=Paper=
{{Paper
|id=Vol-2846/paper2
|storemode=property
|title=Anomaly Detection in Railway Infrastructure
|pdfUrl=https://ceur-ws.org/Vol-2846/paper2.pdf
|volume=Vol-2846
|authors=David Morandi,Stephan Jüngling
|dblpUrl=https://dblp.org/rec/conf/aaaiss/MorandiJ21
}}
==Anomaly Detection in Railway Infrastructure==
https://ceur-ws.org/Vol-2846/paper2.pdf
Anomaly Detection in Railway Infrastructure
David Morandi and Stephan Jüngling
FHNW University of Applied Sciences Northwestern Switzerland, School of Business, Peter Merian-Strasse 86,
4052 Basel, Switzerland,
Abstract
In order to keep complex railway systems fail-safe, sophisticated maintenance of the rolling
stock and infrastructure are most essential. Although AI-based predictive maintenance systems
exist in many different industries, there is still quite large potential for different application
scenarios. The current research shows such an example, where machine learning can be applied
to detect anomalies in the pantograph-catenary system by using a simple convolutional neural
network that is able to detect arc ignitions during train operation. The paper provides some
insights into the process of the system development life cycle. Starting from the initial idea to
use machine learning for anomaly detection, over the system design of a prototype and the
training of the Keras-based machine-learning model, up until the evaluation of the conducted
experiments. The arcVision system prototype provides valuable insights into how a predictive
maintenance process could be established by combining the results from the machine-learning
model with rules and insights from manual inspections.
Keywords 1
Anomaly Detection, Railway Infrastructure, Predictive Maintenance, Arc Ignition, Machine
Learning, CNN, Keras, System Development Life Cycle, Knowledge Engineering
1. Introduction
For many large technical systems, outages are either not tolerable or too costly. Therefore, many
traditional processes exist to maintain complex systems to find anomalies and fix potential defects in
advance. In most of these traditional cases, we use best practice rules, most of them implementing
maintenance intervals, based on parameters such as time, frequency or intensity of use. However, lately
and specifically in the context of the advances in machine learning, many novel types of predictive
analytic solutions are used to further optimize the traditional, mostly time-based maintenance cycles.
Predictive Maintenance (PdM) is one of the fields where AI-based prognostics and health management
(PHM) could successfully be implemented in various industries such as construction-, automotive-,
steel-, or in aeronautics and logistics. Furthermore, with the introduction of IoT, smart factories and
smart cities in the context of Industry 4.0, standardized reference models such as RAMI 4.0 [1]
(Reference Architecture Model Industry) are introduced, which provide guidelines for the
implementation of predictive maintenance systems as described in the use case of aeronautics supply
chains industries [2]. However, a recent systematic literature review of academic papers from the past
five years about PdM by Dalzochio et al. [3] also collected existing challenges in applications of many
different machine-learning techniques, while pointing out a high demand in further research
investigations in the area of applying machine learning and reasoning in the context of Industry 4.0.
Despite the railway industry being an old and traditional industry, digitalization and the potential
of predictive maintenance will not spare this industry. The complex railway system with its security
In A. Martin, K. Hinkelmann, H.-G. Fill, A. Gerber, D. Lenat, R. Stolle, F. van Harmelen (Eds.), Proceedings of the AAAI 2021 Spring
Symposium on Combining Machine Learning and Knowledge Engineering (AAAI-MAKE 2021) - Stanford University, Palo Alto, California,
USA, March 22-24, 2021.
E-mail: david@morandi.me, stephan.juengling@fhnw.ch
© 2021 Copyright for this paper by its authors.
Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR Workshop Proceedings (CEUR-WS.org)
measures, which need to be fail-safe and the sophisticated maintenance of the rolling stock and
infrastructure are most essential. For example, the European Union [4] has the aim to increase the rail
competitiveness and pursues specific goals such as decreasing the life-cycle cost of railways by 50%,
doubling the capacity and increasing the punctuality by 50% achieved by digitalization. It is important
to prevent incidents as efficient as possible, since even a minor local interruption can quickly evolve to
a major disturbance in the operation of the overall railway system. Therefore, the vision for the
conducted master thesis [5] has been created to identify a potential use case for anomaly detection in
the railway industry using artificial intelligence.
