buffer Digital map ATP ATO Tract\Brake
T2G
Movement algorithm Train Interface

Fig. 1: Component of Autonomous Train Control System

ATO works under the protection of ATP system. It is an in interconnection, fusion processing, active learning, and sci-
automatic train control system that realizes the functions of entific decision-making. The benefits of applying autonomous
automatic train running, precise parking, automatic platform railway can be given as: higher speed in trains running;
operation, unmanned return, and automatic train operation shorter time in passengers traveling; a smaller space in lands
adjustment. occupying; a better service in data exchanging; a stricter
ATO adopts a high-reliability hardware structure and soft- safety in the system requiring. The structure of the proposed
ware design to achieve a layered architecture. Based on single autonomous train control system is shown in Fig .1. The main
and multiple mass points dynamic train models, ATO real- modifications different from the current train control systems
izes the organic combination of long-period operating curve are fallen into wayside infrastructure, train operation, on-board
optimization and short-period real-time control. As a result, equipment, and data system.
ATO can significantly reduce the labor intensity of drivers, and
guarantee the rail transit system to be efficient, comfortable, B. System structure and main components
punctual, accurate, and energy-saving.
1) Physical rail and wayside infrastructure: Existing in-
ATO can automatically control the train to complete the
frastructure assembles the basic physical rail. It does not
process of starting, cruising, and precise stopping. Addition-
require any necessary adjustments to the track topology, as
ally, ATO provides drivers with complete driving assistance
shown in Fig. 2. For the wayside infrastructure, a resources
information and guide them to drive. It can automatically open
management unit, which is in charge of resources and commu-
and close the door and control the interlocking between the
nication with trains in its jurisdiction, is required to be laid in
platform door and the safety door. The operation adjustment
section blocks. Similar reports indicated that external safety
commands, such as train jumping, detention, and inter-station
installations and corresponding cost savings in construction
operation adjustment, can be automatically completed by
and maintenance can be saved up to 70% with taking this
applying ATO.
kind of modification [12]. More details about the wayside
infrastructure are discussed in Section III-C.
III. AUTONOMOUS TRAIN CONTROL SYSTEM
2) Train operation: Train operation ensures consistent
A. Definition of autonomous train control system planning and control of train traffic. The creation and ad-
Autonomous high-speed rail has an autonomy characteristic, justment of train schedules are largely automated. If there
which means the train can positively require and occupy the is a deviation, a new optimization timetable will be created
railway resources, such as track sections and switches. Unlike automatically. This ensures that the existing railway system
ATO, an executor, ATCS is more like a decision-maker. By can be used optimally in all conditions. For instance, in one
comprehensive perceiving fixed infrastructure, internal and of our previous publications, deep reinforcement learning is
external environment information, ATCS will efficiently use applied to adjust the timetable. The average total delay could
resources to achieve higher utilization of the ground equip- be reduced by 46.38% in the disturbance scenarios [13].
ment. 3) On-board equipment: On-board equipment undertakes
As a new development direction of intelligent high-speed the resources application, movement authority execution, inter-
railway systems, autonomous train control takes more effect locking verification, and train to train/ground communication,
 4

Resources Management Unit

Analyzing Big/sensitive data application
Train to infrastructure communication

Track Data lake Operation zone ……

Infrastructure Processing
Switch
Controller Onboard equipment zone Resource zone
Light

Train to Train
communication Obtaining Existing data system External data
Train 1 Train 2

