Hierarchical Markov Model for pavement

deterioration forecasting : Potential use in Vietnam

Nam Lethanh
23 September 2011 @Hanoi summer school, Vietnam
Postdoc fellow, Swiss Federal Institute of Technology, Switzerland
Outline
Deterioration process of multi-layers pavement
structure
Basic Markov chain deterioration model
Deterioration forecasting model for multi-layer
pavement structure
Potential of models applicability in Vietnam
2 DBAUG/IMI/IMG
Performance curve
3 DBAUG/IMI/IMG
Surface course
Base course
Condition state
Condition state
Failure (absorbing)
Failure (absorbing)
Interventions
4 DBAUG/IMI/IMG
Surface course
Base course
Condition state
Condition state Time
WHEN renewal should be
carried out for base course?
Optimal time for renewal?
Monitoring/inspection activities
5 DBAUG/IMI/IMG
Surface course
PASCO brochure 2010 (http://www.pasco.co.jp/eng/)
Monitoring/inspection activities
6 DBAUG/IMI/IMG
Base course
Destructive test
Non destructive test
Monitoring/inspection activities
7 DBAUG/IMI/IMG
Falling weight deflectometer (FWD) formula
Source: http://www.uct.edu.vn/utc/data/document/news/08_2010/08_2010_1320.pdf
Monitoring/inspection activities
8 DBAUG/IMI/IMG
Falling weight deflectometer (FWD)
Source: http://www.uct.edu.vn/utc/data/document/news/08_2010/08_2010_1320.pdf
Deterioration process
9 DBAUG/IMI/IMG
Condition state definition
10 DBAUG/IMI/IMG
Condition states
Time
ij
t
i
j
A
t
B
t
Example of condition state
11 DBAUG/IMI/IMG
IRI
(m/km)
Condition
Condition
states
IRI <= 3 Very good 1
3 < IRI <= 4 Good 2
4 < IRI <= 5 Average 3
5 < IRI <= 6 Bad 4
IRI > 6 Very bad 5
Markov transition probabilities
12 DBAUG/IMI/IMG
Conditional Probability
Pr [ ( ) | ( ) ]
B A ij
ob h j h i t t t = = =
Transition Probability Matrix
11 1
0
J
JJ
t t
t
(
(
H =
(
(
1
0
0 (when )
1
ij
ij
J
ij
j
i j
t
t
t
=

>

= >
`

Example
13 DBAUG/IMI/IMG
100
80
64
51
20 32
39
4 9
1
Condition
state
1
2
3
4
2 3
80%
20%
80%
20%
80%
20%
80%
20%
80%
20%
75%
25%
0.80 0.20 0 0
0 0.80 0.20 0
0 0 0.75 0.25
0 0 0 1
| |
|
|
H =
|
|
\ .
0 1
Example
14 DBAUG/IMI/IMG
Estimating Markov transition probabilities
15 DBAUG/IMI/IMG
1 1
1
exp( )
j j k
m m
ij k
k i m i m k
m k m k
Z
u u
t u
u u u u

= = =
+
=

[ [
Hazard rate
' k
i i
x u | =
Characteristic
variables
Unknown parameters
Maximum Likelihood estimation (MLE) (1)
16 DBAUG/IMI/IMG
Maximum likelihood estimation is very very very
very fundemental part of data analysis
1 2 1
( , ,..., | ) ( | ) ( | )
n n
f x x x f x f x | | | =
Observations
Unknown parameters
Join probability density function
Maximum Likelihood estimation (2)
17 DBAUG/IMI/IMG
Maximum likelihood estimation is very very very
very fundemental part of data analysis
1 2 1
( , ,..., | ) ( | ) ( | )
n n
f x x x f x f x | | | =
Observations
Unknown parameters
Join probability density function
MaximumLikelihood estimation (3)
18 DBAUG/IMI/IMG
If we consider observed variable x as a fixed
parameter, and become the functions variable,
which is freely varied likelihood
1 2 1 2
( | , ,..., ) ( , ,..., | )
n n
x x x f x x x | | =
1
( | )
n
i
i
f x |
=
=
[
{ }
1 2