2. Current State of Applications of AI in the Railway Industry
Whenever safety plays an important role, and the resulting cost of manual maintenance effort is high,
there is a potential to apply PdM using automatic anomaly detection. Nevertheless, the possible business
cases, where AI could be used, to overall reduce the amount of manual inspection work are manifold,
and therefore triggered the first two research questions:
• RQ1: Which area contains the biggest potential for anomaly detection in railway
infrastructure?
• RQ2: What is the current state of applications of anomaly detection in railway
infrastructure?
The answer to these two questions can be derived with a literature research. One of the main aspects,
when talking about inspection of railway systems is the borderline of manual work done by humans
versus inspections that can be automated with the help of AI-based systems. Al-Douri et al. [6] state
that railway infrastructure is very complex, covers a large area and consists of many subcomponents,
which are often difficult to maintain. Numerous stakeholders, weather conditions, physical components
as well as administration, traffic situations and new investments need to be considered for the
maintenance of the system. The infrastructure can be further split into technical subsystems, such as the
substructure, track, electrical system, signaling system, and telecom system.
Gibert et al. [7] state that periodical inspection and monitoring is needed to ensure safe
transportation. Safety can be improved by more frequent inspections, reducing human errors and
automating line inspections using computer vision and pattern recognition methods. In contrast, full
automation cannot yet be implemented due to the number of different possible error modes and the
wide range of image variations that can potentially trigger false alarms. In addition, the number of faulty
components is very small, so that to little training data is available for the machine to train a robust
visual anomaly detector. Siemens Ltd. [8] has a system, which detects broken rails. They equip trains
with sensors on one of the bogies. During operations, the sensors scan the tracks and send them to the
control center. If a broken rail is detected, an alert is sent to the subsequent trains and the line can be
blocked by the control center or the existing signaling solution.
Switches are as well critical sub-systems in the railway infrastructure, as they, if malfunctioning can
lead to serious accidents. Either the train can derail or switch to the wrong line. Guzman et al. [9]
developed an anomaly detection model and integrated it in an intelligent workflow, which is used in an
operational environment. The model can calculate an anomaly score which is then aggregated with
meta-data of the corresponding switch (e.g., last maintenance, number of failures) and can take context
information such as different weather conditions or other environmental factors into account. However,
they concluded that the model could be further improved by using more features.
There is a wide range of potential areas for anomaly detection in railway infrastructure, and not only
the rails and switches, but also defects in the catenary system should be considered. One of the key
factors that influence the operation quality of high-speed trains is the electric contact between the
pantograph and the catenary system, as stated by Wu et al. [10]. Combining this fact with the insights
from conducted expert interviews that the manual inspections of rails or switches are much better
accessible for humans than the inspection of the catenary system, the idea of narrowing down the scope
to anomaly detection in the pantograph-catenary system gave the idea for the thesis statement for the
conducted master thesis of Morandi [5]. The thesis hypothesis states that it is possible to design a
prototype (arcVision System) which makes use of machine learning techniques, with the goal to help
detecting arcing in the pantograph-catenary system. According to Wu et al. [10], arcs generate areas of
high temperature on the strips, which causes the material to melt and evaporate, resulting in erosion of
the wire and strip or high thermal temperature gradients, which cause thermal stress and potential
breakage of the strip material. Predictive maintenance processes could be established with the help of
digitally recorded sensor data that detect arcing during train operation. With the help of supervised
learning, an arc detecting system could be trained, that helps to build up a predictive maintenance
system for the pantograph catenary system.
3. Business Case of the arcVision System
With almost 29’000 kilometers, Switzerland has one of the densest public transport networks in Europe
[11]. In consequence, a dense timetable, professional personnel and reliable rolling stock/infrastructure
is needed to meet this high demand and to ensure fast and safe transportation services. Despite its dense
regular interval timetables, Swiss railways are able to use their infrastructure in an optimized way.
