TITLE:Influence of Seed Layer Moduli on FEM Based Modulus Backcalculation Results
Kunihito MATSUI
Professor
Dept. of Civil and Environmental Engineering, Tokyo Denki University
Hatoyama-Cho, Hiki-Gun, Saitama 350-0394, Japan
Tel.: +81-492-96-2911
Fax: +81-492-96-6501
matsui@g.dendai.ac.jp
Yoshitaka HACHIYA
Head
Airport Facilities Division, National Institute for Land and Infrastructure Management, Ministry of Land,
Infrastructure and Transport
1-1, Nagase 3, Yokosuka, 239-0826, Japan
Tel: +81-468-44-5034
Fax: +81-468-44-4471
hachiya@ipc.ysk.nilim.go.jp
James W. MAINA
CSIR, P O Box 395,
Pretoria, 0001, South Africa
Tel.: +27-12-841-3956
Fax: +27-12-841-3232
JMaina@csir.co.za
Yukio KIKUTA
Professor
Dept. of Civil and Environmental Engineering, Kokushikan University
4-28-1 Setagaya, Setagaya-Ku, Tokyo 154-8515, Japan
Tel.: +81-3-5481-3279
Fax: +81-3-5481-3279
kikuta@kokushikan.ac.jp
Tasuku NAGAE
Graduate student
Dept. of Civil and Environmental Engineering, Tokyo Denki University
Hatoyama-Cho, Hiki-Gun, Saitama 350-0394, Japan
Tel.: +81-492-96-2911
Fax: +81-492-96-6501
TOTAL NUMBER OF WORDS (including abstract, figures and tables):7,370
TRB 2006 Annual Meeting CD-ROM
Paper revised from original submittal.
Kunihito MATSUI et al.
1
ABSTRACT
Determination of pavement layer moduli from FWD test data is known as backcalculation analysis. Generally,
backcalculation analysis is an unstable procedure which is greatly influenced by several types of error causes.
These errors may be categorized as modeling error in the forward analysis, deflection measurement error,
numerical computation error due to instability in the backcalculation procedure, etc. Because of all the problems
mentioned, selection of seed values for layer moduli would highly influence backcalculation results.
In order to reduce effects of measurement error, truncated singular value decomposition is utilized in
backcalculation for regularization purpose. Scaling of variables, which is often used in optimization algorithm,
is implemented to improve numerical accuracy. In dynamic backcalculation, Ritz vector reduction method is
employed to efficiently solve a large system of dynamic equations. Various other means are also introduced to
cut down computation time.
This paper presents recent updates of DBALM (Dynamic Back Analysis for Layer Moduli) software whose
solver is based on axi-symmetric FEM and was first developed in 1993. Examples on airfield pavement
application are also presented. The results are compared with results from our static backcalculation software
BALM (Back Analysis for Layer Moduli) where the solver was developed using multilayered linear elastic
theory. From our experience, we believe a dynamic backcalculation is superior to static backcalculation. The
difference between the results from the two methods is presented in this paper.
TRB 2006 Annual Meeting CD-ROM
Paper revised from original submittal.
Kunihito MATSUI et al.
2
INTRODUCTION
Falling weight deflectometer (FWD) is widely used as a standard nondestructive testing device for structural
evaluation of pavements. Several highway and airport agencies have, in recent years, been drawing up their
pavement maintenance and management plans based on FWD test results. A wrong interpretation of the FWD
test results may lead to very high maintenance and management costs. FWD is a dynamic test, which applies
impulsive force to the pavement surface. However, conventional approach to estimate layer moduli is so called
static backcalculation, where peak load and corresponding peak deflections are considered as the responses of a
quasi-static deflection called deflection bowel. Intensive comparison of static backcalculation software may be
found in, for example, (1).
It is well known that the differential equation of motion can be written as,
m
d 2z
dz
+ c + kz = F0 sin ωt
dt
dt 2
(1)
where ω is the frequency of the harmonic excitation. The steady state solution of the above equation becomes,
Z
=
Z0
1
[1 − (ω ω ) ]
2 2
n
(2)
+ [2ς (ω ω n )]
2
where:
ω n = k ω = natural frequency of undamped oscillation in radian per second,
ς = c c0 = damping factor, and
Z 0 = F0 k = zero frequency deflection of the spring-mass system under the action of steady
force F0 .
Eq.(2) is called a magnification factor and is graphically represented as shown in Figure 1. The figure
shows that the response of one degree of freedom system does not in general coincide with static deflection.
Thus, if we apply static backcalculation to dynamic data, we may obtain false results.
FWD is a dynamic test. Generally, depending on the type of the FWD device, a one or two mass force
generating unit is dropped on un-segmented plate with thin hard rubber pad or segmented plate with thick, soft
rubber pad, which is placed on the pavement surface. The impact force generates radially propagating
shockwaves. Waves (deflection forms), in the vertical direction, caused by the impact load are measured at
several points on concentric circles with different radii whose centers are located at the center of the loading
plate. Duration of loading runs from 20 ms to 65 ms depending on the type of FWD device. FWD provides rich
information if it is fully utilized. Considering peak values for each measurement points, it becomes clear that
points farther from the center of loading attain their peak values later than points closer to the center of loading.