Fig. 4: Autonomous high-speed rail data system architecture

Fig. 2: Physical rail and infrastructure

the running data exchanging among nearby trains. Different

as shown in Fig. 3. Unlike the current on-board equipment, from T2G communication, T2T plays an important role for
ATCS will have an on-board interlocking unit. The unit the decentralized self-discipline, which has a shorter time
analyzes the data from train position, track occupation and latency. These two communication functions are implemented
movement authority, and prevents conflicting movements and by different wireless channels and devices, which make the
ensure safety. data processing be parallel. However, the OI procedure is more
complex than the aforementioned two. Large amounts of data
are pouring into the Data buffer, and practical algorithms are
getting involved to generate suitable commands for the train
Optical Sensor Unit
Satellite Based operation.
Relative Positioning
Transmitter control unit (TCU) is applied to control the
On-board Communication Unit communication process, which involves three different com-
Interlocking Unit large/mid range munication targets. These targets are trains, ground relays,
and resources management units. T2T communication can be
implemented by the device to device (D2D) in a short-range,
Fig. 3: On-board equipment components and through relays (D2R) under a long communication range
scenario. TCU is in charge of exchanging train location and
4) Data system: Data system mainly focuses on data ob- operation information between trains, and establishing the T2G
taining, processing, and analyzing as shown in Fig. 4. Precise communication between train and Resources Management
location plays an important role in autonomous train control. Unit.
To provide a precise and continuous train position, various Train to infrastructure communication unit in charge of
sensors are implemented on the train. How to combine the requiring resource with Resources Management Unit; Train to
heterogeneous information and output a safe speed/position train communication unit provides the function of proceeding
data is the main challenge [14]. Hence, some functionalities trains’ identification and data exchanging, and respond to its
can be shifted from the infrastructure to the vehicle. This following train’s request. More details about how the system
allows more precise control of train movements (capacity) and structure and data acquisition of T2T and T2G communication
dismantling the track-release signalling equipment. is implemented can be found in our previous publication [15];
Low latency and high secure data transmission ensure an the hardware structure of train-centric communication unit and
optimized control. After the data distillation process in the availability evaluation are introduced in [16], [17]; the avail-
data lake, high-dimensional and valued data is ready to do the ability and performance analysis of T2T data communication
further analyzing. Note that, normally the on-board equipment system is discussed in [18], [19]. Readers who have an interest
has a high system safe while the computing power is limited. in communication systems can give more information in our
Hence, the data prediction and analyzing within the edge- aforementioned previous publications.
to-cloud environment provide possible solutions, which is
The main function of Resources Management Unit is to
discussed in section V.
record train operation information, provide information query
functions for trains, and release requires to Infrastructure
C. Information flow in the autonomous train control system Controller. The Infrastructure Controller is in charge of switch
The information flow among ground equipment, on-board situation management and commands execution. Infrastructure
units, different trains can be described as shown in Fig. 5. Controller arranges the resources like track, switch, light,
The flow can be divided into train to ground (T2G), train and so on. It will manage the resources’ situation monitoring
to train (T2T), and on-board internal (OI), respectively. T2G and actions executing. Additionally, it will also analyze the
communication is in charge of mission receiving from CTC, requirements from different trains and determine the usage
resources querying with RMU, position data reporting to priority. Depends on the installation places, the Infrastruc-
RBC, and so on. T2T communication pays more attention to ture Controller has two main types: distributed edge and
 5

Radar BTM IMU

Train to infrastructure
Transmitter Transmitter communication unit
Data ATP VOBC
control unit control unit
Train to train buffer
communication unit
Resources
Management Unit Interlocking Movement I/O interface
unit algorithm unit
Infrastructure
Controller Digital map

Fig. 5: Information flow among modules and subsystems

centralized zone ones, which can be implemented in blocks on-board sensors such as Global Navigation Satellite System
and stations, respectively. The distributed edge controller has (GNSS), Balise and inertial measurement unit (IMU), and
a succinct structure and a lower maintenance requirement. information from the ground and other trains. All in all, the
However, the centralized zone one enables train can access Data buffer obtains and processes various data, and outputs an
to all the station resources by establishing one connection, optimal location and speed data for the Movement algorithm
and it is easy to carry maintenance and undertake a complex to do further train movement controlling.
resources authority analysis. With the development of information and embedded system,
The on-board Interlocking unit undertakes three main tasks: the train can complete a more complex mathematical calcula-
• analyzing operation plan; tion. As a result, Movement algorithm provides higher precise
• route resource query and control; and intelligent control, which can be implemented by applying
• route unlock. advanced control algorithms. Importantly, with the assistant of
infusion data, which is further discussed in V, it is possible to
The traditional ground interlocking is to interlock the four
apply a data-based model-free control method to carry a more
objects of the access road, signal, switch, and track sections. A
accurate train operation control.
specific constraint relationship is established to ensure that the
ground equipment is in place and not conflict with other trains
IV. O PERATION SCENARIOS AND TRAFFIC SECURING
when the train is running. The traditional ground interlocking
PRINCIPLE
system as a safety-critical system has logic because of the
resource constraints of the interlocking table. The advantages The operation scenarios of an ATCS can be divided into
are simplicity and high security. But the disadvantages are that resources querying, locking, occupying, and releasing, which
the traditional control method relies on interlocking to a higher are also familiar with the traditional one [21]. These four
degree. The utilization rate of resources and flexibility is low. scenarios mentioned above can be triggered by time, event,
However, the on-board interlocking module can analyze and location. Time triggering means the functions and actions
the driving plan, dynamically or statically search for routes, are activated based on the time table. Event triggering fires
and independently apply for ground resources to complete the the following sequence by a specific event [22]. Location
route arrangement. The advantage of this method is that the triggering indicates the fuse is based on the train position,
train has high autonomy, and can independently apply for which can be carried out by the data system mentioned in
and control the ground resources, effectively. The resource section III-A. The innate character of these three different
utilization and system flexibility are higher than the traditional triggers is that the system state can be modified from one to a
interlocking. While the disadvantage is that the resource specific one. Furthermore, the traffic securing of the following
management changes from a single source to multi-source different scenarios can be validated by applying the principle
control. What is more, the security of the system requires mentioned in our other contribution [23], which will not be
strict resource management processes to ensure. Otherwise, further discussed in this paper.
it is easy to cause problems such as system deadlocks and
live locks [20]. A. Resources querying
Data buffer will provide essential data predicted and edge- As shown in Fig. 6, the train obtains its mission for
based information fusion for the Movement algorithm. As Centralized Train Control (CTC) center based on the timetable.
mentioned in Fig. 4, the data can be obtained from existing Its End of Authority (EOA) stops at the front of a rail
and external data systems. The data resources are distributed in track resource, such as switch and block, as shown in Fig.
 6

