See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/277306564

Global Sensitivity Analysis applied to Model Inversion Problems: A Contribution

to Rail Condition Monitoring

Article in International Journal of Prognostics and Health Management · May 2015

DOI: 10.36001/ijphm.2015.v6i4.2322

CITATIONS READS

11 455

2 authors:

René Schenkendorf Jörn Groos

Hochschule Harz German Aerospace Center (DLR)
82 PUBLICATIONS 790 CITATIONS 62 PUBLICATIONS 446 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by René Schenkendorf on 28 May 2015.

The user has requested enhancement of the downloaded file.

 Global Sensitivity Analysis applied to Model Inversion Problems:
A Contribution to Rail Condition Monitoring
René Schenkendorf1 and Jörn C. Groos1

1
Institute of Transportation Systems, German Aerospace Center (DLR e.V.), Lilienthalplatz 7, 38108 Braunschweig, Germany
{rene.schenkendorf, joern.groos}@dlr.de

A BSTRACT gyroscopes) which can be installed on usual in-line freight

and passenger trains (e.g. (Ward et al., 2011; Molodova, Li,
Rising demands on railroad infrastructure operator by means
& Dollevoet, 2011; Naganuma, Kobayashi, & Tsunashima,
of profitability and punctuality call for advanced concepts of
2013b)). In this article, mathematical models are utilized for
Prognostics and Health Management. Condition based preven-
the purpose of railway track condition monitoring by analyz-
tive maintenance aims at strengthening the rail mode of trans-
ing the dynamic railway track - vehicle interactions. More
port through an optimized scheduling of maintenance actions
precisely, so-called inverse model problems are of special in-
based on the actual and prognosticated infrastructure condi-
terest. An inverse model aims to recalculate previous inputs
tion, respectively. When applying model-based algorithms
(irregularities) of the monitored system (rail surface) which
within the framework of Prognostics and Health Management
had caused recorded system responses (dynamic vehicle reac-
unknown model parameters have to be identified first. Which
tions). These recalculated inputs are the basics of subsequent
of these parameters should be known as precisely as possi-
condition monitoring analyses. As in any model-based context
ble can be figured out systematically by a sensitivity analysis.
uncertainties of the applied model parameters may influence
A comprehensive global sensitivity analysis, however, might
the derived inputs significantly. Thus, a parameter sensitivity
be prohibitive by means of computation load when standard
analysis might help to quantify the impact of parameter im-
algorithms are implemented. In this study, it is shown how
perfections. In particular, it is shown how global parameter
global parameter sensitivities can be calculated efficiently by
sensitivities can be calculated efficiently by combining Polyno-
combining Polynomial Chaos Expansion and Point Estimate
mial Chaos Expansion and Point Estimate Method principles.
Method principles. The proposed framework is demonstrated
In so doing, a subset of model parameters can be identified
by a model inversion problem which aims to recalculate the
which should be known as precisely as possibly to provide
track quality by measurements of the vehicle acceleration, i.e.
credible model-based results.
analyzing the dynamic railway track-vehicle interaction.
The remainder of this paper is organized as follows. In Sec-
1. I NTRODUCTION tion 2 an overview of rail condition monitoring concepts is
given. In Section 3 a mechanistic model of the vehicle dy-
Railway networks are historically grown complex systems
namic is derived. The basics of model inversion strategies are
consisting of various types of infrastructures (sub)components.
summarized in Section 4. Mathematical tools for parameter
Nowadays, about 50 percent of the life cycle costs of railway
sensitivity analysis are introduced in Section 5. Sections 6 and
infrastructures are made up by maintenance costs (Gradinariu
7 show how to apply the Polynomial Chaos Expansion and the
et al., 2008). Here, the condition based preventive maintenance
Point Estimate Method for an efficient calculation of global
aims at cutting costs and increasing availability by an opti-
parameter sensitivities. A numerical example given in Section
mized scheduling of maintenance actions taking into account
8 illustrates the proposed framework descriptively. Finally, the
the actual infrastructure condition and its expected further
conclusion is given in Section 9
degradation. Prerequisite therefore is the almost continuous
condition monitoring for thousands of kilometers of railway
2. R AIL C ONDITION M ONITORING
tracks. One approach to account for this challenge is the
development of low-cost railway track condition monitoring Of high importance for railway operators are rail infrastructure
systems based on MEMS inertial sensors (accelerometers and elements which reduce Reliability, Availability, Maintainabil-
René Schenkendorf et al. This is an open-access article distributed under
ity and Safety, so-called RAMS-killers, as well as drive up
the terms of the Creative Commons Attribution 3.0 United States License, operating costs. A working group of the International Union
which permits unrestricted use, distribution, and reproduction in any medium, of Railways (UIC) identified switches and crossings as most
provided the original author and source are credited.

International Journal of Prognostics and Health Management, ISSN2153-2648, 2015 019 1