This time difference is known as phase difference of deflections and is considered the velocity of propagation of
the shockwaves. Furthermore, it is also well known that velocity of wave propagation in pavement structures
will be different depending on frequency of the shockwaves. These types of time series data contain useful
information that if well utilized would contribute to improved accuracy in structural pavement evaluation. The
use of only peak values implies discarding the rich information. It is essential from now on to find out how to
make use of FWD time series data and what kind of information we can extract from it.
TRB 2006 Annual Meeting CD-ROM
Paper revised from original submittal.
Kunihito MATSUI et al.
3
Static backcalculation can identify pavement layer moduli from the peak loading and the corresponding
peak deflections. It is known that, under the same peak loading, the longer the loading duration the larger the
surface deflection becomes (2). That is why static backcalculation tends to yield larger stiffness. Several authors
have presented dynamic backcalculation of FWD time series data (3-6). They have not stated in detail, methods
of backcalculation and the algorithms used. In the past, some of authors of this paper proposed a method for
dynamic backcalculation of pavement layer moduli using time series FWD data (2, 7-11).
Backcalculation analysis is known to be intrinsically unstable. Its success depends on the selection of
algorithm coupled with regularization technique implemented in the algorithm (7). Authors use the
Gauss-Newton method with truncated singular value decomposition with scaling of variables. Scaling of
variables is recommended for solving optimization problems when parameter values are different in order (12).
It seems the scaling is particularly useful for dynamic backcalculation since damping coefficients are much
smaller than layer moduli. In static backcalculation, which is called BALM (Static Analysis for Layer Moduli),
elastic multilayered analysis software called GAMES (13), is used to compute pavement responses. However,
because FEM is employed to compute pavement responses in our dynamic backcalculation (DBALM),
equations of motion result in a large system of differential equations. If we solve the system of equations,
computational time becomes enormous and dynamic backcalculation impractical. A matrix reduction method
based on Ritz vectors (14) is introduced in the analysis of dynamic system and more than 500 equations of
motion are reduced to 30 equations, which yields a drastic reduction of computational time.
Average surface deflections of three or four sets of test data are conventionally used in static
backcalculation in order to reduce effects of measurement error. However, averaging of time series data for the
backcalculation use is questionable. We have developed backcalculation algorithm, which can handle multiple
sets of data simultaneously with little increase of computation time (10). Although the dynamic backcalculation
is still time consuming as compared with the static backcalculation, it conforms well to the characteristics of
FWD test and the results obtained seem to be more reliable than the results from static method.
The objectives of this research are, 1) to evaluate influence of seed moduli on backcalculation results, 2)
to apply the DBALM to the airfield data taken at test site, 3) to compare results from static and dynamic
backcalculation analyses and 4) to estimate layer damping coefficients.
EQUATIONS OF MOTION AND SENSITIVITY ANALYSIS
Since FWD test applies an impulsive load on the surface of a pavement structure, dynamic and not static
backcalculation should ideally be used to determine pavement layer moduli as well as damping coefficients.
Dynamic forward analysis is required in order to perform dynamic backcalculation. Equation of motion for the
impulsive FWD loading can be presented as follows:
M
dz (l )
+ Kz (l ) = fg (l ) (t )
dt
dt 2
(l )
dz
(l )
z ( 0) = 0 ,
( 0) = 0
dt
d 2z
(l )
+C
TRB 2006 Annual Meeting CD-ROM
(3)
Paper revised from original submittal.
Kunihito MATSUI et al.
4
where; M , C , K represent mass, damping and stiffness matrices, respectively. K is a function of layer
moduli, E = E j , ( j = 1, 2, ..., M ) , and C is a function of layer damping coefficients, Q = Q j ,
( j = 1, 2, ..., M ) . M is a total number of pavement layers, f is nodal load distribution vectors, g (l ) (t ) is a
{ }
{ }
scalar representation of load as a function of time for the l th time series loading measurement. f can be
formulated considering uniform distribution of load over a loading plate of radius a . Eight nodes
iso-parametric elements are utilized in the finite element analysis. By assembling element stiffness matrices, a
layer stiffness matrix can be obtained. The stiffness matrix K j for the j th layer is a function of E j and can
be written as E j H j , ( j = 1, 2, ..., M ) . Similarly, damping matrix C j , which is a function of Q j can be
written as Q j H j . This implies that E j in stiffness matrices of all elements in every pavement layer can be
replaced by Q j to obtain damping matrices. H j is composed of only nodal coordinates in j th layer. M ,
H j as well as f can be prepared during the first iteration step and remain unchanged in subsequent iterations.
A global stiffness matrix, K is constructed from layer stiffness matrix, K j = E j H j while a global damping
matrix, C is from C j = Q j H j . Since every term in Eq. 1 has a unit of force and the unit of E j is Pa or
N/m 2 , the unit of Q j will be N ⋅ s m 2 .
