You are on page 1of 12

Struct Multidisc Optim (2011) 44:137148

DOI 10.1007/s00158-010-0613-8
INDUSTRIAL APPLICATION
Probability-based multiple-criteria optimization
of bridge maintenance using monitoring and expected
error in the decision process
Andr D. Orcesi Dan M. Frangopol
Received: 13 February 2010 / Revised: 31 October 2010 / Accepted: 10 December 2010 / Published online: 3 February 2011
c Springer-Verlag 2011
Abstract In the context of limited financial resources under
uncertainty, an important challenge is to introduce monitor-
ing concepts in the general assessment, maintenance and
repair frameworks of highway bridges. Structural health
monitoring (SHM) provides new information on structural
performance. As the monitoring duration increases, addi-
tional information becomes available. The knowledge of
the structure is more accurate, but the monitoring costs are
greater. Based on the accuracy of monitoring, different deci-
sions can be made. These decisions involve uncertainties,
and, consequently, are expressed in terms of probabilities.
A probability-based approach for multiple criteria optimiza-
tion of bridge maintenance strategies based on SHM is pro-
posed. A measure of the error in the decision process, based
on monitoring occurrence and duration, is proposed. Opti-
mal solutions are obtained considering multiple criteria such
as expected failure cost, expected monitoring/maintenance
cost, expected accuracy of monitoring results, and ensuring
that all constraints are satisfied.
Keywords Optimization Structural health monitoring
Uncertainty Maintenance Cost
A. D. Orcesi (B)
LCPC, Bridges and Structures Department,
Safety and Durability of Structures, 58 boulevard Lefebvre,
75732 Paris Cedex 15, France
e-mail: andre.orcesi@lcpc.fr
D. M. Frangopol
Department of Civil and Environmental Engineering,
ATLSS Engineering Research Center, Lehigh University,
117 ATLSS Dr., Bethlehem, PA 18015-4729, USA
e-mail: dan.frangopol@lehigh.edu
1 Introduction
With aging of civil infrastructure systems, limited financial
funds and ever increasing structural performance demands,
the optimization of assessment, maintenance, and rehabil-
itation strategies becomes crucial. To determine optimal
long-term maintenance strategies, stakeholders require first
to assess the performance of their structural facility in
an optimal way (Frangopol and Liu 2007). In parallel to
in-depth inspections, monitoring techniques are now able
to effectively assess structural performance and demands
on a structure (Ni et al. 2006a, b; Hua 2006; Hess 2007;
Frangopol et al. 2008a, b; Strauss et al. 2008; Liu et al.
2009). SHM was largely used in past years to detect struc-
tural damage. It becomes now a challenge to provide con-
cepts and methodologies integrating SHM information in a
life-cycle analysis to determine optimal maintenance strate-
gies (Klinzmann et al. 2006; Frangopol and Messervey
2009). Orcesi et al. (2010) proposed a general framework
that includes strain monitoring data in a global reliability-
based optimization analysis and considers various limit
states under uncertainties. They have used this information
to update some parameters in the limit state functions and, in
turn, improve the accuracy of the reliability analysis. Also,
they have shown that considering limit states separately
or simultaneously can lead to different optimal mainte-
nance solutions. However, the question of the monitoring
duration/occurrence impact on the maintenance decision is
not fully addressed and providing analyses that (1) include
monitoring information in a life-cycle analysis under uncer-
tainties, and (2) quantify the error in the decision process
due to the monitoring duration/occurrence, is still a challenge.
A probability-based methodology is proposed in this
paper, to capture the uncertainties inherent to the structural
degradation and decision processes. The possible decisions
138 A.D. Orcesi, D.M. Frangopol
in the lifetime of the structure are presented in form of
an event tree (Thoft-Christensen and Srensen 1987; Estes
and Frangopol 2001; Kim and Frangopol 2010; Orcesi and
Cremona 2010). The originality of the proposed approach
consists in the quantification of the probability of mainte-
nance decisions in the event tree by using SHM information.
The proposed approach is described and illustrated by
using the I-39 Northbound B-37-75 Bridge, over the
Wisconsin River, located near Wausau, WI, USA (denoted
herein as the Wisconsin Bridge), long-term monitored by
the Engineering Research Center for Advanced Technology
for Large Structural Systems (ATLSS) at Lehigh University,
Bethlehem, PA (Mahmoud et al. 2005). Strain monitoring
was performed on this bridge during 84 days. The moni-
toring program on I-39 Bridge is used herein to investigate
the impact of monitoring occurrence and duration in a cost-
based maintenance optimization framework. To investigate
the impact of monitoring occurrence, the set of monitor-
ing information associated with the 84 days of the monitor-
ing program is split into two monitoring sets
1
and
2
by
considering the 42 first and last days, respectively. Depend-
ing on the monitoring occurrence, the available information
might change which, in turn, might impact maintenance
strategies. To analyze the impact of monitoring duration,
a fraction of the monitoring information is assumed to be
available, both when considering
1
and
1
. The shorter
the duration of monitoring programs, the less information
to make decisions and, consequently, the greater the error
associated with the decision. A threshold for essential main-
tenance decisions such as rehabilitation or replacement is
optimized in the proposed approach. This probability-based
approach provides stakeholders with optimal maintenance
solutions that minimize maintenance and failure costs while
satisfying constraints at the end of the time horizon. An
optimization procedure using genetic algorithms (GA) is
introduced to determine optimal solutions. This paper is
organized as follows: Section 2 introduces general relia-
bility concepts, the safety and performance margins, and
details the event tree approach. Section 3 introduces the
optimization framework by providing first the expressions
of the objective functions and, second, the optimization
problem formulation. Section 4 presents the main results
obtained when considering the Wisconsin Bridge. Finally,
Section 5 details conclusions and the needs for further
researches.
2 Reliability analysis
The structural performance indicator is associated with
structural reliability. Such indicator is a practical measure
that enables to include uncertainties inherent to resistance
and loading processes in the life of a structure.
2.1 Safety margin
The probability of failure of a component i , for a particu-
lar limit state, is defined as the probability of violating the
associated limit state function. The case of a limit state
function, relating the resistance of the structural compo-
nent i to the load effects acting on this component, is
considered herein. Safety margin at a point in time t is
expressed as M
i
(t ) = R
i
(t ) Q
i
(t ) where M
i
(t ) = instan-
taneous safety margin, R
i
(t ) = instantaneous resistance,
and Q
i
(t ) = instantaneous load effect. The reliability index
is then associated with the probability that M
i
(t ) 0 as
(t ) =
1
(P(M
i
(t ) 0)) where
1
(.) = the inverse
of the standard normal cumulative distribution function ,
and P(M
i
(t ) 0) = probability of occurrence of the
event M
i
(t ) 0. The cross-section S
2
of the Wisconsin
Bridge (Fig. 1) is considered herein. Since this cross-section
is noncomposite and noncompact, the maximum flexural
strength of the steel composite section cannot be computed
as the resultant moment of the fully plastic stress distri-
bution because the steel section will fail with local web
buckling. The ultimate flexural capacity of the steel section
is computed using the composite section modulus S
bc
. The
safety margin M
i
for bending capacity of steel girder G
i
is
M
i
= F
y
S
bc

