Received: 13 May 2020 Revised: 19 September 2020 Accepted: 7 November 2020

DOI: 10.1002/stc.2671

RESEARCH ARTICLE

Determination of flexible pavement deterioration

conditions using Long-Term Pavement Performance
database and artificial intelligence-based finite element
model updating

Yong Deng1 | Xue Luo2 | Yazhou Zhang3 | Shengxin Cai1 | Kai Huang4 |
Xijun Shi5 | Robert L. Lytton1

1
Zachry Department of Civil and
Environmental Engineering, Texas A&M
Summary
University, College Station, TX, USA This paper aims to provide a methodology in determining the deterioration
2
College of Civil Engineering and conditions of flexible pavements using the Long-Term Pavement Performance
Architecture, Zhejiang University,
(LTPP) database and artificial intelligence (AI)-based finite element
Hangzhou, China
3
Structural Engineering Design
(FE) model updating. A new term quantifying the effects of the aging and load
Department of the Fourth Survey and repetitions on the modulus gradient of the asphalt layer was defined. The mod-
Design Branch, CCCC Second Highway ulus gradient change was captured by a two-step calibration process. The pro-
Consultants Co., Ltd, Wuhan, China
4
posed method combines the laboratory and field tests on the characterizations
Texas A&M Engineering Experiment
Station, Texas A&M University, College of the material properties and structural behaviors. Furthermore, it considers
Station, TX, USA the effects of the environmental and loading conditions on the pavement
5
Ingram School of Engineering, Texas behaviors and the gap between the laboratory and field tests on the same mate-
State University, San Marcos, TX, USA
rial characterizations. In this paper, the equivalent frequency in the asphalt
Correspondence layer for typical falling weight deflectometer (FWD) load was determined
Xue Luo, College of Civil Engineering and using the AI-based FE model updating as well. This paper extends the applica-
Architecture, Zhejiang University,
Hangzhou, China. tions of the FE model updating in the pavement structures and discusses the
Email: xueluo@zju.edu.cn performance of the modulus as an indicator of the pavement condition.

KEYWORDS
artificial intelligence, finite element model updating, flexible pavement, kriging model, modulus
gradient, nondestructive testing

1 | INTRODUCTION

Evaluating the pavement condition is a significant process for the pavement design, maintenance, and rehabilitation.
A comprehensive evaluation and a timely advanced warning can effectively avoid impending costly levels of future
distresses.1 In flexible pavements, the deterioration is a gradual process synthetically caused by the environment and
traffic. Two typical distresses are aging and cracking. Aging results from the oil volatilization, oxidation, and steric
hardening of asphalts due to the material exposure.2 The asphalt layer of the pavement becomes stiffer and brittler
with the increasing aging degree. Cracking, based on its contributing factors, can be divided into fatigue cracking
and thermal cracking, which are induced by the load repetitions and temperature cycles, respectively.3 Cracks initi-
ate and propagate in the pavement, which finally result in visible damages and material failures. Pavement

Struct Control Health Monit. 2020;e2671. wileyonlinelibrary.com/journal/stc © 2020 John Wiley & Sons Ltd 1 of 19
https://doi.org/10.1002/stc.2671
2 of 19 DENG ET AL.

deterioration not only reduces the service life of pavements but also affects the riding quality and driving safety.
Therefore, the evaluation and prediction of the pavement deterioration have become enduring topics in pavement
engineering.
Considering the variety of distress modes and the complexity of intrinsic mechanism, studies on the pavement
deterioration can be conducted in terms of individual distress modes and overall performance of the pavement. For
an individual distress mode, the pavement condition can be evaluated either from direct field measurements or from
a mechanistic-empirical (ME) model associating intrinsic mechanism, laboratory tests, and field performance.2,4–9 In
this direction, indicators representing the deterioration condition are specific. The crack depth, for example, is used
in the top–down cracking model to represent the current condition and predict the remaining service life of the
pavement. On the other side, the pavement performance can also be used in the frames of deterioration determina-
tion. Field testing provides much convenience to obtain indirect indicators such as the deflection, friction, and
roughness.10,11
Modulus or modulus-associated factors are commonly used indicators of pavement conditions. As a direct indicator
of the material strength, modulus occurs in many ME models to predict the total number of load repetitions a pavement
can bear.3,12 Deflection, directly controlled by the modulus, is used to calculate indices determining the pavement con-
dition such as the structural number (SN),13,14 the structural condition index (SCI),10 the structural index of a pavement
(SIP),15 and the deflection slope index (DSI).16 The development of nondestructive testing (NDT) devices such as the
falling weight deflectometer (FWD), the rolling wheel deflectometer (RWD), and the traffic speed deflectometer (TSD)
contributes to the obtain and applications of these indices. Pavement responses at critical locations, which consider
both the mechanism and material strength, are used in most ME models for individual distress modes.3 The tensile
strain at the surface of the asphalt layer, for example, is for the top–down cracking while the tensile strain at the bottom
is for the bottom-up cracking.
Different from rigid pavements, the modulus of surface layer (asphalt layer) materials of flexible pavements is
greatly affected by the temperature and loading frequency. Their relationships have been studied comprehensively
in the laboratory tests17 and considered increasingly in field pavements.18,19 Compared with the loading frequency,
effects of the temperature are easier to be implemented into models since the temperature change with time and
pavement layer depth has been characterized thoroughly.20 The temperature can be used to calibrate the modulus7
or serve as an explicit parameter in ME models.6,8 In contrast, traffic loads in these models are typically character-
ized in terms of the axle number and axle weight. It is practical and understandable since the speed variation is
marginal compared with the load magnitude, especially with speed limits in highways. However, for those
structural indices obtained from NDT devices and used for the pavement monitoring over time, effects of the
temperature and loading frequency should be considered equally and characterized accurately for better pavement
evaluations.