It is important to perform sensitivity analysis of deflections with respect to the unknown parameters
during backcalculation analysis. Derivatives of deflections with respect to E j and Q j in Eq. 1 can be written
as follows:
M
d 2 ∂ z (l )
dt 2 ∂E j
(l )
+ C d ∂z
dt ∂E j
(l )
+ K ∂ z = − ∂ K z (l )
∂E j
∂E j
M
d 2 ∂ z (l )
dt 2 ∂Q j
(l )
+ C d ∂z
dt ∂Q j
(l )
(l )
+ K ∂z = − ∂C dz
∂Q j dt
∂Q j
(4)
(5)
Since the differential operators in Eqs. 4 and 5 are similar to Eq. 3, the same method is used to solve for
∂z i(l ) ∂E j and ∂z i(l ) ∂Q j .
The sizes of Eqs. 3, 4 and 5 are reduced by using Ritz vectors (12). The reduced system of equations is
rewritten as a first order system of differential equations, which is solved analytically by using an eigenvalue
analysis (8).
SCALING OF VARIABLES AND BACKCALCULATION ANALYSIS
The process of determining pavement layer moduli and damping coefficients using measured data is referred to
as backcalculation analysis. These parameters are determined such that there is a good much between computed
and measured deflections. It is imperative that measurements of time series loading and deflection data be made
simultaneously. Contrary to static backcalculation, synchronization of the data is essential in dynamic
backcalculation. In order to match both computed and measured deflections, an evaluation function is defined
as:
J=
{
}
2
1 L t1 N (l )
∑ ∑ u (t ) − zi(l ) (X, t ) dt
2 LN l =1∫ t 0 i =1 i
(6)
where J is the least square function and X = ( X1, X 2 , ..., X 2 M )T is a vector of the scaled unknown parameters.
TRB 2006 Annual Meeting CD-ROM
Paper revised from original submittal.
Kunihito MATSUI et al.
5
th
u i(l ) (t ) is the l time series deflection at sensor point i , while z i(l ) ( X, t ) is analytical deflection at sensor
point i due to l th time series load measurement. L is number of data sets and N is number of sensors. The
scaling of variables is considered important when the order of variables is different (12).
Iterative computation is necessary because determination of X is a nonlinear minimization problem.
Taylor expansion may be used as follows:
zi(l ) ( X + dX, t ) = zi(l ) ( X, t ) +
2 M ∂z (l )
i dX
∑ ∂X
j =1
(7)
j
j
Substituting Eq. 7 into Eq. 6, and simplify to obtain Eq. 8 as:
A ∆X = b
(8a)
where
2 M t1 L N ∂z (l ) ∂z (l )
i dt
A = ∑ ∫ ∑∑ i
j =1 t 0 λ=1 i =1 ∂X j ∂X k
{
∆X = ∆X j
b=∫
}
(8b)
( j = 1,2,...,2M )
(8c)
(u (l ) (t) − zi(l ) (X, t )) ∂Xi
t0 ∑ ∑ i
t1 L N
∂z (l )
dt
k
λ=1i =1
(k = 1,2,...,2M )
(8d)
A is M x M square matrix, ∆X and b are 2M x 1 vectors. Eq. 8a represents simultaneous linear equations,
which should be solved with care because there may be some instances when the determinant of A tends to
zero. ∂zi(l ) ∂X j is a function of time and can be determined from Eq. 9 as:
∂z (l )
∂zi(l )
= E 0j i
∂E j
∂X j
(9a)
( j = 1, ..., M )
∂zi(l )
∂z (l )
= Q 0j i
∂X j + M
∂Q j
(9b)
where E 0j and Q 0j are seed modulus and damping coefficient of j th layer. ∂zi(l ) ∂E j and ∂zi(l ) ∂Q j can be
obtained by solving Eqs. 4 and 5.
Considering unstable nature of this set of equations, singular value decomposition is used. A is
decomposed as,
A = UDV T
(10)
in which U T U = VV T = 1 is a unit matrix and D is a diagonal matrix composed of singular values. The value
of maximum singular value divided by minimum singular value is called condition number. By using these
decomposed matrices, the solution of linear set of equations can be written as,
∆X = VD −1 U T b
TRB 2006 Annual Meeting CD-ROM
(11)
Paper revised from original submittal.
Kunihito MATSUI et al.
6
The above equation can be rewritten as,
M
∆X = ∑ a i v i
(12)
i =1
where v i is i th column of V and ai =
1 M
∑ U jib j . U ji is the (i, j ) element of U , b j is the j th
d ii j =1
element of b and d ii is (i, i ) element of D . If dii is smaller than a threshold value, a i is computed such
that 1 / d ii is taken as 0 to prevent a significant influence of measurement error involved in b j . The threshold
value is chosen as 0.001 of maximum singular value in our software.
Assuming initial values for unknown parameters and updating the parameters after each iterative step,
the computations must be repeated until a convergence is achieved. However a convergence will never be
achieved unless some regularization technique is introduced. A truncated singular value decomposition, which
is the simplest and the most efficient one among regularization techniques if a proper threshold value is selected,
is implemented in DBALM.
ACCELERATED PAVEMENT TEST (APT) DATA (13)
Two experimental pavements, Section A and Section B, were constructed in a reinforced concrete vessel as
shown in Figure 2. After construction, the aircraft load simulator, which can simulate the wheel loading of
B747-400 aircraft (910 kN), was used to apply 10,000 load repetitions in a bi-directional mode at a speed of 5
km/h. The position of the wheel path was as shown in Figure 2. After 5,000 load repetitions, it was impossible to
continue the test due to excessive rutting for the operation of the aircraft load simulator. The surfaces of both
sections were, thereafter, repaired using the same materials as that of the surface course in section A before
continuing with the remaining 5,000 load repetitions. In this paper, data for the first 5,000 load repetitions at
Section B is used.
