See discussions, stats, and author profiles for this publication at: [Link]

net/publication/236020323

Network-Level Pavement Roughness Prediction Model for Rehabilitation

Recommendation

Article in Transportation Research Record Journal of the Transportation Research Board · December 2010
DOI: 10.3141/2155-14

CITATIONS READS
80 564

3 authors:

Nima Kargah-Ostadi Shelley Stoffels

Callentis Consulting Group LLC Pennsylvania State University
14 PUBLICATIONS 251 CITATIONS 55 PUBLICATIONS 753 CITATIONS

SEE PROFILE SEE PROFILE

Nader Tabatabaee
Sharif University of Technology
63 PUBLICATIONS 1,176 CITATIONS

SEE PROFILE

All content following this page was uploaded by Nader Tabatabaee on 23 December 2020.

The user has requested enhancement of the downloaded file.

Network-Level Pavement Roughness
Prediction Model for Rehabilitation
Recommendations
Nima Kargah-Ostadi, Shelley M. Stoffels, and Nader Tabatabaee

Pavement performance models are key components of any pavement In a network-level PMS, the relative efficiency of each maintenance
management system (PMS). These models are used in a network-level and rehabilitation alternative is required to be determined to help
PMS to predict future performance of a pavement section and identify choose the most appropriate alternative (3). Pavement performance
the maintenance and rehabilitation needs. They are also used to estimate models are implemented to guide maintenance and rehabilitation
the network conditions after the application of various maintenance and project selection, together with life-cycle cost analysis and engineer-
rehabilitation alternatives and to determine the relative cost effectiveness ing judgment. These models are also used in a network-level PMS
of each maintenance and rehabilitation alternative. Change in pavement to predict future performance of the pavement network, identify
surface roughness over time is one of the most important performance maintenance and rehabilitation needs, and estimate the network con-
indicators in this regard. A model for changes in the international ditions after the application of various maintenance and rehabilitation
roughness index (IRI) over time was developed through artificial neural alternatives (2).
networks (ANNs) pattern recognition, using information from the Change in pavement surface roughness over time is one of the
Specific Pavement Study (SPS)-5 asphalt concrete rehabilitation exper- most important performance indicators in this regard. Pavement
iment extracted from FHWA’s Long-Term Pavement Performance roughness as defined by ASTM Standard E867, is “the deviations of
database. This model can be used to predict and compare pavement a pavement surface from a true planar surface with characteristic
roughness variation trends after various rehabilitation alternatives. An dimensions that affect vehicle dynamics, ride quality, dynamic loads
example illustrates the implementation of the roughness model along and drainage” (ASTM E867-06). One of the main objectives of
with life-cycle cost analysis in making future pavement rehabilitation every road authority is to provide a comfortable ride for users, and
recommendations. Model testing results indicate prediction of IRI with pavement roughness is a good indicator of whether this criterion will
minimal errors, and predicted future roughness trends match perfectly
be fulfilled. In fact, studies following the AASHTO Road Test indi-
with the past performance. These findings indicate that the ANN model
cated that approximately 95% of the information about pavement
performs successfully in predicting IRI trends following each kind of
serviceability is contributed by surface roughness (2).
treatment in the SPS-5 experiment. The ANN model was developed for
Various factors affect pavement surface roughness: traffic, climate,
the SPS-5 flexible pavement rehabilitation sections in a wet–freeze climate
subgrade properties, pavement structure and material type, pavement
and may be applied for similar conditions. The example also shows that
construction quality, maintenance and rehabilitation, and drainage
the detailed model development and implementation framework provided
type and conditions. It is difficult and not cost-effective to collect all
in this study can assist in network-level PMS decision making.
these data for all pavements making up a network. Therefore, con-
sidering all factors affecting pavement roughness in its modeling is
A pavement management system (PMS), as defined by AASHTO, typically not practical.
is “. . . a set of tools or methods that assist decision makers in finding In this study, a deterministic modeling approach was imple-
optimum strategies for providing, evaluating, and maintaining pave- mented in which pavement international roughness index (IRI)
ments in a serviceable condition over a period of time” (1, p. I-31). was considered as the dependent variable and measures of factors
Each network-level pavement management system typically consists affecting roughness were used as independent input variables. Because
of a database for storing inventory and performance data, condition of the large number of variables and the complex way in which they
assessment and needs identification tools, tools for analysis and affect one another, the artificial neural networks (ANNs) pattern
prioritization of maintenance and rehabilitation alternatives, and a recognition approach was chosen to determine the relationship among
feedback mechanism (2). these variables.

N. Kargah-Ostadi and S. M. Stoffels, Department of Civil and Environmental Engi-

neering, Pennsylvania State University, 3127 Research Drive, State College, MODELS FOR ROUGHNESS PROGRESSION:
PA 16801. N. Tabatabaee, Department of Civil Engineering, Sharif University of A BRIEF LITERATURE REVIEW
Technology, Azadi Avenue, P.O. Box 11365-9313, Tehran, Iran. Corresponding
author: N. Kargah-Ostadi, nima@[Link]. Various studies have been conducted regarding roughness progres-
sion trends and factors affecting it. Most of these models have
Transportation Research Record: Journal of the Transportation Research Board,
No. 2155, Transportation Research Board of the National Academies, Washington,
limitations, such as intricacy of the equations, the large number of
D.C., 2010, pp. 124–133. variables used, difficulty of variables data collection (especially for the
DOI: 10.3141/2155-14 mechanistic–empirical models), and shortness of prediction period

