Reliability Engineering and System Safety 80 (2003) 33–40

www.elsevier.com/locate/ress

A reliability centered approach to remote condition monitoring.

A railway points case study
Fausto Pedro Garcı́a Márqueza,*, Felix Schmidb, Javier Conde Colladoa
a
ETSICCP, Universidad de Castilla-La Mancha, Ciudad Real, Spain
b
Department of Mechanical Engineering, University of Sheffield, Sheffield, UK
Received 1 September 2001; accepted 15 August 2002

Abstract
Railway turnouts, consisting of switches and a crossing, are complex electro-mechanical devices which are exposed to severe
environmental influences and which are essential for the operation of any railway bar horizontal lifts. Their safe and reliable operation must
be assured if the rail mode of transport is to flourish. Conventionally, the continuous availability of turnout mechanisms has been assured by
high levels of routine maintenance, to some extent tailored to the criticality of a particular point location. However, traffic increases and
shortened maintenance windows require better approaches to turnout maintenance. The authors of the present paper undertook the
development of algorithms to detect gradual failure in railway turnout which should allow a move to an RCM2 approach to the management
of switch and crossing maintenance. They demonstrate the approach using data from tests on a commonly found point mechanism and
include a discussion of the benefits of adopting a Kalman Filter for pre-processing the data collected during tests.
q 2003 Elsevier Science Ltd. All rights reserved.
Keywords: Remote condition monitoring; Reliability centred maintenance; Kalman Filter; Point mechanism; Failure mode and effect analysis

1. Introduction the railway subsystems. Provision of a reliable infrastruc-

ture plays a very important role in achieving a safe system.
The safety of staff, customers and of the general public The authors of the present paper discuss the development of
in general viewed as one of the most important require- an approach to the application of remote condition monitoring
ments in industry and is of particular importance in the to the reliability centred maintenance of railway turnouts.
railway industry, where passenger rightly expert vary high Section 2 provides the background to the issues, a description
standards of care. In Britain the Railway Regulations were of the test system used in Sections 3 and 4, followed by an
introduced in 1994 and a new safety culture was outline of the criteria in Section 5. A Kalman Filter approach to
established. This was a necessary part of the privatisation the preprocessing of data is given in Section 6, while the
process for British Railways (BR). It was also a results and conclusion from Sections 7 and 8, respectively.
consequence of the realisation that, from 1989 to 1994
alone, 825 members of the general public had been killed 2. Background
on BR. Most of these died as a result of trespass but
substantial numbers by falling from trains [7], due to 2.1. Railway turnouts
design faults and poor maintenance. Any accidents on the
railways though are of serious concern to society, As part of a guided transportation system, a train can
particularly so since the accidents at Shouthall, Ladbroke move from one track to another only in certain places, that
Grave, Hatfield, and, most recently, Potters Bar in Britain, is, where an appropriate mechanical device has been
Eschede and Brühl in Germany and Norway. Ever since installed. This device is known as a turnout.1 The turnout
its inception though, the railway industry has searched has moving parts which are called switches (US: blades) and
for ways to improve the safety and the reliability of which steer the trains in one of two directions, normal
1
* Corresponding author. Some authors refer to turnouts as points, that is, the parts of the turnout
E-mail address: [email protected] (F.P. Garcı́a Márquez). which move to allows trains to change track to enter siding.

0951-8320/03/$ - see front matter q 2003 Elsevier Science Ltd. All rights reserved.
PII: S 0 9 5 1 - 8 3 2 0 ( 0 2 ) 0 0 1 6 6 - 7
34 F.P. Garcı́a Márquez et al. / Reliability Engineering and System Safety 80 (2003) 33–40

