ARTIFICIAL INTELLIGENCE

Artificial intelligence for railroads:

Homepageveröffentlichung unbefristet genehmigt für Reuschling und DLR / Rechte für einzelne Downloads und Ausdrucke für Besucher der Seiten
Potential and challenges for application
Artificial intelligence methods have great potential, but there are also many
challenges. The following article analyzes the current situation.

1. Introduction be used in these three areas and subdivid- Dr.-Ing. Jannis Sinnemann
ed into categories according to their appli- Westfälische Lokomotiv-Fabrik
Artificial intelligence (AI) and machine cation as shown in Figure 1. Reuschling GmbH & Co. KG
learning (ML) methods should be imple- Innovation Manager
mented more efficiently and attractively. 3. Using AI for infrastructure [email protected]
Application is multifaceted and has great
potential, but is repeatedly a topic of discus- Rail transport is an overall system en-
sion. Not long ago, there was a call for tech compassing vehicles, processes, and in-
experts to pause development (see [1]). But frastructure. The great spatial extent of Dr.-Ing. Julian Franzen
that call was contradicted by a number of rail networks and the large number of Westfälische Lokomotiv-Fabrik
Reuschling GmbH & Co. KG
people, including computing science pro- infrastructure elements (tracks, switches,
Head of Innovation
fessor Johanna Börklund at Umeå Universi- signals, etc.) entail correspondingly large
[email protected]
ty in Sweden, who said, “There is no reason maintenance cost and effort.
to hit the brakes.” Instead, transparency re- AI methodologies have the potential to
quirements for developers should become pave the way to condition-based or predic-
more stringent. Safety standards are neces- tive infrastructure maintenance, reducing
Dr.-Ing. Christian Meirich
sary to create rules for the use of the tech- maintenance costs [3]. The key is to detect
Deutsches Zentrum für Luft- und
nology [1]. The article below will consider infrastructure element condition using Raumfahrt e. V. (DLR), Institut für
potential AI fields of application in railroads field data and estimate its development, Verkehrssystemtechnik
and the areas in which AI is already being especially wear, as precisely as possible. Railway Operations Group Leader
used. It will then analyze the open fields of Maintenance measures with fixed and care- in the Design and Assessment
Department in Mobility Solutions
action and the obstacles to developing this fully chosen intervals (such as replacement
[email protected]
potential to its fullest. after a given date) can thus be made more
flexible and linked to condition (such as re-
Dr. Benjamin Baasch
2. State of the art, potential, placement when failure is imminent).
Deutsches Zentrum für Luft- und
and challenges AI requires data to work. Infrastructure Raumfahrt e. V. (DLR), Institut für

genehmigt / © DVV Media Group GmbH

component data sets and associated wear Verkehrssystemtechnik
An extensive overview of the use of AI forecasts must be generated continuously Research assistant in the Informa-
methods in the transport sector can be from sensors installed either in the infra- tion Acquisition and Modelling
Department
found in “A literature review of Artificial In- structure or the vehicle. sensors on in-ser-
[email protected]
telligence applications in railway systems” vice vehicles is a promising approach to
(Tang et. al.) [2], which analyzes about 140 continuous monitoring of the entire track
scientific articles on AI published between network [4]. Repeated transit under similar Dr. Michael Roth
2010 and 2020 and evaluates their applica- conditions is especially valuable for compa- Deutsches Zentrum für Luft- und
tion. rability and development of infrastructure Raumfahrt e. V. (DLR), Institut für
Verkehrssystemtechnik
This evaluation clearly shows that the conditions. The large quantities of data
Research assistant in the Informa-
primary application in rail transport is cur- thus accumulated require automated as- tion Acquisition and Modelling
rently maintenance and inspection. Other sessment. This can be done by supervised Department
sub-domains are safety and security, au- and unsupervised AI (with or without the [email protected]
tonomous driving and control, planning help of reference data). Unsupervised
and management, and passenger flow pre- learning are aimed at detecting anomalies
diction. and thus faults in the infrastructure. Super-
Railway application areas fall roughly vised methods can use reference data to
into the infrastructure, operation, and ve- deliver qualitative and quantitative infor-
hicle categories. Existing AI algorithms can mation about such items as track condition