124
Kargah-Ostadi, Stoffels, and Tabatabaee 125

(4– 6). Consequently, the need for a simpler and more efficient TABLE 1 Nomenclature of Test Sections
empirical model to use in network-level PMS is still an ongoing in LTPP SPS-5 Experiment
pursuit of the pavement engineering community.
Overlay
Some of the most important efforts in developing roughness
Section Thickness
prediction models can be listed as follows: Number Surface Preparation Type of AC (mm)

• World Bank Highway Developmental Management System 0501 Routine maintenance — 0

(HDM)-3 models developed in 1987, which need to be calibrated; 0502 Minimum surface preparation Recycled 50
• Perera et al. models developed using nonlinear regression 0503 Minimum surface preparation Recycled 125
analysis on Long-Term Pavement Performance (LTPP) database in 0504 Minimum surface preparation Virgin 125
1997 (4); 0505 Minimum surface preparation Virgin 50
• Ozbay and Laub models developed using linear regression and
0506 Intensive surface preparation Virgin 50
neural networks analyses on LTPP General Pavement Study (GPS)-1
0507 Intensive surface preparation Virgin 125
and GPS-2 databases in 1999 (7 );
• Perera and Kohn models developed using nonlinear regression 0508 Intensive surface preparation Recycled 125
analysis on LTPP database in 2001 (8); 0509 Intensive surface preparation Recycled 50
• Prozzi and Madanat empirical–mechanistic model for esti-
mating pavement roughness developed by combining experimental
(AASHO Road Test) and field [Minnesota Road Research Project,
(MnROAD)] data using nonlinear regression in 2003 (6); LTPP SPS-5 EXPERIMENT
• AASHTO Mechanistic–Empirical Pavement Design Guide
(MEPDG) models developed using nonlinear regression analysis in There are 18 SPS-5 sites in the LTPP program, each examining the
2004 (NCHRP 1-37A), which are mostly based on pavement distress performance of 8 alternative combinations of rehabilitation treat-
measurements (9); ments on flexible pavements. Except for the rehabilitation alter-
• Bekheet et al. models developed using linear regression analysis native, all factors and conditions affecting pavement performance
on LTPP Specific Pavement Study (SPS)-5 in 2004 (10); are kept the same among sections of a site (15). Rehabilitation
• Chou and Pellinen models developed using neural networks factors that are considered in the design of the SPS-5 experiment
analysis on Indiana Department of Transportation (INDOT) PMS in each site are surface preparation before rehabilitation, asphalt
database in 2005 (11); concrete (AC) overlay material type, and overlay thickness. The
• Madanat et al. models developed using linear regression analysis sections are named accordingly and this nomenclature can be seen
on Washington State PMS database in 2005 (5); and in Table 1 (16).
• Taddesse and Mork roughness progression models using multi- To develop a more homogeneous model, only the eight SPS-5
linear regression and ANNs on Ethiopian Roads Authority data in sites located in the wet–freeze climatic region, which have the most
2009 (12). data available and the most extreme climatic conditions, were con-
sidered for this study. Some of the properties of these sites, such as
These studies have shown that because of the large number of subgrade type, pavement age at the last roughness monitoring after
variables and the complex ways in which they affect one another, rehabilitation, and number of IRI readings after rehabilitation, are
use of simple statistical approaches such as linear regression does listed in Table 2 (17 ). To this point, each SPS-5 pavement section
not appear to be an appropriate means to develop pavement rough- has 11 IRI readings on average, which is anticipated to be adequate
ness prediction models (13). Hence, it is recommended that other for model development. Table 2 also shows that the last IRI reading
approaches including pattern recognition techniques, such as ANNs, for these pavements has been carried out approximately 12 years after
be undertaken (7, 11, 12, 14). overlay, on an average basis. Although this period can be considered

TABLE 2 Properties of Existing LTPP SPS-5 Sites in Wet–Freeze Zone

Age of
Rehab. Project at Number of
Year of Subgrade Date Last Profile Times Profiled
SPS-5 Site State Construction Type (M/D/Y) Date (years) After Rehabilitation

130500 Georgia 1993 Coarse 6/6/1993 11.91 8

230500 Maine 1995 Coarse 6/16/1995 9.13 9
240500 Maryland 1992 Coarse 4/1/1992 14.21 15
270500 Minnesota 1990 Fine 9/15/1990 14.73 13
290500 Missouri 1998 Coarse 8/27/1998 5.53 5
340500 New Jersey 1992 Coarse 8/18/1992 13.82 18
810500 Alberta 1990 Coarse 9/30/1990 15.71 16
830500 Manitoba 1989 Fine 9/12/1989 16.75 15

NOTE: Rehab. = rehabilitation; M/D/Y = month/day/year.