Faulty infrastructure or rolling stock can cause delays, and even minor disturbances could provoke
major disruption to the entire timetable.
There are many causes for train delays. Some of them are predictable and some of them are yet
unpredictable. Regular maintenance can minimize the risk of delays and accidents. However, periodic
maintenance programs are time-consuming and cause unwanted idle times. Furthermore, it is important
to identify the source and root cause of the defects in order to determine appropriate maintenance
activities. Attrition or fatigue of material is one of the main causes for such maintenance.
During maintenance of the rolling stock, train parts, which have passed their operational lifetime
and further parts, which show more attrition of material than expected prior to the expiry of the
guarantee period are replaced. These procedures can be very time and resource intensive. To prevent
additional defects and maintenance, it would be helpful to understand the source of such attrition and
to localize the root causes. With the help of sensors, it is possible to collect plenty of data. However,
manual extraction, evaluation and anomaly detection of the collected sensor data is a highly time-
consuming process. In this case, artificial intelligence could help to minimize the time needed for
anomaly detection. After evaluating the data, the source of attrition can be localized and possibly
increase the operational lifetime of the affected parts on one hand and reduce time and effort for
maintenance on the other.
In order to localize the ignition from the electric spark in the pantograph-catenary system during
train operations, the time of the occurrence, the location (e.g. in a tunnel), the acceleration of the train
(e.g. mechanical forces) or further context information (e.g. weather condition) could have an influence.
As described by Wu et al. [10] the electrical contact between the pantograph and the catenary is rather
complex and variable. The operating performance of the electrical contact system of pantograph and
catenary depends on the contact resistance, contact surface heat, friction and wear. If arcing in the
pantograph and the catenary contact occurs regularly, it can cause unusual wear and even breakage of
the graphite strips, which leads to malfunction power supply respectively of the train. Therefore, it
would be helpful to know all these parameters mentioned above in order to determine why arcs occur
and where the catenary system should be inspected. Frequent arc ignition will accelerate attrition and
faster loss of material results in more frequent maintenance or malfunctioning, which eventually causes
a higher idle time or an interruption in operations.
The conducted literature review has revealed that until now, only limited research has been
conducted on automatic detection of anomalies in the pantograph catenary system using machine
learning techniques. Consequently, the feasibility of such an anomaly detection raised the following
two additional research questions:
• RQ3: Is it possible to design a prototype (arcVision System) which makes use of machine
learning techniques, with the goal to help detecting arcing in the pantograph-catenary
system?
• RQ4: To what extent can arcVision assist a human inspector and where are its limitations?
4. Prototype Design and Architecture – arcVision System
For the envisioned prototype, a supervised learning approach was chosen, for which large amounts of
test data needs to be recorded and labeled. The arcVision System prototype was developed in
cooperation with Regionalverkehr Bern-Solothurn (RBS), a Swiss public transport company in the
region of Bern and Solothurn. RBS operates a fleet of 49 trains, a train network of 53.9 kilometers and
the corresponding infrastructure. The arcVision System contains two different parts. The arcVision
Scanner, which collects the necessary sensor and context data. The arcVision Model consists of a
Convolutional Neural Network (CNN) for the training and arc detection of the recorded pictures from
arcVision Scanner to do a binary classification. Figure 1 shows the envisioned architecture of the
system.
Figure 1: arcVision System Architecture
During train operation, data is collected with the arcVision Scanner. It consists of a single board
computer, a camera and different sensors such as GPS, accelerator, barometer, thermometer and
hygrometer and is mounted on the train’s roof. The camera’s field of view needs to be aligned with the
touching point of the pantograph and the catenary. The collected video sequences could then be used
by the arcVision Model, with an adequate pre-processing such as image extraction and the
corresponding labeling of the images.