mfg

_
M
DLNC,i
+ M
DLC,i
+ M
LL,i
_
(1)
where F
y
= yield strength of steel girder, S
bc
= composite
section modulus,
mfg
= uncertainty factor for model-
ing the flexural capacity of the steel girder (following a
lognormal distribution with mean and standard deviation
(; ) equal to (1.11;0.12), see Akgl and Frangopol 2004),
M
DLNC,i
, M
DLC,i
, and M
LL,i
= moments due to non-
composite and composite dead loads (Akgl and Frangopol
2004), and to live load in section S
2
of girder i , respectively.
It is assumed herein that corrosion is the only degradation
mechanism (deflection, cracking, and delamination of the
concrete deck are excluded) since the formulation presented
in this paper focuses on steel girders only. Corrosion reduces
the original thickness of the webs and flanges of steel gird-
ers as described in Akgl and Frangopol (2004) and affects
the section modulus S
bc
in (1). A power function for the cor-
rosion model is used (Townsend and Zoccola 1982; McCuen
and Albrecht 1995) p = b
0
t
b
1
where b
0
and p = corro-
sion losses after one and t years, respectively, and b
1
is the
slope of the logarithmic transformation of p. A ruralurban
environment is considered for the Wisconsin Bridge. Param-
eters b
0
and b
1
follow a lognormal distribution with mean
and standard deviation (; ) equal to (32.07 m; 2.89 m)
and (0.5; 0.045), respectively (Akgl and Frangopol 2004).
M
LL,i
= D
i
f
_
M
S
2
la,i
+ I
f
M
S
2
trk,i
_
, D
i
f
= distribution factor
in girder i (calculated for interior and exterior girders using
controlled load tests information, see Orcesi and Frangopol
Probability-based multiple-criteria optimization of bridge maintenance 139
Fig. 1 Wisconsin Bridge:
a Top view of spans 1 and 2,
and b plan view of sensors at
detail 1 in (a) (adapted from
Mahmoud et al. 2005)
(a)
G1
G2
G3
G4
SPAN 1 SPAN 2
OF PIER
C
L OF PIER
C
L
P.P.7
DETAIL 1
TOP VIEW
S1
S2
Travel direction
25.0 m 8.40 m 8.40 m 34.23 m
(b)
DETAIL 1 (PLAN VIEW)
38.1 cm
CH 3 OF GIRDER G4
CH 4 OF GIRDER G3
CH 5 OF GIRDER G2
CH 6 OF GIRDER G1
STRAIN GAGES
45.7 cm 7.6 cm
(2010a) for further details), M
S
2
la,i
, M
S
2
trk,i
= moment M
la,i
(90% of the moment associated with a design lane load)
and moment M
trk,i
(90% of the moment associated with two
design trucks) in cross-section S
2
, respectively, according to
AASHTO specifications (AASHTO 2007), I
f
= 1 + IM
and IM = dynamic load allowance fixed at 33%.
2.2 Performance margin
Qualitative inspection results can be defined as information
on the detection or the non detection of an event related to
a particular phenomenon (Thoft-Christensen and Srensen
1987; Cremona and Lukic 1998). Each inspection result
is an event, associated with a probability of occurrence
(e.g., the probability to detect crack sizes larger than an
acceptable threshold). In this paper, the performance mar-
gin, introduced for each component at each decision time,
is not based on visual inspection result but on structural per-
formance estimated from available monitoring information.
This allows taking into account in situ traffic informa-
tion, and more particularly heavy trucks traffic information
demand on the bridge. It is used to quantify the probability
of making a particular decision (rehabilitation is considered
herein) at a decision time (i.e., after an in-depth inspec-
tion or at the end of a SHM program). It is noted that the
deterioration process (corrosion herein) affects the perfor-
mance margin at each decision time. The methodology
to express the performance function H
i
for rehabilitation
of component i is explained in this section by using the
Wisconsin Bridge. The performance margin is associated
with the permanent deformation limit state function intro-
duced by Akgl and Frangopol (2004). This limit state
is formulated for a steel girder section using the bend-
ing moment in this section. Both dead loads according to
AASHTO specifications (AASHTO 2007) and live loads
(directly assessed by SHM results) are considered. Long-
term strain monitoring during a period of days allows
predicting the extreme moment histograms due to heavy
trucks in girders G1 (CH6), G2 (CH5), G3 (CH4), and G4
(CH3) in cross-section S
1
(see Fig. 1). Indeed, the elastic
moment in section S
1
for girder i (i = G1, G2, G3, or G4)
is
b,i
I
1,i
/y
bg,i
, where
b,i
= stress at the bottom flange
of girder i (converted from strain values by using Hookes
law), I
1,i
= moment of inertia of the composite section S
1
in girder i , and y
bg,i
= distance between elastic neutral
axis and the bottom of the flange of girder i (Orcesi and
Frangopol 2010a). The histograms of the extreme moment
are assessed in cross-section S
2
by multiplying the moment
in cross-section S
1
by the ratio between the maximum abso-
lute values of influence lines for moment (by modeling the
bridge with finite elements) in cross-sections S
2
and S
1
(Orcesi and Frangopol 2010a). It is noted that only the worst
case of loading (one truck moving through the bridge on
the right or on the left lane) is considered for each girder
(i.e., when trucks are on the left lane for girders G1 and G2
and when they are on the right lane for girders G3 and G4).
The best fitting values of the Gumbel distribution for each
moment histogram are used. The Gumbel probability distri-
bution (Gumbel 1958) associated with the SHM period of
days is defined as
F
_
M
i