126 Transportation Research Record 2155

as “long term,” future monitoring of LTPP SPS-5 sections will further factors have consistently been deemed as having considerable
foster long-term pavement performance (and particularly roughness) influence. Those factors are initial roughness, pavement age, traffic,
prediction. climatic conditions (average annual precipitation and average annual
On the basis of previous studies on the performance of SPS-5 sites freezing index), pavement structural properties, subgrade properties
in 1997, 2000, and 2002, similar and contrasting remarks regarding (moisture content and percentage passing No. 200 sieve), drainage
roughness progression trends have been suggested (18, 19). The 1997 type and conditions, and maintenance and rehabilitation treatments
study found that there were no relationships between IRI before over- (4, 8, 13, 20–24).
lay and IRI after overlay and that the degree of surface preparation For each IRI reading in each pavement section, a data record
(with or without milling) before overlay placement has no effect on consisted of measures of the input variables at that specific time for
postoverlay IRI (19). Conversely, the 2000 study concluded that the pavement section. Therefore, the number of data records was the
sections with less preoverlay IRI and sections that were milled before same as the number of IRI readings between every two rehabilitation
overlay had lower postoverlay roughness (18). However, all of these efforts on SPS-5 pavement sections in the wet–freeze climatic zone.
studies agreed that sections with thicker overlays have lower post- This total number was 1,072. The required data were extracted from
overlay IRI and slower roughness progression rates. The studies also LTPP Standard Data Release 22.0 DVD and were analyzed using
concurred that overlay material type (virgin or recycled) does not have MATLAB stepwise regression. Of the 19 variables considered, seven
a significant effect on postoverlay IRI or future roughness trends. were found to have a statistically significant relationship with IRI.
A brief description of the variables, their average values and standard
deviations among the data records (not among pavement sections), and
INPUT VARIABLES the p-values regarding the inclusion of seven and exclusion of 12 inputs
in the linear relationship with the output can be found in Table 3.
Factors affecting pavement surface roughness were used as input Because in a PMS, condition data are usually available on a
variables to develop a model that is supposed to predict the output regular basis, and predictions are only desired for the next few time
variable: roughness. As stated before, many factors affect pavement periods, an incremental form is preferred for a model intended to be
roughness. Although most of these factors should be considered in used at the network level (6). Therefore instead of initial roughness,
the model for it to be comprehensive, the large number of variables the previous IRI measurement was considered as an input variable for
makes the process of data acquisition and roughness prediction both each IRI reading. Consequently, incremental and cumulative time and
complicated and expensive. Of course, statistical significance testing traffic factors were also considered as other potential input variables.
and engineering judgment are necessary means to influence the inclu- Table 3 shows that the average and standard deviation (SD) of
sion of various factors as input variables. In most of the previous the time passed previous IRI in the data records used for training
studies regarding factors affecting pavement roughness, several is 1.05 and 0.66 years, respectively. If it is assumed that the data

TABLE 3 Selection of Input Variables

Input Variable Average SD p-Value Significant?a

Previous IRI (m/km) 1.18 0.52 .0000 Yes

Preoverlay IRI (m/km) 1.77 0.61 .1454 No
Age since construction (years) 26.91 5.65 .2960 No
Time passed since rehabilitation (years) 7.50 4.43 .0009 Yes
Time passed since previous IRI (years) 1.05 0.66 .0000 Yes
Traffic since opened (KESAL) 36,600.98 80,477.77 .4814 No
Traffic since rehabilitation (KESAL) 4,986.53 10,136.75 .5601 No
Traffic since previous IRI (KESAL) 904.97 2,008.03 .7547 No
Average annual freezing index (°C-days) 643.03 615.82 .0001 Yes
Average annual precipitation (mm) 928.69 314.02 .2576 No
Total postoverlay HMA thickness (in.)b 9.24 2.64 .1587 No
Overlay thickness (in.) 3.80 2.04 .0001 Yes
Depth of milling (in.) 0.90 0.98 .9329 No
Total preoverlay HMA thickness (in.) 6.33 2.15 .0498 Yes
Base thickness (in.) 6.18 3.79 .7612 No
Subgrade percent passing sieve No. 200 29.81 13.98 .0084 Yes
Subgrade moisture content (%) 10.32 6.25 .2216 No
Subgrade P200*moisture content (%) 340.45 271.01 .3954 No
M&R (qualitative variable) 5.89 3.21 .2054 No

NOTE: For the linear relationship: goodness-of-fit (R2), .93; root mean square error, 0.1504; and output (IRI)
intercept, 0.0257. KESAL = kilo (1,000) ESAL (equivalent single-axle load); HMA = hot-mix asphalt; M&R =
maintenance and rehabilitation.
a
Yes, if p-value < .05, and no, if p-value > .05 (95% confidence).
b
1 in. = 25.4 mm.
Kargah-Ostadi, Stoffels, and Tabatabaee 127

follow a normal distribution, then the upper limit of a range that The neural network was built with a multilayer feed forward
contains approximately 95% of the data is (average + 2SD), which architecture and was trained using a back propagation algorithm. The
is equal to 2.37. As a result, the model trained on the basis of these MATLAB “Traingdx” function was used for the training, because it
data should be able to predict future IRI in increments of at most is a simple function that uses both an adaptive learning rate, one that
approximately 2 years. is changed in each training epoch (round) to hasten the convergence,
The variable named maintenance and rehabilitation in Table 3 is and a defined momentum constant (mc) that smoothes out the net
a qualitative variable considering the effect of treatment type on weight change (25). According to past experiences in using neural
pavement surface roughness. Each type of rehabilitation designed in networks for nonlinear pattern recognition problems such as this (26),
the SPS-5 experiment is identified with a code according to the section the learning rate, learning rate increase, and momentum constant
names in Table 1. These codes were increased by 8 if a maintenance were set to 0.05, 1.05 and 0.90, respectively.
activity has been carried out corresponding to a data record. As shown To find the optimum network architecture [transfer functions,
in Table 3, this variable was found not to be significantly correlated topology (i.e., number of layers and neurons)], the flow chart shown in
with IRI. As stated before, progressive studies on SPS-5 concluded Figure 1 was followed. Hyperbolic tangent (tansig), logistic (logsig),
that among the three rehabilitation factors of overlay thickness, and no-intercept pure linear (purelin) functions were considered in
surface preparation, and overlay material type, the latter does not this analysis.
have a significant effect on roughness trend. Therefore, in addition to The training results of a neural network, to a large extent, are
the overlay thickness variable, which was already selected, the depth dependent on the initial values allocated randomly in each epoch (26).
of milling was appended to consider the effect of surface preparation Therefore, as shown in Figure 1, the network was trained 10 times
(each with 500 epochs) for each alternative (network architecture),
on roughness progression. Consequently, the total number of input
the best and worst results among these 10 were determined, and then
variables increased to eight variables.
the various alternatives were compared according to these best
and worst results. If p = at + b, where p is the predicted IRI, t is the
actual measured IRI, and a and b are constants of the prediction
MODEL DEVELOPMENT equation, then the parameters used for comparison of the results
are as follows:
The MATLAB ANN Toolbox was used for the modeling. To have
uniformly distributed variables in a defined range of numbers, the • a,
“mapminmax” preprocessing function in the MATLAB program was • b,
used so that all the values were normalized to the range of (−1, +1). • R2 or goodness of fit,
In this way, the problem is easier for the neural network to solve, • Mean square error in IRI prediction, and
and the network outputs will be within the desired range (25). • E or number of test absolute errors more than 0.5 m/km.
From all data records, 80% (858) were randomly selected to be
used for training of the neural network, and the other 20% (214) were A neural network with three layers—input, hidden, and output—
kept as a test data set. The model was developed through the training was sufficient to solve this problem. tansig, logsig, and purelin func-
of the neural network using the training data set and then tested using tions were found to be the most appropriate transfer functions for
the test data set, which had not been previously seen by the network. the input, hidden and output layers, respectively.