(straight through) or reverse. The switches move from maintenance approach to infrastructure and logistic of
normal to reverse or reverse to normal direction. Turnouts railway operation), all of which justify the contribution of
are perhaps the most important infrastructure elements of RCM1 for point mechanism in improving the safety and
the railway system and affect its safety greatly. The Potters reliability of railways.
Bar accident of 10th May, 2002 in England was caused by FMEA is a systematic analysis of the potential failure
a faulty turnout while the consequences of the Eschede modes of a component of a system [9]. It includes the
accident in Germany were aggravated by a point moving identification of possible failure modes, determination of
underneath the train. the potential causes and consequences and an analysis of the
The 55% of railway infrastructure component failures on associated risk. It also includes a record of corrective actions
high speed lines are due to signalling equipment and turnouts or controls implemented resulting in a detailed control plan.
[10], where ‘signalling equipment’ covers signals, track FMEAs can be performed on both the product and the
circuits, interlockings, automatic train protection (ATP) or process. Typically, an FMEA is performed at the component
LZB (track loop-based ATP), and the traffic control centre. level, starting with potential failures and then tracing their
From another point of view, the annual cost of maintaining effects up to the ultimate consequences. The FMEA allows
points is high, about UKP (United Kingdom Pound) 3.4 the identification of the most critical components and the
million per year for about 1000 km of railway, compared to likely failure mechanisms, thus leading to the specification
other infrastructure elements. TC-TCR trade circuits, for of system parameters to be monitored.
example, cost UKP 2.1 million per year for the same area. Of Primary performance parameters of complex mechan-
the points expenditure, UKP 1.2 million is for clamp lock isms, such as railway points, are speed of movement,
(hydraulic) turnout and UKP 1.4 million for electrically vibration, supply voltage, power, throwing time, tempera-
operated turnouts (data provided by a British asset manager). ture, current, force, etc. Based on these performance
Turnouts can also be used to implement flank protection for a parameters, RCM1 can be used to define terms such as
train route allocated to another train. This is achieved by risk, quality, control, comfort, economy, containment,
positioning the blades of the turnout in such a way that a train ergonomy, etc. In practice, condition-based maintenance
driving through the turnout is not directed into a track decisions are based substantially upon assessments of the
segment belonging to the route of the other train. condition of the system obtained at discrete monitoring time
The two safe positions of the moving parts of a turnout, intervals [15]. This type of condition monitoring is called
normal and reverse, are generally detected using switches indirect condition monitoring in contrast to direct monitor-
operated by the blades or their operating mechanisms. In ing which measures the actual condition. The latter employs
order to ensure high availability and reliable and safe advanced electronics, sensors and transducers, computing
operation, points require regular inspection and mainten- and communications technology. Their measurement
ance. Currently, such maintenance is carried out on a time (vibration, supply voltage, power) can be embodied in
basis with allowance being made for the operational remote condition monitoring systems (RCM2) [2 –4,11].
criticality of a particular point. A better and more cost- RCM2 leads to improved reliability and can pay for itself in
effective approach though is reliability centred maintenance terms of cost-effectiveness since staff do not have to visit
which is being adopted by a number of railway undertakings installations as frequently. The integration of the two types
for point mechanism maintenance. of RCMi is called RCM2, with the overall aim of using
advanced electronics, control, computing and communi-
2.2. Turnout failures and maintenance cation technologies to address the multiple objectives of
cost effectiveness, improved reliability and services.
Reliability centred maintenance (RCM1) is a process used In addition to the data collection required for RCM2, it is
to decide what must be done to ensure that any physical asset, also necessary to process the large amount of information to
system or process continues to do whatever its users want it provide a warning when the device moves out of tolerance or
to do [4,8,12]. Therefore, RCM1 provides powerful rules for adjustments. Algorithm design for the detection of trends
deciding whether a failure management policy is technically and failure patterns has been undertaken by many research-
appropriate, providing precise criteria for deciding how often ers but only a few papers dealing with the dynamics of
routine tasks should be carried out. RCM1 identifies ways in railway turnouts have been found in the literature [1,4,13,
which the system can fail to live up these expectations. This 14]. The present paper describes a simple approach to RCM2
must generally be followed by a failure mode and effects as applied to railway turnout mechanisms in a case study.
analysis (FMEA) which allows an assessment of the
consequences of failure. As substantial number of research
projects concerning railway infrastructure have been carried 3. Description of turnout used in experiment
out or are still in progress, for example, REMAIN (reliability
and maintainability in European Rail Transport), ROMAIN Turnouts are assembled from switches and a crossing
(Railway Open Maintenance Tool), INFRACOST (The Cost where the moving parts are often described as the ‘points’.
of Railway Infrastructure), RAIL (reliability centred Most standard point machines contain a switch and lock
 F.P. Garcı́a Márquez et al. / Reliability Engineering and System Safety 80 (2003) 33–40 35