 I NTERNATIONAL J OURNAL OF P ROGNOSTICS AND H EALTH M ANAGEMENT

important RAMS-killers and the track geometry as well as rail jima, Matsumoto, & Mizuma, 2010; Molodova et al., 2011))
head quality as the most important cost drivers (International as well as track geometry (e.g. (Kawasaki & Youcef-Toumi,
Union of Railways (UIC), 2010). The recent approach to in- 2002; Feldmann, Kreuzer, & Pinto, 2000; M. Kobayashi, Na-
crease RAMS and to reduce costs is to establish optimized ganuma, Nakagawa, & Okumura, 2008; Weston, Ling, Good-
condition-based preventive maintenance strategies in the field man, et al., 2007; Weston, Ling, Roberts, et al., 2007)). The
of Prognostics ans Health Management (PHM). The task of a different approaches differ in the applied data analysis method-
modern railway condition monitoring is therefore to provide ologies. The common approaches are based on model-free
a sufficient data base of reliable condition information. Once time domain, frequency domain or time-frequency analyses of
the current infrastructure condition can be reliably monitored the time domain acceleration data (e.g. (Caprioli, Cigada, &
and assessed, the future trend of infrastructure degradation Raveglia, 2007; Molodova et al., 2014; Feldmann et al., 2000;
is of utmost interest in condition based preventive mainte- Naganuma, Kobayashi, & Okumura, 2010)) as well as model-
nance. For instance, by applying suitable degradation mod- based inversion methods (e.g. (T. Kobayashi, Naganuma, &
els the remaining useful life (RUL) of technical devices (e.g. Tsunashima, 2013; Kawasaki & Youcef-Toumi, 2002; Weston,
switch engines, signals) or critical track defects (e.g. track Ling, Goodman, et al., 2007)).
settlement/misalignment, rail fatigue) of rail segments under
study can be prognosticated. Subsequently, ongoing mainte-
nance actions can be re-scheduled optimally taking account of
the actual infrastructure condition. Nowadays, rail condition
monitoring is typically done in fixed time intervals by vi-
sual inspection, with manually operated measurement devices
(hand-held, trolleys) or with measurement systems mounted
on dedicated measurement trains. The recent developments to
improve rail condition monitoring are to replace visual inspec-
tion by automatic procedures and to replace manually operated (a) Squat-type failure (b) Corrugation
measurement systems by autonomous systems mounted on
trains. These improvements reduce significantly the costs and Figure 1. Two snapshots of different rail surface failures.
track possession time for condition monitoring. At the same
time, condition monitoring systems to be installed on in-line
commercial trains are developed to further reduce costs and The analyses by inversion methods are typically applied to
to provide a nearly continuous monitoring of railway track accelerations measured at the bogie or the car body which are
condition. Main foci for in-line condition monitoring systems significantly influenced by the damping systems of the railway
are the two main costs drivers: defects of the rail head / rolling vehicles. Especially the utilization of car body accelerations is
contact fatigue (e.g. head-checks, squats, corrugation (Fig. of high interest, as these can be measured with very low effort
1)) and misalignments of track geometry (vertical and lateral even by small portable measurement systems (e.g. (Mori et al.,
track alignment, gauge). One approach is the migration of 2013)). Nevertheless, inversion analyses depend on a careful
(expensive) measurement systems from dedicated measure- selection of parameters influencing the model outcome. Thus,
ment trains to a small number of commercial in-line trains. a parameter sensitivity analysis should be mandatory in this
This approach typically implies a significant shortening of context and deserves a detailed explanation. This paper is
the maintenance intervals for the equipped in-line vehicles. focused on the application of model-based analyses of the dy-
Another approach is the development of comparably low-cost namic reactions of a rail vehicle to assess the condition of the
railway track condition monitoring systems based on MEMS rail surface (e.g. corrugation, squats). In particular, it is shown
sensors (accelerometers, gyroscopes and microphones) which how global parameter sensitivities can be calculated efficiently
can be installed with low maintenance effort on usual in-line by combining Polynomial Chaos Expansion and Point Esti-
freight and passenger trains (e.g. (Ward et al., 2011; Lee, mate Method principles. Some illustrative, preliminary results
Choi, Kim, Park, & Kim, 2012; Naganuma et al., 2013b; and the conclusion complete this manuscript.
Molodova, Li, Nunez, & Dollevoet, 2014; Mori, Sato, Ohno,
Tsunashima, & Saito, 2013; Bocciolone, Caprioli, Cigada, & 3. M ODELING OF THE V EHICLE DYNAMICS
Collina, 2007; Weston, Ling, Roberts, et al., 2007)). These The dynamic response of the vehicle dynamic might be de-
systems are based on the idea to utilize the dynamic train-track- scribed by mechanical vehicle suspension systems. These
interaction caused by track defects for condition monitoring systems may include the full vehicle-body motion up to single
(e.g. (R. B. Lewis & Richards, 1986; R. Lewis & Richards, axis movements (Fig. 2) known as quarter-vehicle models
1988)). The feasibility of this basic idea is proved by a number (Imine, 2011; Naganuma, Kobayashi, & Tsunashima, 2013a).
of numerical as well as experimental studies for defects of the For instance, the governing equations for a quarter-vehicle
rail (e.g. (Sunaga, Sano, & Ide, 1997; Mori, Tsunashima, Ko- system are given as:

2
 I NTERNATIONAL J OURNAL OF P ROGNOSTICS AND H EALTH M ANAGEMENT

ms ẍs = −ks (xs − xus ) − c(ẋs − ẋus )

0
mus ẍus = ks (xs − xus ) + c(ẋs − ẋus )+ (1)
kus (u(t) − xus )

Track Irregularity in mm
-0.5
Direction of travel
ms

-1
ms
xs
ms

-1.5
ks c mus

-2
mus 0 2 4 6 8 10 12
xus Time in s
mus
(a) Simulated track irregularity, u(t).

kus

0.10
u(t)

-0.10 -0.05 0.00 0.05

Acceleration in m/s2
Figure 2. Mechanical suspension system representing the
quarter-vehicle model. By this simplified model the dynamic
railway track - vehicle interaction can be analyzed, e.g., the
vehicle response when crossing track irregularities, u(t).

Here, the differential equation system includes the sprung

mass, ms , and the unsprung mass of the vehicle, mus , which
are connected by a linear spring and damper with the stiff-
0 2 4 6 8 10 12
ness coefficient, ks , and the damping constant, c, respectively. Time in s
The rail surface is considered as the system input, u(t), and
(b) Acceleration of the sprung mass, ms .
is transmitted by a spring (kus ) to the unsprung mass. This
kind of model, for example, might be applied to analysis vehi-
0.10

cle dynamics in simulation studies as part of vehicle design

phases. In Fig. 3 an example of the dynamic railway track
- vehicle interaction is illustrated. A simulated track irregu-
-0.10 -0.05 0.00 0.05
Acceletation in m/s2

larity (Fig. 3a) causes accelerations of the sprung mass (Fig.

3b) and the unsprung mass (Fig. 3c), respectively. In this
contribution, however, the reverse is of interest. Utilizing ac-
celeration measurements to reconstruct track irregularities by
model inversion concepts.

Table 1. Applied model parameters adapted from (Sira-

Ramirez et al., 2011).

Model parameter Numerical value

ms 9875 [kg] 0 2 4 6 8 10 12
mus 1100 [kg] Time in s
ks 2.13 × 106 [N/m] (c) Acceleration of the unsprung mass, mus .
kus 1.42 × 108 [N/m]
c 1.20 × 105 [N − s/m] Figure 3. Snapshot of the dynamic railway-vehicle interaction
at a vehicle speed of 10 m/s.

3
 I NTERNATIONAL J OURNAL OF P ROGNOSTICS AND H EALTH M ANAGEMENT

Input System (Vehicle + Sensor)

Model
Inversion +
Scheme

Inverse Model
Recalculated
Input Output
(Acceleration Data)

Acceleration in m/s²
0 1 2 3 4 5 6
Time in s

Figure 4. Workflow of Model Inversion: The DLR two-way vehicle RailDriVE is affected by the track quality. The resulting
vehicle response given by acceleration data in combination with the inverse model is used to recalcualted the track quality.

4. M ODEL I NVERSION BY I NVERSE S IMULATION input matrix B, the output matrix C, and the feedthrough
matrix D.
The objective of model-based inverse solutions is to recal-
culate the time history of inputs which had caused a given Applied to the suspension system (Eq. (1)) the corresponding
model output result. These recalculated inputs might be ap- matrices assuming ẋ = [ẍs , ẍus , ẋs , ẋus ]> and y = ẍs are:
plied to vehicles to follow a predefined trajectory (Thomson
& Bradley, 2006) or to force technical systems to desired op-
 c c ks ks
erating conditions (Graichen, Hagenmeyer, & Zeitz, 2005). − ms −m

ms s ms
In this contribution, however, the intention is a different one.
A =
 c
 mus − mcus ks
mus − (ksm+kus ) 
us 
Here, the focus is to recalculate model inputs which are as-  1 0 0 0 
sociated to the rail track quality, i.e. rail surface failures (e.g. 0 1 0 0
squats and corrugations) and rail track misalignments. As rail  
irregularities impact the mechanical train/vehicle suspension 0
 kus 
system, acceleration measurements of the resulting vehicle B =  mus 
 0 
dynamic might be used to reconstruct the actual track quality
by a model inversion strategy. 0
h i
ks ks
C = − mcs c
ms −m s ms
For the special case of Linear Time Invariant (LTI) systems D = [0]
the model can be represented in its state-space form:
ẋ = Ax + Bu Within the inverse simulation concept the original model is
(2) extended by a feedback control loop (Murray-Smith, 2011).
y = Cx + Du
For the purpose of illustration the state-space model is trans-
where u ∈ Rnu and y ∈ Rny are the systems inputs and the fered into the transfer function first, i.e. applying Laplace
outputs, respectively. The system states are given by x ∈ Rnx . Transformation:
The system matrices are known as the dynamic matrix A, the