1 layer and 1 neuron 1 layer and 2 neurons

10 times model development with 500 epochs of training each

Determination of the best and the worst model out of the 10 for each alternative

Comparison of the best and the worst of each alternative with the previous one

Improvement Yes
Number of neurons + 1
of the results?

No
Number of layers + 1

FIGURE 1 Flowchart for determination of optimum ANNs architecture.

128 Transportation Research Record 2155

In two separate studies in 1989, Funahashi (27, p. 191) and Hornik As stated before, these records have not been used for training the
et al. (28, p. 363) proved the Universal approximation theorem, which ANN, yet IRI was predicted with small error.
says, “Standard feed-forward networks with only a single hidden layer With the weights and biases of the neural network layers built for
can approximate any continuous function uniformly on any compact the wet–freeze zone roughness prediction model, its implementation
set and any measurable function to any desired degree of accuracy.” for predicting IRI is straightforward. As stated before, the model
This is true if there are sufficient neurons in the hidden layer (29). This predicts future roughness incrementally; the first prediction can be
theorem was also verified in this study because the ultimate models used as an input for the second one and so on. As an example, let the
had standard feed-forward networks with a single hidden layer. input matrix be as follows:
The optimum numbers of neurons in input, hidden, and output
layers were found to be 8, 5, and 1, respectively. It can be seen that the ⎛ previous IRI ( m km ) ⎞ ⎛ 1.14 ⎞
⎜ ⎟ ⎜
⎜ rehabilitation age ( years )
optimum number of neurons in the input and output layers have been ⎟
⎟ ⎜ 3.20 ⎟
determined to be equal to the number of input and output variables, ⎜ ⎟ ⎜
⎜ Δtime since previous IRI ( years )
⎟
respectively. According to empirical research about artificial neural ⎟ ⎜ 1.50 ⎟
networks, it was recommended in 1997 that in three-layer neural ⎜ ⎟ ⎜ ⎟
networks with a single hidden layer, the first or the input layer should input ⎜ average annual freezing index (°C − days )⎟ ⎜ 2, 200.00 ⎟
=⎜ ⎟ =⎜
have as many neurons as the number of input variables, and the matrix ⎜ overlay thickness ( in.) ⎟ ⎜ 2.50 ⎟
⎟
number of neurons in the third or the output layer should be equal ⎜ ⎟ ⎜ ⎟
to the number of output variables (11). This was also demonstrated ⎜ depth of milling ( in. ) ⎟ ⎜ 0 . 00
⎟
⎜ ⎟ ⎜
in this study.
⎜ surface layer thickness ( in.) ⎟ ⎜ 5.00 ⎟
As suggested by Tsoukalas and Uhrig (30), for a network with a ⎟
⎜ ⎟
single hidden layer, it is common practice to initially set the number of ⎝ subgrade percentage passing #200 (%) ⎠ ⎝ 28.00 ⎠
neurons in the hidden layer equal to approximately two thirds of the
sum of the number of neurons in the input and output layers (11, 31). First, the input data are normalized to the range of (−1, +1).
Experimenting more thoroughly by increasing or decreasing the Then the input layer weights matrix is multiplied by the normalized
number of hidden nodes (neurons) that result from the previously input matrix and added to the input layer biases matrix. The hyper-
mentioned calculation has been recommended. This empirical rule bolic tangent of the resulting matrix will be the first layer output
of thumb has also been verified in this study. (Equation Box 1).
Next, the output from the first layer will be multiplied by the
hidden layer weights and added to the hidden layer biases. The
FINAL MODEL AND ITS IMPLEMENTATION result is fed to the logistic function to form the second layer output
matrix (Equation Box 2).
After the determination of the optimum architecture for the network, The hidden layer output will be multiplied by the output weights
the ANN was trained with 1,000 epochs to develop the final model. and added to the output biases. This is the third layer output, because
Figure 2 indicates the results of testing the final model in a graph of the output function is a no-intercept pure linear one (y = x). Finally, the
predicted versus measured IRI for the 214 records in the test data set. output is converted back from a normalized condition to be the model

3.5

3
Predicted IRI (m/km)

2.5

1.5

1
y = 0.958x + 0.0529
0.5 R2 = 0.9578

0
0 0.5 1 1.5 2 2.5 3 3.5 4
Measured (Actual) IRI (m/km)

FIGURE 2 Final model testing result.