6. Before the Train 1 updated it EOA, it will communicate C. Resources occupying

with other trains (Train 2) and Resources Management Unit Train 1 moves over the locked and authorized resources,
through Train to infrastructure communication unit and Train and the state of locked resources is updated to “Resource
to train communication unit, respectively. Afterward, suppose occupied”, as shown in Fig. 8. Under this situation, the
there is no conflict among trains, track, and switch set. In resource can not be unlocked until it is released. Strictly, the
that case, the Train 1 computes the aforethought resources Infrastructure Controller will deny all the requirements from
needed for its movement extension and update it requires to the management unit. At the same time, T2T communication
Resources Management Unit. Finally, the booking command is continuously carrying out. Other trains establish the con-
is transferred to Train to infrastructure communication unit, nection with Resources Management Unit and submit their
and EOA will also be extended. requirements, which will be satisfied once the resources are
free to book.
Received mission
from CTC
Resources occupied
Train 2
Communicated with
Train 2 Train 2
Denied requires
Track, switch set
booked & locked
Communicated with
Train 1 Train 2
EOA EOA
Train 1 Train 1
Extended destination
EOA
Resources Infrastructure
Management Unit Controller
Track
Switch Resources Infrastructure
Light Management Unit Controller
Track
Switch
Fig. 6: Resources querying Light

Fig. 8: Resources occupying

B. Resources locking
D. Resources releasing
Once a specific resource is ordered by a train, it will be
locked and recorded by Infrastructure Controller, which is As shown in Fig. 9, as soon as the rear of the train 1 is left
in charge of recording, setting, and locking track resources. and no longer need the occupied resources, the Infrastructure
As a result, the Resources Management Unit will deny the Controller will release the track and switch resources. Then,
subsequent connection from other trains in the same priority the footprints of occupied track and switch set recorded in
as Train 1, as shown in Fig. 7. At the same time, Train 1 will Resources Management Unit are deleted and updated as free.
communicate with Train 2 and exchange their operation data, What is more, other trains can start another round of resources
for instance, speed, direction, location, and resolution locking. querying, locking, occupying, and releasing.
Note that, if the train operation safety is guaranteed, the
resource state can be modified from “locked” to “unlocked” Resources released

and re-locked by another train. Take the prime train, which Train 2
has the highest priority, as a case study. Track, switch set booked
& locked

Communicated with
Resources locked Train 2

Train 2 1 Train 1
Denied requires
EOA

Communicated with
Train 2 Resources Infrastructure
Management Unit Controller
Train 1 Track
EOA Switch
Light

Resources
Management Unit
Infrastructure
Controller
Fig. 9: Resources releasing
Track
Switch
Light V. DYNAMIC DATA PREDICTION AND EDGE - BASED
INFORMATION FUSION
Fig. 7: Resources locking In the train autonomous control system, the train’s real-time
running data is obtained by installing on-board sensors. We
 7

can use data fusion algorithms to process real-time dynamic represents the error state of the system. Specifically, the error
data and estimate the train’s speed and position accordingly. state is defined as:
Compared with the traditional train control system, this solu- f g
tion does not need track circuits or balise to locate the train, P = δψ nbn n
|δvnb |δpb |ba |bg (2)
thus reducing the installation of ground equipment [24].
where δψ nb
n , δv n , δp , b , are the attitude error (roll, pitch
nb b a
and yaw), speed error, position error(latitude, longitude and
A. Dynamic Data height), accelerometer biases and gyroscope biases, respec-
The framework of the train autonomous navigation system tively.
is shown in Fig. 10, the autonomous train navigation system Since each sensor has its own character, the same one sys-
uses different sensors to obtain real-time dynamic data of tem equipment should be adjusted with unique measurement
trains such as a global positioning system, inertial navigation equations, which can be given by:
system, and speed sensor, etc [25]. As the data of different
sensors are input to the edge node, the local state estimation mi (k) = κ i (k)P(k) + υi (k) (3)
and covariance are obtained through the local filters. The data
fusion center is used as the master filter for train navigation where mi (k) and κ i (k) are the measurement result and mea-
to fuse each local filter’s outputs. The master filter assigns surement matrix of different local filters respectively. υi (k) is
different weight values to each state variable through the the measurement noise with zero mean and covariance Ri (k).
covariance obtained by different local filters. After the data
fusion center processing, the master filter outputs the global B. Data prediction and analyzing
optimal train state estimation result, which contains real-time In order to accelerate the data analysis procedure and
position and speed values of the train. reduce time delay, the time-update and measurement-update
steps are deployed in the edge node [27]. In the time-update
step, the specific force and angular velocity measured by the
Inertial inertial navigation system are applied into the local filter, for
Navigation instance, Kalman filter. Afterwards, the system state and the
System corresponding covariance matrix are propagated. Equations of
one-step state and error variance predictions can be given as
Global 
x 1(t t )
equations (4) and (5), respectively.
Navigation 