2 | MOTIVATION AND O BJECTIVES

The main purpose of this paper is to provide a methodology in determining the deterioration condition of flexible
pavements considering effects of the

• temperature gradient in the asphalt layer;

• loading frequency distribution in the asphalt layer; and
• difference between material properties obtained from laboratory and field tests.

Taking advantages of the rich and timely information stored in the LTPP database such as the pavement structure and
construction, layer properties, climatic conditions, and pavement performance, the deterioration degree of the asphalt
layer can be obtained from the modulus change over time. The deterioration is assumed to result from the material
aging and load-induced fatigue damage. Finite element (FE) simulations of NDT and FE model updating with AI
algorithms were applied in this paper as well. This paper is organized as follows. Information retrieved from the LTPP
database and used in this paper are introduced in Section 3. FE model updating and its applications in determining the
equivalent loading frequencies and layer moduli are described in Section 4. Definition and analysis of the average
deterioration degree of the pavement are presented in Section 5. Conclusions in this paper and recommendations for
future researches are summarized in Section 6.
DENG ET AL. 3 of 19

3 | I N F O R M A T I O N I N L T P P DA T A B A S E

The basic ideas of this paper are simulating pavements in NDT and calibrating model parameters based on the compari-
son between the simulated and measured pavement responses. Two pavement sections 48-0117 and 48-0167 in the
LTPP database were selected for this study; “48” refers to the state of Texas, and “0117” and “0167” are
section identification numbers. The two sections locate near the U.S. Highway 281 in Texas and belong to the pavement
class “Rural Principal Arterial.” Among the total 2,581 pavement sections across North America, information on the
material properties, constructions, environmental conditions, traffic conditions and pavement performance with the
service time of these two pavement sections are recorded in the LTPP database to provide a long-term and comprehen-
sive system for the pavement evaluation and design. Table 1 shows paths to retrieve available information in the LTPP
database to build FE models, and Figure 1 shows structural configurations and properties of supporting layers except
for the modulus of selected pavement sections. According to the record, such layer settings were maintained until the
first reconstruction occurred.
The dynamic modulus of asphalt concrete (AC) samples with original binders is assumed as undamaged material
properties and reference modulus in this study. In the LTPP database, the dynamic moduli predicted in the laboratory
at the loading frequencies 0.1, 0.5, 1, 5, 10, and 25 Hz and temperatures −10
C, 4.4
C, 21.1
C, 37.8
C, and 54.4
C are
provided. As for the modulus at an arbitrary loading frequency and temperature, the master curve shown in Equation 1
was constructed following the procedures presented in a previous research.21

α
logðjE  ð f r ÞjÞ = δ + ð1Þ
1 + e η− γlogð f r Þ

in which

f r = aT f ð2Þ

logðaT Þ = aT 2 + bT + c ð3Þ

TABLE 1 Summary of information retrieved from the LTPP database

Type Location Information

Structure Pavement structure and Layer type; Layer thickness; Layer order;
construction! Construction type
Pavement layer type and
thickness!
Representative pavement structure
Material Pavement structure and Dynamic modulus of AC mixtures with original binders at different loading
construction! frequencies and temperatures
Material-layer properties and field
sampling!
AC mixture properties (calculated
ESTAR)!
Calculated ESTAR modulus
Climate Performance! Temperatures within the AC layer on the dates/times of field testing
Backcalculation and deflection!
Deflection!
Temperature data
NDT Performance! Load (pressure) magnitudes of the FWD;
results Backcalculation and deflection! Deflection peaks at sensor locations
Deflection!
Deflection raw data
4 of 19 DENG ET AL.

F I G U R E 1 Pavement structural
parameters and material properties

where E*(fr) = the dynamic modulus at the reduced frequency fr in Hz; f = the loading frequency in Hz; T = the temper-
ature in kelvin; aT = the time–temperature shift factor; and δ, α, η, γ, a, b, c = fitting parameters.
According to the record, two AC layers were constructed separately in each pavement section. Figure 2a,b shows
master curves of two AC layers in two pavement sections at the reference temperature using Equations 1–3. It can be
seen that two separated AC layers were made of identical materials in both pavement sections. Therefore, two AC
layers were combined in the corresponding FE models for simplicity.
Table 2 shows one example set of climatic data and FWD test results for each pavement section from the LTPP data-
base and supplementary materials.22 In reality, FWD tests were conducted at different locations of the pavement sur-
face. Moreover, FWD tests with different pressure magnitudes were conducted at each testing location, which was
achieved by adjusting the drop height of the loading plate. In the transverse direction (perpendicular to the vehicle
movement direction on the pavement surface), the location “Flexible pavement outer wheel path” was selected to fully
consider the fatigue damage induced by traffic loads. In the longitudinal direction (vehicle movement direction), ran-
dom locations were selected for two pavement sections of which the “POINT_LOC” (point location) were "0" and
"137.2" for the section 48-0117 and 48-0167, respectively. The numbers stand for the distances (in meter) to the starting
points of the pavement sections and random locations were selected for the insignificant deterioration variation along
this direction. Among three different pressure magnitudes of the FWD, the Level 2 with the nominal pressure magni-
tude 570 kPa was selected. The main reason is that in the LTPP database, there are no explicit records of the loading
duration for the FWD tests. However, the drop height of the loading plate does affect the pressure magnitude as well as
the loading duration, in another word, the loading frequency.23 It is reasonable to select 0.03 s as the loading duration
for pressure magnitudes of Level 2, which is generally acknowledged and applied in most studies.24