mechanism which includes a hand-throw lever and a selector

lever to allow operation by power or hand. The mechanism
is normally divided into three major subsystems: (i) the
motor unit which includes a contactor control arrangement
and a termination area, (ii) a gearbox comprising spur-gears
and a worm reduction unit with overload clutch and the dual
control mechanism as well as (iii) the controller subsystem
with motor cut-off and detection contacts. Generally, there
are also mechanical linkages for detection and locking of the
point. The standard railway point is therefore a complex Fig. 2. Railway turnout.
electro-mechanical device with many potential failure
modes.
In the following description reference is made to a
particular type of turnout which is in wide-spread use in UK split-field and permanent magnet field for AC immune
and elsewhere but which cannot be mentioned for reasons of machines. Where the machine will be operated on AC
confidentiality. The point machine is normally non-handed supplies a silicon bridge rectifier is built into the machine.
with respect to the locking bars and detection rods but it Points failures can occur for a number of reasons, ranging
would be necessary to place the hand-throw and selector from foreign objects blocking the switch rails to loss of
levers on the side furthest from the rails. Machines can be position detection due to microswitches being dislodged
ordered and supplied with these levers on the appropriate because of vibration. Failures can also be caused by
side but conversion can be made during installation. excessive friction on movement of the format in the ballast.
The circuit controller includes detection switches and a
pair of snap-action switches to stop the machine at the end
of its stroke and to regeneratively brake the motor so that the 4. The prototype remote condition monitoring system
mechanism is not subjected to impacts. The detection
As part of a research project, the development of a simple
switches have high pressure wiping contacts made of
RCM2 system for turnout monitoring was undertaken. The
silver/cadmium oxide or gold and they are operated by both
motor current was measured using non-intrusive current
the lockbox and the detection rod. The detection switches
transformers mounted within the point machine housing.
have additional contacts to allow mid-stroke short circuiting
The force in the drive bar was measured by replacing the
of the detection relays to avoid wrong indications in the
bolted connection between the drive bar and the drive rod
signal box (Figs. 1 and 2).
with a load-cell. Trigger signals were taken from adjacent
The control contactor is fixed in the motor compartment
signalling locations utilising spare contacts and wired as
with a plug connection to the wiring. It has heavy duty
part of an instrumentation engineer’s installation. The
silver/cadmium oxide wiping contacts which reverse
power supply for the installation was taken from the
the motor and additional contacts for integration with
110 V supply in an adjacent relay room. Sensors were
detection. The contactor is a two position, polarised, connected locally to a connection box mounted in close
magnetic stiction type requiring 2 W to operate. proximity to the points. This box contained no active
The DC-powered motor is a special heavy-duty design components but was equipped with lightning protection for
developed specifically for point machine use. It is plugged the positional detector. For the initial tests, the data were
into the wiring and is supported by a machined spigot collected using a conventional data logger and then
entering the gearbox and two mounting lugs fixed to the processed on a personal computer.
base. The motor can take two basic forms, series-wound

5. Model criteria

Faults in point mechanisms must be detected quickly and

reliably if the information is to be useful. It is a discrete
dynamic system, where data must be processed in real time.
Any detection algorithm must therefore use a simple model
for detecting faults rapidly by analysing data in real time. It
is thus better to apply statistical approaches rather than
simply comparing new data with ‘good’ and ‘bad’ signals
stored earlier. Also, the model for detecting faults must
adapt to external conditions, i.e. changes in the environment
Fig. 1. Point mechanism. (humidity, temperature climatic, etc.) friction forces, etc.
36 F.P. Garcı́a Márquez et al. / Reliability Engineering and System Safety 80 (2003) 33–40

and the model must detect faults in both directions of the as follows
turnout mechanism movement. This was the reason for the ! n¼t j
authors to choose a reference dynamic system for their
j
tmax Xmax Xj
m¼T