System p 1(t t )

Edge Layers

Train Speed x tt )
(
 2
Sensor
the input for k+1
Measurement Update
p 2(t t )
Outputs at k is
Data
Data obtaining

(correction )
Fusion Data
. . Center

Time Update
Aggregation Node

. . x o(t t ) Initial
(prediction)

 estimates
. . p o(t t ) at k=0

Other x n(t t ) IMU data

Sensors p n(t t )
GPS Data
Sensors

Big Data Cloud

Speed sensor Data

Fig. 10: Framework of train dynamic running data subsystem
Cloud
Other data
According to the framework of the train autonomous nav- IoT

igation system, the inertial navigation system is used as a

reference sensor. Therefore, the system equation of local filters Fig. 11: Data prediction edge-to-cloud environment
can be expressed with the inertial navigation system. The error-
state Kalman filter is usually used to achieve the navigation ∧ ∧
results’ closed-loop correction in the inertial navigation system P (k |k − 1) = Φ I N S (k − 1) P (k − 1) (4)
[26]. The INS error model can be derived from INS pure nav-
igation equations using the perturbation method. The system
equation is expressed as: µ(k |k − 1) = Φ I N S (k − 1) µ(k − 1)ΦTIN S (k − 1) +Q(k − 1) (5)
The state and covariance matrix are corrected in the
P(k) = Φ I N S (k − 1)P(k − 1) + g(t − 1) (1)
measurement-update step whenever the sensor measurements
where Φ I N S (k − 1) is the system matrix, g(k − 1) is the white are available. The Kalman measurement-update step consists
Gaussian noise with zero mean and covariance Q, and P(k) of the following operations.
 8

where W is the external force; F is the driving force; B is the

braking force; C1 , C2 , and C3 indicate the integrated forces
 G(k) = µ(k |k − 1)κ i (k)[κ i (k) µ(k |k − 1)κ i (k) + R
T T −1

 i (k)]
 P̂(k) = P̂(k |k − 1) + G(k)[m(k) − κ i (k) P̂(k |k − 1)] −1 under cruising, coasting, and braking, respectively.
 µ(k) = [I − G(k)κ (k |k − 1)]µ(k |k − 1)
 According to Newtons law and related studies [27], the train
i

(6) dynamic model can be written as follows:
where G(t) is the gain matrix; P̂(k) and µ(k) are the estimated dt 1
state and error variance matrix, respectively. = (10)
dx v
As shown in Fig. 11, after each local filter obtains the state
and corresponding covariance matrix, the calculation result, dv u(x) − w0 (v) − g (x) − wr (x)
= (11)
which will be sent to the main filter immediately for further dx v
calculation, is generated by the edge nodes. x i (i = 1, 2, . . . , l) where w0 (v) is running resistance force per unit mass, g(x)
represents the n-dimension state estimation of each local filter. is the component of gravitational acceleration due to the
It is assumed that pi and p j ,i , j are independent and track gradient considering space occupancy, wr (x) is curve
uncorrelated. What is more, the covariance matrix of each resistance force per unit mass considering space occupancy.
local filter is µii . As a result, the optimal information fusion The running resistance force per unit mass w0 (v) is obtained
estimator with matrix weights is given by: as:
l

p̂ = µ
P
µ−1 w0 (v) = ( A + B · v + C · v 2 ) · g · 10 −3 (12)
ii pi


 g
i=1


l (7) where A, B and C are empirical coefficients fitted from field
µ = µ −1 ) −1
 P

 g
 ( ii data, all three are positive. g is the gravity acceleration. So the
i=1
w0 (v) is strictly increasing function, which means w0 (v1 ) >

where p̂ represents the global optimal train state estimation, w0 (v2 ) for v1 > v2 and the derivative of w0 (v) with respect to
µ g is the corresponding global error variance matrix. v is non-negative.
Since the same priori estimates and process noise, the local
estimates obtained by each local filter are correlated. In order B. Engineering implementation control algorithm
to have a more accurate estimations, the fused results can be
The most applied control algorithm is based on PID (Pro-
advanced by modifying the aforementioned equation.
portion Integration Differentiation) control, which can be de-
The status information such as train position and speed scribed as shown in Fig. 12. As released in the IFAC Newslet-
obtained according to the above process does not depend ter No.2, 2019, 91% Engineers thought the PID controller is
on the help of ground equipment and carries out the train’s the most important one in the current. The PID can provide
autonomous navigation function. The navigation system uses a simple in design, easy to tune, and satisfactory performance
a distributed filter structure. Setting multiple local filters can in engineering implementation scenarios.
effectively reduce the computational burden of the data fusion
center and improve the efficiency of the algorithm [28].
P
Moreover, even though a specific sensor was failed, the system
can continuously output the train running data based on the I
Control
target
data obtained by the remaining sensors. As a result, the train
D
positioning function’s stability can be ensure.