_
= exp
_
exp
_

M
i


i
()
()
__
(2)
140 A.D. Orcesi, D.M. Frangopol
where F =Gumbel cumulative distribution function (CDF);
M
i

= extreme value of random variable M


i
in girder i dur-
ing (i.e., the maximum value of M
i,k
(k = 1, ..., p())
during );
i
(),
i
() = location and scale parameters for
girder i to be determined from the measured data.
The future extreme moment values M
i

(T), in the
next T years for girder i , associated with can be
predicted by rearranging (2) as (Ang and Tang 1984)
M
i

(T) =
i
()
i
() ln(ln(F(M
i

(T)))). Since
M
i
X
(T) = maximum value among n
T
future observa-
tions, F(M
i

(T)) = 1 1/n
T
where n
T
= total num-
ber of the passage of the heavy vehicles in the next T
years. Thus, M
i

(T) =
i
()
i
() ln(ln(1 1/n
T
)).
This expression of M
i

(T) enables to determine the


ratio function (Orcesi and Frangopol 2010a)
i

(T) =
max(M
i

(T)/max(M
i,1
, M
i,2
, ..., M
i, p()
), 1) which may
be used to predict future moment values M
i,
traf f ic
(T) through
the probability density function (PDF) of the extreme values
M
i

(denoted f (M
i

) herein). Hence, M
i,
traf f ic
(T) is defined
as M
i,
traf f ic
(T) =
i

(T) f (M
i

).
Finally, the expression of the performance margin at time
T, associated with monitoring duration , is formulated
for each girder i by introducing the moment M
i,
traf f ic
(T) as
follows
H
i,
(T) = F
y
S
bc

_
M
DLNC,i
+ M
DLC,i
+ M
i,
traf f ic
(T)
_
(3)
where = decision parameter associated with maximum
allowable permanent deformation under overload, before
deciding of a rehabilitation of the girder (see (1) for other
notations). In other words, the moment caused by dead and
live loads shall not exceed the moment at first yielding in
the section multiplied by . This parameter controls the
probability to make the decision of rehabilitation for a com-
ponent and expresses the threshold that stakeholders allow
to reach before deciding of essential maintenance actions.
To take into account uncertainties in the decision process,
is assumed to follow a lognormal distribution with mean
and a coefficient of variation (COV) equal to 10%. is
considered as a decision variable herein. It is noted that the
deterioration process due to corrosion used in (1) is also
considered in (3).
As indicated previously, the performance is assessed
at each decision time and the information on the traffic
extremes values is assumed to be affected only by the ini-
tial monitoring program. It is noted that 716 events were
captured during , of which 506 heavy vehicles crossed the
bridge on the right lane, and 210 heavy vehicles crossed the
bridge on the left lane. As previously mentioned,
1
and
2
are the 42 first and last days of = 84 days, respectively.
Figure 2a shows the moment values, both for
1
and
2
in
cross-section S
2
of girder G4, associated with heavy trucks
crossing the bridge on the right lane. Figure 2b shows the
maximum moment value per day of monitoring. Maximum
values per day, as illustrated in Fig. 2b for girder G4 are used
in the following as information on the traffic extreme values
for girders G1, G2, G3, and G4. To quantify the impact of
monitoring duration on location and scale parameters, the
coefficients

1,Gi,
l
(
l
) =

i
(
l
)
i
(
l
)

i
(
l
)

2,Gi,
l
(
l
) =

i
(
l
)
i
(
l
)

i
(
l
)

i = 1,2,3,4 and l = 1,2


(4)
(a)
0 90 180 270 360 450 506
200
300
400
500
600
700
800
900
1000
1100
1200
Number of heavy truck occurrence
M
o
m
e
n
t

(
k
N
.
m
)
Heavy trucks on the right lane
Girder G4
Monitoring period
1
(42 first days)
Girder G4
Monitoring period
2
(42 last days)
(b)
0 10 20 30 40 50 60 69
200
300
400
500
600
700
800
900
1000
1100
1200
Number of heavy truck occurrence
M
a
x
i
m
u
m

m
o
m
e
n
t

v
a
l
u
e

p
e
r

d
a
y

(
k
N
.
m
)
Heavy trucks on the right lane
Girder G4
Monitoring period
1
(42 first days)
Girder G4
Monitoring period
2
(42 last days)
Fig. 2 Moment values in section S
2
of girder G4 associated with heavy
truck occurrences for
1
and
2
when considering a all heavy trucks,
and b the maximum moment value per day
Probability-based multiple-criteria optimization of bridge maintenance 141
Fig. 3 Effect of monitoring
duration on
1,Gi,
l
(l = 1, 2)
for girders a G1, b G2, c G3,
and d G4
(a) (b)
0 6 12 18 24 30 36 42
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Monitoring duration (days)

1
,
G
1
,

l
l = 1
l = 2
Heavy trucks on the left lane
0 6 12 18 24 30 36 42
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Monitoring duration (days)

1
,
G
2
,

l
l = 1
l = 2
Heavy trucks on the left lane
(c) (d)
0 6 12 18 24 30 36 42
0
0.02
0.04
0.06
0.08
0.1
Monitoring duration (days)

1
,
G
3
,

l
l = 1
l = 2
Heavy trucks on the right lane
0 6 12 18 24 30 36 42
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
Monitoring duration (days)

1
,
G
4
,

l
l = 1
l = 2
Heavy trucks on the right lane
are introduced, where
l
= first days of
l
,
i
(
l
) and

i
(
l
) = location and scale parameters associated with
monitoring duration
l
, and
i
(
l
) and
i
(
l
) = location
and scale parameters associated with maximum available
monitoring duration
l
. The profiles of
1,Gi,
l
with mon-
itoring duration
l
(l = 1, 2) are shown in Fig. 3ad for
girders G1, G2, G3, and G4, respectively. Similarly, the
profiles of
2,Gi,
l
with monitoring duration
l
(l = 1, 2)
are shown in Fig. 4ad for girders G1, G2, G3, and G4,
respectively. It is noted that heavy trucks considered in
Fig. 4 Effect of monitoring
duration on
2,Gi,
l
(l = 1, 2)
for girders a G1, b G2, c G3,
and d G4
(a) (b)
0 6 12 18 24 30 36 42
0
0.2
0.4
0.6
0.8
1
Monitoring duration (days)

