Advances in Engineering Software 57 (2013) 59–64

Contents lists available at SciVerse ScienceDirect

Advances in Engineering Software

journal homepage: www.elsevier.com/locate/advengsoft

Modeling for pavement roughness using the ANFIS approach

Serdal Terzi
Suleyman Demirel University, Engineering Faculty, 32260 Isparta, Turkey

a r t i c l e i n f o a b s t r a c t

Article history: The term ‘‘present serviceability’’ was adopted to represent the momentary ability of pavement to serve
Received 16 December 2010 traffic, and the performance of the pavement was represented by its serviceability history in conjunction
Received in revised form 31 October 2012 with its load application history. Serviceability was found to be influenced by longitudinal and transverse
Accepted 6 November 2012
profile as well as the extent of cracking and patching. The amount of weight that should be assigned to
Available online 19 January 2013
each element in the determination of overall serviceability is a matter of subjective opinion.
In this study, an Adaptive Neural-Based Fuzzy Inference System (ANFIS) method is used in modeling
Keywords:
the International Roughness Index (IRI) of flexible pavements. Data from the LTPP IMS database, namely,
Flexible highway pavements
International Roughness Index (IRI)
age, cumulative Equivalent Single Axle Loads (ESALs), and Structure Number (SN) were used in the mod-
Adaptive neural-based fuzzy inference eling. Results showed that the ANFIS model is successful for the estimation of IRI, and this model can be
system easily applied in different regions. The model can be further developed by combining expert judgment
Pavement performance and newly measured data.
Structure number Ó 2013 Elsevier Ltd. All rights reserved.
Equivalent Single Axle Loads

1. Introduction quarter-car simulation as described in the ‘‘Report on the Calcula-

tion of IRI from Longitudinal Road Profile’’ [11].
The pavement management concept was first conceived in the Prediction of future pavement condition is not only essential for
mid-1960s to organize and coordinate the activities involved in maintenance budget forecasting at the network level but also for
achieving the best value possible for the available funds [5]. In re- determining the most cost-effective rehabilitation strategy at the
sponse to the growing need for highway rehabilitation and mainte- project level [14]. The knowledge of future pavement performance
nance on one hand and shrinking resources on the other, there has is essential to sound pavement design and life-cycle economic
been an increased interest in developing a formal management ap- evaluation at the project level of pavement management. At the
proach to optimize the utilization of highway construction and network level, such information is important to ensure adequate
maintenance resources. The specific component of this approach financial funding and budgeting [10].
related to pavement is termed ‘‘pavement management system’’ Jang [4] first proposed the ANFIS method and applied its princi-
(PMS) [9,14]. ples successfully to many problems. It identifies a set of parame-
Hence, optimizing the current pavement condition evaluation ters through a hybrid learning rule combining the back-
practice will be the first and foremost task of efficient pavement propagation gradient descent error digestion and a least squares
management systems [14]. method. It can be used as a basis for constructing a set of fuzzy
Setting priorities for pavement maintenance and rehabilitation IF–THEN rules with appropriate membership functions in order
depends on the availability of a universal scale for assessing the con- to generate the preliminary stipulated input–output pairs.
dition of every element in the network. The condition of a pavement The main purpose of this paper is to develop an ANFIS method-
section has traditionally been assessed by several condition indexes. ology for estimating the International Roughness Index (IRI) with-
The present serviceability index (PSI) is one common evaluator used out any restrictive assumption by considering age, cumulative
to describe the functional condition with respect to ride quality. Equivalent Single Axle Loads (ESALs), and Structure Number (SN)
Pavement condition index (PCI) is another index commonly used as input variables and IRI as output variable.
to describe the extent of distress on a pavement section [2].
The International Roughness Index (IRI) is a statistic used to
estimate the amount of roughness in a measured longitudinal pro- 2. International roughness index
file. The IRI is computed from a single longitudinal profile using a
Roughness is defined in accordance with [1] as ‘‘the deviation of
a surface from a true planar surface with characteristic dimensions
E-mail address: serdalterzi@tef.sdu.edu.tr that affect vehicle dynamics and ride quality’’.

