Signal processing for non-destructive testing
of railway tracks
Cite as: AIP Conference Proceedings 1949, 030005 (2018); https://doi.org/10.1063/1.5031528
Published Online: 20 April 2018
Thomas Heckel, Ralf Casperson, Sven Rühe, et al.
ARTICLES YOU MAY BE INTERESTED IN
NDT of railway components using induction thermography
AIP Conference Proceedings 1706, 150001 (2016); https://doi.org/10.1063/1.4940613
Advanced methods in NDE using machine learning approaches
AIP Conference Proceedings 1949, 020022 (2018); https://doi.org/10.1063/1.5031519
Monitoring the reflection from an artificial defect in rail track using guided wave ultrasound
AIP Conference Proceedings 1949, 090003 (2018); https://doi.org/10.1063/1.5031566
AIP Conference Proceedings 1949, 030005 (2018); https://doi.org/10.1063/1.5031528 1949, 030005
© 2018 Author(s).
SIGNAL PROCESSING FOR NON-DESTRUCTIVE
TESTING OF RAILWAY TRACKS
Thomas Heckel1,a), Ralf Casperson1,b), Sven Rühe2,c), Gerhard Mook3,d)
1
Federal Institute for Materials Research and –testing (BAM), Berlin, Germany
2
PLR Prüftechnik Linke & Rühe, Magdeburg, Germany
3
OVGU Otto-von-Guericke-University Magdeburg, Magdeburg, Germany
Corresponding author: a) thomas.heckel@bam.de
b) ralf.casperson@bam.de
c) sven.ruehe@plr-magdeburg.de
d) gerhard.mook@ovgu.de
Abstract. Increased speed, heavier loads, altered material and modern drive systems result in an increasing number of rail
flaws. The appearance of these flaws also changes continually due to the rapid change in damage mechanisms of modern
rolling stock. Hence, interpretation has become difficult when evaluating non-destructive rail testing results. Due to the
changed interplay between detection methods and flaws, the recorded signals may result in unclassified types of rail flaws.
Methods for automatic rail inspection (according to defect detection and classification) undergo continual development.
Signal processing is a key technology to master the challenge of classification and maintain resolution and detection quality,
independent of operation speed. The basic ideas of signal processing, based on the Glassy-Rail-Diagram for classification
purposes, are presented herein. Examples for the detection of damages caused by rolling contact fatigue also are given, and
synergetic effects of combined evaluation of diverse inspection methods are shown.
INTRODUCTION
The degeneration of rails is increasing due to the constant increase in loads and amount of overall traffic. In
particular the operation of high speed trains, the alteration of materials used in the wheelsets, and the application of
modern drive system concepts result in an increasing number of flaws in rails. This causes a change in the damage
mechanism by the modern rolling stock and may alter the appearance of different flaw types. To guarantee the safe
operation of rail traffic, mechanized non-destructive inspection techniques are used in large scale to detect damages
on rails. Today we face the following list of flaw types to be handled by non-destructive testing methods:
1. Headchecks
2. Defects in welds
3. SQUATs
4. Wheel burns
5. Corrugation
6. Degradation of joints
7. Kidney shaped flaws
44th Annual Review of Progress in Quantitative Nondestructive Evaluation, Volume 37
AIP Conf. Proc. 1949, 030005-1–030005-7; https://doi.org/10.1063/1.5031528
Published by AIP Publishing. 978-0-7354-1644-4/$30.00
030005-1
INSPECTION METHODS
Visual non-destructive inspection of rails has been performed since the very beginning of rail traffic and is still
one of the most commonly used methods. For the detection of internal flaws in the rail caused by employment or
during production, additional methods have been developed over time. For the detection of volumetric and crack type
reflectors in the rail head, web and foot ultrasonic methods have been used since the 1950s. For the advanced detection
of surface breaking and near surface defects in the rolling contact region of the rail head, eddy-current methods have
been developed and applied since the 2000s.
Systems for manual in-service inspection using ultrasonic and eddy-current methods have been developed and
allow a typical inspection speed of about 1 m/s.
For the mechanized inspection of rails, rail inspection trains have been built that incorporate different non-
destructive inspection methods that operate at inspection speeds of up to more than 10 m/s. In this article we focus on
the detection and evaluation of indications using ultrasonic methods in rail inspection trains.