Fig. 12: PID controller

VI. T RAIN CONTROL METHODOLOGY
PID controller is widely used in industrial process control.
A. Dynamic model of train movement Even though the rapid development of control theory and
technology, PID control still has strong vitality due to its
The dynamic model of train movement can be simplified as
simplicity, good stability, and high reliability. In PID control,
shown in
the key problem is adjusting PID parameters. The traditional
ṡ = v method is to determine the PID parameters based on a certain
v̇ = a principle with the mathematical model of the object. The con-
(8) trol error, which represents the difference between a desired
C
a= setpoint r (t) and a measurable output ariable y(t), is given
M
in equation (13). To minimize the e(t), a continuous control
where s is the train moving distance v is the speed, a is
u(t) is applied as shown in (14), where k p , T1 and TD are
the acceleration, and C is the integrated force, and M is the
the proportional, integral, and derivative terms, respectively.
mass. The integrated force C has the following three different
For a computerized system, sucn as train tracking force, the
scenarios:
 C1 = −W

 equation has to be discreted as given in equation (15).
 C2 = F − W (9)
e(t) = r (t) − y(t)
 C3 = − (B + W )

 (13)
 9

"  t #
1 de (t)
u (t) = k p e (t) + e (x) dx + TD (14) System
Outputs
Measurement
T1 0 dt
System
Inputs
1
 1 X e (k) − e (k − 1) 
u (k) = k p 
 
e (k) + T e ( j) + TD
 MPC controller
 T1 j=0 T  Reference
trajectory
 
(15) Model
- +
However, the mechanism of many controlled processes is Predicted
Future
more complex in actual industrial process control. With char- inputs outputs
acteristics such as high nonlinearity, time-varying uncertainty,
and pure hysteresis, pure PID control is unstable. Considering
Optimizer
the influence of noise, load disturbance, and other factors, the Future
process parameters and the model structure will be changed. errors
Hence, it is necessary to modify the parameters not only based Cost Constraints
on the mathematical model, but also adjusting them on-line. function
As a result, the adaptive PID control will be an effective way
to solve this problem.
Fig. 14: MPC controller
C. Model Predictive Control
As the speed and density of trains increase, the requirements
for precise control are increasingly higher. Therefore, some uk |t − Ff
advanced control algorithms are discussed in recent years, such vk+1|t = vk |t + ( )∆t (17a)
M
as model predictive control (MPC) [29], [30]. s k+1|t = s k |t + vk |t · ∆t (17b)
0 ≤ vk |t ≤ vlim (s k |t ) (17c)
Normal formulas Optimized the controller Fmin ≤ uk |t ≤ Fmax (17d)
 sv Pmin ≤ vk |t · uk |t ≤ Pmax (17e)
 Error State estimated Control Operating status
 Mv  u  Fair  Ff
Defined by
Reference MPC
actions
Train
∀k = t, · · · ,t + N p − 1
model optimized
experience
where vk+1|t and s k+1|t are predicted states of trains. u·|t =
controller

Deep learning {ut |t , · · · ,uk+N p −1|t } is the predicted inputs. J is the cost
prediction model function, which can include tracking errors, jerk and so on.
K v ≥ 0 is the weight coefficient to penalize the output
Fig. 13: Deep learning based Model-predictive control deviation from the reference speed. It is important to choose
the appropriate weight when adding multiple optimization
MPC is an advanced method which relays on dynamic goals. [Fmin , Fmax ] is the range of traction and braking force.
models. The main advantage of MPC is that it can deal with [Pmin , Pmax ] is the power limit range. N p is the prediction
constrained optimization control problems, while considering horizon.
future states. Different from other optimi zation methods, MPC
is closed-loop control in every prediction horizon. Therefore, D. Data-based Control
it can correct the deviation between the predicted and actual In fact, it is difficult to accurately model high-speed trains
behaviours due to disturbances [31]. When used to control during operation. And when designing the controller, parame-
trains, MPC can satisfy the performance requirements with ters are usually chosen based on experience. Therefore, some
keeping the constraints in count, such as the speed limits, model-free control methods are proposed to solve the possible
power and so on. problems caused by inaccurate models [32]. Deep neural net-
The structure of MPC can be described as shown in Fig. works (DNN) can find the correct mathematical manipulation
14. The optimizer finds the optimal control policy which to turn the input into the output [33]. The structure of DNN
minimizes the cost function over a prediction horizon. The can be described as shown in Fig. 15.
constraints and observed states of the system are also con- Historical recorded data are used to train the networks. The
sidered. In every prediction horizon, only the first inputs are mean square error loss function is formulated in (18), which
applied to the system. means the Euler distance between the input and output data.
The control problem of trains can be described as (16-17). 1 X
J (w, b) = min k x̂ t − yt k 22 (18)
Xt+N p vk |t T 0≤t ≤T
min J = Kv 1 − (16)
u |t k=t vr e f where x̂ t = { x̂ 1 , x̂ 2 , · · · , x̂ T } is a set of observed data. After
subject to: given a set of input value x t = {x 1 , x 2 , · · · , x T }, the output
 10