0965-9978/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.advengsoft.2012.11.013
60 S. Terzi / Advances in Engineering Software 57 (2013) 59–64

Roughness and skid resistance measurements are required for od is known as Response Type Road Roughness Measurement Sys-
the project level of assessment. In cases where the pavement is tem (RTRRMS). Response type indicators are dependent on the
in good condition, the roughness value is most valuable piece of vehicle and may give different results when they are re-used on
information; if, however, the pavement condition requires recon- the same road section with the same vehicle. Profile-based indica-
struction, it becomes irrelevant [12]. tors can give the same result for re-measurement on the same road
It has been shown that the serviceability of roads is largely re- section because they use a mathematical model [12].
lated to surface roughness. Road users usually define satisfaction The experimental devices are used in the evaluation of the
or dissatisfaction with the concepts of regularity or roughness in pavement roughness gauge to pavement profile or the vehicle’s
terms of comfort. Roughness rating varies depending on the vehi- reaction to pavement roughness. A common scale is needed to
cle passenger’s or driver’s patience, vehicle speed, and the features evaluate the results of the different experimental devices [3].
[3]. RTRRMS measure the vertical movements of the rear axle of an
There are several factors that cause roughness pavement. These automobile or the axle of a trailer relative to the vehicle frame
factors are traffic loads, environmental impact (climate condi- (Fig. 1). The meters are installed in vehicles with a displacement
tions), pavement roughness, and faulty material used in the con- transducer on the body located between the middle of the axle
struction of the pavement. and the body of a passenger car or trailer. The transducer detects
Even when the pavement is newly constructed there is a little small increments of axle movement relative to the vehicle body.
roughness; a completely smooth pavement surface would not be The output data consists of a strip chart plot of the actual axle body
suitable as it would cause an increase in stopping distance [13]. movement versus the time of travel.
The measurement of roughness became a world-wide practice The disadvantage of an RTRRM is that its measured axle body
after the 1970s. At that time, due to the use of different equipment movement vs. time depends on the dynamics of the particular
and methods in different regions, it was not possible to make a measurement vehicle, which results in two unwanted effects:
common evaluation. To this end, in 1982, the World Bank created
a common calibration standard for the measurement of roughness  Roughness measuring methods have not stabilized with time.
and launched an experimental study in Brazil to develop the rela- Measures made today with road meters cannot be compared
tionship in order to define whether all the measurements could be with confidence to those made several years ago.
developed with a common scale that evaluated roughness using a  Roughness measurements are not transferrable. Road meter
fixed index. As a result, the International Roughness Index (IRI) was measures made by one system are seldom reproducible by
developed [7]. another.
Today, several indicators are used in the assessment of rough-
ness. These indicators, or pavement roughness surface profile, or Because of these two effects, profiling devices are becoming
a vehicle-mounted device, are based on outputs. The second meth- more popular [16].

Fig. 1. The quarter-car model [13].

Table 1
Model data description and ranges [8].

Variable Description Range

Age Measured in thousands of days from the date of construction to the day of the IRI reading 1.2–16.5
Cumulative ESALs The number of ESALs the pavement experienced from construction to the day of the IRI reading. 0.2–20
(millions of ESALs)
Structural number The structural number recorded closest to the day of the IRI reading. Most values are from 2.85–6.6
backcalculations of falling weight deflection readings
IRI International Roughness Index 0.7281–3.1426
 S. Terzi / Advances in Engineering Software 57 (2013) 59–64 61

Fig. 2. Structure of ANFIS (a) fuzzy inference system and (b) equivalent ANFIS.

Table 2
3. Modeling roughness
Model parameters.

3.1. IRI data Model parameters Value

Range of influence 0.5
The Long Term Pavement Project (LTPP) was originally designed Squash factor 1.25
Accept ratio 0.5
as a 20-year project to monitor and gather data on various types of
Reject ratio 0.15
pavements. The Federal Highway Administration (FHWA) is the
current coordinator of the LTPP project and database.
The objectives of the LTPP program were as follows:
Table 3
 To evaluate existing design methods. To develop improved ANFIS models and their regression coefficients.
design methodologies and strategies for rehabilitation of exist- Model input R2
ing pavements.
ESAL 0.0005
 To develop improved design equations for new and recon- SN 0.4222
structed pavements. AGE 0.4910
 To determine the effects of (a) loading, (b) environment, (c) ESAL–SN 0.4356
AGE–SN 0.6556
material properties and variability, (d) construction quality,
AGE–ESAL 0.6719
and (e) maintenance level on pavement distress and AGE–ESAL–SN 0.9727
performance.

Establish a national long-term pavement database to support

Strategic Highway Research Program (SHRP) objectives and future
needs [8]. Table 1 shows model data descriptions and ranges.

3.2. ANFIS modeling

ANFIS applications and properties are investigated, and a num-

ber of methods are proposed for partitioning the input space and
hence, the structure identification problem is addressed. Funda-
mentally, ANFIS is a graphical network representation of Sugeno-
type fuzzy systems, endowed with neural learning capabilities. Fig. 3. The structure of the ANFIS model.
62 S. Terzi / Advances in Engineering Software 57 (2013) 59–64

Fig. 4. Membership functions of (a) SUM ESAL, (b) SN, (c) AGE, and (d) IRI.