4
 I NTERNATIONAL J OURNAL OF P ROGNOSTICS AND H EALTH M ANAGEMENT

to track monitoring issues and related problems. In general,

G(s) = C(sI − A)−1 B + D (3) to choose the best method for model inversion might be deli-
cate. However, the universal problem of imprecise parameter
In general, the transfer function, G(s), represents the in- estimates and their impact onto the simulation results applies
put/output behavior of the system under study: to all model-based approaches universally. To reliably assess
the contribution of parameter imperfections onto simulation
Y (s) results is still challenging and an important field of uncertainty
G(s) = (4) analysis. To the best of the authors knowledge this is the first
U (s)
analysis of global sensitivities for inverse simulation mod-
The corresponding closed-loop system (Fig. 5) reads as: els extending the local sensitivity framework (Murray-Smith,
2013).

Y (s) Kc K c · U (s) 5. S ENSITIVITY A NALYSIS

G∗ (s) = ∗
= c
=
U (s) 1 + K · G(s) U (s) + K c · Y (s)
(5) The large number of recent articles devoted to problems of un-
certainty analysis/management in the field of PHM (e.g. (Saha,
Goeble, Poll, & Christophersen, 2009; Daigle & Goebel, 2010;
U∗ U Y Daigle, Saxena, & Goebel, 2012; Lapira, Brisset, Davari,
− Kc G(s)
Siegel, & Lee, 2012; Williard, He, Osterman, & Pecht, 2013;
Sankararaman & Goebel, 2013; Sankararaman, Daigle, Sax-
ena, & Goebel, 2013; Daigle & Sankararaman, 2013; Kulka-
rni, Biswas, Celaya, & Goebel, 2013; Zhang & Pisu, 2014;
Schenkendorf, 2014)) indicates the significance of this topic.
Figure 5. Control loop: Inverse simulation by a proportional Here, the parameter sensitivity analysis is an important as-
feedback strategy. pect (e.g. (Sudret, 2007; Sankararaman, 2012; Chiachio et al.,
2015)) because of the following reasons. The quality of any
model-based result depends significantly on its applied model
In principle, for a large controller gain, K c , the inverse system parameters. Typically, model parameters cannot be identified
can be derived according to: perfectly by measurements, i.e. measurement imperfections
cause imprecise parameter estimates. Hence model parameters
U (s) should be treated as random variables instead as deterministic
lim G∗ (s) = (6) quantities. Naturally, those parameters have to be identified as
c
K →∞ Y (s)
precisely as possible which impact the model response at most.
When Eq. (6) is transfered back into the state-space notation In any model-building process one should be aware that some
the corresponding LTI system reads as: model parameters can be changed by order of magnitude with-
out any detectable model response variation. Apart from that,
ẋ† = A† x† + B † y slight displacements of other model parameters may have a se-
(7)
u = C † x† + D † y vere impact, i.e. the model response is insensitive or sensitive
to parameter variations/uncertainties. The insensitive group
where can be fixed at literature values (Sobol’, Tarantola, Gatelli,
Kucherenko, & Mauntz, 2007), which may simplify the actual
A† = (A − BK c C)
identification process of the sensitive parameters additionally
B† = BK c as the measurement information is split to a reduced number
C† = −K c C of quantities.
†
D = Kc A systematic quantification and classification of parameter
impacts can be done by a sensitivity analysis. In this study
At this point it should be stressed that the control gain has to essentials for the purpose of sensitivity analysis are briefly
be chosen deliberately to ensure a stable inversion (Murray- recapitulated. The general objective is to demonstrate how
Smith, 2011). Hence a stability analysis should be done in so-called global parameter sensitivities can be derived at low
parallel when designing the control gain matrix, K c . computational costs by implementing Polynomial Chaos Ex-
pansion (PCE) and Point Estimate Method (PEM) algorithms.
In principle, alternative methods might be applied to the pur- The dependency of model-based results to model parameters
pose of model inversion. As summarized in APPENDIX A, can be described generically by the following (non)-linear
some of these methods have even been successfully applied

5
 I NTERNATIONAL J OURNAL OF P ROGNOSTICS AND H EALTH M ANAGEMENT

mapping problem: mean and higher moments (e.g. skewness or kurtosis) are
η = g(θ) (8) neglected. Unsurprisingly, a parameter which cause a strong
In particular, the simulation result, η ∈ Rnη , may express offset/bias onto the model-based result but contributes less to
the vehicle dynamic (forward simulation; η := y) and the rail the response variance might be important for the model per-
quality (inverse simulation; η := u), respectively. The model formance, too. Alternatively, so-called moment-independent
parameters (e.g. stiffness coefficients and damping constants) importance measures have been recently introduced in the field
are given as elements of vector θ ∈ Rnθ . Depending on the of GSA (Borgonovo, 2007) analyzing the entire probability
test case, g(·) represents the classical feedforward model (e.g. density function instead of isolated statistical moments.
Eq. (1)) and the inverse model (e.g. Eq. (7)), respectively. The The moment-independent global sensitivity analysis (MI-GSA)
impact of the ith model parameter onto the simulation result aims to identify the impact of a parameter, [θ]i , onto the model-
can be determined by a global sensitivity analysis. based result and its probability density distribution (PDF),
In most global sensitivity analysis (GSA) methods the uncer- respectively (Fig. 7). Explicitly, the difference between the un-
tainty of model parameters is addressed explicitly, i.e. model- conditional PDF, pdf (η), and a conditional PDF, pdf (η|[θ]i ),
based results and parameters are considered as random vari- is determined according to:
ables. Most often as well as in this study, the uncertainty is Z
represented by general second-order random variables (ran- s([θ]i ) = |pdf (η) − pdf (η|[θ]i )|dη (9)
dom variables of finite variance), i.e. [θ]i ∈ L2 (Ω, F, P). The Ω
set Ω is a sample space, F is an appropriate σ-algebra on
Ω, and P is a probability measure (Grigoriu, 2002). Com- To address the uncertainty of [θ]i the expected value of s([θ]i )
monly, the global sensitivity is quantified by the analysis of has to be analyzed:
variance (ANOVA principle). While operating in a pure prob- Z
abilistic framework, the basic idea of so-called Sobol’ Indices E [s([θ]i )] = pdf ([θ]i )s([θ]i )d[θ]i (10)
(SI-GSA) is to identify the variance contribution that each Ω
parameter, [θ]i , adds to the variance of the model-based result,
η. Finally, the normalized importance measure of MI-GSA (Borgonovo,
2007) is given by:
η
η S[θ]
S[θ] η
1,2,3 1
2,3
S[θ] δi = E [s([θ]i )] (11)
1,3 2
Similar to the Sobol’ Indices, an insensitive factor has an
Higher-order
η Sobol’Indices importance measure close to zero and the overall sum holds
S[θ] 1,2 the inequality:
Overall
Uncertainty
(Normalized nθ
X
Variance η δi ≤ 1 (12)
of η) S[θ]3 i=1
η
S[θ]2
To sum up MI-GSA is an appropriate concept for the pur-
First-order
η Sobol’Indices pose of sensitivity analysis, its numerical calculation, however,
S[θ]1 might be a challenge. In most cases, the required probability
density functions cannot be calculated analytically and have to
be derived numerically, e.g. by Kernel density approximations.
This is achieved by a large number of model runs representing
Figure 6. Sobol’ Indices illustrated by a generic 3-dimensional the variation of the uncertain parameters. In case of cpu-
parameter problem, θ ∈ R3 . intensive model evaluations this might be prohibitive due to
limited cpu-power and time, respectively. A remedy might be
to evaluate handy surrogate functions, ĝ(·), instead. Here, the
To address the joint impact of uncertain parameters (Fig. 6) Polynomial Chaos Expansion comes into play and is applied
higher-order Sobol’ Indices can be derived additionally. For to bypass potential cpu-intensive processes as indicated in Fig.
more details see (Saltelli, Ratto, Tarantola, & Campolongo, 8. The basics of PCE and its efficient parameterization by the
2005; Sobol’, 1993) and references therein. Point Estimate Method are described in the sequel.
Obviously, SI-GSA takes the variance contribution into con-
sideration solely. Hence effects caused by variations of the