Input layer Hidden layer Output layer

model-free control algorithm will also be further discussed
in our future work.
t

vref vt+1 ACKNOWLEDGMENT

st+1 This work is supported by the National Natural Science
vt, st
Foundation of China (No. 61925302, 61790573), Beijing Nat-
vlimit ural Science Foundation under Grant L211021.

R EFERENCES
Fig. 15: Deep neural networks [1] H. Dong, X. Liu, M. Zhou, W. Zheng, J. Xun, S. Gao, H. Song, Y. Li,
and F.-Y. Wang, “Integration of train control and online rescheduling
for high-speed railways in case of emergencies,” IEEE Transactions on
Computational Social Systems, 2021.
in DNN after batch renormalization is yt = {s t+1 , vt+1 }, in [2] D. Trentesaux, R. Dahyot, A. Ouedraogo, D. Arenas, S. Lefebvre,
which s t+1 and vt+1 denotes the predicted train position and W. Schön, B. Lussier, and H. Cheritel, “The autonomous train,” in
2018 13th Annual Conference on System of Systems Engineering (SoSE).
velocity at time t + 1, respectively. w ∗ and b∗ are sets of IEEE, 2018, pp. 514–520.
DNN coefficients, which minimize the value of nonlinear loss [3] S. Oh, K. Kim, and H. Choi, “Train interval control and train-centric
function J (w, b). distributed interlocking algorithm for autonomous train driving control
system,” Journal of the Korea Academia-Industrial cooperation Society,
The trained network is used to generate recommended vol. 17, no. 11, pp. 1–9, 2016.
operating strategies. At the input layer, time, reference speed [4] P. Singh, M. A. Dulebenets, J. Pasha, E. D. S. Gonzalez, Y.-y. Lau, and
R. Kampmann, “Deployment of autonomous trains in rail transportation:
and current states are entered. With the hidden layer, the Current trends and existing challenges,” IEEE Access, 2021.
mapping functions between input vectors and observed data [5] K. Sato, H. Kato, and T. Fukushima, “Development of sic applied
can be characterized. After renormalization, the speed and traction system for next-generation shinkansen high-speed trains,” IEEJ
Journal of Industry Applications, vol. 9, no. 4, pp. 453–459, 2020.
position at the next moment can be achieved at the output [6] J. Brosseau, C. Grimes, M. Oztekin, A. Polivka, P. Vieira et al., “De-
layer. velopment of enhanced overlay positive train control (eo-ptc),” United
States. Department of Transportation. Federal Railroad Administration,
Tech. Rep., 2019.
VII. C ONCLUSION [7] F. Flammini, V. Vittorini, and Z. Lin, “Roadmaps for ai integration in
the rail sector-rails,” ERCIM News, vol. 2020, no. 121, 2020.
[8] R. Consortium, “RAILS: Roadmaps for A.I. Integration in the Rail
The current train control systems are based on train to Sector,” https://https://rails-project.eu/, 2020, [Online; accessed 10-
ground communication, which suffers from having various September-2020].
interfaces and subsystems. What is more, the train can only [9] J. Tan, J. Wang, and D. Lu, “Gnss data driven clustering method
for railway environment scenarios classification,” in 2019 14th IEEE
passively move based on the timetable and ground center Conference on Industrial Electronics and Applications (ICIEA). IEEE,
commands. As a results, the efficiency can be blocked during 2019, pp. 2026–2031.
the information exchanging. With involving the T2T and T2G [10] C. Gely, D. Trentesaux, and A. Le Mortellec, “Maintenance of the au-
tonomous train: A human-machine cooperation framework,” in Towards
communication channels, resources management unit, and User-Centric Transport in Europe 2. Springer, 2020, pp. 135–148.
infrastructure controller, the ATCS can provide the following [11] S. Sai, L. Ping, and Z. Jinzi, “Study on the evaluation index system of
train the information of the preceding one, and maximally china intelligent high-speed railway based on technical characteristic,”
in 2019 IEEE Intelligent Transportation Systems Conference (ITSC).
utilize the ground resources. Based on the benefits of on-board IEEE, 2019, pp. 601–606.
equipment and data system illustrated in this paper, the train [12] S. A. Infrastruktur, “SMARTRAIL 4.0 - HOW
can have a more accurate real-time dynamic data and precise SWISS RAILWAY INDUSTRY GETS SMART,”
https://www.innovativeleaders.world/insights/item/22-smartrail-4-0-
control when compared with the current one. It is clear that the how-swiss-railway-industry-gets-smart.html, 2019, [Online; accessed
ATCS will benefit the next generation of train control systems. 22-August-2020].
The paper presents the structure and principle of an ATCS, [13] L. Ning, Y. Li, M. Zhou, H. Song, and H. Dong, “A deep reinforcement
learning approach to high-speed train timetable rescheduling under dis-
and its potential operation scenarios and data flow logic are turbances,” in 2019 IEEE Intelligent Transportation Systems Conference
discussed. For engineering implementation, data prediction (ITSC). IEEE, 2019, pp. 3469–3474.
and edge-based information fusion algorithms are proposed [14] F. Molano Ortiz, M. Sammarco, L. H. M. K. Costa, and M. Detyniecki,
“Applications and Services Using Vehicular Exteroceptive Sensors: a
to locate the train, and estimate the dynamic train data. With Survey,” IEEE Transactions on Intelligent Vehicles, pp. 1–20, 2022.
the assistant of essential data processing ability, advanced train [15] H. Song, H. Liu, and E. Schnieder, “A train-centric communication-
strategies can be implemented. Since the proposed autonomous based new movement authority proposal for etcs-2,” IEEE Transactions
on Intelligent Transportation Systems, vol. 20, no. 6, pp. 2328–2338,
train control system is still under developed, there is no 2018.
practical experiment data to validate the efficiency of proposed [16] A. Lehner, T. Strang, and P. Unterhuber, “Direct train-to-train com-
system. The real-world data-set (real experimental data) would munications at low uhf frequencies,” IET Microwaves, Antennas &
Propagation, vol. 12, no. 4, pp. 486–491, 2018.
make the research argument much stronger and valid. Hence, [17] D. Zamouche, M. Mohammedi, S. Aissani, and M. Omar, “Ultra-safe
in the future we are going to build a physical or digital twin and reliable enhanced train-centric communication-based train control
platform, which will involve actual channel simulator and system,” Computing, pp. 1–20, 2022.
[18] S. Kim, Y. Won, I.-H. Park, Y. Eun, and K.-J. Park, “Cyber-physical
sensors, to obtain the actual data, and compare the simulation vulnerability analysis of communication-based train control,” IEEE
results with the actual data. What is more, the data-based Internet of Things Journal, vol. 6, no. 4, pp. 6353–6362, 2019.
 11

