Original Article

Comparison of sensors and

methodologies for effective prognostics
on railway turnout systems

Proc IMechE Part F:

J Rail and Rapid Transit
2016, Vol. 230(1) 2442
! IMechE 2014
Reprints and permissions:
sagepub.co.uk/journalsPermissions.nav
DOI: 10.1177/0954409714525145
pif.sagepub.com

Fatih Camci1, Omer Faruk Eker2, Saim Baskan3 and

Savas Konur4

Abstract
Railway turnout systems are one of the most important components of a railways infrastructure. Their geographically
distributed nature makes failure detection, forecasting and maintenance planning extremely important. Prognostics,
forecasting the time to failure in order to achieve effective maintenance planning, has attracted increasing attention
from industry and researchers in recent years. The prognostic approach has great potential to achieve reduced costs and
increased availability. However, the applicability of any engineering model requires economic and practical justifications.
This paper presents an analysis of different prognostic methods for railway turnout systems. Five different sensors,
installed in a real turnout system used on Turkish State Railways, are individually analysed by applying various prognostic
methods. This paper aims to guide practitioners on the application of prognostics and health management technologies
to railway turnout systems by discussing the advantages and disadvantages of using different sensors and prognostic
methods.
Keywords
Prognostics, remaining useful life, railway turnout systems, sensor comparison
Date received: 19 June 2013; accepted: 30 January 2014

Introduction
The number of passengers using railway transportation in the EU is expected to double in the next 10
years.1 The eciency of railways needs to be signicantly improved to be able to handle this increase.
Reliability, safety and availability of railway systems,
highly related to service quality and operation/management costs, should be enhanced to satisfy increasing transportation demands. The enhancement of
these aspects depends on factors such as condition
monitoring, early detection of failures (i.e. diagnostics), forecasting the failure time (i.e. prognostics)
and eective maintenance planning. There exist
many studies on condition monitoring and failure
detection in components of railway infrastructures24
and railway turnouts.58.
Prognostics is dened as predicting the remaining
useful life (RUL) of an asset by employing combinations of condition monitoring data, rst principal
physics, mathematical models and past experience of
the system. Prognostics and health management
(PHM) basically involves all aspects of the processes
used to predict the future health state of a system, this
state is then used to guide eective operations management. It allows making a maintenance plan in advance.

Several prognostic methods have been presented in the

literature for railway turnouts510 and other industrial
systems and components.1119 Even though these studies present some level of accuracy in failure forecasting, they mostly focus on the development of
prognostic technologies based on a set of sensors.
This paper aims to analyse the applicability of
prognostic technologies when combining their analysis with the selection of dierent sensors. The applicability includes two main aspects: accuracy and usage
cost. Accuracy is the eectiveness of the prognostic
technology in forecasting the failure. Usage cost
includes the cost of the sensor as well as ease of installation and usage. High applicability indicates high
accuracy and low usage cost. Four prognostic technologies and ve sensors are used in this study. This
paper may be used as a guide for the deployment of
1

Industrial Engineering Department, Antalya International University,

Turkey
2
Manufacturing Department, Cranfield University, UK
3
Electric/Electrical Engineering Department, Fatih University, Turkey
4
Department of Computer Science, University of Sheffield, UK
Corresponding author:
Fatih Camci, Antalya International University, Antalya 07190, Turkey.
Email: fatih.camci@antalya.edu.tr

Camci et al.

25

PHM technologies in railway turnout systems by providing advantages and disadvantages of using dierent sensors and corresponding prognostic methods.
The work presented in this paper was performed in
several stages, which can be summarized as follows.
1. Initially, sensors were compared based on several
parameters such as cost, ease of installation and
ease of use.
2. The natural failure degradation was mitigated
using discrete failure states.
3. Data (the degradation of dry/contaminated slide
chair failure mode) from a real railway turnout
system was collected.
4. Prognostic studies were performed using various
methodologies.
5. The methodologies and sensors were analysed
based on performance and cost factors.
6. Finally, the possibility of identifying a relationship
between expensive and cheap sensors was investigated using statistical methods.
This paper is organized as follows. The next section
presents information on the railway turnouts, sensors
and data acquisition system used in the studies. The
section Failure simulation presents a discussion on
the simulation of failures. The section Prognostic
analysis presents a discussion on prognostic methods
and their results. Finally, conclusions are drawn.

Railway turnouts, sensors and data

acquisition system
Railway turnouts
Trains change their tracks with the help of railway
turnouts that move rails before they pass. Railway
turnouts consist of several components such as
motor, reduction gear box, drive-detection rods and
other control components. Turnouts are locked into
position after the movement of the rails stops, which
takes several seconds. Not only maintenance costs,
but also accidents caused by railway turnout failures
increase their importance in the railway infrastructure.9 Figure 1 shows an example turnout system.
There are more than 20 dierent failure modes for
railway turnout systems.5 They are mostly related to the
following components: drive rod, slide chair, motor and
detection assembly. The failure mode related to the
drive rod is either due to an out-of-adjustment or defective rod. Slide-chair-related failure modes are mainly due
to dried, contaminated or defective slide chairs. The
motor has its own failure modes related to its sub-components. The defective sub-components lead to failure
modes in the detection assembly. Failure modes can be
created by problems with other sub-components such as
relay, heater, micro-switch, cable, etc. Drive-rod and
slide-chair-related failures correspond to 40% of the failure reports cited in the literature5 (each corresponding

Figure 1. Inner components of a railway turnout.

to approximately 20%). The degradations due to driverod and slide-chair-related failures were simulated and
data was collected from sensors during this project. The
data collection process and simulated failure progression
in the drive-rod-related failure mode is more prone to
human error compared with the slide-chair-related failure mode. Thus, the slide-chair-related failure mode is
selected for study in this paper.