The network is comprised of nodes with specific functions, or du- In order to illustrate ANFIS’s representational strength, the
ties, collected in layers with specific functions [15]. neural fuzzy control systems are considered based on the
 S. Terzi / Advances in Engineering Software 57 (2013) 59–64 63

 h !
Tagaki–Sugeno–Kang (TSK) fuzzy rules whose consequent parts . i2 bji
are linear combinations of their preconditions. The TSK fuzzy rules lAj ðxi Þ ¼ 1= 1 þ xi  mji rji ð7Þ
i
are in the following forms:
or
Rj : IF x1 is Aj1 AND x2 is Aj2 AND . . . AND xn is Ajn ð1Þ
( )
h . 2 bji
j
THEN y ¼ fj ¼ aj0 þ aj1 x1 þ aj2 x2 þ  þ ajn xn ð2Þ lAj ðxi Þ ¼ exp  xi  li rji ð8Þ
i

where xi’s (i = 1, 2, . . ., n) are input variables, y is the output variable n o

j
where mj1 ; rj1 ; bi is the parameter set to be tuned. In fact, contin-
(pavement serviceability ratio), Aji are linguistic terms of the pre-
uous and piecewise differentiable functions, such as commonly
condition part with membership functions lAj1 (xi), (j = 1, 2, . . ., n)
used trapezoidal or triangular membership functions, are also qual-
aj1 —
CR are coefficients of linear equations fi (x1, x2, . . ., xn). To simplify
ified candidates for node functions in this layer. Parameters in this
the discussion it is necessary to focus on a specific Neuro-Fuzzy
layer are referred to as precondition parameters,
Controller (NFC) of this type called ANFIS.
Layer 3: Every node in this layer is labeled G and multiplies the
Let us assume that the fuzzy control system is under consider-
incoming signals lj ¼ lAj ðx1 Þ þ lAj ðx2 Þ and sends the product
ation of two inputs x1 and x2 and one output y and that the rule 1 2
out. Each node output represents the firing strength of a rule.
base contains two TSK fuzzy rules as follows:
Layer 4: Every node in this layer is labeled by N and calculates
R1 : IF x1 is A11 AND x2 is A12 ; THEN y ¼ f1 ¼ a10 þ a11 x1 þ a12 x2 ð3Þ the normalized firing strength of a rule. That is, the jth node cal-
culates the ratio of the jth rule’s firing strength of all the rules’
R2 : IF x1 is A21 AND x2 is A22 ; THEN y ¼ f2 ¼ a20 þ a21 x1 þ a22 x2 ð4Þ firing strengths as,
 
In the TSK fuzzy system, for given input values x1 and x2, the in-
lj ¼ lj = lAj ðx1 Þ þ lAj ðx2 Þ ð9Þ
ferred output y is calculated by the following formula: 1 2

y ¼ ðl1 f1 þ l2 f Þ2 =ðl1 þ l2 Þ ð5Þ Layer 5: Every node j in this layer calculates the weighted con-
sequent value as,

where lj are firing strengths of Rj, j = 1, 2, and they are given by the
lj aj0 þ aj1 x1 þ aj2 x2 ð10Þ
equation below:
n o
lj ¼ lAj ðx1 Þ þ lAj ðx2 Þ; j ¼ 1; 2 ð6Þ where lj is the output of layer 4 and aj0 ; aj1 ; aj2 is the set to be
1 2
tuned. Parameters in this layer are referred to as consequent
If product inference is used, the corresponding ANFIS architec- parameters,
ture is shown in Fig. 2, where node functions in the same layers Layer 6: The only node in this layer is labeled as r, and it adds all
are of the type described below. This is an ANFIS architecture incoming signals to obtain the final inferred result for the whole
where the following meanings can be attached to each layer. system [6].
Table 2 shows values for using parameters in the ANFIS models.
Layer 1: Every node in this layer implies an input and it just The ANFIS model is trained by using approximately 80% of the
passes external signals to the next layer. data from the LTPP IMS database by Ozbay and Laub [8] and tested
Layer 2: Every node in this layer acts as a membership function using the rest of the data. The number of membership functions for
lAj ðxi Þ;and its output specifies the degree to which the given xi each input of ANFIS is set to 3. Prior to execution of the model,
i
satisfies the quantifier Aji . Generally, lAj ðxi Þ is selected as bell- standardization, x1i , on the data, Xi (i = 1, 2, . . ., n) is done according
i
shaped with a maximum membership degree equal to 1 and to the following expression such that all data values fall between
minimum equal to zero, such as zero and 1.