After every specified number of load repetitions, FWD tests were performed along the wheel path as
shown in Figure 2. The diameter of the FWD loading plate was 450 mm and a loading level of 250 kN were used
to obtain deflections at 0, 300, 450, 600, 900, 1,500, 2,500 mm from the center of the loading plate. One
measurement consisted of four FWD tests and time series loading and deflection data were recorded at an
interval of 0.0002 second. Thermocouples embedded in the surface and binder courses were used to measure
internal temperatures in the asphalt concrete during the FWD tests.
BACKCALCULATION OF APT DATA
Time series loading and deflection data measured by FWD were used in a dynamic backcalculation program
DBALM, while by using the peak values of loads and deflections, static backcalculation was conducted using
BALM software.
Influence of seed layer moduli
TRB 2006 Annual Meeting CD-ROM
Paper revised from original submittal.
Kunihito MATSUI et al.
7
Investigation on the influence of seed layer moduli was carried out by performing static and dynamic
backcalculation analyses using 1000 sets of seed layer moduli, (ranges: 1,000 MPa < E1 < 10,000 MPa,
100 MPa < E 2 < 500 MPa and 50 MPa < E 3 < 150 MPa ) which were prepared using uniform random numbers
in order to investigate seed value effects. Backcalculation was performed statically and dynamically. FWD tests
ware conducted four times under the same condition. After modifying the measured deflections to the standard
250 kN, surface deflections ware averaged and used to perform static backcalculation using the averaged values.
However for dynamic backcalculation, new algorithm was developed which has similar to the effect of
averaging static deflections (10). Figure 4a shows the meshing of pavement section and Figure 4b is a set of
measured load and surface deflections data. Time increment used for the dynamic analysis is 0.002 second to
reduce computational time, although measurement is taken at every 0.0002 second. When performing dynamic
backcalculation, the measured surface deflections between t0 and t1 are matched with computed deflections
of the same range, while applied load data of the range between 0 and t1 is used for dynamic analysis.
Both static and dynamic backcalculation results before the application of wheel load are plotted in
Figure 4. It is observed that static results vary in relatively small range compared with dynamic results. The
mean values of results are E1 = 14410 MPa, E2 = 330 MPa, E3 = 80 MPa for static backcalculation and
E1 = 9380 MPa, E2 = 150 MPa, E3 = 70 MPa for dynamic backcalculation. Temperature at the middle of
asphalt concrete is 9.7 °C . Since layer damping coefficients can be estimated from dynamic backcalculation,
they are also described in Figure 4. Static backcalculation results are larger than dynamic backcalculation results.
This is a general trend which has been observed in the past. Since the values of base modulus scatters widely,
two sets of initial values corresponding to the largest and smallest of base moduli in Figure 4(b) are selected, the
iteration processes for static and dynamic backcalculations are plotted in Figure 5. The figure demonstrates
steady convergence. Possible cause of larger scatter for dynamic backcalculation is difference in the number of
unknown parameters, three for static backcalculation and six for dynamic backcalculation.
Figure 6 shows comparison of measured and computed deflections after convergence is achieved. Good
match of deflections is observed. However, a slight phase difference is observed because the time corresponding
to peak measured deflections appear to be different from the time corresponding to peak computed deflections.
Records of loading and deflections must be well synchronized when dynamic backcalculation is performed.
Figure 7(a) shows a sample of measured load and deflections where peak points are marked. Figure 7(b)
illustrates a deflection bowl formed from peak surface deflections in Figure 7(a) and surface deflection obtained
from static analysis by using layer moduli from dynamic backcalculation and peak load. It is obvious that the
deflection bowl is much smaller than the static deflection. Thus, if the deflection bowl is considered as static
deflection, backcalculated moduli are expected to be larger. This is the reason why static backcalculation results
become greater than dynamic results.
During the accelerated pavement loading test, FWD tests were performed at 0, 50, 100, 200, 500, 1000,
2000, 3000, 5000 load repetitions. Maximum, mean and minimum backcalculation results are illustrated in
Figure 8. As number of repeated loading increases, layer moduli decreases. It is observed that scatter in static
backcalculation results is much smaller than dynamic one. The scatter for static case decreases with increasing
number of loading, while that for dynamic backcalculation remains nearly same.
APPLICATION TO OTHER FWD DATA
LTPP data at Site 3, State highway 281, Texas (16)
TRB 2006 Annual Meeting CD-ROM
Paper revised from original submittal.
Kunihito MATSUI et al.
8
More evaluations, with the objective of confirming the accuracy and suitability of the software used for dynamic
analysis in this research, were performed also using data obtained from the webpage under TRB committee
A2B05. Among the data released to the public through Nonlinear Pavement Analysis Project under the
Committee, only data from Site 3, which is located on State Highway 281, Texas were used.