BASIC SETUP
The basic setup for the non-destructive ultrasonic rail inspection using rail inspection trains contains a number of
straight beam probes and angle beam probes operated in parallel to cover the areas of interest in the different zones of
the rail. In actual setups used in the inspection trains operated by Deutsche Bahn AG, typically two straight beam
probes emitting longitudinal waves are used to detect reflectors that are orientated in parallel to the running surface of
the rail and to detect the rails height. One of the two probes features a dual element setup and focuses on the rail head
and reflectors close to the running surface while the other one focuses on the rail web and foot. For the detection of
indications that feature an orientation non-parallel or perpendicular to the running surface, angle beam probes emitting
transverse waves are applied. Typically two sets of three angle beam probes with different angles of incidence are
used in the setups to cover the angular range from +35° to +70° and -35° to -70°. For example the setup used in
inspection trains operated in Germany is 35°, 55° and 70°. While the 70° probes mainly focus on the rail head, the 55°
probes focus on the rail head and the web, and the 35° probe covers the whole rail from the rail head to the rail foot.
Supplementary ultrasonic probes may be used to cover optional areas, e.g., angle beam probes focusing on certain
areas in the rail head. Figure 1 shows the setup used in an SPZ1 and SPZ2 train operated by Deutsche Bahn AG.
FIGURE 1. Typical ultrasonic probe setup for rail inspection trains.
SOUNDFIELD SIMULATION
To obtain optimal performance during inspection, sound field parameters for each probe have been simulated using
semi-analytical and finite elements methods. Semi-analytical models have been applied to evaluate signal response
and distance amplitude curves while finite element methods have been used to evaluate wave propagation in the rail
and interaction with the geometric boundary conditions of the rail. Figure 2a shows the sound field of a straight beam
probe in the cross section of a rail perpendicular to the rail axis simulated using ArrayCalculus3D. Figure 2b shows
the correspondent sound field of a 35° angle beam probe in the cross section along the rail axis.
030005-2
FIGURE 2a. Normal beam probe. FIGURE 2b. 35° angle beam probe.
For optimal probe positions and rail geometries, wave propagation has been simulated to identify the expected
geometric positions using the simulation tools AnSys and CIVA. During inspection the probes are aligned as close as
possible to the geometrical center of the rail by the mechanical guiding system of the measurement boogie.
Nevertheless, dependent on the wear of the inspected rail and in curves, a misalignment of the probes may occur. This
misalignment causes a shift in sound travel time and echo amplitude of the received signals. For a misalignment from
0 mm to 15 mm, Figs. 3a, 3b and 3c show wave propagation in the rail head at 6 Ps, 12 Ps and 32 Ps. The relative
acceleration of the elements is shown.
FIGURE 3a. Wave propagation in rail head after 6 Ps.
FIGURE 3b. Wave propagation in rail head after 12 Ps.
030005-3
FIGURE 3c. Wave propagation in rail head after 32 Ps.
It can clearly be seen that for larger misalignments, the reflections from the region of the transition from the rail
head to the web become asymmetric and cause echo signals different in amplitude and travel time. Also mode
conversion from longitude to transverse mode can be identified indicated by the different wavelengths and propagation
velocities of the reflected pulse packets. The behavior of the indications from the rail geometry has been well
understood by the use of these models.
The output from the simulations has been used to optimize the parameters for the setup of the ultrasonic system
and for the setup of the data recording.
DATA RECORDING
The sound velocity in the rail and the geometry of the rail limit the maximum achievable pulse repetition rate, at
least by the sound travel time elapsed in the rail and probe. The lateral resolution of ultrasonic measurements is directly
affected by the train speed. Therefore inspection trains are typically operated at inspection speeds as close as possible
to the physical limits for the ultrasonic testing.
The high inspection speed causes a large amount of incoming data from the probes. This poses a challenge to the
processing and the evaluation of the collected data, e.g., the SPZ1 operated by Deutsche Bahn AG continuously
records measured ultrasound data at a repetition frequency of about 4650 Hz independent of operation speed merged
with additional information, e.g., GPS, time stamps and position markers. This results in raw data volume of about
300 MB per kilometre.
GLASS- RAIL-DIAGRAM
The detectability of defects decreases with the increase of speed. To maintain resolution and detection quality over
a wide inspection speed range independent of operation speed, signal processing algorithms have to be applied.
Therefore real time algorithms and the Glassy-Rail-Diagram have been developed and tested. An example image for
a Glassy-Rail-Diagram is shown in Fig. 4.
