Professional Documents
Culture Documents
The manuscript was received on 14 September 2009 and was accepted after revision for publication on 18 February 2010.
DOI: 10.1243/09544097JRRT313
Abstract: INNOTRACK is a project funded under the European Commission Sixth Framework
research programme. The project aims to develop approaches capable of achieving a 30 per cent
reduction in track life-cycle costs (LCCs). As part of a cost consolidation exercise within the
project, it was identified that switch and crossing maintenance and inspections account for
around 19 per cent of the total maintenance costs. Improved condition monitoring can be used
as part of a condition-based maintenance regime, which saves money over traditional periodic
maintenance. This paper presents a novel algorithm that has been developed, which uses quali-
tative trend analysis (QTA) to detect and diagnose incipient faults in switches, which have been
difficult to detect using current commercial methods. The algorithm is demonstrated using fault
simulation data collected from DC electric switch actuators of a type widespread in the UK.
The increased fault diagnosis capability has the potential to contribute significantly towards the
achievement of the 30 per cent reduction in track LCCs.
Keywords: fault diagnosis, condition monitoring, qualitative trend analysis, railway switch
actuator, incipient
JRRT313 Proc. IMechE Vol. 224 Part F: J. Rail and Rapid Transit
Downloaded from pif.sagepub.com at PENNSYLVANIA STATE UNIV on March 6, 2016
294 J A Silmon and C Roberts
the condition of the assets. This implies the use of reli- 4. The system shall require minimal pre-progra-
able automatic condition monitoring systems, which mming and training data.
can detect and diagnose most, if not all, possible faults. 5. The system shall be compatible with existing com-
This capability has not yet been achieved. munication systems.
INNOTRACK subproject 3.3 has been concerned 6. The system shall not generate false alarms.
with the development of innovative monitoring sys-
tems for switches and crossings (railway points). This 2.2.1 Fault detection
package of work has determined the key parame-
ters for switch system monitoring [1], proved exper- An incipient fault is one which develops gradually over
imentally that the parameters reflect simulated fault a period of time. Conversely, an abrupt fault is one
conditions [2], suggested an open standard for the which occurs all at once, with no prior warning. Cur-
specification of switch monitoring systems [3], and rently, condition monitoring systems on the railways
demonstrated the use of advanced algorithms to have used fairly simple methods such as thresholds
detect incipient faults in measured data [4]. to detect when values fall out of tolerance. Although
this may happen due to an incipient fault, the fault
indication when a threshold is breached is an abrupt
1.3 Structure of this paper one. These implemented systems have limited diag-
nosis capabilities; usually, an alarm is raised due to
A discussion of monitoring requirements is in section a breached threshold and human staffs examine the
2. Section 3 outlines the process followed to deter- data for clues as to the fault’s nature before heading
mine the key parameters for condition monitoring of out to perform maintenance.
switches. Using the key parameters, a novel approach
was developed after an extensive review of current 2.2.2 Fault diagnosis
methods (summarized in section 4). The novel method
is described in section 5. Section 6 describes a case The diagnosis of abrupt faults is also of benefit to main-
study carried out using the HW switch actuator, a tainers because it gives them advance warning of the
DC electric actuator used predominantly in the UK. procedures required and therefore the amount of time
Conclusions (section 7) are presented at the end. and the tools needed. By diagnosing the most likely
condition of the monitored asset at all times, the sys-
tem will be able to provide maintainers with an idea
2 REQUIREMENTS FOR MONITORING SYSTEMS of what condition the asset is suffering, and this will
make the process of planning corrective maintenance
2.1 Purpose more efficient.
The purpose of a condition monitoring system is
to provide accurate and timely warning to mainte- 2.2.3 Time-to-failure estimation
nance staff of any deterioration in the condition of An ideal condition monitoring system would pick up
the monitored asset. This allows corrective action to the first small signs of an incipient fault, diagnose its
be carried out before a failure occurs, which means nature, and determine how long maintenance staff can
that the tasks can be planned to cause minimum afford to wait before taking action. This allows main-
disruption to railway traffic. This approach is called tainers to prioritize tasks correctly and to arrange the
‘condition-based maintenance’ and is more efficient most convenient times to carry out maintenance: for
in terms of disruption and staff hours than the present example, a low priority fault could be dealt with dur-
periodic maintenance regimes, where certain mainte- ing the dead of night, rather than during the day when
nance tasks are carried out at fixed intervals calculated trains must be stopped for the work to be done.
from risk mitigation exercises.