4 | F E MO D EL U P D ATI N G

In this study, the FE model updating was used in determinations of the equivalent frequency of the FWD load in the
AC layer and moduli of pavement layers. The equivalent frequency in the AC layer was proposed for the sensitivity of
AC to the loading frequency21 and nonuniform predominant frequencies observed in a previous research.19 In addition
DENG ET AL. 5 of 19

F I G U R E 2 Comparison of master curves

of two layers at the reference temperature

to directly using viscoelastic models to characterize AC layers in analytical and numerical flexible pavement
models,1,24–27 the equivalent frequency can contribute to accurate flexible pavement responses. It has been found that
the equivalent frequency of a moving load applied on the pavement surface depends on the pavement layer depth,
moving speed, environmental conditions, and pavement responses.19 In this study, such concept was expanded to
stationary dynamic loads, for example, the FWD load. Moreover, a determined equivalent frequency can be applied in
Equation 1 for the magnitude of the dynamic load, which is a more straightforward indicator of the material strength
than viscoelastic model parameters such as the Prony series coefficients. Current procedures to determine the equiva-
lent frequency are selecting a certain pavement response, characterizing the pulse time of the response, and calculating
the corresponding loading frequency via an inverse process.19,28
As for moduli of pavement layers, the backcalculation of layer properties from pavement responses is another endur-
ing topic in pavement engineering. Researchers have been making continuous developments in areas such as analytical
and numerical models of pavements, techniques of data measuring and processing, and optimization methods for
parameter backcalculation. Pavement models are responsible for connecting pavement parameters to be backcalculated
and pavement responses measured from either sensors in instrumented pavements or NDT devices. Typical analytical
and numerical flexible pavement models include elastic, quasi-elastic and viscoelastic layered systems,25,29–32 and FE
models.33,34 Optimization methods aim in inversely and iteratively solving and calibrating equation systems31,35 and
data-driven mathematical models,33,34 which are provided by analytical and numerical pavement models, respectively.
 6 of 19

TABLE 2 Climate and NDT results retrieved from the LTPP database

Sensor distance to the loading center (Line 1) (mm) and deflection peak
(Line 2) (mm)
Testing Layer Temperature, Pressure magnitude of the


Section date depth, mm C FWD, kPa 1 2 3 4 5 6 7 8
48-0117 1999/8/18 25.4 57.2 574 0 203 305 457 610 914 1,524
190.5 45.6 0.194 0.109 0.076 0.058 0.049 0.036 0.023
2000/5/9 25.4 30.0 548 0 203 305 457 610 914 1,219 1,524
190.5 32.0 0.105 0.072 0.058 0.047 0.040 0.030 0.025 0.020
2002/2/12 25.4 27.4 598 0 203 305 457 610 914 1,219 1,524
188.0 21.0 0.091 0.070 0.062 0.054 0.047 0.036 0.029 0.023
48-0167 1999/8/19 25.4 54.4 567 0 203 305 457 610 914 1,524
63.5 51.9 0.305 0.177 0.124 0.082 0.058 0.037 0.022
2000/5/9 25.4 46.8 532 0 203 305 457 610 914 1,219 1,524
63.5 42.4 0.239 0.149 0.113 0.078 0.056 0.036 0.028 0.022
2002/2/13 25.4 21.8 583 0 203 305 457 610 914 1,219 1,524
63.5 21.1 0.223 0.182 0.151 0.112 0.081 0.048 0.031 0.023
DENG ET AL.
DENG ET AL. 7 of 19

FIGURE 3 Flow chart of the FE model updating

Generally, the method using FE models in the inverse analysis such as calibrating structural and material parame-
ters of the infrastructure from its responses is called FE model updating. The FE model updating has wide applications
in the structural health monitoring (SHM) and structural optimization design, especially when it was implemented with
AI algorithms as its optimization (or updating) methods.36–38 Such a fast and targeted calibration process contributes to
an accurate pinpoint of the weakest location of a structure or a timely evaluation of the deterioration condition of a
structure.
The reasons to select FE model updating in determining both the equivalent frequency of the FWD load in the AC
layer and moduli of pavement layers can be summarized as follows. First, equivalent frequencies generated from
previous studies cannot be applied in this study since they were supposed to match their own pavement structures,
pavement responses, and so forth. Accidentally, the determination of the equivalent frequency (model input) from
pavement responses (model output) is also a typical inverse problem as the layer moduli calibration. Besides, it is preva-
iling to use FE models in simulating and analyzing pavements in NDT tests for their accuracy,1 convenience of coupling
and adjusting environmental, loading and material models,21,39 and low expenses. Moreover, a comparison was
conducted in a previous study34 which showed that pavement parameters backcalculated from AI-based FE model
updating were more accurate and representative than those from traditional backcalculation methods using analytical
pavement models.
Figure 3 shows basic procedures of the FE model updating used in this study. Details of the kriging model, particle
swarm optimization (PSO) algorithm, model inputs, and outputs are introduced in the following sections.