6
 I NTERNATIONAL J OURNAL OF P ROGNOSTICS AND H EALTH M ANAGEMENT

Start
Probability Density Function

0.6
pdf (η)
pdf (η|[θ]i ) Bypass Processing
Acceleration
0.4

s([θ]i )
Data
Default Processing
0.2

Model Inversion PCE Model

e.g. Inverse Simulation
(Eq. (14))
(Eq. (7))
0.0

-5 0 5 10 15
Track
Quality
Model-based result, η

Figure 7. Difference, s([θ]i ), between the unconditional PDF,

pdf (η), and the conditional PDF, pdf (η|[θ]i ), is highlighted. Stop

Figure 8. The cpu-intensive process of model inversion (dark

gray) is bypassed by a fast to evaluate PCE surrogate model
6. P OLYNOMIAL C HAOS E XPANSION (light gray).
To ensure a computationally tractable calculation, Polynomial
Chaos Expansion (PCE) is used to propagate the probabilistic
parameter uncertainties onto the model-based results. In detail, application is prohibitive in most cases. In this study, the
PCE aims to represent the model response by a weighted super- Point Estimate Method (PEM) is implement as a practicable
position of deliberately chosen basis functions, i.e. orthogonal alternative. As explained in Sec. 7, by combining PCE with
polynomials (Maitre & Knio, 2010), Ψi (·), according to: PEM the overall required sample number scales polynomially
(quadratic or cubic) with the dimension of ξ (the number of
∞
X random elements / parameters). Obviously, this corresponds
η = g(ξ) = ai Ψi (ξ) (13) to a significant reduction in comparison to standard numerical
i=0 integration methods, e.g. Gaussian Quadrature (exponential
where [ξ]i ∼ N (0, 1), i.e. the elements of ξ follow a standard growth) or Monte Carlo simulations (educated guess: sample
Gaussian distribution and are uncorrelated to each other. number ≥ 10,000; and challenging to quantify in general)
(Maitre & Knio, 2010). Once parameterized (Eq. (15)), the
For practical applications the expansion is implemented in
surrogate model response, η̂, can be derived at relative low
truncated form (lpce << ∞) as:
computational costs as the cpu-intensive process of model
inversion is bypassed by PCE (Fig. 8). Only now, Monte
lpce
X Carlo simulations propagating the uncertain parameters onto
η ≈ η̂ := ai Ψi (ξ) = a> Λ(ξ) (14) the model response become feasible. Subsequently, associated
i=0
PDFs can be derived from these accelerated Monte Carlo sim-
 >  >
where a := a0 , . . . , alpce and Λ := Ψ0 , . . . , Ψlpce . ulations by applying a Gaussian Kernel density approximation
(Bishop, 2008) as:
The model response is characterized by the PCE coefficients,
ai , and can be derived via:
N  
1 X η − η̂i
pdf (η) = K (16)
N h i=1 h
R
g(ξ) Ψi pdf (ξ) dξ
ai = RΩ (15)
Ψ (ξ)2 pdf (ξ) dξ
Ω i
with the bandwidth parameter h, the sample number N , and
The denominator may be solved analytically (Maitre & Knio, the Gaussian Kernel function:
2010), the numerator of Eq. (15), however, has to be calcu-
 
lated numerically. To keep the computational load tractable the η − η̂i 1 1 2

integrals have to be evaluated by efficient sampling strategies. K = K(z) := √ e− 2 z (17)

h 2π
Commonly, standard Monte Carlo simulations and Quadra-
ture methods involve an infeasible sample number, i.e. their So far, only standard Gaussian random variables, [ξ]i , have

7
 I NTERNATIONAL J OURNAL OF P ROGNOSTICS AND H EALTH M ANAGEMENT

been assumed. A way to incorporate general Gaussian random

variables, which may describe problem dependent parameter nξ
uncertainties ([θ]i ∼ N (µi , σi )), reads as: w0 = 1−
ϑ2
1
[θ]i := qi ([ξ]i ) = µi + σi [ξ]i (18) w1 =
2ϑ2
√
Similar transformation functions, qi (·), can be derived for ϑ = 3
different types of parametric PDFs, too, as summarized in Tab.
2.

Table 2. Probability density function transformation formulas 4

adapted from (Isukapalli, 1999). Here, the term erf means the Contour Plot
error function.