In a second step, the video sequence is divided into single frames. As no pre-trained model or pre-
labelled data set can be used at the beginning, a screening of all video/image data is needed to search
for potential pictures with arc ignitions to be labelled. The filename of the images and the corresponding
label is then stored in a file-based database. The high effort for labeling the pictures is well known in
literature [12], and different techniques such as data subset selection and active labeling [13] or self-
supervised learning [14] have been developed, to reduce the labeling effort. The sparsity of the pictures
with arcing turned out to be only 0.02% and the task resembled to the well-known situation of finding
the needle in a haystack. However, due to an iterative approach, where the first arc picture could have
been found, the initially badly trained model could never the less help to find further candidates of
pictures with arc ignitions.
The arcVision Model is a CNN, based on Keras and TensorFlow, which allowed for a very fast
experimentation with the before-mentioned iterative approach, where the entire dataset could be
labeled, and the model be trained more effectively. The training was performed on a regular personal
computer with a dedicated GPU, which turned out to have sufficient computing power for the machine
learning task at hand.
During the fourth step, the pre-trained model was then used to predict new data sets that were
collected in subsequent test-runs. The data collection is described in further detail in the next chapter.
In the very last step, the images, which showed arc ignition can be combined with the additional
sensor data (e.g. temperature, GPS data) where additional context data can be recorded and analyzed
for the anomaly detection process, which in the end, allows domain experts to draw conclusions for the
PdM tasks.
4.1. Data Collection – arcVision Scanner
As mentioned before, to collect the necessary data (video and sensor), a corresponding device such as
the arcVision Scanner was designed during the master thesis. It consists of a Single Board Computer
(Raspberry Pi 4 Model B), a GPS receiver, a Raspberry Pi camera module and external temperature and
humidity sensors.
As can be seen in figure 2, the prototype device was designed iteratively until the final packaging
and mounting box was ready that resisted the various weather conditions. The sensitive electronic parts
needed to be protected not only from rain, but also from fine dust. On the other hand, a continuous
power supply is necessary, which can be taken most of the time from the train via the compressor
control system. Since the train is in some situations decoupled from the catenary system, an additional
power bank is used that can provide power for up to 40 hours. Furthermore, different adjustments were
necessary to find the optimal angle and distance to the pantograph, in order to cover the entire contact
range with the catenary system, where potential arcs can be detected. With the help of a Python script,
all sensor data is stored in a file-based database.
Figure 2: arcVision Scanner (a) Temporary mount during first field test, (b) Packaging during first field
test (without external power supply), (c) Final mount box, (d) Final packaging, (e) Final mounting
4.2. Model Creation – arcVision Model
For the model creation and training, a CNN model, based on the frameworks TensorFlow, Keras and
several other state-of-the-art tools such as numPy, scikit-learn, pandas, and opencv was used. Keras
provides a deep learning API, running on top of the machine learning platform TensorFlow. When
developing Keras, the focus was laid on allowing fast experimentation, being able to go from the idea
to result as fast as possible, which is visualized and suggested by the Keras Special Interest Group [15]
and shown in figure 3.
With increasing popularity of machine learning, many different pre-trained models exist. A recent
survey pointing out the importance of transfer learning analysed the different approaches how existing
pre-trained CNN networks can be re-used [16]. However, a pre-trained model (e.g., one for lightning
detection), which would be potentially able to recognize arc ignition, could not be found. Furthermore,
in our specific case of detecting arcs, the advantage of transfer learning, where lots of labelling effort
and CPU hours from previous trainings could be reused to extend large sets of existing classes by this
additional class, does not provide any advantage. Therefore, the CNN for the arcVision Model was built
from scratch.
Figure 3: - The loop of progress, from Keras Special Interest Group [15]
A sequential Keras model was created and its layers are visualized using
TensorBoard (see figure 4) as described in the loop of progress in figure 3.
The model follows suggestions from O’Shea & Nash [17] and uses a plain
stack of alternating convolutional and pooling layers followed by a flattening
and dropout layer in the end before eventually feeding to the dense, fully
connected layer.