2
,
G
1
,

l
l = 1
l = 2
Heavy trucks on the left lane
0 6 12 18 24 30 36 42
0
0.2
0.4
0.6
0.8
1
Monitoring duration (days)

2
,
G
2
,

l
l = 1
l = 2
Heavy trucks on the left lane
(c) (d)
0 6 12 18 24 30 36 42
0
0.2
0.4
0.6
0.8
1
Monitoring duration (days)

2
,
G
3
,

l
l = 1
l = 2
Heavy trucks on the right lane
0 6 12 18 24 30 36 42
0
0.2
0.4
0.6
0.8
1
Monitoring duration (days)

2
,
G
4
,

l
l = 1
l = 2
Heavy trucks on the right lane
142 A.D. Orcesi, D.M. Frangopol
Figs. 3 and 4 are always on the left lane for girders G1
and G2 and on the right lane for girders G3 and G4. It
is observed that monitoring occurrence and duration affect
these parameters in different ways for components of the
Wisconsin Bridge.
2.3 Event tree approach
Since maintenance decisions depend on assessment strate-
gies, the optimal maintenance decision times should be
determined by taking into consideration all the possible
outcomes of the performance assessment. The concept
of event tree for maintenance decision is schematically
illustrated with the example of Fig. 5 for component i
by considering the two options Do nothing (DN) and
rehabilitation (R).
2.3.1 Probability of failure
The probability of failure of a component i at time t is a
function of the past history of this component. For t t
i,1
,
the probability of failure only depends on the value of the
safety margin M
i
(t ) (defined in (1)) at this time t since there
is no previous event. Therefore, the probability of failure
P
f
i
is
P
f
i
(t ) = P(F) t t
i,1
(5)
where F = {M
i
(t ) 0}. For t
i,1
< t t
i,2
, the probability
of failure of component i ,
P
f
i
(t ) = P(F|h
i,1
)P(h
i,1
) + P(F|

h
i,1
)P(

h
i,1
)
t
i,1
< t t
i,2
(6)
depends on all the outcomes at the decision time t
i,1
, where
h
i,1
= {H
i,
l
(t
i,1
) 0},

h
i,1
= {H
i,
l
(t
i,1
) > 0}, and
Decision time
time
Component i
f
T
0
T ,1 i
t
,4 i
t
DN
R
,3 i
t
DN
DN
R
R
DN
R
DN
R
DN
DN
R
R
,2 i
t
DN = Do Nothing
R = Rehabilitation
Fig. 5 Performance-based event tree decisions
H
i,
l
(t
i,1
) = performance margin (see (3)) associated with
available monitoring days
l
. The following convention is
chosen: the rehabilitation of the component, denoted as
(R), is decided if H
i,
l
(t
i,1
) 0 whereas the action do
nothing, denoted (DN), is decided if H
i,
l
(t
i,1
) > 0.
{F|h
i,1
} = failure associated with safety margin at time t
knowing that rehabilitation of component i was performed
at time t
i,1
. Conversely, {F|

h
i,1
} = safety margin at time t
knowing that the decision was to do nothing at time t
i,1
. It is
noted that a first-order reliability analysis is associated each
time with safety and performance margins.
More generally, the sequence of events is introduced
to take into account all past events for a component. For
example, the sequence of events = [R, DN] means that
rehabilitation was performed for component i at time t
i,1
and that nothing was done at time t
i,2
. By using these
notations, H

i,
l
(t
i,k
) is the kth performance margin at time
t
i,k
associated with available monitoring days
l
, knowing
that the sequence of events occurred in the past. With
these notations, the probability of failure, using the total
probability theorem, is
P
f
i
(t ) = P(F|h
i,1
h
i,2
)P(h
i,1
h
i,2
)
+ P(F|h
i,1

h
i,2
)P(h
i,1

h
i,2
)
+ P(F|

h
i,1
h
i,2
)P(

h
i,1
h
i,2
)
+ P(F|

h
i,1

h
i,2
)P(

h
i,1

h
i,2
) (7)
for t
i,2
t < t
i,3
, where notations are those of (6)
and where h
i,1
h
i,2
= {H
i,
l
(t
i,1
) 0}
_
H
R
i,
l
(t
i,2
) 0
_
;
h
i,1

h
i,2
= {H
i,
l
(t
i,1
) 0}
_
H
R
i,
l
(t
i,2
) > 0
_
;

h
i,1
h
i,2
=
{H
i,
l
(t
i,1
) > 0}
_
H
DN
i,
l
(t
i,2
) 0
_
; and

h
i,1

h
i,2
=
{H
i,
l
(t
i,1
) > 0}
_
H
DN
i,
l
(t
i,2
) > 0
_
. The probabilities
P(h
i,1
h
i,2
), P(h
i,1

h
i,2
), P(

h
i,1
h
i,2
), and P(

h
i,1

h
i,2
) in (7)
are those of a system in parallel. For m limit states of a sys-
tem with reliability indices
i
(i = 1, 2, ..., m) and matrix
(m m) of correlation among limit states, the system
probability of failure is approximated as
m
(, ) for a
system in parallel, where
m
is the multivariate normal dis-
tribution. The improved version I-PCM, proposed by Yuan
and Pandey (2005) to calculate the product of conditional
marginal (PCM), is used herein to compute the integration
associated with
m
. Finally, by applying this methodology
to the overall decision planning, the probability of failure
can be determined for component i at every time t until time
horizon T
f
is reached.
2.3.2 Probability of maintenance intervention
The probability of maintenance intervention of component
i at decision time t
i,k
(Fig. 5) is associated with the perfor-
mance margin of (3) at time t
i,k
but also with all previous
Probability-based multiple-criteria optimization of bridge maintenance 143
maintenance action decisions. For example, the probability
of maintenance intervention at time t
i,1
is
P
r
i
(t
i,1
) = P(h
i,1
) (8)
and is
P
r
i
(t
i,2
) = P(h
i,1
h
i,2
) + P(

