Icossar 09

ICOSSAR 09
13 September 17 September 2009,

Osaka, Japan
Colin C. Caprani & Eugene J. OBrien
Dublin Institute of Technology & University College Dublin
Estimating Extreme Highway Bridge Traffic Load Effects
Introduction
Estimating Extreme Highway Bridge Traffic Load Effect
C.C. Caprani & E.J. OBrien
Statistics in the hands of an engineer are like a lamppost to a
drunk they're used more for support than illumination.
- A. E. Housman.
Extreme Value Methods I
Block maxima approach data modelled using GEV distribution:
Time
Load Effect
Time
Load Effect
Block
Size
Time
Load Effect
Block
Size
Extreme Value Methods II
Peaks Over Threshold (POT) approach data modelled using GPD distribution:
Time
Load Effect
Time
Load Effect
T
h
r
e
s
h
o
l
d
Time
Load Effect
T
h
r
e
s
h
o
l
d
Extreme Value Methods III
Differences in the approaches:
Time
Load Effect
Extreme Value Methods IV
Choice between the approaches:
Gives different results
Often subjective
Difficult to assess best model
As the GEV and GPD are different distributions:
They are non-nested models
Cannot calculate a statistical significance of differences
Box-Cox GEV Distribution I
The BCGEV distribution:
Introduced by Bali (2003) for use in economic modelling
Includes both GEV and GPD distributions through a model parameter,
Maintains the usual GEV/GPD parameter set, ( , , ):
( ) ( )
{ }
1/
1
exp 1 1 H s h s
+

= +

( )
1
s
h s

=

where
Thus, as:
, BCGEV GEV distribution
, BCGEV GPD distribution (by LHopitals rule)
1
0
Benefit: GEV and GPD are now nested models and can be compared statistically.
Box-Cox GEV Distribution II
Application of the BCGEV model:
A high threshold is set - about 2 standard deviations above the mean of the
parent data
Data arranged sequentially: s
1
s
r
s
n
1 1
log log 1 1 log 1
1
r
s r
n

+ = +

+

Minimize the sum of the squares of the residuals (SSR),
2
Estimation of BCGEV:
Maximum likelihood estimation not robust, so
Non-linear regression estimation used:
Residual
Bridge Traffic Loading I
0
2
4
6
8
10
0 10 20 30 40
Distance
I
n
f
l
u
e
n
c
e

O
r
d
i
n
a
t
e
Effect 1
Effect 2
Using real traffic measured using Weigh-In-Motion
Traffic characteristics are statistically modelled
Monte Carlo simulation allows more traffic to be studied
Load effects are calculated using influence lines of interest
Bridge Traffic Loading II
3 load effects considered:
LE1 moment at B;
LE2 moment at E;
LE3 shear at A.
5 days of data from the A6 Paris-Lyon motorway is used as basis
A 1000-day Monte Carlo traffic sample is generated
Thus 1000 daily maximum static load effects
Consider 5 bridge lengths of 20, 30, 40, 50, 60 m
The optimal statistical extrapolation of this data set to determine lifetime load
effect is what is considered in this work.
Bridge Traffic Loading III
In bridge traffic loading, different loading event types occur:
1-truck
2-truck
3-truck
These loading events have different statistical distributions
Use a composite distribution of load effect (Caprani et al 2008):
1
( ) ( )
N
C i
i
G z G z
=
=
Composite
Distribution
Individual Event-type
Distribution
Basis of BCGEV Analysis
For 3 load effects, 5 bridge lengths and each loading event type,
There are 41 data sets to be modelled.
11 thresholds are applied to the daily maximum data:
In 0.5 standard deviation steps
From k = -2.5 to k = +2.5 standard deviations about the mean
BCGEV model:
Estimation of model parameter, is not robust
Thus varied from 0 to 1 in 0.01 steps
Best fit of remaining parameters then found for each .
Optimum BCGEV Parameters
0
0.5
1
-2
0
2
0
0.1
0.2
0.3
Threshold:
+k
Model Parameter,
M
e
a
n

S
S
R
The mean SSR of the 41 data sets for each and threshold are taken to give:
Thus best fit on average is:
Threshold, k = -1.5
Model Parameter, = 0.98
Also:
Best fit model parameter
always 0.9 < < 1.0
Thus GEV better than GPD
for bridge traffic loading?
Likelihood Ratio Test I
Using the LR test which applies to nested models:
Determine of GEV or GPD (or neither) better represents the data
Calculate the statistical significance of the representation
Calculate:
Standard Error of Regression (SER) the mean error per data point: SSR/n
The LR statistic then is:
( )
LR logSER logSER
P F
n =
Where:
P SER of partial model fit (GEV or GPD)
F SER of full model fit (BCGEV)
Likelihood Ratio Test II
This LR statistic is approx.
2
- distributed with 1 degree of freedom:
For 95% significance level - critical value is 3.842
For 99% significance level - critical value is 6.635
Hypothesis: partial model adequately represents data:
Reject if LR statistic greater than critical value at chosen significance level
-3 -2 -1 0 1 2 3
10
-2
10
0
10
2
10
4
Threshold:
+k
L
i
k
e
l
i
h
o
o
d