Sensors
Before the data collection process was started, the failure modes were analysed to identify the best sensors for
eective data collection, they were subsequently purchased. Some of the failure modes, especially within the
motor, lead to an inability to start the movement of the
rods. Most of the failure modes result in making the
movement of the rods harder, which would lead to an
increased required force, reduced rod speed, change in
the current and voltage to produce the required power.
A proximity sensor was recommended by maintenance
operators, since they see the changes on the ground
surface that cause improper approximation of the
rods to the blades. Hence, force, linear ruler to calculate the speed, current, voltage, and proximity sensors
were selected and bought.
Failure eects can be observed to dierent degrees
in each sensor. Even though the sensitivity of these
parameters to failure is important, there exist other
factors, such as cost, ease of installation and ease of
use that should be considered in selecting the best
sensor to deploy among many alternatives. Table 1
compares some features of the sensors used in the
prognostic analysis of railway turnouts. The values
presented in Table 1 are based on our ndings
obtained during the search for appropriate sensors
at the start of the project. The eectiveness of the
chosen sensors will be discussed in a later section.
The ease of installation column in Table 1 refers to
after purchase of the sensor. If the sensor can be

26

Proc IMechE Part F: J Rail and Rapid Transit 230(1)

Table 1. Initial comparison of sensors for application to prognostics.

Sensor

Cost (E)

Ease of installation

Reliability

Sensitivity
to failure

Force
Current
Voltage
Proximity
Linear ruler
Rotary encoder motor speed position
String encoder Gearbox output
position speed

360 installation cost

30
50
20
250
50 installation
50 installation cost

Difficult
Easy
Easy
Easy
Moderate
Difficult
Difficult

Reliable
Reliable
Reliable
Reliable for short distance
Reliable
Reliable
Reliable

High
Low
Low
Limited
Moderate
Low
Low

installed by a technician in an hour without needing

extra modication, then it is considered to be an easy
installation. If the sensor is installed in 2 h and needs
extra mounting xtures, then it is considered to be a
moderate installation. If the sensor requires extra
machine shop work requiring additional time and
labour before the 2 h required by the technician for
its installation, then it is considered to be a dicult
installation.
The sensitivity column in Table 1 indicates the
change in obtained data when failures are articially
created. If the data changes by 10% or more, this is
considered to be a high change. If it changes by
between 3 and 10%, it is considered to be a moderate
change. If the change is less than 5%, then it is considered to be a low change. If the sensor cannot
monitor the whole failure range, it is considered to
display a limited sensitivity to failure.
Force sensor. This sensor measures the force
applied by the load cells to drive the arms during
their operation. This force is limited to a maximum
value of 600 kgf so as to not damage the arm or other
components. S-type strain-gauge-based sensors,
placed in the drive arm by inserting screws from
both sides, were used to measure the applied force
up to a 1000 kg force as shown in Figure 2. The
force sensor comes with an amplier that provides a
10 V analogue output measuring force in two directions (tension and compression). A force sensor was
installed on each drive arm. Pin-type load cells are
also available in the marketplace. Their installation
is easy; on the other hand their cost is 10 times
higher. Due to nancial considerations, a lower cost
is preferred over installation diculty.
Current and voltage sensors (DC motor voltage /
DC motor current). Current and voltage sensors were
used to measure the current and voltage in the DC
motor during its operation. Two Hall-eect voltage
sensors were used to measure the rail movement in
each direction. A DC Hall-eect current sensor that
has a hole in its centre was used to measure the
change in motor current in each direction. The current sensor gives an analogue 10 V isolated output
between 0 and 10 A in each direction. The voltage
sensor can measure between 0 and 1000 V, giving

Figure 2. Force sensors installed on drive arms.

Figure 3. Current and voltage sensors.

5 V isolated output. Current and voltage sensors

were installed at the terminals of the motor, as
shown in Figure 3.
Linear ruler. The linear ruler measures the time and
speed of the turnout blade, as shown in Figure 4, this
allows movements of up to 25 cm of the drive arm to
be detected. A moving magnet is attached to the drive

Camci et al.
arm that slides over the ruler mounted on the traverse.
The two linear rulers can each measure a 010 V analogue DC output. A drive arm equipped with a
magnet might move slower than normal due to a
resistance increase caused by an incipient failure.
There are two drive arms in the turnout system and
thus it requires two linear rulers: one is close to the
rail, and the other moves away from the other rail.
Proximity sensor. The proximity sensor measures
the distance between the railway and turnout blades,
which is required to be within 24 mm. A proximity
sensor, as shown in Figure 5, is installed on each side
of the railway. These sensors can detect the distance
of the moving steel blade to the main blade to within
10 mm. This sensor produces 010 V analogue DC
output for the 010 mm distance. If a selected
moving blade of a track is more than 4 mm away
from its main rail, this indicates that the turnout
system is in a failed state.

27
Rotary encoder. This sensor was placed next to the
shaft that transmits the movement from the DC
motor to the reduction gear, which then transmits
the movement to the arms to measure its acceleration
and rotational velocity. The encoder produces digital
pulses as an output that can be counted for speed and
position information. A two-channel digital output
incremental encoder was used to detect missing signals during the transmission of data on the movement
to the arms. This sensor was also used to obtain friction and space parameters that were used in the creation of the physical model of a railway turnout.
Figure 6(a) shows the circular position, velocity
sensor and connected metal arm. A magnetic encoder
(non-contact circular magnet), as shown in
Figure 6(b), was installed on the motor shaft. Figure
6(a) and (b)
A string-pulled incremental encoder. This encoder,
shown in Figure 7, was installed to measure the position and velocity of the pinion located after the gearbox, whose movement does not complete a full
revolution. The reason for using this sensor was to
calculate the torque and inertia between gearbox
and load, used to build the physics-based model of
the railway turnout. The string of the encoder can
be pulled a distance of up to 1 m and it produces
digital pulses as an output that can be counted for
speed and position information on the constrained
movement.

Data acquisition system

Figure 4. Detection of the position of the drive arm using a

linear ruler sensor.