Type of PDF Transformation: qi ([ξ]i ) :=

Normal(µ, σ) µ + σ[ξ]i 2
√

Uniform(a, b) a + (b − a) 12 + 12 erf([ξ]i 2)
Log-normal(µ, σ) exp(µ + σ[ξ]i )
 q 3
1 1
Gamma(a, b) ab [ξ]i 9a + 1 − 9a

[ξ]2
   0
Exponential(λ) − λ1 log 12 + 12 erf [ξ]
√i
2

−2
Such an isoprobabilisitic transform (Sudret, 2007) can also
be applied to incorporate non-parametric empirical PDFs
(Schöniger, Nowak, & Franssen, 2012) as well. Thus an effi-
cient sampling strategy for (standard) Gaussian distributions is −4
−4 −2 0 2 4
of utmost significance within the proposed framework of sen-
[ξ]1
sitivity analysis. For this purpose the Point Estimate Method
is tailor-made and explained in the sequel.
Figure 9. Bivariate standard Gaussian distribution: Contour
plot superimposed by sample points generated by PEM3( , )
7. P OINT E STIMATE M ETHOD and PEM5( , , ), respectively.
In the last two decades the so-called Unscented Transforma-
tion (UT) introduced by Julier and Uhlmann in 1994 (Julier &
Uhlmann, 1994) has become the gold standard in non-linear
filtering. The roots of UT, however, date back more than 60 Now, an nξ -dimensional integration problem can be solved
years in time (Tyler, 1953). In detail, Point Estimate Meth- approximatively by a weighted superposition of function eval-
ods (PEM) had been introduced to solve multi-dimensional uations at these deliberately chosen sample points according
integration problems over symmetrical regions, e.g. symmet- to:
ric probability density functions (Evans, 1967, 1974). Due Z 2nξ
to this symmetry, numerical integration techniques can be X
g(ξ)pdfξ dξ ≈ w0 g(ξ0 ) + w1 g(ξk ) (20)
derived which scale polynomially (quadratic or cubic) to n- Ω k=1
dimensional integration problems (e.g. Eq. (15)). Essentials of
PEM are briefly reviewed in the sequel following annotations As only a finite number of raw moments of the input random
given in (Tyler, 1953; Lerner, 2002). variable, ξ, is considered, a non-linear function, g(·), is approx-
The main idea of PEM as a sampling approach is to generate imated by monomials of finite degree (Evans, 1967; Lerner,
sample points, ξk , by permuting at most one element of ξ ∈ 2002). For the proposed sampling strategy (Eqs. (19)-(20))
Rnξ at once as: monomials up to order three are approximated correctly and
[ξk ]i := [ξ0 ]i ± ϑ (19) consequently labeled as PEM3 in what follows.
Due to combinatorics 2nξ sample points can be derived addi- Moreover, the general precision of the PEM approach can
tionally to the original random vector, ξ0 = E[ξ]. To deter- be increased gradually by incorporating an increased sample
mine the permutation constant, ϑ, and the associated sampling number and considering higher order raw moments of ξ, re-
weights, wi , the first raw statistical moments of ξ are utilized spectively. Hence an approximation scheme can be applied
(Lerner, 2002) and read as: which represents monomials of g(·) correctly up to the preci-

8
 I NTERNATIONAL J OURNAL OF P ROGNOSTICS AND H EALTH M ANAGEMENT

sion of 5 via: vehicle response (Eq. (3)) by means of the frequency response,
Z Bode Diagram (Fig. 10), it can be seen that: i) Excluding sta-
g(ξ)pdfξ dξ ≈ w0 g(ξ0 )+ tionary changes (very low frequencies) both sensor setups, i.e.
Ω
acceleration data of the unsprung mass and the sprung mass,
2nξ 2nξ (nξ −1) (21)
X X respectively, are feasible configurations. ii) Track irregularities
w1 g(ξk ) + w2 g(ξf ) impact the unsprung mass at most, i.e. associated acceleration
k=1 f =1
data are more informative. Thus data of the unsprung mass
The approximation scheme given in Eq. (21) is labeled as might be preferable in theory.
PEM5 subsequently. In this case, the number of generated In this study, however, the focus is on acceleration data of the
sample points correlates to 2n2ξ + 1 for a nξ -dimensional sprung mass for the simple reason that the sensor hardware can
integration problem. For the purpose of parametrization of wi be installed inside of the vehicle body and maintained easily -
and ϑ an equation system can be derived taking into account a basic requirement for the intended purpose of a continuously
monomials of degree 5 or less (Lerner, 2002). operating rail condition monitoring system on a large number
The four unknowns can be uniquely determined as: of commercial in-line trains.

Bode Diagram
√
ϑ = 3 100
Gs
n2ξ − 7nξ

Magnitude (dB)
Gus
w0 = 1+
18 50
4 − nξ
w1 =
18
1 0
w2 =
36
100 101 102 103 104
180
Resulting sample points, for example, are illustrated in Fig.
135
(9) for a bivariate standard Gaussian distribution.
Phase (deg)

90
In summary, the proposed PEM5 framework provides a trade- 45
off between accuracy and computational demands and is ap- 0
plied in subsequent considerations for this very reason. The −45
deliberately chosen sample points can be applied to determine −90
the PCE coefficients (Eq. (15)) efficiently. In so doing, Sobol’ 100 101 102 103 104
Indices might be derived immediately by the PCE coefficients Frequency (rad/sec)
(Sudret, 2007; Alexanderian, 2013) at a total cost of 2n2ξ + 1
function evaluations. As conditional probabilities have to be
incorporated in case of MI-GSA explicitly, the overall sample Figure 10. Bode Diagram of the proposed quarter-vehicle
model assuming the acceleration of the sprung mass (Gs ) and
number scales cubically (while eliminating redundant identical the unsprung mass (Gus ) as the system output, respectively.
samples):

8.1. Scenario 1: Inverse Simulation

Sample Number(P EM 5) = 2n3ξ − 6n2ξ + 10nξ − 3 (22)
In this scenario simulated acceleration data of the sprung mass
Thus, the added value of MI-GSA has to be paid for by an ms are utilized to recalculate the rail track quality. To do so,
increased (but manageable) computational demand. An effort the quarter-vehicle model (Eq. (1)) has to be inverted. As the
which is worth to take into consideration in many practical input (rail surface) is indirectly linked to the sprung mass, an
studies of sensitivity analysis. The numerical results given analytical inversion (Sec. 4) fails. When measuring the sprung
in Sec. 8 confirm the usefulness of the proposed concept mass acceleration, the system output, y sim (t) = d2 xs /dt2 ,
convincingly. is not a flat output (y f (t) 6= y sim (t)). In consequence, also
the flatness approach (Sec. A.2) is not directly applicable. As
8. N UMERICAL R ESULTS measurement imperfections are neglected, the Inverse Simula-
tion (Sec. 4) is favored over Kalman filtering (Sec. A.3) for
First, the general applicability of acceleration data to track ir- simplicity and implemented for subsequent analysis.
regularity detection is assessed. When comparing the dynamic
Assuming a perfect model (model imperfections will be ex-

9
 I NTERNATIONAL J OURNAL OF P ROGNOSTICS AND H EALTH M ANAGEMENT

plicitly addressed in ongoing work) as well as perfect model

parameters rail irregularities can be reconstructed perfectly by

10000 15000 20000 25000

the inverse simulations approach (Fig. 11). Here, a control
gain of K c = 1000 is applied.