Kargah-Ostadi, Stoffels, and Tabatabaee 129

EQUATION BOX 1

input layer weights × normalized input + biases

⎛ −0.448 0.83 0.22 − 0.26 0.70 − 0.89 − 0.05 − 0.99 ⎞ ⎛ − 0.49 ⎞ ⎛ 1.84 ⎞ ⎛ 2.03 ⎞
⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎜ − 0.57 − 0.57 − 0.20 − 0.60 − 0.83 − 0.76 0.88 − 0.10 ⎟ ⎜ − 0.62 ⎟ ⎜ 1.31 ⎟ ⎜ 2.28 ⎟
⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎜ − 0.80 − 0.08 0.54 0.27 − 0.07 − 0.22 0.31 0.06 ⎟ ⎜ − 0.33 ⎟ ⎜ 0.83 ⎟ ⎜ 1.49 ⎟
⎜ 0.08 0.26 − 0.57 − 0.28 − 0.50 0.20 0.68 0.98 ⎟ ⎜ 0.94 ⎟ ⎜ 0.15 ⎟ ⎜ − 0.48 ⎟
⎜ ⎟ ×⎜ ⎟ +⎜ ⎟ =⎜ ⎟
⎜ 0.67 −1.11 0.29 − 0.33 − 0.35 0.16 0.49 − 0.05 ⎟ ⎜ − 0.336 ⎟ ⎜ 0.12 ⎟ ⎜ − 0.10 ⎟
⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎜ 0.91 − 0.07 0.29 0.11 − 0.04 − 0.07 0.03 − 0.03 ⎟ ⎜ −1.00 ⎟ ⎜ 0.19 ⎟ ⎜ − 0.13 ⎟
⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎜ 0.15 0.02 0.09 − 0.27 0.51 − 0.45 1.09 − 0.16 ⎟ ⎜ − 0.28 ⎟ ⎜ 1.80 ⎟ ⎜ 1.42 ⎟
⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎝ 1.06 0.05 0.27 0.97 − 0.73 0.18 − 0.03 0.72 ⎠ ⎝ − 0.13 ⎠ ⎝ 1.76 ⎠ ⎝ 2.02 ⎠
input layer output
⎛ 2.03 ⎞ ⎛ 0.97 ⎞
⎜ ⎟ ⎜ ⎟
⎜ 2.28 ⎟ ⎜ 0.98 ⎟
⎜ ⎟ ⎜ ⎟
⎜ 1.49 ⎟ ⎜ 0.90 ⎟
⎜ − 0.48 ⎟ ⎜ − 0.44 ⎟
⎛ 2 ⎞
tansig ⎜ ⎟ =⎜ ⎟ tansig ( n ) = ⎜ −1
⎜ − 0.10 ⎟ ⎜ − 0.10 ⎟ ⎝ 1 + e −2 n ⎟⎠
⎜ ⎟ ⎜ ⎟
⎜ − 0.13 ⎟ ⎜ − 0.13 ⎟
⎜ ⎟ ⎜ ⎟
⎜ 1.42 ⎟ ⎜ 0.89 ⎟
⎜ ⎟ ⎜ ⎟
⎝ 2.02 ⎠ ⎝ 0.96 ⎠

EQUATION BOX 2

hidden layer weights × input layer output + biases

⎛ 0.97 ⎞
⎜ ⎟
⎜ 0.98 ⎟
⎛ 1..78 −1.10 1.07 1.23 −1.04 −1.08 0.52 1.89⎞ ⎜ ⎟ ⎛ −3.12 ⎞ ⎛ 0.47 ⎞
⎜ ⎟ ⎜ 0.90 ⎟ ⎜ 1.87 ⎟ ⎜ 0.08 ⎟
⎜ −1.08 1.23 −1.22 −1.12 1.81 1.24 − 0.27 − 0.79⎟
⎜ − 0.44 ⎟ ⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎟ + ⎜ 0.28 ⎟ = ⎜ 0.94 ⎟
⎜ 0.30 0.07 1.38 − 0.24 0.47 −1.88 −1.37 − 0.02⎟ ×⎜
⎜ − 0.10 ⎟ ⎜ ⎟ ⎜ ⎟
⎜ − 0.86 −1.69 − 0.87 − 0.89 0.62 −1.87 1.21 0.97⎟ ⎜ ⎟ ⎜ −1.80 ⎟ ⎜ −2.48 ⎟
⎜ ⎟ ⎜ − 0.13 ⎟ ⎜ ⎟ ⎜ ⎟
⎜⎝ −2.04 0.06 0.64 − 0.63 0.07 −1.89 1.08 −1.31⎟⎠ ⎜ ⎟ ⎜⎝ −3.50 ⎟⎠ ⎜⎝ − 4.62 ⎟⎠
⎜ 0.89 ⎟
⎜ ⎟
⎝ 0.96 ⎠
hidden layer output
⎛ 0.47 ⎞ ⎛ 0.62⎞
⎜ ⎟ ⎜ ⎟
⎜ 0.08 ⎟ ⎜ 0.52⎟
⎜ ⎟ ⎜ ⎟ 1
logsig ⎜ 0.94 ⎟ = ⎜ 0.72⎟ logsig ( n ) =
1 + e−n
⎜ −2.48 ⎟ ⎜ 0.08⎟
⎜ ⎟ ⎜ ⎟
⎜⎝ − 4.62 ⎟⎠ ⎜⎝ 0.01⎟⎠
130 Transportation Research Record 2155