[19] H. Song and E. Schnieder, “Availability and performance analysis Shigen Gao received the B.S. degree in Electrical
of train-to-train data communication system,” IEEE Transactions on Engineering and Automation from Tianjin Univer-
Intelligent Transportation Systems, vol. 20, no. 7, pp. 2786–2795, 2019. sity of Technology in 2009, and the Ph.D. degree
[20] T. Zhao, E. Yurtsever, J. Paulson, and G. Rizzoni, “Formal certification in Traffic Information Engineering and Control from
methods for automated vehicle safety assessment,” IEEE Transactions the State Key Laboratory of Rail Traffic Control and
on Intelligent Vehicles, pp. 1–18, 2022. Safety, Beijing Jiaotong University in 2016.
[21] P. Gurnı́k, “Next generation train control (ngtc): More effective railways He is currently an Associate Professor with the
through the convergence of main-line and urban train control systems,” State Key Laboratory of Rail Traffic Control and
Transportation Research Procedia, vol. 14, pp. 1855–1864, 2016. Safety, Beijing Jiaotong University. His primary
[22] Y. Wang, Y. Yan, T. Shen, S. Bai, J. Hu, L. Xu, and G. Yin, “An event- research interest is nonlinear control.
triggered scheme for state estimation of preceding vehicles under con-
nected vehicle environment,” IEEE Transactions on Intelligent Vehicles,
pp. 1–10, 2022.
[23] H. Song and E. Schnieder, “Development and evaluation procedure of
the train-centric communication-based system,” IEEE Transactions on
Vehicular Technology, vol. 68, no. 3, pp. 2035–2043, 2018.
[24] W. Jiang, S. Chen, B. Cai, J. Wang, W. ShangGuan, and C. Rizos,
“A multi-sensor positioning method-based train localization system for
Yidong Li (Senior Member, IEEE) received the
low density line,” IEEE Transactions on Vehicular Technology, vol. 67,
B.Eng. degree in electrical and electronic engineer-
no. 11, pp. 10 425–10 437, 2018.
ing from Beijing Jiaotong University, Beijing, China,
[25] Y. Xiong, Y. Zhang, X. Guo, C. Wang, C. Shen, J. Li, J. Tang, and
in 2003, and the M.Sc. and Ph.D. degrees in com-
J. Liu, “Seamless global positioning system/inertial navigation system
puter science from The University of Adelaide, Ade-
navigation method based on square-root cubature kalman filter and
laide, SA, Australia, in 2006 and 2010, respectively.
random forest regression,” Review of Scientific Instruments, vol. 90,
He is currently a Professor with the School of Com-
no. 1, p. 015101, 2019.
puter and Information Technology, Beijing Jiaotong
[26] A. Renner, H. Wind, and O. Sawodny, “Loosely coupled integrated
University. His current research inter- ests include
navigation system for excavators,” in 2020 European Control Conference
artificial intelligence, privacy-preserving data min-
(ECC). IEEE, 2020, pp. 1192–1197.
ing, intelligent transportation, and high- performance
[27] N. Hassan, S. Gillani, E. Ahmed, I. Yaqoob, and M. Imran, “The role of
computing. He has published over 60 papers in international journals and
edge computing in internet of things,” IEEE communications magazine,
conferences. Dr. Li serves on the program committees of over 15 international
vol. 56, no. 11, pp. 110–115, 2018.
conferences.
[28] Z. Zhang, J. Zhao, C. Huang, and L. Li, “Learning visual semantic map-
matching for loosely multi-sensor fusion localization of autonomous
vehicles,” IEEE Transactions on Intelligent Vehicles, pp. 1–1, 2022.
[29] J. Rodriguez, C. Garcia, A. Mora, S. A. Davari, J. Rodas, D. F. Valencia,
M. Elmorshedy, F. Wang, K. Zuo, L. Tarisciotti et al., “Latest advances
of model predictive control in electrical drivesłpart ii: Applications and
benchmarking with classical control methods,” IEEE Transactions on
Power Electronics, vol. 37, no. 5, pp. 5047–5061, 2021. Ling Liu received the Doctor of Engineering degree
[30] Z. Zuo, X. Yang, Z. Li, Y. Wang, Q. Han, L. Wang, and X. Luo, in electronic and information engineering at Peking
“Mpc-based cooperative control strategy of path planning and trajectory University, Beijing, China. He is a professor-level
tracking for intelligent vehicles,” IEEE Transactions on Intelligent senior engineer and Vice Director of the General
Vehicles, vol. 6, no. 3, pp. 513–522, 2021. Technology Research Institute, National Railway
[31] F. Mohseni, E. Frisk, and L. Nielsen, “Distributed cooperative mpc for Research and Design Institute of Signal and Com-
autonomous driving in different traffic scenarios,” IEEE Transactions on munication, Beijing, China. He has been involved
Intelligent Vehicles, vol. 6, no. 2, pp. 299–309, 2021. with the design and implementation of several mod-
[32] J. Zhan, Z. Ma, and L. Zhang, “Data-driven modeling and distributed els for onboard systems of Chinese Train Control
predictive control of mixed vehicle platoons,” IEEE Transactions on System (CTCS), which has been widely used in
Intelligent Vehicles, pp. 1–11, 2022. China high-speed railway. He has authored several
[33] J. Nie, J. Yan, H. Yin, L. Ren, and Q. Meng, “A multimodality fusion papers and patents on CTCS and related technologies. His research interests
deep neural network and safety test strategy for intelligent vehicles,” include system design of train control system, safety related software tech-
IEEE Transactions on Intelligent Vehicles, vol. 6, no. 2, pp. 310–322, nology, and intelligent technologies and applications.
2021.