h
i,1
h
i,2
) (9)
at time t
i,2
, where notations are those used in (6) and (7).
It is noted that the maintenance intervention (i.e., rehabili-
tation) considered herein for each girder is its replacement.
Therefore, if the rehabilitation occurs at t = T, the compo-
nent section modulus in (1) and (3) is restored to its as new
condition value and the corrosion process restarts with the
same corrosion model p = b
0
t
b
1
0
with t
0
= t T.
3 Optimization framework
The main objective of the optimization procedure is to
determine optimal maintenance decision times. First, the
objective functions considered in the optimization process
are introduced and their formulation is provided. Second,
the optimization problem formulation is introduced and
discussed.
3.1 Objective functions
3.1.1 Expected error in the decision process
SHM duration is generally constrained due to limited finan-
cial resources. Answering to the questions howlong should
a bridge be monitored? to effectively capture the structural
demand of traffic and which is the impact of monitoring
occurrence on the maintenance optimization results? is an
important issue. The approach proposed in this section is
devoted to answer to these questions by the introduction of
an expected error in the decision process. The expected error
associated with a decision is defined as the probability to
make this decision multiplied by the associated error. Know-
ing the extreme events for the overall monitoring duration

l
(l = 1 or 2) of the Wisconsin Bridge, the change in the
performance margin, when only a fraction of this monitor-
ing period is available, enables to define an error for each
decision of the event tree (Fig. 5). Hence, the expected error
at a decision time t
i,k
is defined as the sum of all expected
errors associated with all decisions at t
i,k
in the event tree.
For instance, the expected error at time t
i,1
is defined as
E[
i
(t
i,1
)] = P(h
i,1
)
i
(h
i,1
) + P(

h
i,1
)
i
(

h
i,1
) (10)
where
i
(h
i,1
) = error associated with the decision
{H
i,
l
(t
i,1
) 0}. This error is defined as

i
(h
i,1
) =

P(H
i,
l
(t
i,1
) 0) P(H
i,
l
(t
i,1
) 0)
P(H
i,
l
(t
i,1
) 0)

(11)
where P(H
i,
l
(t
i,1
) 0) = probability of making the deci-
sion associated with H
i,
l
(t
i,1
) 0 if the information with
the maximum monitoring duration
l
was available. The
same approach than in (11) enables to define
i
(

h
i,1
) and
then the expected error E[
i
(t
i,1
)] in (10). In a similar way,
the error at decision time t
i,2
is
E[
i
(t
i,2
)] = P(h
i,1
h
i,2
)
i
(h
i,1
h
i,2
)
+ P(

h
i,1
h
i,2
)
i
(

h
i,1
h
i,2
)
+ P(h
i,1

h
i,2
)
i
(h
i,1

h
i,2
)
+ P(

h
i,1

h
i,2
)
i
(

h
i,1

h
i,2
) (12)
where
i
(h
i,1
h
i,2
) = error associated with decision
_
H
R
i,
l
(t
i,2
) 0
_
and expressed as

i
(h
i,1
h
i,2
) =

P
_
H
R
i,
l
(t
i,2
) 0
_
P
_
H
R
i,
l
(t
i,2
) 0
_
P
_
H
R
i,
l
(t
i,2
) 0
_

(13)
where P
_
H
R
i,
l
(t
i,2
) 0
_
= probability to make the deci-
sion associated with H
R
i,
l
(t
i,2
) 0 if the maximum moni-
toring duration
l
is available. The same principles and
notations are used to express errors associated with other
decisions in (12). Finally, the expected error for all compo-
nents and for all decision times is
E[] =
N
c

i =1
N
d,i

j =1
E[
i
(t
i, j
)] (14)
where N
c
= number of components, N
d,i
= number of deci-
sion times for component i , and E[
i
(t
i, j
)] = expected error
of decision at the jth decision time t
i, j
of component i (see
(10) and (12)). It is considered in the following that deci-
sion maintenance times are the same for all components of
the bridge.
3.1.2 Expected management cost
The monitoring cost depends herein on the monitoring pro-
gram duration. It is composed of fixed costs due to the
preparation and analysis of monitoring (e.g., installation and
wiring of sensors), and of variable costs that depend on the
monitoring duration. The value of the monitoring cost is
C
mon
= C
mon,ref

l
d
0
+ C
f
mon
(15)
144 A.D. Orcesi, D.M. Frangopol
where C
mon,ref
= reference cost of monitoring the structure
during d
0
days,
l
= monitoring program duration (in days),
and C
f
mon
= fixed costs for each monitoring program.
The present value of the maintenance cost is
E[C
r
] =
N
c

i =1
N
d,i

j =1
c
r,i
P
r,i
(t
i, j
)
(1 + )
t
i, j
(16)
where P
r,i
(t
i, j
) = probability of repair at the jth decision
time t
i, j
of component i , c
r,i
= maintenance cost for com-
ponent i , and = yearly discount rate of money (see (8),
(9) and (14) for other notations). Finally, the management
cost is defined as
E[C
m
] = C
mon
+ E[C
r
] (17)
3.1.3 Expected failure cost
The present value of the expected failure cost is
E[C
f
] =
n
0

k=1
c
f,s
P
f,s
(
0,k
)
(1 + )