2.2.4 Training data and models
2.2 Requirements Advanced fault detection and diagnosis methods usu-
An ideal set of requirements for a condition monitor- ally require some form of configuration in order to
ing system (neglecting practical considerations such produce useful outputs. This may be a set of param-
as robustness) might read as follows. eters for a mathematical model, which would be
constructed through detailed analysis, or it may be a
1. The system shall detect incipient faults before they set of training data measured from experimental fault
become severe enough to cause failures. simulations.
2. The system shall diagnose the most likely condition It is desirable to minimize the amount of prior
of the monitored asset at all times. knowledge required in order to make the system work
3. The system shall estimate the time remaining satisfactorily, because the gathering of such knowl-
before the asset fails. edge is time-consuming and costly. It is not practical to
Proc. IMechE Vol. 224 Part F: J. Rail and Rapid Transit JRRT313
Downloaded from pif.sagepub.com at PENNSYLVANIA STATE UNIV on March 6, 2016
Improving railway switch system reliability 295
train a monitoring system by simulating faults on every and would negate any benefits of installing automatic
asset it must monitor. This is because each asset has monitoring.
slightly different characteristics, which cannot always
be assumed to be of negligible variance.
3 KEY PARAMETERS FOR CONDITION
2.2.5 Compatibility MONITORING OF SWITCH SYSTEMS
It is desirable that a monitoring system adds little
Failure statistics were examined for a number of differ-
to the overall complexity of railway infrastructure.
ent types of switch actuators. It was determined that
Therefore, the system should be capable of function-
the most common failure modes are generally linked
ing within existing or planned telecommunications
to adjustment faults in the drive of the actuator. In
networks, rather than requiring its own bespoke com-
order to determine the key parameters for monitoring,
munications lines. However, the communications net-
experiments were carried out on actuators at train-
works on most European railways are not unified
ing schools, where faults could be simulated without
or standardized, and one of the advances suggested
risking disruption to train services.
in INNOTRACK was the replacement of the many
During the fault simulations, the displacement of
cables required for railway control with a standardized
the switch, the force in the drive, and the current in
interface, as shown in Fig. 1.
the motor were all measured. These parameters were
Under such a scheme, the monitoring system would
affected to different extents by the introduction of the
share the communications network with command
fault, as can be seen in Fig. 2, which comes from an
and control signals for all types of signalling apparatus.
HW switch actuator. More details on this actuator can
However, monitoring is not safety-critical as signalling
be found in section 6. The displacement was almost
is, so monitoring data would be assigned a lower pri-
unaffected, but clear differences can be seen in the
ority on this network, so that signalling commands are
current waveform, and most prominently of all in the
not delayed. When it is considered that values x, y, and
force waveform.
z are potentially more than ten each, it becomes clear
It can be concluded that force, current, and perhaps
very quickly that the use of a networked command and
displacement are the key parameters to be measured
control system with a common power bus can reduce
if the major adjustment faults in switches are to be
the amount of lineside cabling dramatically. With an
detected. It then becomes possible to make tradeoffs
open standard for the interface, compatibility of mon-
between the delays saved by the detection of more
itoring systems, control systems, and other equipment
unusual faults, and the extra sensors required to detect
can be ensured.
them.
False alarms cause maintenance crews to waste time 4 A NOVEL APPROACH TO THE DETECTION OF
and money maintaining equipment, which does not INCIPIENT FAULTS IN SWITCH SYSTEMS
require attention. This is a bad thing economically, but
it also has a detrimental effect on the faith that main- 4.1 Current methods in railway fault diagnosis
tenance crews will put in the system. If it is not seen
by the technicians as reliable, they may feel that they There have been several pilot installations of condi-
should ignore it or find a work-around to avoid rely- tion monitoring equipment on railway infrastructure
ing on its results. This could have safety implications throughout Europe. In most cases, these consist of
data collection equipment by the monitored asset,
with monitoring taking place either in the nearest
interlocking room, with a computer workstation, or
manually (by trained operators) at a central location.
Thresholds are the most common format of rule for
fault detection: if measured values exceed a preset
threshold, an alarm is generated and maintenance
tasks are performed.
Thresholds are adequate for the detection of some
faults, especially abrupt ones (which are not easily pre-
dictable), but for faults in adjustment, it is possible for
assets to perform within thresholds and yet deteriorate
over a period of time, before failing. Incipient faults,
Fig. 1 Configuration of traditional and proposed sig- which develop gradually, need more sensitive analysis
nalling wiring [3] to be detected when they are in their early stages [5].