4.1 | FE models of FWD tests

A precise FE model is a necessity for the model updating. In this study, 2-D axisymmetric models were selected to
simulate pavements in FWD tests of which the loading configuration can be totally replicated.24 In addition to those
listed in Figure 1 and Table 2, sensitivity analysis is required to determine model parameters such as the radius, mesh
size, and depth of subgrade.
Following previous researches,24,26 the purpose of the sensitivity analysis is to strike a balance between the accuracy
and computational time. For example, the subgrade depth should be increased gradually until model outputs reach a
convergence. Figure 4 shows the 2-D axisymmetric FE model of Section 48-0167 built in the software ABAQUS40 and
results of the sensitivity analysis to determine the subgrade depth. In the sensitivity analysis, the pressure magnitude
was 565 kPa, and layer moduli were averaged backcalculation moduli recorded in the LTPP database. The vertical
deflection at the loading center was selected as the model output. The model outputs were compared in the models with
different subgrade depths and element types in which the infinite element was typically applied at boundaries to reduce
the element number without losing significant accuracy.24 It can be seen from Figure 4b that variations of results
caused by the subgrade depth decreased as the subgrade depth increased. The differences between the linear element,
quadratic element, and infinite element dropped below 0.5% when the subgrade depth was 7,238 mm. Finally, the
8 of 19 DENG ET AL.

FIGURE 4 2-D axisymmetric FE model

model radius, subgrade depth, and element number were 5.0 m, 7,238 mm and 4,200 for Section 48-0167. The numbers
were 5.0 m, 7,315.2 mm and 5,760 for Section 48-0117.
As described above, one goal of the FE model updating is to calibrate layer moduli based on the measured pavement
responses in the FWD tests, which are the deflection peaks at sensor locations in Table 2. A multistep calibration pro-
cess was proposed based on different sensitivities of surface vertical deflections to layer moduli according to their dis-
tances to the loading center.34 Figure 5 shows the results of a similar sensitivity analysis for Section 48-0117. In each
case, the modulus of one layer was adjusted, and the surface vertical deflection at each sensor location was recorded
and compared with the original case. It can be seen that as the distance to the loading center becomed greater, the sur-
face vertical deflection was more affected by the lower layers. The extreme cases were that the vertical deflection at the
loading center was sensitive to all layer moduli, and the vertical deflection at Sensor 8 depended merely on the
DENG ET AL. 9 of 19

FIGURE 5 Sensitivity analysis of layer moduli on surface deflections in Section 48-0117

subgrade modulus. In pavements, different sensitivities were mostly considered in the system identification (SID)
inverse process in terms of the sensitivity matrix.35 In this study, different sensitivities were explicitly reflected in a
multistep calibration process, which calibrated layer moduli based on corresponding vertical deflections. According to
the results in Figure 5, vertical deflections at Sensors 4–8 were used to calibrate moduli of supporting layers first.
Vertical deflections at Sensors 1–3 along with the calibrated supporting layer moduli were used to calibrate the modulus
gradient of the AC layer then.

4.2 | Kriging model

The kriging model served as a surrogate model in the FE model updating. It is a neat alternative to analytical solutions
for those fuzzy problems such as the equivalent frequency of the FWD load in the AC layer. Moreover, the introduction
of surrogate models saved large computational cost and time than the traditional FE model updating.41 As a data-driven
model, it can efficiently connect implicitly related parameters and predict results via multiple linearly combined regres-
sion models and a stochastic process, which considers the residuals between the regression models and observations.42
The general form of the kriging model is presented as Equation 4,

X
N
^ ð XÞ =
Y β i f i ð XÞ + Z ð XÞ ð4Þ
i=1
10 of 19 DENG ET AL.

where X = model inputs; Y ^ = predicted outputs; f = regression functions; β = model coefficients; and Z = a model of a
Gaussian and stationary random process with zero mean. The kriging model is popular in geostatistics and spatial
statistics, which sketch a domain from sampled data. Similarly, we assumed the kriging model trained with data which
were sampled from the solution domain can be used for prediction. The sampling method used in this study was Latin
hypercube sampling (LHS), which was first proposed by McKay in 1979.43 It proved to be a good method to select
representative input values in computer experiments since it considers the total number of samples and their relative
positions to avoid data concentration as random sampling.37,43