The training and validation of the model were performed on PC with a
decent graphics card (Nvidia GeForce GTX 980 Ti) and needed at least
500GB free space (SSD). Despite slightly outdated hardware, the task could
be still performed successfully. During training, the model analysed 1’319
images with arc ignition and 4’194 images without arc ignition (total 5’513).
As the numbers show, the classes were quite imbalanced. However, as stated
by Kotsiantis et al. [18] such high imbalances occur in different real-world
domains such as detecting oil spills in satellite radar images or detecting
fraudulent telephone calls, where the aim is to detect an occasional but
important case or event.
Figure 4:
Model Structure
4.3. Experiments
As stated in [19], the characteristic system development life cycle (SDLC) methods from projects with
pure classical software development tasks versus projects that include applications of machine learning
might be different not only in terms of the skillset of the engineers, but also in terms of the suitability
of the applied SDLC methodology. When talking about software engineering, test-driven development
is considered as one of the “best practices” of software engineering. In the case of hardware related
engineering tasks, where physical and environmental issues influence the experiments, testing is even
more essential. In software development processes, APIs (application programming interfaces) can
mimic real behavior and the code under development can be isolated from the environment. This is not
as easy for the engineering and research task of designing the arcVision Scanner prototype.
Nevertheless, partial testing can be implemented by breaking down experiments which are close to the
reality into smaller experiments, which can at least provide partial insight into facts and constraints
from the technical environment.
Figure 5: Field test maps (left: bike test, right: train)
To test the very basic functionality the first prototype iteration was equipped with a GPS receiver,
different sensors (accelerator, barometer and thermometer) and a power supply in the form of a power
bank. This test was conducted with a small bike tour, which started in the city of Bern, as can be seen
on the left of figure 5. This test revealed the necessity of having a “time to first fix” (TTFF), which is
necessary for the GPS receiver for an initial position fix. However, this first fix was achieved exactly
at the entrance of the forest (northbound), which was approximately 5 minutes after the prototype was
powered on and started. In the forest, the position could not be determined anymore, since the trees
blocked the reception of GPS data from almost every angle under or over 90°. After leaving the forest,
outlier positions were recorded. But shortly after about a minute the position could be adjusted thanks
to the environment with an open sky. The rest of the track was correctly recorded until the small power
bank ran out of power.
The second test iteration the arcVision Scanner was tested closer to real conditions and mounted on
a train. In order to be optimally prepare for this setup, the capacity of the power bank was increased
from approximately 1’000mAh to 21’400mAh. On the other hand, all devices were placed into an
enclosure in order to ensure that the components are protected against spray water, strokes or dust. The
ride was declared as a “not in service”-ride and therefore, no regular passengers were on the train. The
test route started from the Solothurn train station to Lohn-Ammannsegg and went back to Solothurn
with a total of approximately 30 minutes test time (see figure 5 – right side).
In order to collect all the necessary data in a final reality check, the arcVision Scanner was integrated
into the regular daily operation time schedule for a few days on the regional express line Bern-
Solothurn. It turned out, that the arcVision Scanner was always operational and even after rainfall, no
damage could be found. Nevertheless, the roof of a train is very filthy, so the camera lens became quite
dusty and needed to be cleaned. Nevertheless, the recorded data sets could be used for the validation
and testing of the arcVision Model.
4.4. Model Validation
The validation data set contains 200 selected images with a balanced representation of the classes, where
92 had arcs and got the label “1” and 108 with no arc the label “0”, whereas images with features and
circumstances that rarely occurred were also taken into account. The accuracy of the model was
optimized over several iterations by experiments with tuning the different training parameters such as
batch size, number of epochs or different optimizers. At some point, the accuracy was considered as
“good enough” and an accuracy of 74.5% could be achieved which is determined by the confusion
matrix in table 1.
n=200 Predicted: Positive (arc) Predicted: Negative (no arc)
Actual: 76 16
Positive (arc) (True Positives) (a) (False Positives) (b)
Actual: 35 73
Negative (no arc) (False Negatives) (c) (True Negatives) (d)
Table 1: Confusion Matrix
The data validation showed that the model is in general able to detect arc ignition with an adequate
confidence as shown by some typical cases in figure 6.
