Professional Documents
Culture Documents
The manuscript was received on 23 December 2010 and was accepted after revision for publication on 8 July 2011.
DOI: 10.1177/1748006X11418422
Abstract: Travelling safely and comfortably on high-speed railway lines requires excellent con-
ditions of the whole railway infrastructure in general and of the railway track geometry in par-
ticular. The maintenance process required to achieve such excellent conditions is complex
and expensive, demanding a large amount of both human and technical resources. In this
framework, choosing the right maintenance strategy becomes a critical issue. A reliable simu-
lation of the railway geometry ageing process would offer a great advantage for the optimiza-
tion of planning and scheduling of maintenance activities. A fundamental requirement for
such simulation is a statistical model describing the behaviour of the railway track geometry
deterioration as well as the effects of maintenance activities. The French railway operator
SNCF has been periodically measuring the geometrical characteristics of its high-speed net-
work since its commissioning (i.e. for more than 20 years now). These records are an excellent
data source to achieve a sound statistical description of the process.
In this paper a new system identification method to obtain such simulations is presented.
The proposed method uses a grey-box model: a model structure and its constraints are speci-
fied basing on previous knowledge of the process to be identified, and then the set of para-
meter values which best fits the signal measurements is searched. As previous knowledge
indicates that the process is non-linear, parameter values are searched by means of the
Levenberg–Marquardt algorithm, an iterative technique that finds a local minimum of a func-
tion that is expressed as the sum of squares of non-linear functions. Furthermore, the pre-
sented model is extended in order to analyse the effect of the variation of factors influencing
the ageing process (e.g. operational speed). Finally, the method is applied and validated with
real data of a French high-speed TGV line.
the measurement units are not shown. Degradation The available models can be classified into two
decrements take place only when some mainte- types: (i) deterministic and (ii) stochastic. Next
nance activity is performed. In Fig. 1 the mainte- some relevant models of both types are reviewed.
nance activities most relevant for track geometry are
included: tamping interventions. The bar heights 2.1 Deterministic models
represent the fraction of the railway sector affected
by the maintenance activity. Tamping yields a Deterministic models use mainly polynomial [5],
visually obvious effect, generating a sudden drop in exponential [6, 7] or multi-stage linear [8] models.
NL. Technical and human resources required By means of system identification techniques (typi-
for performing tamping interventions are a major cally least squares) the parameter set that best fits
cost factor in high-speed railway systems [2]. the measurement series is found. Figure 2 shows
Furthermore, due to high logistical cost constraints, schematically an example of track quality course
most track geometry maintenance activities need to using the exponential model.
be planned up to one year in advance. In reference [10] different approaches are used
These facts make it clear that the maintenance for forecasting the railway geometry course on a
process of railway track geometry is highly complex, French TGV line. The best results are obtained by
so predicting the effect of introducing changes in the model which assumes an exponential growth of
the process is a very difficult task. The authors the NL indicator. Finally, in reference [11], the
dependency between the value of NL after a tamp-
believe that the possibility of simulating the degra-
dation–restoration process would be of great advan- ing intervention and the number of tampings per-
formed since the last ballast renewal is investigated
tage in this sense.
and confirmed.
This paper is organized as follows: in section 2 an
overview of the track geometry ageing and degrada-
tion models found in the literature is given. In sec- 2.2 Stochastic models
tion 3 a simulation method for the railway track
geometry degradation and restoration process is pre- The irregularities observed over the years have led
sented. It comprises an ageing and restoration model to modelling track geometry ageing as a stochastic
which statistically describes the process based on process. Reference [12], for instance, has modelled
historical measurements, and then uses the Monte an indicator of the track geometry quality by means
Carlo method to obtain simulated process realiza- of a Markov chain [13]. Without maintenance inter-
tions. Then, in section 4, the model is applied in a ventions, the indicator can either grow or sink, but
case study based on real data from a French high- the net ageing speed is always positive. The values
speed line, where the process characteristics of the transition rates between states are calculated
assumed in the model are confirmed. Finally, section basing on the measured data. Three state groups are
5 presents some concluding remarks. defined: for Z1, ., Za with a . 1 the track quality is
The model here proposed will be part of a holistic good, for Za + 1, ., Zb with b . a it is in alert state,
maintenance strategy optimization method for rail- and for Zx with x . b a maintenance intervention is
way infrastructures currently under development at required.