4.3 | Particle swarm optimization

With a trained kriging model, the next step for an inverse problem is to find the corresponding inputs based on the
given outputs. AI algorithms are able to provide an efficient solution search from an approaching process rather than
directly solving Equation 4. Among various AI algorithms, the PSO algorithm has advantages of the simple concept,
easy implementation, and short computational time.34 Potential inputs are simulated as individuals in a swarm. They
have the abilities of exchanging information, which is the quality of the output and adjusting their moving directions
and distances in the domain for the optimal location, which is the solution. The mathematical expressions are
presented in Equations 5 and 6,44

xki + 1 = xki + v ki + 1 ð5Þ

   k k
v ki + 1 = wv ki + c1 r k1,i pki −xki + c2 r 2,i
k
gi −xi ð6Þ

where xki and v ki = the position vector and velocity vector of the ith particle in the kth iteration; pki = the best position
of the ith particle in the past k iterations; gki = the best position of all particles in the past k iterations; w = the inertial
weight; c1 and c2 = the cognition and social coefficients; and r k1,i and r k2,i = random numbers in the interval [0,1] for the
ith particle in the kth iteration. Each type of inputs is contained in the position vector as an individual variable. The
distance and direction of each particle's movement are contained in the velocity vector, which depends on its previous
velocity vector and qualities of historical locations of the swarm.
To evaluate the quality of each particle's location, an objective function was used as Equation 7,

k y −y y^ k
2
F ðx Þ = ð7Þ

where y = actual outputs in the position x; y^ = predicted outputs in the position x by the kriging model.

4.4 | Implementation of FE model updating

A step-by-step procedure for all FE model updating in this study is presented as follows.

1. Obtain the equivalent frequency of the FWD load in the AC layer.

1.1 Use the FE model built in Section 4.1 in which the supporting layers are elastic and the AC layer is viscoelastic
characterized by Equation 1. Apply the FWD load with the duration 0.03 s and extract vertical deflections under
the loading center along the AC layer, which are “Measured Outputs” in Figure 3;
1.2 Change the material model of the AC layer from Equations 1 to 8, which is elastic but can characterize the
modulus variation with depth. The model coefficients Eb, Es, and n are “Inputs” in Figure 3;
1.3 Train the kriging model with sampled inputs and corresponding outputs obtained from the FE model in Step
1.2. With the trained kriging model and “Measured Outputs”, find “Calibrated Inputs” via the PSO algorithm;
1.4 From “Calibrated Inputs” and the master curve expressed in Equation 1, calculate the corresponding frequencies
along the AC layer, which form the equivalent frequency of the FWD load for this pavement section.
DENG ET AL. 11 of 19

2. Obtain the moduli of supporting layers.

2.1 Use the FE model built in Section 4.1 in which the modulus of the AC layer is calculated from Equation 1 with
temperatures in Table 2 and frequencies from Step 1. Apply the FWD load with the duration 0.03 s and pressure
magnitude in Table 2;
2.2 Treat supporting layer moduli as inputs and vertical deflections at Sensors 4–8 as outputs. Train the kriging
model with sampled inputs and corresponding outputs from the FE models in Step 2.1.
2.3 Calibrate supporting layer moduli with the trained kriging model and “Measure Outputs” in Table 2.
3. Obtain the modulus of the deteriorated AC layer.
3.1 Use the FE model built in Section 4.1 in which the moduli of supporting layers are obtained from Step 2 and
the modulus model of the AC layer is Equation 8. Apply the FWD load with the duration 0.03 s and pressure
magnitude in Table 2;
3.2 Treat model coefficients Eb, Es, and n as inputs and vertical deflections at Sensors 1–3 as outputs. Train the
kriging model with sampled inputs and corresponding outputs from the FE models in Step 3.1.
3.3 Calibrate model coefficients of the AC layer modulus with the trained kriging model and “Measure Outputs” in
Table 2.

 n
d−z
E ðzÞ = E b + ðE s −E b Þ ð8Þ
d

where E(z) = the modulus at depth z; Eb and Es = the modulus at the bottom and top of the layer; d = the thickness of
the layer; and n = the model parameter representing the shape of the gradient.
The deterioration degree of the pavement was determined from the comparison between the moduli of the AC layer
obtained from Steps 1 and 3, which came from undamaged and damaged pavements. Details of the method are
presented in the next section.

5 | R ES U L T S A N D D I S C U S S I O N

In this section, results obtained from the previous section such as the equivalent frequency of the FWD load and layer
moduli are presented and validated. The average deterioration degree of the pavement is introduced and discussed.

5.1 | Equivalent frequency of the FWD load

The vertical deflection was selected as the indicator for the equivalency validation since it was used for the modulus
calibration in this study.
Figure 6 shows obtained frequencies at 25
C in two pavement sections. Obvious gradients can be observed with
the depth in both sections, and the descending trend is similar with the previous research.19 Specifically, the gradi-
ent varies with the pavement structure, and equivalent frequencies in Section 48-0117 are significantly different
from the commonly used value 10 Hz. A more straightforward comparison was conducted and presented in
Figure 7 in which obtained frequencies, 10 Hz and 33 Hz (1/T), were used, and vertical deflections were compared.
It can be seen that cases with the equivalent frequencies have the optimal accuracy. Moreover, it indicates any
constant value cannot replace loading frequencies of the FWD load in the AC layer, at least in terms of the vertical
deflection.
Furthermore, equivalent frequencies obtained at 25
C were applied in different temperatures. Temperatures 40
C
and 60
C with frequencies in Figure 6 were substituted into Equation 1, and vertical deflections were compared with
cases using viscoelastic moduli at corresponding temperatures. One major conclusion can be drawn from Figure 8 that
the equivalent frequency for a certain load is independent of the temperature. Therefore, it is reasonable and practical
to apply the equivalent frequency obtained at one temperature to any other temperature or temperature gradient, which
is exactly the method in this study.
12 of 19 DENG ET AL.