Frequency
0
Track Irregularity in mm
-0.5

Original
Recalculation

5000
-1

0
-2 -1.95 -1.9 -1.85 -1.8 -1.75
-1.5

Track Irregularity in mm
Figure 12. Histogram of the recalculated track irregularity (t =
7s) representing the uncertainty induced by model parameter
-2

imperfections.
0 2 4 6 8 10 12
Time in s
Figure 11. Benchmark of the original and the recalculated In a similar way, the associated model parameter sensitivities
track irregularity: Almost no differences can be seen. (MI-GSA) are derived at a total cost of 27 (Eq. (22)) cpu-
intensive inverse simulations (Eq. (7)). Any subsequent cal-
culation is based on easily to evaluate PCE model surrogates
To be more realistic, it is assumed that the stiffness constants (Eq. (14)). For instance, parameter sensitivities associated to
and the damping constant are only vaguely known, i.e. these 3 the recalculated track quality are illustrated in Fig. 13 for time
model parameters are described by random variables (Tab. 3). point t = 7s.
Obviously, the stiffness constant ks has the strongest impact
Table 3. Assumed parameter uncertainties following a Gaus-
sian distribution, [θ]i ∼ N (µi , σi ). onto the recalculated track quality at t = 7s. Minor contribu-
tions of the parameters kus , c, and parameter combinations
Mean Value Standard Deviation {ks , kus , c} can also be detected. In consequence, the model
µks = ks σks = 1/3ks parameter ks has to be known as precisely as possible to ensure
µkus = kus σkus = 1/3kus a reliable recalculation of track irregularity maximum.
µc = c σc = 1/3c
8.2. Scenario 2: Classical Simulation
In this case, the uncertainty of the model parameters is trans- The proposed concept of an efficient global parameter sensitiv-
fered onto the model-based results. For instance, in Fig. 12 a ity analysis can also be applied to classical simulations, e.g. to
histogram of the recalculated track quality is illustrated at the simulate the vehicle response (Eq. (1)) for given artificial track
time point of maximum displacement (t = 7s) shown in Fig. quality configurations. Here, the same parameter uncertainties
11. This histogram is based on 100.000 accelerated Monte as in the previous study and the same track profile as shown
Carlo simulations, i.e. a truncated third-order PCE model in Fig. 11 are assumed. In Fig. 14 parameter sensitivities are
(Eq. (14)) is evaluated instead of the inverse model (Eq. (7)). shown of the simulated sprung mass acceleration at time point
Here, the parameterization of the PCE model was done by 9 t = 7s. Compared to the sensitivities derived for the inverse
(2n2ξ + 1; nξ = 3) evaluations of the original inverse model simulation, a different range of parameter sensitivities can be
which means a significant reduction in computational costs seen. Here, the stiffness parameter ks determines the simu-
(i.e. 100.000 runs of the cpu-intensive inverse model). In prin- lation result almost exclusively. In principle, it may happen
ciple, advanced sampling methods like Importance Sampling that a sensitive model parameter for the classical simulation is
and MCMC can be applied to the surrogate model as well to insensitive for the inverse simulation and vice versa. Thus in-
improve the overall computational time additionally. dividual sensitivity analysis studies for the inverse simulation

10
 I NTERNATIONAL J OURNAL OF P ROGNOSTICS AND H EALTH M ANAGEMENT

c,ks ,kus

c,ks ,kus
kus

kus
Parameters

Parameters
ks

ks
c

c
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Importance Measure, δi Importance Measure, δi
Figure 13. Sensitivity Analysis applied to the inverse simula- Figure 14. Sensitivity Analysis applied to the classical simu-
tion. lation.

A. A PPENDIX

have to be performed. Alternative mathematical tools which can be applied to model

inversion are shortly summarized below following notations
9. C ONCLUSION given in (Buchholz & v. Grünhagen, 2007; Murray-Smith,
2011).
In this study it was shown how global parameter sensitivities
can be derived efficiently by combining Polynomial Chaos A.1. Analytical Inversion
Expansion and Point Estimate Method principles. Considera-
tions which are relevant for any model-based analyses, thus Assuming a regular D matrix the input can be determined by
for model-based concepts in PHM, too. In this regard, the given outputs as:
general idea of a model-based rail condition monitoring setup
u = D−1 (y − Cx) = −D−1 Cx + D−1 y
was introduced. A model inversion strategy enabled the re-
calculation of the track quality by incorporating acceleration Inserted in Eq. (2) the corresponding LTI reads as:
data of the vehicle dynamic. Which model parameters have to
be determined as precisely as possible to ensure meaningful ẋ∗ = A∗ x ∗ + B ∗ y
calculations could be identified efficiently by the proposed ap- u = C ∗ x∗ + D ∗ y
proximation and sampling scheme. It is common knowledge
that a comprehensive sensitivity analysis contributes to model where
calibration and refinement. Improved models in PHM ensure
A∗ = (A − BD−1 C)
more credible model-based prognostic results in consequence.
∗
In future, sensitivity studies of more complex inverse models, B = BD−1
e.g. full-vehicle models, will be analyzed. C∗ = −D−1 C
D∗ = D−1
ACKNOWLEDGMENT
In the common case of non-regular D matrices, however, al-
We thank apl. Prof. Dr.-Ing. Michael Mangold for his helpful
ternative methods (e.g. inverse simulation) have to be applied.
comments on a previous version of the paper. Moreover, the
anonymous reviewers deserve special thanks for their valuable
A.2. Flatness Approach
input. The financial support by Deutsches Zentrum fur Luft-
und Raumfahrt (DLR e.V., Cologne) under grant TrackScan is The concept of differential flatness (Fliess, Lévine, Martin, &
gratefully acknowledged. Rouchon, 1992, 1995) is applicable for linear as well as non-

11
 I NTERNATIONAL J OURNAL OF P ROGNOSTICS AND H EALTH M ANAGEMENT

linear systems and aims to represent the system states and the timate the track geometry (Naganuma et al., 2013a). Different
inputs by so-called flat outputs, y f (t), and derivatives thereof. types of (non)-linear observer strategies might be applicable
The determination of these flat outputs might be challenging. in a similar way, too. For instance, a Sliding Mode Observer
In most cases, flat outputs do not agree with originally given framework has been proposed reconstructing the road profile
output configurations, y f (t) 6= y sim (t), but have to be defined by sensors mounted on a heavy vehicle (Imine & Fridman,
appropriately. If a flat output exists the states and inputs can 2008).
be expressed as:

x(t) = Ψx (y f (t), Dy f (t), . . . , Dn−1 y f (t)) R EFERENCES

f f n f
u(t) = Ψu (y (t), Dy (t), . . . , D y (t)), Alexanderian, A. (2013). On spectral methods for variance
i
where D represents the operator notation of the ith derivative, based sensitivity analysis. Probability Surveys.
di /dti . Bishop, C. M. (2008). Pattern recognition and machine
learning. Springer.
Recently, the flatness approach has been applied to an active Bocciolone, M., Caprioli, A., Cigada, A., & Collina, A. (2007).
regulation concept of a railway vehicle suspension system A measurement system for quick rail inspection and ef-
(Sira-Ramirez et al., 2011). In general, however, a successful fective track maintenance strategy. Mechanical Systems
application depends on the relative degree of the system under and Signal Processing, 21(3), 1242–1254.
study. In case of single input single output LTI systems this
Borgonovo, E. (2007). A new uncertainty importance measure.
corresponds to the difference of the numerator and denomina-
Reliab. Eng. Sys. Safety, 92, 771–784.
tor polynomials of the associated transfer function. A system
Buchholz, J., & v. Grünhagen, W. (2007). Inversion impossible.
is called flat when this difference is equal to the number of the
GRIN Publishing GmbH.
system states, i.e. a flat output exists in principle.
Caprioli, A., Cigada, A., & Raveglia, D. (2007). Rail in-
In case of noisy measurements and model imperfections the spection in track maintenance: A benchmark between
Kalman Filter is a reasonable alternative. the wavelet approach and the more conventional fourier
analysis. Mechanical Systems and Signal Processing,
A.3. Kalman Filtering 21(2), 631–652.
The underlying idea is to extend the system states, x, by the Chiachio, J., Chiachio, M., Saxena, A., Sankararaman, S., Rus,
system inputs, u, according to: G., & Goebel, K. (2015). Bayesian model selection and
parameter estimation for fatigue damage progression
>
x‡ = [x, u] models in composites. International Journal of Fatigue.
Daigle, M., & Goebel, K. (2010). Improving computational
In the field of unknown input reconstruction (Witczak, 2014), efficiency of prediction in model-based prognsotics us-
the inputs are frequently treated as Gaussian random variables, ing the unscented transformation. Procceedings of the
i.e. [u]i ∼ N ([u]i , σu2 ). The time-discrete version of the Annual Conference of the Prognostics and Health Man-
corresponding state-space system is given by: agement Society.
x‡k = A‡ x‡k−1 + wk−1 Daigle, M., & Sankararaman, S. (2013). Advanced methods
for determining prediction uncertainty in model-based
yk = C ‡ x‡k + vk prognostics with application to planetary rovers. Annual
Conference of the Prognostics and Health Management
where A‡ and B ‡ are the adapted dynamic matrix and output
Society.
matrix. Moreover, wk and vk represent the process noise and
the measurement noise (Gelb, 1974), respectively. Within Daigle, M., Saxena, A., & Goebel, K. (2012). An efficient
D
the Kalman Filter framework recent measurements, yk+1 , are deterministic approach to model-based prediction uncer-
‡,(+) tainty estimation. Annual Conference of the Prognostics
utilized to reconstruct the extended state vector, xk+1 , based and Health Management Society.
‡,(−)
on a preliminary model-based prediction, xk+1 , as: Evans, D. H. (1967). An application of numerical integration
  techniques to statistical tolerancing. Technometrics, 9,
‡,(+) ‡,(−) ‡,(−)
xk+1 = xk+1 + Kk+1 yk+1 D
− C ‡ xk+1 441-456.
Evans, D. H. (1974). Statistical tolerancing: The state of the
Details about the Kalman gain, Kk+1 , can be found among art. Journal of Quality Technology, 6, 188-195.
others in (Stengel, 1994). In consequence, the unknown sys- Feldmann, U., Kreuzer, E., & Pinto, F. (2000). Dynamic
tem inputs, u, are estimated in parallel with the system states, diagnosis of railway tracks by means of the karhunen–
x, and can be evaluated subsequently for the purpose of PHM. loève transformation. Nonlinear Dynamics, 22(2), 183–
Recently, the Kalman Filter has been successfully applied to es- 193.

12
 I NTERNATIONAL J OURNAL OF P ROGNOSTICS AND H EALTH M ANAGEMENT

Fliess, M., Lévine, J., Martin, P., & Rouchon, P. (1992). On dif- Lee, J. S., Choi, S., Kim, S.-S., Park, C., & Kim, Y. G. (2012).
ferentially flat non-linear systems. C.R. de l’Académie A mixed filtering approach for track condition moni-
des Sciences, 315, 619-624. toring using accelerometers on the axle box and bogie.
Fliess, M., Lévine, J., Martin, P., & Rouchon, P. (1995). IEEE Transactions on Instrumentation and Measure-
Flatness and defect of nonlinear systems: Introductory ment, 61(3), 749–758.
theory and examples. Int. J. Control, 61, 1327-1361. Lerner, U. N. (2002). Hybrid bayesian networks for rea-
Gelb, A. (1974). Applied optimal estimation. The M.I.T. soning about complex systems (Unpublished doctoral
Press. dissertation). Stanford University.
Gradinariu, T., Tursø-Finnich, K., Juntti-Espling, U., Mangan, Lewis, R., & Richards, A. (1988). A compensated accelerom-
C., Thill, L., Gebhardt, M., . . . Gardin, D. (2008). Last- eter for the measurement of railway track crosslevel.
ing infrastructure cost benchmarking (licb): Summary IEEE Transactions on Vehicular Technology, 37(3), 174–
report december 2008. 178.
Graichen, K., Hagenmeyer, V., & Zeitz, M. (2005). A new Lewis, R. B., & Richards, A. N. (1986). A new method for
approach for inversion-based feedforward control de- the routine measurement of railhead corrugations. Rail
sign of non-linear systems. Trans. Autom. Control, 41, International(2), 37–41.
2033-2041. Maitre, O. P. L., & Knio, O. M. (2010). Spectral methods for
Grigoriu, M. (2002). Stochastic calculus: Applications in uncertainty quantification. Springer.
science and engineering. Birkhäuser. Molodova, M., Li, Z., & Dollevoet, R. (2011). Axle box
Imine, H. (2011). Sliding mode based analysis and identifica- acceleration: Measurement and simulation for detection
tion of vehicle dynamics. Springer. of short track defects. Wear, 271(1-2), 349–356.
Imine, H., & Fridman, L. (2008). Road profile estimation Molodova, M., Li, Z., Nunez, A., & Dollevoet, R. (2014).
in heavy vehicle dynamics simulation. International Automatic detection of squats in railway infrastructure.
Journal of Vehicle Design. IEEE Transactions on Intelligent Transportation Sys-
International Union of Railways (UIC). (2010). Monitoring tems, 15(5), 1980–1990.
track condition to improve asset management: Synthesis Mori, H., Sato, Y., Ohno, H., Tsunashima, H., & Saito, Y.
report of the uic track condition monitoring working (2013). Development of compact size onboard device
group. for condition monitoring of railway tracks. Journal of
Isukapalli, S. S. (1999). Uncertainty analysis of transport- Mechanical Systems for Transportation and Logistics,
transformation models (Unpublished doctoral disserta- 6(2), 142–149.
tion). Rutgers, The State University of New Jersey. Mori, H., Tsunashima, H., Kojima, T., Matsumoto, A., &
Julier, S. J., & Uhlmann, J. K. (1994). A general method Mizuma, T. (2010). Condition monitoring of railway
for approximating nonlinear transformations of proba- track using in-service vehicle. Journal of Mechanical
bility distributions (Tech. Rep.). Dept. of Engineering Systems for Transportation and Logistics, 3(1), 154–
Science, University of Oxford. 165.
Kawasaki, J., & Youcef-Toumi, K. (2002). Estimation of rail Murray-Smith, D. J. (2011). Feedback mehtods for inverse
irregularities. In 2002 american control conference (pp. simulation of dynamic models for engineering systems
3650–3660). applications. Mathematical and Computer Modelling
Kobayashi, M., Naganuma, Y., Nakagawa, M., & Okumura, T. of Dynamical Systems, 17, 515-541.
(2008). Digital inertial algorithm for recording track ge- Murray-Smith, D. J. (2013). The application of parameter sen-
ometry on commercial shinkansen trains. In Computers sitivity analysis methods to inverse simulation models.
in railways xi (Vol. 103, pp. 683–692). Mathematical and Computer Modelling of Dynamical
Kobayashi, T., Naganuma, Y., & Tsunashima, H. (2013). Systems, 19, 67-90.
Condition monitoring of shinkansen tracks based on Naganuma, Y., Kobayashi, M., & Okumura, T. (2010). Inertial
inverse analysis. Chemical Engineering Transactions, measurement processing techniques for track condition
33, 703–708. monitoring on shinkansen commercial trains. Journal of
Kulkarni, C. S., Biswas, G., Celaya, J. R., & Goebel, K. (2013). Mechanical Systems for Transportation and Logistics,
Physics based degradation models for electronic capac- 3(1), 315–325.
itor prognostics under thermal overstress conditions. Naganuma, Y., Kobayashi, T., & Tsunashima, H. (2013a).
International Journal of Prognostics and Health Man- Track geometry estimation for car-body motions of rail-
agement. way vehicle. Journal of Mechanical Systems for Trans-
Lapira, E., Brisset, D., Davari, H., Siegel, D., & Lee, J. portation and Logistics.
(2012). Wind turbine performance assessment using Naganuma, Y., Kobayashi, T., & Tsunashima, H. (2013b).
multi-regime modeling approach. Renewable Energy, Track geometry estimation from car-body motions of
45, 86-95. railway vehicle. Journal of Mechanical Systems for