j dnj < dmj ; ð3Þ

analysis. This is based on calculating individually for T j 2 tmax n¼0 j
m¼tmax
normal to reverse and reverse to normal operations,
the arithmetic mean between the previous reference data j
where tmax and T j are the maximum position and total time,
set and the latest new data set for operation in that direction, P max j
j
tmax j
=ðT j 2 tmax Þ is the area relationship, n¼t j
n¼0 dn and
once I has been proved to be fault free (Eq. (1)). Pm¼T j j j
j dm are the areas under and above tmax in the j actual
FMEA was employed for the identification of possible m¼tmax
failure modes, determination of the potential causes and curve of d. Any set of test data that satisfies the three criteria
consequences and an analysis of the associated risk. A will be considered to be the result of a test without fault and
sample list of faults is shown in Appendix A. It also includes is defined ‘as commissioned’. This is shown in Fig. 3(c).
a record of corrective actions or controls implemented Although using these criteria has allowed a significant
resulting in a detailed control plan improvement in detecting faults by analysing changes from
data set to data set, sensitivity values can be applied for
xs21 þ xi j further improvement. To increase, the reliability of the last
xsi ¼ i
; ð1Þ
2 criterion, we have employed a Kalman Filter that is
where s 2 1 is the previous reference data set; j, the new described in Section 6.
data set (fault free), s, the new reference data set with xi is
the data value at time-step i.
The data collected refers to force (N) versus time (s). The 6. Kalman Filter approach to data in RCM2
first conclusion after studying these curves was that we
could not detect the faults directly by analysing the curves. The Kalman Filter [6] approach is based on a set of
However, if we analyse the difference between the actual mathematical equations that provide an efficient compu-
data xj and the reference data xs in the form of absolute tational solution using the least squared method. The
values ðdj Þ; we can detect the majority of faults as they Kalman Filter addresses the general problem of trying to
develop estimate the state x of a discrete-time controlled process that
is governed by a linear stochastic difference equation. The
dij ¼ lxji 2 xsi l: ð2Þ objective in using Kalman filtering in this study was to
The tests carried out where no faults are present, that is, the ‘as increase the reliability of the model presented to the rule-
commissioned’ tests result in curves d which are very similar. based decision mechanism. The Kalman Filter estimates a
They are symmetric with respect to the maximum position process by using a form of feedback control: the filter
and do not display major irregularities, although their estimates the process state at some time and then obtains
amplitudes are different, and the maximum position is similar feedback in the form of measurements.
but not equal in all cases. These are the signal characteristics The equations for the Kalman Filter fall into two groups:
considered to detect faults, described as follows. time update equations and measurement update equations.
A first criterion is based on whether the shape of the test The time update equations are responsible for projecting
data curve is irregular. If this is the case it is assumed to be forward the current state and error covariance estimates to
the consequence of a fault. The detection of irregularity in obtain the a priori estimates for the next time step. The
the curves requires the application of a sensitivity value for measurement update equations are responsible for the
perturbations in the signal (Fig. 3(a)). feedback. The time update equations can also be thought
If the data is not sufficiently distinctive to detect a fault of as predictor equations, while the measurement update
with the first criterion we can check the data gathered in test equations can be thought of as corrector equations. Indeed,
j to find the maximum position tmax j
of the curve dj : We may the final estimation algorithm resembles that of a predictor-
find that it is not the same as the maximum position tmax i
of corrector algorithm for solving numerical problems [2,3].
the current reference curve xr : Allowing for the margin tmg,
s The state d and measurement m of the process are
a signal is considered to be indicative of a fault if the governed by the linear stochastic difference equations:
position of the maximum is outside the band (Fig. 3(b)). diþ1 ¼ Ai di þ Bi ui þ wi ; ð4Þ
Finally, if the first two criteria have not resulted in the
detection of a fault, then a third criterion is applied whether mi ¼ Hi di þ vi ; ð5Þ
the curve is symmetric with respect to the maximum where vi and wi are the process and measurement noises,
position, again with a margin of a given width. If this is the respectively. They are both assumed to have the character-
case then it will be assumed to be a fault. This criterion istic of Gaussian white noise. Their mean and covariances
cannot be used for detecting a fault in real time. are
This supposition has been demonstrated with numerous
simulations and experiments. The mathematical analysis is E½wi  ¼ 0; E½w2i  ¼ Qi ; ð6Þ
 F.P. Garcı́a Márquez et al. / Reliability Engineering and System Safety 80 (2003) 33–40 37