Figure 5. Outer and inner side views of a proximity sensor.

The signals from the sensors were acquired by a computer using a NI-USB-6259 data collection system
from National Instruments. The NI-USB-6259 has a
multi-channel analogue/digital input and output and
a 1.5 MS/s data transmission speed. The data acquisition (DAQ) system shown in Figure 8 was built to

28
ease the installation process for each data acquisition
process. The DAQ system can acquire continuous
analogue signals from the sensor, current sensor, voltage sensor, proximity sensor and linear ruler with an
adjustable sampling rate that is specied on the specication sheet at the manufacturers web site.
Simultaneously, the digital counter in the DAQ
system counts digital pulses. All measured data can
be seen on the front panel of a LABVIEW program,
also the data is recorded in Excel or text format with
preferred sampling. Power supplies at 24 V, 15 V
and 5 V DC were used to power the sensors; they produce analogue or digital outputs that are within
10 V. LABVIEW data acquisition software was
used to acquire the data. Figure 9 shows the graphical
programming part of the software.
The customized DAQ system was used rst to
search for a suitable sensor and then the DAQ
board selected the type and range of signals for that
sensor. There were several cycles of modications
before the nal system setup was achieved. The sampling speed of the DAQ board can reach 1.5 MS/s as
the sum of all channels, which is more than enough
for this project. The sensory data used in this paper is
publicly available at http://www.aiu.edu.tr/sta/
fatih.camci/datasets.html.

Proc IMechE Part F: J Rail and Rapid Transit 230(1)

lie between the diverging outer rails. The rail blades
move back and forth on metal platforms called slide
chairs that are mounted on wooden oors called traverses (or sleepers) as shown in Figure 10. A rail blade
slides on 12 slide chairs. The sliding distance is dierent for each traverse (minimum value at traverse 12
and maximum value at traverse 1) and the movement
of rail blades is circular with a very small angle and
with one end of the blade pinned to a xed point near
traverse 12. The sliding distance increases as the slide
chair gets closer to the railway turnout. Sliding chairs
are periodically lubricated to achieve a smooth
movement.
Initially, all 12 slide chairs were lubricated, which
represents the failure-free state. In the natural failure
degradation process, all slide chairs are dry or contaminated to a similar extent. However, a simulated
failure was started by contaminating the three furthest
slide chairs (10th, 11th and 12th) with sand. This state
was dened as the rst failure state. Note that the
sliding distance in the furthest slide chairs is very
small and increases for slide chairs closer to the railway turnout. The second failure state was obtained by

Simulation of failure events

In this section, the process of data collection on degradation due to failures is discussed. Data relating to
the degradation resulting from a dry/contaminated
slide chair failure of a railway turnout was collected.
The collection of data during a natural failure is very
dicult, in not in fact impractical, to obtain in most
cases, since a natural degradation event requires an
extensive amount of time. Thus, the failure was simulated by creating discrete failure states.
A railway turnout is a mechanical system that
enables trains to switch between tracks. The turnout
contains a pair of linked rails, called rail blades that

Figure 7. String-pulled encoder for the gearbox output.

Figure 6. (a) Incremental encoder and mounting arm and (b) magnetic-drive incremental encoder.

Camci et al.

29

Figure 8. The DAQ system.

Figure 10. Sliding chairs and turnout blades.

contaminating the ninth slide chair in addition to the

previously contaminated ones. This procedure was
repeated until all slide chairs were contaminated,
which was dened as the failure state. Modelling of
the failure states is depicted in Figure 11.
For each failure state, at least 20 sets of back-andforth sensor data were collected. Each movement took
up to 8 s, and the data collection process was performed during this period. Figure 12 shows the
force signals with the x-axis representing time and
the y-axis representing the amount of force applied
during the movement.
The time spent in each state is an important parameter to model the failure degradation. In order to
estimate the failure time, extrapolations can be performed using either: a linear, exponential, power or

30

Proc IMechE Part F: J Rail and Rapid Transit 230(1)

Figure 11. Failure degradation modelling and data collection.

Figure 12. Force sensor signals collected from 10 different health states.

logarithmic model.10 The linear and exponential

models are recommended when it is known that no
prior degradation exists.20 This is the situation in our
case. Since usage, operation and environment aect
the system health more in degraded systems compared

with brand new systems, exponential degradation was

selected since it better represents the real degradation.
This is also considered in the failure simulation by
starting the contamination process from the furthest
slide chair. The travel distance of the rail blade on a

Camci et al.

31

Figure 13. Total life of a railway turnout with failure degradation.

Figure 14. The prognostic process.

slide chair close to the turnout system is more than the

travel distance on a distant slide chair. The contamination on the closer slide chair creates more resistance due to the longer travel distance. Thus, the
failure degradation increases as the failure develops.
The total life of a railway turnout was dened after
identication of the time spent in each state. The data
collected from various sensors over the life of the turnout life were visualized. The eectiveness of the sensors for degradation was gauged based on the RUL
estimates. However, the degradation can be estimated
by visually comparing sensor data. The force sensor
clearly shows the presence of degradation. The degradation can also be seen in the data from the current
sensor, but less clearly. The remaining sensors also
indicate some level of degradation; however, visual

ranking is not easy. An example of a full failure for

force and current sensors are given in Figure 13.

Prognostic analysis
Prognostic methods
The concept of prognostics was introduced in the
early-1990 s. However, solid results have been presented only in the last few years; nevertheless, most
of the methods are specic applications rather than a
generic method.21
Prognostics process can be separated into two
phases as shown in Figure 14. The rst phase of prognostics, which could also be considered in diagnostics,
aims to assess the current health state of a system.