JRRT313 Proc. IMechE Vol. 224 Part F: J. Rail and Rapid Transit
Downloaded from pif.sagepub.com at PENNSYLVANIA STATE UNIV on March 6, 2016
296 J A Silmon and C Roberts
Fig. 2 Progressive introduction of an overdriving fault towards the normal side in a HW electric
switch actuator
There exist, therefore, some opportunities for be matched to prior knowledge about the effects of
the improvement of railway condition monitoring certain types of fault.
through advanced methods: If an accurate mathematical model of a monitored
asset is available, quantitative modelling can provide
(a) increased sensitivity to incipient faults;
extremely accurate and reliable fault diagnosis. In the
(b) intuitive methods that technicians can under-
case of railway actuators, however, there are problems
stand;
that hinder modelling, such as the relatively high num-
(c) automatic diagnosis and time-to-failure predic-
ber of adjustable parts, non-linear loads, and rapidly
tion.
changing operational conditions.
Qualitative model-based diagnosis uses a variety
4.2 Advanced fault diagnosis methods of techniques, including state machines [10], qual-
Industries other than railways have been working with itative physics [11], and digraphs [12], to predict
sophisticated supervisory control and data acquisi- the qualitative behaviour of a system from qualita-
tion equipment for many years. Numerous methods tive observations or quantized measurements. These
have been used to predict and diagnose faults in fields models are useful in applications in which it is not
such as electricity transmission and distribution [6], possible to obtain quantitative measurements or in
gas turbines [7], and chemical plants [8]. which there is large uncertainty in the behaviour of
There are three main types of fault diagnosis the monitored system [13]. However, the accommo-
method, into which most individual approaches can dation of such uncertainty reduces the sensitivity of
be gathered [9]: a qualitative model-based diagnosis system to faults,
which manifest themselves with small symptoms or
(a) quantitative model-based; gradually over time.
(b) qualitative model-based; All model-based methods require some analysis in
(c) process history-based. order to set a model correctly before monitoring can
Quantitative model-based methods work by using begin. Railway actuators, especially switch actuators,
mathematical relations to predict or estimate the per- may have very diverse local conditions, even though
formance of a monitored system, from measurable the driving actuator is of a common type. This means
inputs. By comparing the predicted performance with that the variations in performance can be quite large,
the actual (measured) performance, a set of residuals even between actuators of the same type. For a model-
can be calculated. The nature of these residuals can based method to succeed, the model must either
Proc. IMechE Vol. 224 Part F: J. Rail and Rapid Transit JRRT313
Downloaded from pif.sagepub.com at PENNSYLVANIA STATE UNIV on March 6, 2016
Improving railway switch system reliability 297
reject these variations or predict them accurately. Process history-based methods use data collected
For a railway asset base which may number in the from the monitored asset to establish rules for
thousands, the amount of analysis involved would be diagnosing faulty behaviour. The collected data can
impractical. be analysed with statistical methods such as wavelets
JRRT313 Proc. IMechE Vol. 224 Part F: J. Rail and Rapid Transit
Downloaded from pif.sagepub.com at PENNSYLVANIA STATE UNIV on March 6, 2016
298 J A Silmon and C Roberts
[14] and principal component analysis [15] or be used 5.2 QTA procedure
to train neural networks [16]. These methods can
be extremely useful because they only require data, Because each measured waveform comes from an
not analysis, and as data collection equipment forms actuator that is of a common design, it is fair to assume
part of the hardware needed for railway condition that there will be some common characteristics of the
monitoring, this is not a difficult thing to provide. waveform. The purpose of using QTA is to isolate these
However, these complex methods are often not common characteristics and to form rules, which
transparent in the way they work, particularly in the describe their behaviour under fault conditions.
case of neural networks. One key requirement for Figure 3 shows how a waveform of noisy data
a railway condition monitoring system is that the is converted into a sequence of episodes, which
method is simple to understand and mimics human describe the qualitative and quantitative characteris-
analysis, because it then becomes easier for technical tics of the waveform. When this method was imple-
staff to understand the reasons that the system has mented, the waveforms were first filtered to remove
come to its conclusion. This aids the use of the system high-frequency components. The filtered waveforms
and mitigates any false alarms. One process history- were quite different from the originals, but the effects
based method that is quite intuitive is QTA [17]. This of each fault were equally visible, allowing the system
process mimics human shape analysis by tracing the to identify more common episodes.
important trends in a measured waveform and repre- Figure 4 shows the nine shapes, which can be
sents them as a profile of episodes, each representing assigned to a partition in the waveform.
a trend with a particular shape.