The pavement structure on this section was as follows: 203mm surface layer (asphalt concrete), 308 mm
flexible base (no treatment) and the subgrade soil, which was clayey and moderately stress-dependent. In this
site, bedrock was located at a depth of 1.9 m below the pavement surface. Four types of loading (27, 40, 53, 71
kN) were applied during FWD testing, with 3 drops per each load. Furthermore, 6 deflection sensors located at 0,
305, 610, 915, 1220, 1525 mm from the center of the loading plate were used to record vertical deflections. Seed
moduli used here are same as those of the aircraft load simulator data in the previous section. Backcalculation
results presented here are those corresponding to 71kN. The FWD load plate was positioned at approximately
220mm from the multi-depth deflectometer (MDD) cap. The MDD recorded the pavement’s deflections at three
different depths, which were approximately 95 mm (AC layer), 314mm (base course layer) and 594 mm
(subgrade layer) from the surface of the pavement. Figure 9 shows the results from static and dynamic
backcalculation. Figure 10 presents the comparison of measured and computed surface deflections and the
comparison of measured and computed vertical displacements at MDD sensor locations. The vertical
displacements at MDD sensor locations are computed by DynaPave3 (11) which is dynamic analysis solver for
pavement implemented in DBALM.
Road Test Sections in Japan
Figure 11 shows 2 types of pavement sections where FWD tests were conducted. The two sections were
constructed side by side with similar base and subgrade materials. Thickness of asphalt concrete layer for
section C was 5.1 cm while that of section D was 24.6 cm. Past experience with backcalculation analysis has
shown that different seed moduli almost always give different backcalculation results and this trend is very
prominent in case the thickness of surface course is less than 7.5 cm. It is with this respect that experience is very
important for good selection of seed moduli in backcalculation analysis. In this research, random numbers were
used to generate seed values for layer moduli and the influence thereof was investigated.
By using the 1000 sets of seed values, E1 between 1,000 MPa and 10,000MPa, E2 between 100 MPa
and 500 MPa, E3 between 80 MPa and 300 MPa, and E4 between 50 MPa and 150 MPa, static and dynamic
backcalculation were conducted and the results are described in Figures 12 and 13 respectively. Figure 12(a) and
(b) are static backcalculation results for section C and section D. Although both sections are composed of similar
materials, asphalt concrete layer modulus of section C ( E1 = 8,000 MPa) is greater than that of section D
( E1 = 5,000 MPa), while subgrade modulus of section C ( E4 = 80 MPa) is smaller than that of section D
( E4 = 100 MPa). By comparing layer modulus in Figures 13(a) and 13(b), it is found that the corresponding
layers of sections C sand D are nearly same. Layer damping coefficients for section C run between 17.8 kN ⋅ s/m
and 58.9 kN ⋅ s/m for C1 , between 1.82 kN ⋅ s/m and 2.04 kN ⋅ s/m for C2 , between 1.22 kN ⋅ s/m and
1.66 kN ⋅ s/m for C3 and between 0.217 kN ⋅ s/m and 0.247 kN ⋅ s/m for C4 . And the damping coefficients
for section D are between 28.2 kN ⋅ s/m and 43.2 kN ⋅ s/m for C1 , between 1.25 kN ⋅ s/m and 2.61kN ⋅ s/m
for C2 ,between 0.832 kN ⋅ s/m and 1.922 kN ⋅ s/m for C3 and between 0.158 kN ⋅ s/m and 0.351kN ⋅ s/m
for C4 .
TRB 2006 Annual Meeting CD-ROM
Paper revised from original submittal.
Kunihito MATSUI et al.
9
CONCLUSIONS
In backcalculation analysis of FWD data, it is common practice to perform static backcalculation even though
the FWD test itself is a dynamic test. This research was, therefore, performed based on the “dynamic analysis
should be performed on FWD data” point of view. The following conclusions were drawn based on the results
obtained:
1.
2.
3.
4.
5.
6.
Static backcalculation in general yields estimates of layer moduli larger than dynamic backcalculation
does.
Dynamic backcalculation tend to yield larger scatter than static backcalculation when asphalt concrete is
thicker. The possible reason is that number of unknown parameters for dynamic backcalculation is as twice
as that of static backcalculation
Because dynamic backcalculation simulate FWD test, dynamic backcalculation analysis were more
reliable than results from static method. However further updates of DBALM is necessary to reduce effects
of seed values.
Backcalculated results, especially for upper layers in a pavement structure, were highly affected by the
seed moduli.
Results from TRB data indicate that DBALM and DynaPave3 give sufficiently accurate and acceptable
results.
Layer damping coefficients also can be identified by DBALM.
At this stage, it is assumed that layer damping coefficients relate to distress of pavement as well as
moisture contents. However, further examinations must be made to understand what physically they mean.
ACKNOWLEDGMENTS
Authors of this research study would like to acknowledge the supports by Research Institute of Technology,
Tokyo Denki University (Project No. Q02M-05) and Program for Promoting Fundamental Transport (Project
No. 2000-03). Authors are also very thankful to Professor Waheed Uddin, University of Mississippi for his
advise.
TRB 2006 Annual Meeting CD-ROM
Paper revised from original submittal.
Kunihito MATSUI et al.