Haifeng Song (M’19) received his B.Sc. degree in

2011 from Beijing Jiaotong University (China), the Hairong Dong (M’12-SM’12) earned her Ph.D.
Master degree in Traffic Information Engineering degree from Peking University in 2002. She is
and Control from the same university in 2014. He the deputy director of the National Engineering
received his Dr.-Ing degree from Technische Univer- Research Center for Rail Transportation Operation
sität Braunschweig, Germany in 2018. He has joined Control Systems and a professor in the State Key
the Institute for Traffic Safety and Automation Engi- Laboratory of Rail Traffic Control and Safety, Bei-
neering, Technische Universität Braunschweig (Ger- jing Jiaotong University, Beijing, 100044, China.
many) as a visiting scientist since 2014. He is an Her research interests include intelligent transporta-
Associate Professor with the School of Electronic tion systems, automatic train operation, intelligent
and Information Engineering, Beihang University, dispatching, and complex network applications. She
China. serves as an associate editor for IEEE Transactions
He specializes in safety and security of transportation systems, his current on Intelligent Transportation Systems, IEEE Intelligent Transportation Sys-
research interests include railway control system, formal method, intelligent tems Magazine, and Journal of Intelligent and Robotic Systems. She is a
control, and transportation modeling. He has been involved in several national fellow of the Chinese Automation Congress, a Senior Member of IEEE, and
and international research projects dealing with system safety and system eval- co-chair of the IEEE Intelligent Transportation Systems Society Technical
uation. He is a member of IEEE Intelligent Transportation Systems Society, Committee on Railroad Systems and Applications.
Chinese Association of Automation, China Institute of Communications and
a reviewer for international journals.

View publication stats