Fig. 3. Criteria employed for detecting faults in a point mechanism. (a) First criterion, (b) second criterion and (c) third criterion.
38 F.P. Garcı́a Márquez et al. / Reliability Engineering and System Safety 80 (2003) 33–40

Fig. 4. Difference between the actual reference curve and the new curve in absolute values with and without Kalman Filter. (a) Normal to reverse direction and
(b) reverse to normal direction.

to normal, respectively. Each diagram shows both the actual

E½vi  ¼ 0; E½v2i  ¼ Ri ; ð7Þ
measurement and a filtered measurement.
E½wi vi  ¼ 0; ð8Þ
where E is a statistical averaging operator. 7. Results
The values of the a posteriori state estimate d^ 2
i and error
covariance Pi are as follows The reported RCM2 was developed as part of a research
d^ i ¼ d^ 2 ^2 project between the Department of Mechanical Engineering
i þ Kðmi 2 Hi x i Þ; ð9Þ
at the University of Sheffield, England, and an industrial
Pi ¼ E½ðdi 2 d^ i Þ2 ; ð10Þ partner. The data collection and algorithm developments
were carried out during the year 2000 with a number of
where K is the blending factor that minimises Pi (the valour experiments at a test site.
K in Eq. (9) is calculated in Ref. [5]), and d^ 2
i is the a priori The partner company arranged for a suitable turnout
state estimate. The priori estimate error covariance is as assembly, fitted with a point machine of a type common in
follows: the UK and elsewhere, and made it available in a fully
Pi ¼ E½ðdi 2 d^ 2 2 instrumented form for the collection of test data. Test data
i Þ : ð11Þ
were collected with the mechanism both in perfect working
The output from a Kalman Filter application to the problem order, and also with a number of previously encountered
in hand is shown in Fig. 4. Fig. 4(a) and (b) shows the forces real life fault conditions created artificially in turn. Data
measured for transitions from normal to reverse and reverse were recorded for each fault condition at a variety of levels
 F.P. Garcı́a Márquez et al. / Reliability Engineering and System Safety 80 (2003) 33–40 39

of severity (Appendix A). A total of 151 experiments were the rule-based decision mechanism. The Kalman Filter
carried out, 79 in the reverse to normal direction and 72 in estimates a process by using a form of feedback control: the
the normal to reverse direction. filter estimates the process state at some time and then obtains
The most important results are as follows. With a Kalman feedback in the form of measurements.
Filter, we could detect 100% of faults in the reverse to normal With a Kalman Filter, the authors can currently detect
direction in the 79 experiments. Without the Kalman Filter 100% of faults in the reverse to normal direction, and
this drops to 97.33%. In this direction, the margin employed without the Kalman Filter this drops to 97.33%. In the other
for detecting the maximum position is 0.3 s less when using direction, we can currently detect only 97.1% of faults with
the Kalman Filter, and the margin considered for detecting the Kalman Filter and without it only a 94.2%. In general,
irregularities in curves is reduced to 91.3%. In the other employing Kalman Filter has improved the margins of
direction, we can currently detect only 97.1% of faults when criteria in both directions.
using the Kalman Filter and without it only 94.2%. The The authors cannot explain why detection in the normal
margin for detecting irregularities is 85.71% better, and the to reverse is not as successful as for reverse to normal, but in
margin in maximum position is 0.3 s worse. their opinion they have achieved a minimum rejection rate
but they continue to improve their methods.