22 ETR  |  International Edition  |  2023 www.eurailpress.de/etr

ARTIFICIAL INTELLIGENCE

or track fault classification. Vibration data

provide especially relevant information. Ex-

Homepageveröffentlichung unbefristet genehmigt für Reuschling und DLR / Rechte für einzelne Downloads und Ausdrucke für Besucher der Seiten
amples are axle bearing acceleration to de-
termine track geometry or defects [5] and
camera images for automated detection of
visible faults in the track infrastructure [6].
Position information (from satellite naviga-
tion and inertial sensors) are important for
establishing the link to the infrastructure
transited [7].
Track-side sensor data can provide con-
tinuous monitoring of especially critical in-
frastructure elements such as switches. For
instance, AI can analyze switching curve
waveforms to detect defects [8]. Vehicle Operation
4. AI used in rail vehicles

Across a rail vehicle’s life cycle, AI will be

used in various phases with current activi-
ty focused on maintenance and operation. Infrastructure
Some of these activities will be considered
below.
In the area of vehicle use, AI constitutes
a key technology for automating driving
functions.
For automation, including partial au-
tomation, of vehicle functions, AI meth-
ods for monitoring vehicle environment
predominate. They use machine vision
methods to identify relevant objects with 1: Potential fields of application in the railway area and application categories for AI algorithms
camera, lidar, radar, and ultrasonic sensors
in the monitored environment. Depending
on the type of detected object, this infor-
mation is made available to the vehicle op-
erator in lightly automated processes (see simplify the associated processes and make ation. So AI should be used in scheduling and
[9]) or used for decision-making during au- them more reliable. If the dimensions of disposition to identify potential or actual con-
tonomous driving (see [10, 11]). components at hard-to-reach places on the flicts and prevent or correct them and mini-
Much like predictive maintenance for vehicle required, they can be detected at a mize delays that arise as quickly as possible
infrastructure components, vehicle and ve- spoken command and transferred to the and avoid new delays or delay transfer.
hicle component condition monitoring has maintenance documentation. Machine vi- AI-supported processes for predictive

genehmigt / © DVV Media Group GmbH

tremendous potential for increasing availa- sion also provides the option of using aug- travel can also prevent unnecessary acceler-
bility, cost efficiency, and safety for rail ser- mented reality, such as the use of a smart ation and braking. The speed harmonization
vice provision. Vehicles can be equipped device to deliver information that will sup- thus achieved has a positive effect on delays
with sensors to produce large quantities of port inspection or repair processes [14]. by reducing waiting times and enhancing the
data for analysis with AI methods to deter- infrastructure’s capacitive loads (such as [15]).
mine a vehicle condition forecast. Depend- 5. AI applications in railway operations Machine learning processes are also used
ing on application, cloud or edge applica- for passenger information such as forecast
tions are used [12]. Artificial intelligence is also used for railway processes based on past real-time data for
The core of data analysis is systems for operations. At the moment, scheduling more reliable arrival and departure times and
monitoring rotating components such as and disposition is the focus of research and train and station capacity usage.
wheelsets and wheelset bearing compo- development. Sub-categories here are stra- Most AI methodologies used in rail-
nents based on acceleration and vibration tegic and tactical planning, analysis of com- way operations are formulated as op-
data (see [13]). Newer approaches that use pleted transport, and dispositive resched- timization problems based on the fact
knowledge of vehicle component condi- uling following delays or schedule changes that they are applied to problems that
tions to achieve resilient operation are cat- to enhance operational efficiency (see [2]). are currently very difficult or impossible
egorized as prescriptive maintenance. A major advantage of AI-supported to solve or greatly increase the size of the
For performing the maintenance itself, methods in railway operations is that they described problem. But they are primarily
AI offers the potential of using language can identify sufficient time gaps when there used for data-based forecast approaches
recognition and generation methods to are delays and thus facilitate predictive oper- (passenger forecasts or conflict avoid-

www.eurailpress.de/etr ETR  |  International Edition  |  2023 23

 ARTIFICIAL INTELLIGENCE

ance) that use historical data to generate But the behavior of various actors (2019).
better forecasts for the future. ([16], for example) indicates efforts toward [4] Alten, K.; Fuchs, A.; Struhar, J.; Baasch, B.; Heusel, J.;