13
 I NTERNATIONAL J OURNAL OF P ROGNOSTICS AND H EALTH M ANAGEMENT

Transportation and Logistics, 6(2), 133–141. Ward, C., Weston, P., Stewart, E. J. C., Li, H., Goodall, R.,
Saha, B., Goeble, K., Poll, S., & Christophersen, J. (2009). Roberts, C., . . . Dixon, R. (2011). Condition monitoring
Prognostics methods for battery health monitoring using opportunities using vehicle-based sensors. Proceedings
a bayesian framework. IEEE TRANSACTIONS ON of the Institution of Mechanical Engineers, Part F: Jour-
INSTRUMENTATION AND MEASUREMENT, 52(2). nal of Rail and Rapid Transit, 225(2), 202–218.
Saltelli, A., Ratto, M., Tarantola, S., & Campolongo, F. (2005). Weston, P., Ling, C., Goodman, C., Roberts, C., Li, P., &
Sensititivity analysis for chemical models. Chemical Goodall, R. (2007). Monitoring lateral track irregularity
Reviews, 105, 2811–2828. from in-service railway vehicles. Proceedings of the
Sankararaman, S. (2012). Uncertainty quantification and inte- Institution of Mechanical Engineers, Part F: Journal of
gration in engineering systems (Unpublished doctoral Rail and Rapid Transit, 221(1), 89–100.
dissertation). Vanderbilt University. Weston, P., Ling, C., Roberts, C., Goodman, C., Li, P., &
Sankararaman, S., Daigle, M., Saxena, A., & Goebel, K. Goodall, R. (2007). Monitoring vertical track irregular-
(2013). Analytical algorithms to quantify the uncer- ity from in-service railway vehicles. Proceedings of the
tainty in remaining useful life prediction. In Aerospace Institution of Mechanical Engineers, Part F: Journal of
conference. Rail and Rapid Transit, 221(1), 75–88.
Sankararaman, S., & Goebel, K. (2013). Uncertainty quan- Williard, N., He, W., Osterman, M., & Pecht, M. (2013).
tification in remaining useful life of aerospace compo- Comperative analysis of features for determining state
nents using state space models and inverse form (Tech. of health in lithium-ion batteries. International Journal
Rep.). NASA Ames Research Center; Moffett Field, of Prognostics and Health Management.
CA, United States. Witczak, M. (2014). Fault diagnosis and fault-tolerant control
Schenkendorf, R. (2014). A general framework for uncertainty strategies for non-linear system: Analytical and soft
propagation based on point estimate methods. In Sec- computing approaches. Springer.
ond european conference of the prognostics and health
Zhang, X., & Pisu, P. (2014). Prognostic-oriented fuel cell
management society, phme14. Nantes, France.
catalyst aging modeling and its application to health-
Schöniger, A., Nowak, W., & Franssen, H.-J. H. (2012). Pa- monitoring and prognostics of a pem fuel cell. Interna-
rameter estimation by ensemble kalman filters with tional Journal of Prognostics and Health Management,
transformed data: Approach and application to hy- 5.
draulic tomography. Water Resources Research, 48.
Sira-Ramirez, H., Matamoros-Sanchez, A., & Goodall, R. M.
(2011). Flatness based control of a suspension system: B IOGRAPHIES
A gpi observer approach. In 18th ifac world congress. René Schenkendorf René received his
Sobol’, I. M. (1993). Sensitivity analysis for nonlinear mathe- Dipl.-Ing. and Dr.-Ing. in Engineering
matical models. Mathematical Modeling and Computa- Cybernetics from the Otto-von-Guericke-
tional Experiment, 1, 407–414. University Magdeburg, Germany in 2007
Sobol’, I., Tarantola, S., Gatelli, D., Kucherenko, S., & and 2014, respectively. From 2007 until
Mauntz, W. (2007). Estimating the approximation 2012 he had been a Ph.D. student at the Max
error when fixing unessential factors in global sensitiv- Planck Institute for Dynamics of Complex
ity analysis. Reliability Engineering & System Safety, Technical Systems in Magdeburg, Germany.
92(7), 957 - 960. Since 2013 he is with the German Aerospace Center at the
Stengel, R. F. (1994). Optimal control and estimation. Dover Institute of Transportation Systems in Brunswick, Germany.
Publications. His current research interests include data analysis, statisti-
Sudret, B. (2007). Uncertainty propagation and sensitivity cal process control, uncertainty propagation, and modeling in
analysis in mechanical models. Universite BLAISE PHM.
PASCAL - Clermont II.
Sunaga, Y., Sano, I., & Ide, T. (1997). A method to control the Jörn C. Groos Jörn received is diploma in
short wave track irregularities utilizing axlebox acceler- geophysics from the University of Karlsruhe,
ation. Railway Technical Research Institute, Quarterly Germany, in 2007 and his Dr. rer. nat. from
Reports, 38(4), 176–181. the faculty of physics of the Karlsruhe In-
Thomson, D., & Bradley, R. (2006). Inverse simulation stitute of Technology in 2010. Since 2014
as a tool for fight dynamics research - principles and he is with the German Aerospace Center at
applications. Progress in Aerospace Sciences, 42, 174- the Institute of Transportation Systems in
210. Brunswick, Germany. His current research
Tyler, G. W. (1953). Numerical integration of functions of interests include data acquisition, data analysis, and pattern
several variables. Canadian Jn. Math., 5, 393-412. recognition for condition monitoring of railway tracks.

14

View publication stats