32
Other terms such as condition identication, severity detection or degradation detection are used in
the literature. Typically, pattern recognition techniques (classication and clustering) are employed in
this phase to identify the health condition of the
system. In the second phase, the health state is extrapolated to estimate the RUL of the system using forecasting techniques such as time series or physical
degradation methods.22
Prognostic methods are divided into two major
categories: data-driven (empirical) and physics-based
(model-based) prognostics.23 In the data-driven
approach, the failure degradation is modelled based
on past failure degradation data collected from similar equipment. Data representing the failure degradation is the key to achieving a data-driven prognostic
model. The physics involved in failure degradation is
not considered. The model, which is created using the
failure degradation of similar equipment, is then used
to forecast the failure degradation of the equipment
under observation. Data-driven models have been
considered to be a black-box operation since they possess no detailed information related to the physics or
mechanics of the system.24
Physics-based models, on the other hand, typically
involve the physics of the failure mechanism and/or
degradation. Physical relationships between parameters that cause and propagate the failure are
mathematically modelled.25 In order to provide
knowledge-rich prognostic output; physics-based
models attempt to combine defect growth formulas,
system-specic mechanistic knowledge and condition
monitoring data.26 In physics-based models, the failure degradation is either directly formulated as in the
Paris Law applied in structural health monitoring, or
any deviation from a formulated healthy system is
tracked as the failure degradation. The ability to
model the physics of the failure and knowledge of
the parameter values are essential to physics-based
models. An advantage over the data-driven models
is that historical data, representing the failure degradation, is not needed. Data-driven and physics-based
models have been applied to railway turnouts.
Both physics-based and data-driven methods were
used in this study. Even though the physics-based
method consumed large amounts of eort and time,
it did not lead to good results. Thus, only details of
the data-driven methods are discussed below. The failure degradation is assumed to be discrete as discussed
in the previous section. Among the most well-known
clustering algorithms e.g. k-means, Fuzzy c-means,
Hierarchical Clustering and Mixture of Gaussians,
k-means was used in Phase I of the data-driven prognostics. In the k-means algorithm, data is initially randomly grouped into k clusters. The mean and member
of each cluster is recursively updated based on the
nearest mean principal until convergence. For more
information about k-means clustering, readers are
referred to Gan et al.27

Proc IMechE Part F: J Rail and Rapid Transit 230(1)

Identication of the best number of clusters (k) is
one of the challenges in clustering problems. Several
metrics have been dened to identify the best value for
k. A silhouette cluster evaluation function is used in
this paper. Note that 10 discrete health states were
used during the failure degradation modelling.
However, due to the small changes in the early
phase of failure degradation, it was dicult to dierentiate initial health states based on the generated
data. Thus, the best number of clusters was identied
to be eight, based on the evaluation of the clustering.
After the means of the clusters were dened based
on the training data, the health state of the system
under observation was identied using the nearestneighbour principal on the condition indicator.
Then, the health state was extrapolated and the
RUL was calculated based on the following methods.
Simple state-based prognostic method (SSBP). This
method can be used for the stochastic modelling of the
transition between health states from brand new
through to failure. In this stochastic model, the failure
degradation occurs by transitions between healthy
states. A transition from one healthy state to another
is dened by a probability value. Transition probabilities between healthy states are calculated using training samples. The expected RUL of a turnout is
estimated by calculating the expected number of transitions from a current state that reach a failure state.
More details about the method can be found in Eker
et al.9
State-based prognostic with state duration information (SBPD). SSBP does not incorporate the duration
spent in a state in the transition probability calculation. In other words, the RUL of a system that
recently entered a healthy state is the same as that
for a system that has been in that state for a long
time. It is obvious that more time spent in a state
means that it is more likely to leave that state due
to degradation. SBPD extends SSBP to incorporate
the duration information in the RUL calculation.
The transition probability from the current state to
the next state is calculated based on the time spent
in the current state. The increase in transition probability to the next state with time spent in a state is
incorporated using past examples. Results indicate that the accuracy of RUL estimation considerably increases by incorporating the duration
information.10
Time delay neural network (TDNN). This is a specic type of neural network that incorporates inputs
from dierent time slots in the estimation process.
The inputs could be data from current and previous
times as well as its own output obtained from previous time slots. TDNN creates a nonlinear relationship
between parameters from dierent time slots and its
own past predictions and uses this relationship to estimate the parameters in following time slots. Once the
current state is identied, previous states, as well as
previous predicted states, are used as input to estimate

Camci et al.
the next state. The process is repeated assuming that
the predicted state is the real state to estimate
the following states until the predened state is
reached. The number of estimations is counted as
the estimated RUL. Detailed information can be
found in Yilboga et al.28
The state-based prognostic approach is one of the
major data-driven prognostic approaches. SSBP and
SBPD were selected for study in this paper since they
have advantages over other state-based prognostic
methods, e.g. hidden Markov model29 including computational complexity, simplicity and consistency of
results with dierent starting points in training.10
Neural networks are mature methods used in many
classication, clustering and forecasting problems. A
time-delayed neural network was selected for further
study, it is a type of neural network that is specically
used in forecasting studies.

RUL estimation accuracy (performance)

In this section, the RUL performances of the methodologies are discussed.
We were not able to obtain satisfactory results using
the physics-based method. Even though we tried hard
and spent most of our time on the physics-based model,
a full physics-based method could not be obtained. The
goal of using such a model was to obtain the expected
value of the sensor parameters (force, linear ruler, etc.)
using the current reading of the sensor. The residual
between expected and observed values was planned to
be used to forecast the failure. A gear box is located
right after the motor to transfer the circular movement
to linear movement. We needed the design of the gear
box, structure of the gears, and their parameters to calculate the relationship between circular and linear
movement. We were not allowed to open the gear box
due to the regulations and requirements of Turkish
State Railways. Thus, we had to approximate the relationship with extra sensors. However, the approximation was not good enough. We believe that the main
reason for the poor results obtained with the physicsbased model was an inability to obtain the full physical
equations and parameter values required in the equations to model the physical eect of the gear box in the
movement.
Data-driven models give good results. Acquired
data from each sensor was used as input to determine
the health status of the turnout system. Then, health
status was extrapolated until the failure state. The
RUL results of each sensor (ve sensors) and each
method (three methods) for 10 turnouts were
obtained. The rst eight turnouts were used for training and the other two were reserved for testing.
Figure 15(a) to (d) displays RUL estimations for
force, linear ruler, proximity, voltage and current sensors of all 10 turnout systems. The x-axis represents
the life span of a turnout system and the y-axis represents the estimated RUL value. Since the failure is