Homepageveröffentlichung unbefristet genehmigt für Reuschling und DLR / Rechte für einzelne Downloads und Ausdrucke für Besucher der Seiten
Groos, J.: Zustandsüberwachung mit Regelfahrzeugen.
progress.
In: Eisenbahn Ingenieur Kompendium, 2017,
6. Assessment The AI methodologies used in the con- S. 111–127.
text of railway operations are largely in the [5] Niebling, J.; Baasch, B.; Kruspe, A.: Analysis of Railway
For detecting the condition of vehicle com- testing and trial phase. They are being used Track Irregularities with Convolutional Autoencoders
ponents and infrastructure, it is clear that primarily for non-safety-critical forecasts and Clustering Algorithms, Bd. 1279. In: Bernardi, S.;
Vittorini, V.; Flammini, F.; Nardone, R.; Marrone, S.; Adler,
one challenge of moving applications from and passenger information, such as arriv-
R.; Schneider, D.; Schleiß, P.; Nostro, N.; Løvenstein Olsen,
research to practice is the low availabili- al and departure times, and for train and R.; Di Salle, A.; Masci, P. (Hrsg.): Dependable Computing
ty of data, especially for existing vehicles. station capacity usage [14]). Another focus - EDCC 2020 Workshops. Cham: Springer International
One problem is the high cost of data gen- of AI trials is the solution of optimization Publishing, 2020 (Communications in Computer and
eration solutions. Another is the lack of IT problems having largely to do with analyz- Information Science), S. 78–89.
[6] Jahan, K.; Umesh, J. P.; Roth, M.: Anomaly Detection
capacity on existing vehicles. This results in ing historical traffic data. It is relatively early
on the Rail Lines Using Semantic Segmentation and
a challenge for broader application, espe- going in the implementation and testing of Self-supervised Learning. In: 2021 IEEE Symposium
cially for adapting AI approaches that can applications in operational conflict avoid- Series on Computational Intelligence (SSCI) : IEEE, 2021,
manage with little data (few-shot learning, ance as part of rescheduling after opera- S. 1–7.
for example), and for implementing edge tional disruptions and in strategic and tac- [7] Roth, M.; Baasch, B.; Havrila, P.; Groos, J.: Map-Sup-
ported Positioning Enables In-Service Condition
concepts for partially decentralized and, as tical network planning aimed at improving
Monitoring of Railway Tracks. In: 2018 21st Internation-
necessary, energy-optimized data analysis performance. After the assessment of [2], al Conference on Information Fusion : 10-13 July 2018,
in existing vehicles. AI methodologies have so far been used Cambridge, 2018, S. 2346–2353.
This is also true of AI-supported railway only for decision-making and support for [8] Guzman, D. N.; Hadzic, E. Baasch, B.; Heusel, J.; Neu-
infrastructure maintenance. AI results must dispatchers and planners. mann, T.; Schrijver, G.; Buursma, D.; Groos, J. C.: Anomaly
Detection and Forecasting Methods Applied to Point
be robust, and the approach must be com-
Machine Monitoring Data for Prevention of Railway
pared with existing ones. This cannot be 7. Summary Switch Failures, Bd. 166. In: Ball; Vyas (Hrsg.): Advances
managed in some places, especially in new in Asset Management and Condition Monitoring. [S.l.]:
applications that have been made possible It is clear that AI methods have a great deal Springer International Publishing, 2020 (Smart Innova-
by AI in the first place. It is important to to contribute to the solution of relevant tion, Systems and Technologies), S. 307–318.
[9] Samira: Forschungsprojekt. URL https://samira-rangi-
give AI and other data-driven approaches a problems. Realistic expectation manage-
er-assistent.de/. – Aktualisierungsdatum: 2023-04-26.
chance to make a contribution. Preparation ment for AI solutions is important for a pro- [10] Rangierterminal4.0: Forschungsprojekt. URL https://
should also be made to make data availa- ductive compromise between the extreme verkehrsforschung.dlr.de/de/projekte/rangiertermi-
bility (with sustainable data management, “AI can solve everything” and “AI is unusable” nal40. – Aktualisierungsdatum: 2023-04-26.
for instance). positions. AI tools should therefore not be [11] Sensors4Rail: Forschungsprojekt. URL https://digi-
tale-schiene-deutschland.de/Sensors4Rail. – Aktualisi-
For environment motoring in automat- considered in isolation, but approached as
erungsdatum: 2023-04-26.
ed driving, computer vision algorithms an effective tool that can harmonize with [12] Bohnet, J.: Bahnverkehr: Thank you for traveling
should be considered. A significant obsta- other methodologies. This secures a con- with AI. URL https://safe-intelligence.fraunhofer.de/
cle to high degrees of automation is cur- nection to established procedures and en- artikel/bahnverkehr-thank-you-for-traveling-with-ai –
rently less the availability of sensors and hances acceptance by stakeholders. It is also Überprüfungsdatum 2023-04-26.
[13] Baasch, B.; Heusel, J.; Roth, M.; Neumann, T.: Train
assessment than the monitoring of the clear that there is a need of guidelines and
Wheel Condition Monitoring via Cepstral Analysis of
unmanned driver’s cab and the approval of procedures for approving AI algorithms, es- Axle Box Accelerations. In: Applied Sciences 11 (2021),
the relevant systems for safety-relevant ap- pecially in safety-critical applications. Since Nr. 4, S. 1432.
plications because of such concerns as vari- the capabilities and reliability of AI methods [14] Siemens Mobility GmbH: Künstliche Intelligenz