This approach was adopted for the algorithm devel- 5.2.1 Rule formation
oped in this project, because it uses a combination When faults were introduced (in the TA data set),
of qualitative and quantitative aspects to describe the the episode sequence (in terms of shapes) remained
performance of actuators. This is useful because the largely the same, but the quantities associated with
qualitative aspects can overcome the differences in each episode changed. The differences between faulty
performance between actuator instances, whereas the and fault-free values were used to construct fuzzy sets,
quantitative aspects ensure sensitive performance for which act as rules, testing the presence of each fault.
incipient faults. An example is shown in Fig. 5.
Proc. IMechE Vol. 224 Part F: J. Rail and Rapid Transit JRRT313
Downloaded from pif.sagepub.com at PENNSYLVANIA STATE UNIV on March 6, 2016
Improving railway switch system reliability 299
Fig. 5 Fuzzy membership functions used to create rules identifying the symptoms of incipient
faults
6.1 Background
The HW is a switch actuator widely used in the UK. It is
driven by an 110 V DC permanent magnet motor con-
nected through a reduction gearbox, and a mechanical
or magnetic clutch, to the switch drive. Locking and
detection are integral to the actuator, although it is
common for supplementary detection to be provided
on long switches. A sketch of the actuator’s layout is
shown in Fig. 7.
The most common cause of failure in switches
driven by HW actuators is incorrect adjustment of the
many mechanical parts in the switch itself. This can
occur during maintenance, installation, or by a nat-
ural process as nuts move along screw threads. The
Fig. 6 Ideal output of the fault detection and diagnosis result is that the forces in the drive change, eventually
system resulting in a throw failure.
JRRT313 Proc. IMechE Vol. 224 Part F: J. Rail and Rapid Transit
Downloaded from pif.sagepub.com at PENNSYLVANIA STATE UNIV on March 6, 2016
300 J A Silmon and C Roberts
Proc. IMechE Vol. 224 Part F: J. Rail and Rapid Transit JRRT313
Downloaded from pif.sagepub.com at PENNSYLVANIA STATE UNIV on March 6, 2016
Improving railway switch system reliability 301
Fig. 9 Introduction of overdriving towards the reverse side in the HW switch actuator
both end positions to be varied. Figure 9 shows the Table 1 Test results for normal-to-reverse movements
data acquired when the drive was adjusted to drive
OK %
too far towards the reverse side. Spec. Condition combinations effectiveness
JRRT313 Proc. IMechE Vol. 224 Part F: J. Rail and Rapid Transit
Downloaded from pif.sagepub.com at PENNSYLVANIA STATE UNIV on March 6, 2016
302 J A Silmon and C Roberts
REFERENCES
BIBLIOGRAPHY
1 Silmon, J. and Roberts, C. List of key parameters for
switch and crossing monitoring. INNOTRACK deliver- Lehrasab, N. and Fararooy, S. Formal definition of sin-
able 3.3.1. 2007. gle throw mechanical equipment for fault diagnosis. IEE
2 Roberts, C. and Silmon, J. Available sensors for railway Elect. Lett., 1998, 34(23), 2231–2232.
environments for condition monitoring. INNOTRACK Silmon, J. A. and Roberts, C. A systems-engineered intuitive
deliverable 3.3.2. 2008. adaptive failure prediction system. In Proceedings of the
3 Ziemann, A. and Silmon, J. Requirements and func- 2nd IET International Conference on Railway condition
tional description for S&C monitoring. INNOTRACK monitoring, Derby, UK, 18–20 June 2008.
deliverable 3.3.3. 2008. Silmon-Monerri, J. A. and Roberts, C. A systems approach
4 Silmon, J. Algorithms for detection and diagnosis of to fault detection and diagnosis for condition-based
faults on S&C. INNOTRACK deliverable 3.3.4. 2009. maintenance. In Proceedings of the 1st IET Interna-
5 Roberts, C., Dassanayake, H. P. B., Lehrasab, N., and tional Conference on Railway condition monitoring,
Goodman, C. J. Distributed quantitative and qualitative Birmingham, UK, 29–30 November 2006.
Proc. IMechE Vol. 224 Part F: J. Rail and Rapid Transit JRRT313
Downloaded from pif.sagepub.com at PENNSYLVANIA STATE UNIV on March 6, 2016