prediction. This process can be easily handled with a Microsoft assumed that the postoverlay IRI is 0.9 m/km. As shown in Figure 4,
EXCEL Visual Basic macro or any other software. the thicker overlay (6 in.) has been predicted to perform better than
the thin overlay (2 in.), which is in line with past studies on the
output layer weights × hidden layer output + biases performance of SPS-5 sections (18, 19). Figure 4 also shows that
⎛ 0.62⎞ there is not much difference between the performance of rehabilitation
⎜ ⎟ treatments with or without milling, as was expected on the basis of the
⎜ 0.52⎟ results of stepwise regression.
( 0.39 0.62 − 1.99 − 0.36 − 0.18) × ⎜⎜ 0.72⎟⎟ + ( 0.38) = ( − 0.52) If each road authority defines the maximum acceptable IRI value
⎜ ⎟ according to ride quality and pavement management policies, then
⎜ 0.08⎟ the age at which pavement sections will have IRI values more than
⎜⎝ 0.01⎟⎠ the defined threshold can be determined by using appropriate models
such as the one developed in this study. The agency can then decide
pureliin ( − 0.52 ) = − 0.52 purelin ( n ) = n when and where to execute maintenance and rehabilitation on
prediction = normalization reversed ( − 0.52 ) = 1.26 m km pavement sections in a network. Besides the maintenance and reha-
bilitation needs analysis, performance models can be used to predict
pavements’ future performance after each treatment type. According
ROUGHNESS TRENDS to this prediction, the maintenance and rehabilitation treatment
type that is most successful in postponing the time of arrival at the
In the same way, the model was used to predict the future performance IRI threshold can be selected as the most appropriate type.
of SPS-5 sites under various scenarios. Considering that there will One suitable method for setting roughness trigger values for rehabil-
be no future treatments, the past and predicted future IRI trend for itation is to determine the level at which at least 50% of highway users
section 0506 of Maryland SPS-5 site is illustrated in Figure 3. This decide that the pavement ride is unacceptable for a given functional
section was milled 0.6 in. before rehabilitation with a thin 2.8-in. class and pavement type (32). In a study of 2,180 in-vehicle ratings and
overlay. It has demonstrated good performance for 14 years after the Washington state PMS database, nearly half of all “acceptable”
rehabilitation, with its IRI increased from 0.76 m/km to 1.32 m/km. ratings were made on segments with an IRI of less than 1.42 m/km (33).
The predicted future trend also shows that the good performance This value was chosen as the rehabilitation trigger value to be used
continues up to 19 years after rehabilitation when IRI is 1.64 m/km in this example. On the basis of an IRI trigger value of 1.42 m/km, the
and the linearly increasing roughness trend changes to an expo- number of rehabilitation treatments and the roughness trend following
nentially increasing one. This finding might be an indication for a
each scenario in Figure 4 has been indicated in Figure 5. As noted
treatment need.
earlier, Table 2 shows that the average maximum rehabilitation
age at IRI measurements of pavement sections used to develop the
USING MODEL FOR REHABILITATION model was 12 years. Therefore, it is assumed that the model predicts
RECOMMENDATIONS future roughness most reliably up to 12 years after rehabilitation,
and accordingly, an analysis period of 12 years was selected for this
The predicted performance of section 0509 of Manitoba, Canada example. The analysis period could have been extended here, if
SPS-5 site under various scenarios has been illustrated in Figure 4. necessary, because each alternative was predicted to have an effective
On the basis of the average IRI after overlay in SPS-5 sections, it was life before the trigger value of less than 12 years.

4.0
Past Performance Predicted Future Trend
3.5

3.0

2.5
IRI (m/km)

2.0

1.5

1.0

0.5

0.0
0 5 10 15 20 25
Time After First Rehabilitation (yrs)

FIGURE 3 The past and predicted future IRI trend for Section 24-0506.
Kargah-Ostadi, Stoffels, and Tabatabaee 131

2.4
Past Performance
2.2 Thin Overlay
Thin Overlay with Milling
2.0
Thick Overlay
Thick Overlay with Milling
1.8
IRI (m/km)

1.6

1.4

1.2

1.0
0.9 m/km
0.8
14 16 18 20 22 24 26
Time After First Rehabilitation (yrs)

FIGURE 4 Predicted performance of Section 83-0509 under alternative scenarios.

As illustrated in Figure 5, IRI of the pavement section following for that alternative. (Although equivalent annual cost is sometimes
both types of thin overlay regardless of milling has been predicted used to attempt to avoid this problem, that approach is only viable
to reach the trigger value after 5 years. Meanwhile, the service life of if it is reasonable to assume that the alternatives will be repeated
thick overlay before another treatment is required has been predicted identically for the least common multiple of lives—in this case, that
to be 10 years. This period increases to 12 years if milling was carried period was too long to be reasonable.)
out before thick overlay placement. Table 4 shows that thick overlay with milling has the least life-cycle
For this example, the cost of thin overlay, thick overlay, and milling cost. The life-cycle cost of thick overlay without milling is also
are assumed to be $60,000, $85,000, and $16,000 per lane mile, comparable. If the first year capital investment has more influence
respectively. Considering an effective discount rate of 4%, the present on decision making, then the thick overlay only alternative might
worth of each of the four alternative scenarios in 12 years can be be chosen, because it has a lower initial cost. For the purpose of
calculated as in Table 4. Because the thick overlay alternative was simplicity, other important factors such as user costs, social factors,
the only scenario with IRI significantly better than the trigger value and environmental impacts were not considered. Inclusion of any of
at 12 years, a condition–proportionate residual value was included those factors might change the ultimate decision dramatically. For

2.2

Thin Overlay
2.0
Thin Overlay with Milling
Thick Overlay
1.8
Thick Overlay with Milling

1.6
IRI (m/km) 1.42 m/km
1.4

1.2

1.0
0.9 m/km

0.8
0 2 4 6 8 10 12
Time (yrs)

FIGURE 5 Analysis of different rehabilitation scenarios on Section 83-0509.