33
simulated, the unit of the life is not xed and can be
assumed as day, month or year depending on the realization of the simulated failure. Dashed linear black
lines represent the real RUL values. The remaining
three lines represent TDNN, SSBP and SBPD predictions. A closer match to the real RUL value means a
better prediction result of a method.
As seen from the gure, the accuracy of RUL estimation increases as the failure gets closer. This is obvious since the failure symptoms increase when a failure
approaches. However, in some of the results of TDNN,
a huge uctuation is observed. This indicates a lack of
robustness of the method. Two types of comparison
can be performed using the gure: methodology and
sensor comparison. In a methodology comparison, the
SBPD RUL predictions gave the closest matches to
real RUL values of all the methods. In a sensor comparison, the force sensor gives the best results.
Visual evaluation of the ranking of the eectiveness
of the methods and sensors is not sucient. Several
metrics have been proposed to evaluate the eectiveness of the prognostic technique.30 Three metrics are
used in this paper: prognostic horizon, a- accuracy
and cumulative relative accuracy. The prognostic
horizon indicates the dierence between the failure
time and the time that the estimation enters and
does not leave the desired specications. The a -
accuracy metric quanties the prediction accuracy in
RUL estimation at given time instances. a indicates
the desired accuracy and  indicates the time instance
as a percentage (one is failure time, one-half is half
way to the failure, zero is full life before failure). As
an example, Figure 15 shows the bounds of a as a
shaded area used in the a- accuracy metric for
RUL estimations for testing turnouts 9 and 10 using
ve sensors. The cumulative relative accuracy is the
normalized sum of the relative accuracies at given
time instances. Table 2 displays the mean values of
the discussed metrics for the turnouts used in the testing procedure. In addition to these parameters, The
RMSE and r-square indicators are used to assess the
performance of the RUL results. The r-square measures the similarity of the time series of the estimated
and real RUL time units. Real RUL values are known
since the data were collected from known discrete
health states. RMSE is the root mean square error
of the dierences between the real (yi) and the estimated (fi) RUL values. The equations of the r-square
and the RMSE are shown in equations (1) and (2)
where, y is the mean of the real RUL values and is
the number of observations used in equation (2)
P
yi  fi 2
r  square 1  Pi
2
i yi  y
s
Pn
2
i1 yi  fi
RMSE
n

34

Proc IMechE Part F: J Rail and Rapid Transit 230(1)

Figure 15. RUL plots for: (a) force sensor; (b) proximity sensor; (c) linear ruler sensor; and (d) current sensor.

As seen from the table, the force sensor leads to the

best results. The current and voltage sensors give reasonably good results. The proximity sensor is not a
good option for prognostics. As for the methods,
SBPD gives the best results.

Economic analysis versus performance

Even though the performance of a sensor (RUL estimation accuracy) is an important factor, the applicability of any engineering method needs economic
justication. Technological feasibility is not sucient

Camci et al.

35

Figure 15. Continued.

Table 2. Performance comparison of prognostic approaches for all the used methods and sensors.
SBPD

SSBP

TDNN

Sensor/
method

PH
(%)

a-
(%)

CRA

RMSE

r-sqr

PH
(%)

a-
(%)

CRA

RMSE

r-sqr

PH
(%)

a-
(%)

CRA

RMSE

r-sqr

Force
Linear Ruler
Proximity
Voltage
Current

94
94
94
94
94

79
44
65
68
74

0.96
0.91
0.90
0.93
0.92

0.69
0.63
0.61
0.65
1.44

0.99
0.84
0.69
0.80
0.96

94
6.3
53
6.3
50

29
0
3
0
3

0.87
0.60
0.62
0.59
0.68

2.10
6.10
5.80
4.58
6.41

0.88
0.67
0.49
0.71
0.80

94
65
53
41
28

18
21
18
15
15

0.79
0.76
0.45
0.74
0.71

4.05
3.50
4.35
4.30
5.20

0.91
0.94
0.44
0.60
0.66

for the application of a technology in real systems.

Economic feasibility is another factor that aects
the applicability of the technology. For example,
prognostics using the force sensor might give very
accurate results; however, it may not worth spending
a lot of money on the sensor. On the other hand, the
current sensor may give less accuracy in prognostics.
The small value that the sensor is adding may be

worth its application in a real system due to its low

cost. In this section, a methodology to analyse the
performance and cost factors of sensors is discussed.
For the bulk installation and multiple turnout systems the same sensor type with a dierent brand and
quality can be found and this might be adequate.
The two-encoder data could not be used for prognostic applications due to the ineectiveness of the

36

Proc IMechE Part F: J Rail and Rapid Transit 230(1)

Table 3. Pre-emptive optimization.

SBPD

1
2
3
4
5

SSBP

TDNN

Cost

r-square

RMSE

r-square

RMSE

r-square

RMSE

Proximity
Current
Voltage
Linear ruler
Force

Force
Current
Linear ruler
Voltage
Proximity

Proximity
Linear ruler
Voltage
Force
Current

Force
Current
Voltage
Linear ruler
Proximity

Force
Voltage
Proximity
Linear ruler
Current

Linear ruler
Force
Current
Voltage
Proximity

Linear ruler
Force
Voltage
Proximity
Current

Table 4. Sensor cost and r-square and RMSE combined as FP.