Figure 6: Typical classification samples – (a) true positives, (b) false positives, (c) false negatives, (d)
true negatives
Nevertheless, the confusion matrix indicates that the model has problems with false positives and
predicts an arc ignition where no arc ignition happened. Furthermore, the model shows a tendency of
overfitting since it was trained with a small data set and possibly learned irrelevant features.
During the validation, the authors noticed some problematic cases where the prediction was wrong,
as shown in figure 7. Especially rainy images, lamps, reflections, and small artefacts of arc ignition
were problematic. In addition, the training data was captured in a short timeframe during a few days in
autumn weather.
Figure 7: Problematic cases
Based on the conducted experiments and the model validation, the RQ3 can be answered as the follows:
Yes, it is possible to design a system, which makes use of machine learning techniques, and detects
arcing in the pantograph-catenary system.
However, the prototype needs further improvements in terms of data collection and optimization of the
model in terms of accuracy. As we have seen in figure 6, the predictions differ heavily on the quite
diverse situations and circumstances. Nevertheless, railway infrastructure domain experts evaluated and
assessed the current prototype and there was a common understanding, that the arcVision System can
provide very valuable insights in terms of finding potential defects early by analyzing the arc ignition
pattern along the railway infrastructure system and the system seems to be a valid starting point for
PdM. However, there was also a common understanding, that drawing conclusions out of such arc
ignition patterns is still a task to be conducted by human experts. The potential overall improvement
with a combination of ML and KE is addressed below. Reflecting these findings, the following chapter
further elaborates on the combination of machine learning with knowledge engineering.
5. Combining Machine Learning with Knowledge Engineering
Based on the data of the arcVision System, a map with all arc ignitions can be generated. For all these
data points, additional sensor data can complement the recorded data. Based on the current primary data
of the sensors, such as time, x/y-coordinates, and acceleration in x/y-direction, humidity, air pressure,
temperature and humidity, secondary data such as real speed and acceleration/deceleration of the train
can be calculated.
Furthermore, the perceived intensity of the arc ignition is heavily dependent on static as well as
dynamic context, as shown in figure 8. With static context, environmental situations of the train such
as entering a tunnel, distance from the tunnel entry, entering or exiting stations, at the location to a
switch/catenary split etc. can be determined and added to the dataset. Dynamic context information
such as the brightness of the sky (e.g. on cloudy or sunny days) could be added, in order to determine
the absolute intensity of the arc rather than the relative perception of the intensity, which depends on
the background of the picture. Based on a hypothesis, that the intensity of the arc caused by a defect at
a particular location in a tunnel has the same intensity for all trains could lead to the conclusion that the
intensity of an arc outside of tunnels is equal for different trains as well. Given this reproducibility, the
effect of different weather conditions (e.g. sunny/cloudy sky) on the perceived relative intensities on
the recorded pictures could be learned and the absolute intensities be calculated from the perceived,
relative intensities.
Figure 8: PdM Model
This new set of calculated secondary data could serve as new features to train a final PdM Model, which
directly calculates a measure for presumed material deterioration based on the insights from the
arcVision System.
Another option could be to combine the predictive PdM model with Case Based Reasoning (CBR),
as demonstrated by Zhao et. al. [20] for PdM of railway turnout systems, which are most essential for
the current high speed railway infrastructure in China. Suggestions from the machine-learning model
could be compared to the results from existing cases in the case database. It is left to the humans to
decide which suggestion they will follow in case of diverse results from the ML and the CBR model.
Based on the current state of insights from the ML model, which certainly needs more training data to
be improved over time, some facts, rules can already be stated:
• The most intense arc ignition events are in tunnels (e.g. in the current set of the 22 most
intense arcs in terms of brightness were 15 in a tunnel).
• Arc intensities seem to correlate with acceleration and deceleration of the train.