FIGURE 4. Glassy-Rail-Diagram of fish-plate junction.
030005-4
The Glassy-Rail-Diagram developed by BAM gives a side view from the rail like conventional B-scans. The
algorithms consider the position of probes, angles of incidence of the probes and sound paths as well as the rail
geometry to generate a geometry corrected diagram incorporating all A-scan and gate data from all recorded channels.
The diagrams resolution respectively the size of a pixel is 3 mm by 3 mm. Two types of information are displayed in
the Glassy-Rail-Diagram. A gray scale image represent the maximum amplitude recorded for each pixel while a
coloured image represents the probes which recorded a significant amplitude for each pixel. For a combined display
both images are overlayed.
IMAGE PROCESSING AND DATA EVALUATION
To support evaluation of data by the operator, automated indication classification algorithms featuring ten classes
have been developed and adapted, which allow preselection of data displayed during evaluation. For each of the three
regions (rail head, web and foot) three classes have been implemented, respectively. Unascertainable indications are
handled in an additional class for unknown indication types.
The implemented data post processing on the recorded raw data for each rail uses a three step algorithm based on
statistic methods, a neuronal network and fuzzy logic.
During processing the recorded data are segmented to 2D-clusters with a size of 64 by 64 pixels, which equals an
area of 192 mm by 192 mm. Overlapping of the segments is set to 50 percent. For each kilometer and each rail, 10417
clusters of amplitude and probe gate data have to be evaluated.
Main foci of the implemented algorithms are the identification of indications caused by acoustic and electric noise
as well as the identification of non-generic indication patterns and indication patterns caused by drill holes and welds.
Rail type can be evaluated by measuring rail height.
In a first step, the 2D-cluster is evaluated by statistic methods to extract the features describing this cluster. Due
to the harsh environment and the boundary conditions for the inspection, recorded data will typically be incomplete
(up to a certain amount), and will lack some features or will contain unwanted information, e.g., noise. In this case
using an evaluation based on statistic methods is a good choice to become stable against partial signal loss or noise
from some of the sensors. For the definition of the features to be extracted from the cluster, detailed a-priori
information on the behavior of the ultrasound and its interplay with defects in the rail head, web and foot region has
been a mandatory input. Dependent on the region, features listed below are evaluated for each cluster.
1. Histogram of amplitudes
2. Neighborhood criteria
3. Local distributions of indications
4. Multiple indications
5. Multiple pixel hits
6. Rail height
7. Signal loss
The evaluated features will be input to the next processing step.
Based on a neuronal network, pattern recognition is performed in a second step. For each 2D-cluster, the feature
list outputted by the statistical evaluation is analyzed by a supervised trained neuronal network to identify significant
patterns. The pattern descriptors are subdivided into indication groups from geometry, indication groups from noise,
indication groups from forms (e.g., drill holes and welds), and flaw type indication groups. Caused by the nature of
the recorded data, multiple findings of patterns in one cluster will occur.
The training of the neuronal network has been stopped at a certain stage to maintain a stable and consistent reaction
of the analyzing process during inspection.
The third step of the signal processing is done by means of fuzzy logic. The pattern descriptors outputted by the
neuronal network are weighted with fuzzy logic to decide on the final classification. Different rail geometries (e.g.,
different rail heights), unstable signal quality, incomplete data sets and multiple findings have to be covered during
evaluation. These boundary conditions will not allow sharp detection thresholds. For the avoidance and reduction of
false calls, decision rules on the pattern descriptors for the final classification have to be rather soft and floating.
030005-5
Therefore the applied fuzzy logic algorithm is designed to balance the probabilities of selected options for each class.
Design of the algorithm has been done incorporating a-priori information based on expert knowledge.
CLASSIFICATION EXAMPLES
To demonstrate the performance of the algorithms three examples are given. Each Glassy-Rail-Diagram image
shows about one meter of one rail. The first example in Figure 5 shows a typical thermite weld flanked by two drillings
one to the left and one to the right at the ends of the rail. The data evaluation has detected at least four relevant
indication patterns for classification. Indication one and three have classified as drilling in the rail web with 75% of
possible features detected. Indication two has been classified as weld with 66% of possible features detected.
Indication 4a and 4b have been identified as noise with a threshold of 25%. The proposal for evaluation to the operator
is drilling, weld, drilling for the affected positions in the clusters.
FIGURE 5. Example Glassy-Rail-Diagram evaluation drillings and weld.