1
2
3
4
5

Sensor

Cost (E)

SBPD FP

SSBP FP

TDNN FP

Proximity
Current
Voltage
Linear ruler
Force

20
30
50
250
360

0.738,717
0.731,05
0.787,793
0.809,954
0.874,885

0.519,229
0.649,15
0.677,064
0.595,133
0.860,579

0.512,865
0.582,35
0.592,319
0.799,481
0.759,888

physics-based model. Thus, an analysis is performed

for the ve sensors, i.e. force, current, proximity,
linear ruler, and voltage sensors. In large- scale applications having ve sensors in each turnout system is
not feasible because of the cost, installation and operational diculties. The best sensor in terms of performance and cost should be selected. The analysis
of performance and cost is essentially a multi-objective optimization problem with the objectives of, rst,
lowest sensor cost and second, better prognostic
result.
If the optimization is performed pre-emptively by
considering cost as the priority, Table 3 is obtained (1
to 5: best to worst choice for the particular objectives).
Considering the cost dierence between a proximity
sensor and a force sensor it is not easy to perform the
sensor selection considering both the performance
and cost. Therefore, the multi-objective optimization
scheme can be used to make an optimal sensor
selection.
Multi-objective optimization (MOO) optimizes
conicting objectives such as cost, mass and deformation. Performance and cost could have an equal
weight in system design optimization or depending
on the importance level one may require more
weight than other. An objective function represents
designer preference. Multi-disciplinary design optimization involves several disciplines e.g. structures,
control, aero, manufacturing.
Unlike the single-objective optimization, a solution
to a MOO problem is more of a concept than a definition. There is no single global solution, and it is
crucial to determine a set of points that all t a prearranged denition for an optimum. The concept in
dening an optimal point is Pareto optimality. Later
Pareto optimum was adapted to engineering

problems. Applications of MOO to engineering

design grew over the following decades. A survey of
MOO methods for engineering31 provides a comprehensive review of methods and their variants for
engineering applications. The weighted global criterion method combines all objective functions to form a
single scalar function. The weights are related to a
decision-makers preference function. Graphical interpretation of the weighted sum method is provided for
two-objective problems to explain some of its deciencies. The Pareto curve provides a general picture for
MOO. The weighted sum approach nalizes the
decision.
An RMSE is normalized and converted to maximize the objective (second term) then r-square and
RMSE are combined to create a new term tness performance (FP) in equation (3). Table 4 indicates sensor
cost and tness performance for all three prognostic
methods.


1
1
RMSE
FP r  square
1
2
2
2RMSEmax

The order of the sensors in Table 3 and Table 4 is

from least expensive to most expensive. The datadriven prognostic methods (SBPD, SSBP, TDNN)
have dierent RMSE and r-square orders for the
same method. The parameter (FP) combines the
result of RMSE and r-square orders for SBPD,
SSBP and TDNN, separately.
Pareto optimum curve and weighted sum have
been applied as the MOO methods.
Pareto optimum curve. The cost versus tness performance for SBPD was plotted using the data in
Table 4. Figure 16 displays the Pareto curve that

Camci et al.

37

Figure 16. Pareto curve for SBPD fitness performance and cost.

starts from zero-performance and zero-cost and continues with proximity sensor (cost 20 and performance
0.738,717). The current sensor increases the cost to 30,
which increases the tness performance; therefore, the
second point is not included in the Pareto curve.
Pareto optimum points for SBPD are 1, 3, 5 (proximity, voltage, force sensors). Non-optimum Pareto
points are 2 and 4 (current sensor and linear ruler).
Non-optimum points are not shown on the Pareto
curve in Figure 16. By using a similar analogy,
Pareto curves for SBPD, SSBP, TDNN tness performances and costs are plotted in Figure 17.
Pareto curves provide a general picture for optimum sensor selection; however, there is no unique
optimum sensor. The weighted sum approach is
used in order to ease the problem of sensor selection
by decision-makers.
Weighted sum. Objective value zi (i 1, . . . , 5,
sensor number) is formulated in equation (4) where
w1, w2 are the weight of the cost and FP tness performance, respectively and w1 w2 1.
For equal importance w1 0.5, w2 0.5. A decision-maker can change the weights based on the
importance level. Cn is the normalized cost formulated
in equation (5). The number 360 in equation (5) is the
maximum sensor cost limit which can be varied by
decision-makers
z1 w1 Cn1 w2 FP1
Cn

360  cos t
360

4
5

The importance of cost versus importance of tness

performance is adjusted by w1 and w2. After the
weights are arranged, the weighted sum objective

Figure 17. Pareto curves for SBPD, SSBP, TDNN fitness

performances and costs.

values are calculated. The sensor with the highest

objective value is the best choice.
Example 1. When w1 0.5 w2 0.5 the SBPD prognostic method sensor weights are displayed in
Figure 18. Sensor 1 (proximity sensor) has the highest
weight (0.841,581), and is thus the best choice for the
given conditions.
When w1 is reduced and w2 is increased (w1 0.3,
w2 0.7), sensor 3 (voltage) becomes the best choice
(0.809,788). Note that these results are compatible
with those of the Pareto curve shown in Figure 17.
For example, non-optimum point 2 never appears

38

Proc IMechE Part F: J Rail and Rapid Transit 230(1)

Figure 18. Weighted sum for SBPD (w1 0.5, w2 0.5) and (w1 0.3, w2 0.7).

with highest weight no matter the values of w1 and w2.

Also, note that the results are highly dependent on the
weight values. w1 indicates the importance of the cost
whereas w2 indicates the importance of the accuracy
in RUL estimation. w1 is associated with the required
investment cost of the PHM technology and w2 is
associated with the consequence of under- or overprepared failure. Even though data exist to analyse
the investment cost for sensors, the latter is more difcult to estimate and depends on many factors such as
labour cost, unavailability cost, inventory cost, failure
cost, maintenance cost, etc. Thus, optimization of the
weight parameters is beyond the scope of this paper.