genehmigt / © DVV Media Group GmbH

ous operating scenarios (harbors, industrial will continue to grow in the future, it is im- im Schienenverkehr: SBB Cargo und Siemens Mo-
bility bringen künstliche Intelligenz auf die Schiene.
parks, shunting yards, etc.). portant to start preparing for their use by
URL https://www.mobility.siemens.com/ch/de/
initiating suitable approval procedures.  unternehmen/newsroom/news-archiv/kuenstliche-in-
telligenz-im-schienenverkehr-von-sbb-cargo.html. –
Aktualisierungsdatum: 2023-04-26.
[15] Meirich, C.: Berechnung und Bewertung der Ge­
samtleistungsfähigkeit von Eisenbahnnetzen. Aachen,
Since the capabilities and Literature
Universitätsbibliothek der RWTH Aachen. Dissertation.
2017.
reliability of AI methods will [1] Schreiber, S.: Experten fordern Pause bei KI-Entwick- [16] Siemens Mobility GmbH: Wir stehen in den
Startlöchern. URL https://www.mobility.siemens.com/
lung. URL https://www.tagesschau.de/wissen/musk-
continue to grow in the future, tech-pause-ki-entwicklung-101.html. – Aktualisierungs- global/de/portfolio/schiene/storys/wir-stehen-in-den-
startloechern.html – Überprüfungsdatum 2023-04-26.
datum: 2023-04-26 .
it is important to start preparing [2] Tang, R.; Donato, L. de; Besinović, N.; Flammini, F.;
Goverde, R. M.P.; Lin, Z.; Liu, R.; Tang, T.; Vittorini, V.; Wang,
for their use by initiating suitable Z.: A literature review of Artificial Intelligence applica-
tions in railway systems. In: Transportation Research
approval procedures Part C: Emerging Technologies 140 (2022), S. 103679.
[3] Kornfeld, N.; Luber, A.; Leich, A.; Kaiser, M.; Schubert,
L.; Groos, J.: Zustandsüberwachung der Bahninfrastruk-
tur mit KI. In: Eisenbahn Ingenieur Kompendium 2020

24 ETR  |  International Edition  |  2023 www.eurailpress.de/etr