10
REFERECES
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
Deusen, D.V. : Selection of Flexible Backcalculation Software for the Minnesota road research project,
Minnesota Department of Transportation, Aug. 1996.
Matsui, K., K. Kikuta, T. Nishizawa, and A. Kasahara. Comparative Studies of Backcalculated Results
from FWDs with Different Loading Duration, Nondestructive Testing of Pavement and Backcalculation of
Moduli, (Edtors: Tayabji, S. D. and E.O. Lukanen) Third Volume, pp. 470-483, March 2000.
Uzan, J.: Dynamic Linear Backcalculation of Pavement Material Parameters, Journal of Transportation
Engineering, ASCE ,Vol. 120(1), pp. 109-126, 1994.
Kang, Y. V. Multifrequency Back-Calculation of Pavement-Layer Moduli, Journal of Transportation
Engineering, vol. 124 (1), pp. 73-81, 1998.
Al-Khoury, R., Scarpas, A., Kasbergen, C., and Blaauwendraad, J.: Spectral Element Technique for
Efficient Parameter Identification of Layered Media, Part II: Inverse Calculation, International Journal of
Solids and Structures, Vol. 38, pp. 8753-8772, 2001.
Chatti, K., S. W. Haider, H.S. Lee, Y. Ji, and H, Salama. Evaluation of Nonlinear and Dynamic Effects on
Asphalt Pavement Response under FWD Loading, International Journal of Pavements, Vol. 2, No.1-2, pp.
88-99, 2003.
Matsui,K., Nishizawa, T., and Kikuta, Y. Time Domain Backcalculation of Pavement, Proceedings of
Nondestructive Evaluation of Bridges and Highways, SPIE The International Society for Optical
Engineering, March 31- April 3, 1998, San Antonio, Texas, USA.
Dong, Q.X., Hachiya, Y., Takahashi, O., Tsubokawa, Y., and Matsui, K. An efficient backcalculation
algorithm of time domain for large-scale pavement structures using Ritz vectors, Finite Element in
Analysis and Design 38 (2002), pp. 1131-1150.
Matsui, K., Maina, J. W., Dong, Q., and Sasaki, Y. A Fast Dynamic Backcalculation of Layer Moduli using
Axi-Symmetric Approach, IJP 2003 Volume 2 Number 1-2 January – May 2003, pp. 75 – 87.
Maina, J. W., Matsui, K. Kikuta, Y., and Inoue, T. Dynamic Backcalculation of Pavement Structure using
Multiple Sets of Time Series data, ASCE Geo Institute Conference on Geotechnical Engineering for
Transportation Projects, USA, 2004.
Maina, J. W., S. Higashi, Y. Kikuta, and K. Matsui. Influence of Seed Values on Dynamic
Backcalculation Results, MAIREPAV4, Belfast, Aug. 18-20, 2005.
Arora, J. S. Introduction to Optimization, McGraw Hills, 1989.
Maina, J. W. and Matsui, K.: Development Software for Elastic Analysis of Pavement Structure Responses
to Vertical and Horizontal Surface Loading, In Transportation Research Record 1896, TRB, National
Research Council, Washington, D.C., 2004, pp. 107-118.
Wilson, E. L., Yuan, M. W., and Dicken, J. M. Dynamic analysis by direct superposition of Ritz vectors,
Journal of Earthquake Engineering and Structural Dynamics 10 (1982), pp. 813-821.
Tsubokawa, Y., Hachiya, Y., Maina, J.W., Nishizawa, T., and Matsui, K. Accelerated Tests on Applicability
of Large Stone Asphalt Concretes in Surface Course for Airports, 2nd International Symposium for
Accelerated Pavements Test, Minneapolis, Minn. 2004.
http://www.clrp.cornell.edu/A2B05/
TRB 2006 Annual Meeting CD-ROM
Paper revised from original submittal.
Kunihito MATSUI et al.
11
LIST OF FIGURES
FIGURE 1 Magnification factor for the vibration of a viscously damped system
FIGURE 2 Experimental pavement for airfields (aircraft load simulator)
FIGURE 3 Example of mesh and interval of evaluation function (aircraft load simulator)
FIGURE 4 Comparison of static and dynamic results for airfield pavement (aircraft load simulator)
FIGURE 5 Iteration process (aircraft load simulator)
FIGURE 6 Comparison of measured and computed deflections (aircraft load simulator)
FIGURE 7 Measured FWD data, and comparison of deflection bowl and static deflection (aircraft load
simulator)
FIGURE 8 Maximum, mean and minimum modulus with respect to loading repetitions (aircraft load simulator)
FIGURE 9 Frequency distributions of static and dynamic backcalculation results (Texas LTPP data)
FIGURE 10 Surface deflections and vertical displacements at MDD sensor locations (Texas LTPP data)
FIGURE 11 Road test sections in Japan
FIGURE 12 Frequency distributions of static backcalculation (test site in Japan)
FIGURE 13 Frequency distributions of dynamic backcalculation (test site in Japan)
TRB 2006 Annual Meeting CD-ROM
Paper revised from original submittal.
Kunihito MATSUI et al.