0,k
(18)
where c
f,s
= failure cost of the system, n
0
= total number
of decision times,
0,k
= kth decision time for one/several
component(s) of the bridge, P
f,s
(
0,k
) = probability of fail-
ure of the system at the kth decision time
0,k
. P
f,s
is
assessed by considering the probabilities of failure P
f
i
of
components i = 1,..., N
c
at
0,k
(see (5), (6) and (7) for
expressions of P
f
i
) in a system analysis.
3.2 Optimization problem formulation
The design variables in this problem are: the interval
d
(in
years) between two maintenance decision times; the mon-
itoring duration l (in days) considered in the optimization
framework, and the decision criterion , mean of the ran-
dom variable introduced in (3). It is noted that time
0,k
in (18) is determined by the interval
d
and the end of the
service life T
f
(fixed at 75 years herein).
The following parameter
K
i
(T
f
) =
1
(P
f
i,0
(T
f
) P
f
i
(T
f
)) (19)
is used as an efficiency indicator of the event tree strategy,
where P
f
i,0
(T
f
), and P
f
i
(T
f
) = probabilities of failure at
T
f
for component i if no decision is made for this com-
ponent until the end of the time horizon (i.e., with never
making any rehabilitation), and if the event tree procedure
described in (5), (6), and (7) is applied. Obviously, the less
K
i
(T
f
) the larger the difference P
f
i,0
(T
f
) P
f
i
(T
f
) is and,
consequently, the more efficient the event tree strategy is.
Therefore, the purpose is to have the parameter K
i
(T
f
) the
smallest possible.
The optimal solution (20) is searched by simultaneously
minimizing the expected management cost E[C
m
] (21), the
expected failure cost (22), and the expected error of deci-
sion (23), such that constraints (24) are satisfied. The
optimization problem is formulated as
Find
d
, l, and (20)
to Minimize E[C
m
] (21)
and Minimize E[C
f
] (22)
and Minimize E[] (23)
such that K
i
(T
f
) < K
0
i = G1, G2, G3, and G4 (24)
where K
0
= threshold associated with K
i
(T
f
). As men-
tioned previously, the purpose is to have the parameter
K
i
(T
f
) the smallest possible or, as it is the case herein,
to ensure that it is below a threshold K
0
fixed in advance.
It is noted that other optimization frameworks can be con-
sidered. For instance, the goal can be to search
d
, l,
and to simultaneously minimize K
i
(T
f
) and E[] such
that E[C
m
] + E[C
f
] stays below a fixed budget. However,
only the framework introduced in (20)(24) is considered
herein. The general framework for assessment of E[C
m
],
E[C
f
] and E[], for a particular solution
d
, l, and is
shown in Fig. 6. It is noted that an event tree is built for each
component. NSGA-II (Non-Dominated Sorting Genetic
d
, l ,
[ ]
m
E C ,
f
E C , [ ]
c
i N ? =
Yes
No
Event tree for component i :
Assessment of
i
f
P ,
i
r
P , ( )
i
at each of
the
i
d
N decision time(s) by using the
event tree based-approach
1 i i = +
Determination of the
probability of failure
f ,s
P
i = 1
Fig. 6 General framework for determination of E[C
m
], E[C
f
] and
E[], for a particular solution
d
, l, and
Probability-based multiple-criteria optimization of bridge maintenance 145
Algorithm) program developed by Deb et al. (2002) is used
to determine optimal solutions set of this multi-objective
optimization problem.
4 Results
In this section, the reference cost of monitoring C
mon,ref
(see (15)) is defined as $1,500 for remote monitoring dur-
ing 7 days and the fixed monitoring cost C
f
mon
is defined
as $8,500 for general preparation, data acquisition and
other activities (adopted from Connor and Fisher 2002). In
(16), the essential maintenance action considered herein is
replacement of a girder and the associated cost is fixed at
$114,600 (adopted from Estes 1997). In (18), the cost of
failure of the system is assumed to be $10
6
. All girders
G1, G2, G3, and G4 are assumed to be a series system
(considering herein that failure of any girder causes the
bridge failure). The probability of failure of the system
P
f,s
(
0,k
) at
0,k
is assumed herein to be the mean of the
Ditlevsens bounds (Ditlevsen 1979) for the system of gird-
ers G1, G2, G3, and G4 considered in series (lower and
upper Ditlevsens bounds are close herein). It is noted that a
similar analysis was lead by directly considering Ditlevsens
upper bound and that similar results were obtained. Finally,
the yearly discount rate of money is fixed at 4%. The
resulting three dimensional Pareto-optimal sets of the opti-
mization problem (20)(24), when considering K
0
= 6.5
for monitoring periods
1
and
2
, are shown in Figs. 7, 8
and 9 when
d
= 8, 9, and 10 years, respectively. Table 1
details some solutions of the Pareto front shown in Fig. 7.
Table 1 shows that an increase of the monitoring dura-
tion tends to decrease the expected error E[]. Indeed, it
is clearly shown that the expected error decreases with an
increase of monitoring duration both for monitoring periods
1
1.5
2
2.5 27.4
27.6
27.8
28
28.2
0
2
4
6
8
10
E[C
f
] ($)
E[C
m
] ($10
4
)
E
[

]
Monitoring period
1
Monitoring period
2
=4%

d
=8 years
A
Fig. 7 Three dimensional Pareto-optimal set of the optimization
problem when K
0
= 6.5 and
d
= 8 years
1
1.5
2 28.4
28.45
28.5
28.55
0
2
4
6
8
10
E[C
f
] ($)
E[C
m
] ($10
4
)
E
[

]
Monitoring period
Monitoring period

2
=4%

d
=9 years
B
Fig. 8 Three dimensional Pareto-optimal set of the optimization
problem when K
0
= 6.5 and
d
= 9 years

1
and
2
. In addition, the parameter greatly impacts
the expected error (compare solutions {8, 7, 59} and {8,
7, 73} in Table 1). Indeed, this parameter might change
the probability of maintenance, and in turn, the error in the
decision process (see (11) and (13)). Also, it is shown that
the expected failure cost E[C
f
] increases with
d
(compare
Pareto solutions in Figs. 7, 8, and 9 when
d
= 8, 9, and
10 years, respectively). Finally, it is noted that a decrease in
parameter results in lower expected failure cost and larger
expected management cost (Figs. 7, 8, and 9). To further
illustrate optimization results, three solutions A, B, C, are
selected in Figs. 7, 8, and 9, respectively. These solutions
are A = {8, 36, 52}, B = {9, 36, 63}, and C = {10, 36, 73}.
It is noted that 36 days of monitoring are considered for
these three solutions and that 8 years (solution A), 9 years
(solution B), and 10 years (solution C) are considered as
1
1.2
1.4
1.6
1.8 29
29.05
29.1
29.15
29.2
0
2
4
6
8
E[C
f
] ($) E[C
m
] ($10
4
)
E
[