Sensor relationship approximations

In the section RUL estimation accuracy (performance), it is shown that force sensors are the best
options to estimate the RUL accuracy; this is because
force sensors are more sensitive to the health of the
turnout system. However, force sensors are more
expensive and dicult to use. Can we mathematically
obtain the value of the force sensor through a function using the current, voltage or proximity sensor
values as input? If so, we may obtain good results at
a lower cost. If such a relationship is possible, then
installation of cheap sensors will be sucient. This
section presents a study to obtain a reliable model
to estimate the force values in terms of the current,
voltage and/or proximity signals.
Figure 19 illustrates the relationship between the
force sensor and other sensors. The gure suggests
that there is no direct relationship if the sensors are
individually compared.
We tried to nd if the force sensor could be represented as a function of the other sensors. Here, the
regression technique was used to search for such a
relationship. The idea was that each sensor is considered as a variable and the corresponding sensor
reading, i.e. time series was considered as a value of
this variable. Any relation between these variables
was then analysed. We rst applied the linear regression method to check if the force sensor is a linear
function of the current, voltage and proximity

sensors. Based on the regression results, the following

transformation function was obtained
ForceEstimatedLinear 0:032Current 0:183
Proximity1 0:151
Proximity2 0:0764
Voltage1 0:063
Voltage2 2:289
Note that the regression method does not include
any insignicant variable in the regression function,
so we could not remove any of the parameters from
the equation. In our case, the current, voltage and
proximity sensors are found to be signicant variables, so they all have a relationship with the behaviour of the force sensor, and are included in the
regression function.
All the ve explanatory variables are highly statistically signicant with p-values of less than 1% indicating their strong explanatory powers in the
regression. In addition the p-value of the overall
regression measured as the p-value of the F-statistic
is very low. The p-value of the F-statistic measures the
overall signicance of the regression. The lower the
p-value, the greater is the likelihood that the reported
relationship is real and has not occured at random.
Hence, the zero p-value to three decimal points indicates that the variables in the equation do explain
some of the variation of the ForceEstimatedLinear.
In fact, according to r-squared value of the regression,
the ve variables explain 81.89% of the variation in
ForceEstimatedLinear.
Figure 20 plots the exact force sensor
values (Force) and the estimated force values
(ForceEstimatedLinear). Although the estimated
force sensor values give a better representation compared with the individual sensors in Figure 19, this
model does not provide the clear mathematical relationship for which we are searching.
The issue with the method we have applied is that
we have restricted our search to linear models.
However, the relationship between the force and
other sensors could be arbitrary. We applied another

Camci et al.

39

Figure 19. Relationship between force sensor and current, voltage and proximity sensors.

regression technique, called nonlinear regression. In

this method, the relationship between dependent and
independent variables is estimated by mathematically
arbitrary models.
To nd such a model, we rst applied the nonlinear
regression method instead of the linear regression
method. However, we could not obtain a better
result than Figure 20, because we analysed several
sensors as independent variables at the same time
and therefore the estimation algorithms employed in
this technique could not easily guess the relationship
between the dependent variable and each independent
variable.
We therefore applied the regression in two stages.
In the rst stage, we rst ran the linear regression as
dened above. Since some of the estimated force
sensor values are negative and very small, we normalized the estimated force sensor values by taking their
power as a function of 2, i.e. 2ForceEstimatedLinear.
We then ran the nonlinear regression in the second
stage. The dependent variable ForceEstimatedLinear
of Stage 1 became the independent variable of Stage
2. That is, we guessed a nonlinear model for the new

variable ForceEstimatedNonLinear in terms of

ForceEstimatedLinear. We guessed the following
function
ForceEstimatedNonLinear

1
ForceEstimatedLinear
b0 b1  b22

We then applied the nonlinear regression method.

The nonlinear regression analysis returned the following values: b0 0, b1 1:16 and b2 0:676.
Finally, we found the relation between Force and
ForceEstimatedNonlinear. Figure 21 plots these two
variables.
We guessed a function for the estimated force
values with the following form

ForceEstimated c0 c1 

1
ForceEstimatedNonLinear

We found the values c0 2:98 and c1 3:8 with

an r-squared value of 93%, which implies a very high
accuracy.

40

Proc IMechE Part F: J Rail and Rapid Transit 230(1)

Figure 20. Force versus ForceEstimatedLinear.

Figure 21. Force versus ForceEstimatedNonLinear.

We can now easily justify the accuracy of the

regression model by plotting the measured and estimated force values. As illustrated in Figure 22, we
have a very precise tting.
Thus, by applying linear and nonlinear regression
in two stages (along with variable normalization), we

could obtain a very high accuracy when tting our

model. This suggests that rather than using the raw
data from the current, voltage and proximity sensors,
we process them using the above functions, and generate a new data set, which simulates the force sensor
and has a better representation of failure degradation.

Camci et al.

41

Figure 22. Exact force values versus estimated force values.

Conclusions
This paper compares various sensors and dierent
prognostics methods for railway turnout systems.
Our results can be summarized as follows: First, sensors have been compared based on several features
used in the prognostics analysis of railway turnouts.
Natural failure degradation by creating discrete failure states has been modelled. RUL performances of
sensors have been evaluated. Our results show that
force sensors have better representation of failure degradation, so it is the best option to use for prognostic
input. Based on these results, it is observed that the
linear ruler and current sensors are the second-best
option, and the proximity sensor is not a good
option for prognostics. The performance and cost factors of sensors are also analysed using multi-objective
optimization to make an economic justication of the
optimal sensor selection. In addition, the possibility of
approximating to the values of expensive sensors is
investigated using the values of cheaper sensors
obtained from a prototype system. Thus, rather than
using the raw data from the current, proximity and
voltage sensors, they are processed using a function
that converts these values into force sensor values,
and generate a new data set, which simulates the
force sensor and has a better representation of failure
degradation. The function approximation methodologies might then avoid deployment of expensive sensors. This area should be further explored in future
studies.