FIGURE 6 Equivalent frequency of the FWD load at 25
C

F I G U R E 7 Comparison of the equivalent

frequency with constant frequencies
DENG ET AL. 13 of 19

F I G U R E 8 Application of the equivalent

frequency in different temperatures

5.2 | Moduli of supporting layers

Values of supporting layer moduli are out of interest in this study. Main contents of this section are validations of the
FE model updating and two-step calibration method.
Desirable performance of the kriging model and PSO algorithm can be seen from Figure 9a,c. Measured vertical
deflections at Sensors 4–8 were compared with those predicted by the FE models in which supporting layer moduli
were calibrated by the FE model updating. Comparisons between Figure 9a,b and Figure 9c,d validate the two-step
calibration method. The only difference between FE models in Steps 2 and 3 is the modulus of the AC layer. Adjust-
ments in this layer do not affect deflections at Sensors 4–8 much as indicated by Figure 9.

5.3 | Modulus of the AC layer

The modulus of the AC layer used in Step 2 was illustrated in Figure 2. Figure 2 is the master curve constructed from
laboratory tests on asphalt mixtures with original binders. These undamaged moduli were supposed to result in
14 of 19 DENG ET AL.

FIGURE 9 Validation of supporting layer moduli from the FE model updating

predictions with lower accuracy, which can be observed in Figure 10. The deterioration of the AC layer reflected on the
modulus and then on the surface deflection. The FE model updating took advantages of these variations in deflections
and calibrated to damaged moduli.

5.4 | Average deterioration degree

The average deterioration degree was determined from the modulus change in the AC layer between undamaged and
damaged moduli. Figure 11 shows modulus gradients of undamaged and damaged AC layers under the FWD load and
field temperatures. With effects of the temperature and frequency, it is meaningless to directly use or compare these
values. The average deterioration degree was defined in Equation 9, which averages deterioration degrees within the
AC layer,

z=d
1 E d ðzÞ −E u ðzÞ
φ= dz × 100% ð9Þ
d ð z=0 E u ðz Þ
DENG ET AL. 15 of 19

FIGURE 10 Validation of AC layer moduli from the FE model

updating

where Ed(z) = the damaged modulus at depth z; Eu(z) = the undamaged modulus at depth z; and d = the thickness of
the layer. The average deterioration degree

• considers opposite effects of the aging and load-induced fatigue damage on the modulus by considering and keeping
the sign of the difference between the damaged and undamaged moduli and
• eliminates effects of the temperature and frequency on the modulus by dividing the undamaged modulus for the
relative difference.

Additionally, the undamaged modulus was calculated from Equation 1, which works for the laboratory tests, while the
damaged modulus was directly calibrated from the field tests. Differences between the material properties from the
laboratory and field tests30 can also be eliminated by using the relative difference as Equation 9.
Figure 12 shows the average deterioration degrees of two pavement sections at different service times. Overall, the
major deterioration resulted from the load repetitions, which reduced the modulus of the AC layer. Meanwhile,
the aging effects can still be explicitly detected from the average deterioration degree of Section 48-0167 in 1999 and the
difference between average deterioration degrees of Section 48-0117 in 2000 and 2002. Literatures introduced in
Section 1 show the modulus and associated terms are popular in the pavement evaluation and management. However,
results in Figure 12a indicate the modulus change with the service time is not always consistent. The modulus or associ-
ated terms as indicators seem not enough for a comprehensive evaluation of flexible pavements. Considering the
limited case number in this study; a more general relationship between the modulus and deterioration degree of the
AC layer is required in the future.
16 of 19 DENG ET AL.

FIGURE 11 Comparison of undamaged and damaged AC layer moduli

FIGURE 12 Average deterioration degree with service time

DENG ET AL. 17 of 19

6 | C ONCLUSIONS A ND F UTURE WORK

This paper applied the FE model updating in calibrating layer moduli of damaged flexible pavements and determined
their deterioration degrees at different service times. The conclusions drawn from this paper and recommendations for
future work are listed as below.

• The FE model updating with the kriging model and PSO algorithm had desirable performance in the determination
of the equivalent frequency of the FWD load and calibrations of layer moduli;
• The equivalent frequency of the FWD load decreases with the layer depth. For a certain dynamic/moving load, the
equivalent frequency in the AC layer is temperature-independent and structure-dependent;
• A multistep calibration method was proposed and applied successfully based on different sensitivities of vertical
deflections at sensor locations to layer moduli;
• The average deterioration degree defined from undamaged and damaged moduli can capture effects of the aging and
fatigue damage on the modulus. It also addresses the strength and weakness of the modulus as an indicator of the
pavement condition.