• The phenomenon of sparking can be observed throughout the line.
• Arc ignition happens more frequently while passing a switch, which is reasonable, since this
always corresponds with a split of the wire in the catenary system, which causes the arc due
to an air gap between the pantograph and the catenary.
However, this list of facts are currently assumption, which could not yet be proven based on the current
pictures from the experiments. However, when more data is collected, these patterns can be validated,
and rules for maintenance activities can be derived.
During the evaluation with domain experts the statement was made that the arcVision System offers
an increased reliability in detecting existing arc ignitions, which were until now, only occasionally
observed and reported by the train crew in cases when, they considered it to be important. So far,
humans were not even able to detect smaller ignition events. Furthermore, the accuracy of the location
of the arc ignitions reported by humans can be substantially improved. In cases of many arc ignitions
on specific route sections, the depots have more time to plan the maintenance intervals and get more up
to date information on wear conditions of the train engines. Up until now, the visual inspections were
on a monthly base only.
The arcVision System can detect arc ignition patterns and provides some additional data about static as
well as dynamic context information. The combination and aggregation of the available data may result
in PdM tasks and activities that over time will provide more insights for human inspectors to draw
conclusions and initiate appropriate measures. As such, human knowledge and expertise is still needed
to properly categorize the cases and optimize the necessary maintenance activities. The arcVision
System can assist humans by collecting enormous amount of data, detect anomaly patterns, visualize
and analyze them in order to reduce the manual inspection work and increase the quality and reliability
of the railway infrastructure. In consequence, the RQ4 can be answered in terms of that the arcVision
System can assist a human inspector in terms of providing and processing data on a large scale.
6. Conclusion and Outlook
It is important to detect potential incidents as fast as possible and prevent defects in the railway
infrastructure, where even a minor interruption can evolve to a major disturbance. Predictive
maintenance is possible to some extent with artificial intelligence, especially when using results from
machine learning techniques to derive appropriate rules for the maintenance tasks with the help of
knowledge engineering.
The development of the arcVision System has shown that it is possible to detect arc ignitions in the
pantograph-catenary system, which can assist human inspectors to analyze the state of the health of the
current railway infrastructure and draw the right conclusions when predictive maintenance is necessary.
Although the current arcVision System prototype demonstrated that in principle, the model is able to
detect arc ignitions with an adequate accuracy. Nevertheless, the confusion matrix shows that the model
has still some problems with false positives and predicts arc ignitions, which do not exist. The model
also shows signs of overfitting since it was trained with a small data set and possibly learned irrelevant
features. By using Keras, the design of the CNN is easily possibly and the application of AI and machine
learning is feasible within the context of a typical system development lifecycle. The amount of time
for the design, implementation and testing of the prototype was quite reasonable and the return on
investment of an AI-based solution development is feasible even with little prior knowledge about
machine learning. Some valuable insights about the arc ignition patterns could be derived, which were
out of reach before.
During the evaluation of the arcVision System, several railway operators were interviewed in a
qualitative way. Some of their key findings were that today, only little effort was raised to address
anomaly detection with machine learning techniques. Many hurdles need to be considered but the main
message was the importance of the cost-benefit-analysis. Some concerns were mentioned about
technology replacing humans. When AI-based systems such as arcVision can assist humans to reach a
higher quality and better reliability in collecting and processing data, they are very valuable. When
drawing conclusions, human knowledge and human wisdom were still considered superior.
7. References
[1] Reference Architecture Model Industrie 4.0 (RAMI4.0), https://www.beuth.de/en/technical-
rule/din-spec-91345/250940128
[2] Hribernik, K., von Stietencron, M., Bousdekis, A., Bredehorst, B., Mentzas, G., Thoben, K.D.:
Towards a unified predictive maintenance system - A use case in production logistics in
aeronautics. Procedia Manuf. 16, 131–138, 2018. doi:10.1016/j.promfg.2018.10.168
[3] Dalzochio, J., Kunst, R., Pignaton, E., Binotto, A., Sanyal, S., Favilla, J., Barbosa, J.: Machine
learning and reasoning for predictive maintenance in Industry 4.0: Current status and challenges.