]
Monitoring period
1
Monitoring period
2
=4%

d
=10 years
C
Fig. 9 Three dimensional Pareto-optimal set of the optimization
problem when K
0
= 6.5 and
d
= 10 years
146 A.D. Orcesi, D.M. Frangopol
Table 1 Comparison of Pareto solutions (see Fig. 7) when = 8 and
monitoring sets
1
and
2
are considered, respectively
Solutions {
d
, l, } E[C
m
] ($10
4
) E[C
f
] ($) E[]
Monitoring period
1
{8, 7, 59} 1.0554 27.8564 5.6494
{8, 7, 73} 1.0007 27.8844 7.7419
{8, 12, 50} 2.0040 27.4999 3.4726
{8, 30, 53} 1.8001 27.7383 0.8683
{8, 33, 57} 1.6428 27.8411 0.3876
{8, 36, 52} 2.0151 27.7013 0.1102
Monitoring period
2
{8, 7, 59} 1.0510 27.8580 6.3183
{8, 7, 73} 1.0006 27.8844 8.7912
{8, 12, 50} 1.9603 27.5122 4.9752
{8, 30, 53} 1.7952 27.7399 3.0813
{8, 33, 57} 1.6400 27.8424 2.6206
{8, 36, 52} 1.9816 27.7162 0.5983
interval time between each decision maintenance. Consid-
ering that the end of the service life T
f
= 75 years, the end
of the time horizon is 32 years after 2004 (when the bridge
was 43 years old). Hence, there will be maintenance deci-
sions at years 1, 9, 17 and 25 after year 2004, for solutions
A, at years 1, 10, 19, and 28 after year 2004 for solution B,
and at years 1, 11, 21, and 31 after year 2004 for solution
C. The cumulative expected management cost, failure cost,
and error profiles are provided in Fig. 10ac, respectively
(the value at T
f
is E[C
m
], E[C
f
] and E[], respectively). It
is shown that solution C is the less expensive (Fig. 10a) but
is associated with the highest cost of failure at T
f
(E[C
f
] =
$27.71, $28.53, and $29.17 for solutions A, B, and C,
respectively) (Fig. 10b). The significant difference in man-
agement cost between solutions A and B (see Fig. 10a) is
mainly due to the parameter that dramatically influences
the decision of applying or not some essential maintenance
actions. It is shown in Fig. 10c that the expected error is the
lowest for solution A and the largest for solution C, which
can be explained by the value of the parameter . Indeed,
as previously mentioned, this parameter changes the prob-
ability of maintenance, and, consequently, the error in the
decision process. Finally, solution B offers a good com-
promise between expected management cost that remains
affordable (Fig. 10a), expected failure cost (Fig. 10b), and
the error of decision (Fig. 10c).
Through these examples, it is shown that the event tree
based approach allows considering all the possible out-
comes when SHM information is considered. According to
allowable serviceability and safety thresholds, this frame-
work sheds light on the link between SHM strategies and
their use in optimization of maintenance decisions.
(a)
5 10 15 20 25 30 1 32
1.6
1.65
1.7
1.75
1.8
1.85
1.9
1.95
2
Time after year 2004 (years)
E
[
C
m
]

(
$
1
0
4
)
T
f
=75 years
=4%
A
B
C
(b)
5 10 15 20 25 30 1 32
3
10
17
24
30
Time after year 2004 (years)
E
[
C
f
]

(
$
)
=4% T
f
=75 years
B
C
A
(c)
5 10 15 20 25 30 1 32
0
0.2
0.4
0.6
0.8
1
Time after year 2004 (years)
E
[