argmax ( | , ,..., )
mle n
x x x
|
| |
eO
=
Finding that maximize
Models formulation (Hidden Markov model)
19 DBAUG/IMI/IMG
Conditional transition
20 DBAUG/IMI/IMG
Base course
Surface course
Hazard rates
21 DBAUG/IMI/IMG
Base course
Surface course
Complete likelihood function
22 DBAUG/IMI/IMG
Can we use conventional Maximum
likelihood estimation approach?
Bayesian estimation
23 DBAUG/IMI/IMG
Data generation: Modeled by parametric likelihood
MLE: Considers only maximizer of the likelihood
Bayesian idea: Consider all possible value of |
a) To rank them against one another, we need their distribution
b) To take the data into account, we need a conditional
distribution of ||x
How can we obtain p(||x) from likelihood?
Posterior distribution
24 DBAUG/IMI/IMG
Posterior: We can begin with p(||x) and plug in definition of
conditional distribution
In MLE: one model assumption is needed (likelihood).
But in order to work with a full distribution of the
parameter, we need a second model assumption (prior)
( | ) ( )
( | )
( )
p x p
p x
p x
| |
| =
Consequence: To obtain data conditional distribution of |
(posterior) from likelihood p(x||), we need p(|) as prior
Glossary
25 DBAUG/IMI/IMG
Bayesian estimation: the distribution of ||x follow Bayess
formula, thus, estimation based on posterior is called Bayesian
estimation
( | ) ( )
( | )
( )
p x p
p x
p x
| |
| =
Evidence p(x): p(x) is normalization constant since data is
assumed to be given (x)
likelihood x prior
posterior =
evidence
MCMC
26 DBAUG/IMI/IMG
Markov chain Monte Carlo (MCMC) methods (which
include random walk Monte Carlomethods) are a class
of algorithms for sampling from probability
distributions based on constructing a Markov chain that
has the desired distribution as its equilibrium distribution.
The state of the chain after a large number of steps is then
used as a sample of the desired distribution. The quality of
the sample improves as a function of the number of steps.
Source: http://en.wikipedia.org/wiki/Markov_chain_Monte_Carlo
Prior distribution for the hierarchical model
27 DBAUG/IMI/IMG
Assumption for parameter of base course
Assumption for parameter of surface course
Complete likelihood function
28 DBAUG/IMI/IMG
Empirical study
29 DBAUG/IMI/IMG
Numbers of the map are the names of national road
Source: http://hdl.handle.net/2433/85384
- 2 inspection times (2003) and (2004)
- >10000 road sections were monitored

Condition state definition
30 DBAUG/IMI/IMG
IRI
(m/km)
Condition
Condition
states
IRI <= 3 Very good 1
3 < IRI <= 4 Good 2
4 < IRI <= 5 Average 3
5 < IRI <= 6 Bad 4
IRI > 6 Very bad 5
Asphalt surface layer
FWD (mm)

Condition
Condition
states
FWD <= 400 Very good 1
400 < FWD <= 800 Good 2
800 < FWD <= 1200 Average 3
1200 < FWD <= 1600 Bad 4
FWD > 1600 Very bad 5
Base course layer
? ? ? sections ~10000 sections
Deterioration curves
31 DBAUG/IMI/IMG
5 10 15 20
1
2
3
4
5
Condition states
0
Surface
course
Base
course
? ? ?
Interventions
32 DBAUG/IMI/IMG
5 10 15 20
1
2
3
4
5
Condition states
0
Surface
course
Base
course
Define time to intervene,
and type of intervention
for surface course
Define time to intervene,
and type of intervention
for base course
Example
33 DBAUG/IMI/IMG
Description Intervention times Cost
1st 2nd 3rd 4th 5th 6th
Intervention year 2011 2021 2028 2033 2038 2043
Top asphalt course
- Owner cost 29.04
+ set up cost 2 2 2 2 2 2 3.54
+ intervention cost 10 11 12 13 14 15 19.23
- Social cost (stakeholders) 4 3.5 3 2 2 2 6.27
Base course
- Owner cost 0.00
+ set up cost 0 0 0 0 0 0 0.00
+ intervention cost 0 0 0 0 0 0 0.00
- Social cost (stakeholders) 0 0 0 0 0 0 0.00

Discount factor 0.1
Total cost 29.04
Example
34 DBAUG/IMI/IMG
Description Intervention times Cost
1st 2nd 3rd 4th
Intervention year 2011 2021 2028 2033 2038 2043
Top asphalt course
- Owner cost 25.84
+ set up cost 2 2 2 2 0 0 3.32
+ intervention cost 10 11 12 0 0 0 16.24
- Social cost (stakeholders) 4 3.5 3 4 0 0 6.28
Base course
- Owner cost 2.44
+ set up cost 0 0 0 2 0 0 0.22
+ intervention cost 0 0 0 20 0 0 2.22
- Social cost (stakeholders) 0 0 0 0 0 0 0.00

Discount factor 0.1
Total cost 28.28
THANK YOU FOR YOUR ATTENTION!
35 DBAUG/IMI/IMG