12
4.0
← ζ = 0.00
Magnification factor
3.0
← ζ = 0.15
2.0
ζ = 0.25→
1.0
↑
ζ = 1.00
0.0
0.0
1.0
2.0
3.0
4.0
5.0
Frequency ratio ω / ω
FIGURE 1 Magnification factor for the vibration of a viscously damped system.
TRB 2006 Annual Meeting CD-ROM
Paper revised from original submittal.
Kunihito MATSUI et al.
13
8,500
Section B
Section A
3,000
2,000
500
4,250
1,118
4,250
8,500
3,000
Wheel path
Measurement line of transverse surface profile
Loading plates of FWD and static loading test
Displacement meter
(Unit: mm)
(a) Plan
Section B
0
80
160
Section A
Surface Course
Surface Course
Binder Course
Binder Course
0
50
160
Asphalt Stabilized Base Course
310
310
Granular Material Subbase
1,060
1,060
Subgrade
(Unit: mm)
(b) Pavement Section
FIGURE 2 Experimental pavement for airfields (aircraft load simulator).
TRB 2006 Annual Meeting CD-ROM
Paper revised from original submittal.
Kunihito MATSUI et al.
14
CL
r =5 m
Loading radius =0.225 m
Loading level =250 kN
Surface course
Base course
6m
Subbase
Subgrade
(a) Mesh
200
Load (kN)
100
0
-100
-200
-300
t1
t0
-400
0
10
20
30
40
Load
Deflection (mm)
0.6
0.5
0.4
0.3
0.2
0.1
0.0
-0.1
-0.2
-0.3
-0.4
-0.5
-0.6
-0.7
-0.8
300
D0
D30
D45
D60
D90
D150
D250
50
Time (ms)
(b) Interval of evaluation function (t0-t1)
FIGURE 3 Example of mesh and interval of evaluation function (aircraft load simulator).
TRB 2006 Annual Meeting CD-ROM
Paper revised from original submittal.
Kunihito MATSUI et al.
15
18000
14000
12000
10000
Mean: 14414 MPa
8000
CV: 2.6 %
6000
E1
16000
Layer modulus(MPa)
Layer modulus(MPa)
18000
E1
16000
4000
14000
12000
10000
8000
6000
4000
Mean: 9379 MPa
CV: 5.0 %
2000
2000
0
0
0
200
400
600
Random number
800
0
1000
500
200
400
600
Random number
800
1000
800
1000
E2
400
Layer modulus(MPa)
Layer modulus(MPa)
1000
500
E2
300
Mean: 328 MPa
CV: 4.8 %
200
100
0
400
Mean: 150 MPa
CV: 20.8 %
300
200
100
0
0
200
400
600
Random number
800
1000
0
100
200
400
600
Random number
100
E3
E3
80
Layer modulus(MPa)
Layer modulus(MPa)
800
60
Mean: 81.2 MPa
40
CV: 1.1 %
20
0
80
60
40
Mean: 69.0 MPa
20
CV: 6.7 %
0
0
200
400
600
Random number
800
1000
0
200
400
600
Random number
Layer damping coefficients
Max (kN s/m)
Min (kN s/m)
Mean (kN s/m)
CV (%)
(a) Static
C1
65.6
32.9
50.7
15.0
C2
2.98
1.99
2.59
9.4
C3
0.325
0.179
0.253
13.0
(b) Dynamic
FIGURE 4 Comparison of static and dynamic results for airfield pavement (aircraft load simulator).
(CV : Coefficient of Variation)
TRB 2006 Annual Meeting CD-ROM
Paper revised from original submittal.
Kunihito MATSUI et al.
Layer modulus (MPa)
E1
Initial value set #1
Initial value set #2
1
4
7 10 13 16 19 22 25 28 31
Iteration
400
300
200
Initail value set #1
Initial value set #2
100
160
140
120
100
80
60
40
20
0
1
4
7 10 13 16 19 22 25 28 31
Iteration
(a) Static backcalculation
E2
400
Initial value set #1
Initial value set #2
300
200
100
1
E3
Initial value set #1
Initial value set #2
7 10 13 16 19 22 25 28 31
Iteration
500
7 10 13 16 19 22 25 28 31
Iteration
Layer modulus (MPa)
Layer modulus (MPa)
4
4
0
0
1
Initial value set #1
Initial value set #2
600
E2
500
E1
1
Layer modulus (MPa)
Layer modulus (MPa)
600
16000
14000
12000
10000
8000
6000
4000
2000
0
Layer modulus (MPa)
16000
14000
12000
10000
8000
6000
4000
2000
0
16
4
160
140
120
100
80
60
40
20
0
7 10 13 16 19 22 25 28 31
Iteration
E3
Initial value set #1
Initial value set #2
1
4
7 10 13 16 19 22 25 28 31
Iteration
(b) Dynamic backcalculation
FIGURE 5 Iteration process (aircraft load simulator).
TRB 2006 Annual Meeting CD-ROM
Paper revised from original submittal.
Kunihito MATSUI et al.