Fig. 5. A typical input–output surface between SUM ESAL, AGE, and IRI.
64 S. Terzi / Advances in Engineering Software 57 (2013) 59–64

The high success rate with one of these models has estimated
the value of the IRI.
As a result, SUM ESAL, SN, and AGE using the data of the ANFIS
method gave a predictably high success rate with the value of the
IRI. Many highway agencies have the necessary equipment for the
measurement of the IRI but are unable to find enough personnel
and time to do the measurements. With this model, the current
state of pavement performance can be estimated at the network le-
vel as well as foreseeing future pavement condition, and mainte-
nance scheduling can be made accordingly.

Acknowledgments
Fig. 6. Comparison of the ANFIS and target data for a testing set.
The data used in this study was obtained from the LTPP IMS
database by Ozbay and Laub [8].

x1i ¼ ðX i  X min Þ=ðX max  X min Þ ð11Þ References

where Xi is the actual value and Xmax and Xmin are the maximum [1] ASTM E867-06. Standard terminology relating to vehicle–pavement systems.
and minimum of the measurement values. Such standardization West Conshohocken, PA: ASTM International; 2003 doi: 10.1520/E0867-06.
procedure also turns the data into a dimensionless form. Further- [2] Bandara N, Gunartne M. Current and future pavement maintenance
prioritization based on rapid visual condition evaluation. J Transport Eng
more, standardization removes the arbitrary effects of similarity be-
2001;127(2):116–23.
tween objects or variables. Membership function types for inputs [3] Haas R, Hudson WR, Zaniewski J. Modern pavement management
are selected as Gauss-bell, whereas it is linear for the output. systems. USA: Krieger Publishing Company; 1994.
[4] Jang J-SR. ANFIS: adaptive-network-based fuzzy inference system. IEEE Trans
Syst Manage Cybern 1993;23(3):665–85.
3.3. Model results [5] Karan, MA, Haas, R, Walker, T. Illustration of pavement management: from
data inventory to priority analysis. Transportation Research Record
The model incorporates the effects of three input parameters; 1981;814:22–8.
[6] Lin CT, Lee CSG. Neural fuzzy systems. New Jersey: Prentice Hall PTR 797;
SUM ESAL, SN, and AGE, were used to simulate the International 1995.
Roughness Index (IRI). For this purpose, they were entered one [7] Sayers MW, Karamihas SM. The little book of profiling, UMTRI, 1995, 85 pM. In:
by one and brought together as has been established using the se- Sayers W, editor. On the calculation of IRI from longitudinal road profile’’.
Transportation Research Record 1501, Transportation Research Board, National
ven different models. Table 3 shows ANFIS models and their Research Council, Washington, DC; 1995, p. 1–12.
regression coefficients. As seen in the table, the high success rate [8] Ozbay K, Laub R. Models for pavement deterioration using LTPP. Washington
with the last of these models has estimated the value of the IRI. (DC): Federal Highway Administration, U.S. Department of Transportation
FHWA-NJ-1999-030; 2001.
The schematic of the architecture of ANFIS obtained the best [9] Rada GR, Perl J, Witczak W. Integrated model for project-level management of
regression coefficient shown in Fig. 3. Overall-input–output sur- flexible pavements. J Transport Eng 1985;112(4):381–99.
faces of SUM ESAL, SN, and AGE, input data were computed using [10] Saraf CL. Pavement condition rating system. Federal Highway Administration.
Report no. FHWA/OH-99/004; 1998.
IRI.
[11] Sayers MW. On the calculation of international roughness index from
The membership function plots of SUM ESAL, SN, AGE, and IRI longitudinal road profile. Transportation Research Record. Issue Number:
are shown in Fig. 4, respectively. A typical input–output surface 1501; 1995. ISSN: 0361-1981.
of the training phase is plotted in Fig 5. In the testing process, [12] Shahin MY. Pavement management for airports, roads, and parking
lots. London: Kluwer Academic Publishers; 2002.
the ANFIS model’s regression coefficient was obtained as 0.97 [13] Gillespie TD. Everything you always wanted to know about the IRI. But were
(Fig. 6). afraid to ask! In: Presented at the road profile users group meeting. Lincoln,
Nebraska, September 22–24; 1992.
[14] Tavakoli A, Lapin ML, Ludwig F. PMSC: pavement management system for
4. Conclusion small communities. J Transport Eng 1992;118(2):270–81.
[15] Tsoukalas, LH, Uhrig, RE. Fuzzy and Neural Approaches in Engineering. New
York: A Wiley-Interscience Publications, John Wiley & Sons, Inc; 1997. p. 587.
In this study, pavement performance indicators, the IRI, SUM,
[16] UMTRI. LTPP profiles. Transportation Research Institute, Ann Arbor: University
ESAL, SN and AGE data were estimated using the ANFIS method. of Michigan. <http://www.umtri.umich.edu/red/roughness>; 1998.
For this purpose, they were entered one by one and brought to-
gether as has been established using the seven different models.