the TU Braunschweig [3]. In this framework, the
model is used to perform simulations of the com- 1
Meas. Values
plete life cycle of the track adopting different main- Tampings
0.9
tenance strategies.
0.8
Long. levelling
0.7
2 STATE OF THE ART
0.6
ity. This reveals that a universal characterization of Fig. 1 Course of longitudinal levelling degradation for
the track quality course has not been feasible so far. a railway sector
The model hereby presented relies on measure- NLinitn ; LN (mNLinit (n), s2NLinit (n)) (1)
ment data.
where m is the mean value and s2 is the variance.
According to these assumptions, for the model to next step is to find the functions mNLinit(n),
be applied it is necessary to find expressions for sNLinit(n), mb(n), and sb(n) which best fit the esti-
mNLinit(n), sNLinit(n), mb(n), and sb(n). To obtain mated values.
these functions, it is necessary to have a database
with track geometry measurements on many railway 3.4 Monte Carlo simulation process
sectors for many years, including tamping activities
The proposed process model can be used to per-
performed. In section 4 this identification is made
form Monte Carlo simulations following the flow-
with the Levenberg–Marquardt process.
chart of Fig. 3. Briefly described, the simulation
process consists of dicing NLinitn, bn and the mea-
3.3 Model parameterization
surement noise e(t) and then calculating the process
Let NLi,n = {NLi(tinit), ., NLi(tend)} be the series value as NLinitn ebn (ttls ) + e(t) at each simulation
of available measurements of NL at sector i between step, until the tamping condition is reached. At this
th nth and the (n + 1)th tamping interventions. point, the number of accumulated tampings is
Then the minimization problem consists of incremented and the process is repeated until the
finding NLiniti,n and bi,n such that the function simulation end condition is reached.
c = NLiniti, n ebi, n t best fits the measured values;
NL(t) According to this the simulation only needs the
that is, minimizes the total square error represented following input parameters:
by
desired simulation start and end times;
X
tend
simulation step; and
c
(NLt 2 NL(t)) 2
(5) either
t = tinit
– the set of times at which tamping activities
should take place, or
This is done at each 1-km-long sector and for each
– a degradation threshold at which, when
period between tampings, obtaining for each sector
achieved, a maintenance activity should take
i and each period between tampings n the para-
place.
meters NLiniti,n and bi,n. Additionally, the mean
quadratic value of the error signal, (i.e. the differ-
These parameters are set by the user according to
ence between the measured values and the fitted
the objectives of the simulation. In section 4 the
curve) is recorded and then used to simulate the
method is illustrated by means of an example.
measurement noise en(t).