132 Transportation Research Record 2155

TABLE 4 Present Worth Calculation for Rehabilitation Alternatives on 83-0509

Alternative Initial Cost Future Treatment (years) Present Worth

60 60
Thin overlay 60 5 60 + + = 51
9 (1.04 )5
(1.04 )9
76 76
Thin overlay + milling 60 + 16 = 76 5 76 + + = 192
9 (1.04 )5 (1.04 )9
85 0.6 × 85
Thick overlay 85 10 85 + − = 111
12 (residual value) (1.04 )10 (1.04 )12
Thick overlay + milling 85 + 16 = 101 N/A 101

NOTE: All costs are for $1,000 per lane mile.

example, alternatives with milling might impose higher user delay REFERENCES
costs because of excessive change in traffic patterns for longer con-
struction periods, and, therefore, thick overlay without milling might 1. Guide for Design of Pavement Structures. AASHTO, Washington, D.C.,
be recommended instead. 1993.
2. Haas, R., W. R. Hudson, and J. Zaniewski. Modern Pavement Manage-
ment. Krieger Publishing Company, Malabar, Fla., 1994.
3. Shahin, M. Y. Pavement Management for Airports, Roads, and Parking
CONCLUSIONS Lots, Chapman and Hall, New York, 1994.
4. Perera, R. W., C. Byrum, and S. D. Kohn. Investigation of Development
A simple empirical model for IRI progression was developed through of Pavement Roughness. Publication FHWA-RD-97-147. Office of
Infrastructure Research and Development, FHWA, U.S. Department of
ANN pattern recognition using the information from the LTPP SPS-5 Transportation, 1997.
database. Using an example, it was shown how straightforward the 5. Madanat, S., Z. Nakat, and N. Sathaye. Development of Empirical–
application of an ANN model is. This model can be used in network- Mechanistic Pavement Performance Models using Data from the
Washington State PMS Database. Publication UCPRC-RR-2005-5.
level PMS to predict and compare pavement roughness variation Pavement Research Center, University of California, Davis and Berkeley,
trends after various rehabilitation alternatives. 2005.
It was shown how an IRI model can be used to determine when 6. Prozzi, J. A., and S. M. Madanat. Empirical–Mechanistic Model for
treatments are required and what the performance trend is going Estimating Pavement Roughness. Presented at 82nd Annual Meeting of
Transportation Research Board, Washington, D.C., 2003.
to be after each kind of rehabilitation treatment. An example illus- 7. Ozbay, K., and R. Laub. Models for Pavement Deterioration Using
trated the use of the IRI model along with life-cycle cost analysis LTPP. Publication FHWA-NJ-1999-030. FHWA, U.S. Department of
in making future pavement rehabilitation recommendations. Under Transportation, 2001.
8. Perera, R. W., and S. D. Kohn. NCHRP Web Document 40: LTPP Data
the specific assumptions and considering only the construction
Analysis: Factors Affecting Pavement Smoothness. National Coopera-
costs, it was found that the life-cycle cost of thick overlay with tive Highway Research Program, Transportation Research Board of the
milling is the least among rehabilitation alternatives for the example National Academies, Washington, D.C., 2001.
pavement section. 9. ARA, Inc., ERES Consultants Division. Guide for Mechanistic–Empirical
Design of New and Rehabilitated Pavement Structures. Final report,
The predicted IRI trends illustrated that thick overlays outperform NCHRP Project 1-37A. Transportation Research Board of the National
thin overlays regarding roughness performance. These trends also Academies, Washington, D.C., 2004. [Link]
indicated that there is not much difference between the performance [Link]. Accessed April 26, 2010.
of rehabilitation treatments with or without milling, as it was expected 10. Bekheet, W., K. Helali, A. Abd El Halim, and J. Springer. Comprehen-
sive Approach for Development of Performance Models for Network-
on the basis of the results of stepwise regression. These findings Level Pavement Management System Using LTPP Data. Presented at
agree with previous studies on the roughness progression of SPS-5 84th Annual Meeting of Transportation Research Board, Washington,
pavements. D.C., 2005.
11. Chou, S.-F., and T. K. Pellinen. Assessment of Construction Smoothness
Predicting IRI values for the test data set with minimal errors Specification Pay Factor Limits Using Artificial Neural Network
and similarity between predicted future roughness trends and past Modeling. Journal of Transportation Engineering, Vol. 131, No. 7, 2005,
performance for various rehabilitation alternatives shows that pp. 563–570.
the ANN model performs successfully in predicting IRI trends 12. Taddesse, E., and H. Mork. Roughness Progression Models by Regres-
sion and Artificial Neural Network Techniques. In Bearing Capacity of
following each kind of treatment in SPS-5 experiment. Although Roads, Railways and Airfields, Proc., 8th International Conference.
the model performed well for the test data set and may be applied Champaign, Ill. CRC Press, Taylor and Francis Group, New York, 2009,
for similar conditions, it might not perform the same way on other pp. 521–530.
databases. This is because it was trained using SPS-5 wet–freeze 13. Hunt, P. D., and J. M. Bunker. Study of Site-Specific Roughness Pro-
gression for a Bitumen-Sealed Unbound Granular Pavement Network.
experiment data. Therefore, it is recommended that data from In Transportation Research Record: Journal of the Transportation
each PMS database be used to train agency-specific ANN perfor- Research Board, No. 1819, Transportation Research Board of the National
mance models before implementing them in the decision making Academies, Washington, D.C., 2003, pp. 273–281.
14. Lee, Y.-H., Y.-B. Liu, and H.-W. Ker. Application of Modern Regres-
process of that PMS. The methodology in this paper provides sion Techniques and Artificial Neural Networks to Pavement Prediction
a detailed framework for such ANN model development and Modeling. Presented at 86th Annual Meeting of Transportation Research
implementation. Board, Washington, D.C., 2007.
 Kargah-Ostadi, Stoffels, and Tabatabaee 133