Example two (Fig. 6) shows the typical indication of one defect of type SQUAT in the very middle position of the
image. Maybe there is a second SQUAT in the right area of the image, but signal amplitude has been too low to trigger
the hardware gates so the signal is only displayed in the grey-scale image and no signal has been recorded in the gate
data set. The data evaluation has detected two relevant indication patterns. Indication one has been classified as defect
in the rail head with 100% of possible features. Indication two has been classified as unascertainable indication. The
probability for classification weld and for classification defect in rail head has been 50%. Caused by the lack of gate
data for this region in the rail head and a loss of backwall indication, the fuzzy logic decided to sort the indication to
unascertainable indication for the affected cluster, which has to be decided by the operator.
FIGURE 6. Example Glassy-Rail-Diagram evaluation SQUAT.
Example three (Fig. 7) shows a very noisy section of data. This typically occurs when the train wheels are
generating audible high frequency noise, which generates surface waves in the rails overlaying the ultrasound signals
received by the probes. In this section there are two drillings present. These drill holes are typically used for the
mounting of ground connections. In all clusters, noise indication pattern has been detected and classified with 100%
of possible features. Indication two and three have been classified as drilling in the rail web with 87% of possible
features for indication two and 75% for indication one. Both drillings two and three are marked with a cursor for
distance measurement in the affected clusters.
FIGURE 7. Example Glassy-Rail-Diagram evaluation noise and drillings.
030005-6
CONCLUSION
Rail defects and their detection have posed a challenge to the save operation of rail traffic since the beginning of
the railway age. Damage mechanisms and their appearance have been altered over time due to different materials,
loads and gear used.
Mechanized ultrasonic rail inspection carried out with rail inspection trains acquires huge amounts of inspection
data, which has to be processed and evaluated by the means of signal processing. Caused by the large speed span of
the train and harsh environmental conditions, the recorded data have a variation in signal quality. Using a three-step
evaluation based on statistical methods, a neuronal network and fuzzy logic, algorithms become stable against
variation. For the classification of the recorded data on clusters taken from the Glassy Rail Diagram in a first step,
statistic methods are applied to extract features for further evaluation. Data analysis and classification algorithms
based on a neuronal networks followed by fuzzy logic, implemented with expert knowledge reporting ten classes, give
support to the operator when evaluating measured data with the Glassy-Rail-Diagram. Three examples have been
shown and the performance of the algorithm has been demonstrated.
For the future the next step shall be an estimation for the probability of detection (POD) and for the probability of
classification (POC) calculated by the algorithm itself based on the evaluated data set. Due to the number of different
rail types and rail profiles as well as the constant change of operating conditions, the adaptation and optimization of
system setups and algorithms is still an ongoing process.
ACKNOWLEDGMENTS
Special thanks goes to Rainer Boehm and Yannick Wack from BAM who supported this work with their
simulations using ArrayCalculus3D, AnSys and CIVA.
REFERENCES
1. H.-M. Thomas, T. Heckel and G. Hanspach; “Advantage of a Combined Ultrasonic and Eddy Current
Examination for Railway Inspection Trains,” Insight, 49:6, 341-344, (2007).
2. T. Heckel, H.-M. Thomas, M. Kreutzbruck, S. Rühe, “High Speed Non Destructive Rail testing with Advanced
Ultrasound and Eddy-Current Testing Techniques,” Proceedings NDT in Progress, 5th International Workshop
of NDT Experts. Prague, ed. P. Mazal (2009).
3. 3. R. Casperson, T. Heckel, “New Evaluation Methods for Non-Destructive Rail Inspection Using Eddy Current
and Ultrasound,” Railway Engineering, Edinborough, (2017).
4. R. Krull, H. Hintze, H.-M. Thomas, T. Heckel, “Nondestructive testing of Rails today and in the Future,” ZEVrail
Glasers Annalen, 127, 286-296, (2003).
5. DIN EN 16729-1:11/2016, “Railway applications. Infrastructure. Non-destructive testing on rails in track.
Requirements for ultrasonic inspection and evaluation principles,” (2016).
6. Rene Heyder, “The new UIC catalogue of rail defects,” Der Eisenbahningenieur, 52:9, 94-109, (2001).
7. E. Martin, K. Werner, “Schienenprüfung mit Ultraschall und der Ultraschall-Schienenprüfwagen der Deutschen
Bundesbahn,” Eisenbahntechnische Rundschau, Heft 12, (1956).
030005-7