It is well known from the literature [1, 10, 17] and
3.5 Model adaptation to variable influencing
from practical experience that track geometry exhi-
factors
bits a transient behaviour in the first months after a
tamping intervention, so measurements taken in the The track railway geometry ageing process depends
first three months after an intervention are dis- on a number of factors. The process model used in
carded. Additionally, to rest assured that the real this work is based on the assumption that these fac-
behaviour of degradation is being fitted and not just tors remain constant in the course of the simulation
measurement noise, it is necessary to only fit data time. In the literature numerous papers on this
sets where the time span between the first and last topic can be found. They are reviewed in reference
sample is at least one complete year. [4], according to which the effect of the change of a
c
Since NL(t) is non-linear, for solving these mini- parameter P on the geometry deviation can be
mization problems [18], an open-source implemen- described with a function of the form (7). Equation
tation of the Levenberg–Marquardt algorithm (LMA) (7) means that knowing the degradation law for a
[19], is used. The LMA is an iterative algorithm that given value of P = p0 and the scaling factor aP, it is
finds a local minimum of a function that is possible to calculate the law that describes the age-
expressed as the sum of squares of non-linear func- ing process for P = p1. The value of aP has been esti-
tions. It can be regarded as a combination of stee- mated for the most relevant influencing factors by
pest descent and the Gauss–Newton method. When different railway companies [4]
the current solution is far from the local minimum,
DNL(t) = NL(t)2NL(tn ) (6)
the algorithm behaves like a steepest descent
method (i.e. slow) but guaranteed to converge. aP
When the current solution is close to the local mini- DNL(tjP = p1 ) p1
= (7)
mum, it becomes a Gauss–Newton method. The DNL(tjP = p0 ) p0
aP
p1
Start = (NL(tjP = p0 )2NL(t = tn jP = p0 )) (12)
p0
According to this assumption, equation (13) can 4.1 Analysis of the statistical independence of
be used for the conversion of a simulation made NLinit and b
with P = P0 into one with P = P1. This means that, for
As stated in section 3, one assumption of the pro-
instance, with the data of a track where trains have
cess model is the statistical independence of NLinit
been running at 250 km/h, a simulation with
and b. Figure 4 shows all (NLinit, b) pairs in the pre-
300 km/h can be made, as shown in section 4.4
viously mentioned database (i.e. the NL achieved
DNL(tjP = p1 ) : = NL(tjP = p1 )2NL(t = tn jP = p1 ) after a tamping intervention and the parameter b of
(9)
the ensuing curve NL(t) = NLinit ebt). No correlation
3:8 (10) can be recognized visually. Furthermore, the corre-
= NL(tjP = p1 )2NL(t = tn jP = p0 ) lation coefficient rNLinit,b using equation (15) yields
aP 21.1031022. In view of these facts, the assumption
3:7 p1 (11) of statistical independence of NLinit and b seems to
= DNL(tjP = p0 )
p0 be reasonable
n N da dmax da/dmax
0.6
1 112 0.0718 0.128 55.8
0.4 2 190 0.0544 0.986 55.2
3 177 0.0559 0.102 54.7
0.2 4 137 0.113 0.116 97.4
5 96 0.0863 0.138 62.2
0 6 76 0.0845 0.156 54.2
-0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 7 51 0.0893 0.190 46.8
b 8 31 0.0804 0.244 32.9
9 20 0.158 0.304 51.9
Fig. 4 NLinit versus the parameter b of the ensuing 10 13 0.234 0.377 62.1
curve NL(t) = NLinit ebt
0.1
0 2 4 6 8 10 12 14
on the measurements database are presented. Accumulated tamping interventions
Figure 5 shows the estimated mean values and var-
Fig. 6 Estimated and fitted mean values of bn; error
iance of NLinitn, and the fitted mean values (dotted bars represent the variance
curve). Additionally, the amplitudes of the bars rep-
resent the estimated variances. In the same fashion,
Fig. 6 shows estimated and fitted mean values and 0.08
σNLinit
estimated variances for the parameters bn. Figures 7 σNLinit (adjusted)
0.07
and 8 show the variances (they are also shown in
Figs 5 and 6 as bar amplitudes) and the curves fit- 0.06
ting them. For the function fitting procedures, the
0.05
estimations of NLinitn and bn are discarded for n .