Several directions are recommended for future researches. The FE model updating with AI algorithms has potentials in
solving other problems in pavement materials and structures in which such applications are limited so far. Meanwhile,
the average deterioration degree can be applied in more pavement sections for general correlations with other indica-
tors. Moreover, effects of the aging and fatigue damage on flexible pavements can be separated with the aid of the
method in this study and those from previous researches.2,6,45

ORCID
Xue Luo https://orcid.org/0000-0002-0181-874X

R EF E RE N C E S
1. Deng Y, Luo X, Gu F, Zhang Y, Lytton RL. 3D simulation of deflection basin of pavements under high-speed moving loads. Construct
Build Mater. 2019;226:868-878.
2. Luo X, Gu F, Lytton RL. Kinetics-based aging prediction of asphalt mixtures using field deflection data. International Journal of Pave-
ment Engineering. 2019;20(3):287-297.
3. ARA-ERES (Applied Research Associates, Inc., ERES Consultants Division). Guide for mechanistic–empirical design of new and reha-
bilitated pavement structures, Final report, NCHRP Project 1-37A, Transportation Research Board, Washington, D.C. 2004.
4. Zhou F, Newcomb D, Gurganus C, et al. Experimental design for field validation of laboratory tests to assess cracking resistance of
asphalt mixtures. No. NCHRP Project 9–57, National Cooperative Highway Research Program, Transportation Research Board of the
National Academies, Washington, D.C. 2016.
5. Lytton RL, Luo X, Ling M, Chen Y, Hu S, Gu F. A mechanistic–empirical model for top–down cracking of asphalt pavements layers,
No. NCHRP Project 01–52, National Cooperative Highway Research Program, Transportation Research Board of the National Acade-
mies, Washington, D.C. 2018.
6. Luo X, Gu F, Zhang Y, Lytton RL, Birgisson B. Kinetics-based aging evaluation of in-service recycled asphalt pavement. J Clean Prod.
2018;200:934-944.
7. Ling M, Chen Y, Hu S, Luo X, Lytton RL. Enhanced model for thermally induced transverse cracking of asphalt pavements. Construct
Build Mater. 2019;206:130-139.
8. Ling M, Luo X, Chen Y, Hu S, Lytton RL. A calibrated mechanics-based model for top-down cracking of asphalt pavements. Construct
Build Mater. 2019;208:102-112.
9. Ling M, Luo X, Chen Y, Gu F, Lytton RL. Mechanistic-empirical models for top-down cracking initiation of asphalt pavements. Interna-
tional Journal of Pavement Engineering. 2020;21(4):464-473.
10. Elbagalati O, Elseifi MA, Gaspard K, Zhang Z. Prediction of in-service pavement structural capacity based on traffic-speed deflection
Measurements. Journal of Transportation Engineering. 2016;142(11):04016058. https://doi.org/10.1061/(asce)te.1943-5436.0000891
11. Susanna A, Crispino M, Giustozzi F, Toraldo E. Deterioration trends of asphalt pavement friction and roughness from medium-term sur-
veys on major Italian roads. International Journal of Pavement Research and Technology. 2017;10(5):421-433.
12. Sadeghi JM, Fathali M. Deterioration analysis of flexible pavements under overweight vehicles. Journal of Transportation Engineering.
2007;133(11):625-633.
13. Martin TC. Experimental estimation of the relative deterioration of flexible pavements under increased axle loads. International Journal
of Pavement Engineering. 2011;12(1):37-45.
14. Crook AL, Montgomery SR, Guthrie WS. Use of falling weight deflectometer data for network-level flexible pavement management.
Transportation Research Record. 2012;2304(1):75-85.
18 of 19 DENG ET AL.