8. Conclusions
Appendix A. Sample list of faults
Turnouts are probably the most important infrastructure
elements of the railway system and affect its safety greatly. † 15 mm obstruction at second bearer on normal side of
The standard railway turnout is a complex electro- points;
mechanical device with many potential failure modes. In † 13 mm obstruction at eighth bearer on reverse side of
order to ensure high availability and reliable and safe points;
operation, points require regular inspection and † 12 mm obstruction at toe on normal side of points;
maintenance. † Back drive overdriving at heel on normal side with dry
Reliability-centred maintenance (RCM1) provides slide chairs;
powerful rules for deciding whether a failure management † Back drive slack end off at toe end;
policy is technically appropriate, providing precise criteria † Back drive slack end off at toe end (LHS side drive
for deciding how often routine tasks should be carried out. basket slack end off);
The technique employs advanced electronics, sensors and † Diode snubbing block disconnected;
transducers, computing and communications technology † Dry slide chairs;
embodied in remote condition monitoring systems (RCM2). † Low tension on motor brush;
RCM2 leads to improved reliability and can pay for itself in † Operational contact in original position;
terms of cost-effectiveness since staff do not have to visit † Tight lock on reverse side;
installations as frequently. The integration of the two types † Tight lock on reverse side (sand on all bearers on both
of RCMi is called RCM2. The authors of the present paper sides).
have described a simple approach to RCM2 as applied to
railway turnout mechanisms, based on a case study.
Faults in turnout mechanisms must be detected quickly References
and reliably if the information is to be useful. It is a discrete
dynamic system, where data must be processed on line. Any [1] Andersson C, Dahlberg T. Wheel/rail impacts at a railway turnout
detection to be adopted must use a simple model for crossing. J Rail Rapid Transit 1998;212(2):123–34.
[2] Christer AH, Wang W. A simple condition monitoring model for a
detecting faults quickly by analysing data in real time. The
direct monitoring process. Eur J Oper Res 1995;82:258– 69.
model for detecting faults must adapt to external conditions [3] Christer AH, Wang W, Sharp JM. A state space condition monitoring
and must detect faults in both directions of the turnout model for furnace erosion prediction and replacement. Eur J Oper Res
mechanism movement. This was the reason for the authors 1997;101:1 –14.
to choose a reference dynamic system for their analysis. [4] Garcı́a Márquez FP, Schmid F, Conde J. Mantenimiento Centrado en
la Fiabilidad y Monitorización Remota Basada en la Condición,
The RCM2 implementation was developed as part of a
RCM2: Un caso de Estudio. Gestión de Activos Industriales 2002; in
research project. The data collection and algorithm devel- press.
opments were carried out during the year 2000 with a number [5] Jacobs OLR. Introduction to control theory, 2nd ed. New York:
of 151 experiments at a test site. The data collected refers to Oxford University Press; 1993.
force (N) versus time (s). If we analyse the difference [6] Kalman RE. A new approach to linear filtering and prediction
problems. Trans ASME J Basic Engng 1960;35–45.
between the actual data and the reference data in the form of
[7] Kennedy A. Risk management and assessment for rolling stock safety
absolute values, we can detect the majority of faults as they cases. Proc Inst Mech Engrs: Part F 1997;211:67 –72.
develop. The objective for using Kalman filtering in this [8] Moubray J. Reliability-centred maintenance. Oxford, UK: Butter-
study was to increase the reliability of the model presented to worth/Heinemannn; 1997.
40 F.P. Garcı́a Márquez et al. / Reliability Engineering and System Safety 80 (2003) 33–40

[9] Price CJ, Pragh DR, Wilson MS, Snooke N. The flame system: USA: International Society of Automotive Engineers Department;
automating electrical failure mode and effects analysis (FMEA). Proc 1999.
Reliab Maintain Symp 1995;90–5. [13] Shimonae T, Kawakami T, Miki H, Matsuda O, Tekeuchi H.
[10] REMAIN Consortium. Modular system for reliability and maintain- Development of a monitoring system for electric point machines.
ability management in European rail transport. Final Report, IITB; IRSE Aspect Int Conf 1991;395–401.
1998. [14] Stott PF. Automatic open level crossing a review of safety. London,
[11] Roberts C, Fararooy S. Remote condition monitoring into the next UK: Her Majesty’s Stationery Office; 1987.
millennium. Proc Comput Aid Des Manuf Oper Railway 1998;. [15] Wang W. A model to determine the optimal critical level and the
[12] SAE Standard JA1011, Evaluation criteria for reliability-centered monitoring in condition based maintenance. Technical Report CMS-
maintenance (RCM) process. Commonwealth Drive Warrendale, 99-07; 1999.