10 because there are only very few sectors in the 0.04
measurements database with more than ten tamp- 0.03
ing interventions, so they are not statistically
significant. 0.02
0.25
0 2 4 6 8 10 12 14 a2be2un (19)
Accumulated tamping interventions
Fig. 5 Estimated and fitted mean values of NLinitn; and for fitting b(n) and the variances a function of
error bars represent the variance the form
0.045 0.8
σb NL
0.04 σb (adjusted) Tampings
0.7
0.035
0.03 0.6
Long. levelling
0.025
0.5
0.02
0.4
0.015
0.01 0.3
0.005
0.2
0 1985 1990 1995 2000 2005 2010
0 2 4 6 8 10 12 14 Time [Years]
Accumulated tamping interventions
Fig. 9 Deterministic simulation with intervention
Fig. 8 Estimated and fitted variance values of bn threshold set to 0.7
a + (bn)u and after that the curve with v = 300 km/h grows
(20)
always faster. This means that to keep NL under a
Thus applying once again the previously mentioned given threshold, with v = 300 km/h interventions will
error minimization technique values are found for be necessary more often than at v = 250 km/h.
the parameters a, b, and u, and achieve the
dotted curves shown in Figs 5 to 8, as well as an
estimation of the measurement noise. Having calcu- 5 CONCLUSIONS
lated these four functions, it is possible to perform
simulations. This paper has presented a method for modelling
and simulation of the track geometry ageing and
restoration process. It models both the exponential
4.4 Process realizations
growth constant and the value of NL after tamping
In this section some process realizations basing on interventions as log-normally distributed variables
the data identification of section 4.3 are presented which are dependant on the number n of accumu-
and analysed. In order to get a glimpse of what a typi- lated tamping interventions. These dependencies
cal process would look like, Fig. 9 shows the achieved become evident in the case study in section 4,
simulation when setting all variances and the noise to where it is verified that the exponential growth con-
0; that is, shutting down the whole stochastic charac- stant and the value of NL achieved after a tamping
teristics of the process, and setting the degradation intervention tend to rise with the number of accu-
limit to 0.9. This process realization can be considered mulated tamping interventions. This leads ulti-
as a kind of mean of all process realizations available mately to the end of the track life cycle and the
in the database used to tune the simulator. need for track renewal. Since the model is suitable
A parameter with significant influence on the
ageing speed of the track geometry is the maximal NL(t|ν=250)
1.2 NL(t|ν=300)
travelling speed v. In reference [4] the value of av Tampings
on the high-speed line Madrid–Seville is 1.27. 1
If v1 = 300 km/h and v0 = 250 km/h then, using
Long. levelling
0.6
NL(tjv = 300) = (NL(tjv = 250)2NL
300 1:27 (21) 0.4
(t = tn jv = 250)) + NL(t = tn jv = 250)
250
0.2
for the simulation of the whole life cycle of the high-speed railways: influence of stochastic para-
track, the authors believe that it can be regarded as meters. In Proceedings of the 8th International
a very useful tool to evaluate potential benefits of Conference on Railway engineering, University of
modifications on the track maintenance strategy. Westminster, London, 2005, CD-ROM.
5 Jovanovic, S. Railway track quality assessment and
The approach hereby presented is intended to be
related decision making. Syst. Man Cybernet., 2004,
used by the tamping scheduling optimization system 6, 5038–5043.
under development at the Institute for Traffic Safety 6 Veit, P. Track Quality – Luxury or Necessity? Rail-
and Automation Technologies of the TU Braunschweig way Technical Review Special: Maintenance and
in cooperation with the SNCF. In this context, process Renewal, July 2007, pp. 8–12.
simulation would be useful to assess the sensibility of 7 UIC IDTEG. Best practice guide for optimum track
the maintenance process to the stochastic characteris- geometry durability, 2008 (UIC Railway Technical
Publications).
tics of the degradation and restoration process.
8 Chang, H., Liu, R., and Li, Q. A multi-stage linear
The simulation method is intended to be applied prediction model for the irregularity of the longitu-
on complete high-speed lines. Reference [14] shows dinal level over unit railway sections. In Computers
that the degradation behaviour along a complete in railways XII (Eds B. Ning and C. Brebia), 2010,
line is not homogeneous. Hence, in order to simu- pp. 641–650 (WIT Press, Southampton).
late complete lines, it would be necessary to divide 9 Lichtberger, B. Das System Gleis und seine Instand-
each of them into clusters of homogeneous sectors. haltung. Eisenbahningenieur, 2007, 58(1), 10–19.