Comput. Ind. 123, 103298, 2020. doi:10.1016/j.compind.2020.103298
[4] European Parliament: Digitalisation in railway transport A lever to improve rail competitiveness,
2019
[5] Morandi, D.: Anomaly Detection in Railway Infrastructure, Master Thesis University of Applied
Sciences and Arts Northwestern Switzerland, School of Business, 2020
[6] Al-Douri, Y.K., Tretten, P., Karim, R.: Improvement of railway performance: a study of Swedish
railway infrastructure. J. Mod. Transp. 24, 22–37 (2016). doi:10.1007/s40534-015-0092-0
[7] Gibert, X., Patel, V.M., Chellappa, R.: Deep Multitask Learning for Railway Track Inspection.
IEEE Trans. Intell. Transp. Syst. 18, 153–164 (2017). doi:10.1109/TITS.2016.2568758
[8] Siemes Ltd.: Broken Rail Detection,
https://assets.new.siemens.com/siemens/assets/api/uuid:246cb11523d1a7b06f5423baa17f4479cf
93f951/version:1525252841/datasheet-mrx-brd-audisclaimer.pdf
[9] Guzman, D.N., Hadzic, E., Schuil, R., Baars, E., Groos, J.C.: Turning data driven condition now-
and forecasting for railway switches into maintenance actions. (2018)
[10] Wu, G., Gao, G., Wei, W., Yang, Z.: The Electrical Contact of the Pantograph- Catenary System.
Springer Nature Singapore Pte Ltd (2019)
[11] VÖV: Facts & Figures Swiss Public Transport 2016/2017. (2017)
[12] Culotta, A., McCallum, A.: Reducing labeling effort for structured prediction tasks. Proc. Natl.
Conf. Artif. Intell. 2, 746–751 (2005)
[13] Kaushal, V., Iyer, R., Kothawade, S., Mahadev, R., Doctor, K., Ramakrishnan, G.: Learning from
less data: A unified data subset selection and active learning framework for computer vision. Proc.
- 2019 IEEE Winter Conf. Appl. Comput. Vision, WACV 2019. 1289–1299 (2019).
doi:10.1109/WACV.2019.00142
[14] Zhao, H., Chen, H., Dong, W., Sun, X., Ji, Y.: Fault diagnosis of rail turnout system based on case-
based reasoning with compound distance methods. Proc. 29th Chinese Control Decis. Conf. CCDC
2017. 4205–4210 (2017). doi:10.1109/CCDC.2017.7979237
[15] Chollet, F.: Keras API Special Interest Group (SIG)
[16] Ribani, R., Marengoni, M.: A Survey of Transfer Learning for Convolutional Neural Networks.
Proc. - 32nd Conf. Graph. Patterns Images Tutorials, SIBGRAPI-T 2019. 47–57 (2019).
doi:10.1109/SIBGRAPI-T.2019.00010
[17] O’Shea, K., Nash, R.: An Introduction to Convolutional Neural Networks. 1–11 (2015),
ArXiv:1511.08458 [Cs]. http://arxiv.org/abs/1511.08458
[18] Kotsiantis, S., Kanellopoulos, D., Pintelas, P.: Handling imbalanced datasets : A review. Science
(80-. ). 30, 25–36 (2006)
[19] Jüngling, S., Peraic, M., Martin, A.: Towards AI-based Solutions in the System Development
Lifecycle. Proc. AAAI 2020 Spring Symp. Comb. Mach. Learn. Knowl. Eng. Pract. (AAAI-
MAKE 2020) - Vol. I. 2600, (2020)
[20] Zhao, H., Chen, H., Dong, W., Sun, X., Ji, Y.: Fault diagnosis of rail turnout system based on case-
based reasoning with compound distance methods. Proc. 29th Chinese Control Decis. Conf. CCDC
2017. 4205–4210 (2017). doi:10.1109/CCDC.2017.7979237