15. Zhang Z, Manuel L, Damnajanovic I, Li Z. Development of a new methodology for characterizing pavement structural condition for
network level applications. Report No. FHWA/TX-04/0-4322-1, Center for Transportation Research, University of Texas at Austin,
Austin, TX, 2003.
16. Nasimifar M, Thyagarajan S, Siddharthan RV, Sivaneswaran N. Robust Deflection indices from traffic-speed deflectometer measure-
ments to predict critical pavement responses for network-level pavement management system application. Journal of Transportation
Engineering. 2016;142(3):04016004. https://doi.org/10.1061/(asce)te.1943-5436.0000832
17. Bonaquist R. Refining the simple performance tester for use in routine practice. No. NCHRP Project 9-29, National Cooperative Highway
Research Program, Transportation Research Board of the National Academies, Washington, D.C. 2008.
18. Nazarian S, Alvarado G. Impact of temperature gradient on modulus of asphaltic concrete layers. Journal of Materials in Civil Engineer-
ing. 2006;18(4):492-499.
19. Ulloa A, Hajj EY, Siddharthan RV, Sebaaly PE. Equivalent loading frequencies for dynamic analysis of asphalt pavements. Journal of
Materials in Civil Engineering. 2013;25(9):1162-1170.
20. Han R, Jin X, Glover CJ. Modeling pavement temperature for use in binder oxidation models and pavement performance prediction.
Journal of Materials in Civil Engineering. 2011;23(4):351-359.
21. Zhang Y, Birgisson B, Lytton RL. Weak form equation–based finite-element modeling of viscoelastic asphalt mixtures. Journal of
Materials in Civil Engineering. 2016;28(2):04015115. https://doi.org/10.1061/(asce)mt.1943-5533.0001395
22. Schmalzer PN. LTPP manual for falling weight deflectometer measurements, version 4.1. Report No. FHWA-HRT-06-132, Federal
Highway Administration, US Department of Transportation, 2006.
23. Sebaaly BE, Mamlouk MS, Davies TG. Dynamic analysis of falling weight deflectometer data. Transportation Research Record. 1986;
1070:63-68.
24. Wang H, Li M. Comparative study of asphalt pavement responses under fwd and moving vehicular loading. Journal of Transportation
Engineering. 2016;142(12):04016069. https://doi.org/10.1061/(asce)te.1943-5436.0000902
25. Fernando EG, Liu W. User's manual for pavement dynamic back-calculation procedure with systems identification method (DBSID).
College Station: Texas Transportation Institute, Texas A&M University; 1999.
26. Deng Y, Luo X, Zhang Y, Lytton RL. Evaluation of flexible pavement deterioration conditions using deflection profiles under moving
loads. Transportation Geotechnics. 2020;100434. https://doi.org/10.1016/j.trgeo.2020.100434
27. Ling M, Deng Y, Zhang Y, Luo X, Lytton RL. Evaluation of complex Poisson’s ratio of aged asphalt field cores using direct tension test
and finite element simulation. Construction and Building Materials. 2020;261:120329. https://doi.org/10.1016/j.conbuildmat.2020.120329
28. Al-Qadi IL, Xie W, Elseifi MA. Frequency determination from vehicular loading time pulse to predict appropriate complex modulus in
MEPDG. Asphalt Paving Technology-Proceedings. 2008;77:739–772.
29. Huang YH. Pavement analysis and design. 2nd ed. Upper Saddle River, NJ: Pearson; 2003.
30. Kutay ME, Chatti K, Lei L. Backcalculation of dynamic modulus mastercurve from falling weight deflectometer surface deflections.
Transportation Research Record. 2011;2227(1):87-96.
31. Ma X, Dong Z, Yu X, Chen F, Cao C, Sun J. Monitoring the structural capacity of airfield pavement with built-in sensors and modulus
back-calculation algorithm. Construct Build Mater. 2018;175:552-561.
32. Ma X, Dong Z, Chen F, Xiang H, Cao C, Sun J. Airport asphalt pavement health monitoring system for mechanical model updating and
distress evaluation under realistic random aircraft loads. Construct Build Mater. 2019;226:227-237.
33. Li M, Wang H. Development of ANN-GA program for backcalculation of pavement moduli under FWD testing with viscoelastic and
nonlinear parameters. International Journal of Pavement Engineering. 2019;20(4):490-498.
34. Deng Y, Luo X, Zhang Y, Lytton RL. Determination of complex modulus gradients of flexible pavements using falling weight
deflectometer and artificial intelligence. Materials and Structures. 2020;53(4):1-17.
35. Lytton RL. Backcalculation of pavement layer properties. In: STP 1026. In: Nondestructive testing of pavements and backcalculation of
moduli. Philadelphia: ASTM; 1989:7-38.
36. Cao H, Qian X, Chen Z, Zhu H. Layout and size optimization of suspension bridges based on coupled modelling approach and enhanced
particle swarm optimization. Eng Struct. 2017;146:170-183.
37. Qin S, Zhang Y, Zhou Y, Kang J. Dynamic model updating for bridge structures using the kriging model and pso algorithm ensemble
with higher vibration modes. Sensors. 2018;18(6):1879. https://doi.org/10.3390/s18061879
38. An Y, Chatzi E, Sim S, Laflamme S, Blachowski B, Ou J. Recent progress and future trends on damage identification methods for bridge
structures. Structural Control and Health Monitoring. 2019;26(10). https://doi.org/10.1002/stc.2416
39. Dong Z, Ma X, Gong X, Oeser M. Theoretical evaluation of the measurement accuracy of fiber Bragg grating strain sensors within
randomly filled asphalt mixtures based on finite element simulation. Structural Control and Health Monitoring. 2018;25(1):e2057.
https://doi.org/10.1002/stc.2057
40. ABAQUS version 6.12 [Computer software]. Dassault Systemes, Waltham, MA, 2012.
41. Xia Q, Xia Y, Wan H, Zhang J, Ren W. Condition analysis of expansion joints of a long-span suspension bridge through
metamodel-based model updating considering thermal effect. Structural Control and Health Monitoring. 2020;27(5). https://doi.org/10.
1002/stc.2521
42. Martin JD, Simpson TW. Use of kriging models to approximate deterministic computer models. AIAA Journal. 2005;43(4):853-863.
43. McKay MD, Beckman RJ, Conover WJ. Comparison of three methods for selecting values of input variables in the analysis of output
from a computer code. Dent Tech. 1979;21(2):239-245.
DENG ET AL. 19 of 19

44. Shabbir F, Omenzetter P. Particle swarm optimization with sequential niche technique for dynamic finite element model updating.
Comput Aided Civ Inf Eng. 2015;30(5):359-375.
45. Luo X, Li H, Deng Y, Zhang Y. Energy-based kinetics approach for coupled viscoplasticity and viscofracture of asphalt mixtures. Journal
of Engineering Mechanics. 2020;146(9):04020100. https://doi.org/10.1061/(asce)em.1943-7889.0001836

How to cite this article: Deng Y, Luo X, Zhang Y, et al. Determination of flexible pavement deterioration
conditions using Long-Term Pavement Performance database and artificial intelligence-based finite element
model updating. Struct Control Health Monit. 2020;e2671. https://doi.org/10.1002/stc.2671