10 Quiroga, L. and Schnieder, E. Modelling high speed
This would lead to an increased simulation verisimi-
railroad geometry ageing as a discrete-continuous
litude, and enhance the feasibility of an integral process. In Proceedings of the International Confer-
optimization of the maintenance process on the ence on Stochastic modeling techniques and data
complete line. This has special relevance consider- analysis (SMTDA), 2010. CD-ROM.
ing that in most real cases the same human and 11 Antoni, M. Degradation laws of railway ballasted
technical resources are deployed for the mainte- track. Estimation of the need of maintenance. In Pro-
nance of a whole line. ceedings of the 3rd International Conference on Accel-
erated life testing, Clermont-Ferrand, France, 2010.
12 Lyngby, N., Hokstad, P., and Vatn, J. RAMS man-
agement of railway tracks. In Handbook of per-
ACKNOWLEDGMENTS formability engineering, 2008, pp. 1123–1145
The authors would like to express their gratitude to (Springer, Berlin).
Marc Antoni, Chief of the Cellule Emergence et 13 Markov, A. Extension of the limit theorems of
Prospective of the SNCF Direction de l’Infrastructure, probability theory to a sum of variables connected
for taking the initiative and providing all necessary in a chain. In Dynamic probabilistic systems, I:
Markov chains (Ed. R. Howard), 1971, pp. 552–576
information and expertise, and to the anonymous
(Wiley, New York).
referees for helping to improve the paper with their 14 Meier-Hirmer, C., Senée, A., Riboulet, G.,
sound reviews. Sourget, F., and Roussignol, M., et al. A decision
support system for track maintenance. In Comput-
FUNDING ers in railways X, 2006, pp. 217–226 (WIT Press,
This work was supported by the German Academic Southampton).
15 Sato, K. Lévy processes and infinitely divisible distribu-
Exchange Service, grant number A/08/74683.
tions, 1999 (Cambridge University Press, Cambridge).
Ó IMechE 2011 16 Breiman, L., Friedman, J. H., Olshen, R. A., and
Stone, C. J. Classification and regression trees, 1984
(Chapman and Hall, London).
REFERENCES 17 Lichtberger, B. Handbuch gleis, 3rd edn, 2010 (Eur-
ail Press, Hamburg).
1 Meier-Hirmer, C. Modèles et techniques probabil- 18 Lourakis, M. I. A. LEVMAR: Levenberg–Marquardt
istes pour l’optimisation des stratégies de mainte- nonlinear least squares algorithms in C/C ++
nance. Application au domaine ferroviaire. PhD [Open Source library on the Internet], 2004,
Thesis, Université de Marne-la-Vallée, 2007. available from www.ics.forth.gr/~lourakis/levmar/
2 Esveld, C. Modern railway track, 2001 (MRT Pro- [accessed 31 Jan 2010].
ductions, Zaltbommel). 19 Marquardt, D. An algorithm for least-squares esti-
3 Quiroga, L. and Schnieder, E. A heuristic approach mation of nonlinear parameters. SIAM J. Appl.
to railway track maintenance scheduling. In Com- Math., 1963, 11, 431–441.
puters in railways XII (Eds B. Ning and C. Brebia), 20 Stuart, A., Ord, K., and Arnold, S. Kendall’s
2010, pp. 687–699 (WIT Press, Southampton). advanced theory of statistics, 2A: Classical inference
4 Ubalde, L., López Pita, A., Teixeira, P., and the linear model, 6th edition, 1998 (Hodder
Bachiller, A., and Gallego, I. Track deterioration in Arnold, London).