R
a
t
i
o

S
t
a
t
i
s
t
i
c
99% Significance
95% Significance
Mean GEV LR
Significance Testing I
For the GEV distribution:
NB: Reject hypothesis if
LR statistic > critical value
Thus:
GEV not statistically
significant for most
thresholds
For about k = +1.5 and
above, GEV is
significant (shaded area)
Significance Testing II
For the GPD distribution:
NB: Reject hypothesis if
LR statistic > critical value
Thus:
GPD not statistically
significant for all
thresholds
-3 -2 -1 0 1 2 3
10
-2
10
0
10
2
10
4
Threshold:
+k
L
i
k
e
l
i
h
o
o
d

R
a
t
i
o

S
t
a
t
i
s
t
i
c
99% Significance
95% Significance
Mean GPD LR
Load Effect Prediction I
For each span and load effect, extrapolate the BCGEV fit:
0 1000 2000 3000 4000 5000
-2
0
2
4
6
8
10
12
14
Load Effect 2 (kNm)
-
l
o
g
(
-
l
o
g
(
H
(
s
)
)
)
1-tr 2-tr 3-tr 4-tr
CDS
Return Period
Return Level
Bridge Length 40 m
Load Effect Prediction II
The BCGEV distribution predictions of lifetime load effect by threshold:
-2 -1 0 1 2
10
3
10
4
10
5
10
6
Threshold:
+k
L
o
a
d

E
f
f
e
c
t

(
k
N
m
)
20 m
30 m
40 m
50 m
-2 -1 0 1 2
10
3
10
4
10
5
Threshold:
+k
L
o
a
d

E
f
f
e
c
t

(
k
N
m
)
-2 -1 0 1 2
10
2
10
3
10
4
10
5
Threshold:
+k
L
o
a
d

E
f
f
e
c
t

(
k
N
)
(a) (b) (c)
Load Effect 1 Load Effect 2 Load Effect 3
Load Effect Prediction III
Comparison of different prediction methods:
Conventional: GEV model, ignoring different loading event types
GEV: using CDS to account for different loading event types
BCGEV, k = -2.5: considers all data and uses CDS
BCGEV, k = -1.5: the global optimum threshold identified previously
Comparison with GPD not included as the best fit model parameter was never
found to be close to zero for this data.
Load Effect Prediction IV
Comparison of different prediction methods:
20 30 40 50
0.4
0.6
0.8
1
1.2
1.4
1.6
x 10
4
Bridge Length (m)
L
o
a
d

E
f
f
e
c
t

(
k
N
m
)
20 30 40 50
1000
1500
2000
2500
3000
3500
4000
4500
5000
Bridge Length (m)
L
o
a
d

E
f
f
e
c
t

(
k
N
m
)
20 30 40 50
900
950
1000
1050
1100
1150
1200
1250
1300
Bridge Length (m)
L
o
a
d

E
f
f
e
c
t

(
k
N
)
(a) (b) (c)
Conventional GEV BCGEV, k =-2.5 BCGEV, k =-1.5
Load Effect 1 Load Effect 2 Load Effect 3
Conclusions I
The Box-Cox-GEV model allows the data to determine the appropriate form
of extreme value analysis.
The BCGEV model has been extended with Composite Distribution Statistics
(CDS) to account for the different loading event types.
The BCGEV model is a better fit than the GEV and GPD models with
considerable statistical significance, for almost all thresholds considered.
Bridge traffic load effect data lies strongly in the domain of the GEV
distribution.
Conclusions II
An optimum threshold level to apply to daily maximum load effect has been
identified, k = -1.5.
The BCGEV model is stable for k < 0, i.e. thresholds below the mean daily
maximum load effect.
The BCGEV model gives slightly higher lifetime load effect predictions that
other methods.
The BCGEV model predictions were found to be more sensitive to different
loading event types than other models.
Overall Conclusion:
The BCGEV model is more flexible and so more sympathetic to the data,
giving increased confidence to load effect predictions.
ICOSSAR 09
13 September 17 September 2009,
Osaka, Japan
Colin C. Caprani & Eugene J. OBrien
Dublin Institute of Technology & University College Dublin
Estimating Extreme Highway Bridge Traffic Load Effects
Blank

Icossar 09

Uploaded by

Copyright:

Available Formats

Icossar 09

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Icossar 09

Uploaded by

Copyright:

Available Formats

ICOSSAR 09

13 September 17 September 2009,

You might also like