Funding
This research was partially supported by The Scientic and
Technological Research Council of Turkey (TUBITAK)
under project 108M275.

References
1. European Union. EU corporate responsibility. Report
no., 2005. Brussels: European Union.
2. Xia HW, Ni YQ, Wong KY, et al. Reliability-based condition assessment of in-service bridges using mixture distribution models. Comput Struct 2012; 106: 204213.
3. Hada A, Soga K, Liu R, et al. Lagrangian heuristic
method for the wireless sensor network design problem
in railway structural health monitoring. Mech Syst
Signal Process 2012; 28: 2035.
4. Walentine DT, Kube CM, Turner JA, et al. Diffuse
ultrasonic backscatter measurements for monitoring
stress in rail. J Acoust Soc Am 2012; 132(3): 1961.
5. Marquez FPG, Lewis WR, Tobias MA and Roberts C.
Life cycle cost for railway condition monitoring.
Transportation Research Part E: Logistics and
Transportation Review 2008; 4(6): 11751187.
6. Marquez FPG, Schmid F and Collado JC. A reliability
centered approach to remote condition monitoring: a
railway points case study. Reliab Engng. Syst Saf 2003;
80(1): 3340.
7. Marquez FPG, Tercero DJP and Schmid F. Unobserved
component models applied to the assessment of wear in
railway points: a case study. Eur J Oper Res 2007; 176(3):
17031712.
8. Marquez FPG, Roberts C and Tobias AM. Railway point
mechanisms: condition monitoring and fault detection. Proc
IMechE, Part F: J Rail Rapid Transit 2010; 224(1): 3544.

42
9. Eker OF, Camci F, Guclu A, et al. A simple state-based
prognostic model for railway turnout systems. IEEE
Trans Ind Electron 2011; 58(5): 17181726.
10. Eker OF and Camci F. State based prognostics with
state duration information. Qual Reliab Engng Int
2013; 29(4): 465476.
11. Camci F and Chinnam RB. Health-state estimation and
prognostics in machining processes. IEEE Trans Autom
Sci Engng 2010; 7(3): 581596.
12. Ming D and Ying P. Equipment PHM using non-stationary segmental hidden semi-Markov model. Robot
Comput-Integr Mfg 2011; 27(3): 581590.
13. Ying P and Ming D. A prognosis method using agedependent hidden semi-Markov model for equipment
health prediction. Mech Syst Signal Process 2011;
25(1): 237252.
14. Peng Y and Dong M. A hybrid approach of HMM and
grey model for age-dependent health prediction of
engineering assets. Expert Syst Appl 2011; 38:
1294612953.
15. Dong M and He D. A segmental hidden semi-Markov
model (HSMM)-based diagnostics and prognostics
framework and methodology. Mech Syst Signal
Process 2007; 21: 22482266.
16. Jardine A, Lin D and Banjevic D. A review on machinery diagnostics and prognostics implementing condition-based maintenance. Mech Syst Signal Process
2006; 20(7): 14831510.
17. Kothamasu R, Huang SH and VerDuin WH. System
health monitoring and prognostics a review of current paradigms and practices. Int J Adv Mfg Technol
2006; 28(910): 10121024.
18. Zhang L, Li X and Yu J. A review of fault prognostics
in condition based maintenance. Proc SPIE 2006; 6357:
635752.
19. Uckun S, Goebel K and Lucas PJF. Standardizing
research methods for prognostics. In: The international
conference on prognostics and health management,
Denver CO, 69 October 2008. PHM 2008.
20. Gebraeel N, Elwany A and Pan J. Residual life predictions in the absence of prior degradation knowledge.
IEEE Trans Rel 2009; 58(1): 106117.
21. Peng Y, Dong M and Zuo MJ. Current status of
machine
prognostics
in
condition-based

Proc IMechE Part F: J Rail and Rapid Transit 230(1)

22.

23.

24.

25.

26.

27.

28.

29.

30.

31.

maintenance: a review. Int J Adv Mfg Technol 2010;

50(14): 297313.
Xiao-Sheng S, Wenbin W, Chang-Hua H, et al.
Remaining useful life estimation - a review on the statistical data driven approaches. Eur J Oper Res 2011;
213(1): 114.
Sun B, Zeng S, Kang R and Pecht MG. Benefits and
challenges of system prognostics. IEEE Trans Reliab
2012; 61(2): 323335.
Zhang H, Kang R and Pecht M. A hybrid prognostics
and health management approach for condition-based
maintenance. In: The IEEE international conference on
industrial engineering and engineering management, 811
December 2009, p.1169. Piscataway, NJ: IEEE Press.
Tingan T. Application of physical failure models to
enable usage and load based maintenance. Reliab
Engng Syst Saf 2010; 95: 10611075.
Heng A, Zhang S, Tan A and Mathew J. Rotating
machinery prognostics: state of the art, challenges and
opportunities. Mech Syst Signal Process 2009; 23(3):
724739.
Gan G, Ma C and Wu J. Data clustering: theory, algorithms, and applications. Philadelphia, PA: Society for
Industrial and Applied Mathematics, 2007.
Yilboga H, Eker OF, Guclu A and Camci F. Failure
prediction on railway turnouts using time delay neural
networks. In: The IEEE international conference on
computational intelligence for measurement systems and
applications, Italy, 2010, pp.134137. Piscataway, NJ:
IEEE Press.
Camci F and Chinnam RB. Dynamic Bayesian networks for machine diagnostics: hierarchical hidden
Markov models vs. competitive learning. In: The international joint conference on neural networks, Montreal,
Canada, 31 July 4 August, 2005, vol. 3, p.1752.
Saxena A, Celaya J, Balaban E, et al. Metrics for evaluating performance of prognostic techniques. In: The
international conference on prognostics and health management, Denver, CO, 69 October 2008, pp.117.
Marler RT and Arora JS. Survey of multi-objective
optimization methods for engineering. Struct
Multidiscip Optim 2004; 26: 369395.