15. Elkins, G. E., P. Schmalzer, T. Thompson, and A. Simpson. Long-Term 24. Zhou, G., W. Linbing, and L. Yang. International Roughness Index Model
Pavement Performance Information Management System: Pavement Enhancement for Flexible Pavement Design Using LTPP Data. Presented
Performance Database User Reference Guide. Publication FHWA- at 87th Annual Meeting of Transportation Research Board, Washington,
RD-03-088. Office of Infrastructure Research and Development, FHWA, D.C., 2008.
U.S. Department of Transportation, 2007. 25. MATLAB help v.7.6 R2008a. MathWorks, Inc., Natick, Mass.
16. Hanna, A. N., S. D. Tayabji, and J. S. Miller. SHRP-LTPP Specific 26. Samarasinghe, S. Neural Networks for Applied Sciences and Engineering:
Pavement Studies: Five-Year Report. Publication SHRP P-395. Strategic From Fundamentals to Complex Pattern Recognition. Auerbach Publi-
Highway Research Program, National Research Council, Washington cations, Taylor and Francis Group, New York, 2007.
D.C., 1994. 27. Funahashi, K. On the Approximate Realization of Continuous Map-
17. Long-Term Pavement Performance (LTPP) Standard Data Release 22.0.
pings by Neural Networks. Neural Networks Journal, Vol. 2, 1989,
DVD Version. FHWA, U.S. Department of Transportation, 2008.
18. Von Quintus, H. L., A. L. Simpson, and A. A. Eltahan. Rehabilitation of pp. 183–192.
Asphalt Concrete Pavements: Initial Evaluation of the SPS-5 Experiment— 28. Hornik, K., M. Stinchcombe, and H. White. Multilayer Feed-Forward
Final Report. Publication FHWA-RD-01-168. Office of Engineering Networks Are Universal Approximators. Neural Networks Journal,
Research and Development, FHWA, U.S. Department of Transportation, Vol. 2, 1989, pp. 359–366.
2006. 29. Yang, J., J. J. Lu, M. Gunaratne, and Q. Xiang. Forecasting Overall
19. Hall, K. T., C. E. Correa, and A. L. Simpson. Performance of Flexible Pavement Condition with Neural Networks: Application on Florida
Pavement Rehabilitation Treatments in the LTPP SPS-5 Experiment. In Highway Network. In Transportation Research Record: Journal of the
Transportation Research Record: Journal of the Transportation Research Transportation Research Board, No. 1853, Transportation Research Board
Board, No. 1823, Transportation Research Board of the National Acad- of the National Academies, Washington, D.C., 2003, pp. 3–12.
emies, Washington, D.C., 2003, pp. 93–101. 30. Tsoukalas, L. H., and R. E. Uhrig. Hypermedia Integration of Information
20. Chatti, K. Effect of Design and Site Factors on Long-Term Performance Resources for Nuclear Plant Operations. Journal of Nuclear Technology,
of Flexible Pavements in SPS-1 Experiments. Presented at 85th Annual Vol. 119, No. 1, 1997, pp. 48–62.
Meeting of Transportation Research Board, Washington, D.C., 2006. 31. Neuralware. [Link] Accessed September 14, 2007.
21. Stoffel, S., S. W. Lee, and A. Bae. Observed Effects of Subgrade Mois- 32. Janoff, M. S. NCHRP Report 308: Pavement Roughness and Rideability
ture on Longitudinal Profile. Presented at 85th Annual Meeting of Field Evaluation. TRB, National Research Council, Washington, D.C.,
Transportation Research Board, Washington, D.C., 2006. 1988.
22. Kutay, M. E. Spectral Analysis of Factors Affecting Roughness in
33. Shafizadeh, K., and F. Mannering. Acceptability of Pavement Roughness
Flexible Pavements. FHWA, Turner-Fairbank Highway Research Center,
on Urban Highways by Driving Public. In Transportation Research
McLean, Va. Submitted to PIARC Paper Contest, 23rd World Road
Congress, Paris, 2007. [Link] Record: Journal of the Transportation Research Board, No. 1860, Trans-
MEKutay_2007_PIARC.pdf. Accessed April 21, 2010. portation Research Board of the National Academies, Washington, D.C.,
23. Tighe, S. Evaluation of Subgrade and Climatic Zone Influences on 2003, pp. 187–193.
Pavement Performance in the Canadian Strategic Highway Program’s
(C-SHRP) Long-Term Pavement Performance (LTPP) Study. Canadian
Geotechnical Journal, Vol. 39, No. 2, 2002, pp. 377–387. The Pavement Management Section peer-reviewed this paper.

View publication stats