]
T
f
=75 years
B
C
A
Fig. 10 Profiles of cumulative expected a management cost, b failure
cost, and c error, for solutions A, B, and C
5 Conclusions
In this paper, a probability-based approach is proposed
for multi-criteria optimization of bridge assessment (SHM
programs) strategies and thresholds for essential mainte-
nance decisions under uncertainty. The objectives are to
Probability-based multiple-criteria optimization of bridge maintenance 147
demonstrate that long-term monitoring information can be
integrated in a life-cycle cost analysis and to analyze how
monitoring occurrence and duration affect analysis results.
The Wisconsin Bridge SHM program is used as example
to illustrate the proposed concepts and methodologies. The
following conclusions can be drawn from this paper.
1. SHM information is used to determine a performance
margin at each decision time. Statistics of extremes
are introduced to determine future bending moment
values due to heavy trucks. The monitoring duration
affects parameters assessed with statistics of extremes
in a different way for various structural components.
The performance margin is then used to quantify the
probability of making a maintenance decision at each
decision time.
2. The event-tree procedure enables the consideration
of all possible future outcomes. These outcomes are
affected by the initial SHM program. A measure for
expected error in the decision process is proposed. The
SHM and maintenance costs and the expected fail-
ure cost are also determined and considered in the
optimization process.
3. By using statistics of extremes, it is possible to use
available SHM information at each decision time. By
introducing the error of the decision process in the opti-
mization procedure, it is possible, as demonstrated in
example of the Wisconsin Bridge, to quantify how opti-
mal solutions are affected by monitoring occurrence
and duration. Of course, this analysis of the life-cycle
accuracy is only possible for the 84 days of monitoring
for this bridge. However, it shows how the life-cycle
analysis would be affected by the quality and quantity
of monitoring information. The approach proposed in
this paper provides tools for better answering the ques-
tions how long should be monitored a structure? and
which is the impact of monitoring occurrence on the
maintenance optimization results?
4. Future studies should account for additional degra-
dation mechanisms and maintenance decisions in the
event tree. Indeed, the formulations proposed in this
paper focus on steel girders by considering only es-
sential maintenance actions; preventive maintenance
actions should be also considered. Also, only strain data
are used in the proposed framework. Different types of
SHM information should be used to extend the use of
this framework. The final objective is to explore, under
uncertainty and limited financial resources, the use of
SHM information in life-cycle analysis and to opti-
mize its use as well as maintenance and rehabilitation
strategies. Use of survivor functions of structural com-
ponents and their updating with additional information
provided by SHM, along the lines proposed in Orcesi
and Frangopol (2010b, 2011) and Frangopol (2010),
have also to be further investigated.
Acknowledgments The support from (a) the National Science Foun-
dation through grant CMS-0639428, (b) the Commonwealth of
Pennsylvania, Department of Community and Economic Develop-
ment, through the Pennsylvania Infrastructure Technology Alliance
(PITA), (c) the U.S. Federal Highway Administration Cooperative
Agreement Award DTFH61-07-H-00040, and (d) the U.S. Office
of Naval Research Contract Number N-00014-08-0188 is gratefully
acknowledged. The writers thank Victor Voiry for computational help.
The opinions and conclusions presented in this paper are those of
the writers and do not necessarily reflect the views of the sponsoring
organizations.
References
AASHTO (2007) American Association of State Highway and Trans-
portation Officials (AASHTO) LRFD bridge design specifi-
cations, 4th edn. AASHTO, Washington, DC
Akgl F, Frangopol DM (2004) Lifetime performance analysis of
existing steel girder bridge superstructures. J Struct Eng ASCE
130(12):18751888
Ang AH-S, Tang HW (1984) Probability concepts in engineering
planning and design, vol II. Wiley, New York
Connor RJ, Fisher JW (2002) Report on field inspection, assess-
ment, and analysis of floorbeam connection cracking on the
Birmingham Bridge Pittsburgh PA. ATLSS Report No. 02-10.
Lehigh University, Bethlehem
Cremona CF, Lukic M (1998) Probability-based assessment and main-
tenance of welded joints damaged by fatigue. Nucl Eng Des
182(3):253266
Deb K, Pratap A, Agarwal S, Meyarivan T (2002) A fast and eli-
tist multiobjective genetic algorithm: NSGA-II. IEEE Trans Evol
Comput 6(2):182197
Ditlevsen O (1979) Narrow reliability bounds for structural systems.
J Struct Mech 7(4):453472
Estes AC(1997) Asystemreliability approach to the lifetime optimiza-
tion of inspection and repair of highway bridges. Ph.D. thesis,
University of Colorado, Boulder
Estes AC, Frangopol DM (2001) Minimum expected cost-oriented
optimal maintenance planning for deteriorating structures: appli-
cations to concrete bridge deck. Reliab Eng Syst Saf 73(3):281
291
Frangopol DM (2010) Life-cycle performance, management, and
optimization of structural systems under uncertainty: accom-
plishments and challenges. Struct Infrastr Eng. doi:10.1080/
15732471003594427
Frangopol DM, Liu M (2007) Maintenance and management of civil
infrastructure based on condition, safety, optimization, and life-
cycle cost. Struct Infrastr Eng 3(1):2941
Frangopol DM, Messervey TB (2009) Maintenance principles for civil
structures. In: Boller C, Chang F-K, Fujino Y (eds) Encyclo-
pedia of structural health monitoring, chapter 89, vol 4. Wiley,
Chichester, pp 15331562
Frangopol DM, Strauss A, KimS (2008a) Bridge reliability assessment
based on monitoring. J Bridge Eng ASCE 13(3):258270
Frangopol DM, Strauss A, Kim S (2008b) Use of monitoring ex-
treme data for the performance prediction of structures: general
approach. Eng Struct 30(12):36443653
Gumbel EJ (1958) Statistics of extremes. Columbia University Press,
New York
148 A.D. Orcesi, D.M. Frangopol
Hess PE (2007) Structural health monitoring for high-speed naval
ships. In: Chang F-K (ed) Proceedings of the 6th international
workshop on structural health monitoring. DEStech Publications
(keynote paper)
Hua XG (2006) Structural health monitoring and condition assess-
ment of bridge structures. Department of Civil and Structural
Engineering, Hong Kong Polytechnic University, Hong Kong
Kim S, Frangopol DM (2010) Inspection and monitoring planning
for RC structures based on minimization of expected damage
detection delay. Probab Eng Mech. doi:10.1016/j.probengmech.
2010.08.009
Klinzmann C, Schnetgke R, Hosser D (2006) A framework for
reliability-based system assessment based on structural health
monitoring. In: Proceedings of the 3rd European workshop on
structural health monitoring. Granada, Spain
Liu M, Frangopol DM, KimS (2009) Bridge safety evaluation based on
monitored live load effects. J Bridge Eng ASCE 14(4):257269
Mahmoud HN, Connor RJ, Bowman CA (2005) Results of the fatigue
evaluation and field monitoring of the I-39 Northbound Bridge
over the Wisconsin River. ATLSS Report No. 05-04. Lehigh
University, Bethlehem
McCuen RH, Albrecht P (1995) Composite modeling of atmospheric
corrosion penetration data. Application of accelerated corrosion
tests to service life prediction of materials. ASTM STP 1194,
Philadelphia, pp 65102
Ni YQ, Hua XG, Ko JM (2006a) Reliability-based assessment of
bridges using long-termmonitoring data. Key Eng Mater 217222
Ni YQ, Hua XG, Ko JM (2006b) Reliability-based assessment of
bridges using long-termmonitoring data. Key Eng Mater 321323
Orcesi AD, Cremona CF (2010) Optimal maintenance strategies for
bridge networks using the supply and demand approach. Struct
Infrastr Eng. doi:10.1080/15732470902978566
Orcesi AD, Frangopol DM (2010a) Inclusion of crawl tests and long-
term health monitoring in bridge serviceability analysis. J Bridge
Eng ASCE 15(3):312326
Orcesi AD, Frangopol DM (2010b) The use of lifetime functions
in the optimization of nondestructive inspection strategies for
bridges. J Struct Eng ASCE. doi:10.1061/(ASCE)ST.1943-541X.
0000304
Orcesi AD, Frangopol DM (2011) Optimization of bridge maintenance
strategies based on structural health monitoring information.
Struct Saf 33(1):2641
Orcesi AD, Frangopol DM, Kim S (2010) Optimization of bridge
maintenance strategies based on multiple limit states and moni-
toring. Eng Struct 32(3):627640
Strauss A, Frangopol DM, Kim S (2008) Use of monitoring extreme
data for the performance prediction of structures: Bayesian updat-
ing. Eng Struct 30(12):36543666
Thoft-Christensen P, Srensen J (1987) Optimal strategies for in-
spection and repair of structural systems. Civ Eng Syst 4:94
100
Townsend HE, Zoccola JC (1982) Eight-year atmospheric corrosion
performance of weathering steel in industrial, rural and marine
environments. Atmospheric corrosion of metals. ASTM STP 767,
Philadelphia, pp 4559
Yuan X-X, Pandey MD (2005) Analysis of approximations for multi-
normal integration in system reliability integration. Struct Saf
28:361377

You might also like