17
0.0
D0-measured
D0-computed
D30-measured
D30-computed
D45-measured
D45-computed
D60-measured
D60-computed
D90-measured
D90-computed
D150-measured
D150-computed
D250-measured
D250-computed
Deflection (mm)
-0.1
-0.2
-0.3
-0.4
-0.5
-0.6
16
18
20
22
24
26
28
30
32
Time (ms)
(a) Initial value set #1
0.0
D0-measured
D0-computed
D30-measured
D30-computed
D45-measured
D45-computed
D60-measured
D60-computed
D90-measured
D90-computed
D150-measured
D150-computed
D250-measured
D250-computed
Deflection (mm)
-0.1
-0.2
-0.3
-0.4
-0.5
-0.6
16
18
20
22
24
26
28
30
32
Time (ms)
(b) Initial value set #2
FIGURE 6 Comparison of measured and computed deflections (aircraft load simulator).
TRB 2006 Annual Meeting CD-ROM
Paper revised from original submittal.
18
300
250
200
150
100
50
0
-50
-100
-150
-200
-250
-300
-350
0.6
0.5
0.4
0.3
0.2
0.1
0.0
-0.1
-0.2
-0.3
-0.4
-0.5
-0.6
-0.7
peak load
peak deflection
0
10
20
30
40
Deflection (mm)
Load (kN)
Kunihito MATSUI et al.
Load
D0
D30
D45
D60
D90
D150
D250
50
Time (ms)
(a) Measured FWD data
0
50
Sensor position (cm)
100
150
200
250
300
0.0
-0.1
Surface deflection (mm)
-0.2
-0.3
-0.4
-0.5
-0.6
-0.7
Static deflection
Deflection bowl
-0.8
-0.9
-1.0
(b) Comparison of deflection bowl and static deflection
FIGURE 7 Measured FWD data, and comparison of deflection bowl and static deflection (aircraft load
simulator).
TRB 2006 Annual Meeting CD-ROM
Paper revised from original submittal.
Kunihito MATSUI et al.
19
15000
E1
15000
12000
12000
9000
Max
Mean
Min
6000
3000
0
9000
6000
Max
Mean
Min
3000
Loading repitition
Loading repitition
400
300
200
Max
Mean
Min
100
0
Max
Mean
Min
400
E2
300
200
100
Loading repitition
Loading repitition
E3
80
60
40
Max
Mean
Min
20
0
100
Layer modulus (MPa)
100
10
0
20
0
50
0
10
00
20
00
30
00
50
00
0
10
0
20
0
50
0
10
00
20
00
30
00
50
00
50
0
0
50
Layer modulus (MPa)
E2
Layer modulus (MPa)
500
500
E3
80
60
40
Max
Mean
Min
20
Loading repitition
(a) Static backcalculation
50
0
10
0
20
0
50
0
10
00
20
00
30
00
50
00
50
0
0
Layer modulus (MPa)
10
0
20
0
50
0
10
00
20
00
30
00
50
00
0
50
10
0
20
0
50
0
10
00
20
00
30
00
50
00
50
0
0
10
0
20
0
50
0
10
00
20
00
30
00
50
00
Layer modulus (MPa)
E1
Layer modulus (MPa)
18000
18000
Loading repitition
(b) Dynamic backcalculation
FIGURE 8 Maximum, mean and minimum modulus with respect to loading repetitions (aircraft load
simulator)
TRB 2006 Annual Meeting CD-ROM
Paper revised from original submittal.
Kunihito MATSUI et al.
1200
1000
800
600
400
200
0
Frequency
Mean: 1275 MPa
CV: 1.4 %
1200
1000
800
600
400
200
0
Layer modulus (MPa)
Frequency
E2
Mean: 215 MPa
CV: 0.0 %
1200
1000
800
600
400
200
0
E2
Mean: 144 MPa
CV: 2.4 %
0
0
0
0
0
0
0
12 - 14 - 16 - 18 - 20 - 22 - 24
0
0
0
0
0
0
0
10 12 14 16 18 20 22
10
0
-1
12 20
0
-1
14 40
0
-1
16 60
0
-1
18 80
0
-2
20 00
0
-2
22 20
0
-2
40
Frequency
Layer modulus (MPa)
Layer modulus (MPa)
Layer modulus (MPa)
(a) Static backcalculation
E3
0
-9
90
-8
0
Mean: 119 MPa
CV: 1.5 %
80
10 00
0
-1
11 10
0
-1
12 20
0
-1
13 30
0
-1
40
-1
90
-9
80
70
-8
0
0
Mean: 82 MPa
CV: 0.0 %
1200
1000
800
600
400
200
0
70
Frequency
E3
Frequency
1200
1000
800
600
400
200
0
Layer modulus (MPa)
10 00
0
-1
11 1 0
0
-1
12 2 0
0
-1
13 3 0
0
-1
40
10
50
11 11
00 0 0
11 11
50 5 0
12 12
00 0 0
12 12
50 5 0
13 13
00 0 0
13 13
50 5 0
-1
40
0
Mean: 1251 MPa
CV: 0.0 %
E1
10
50
11 11
00 0 0
11 11
50 5 0
12 12
00 0 0
12 12
50 5 0
13 13
00 0 0
13 13
50 5 0
-1
40
0
Frequency
E1
-1
1200
1000
800
600
400
200
0
20
Layer modulus (MPa)
Layer damping coefficients
C1
C2
Max (kN s/m)
5.609 0.2864
Min (kN s/m)
4.089 0.0052
Mean (kN s/m)
4.396 0.2287
CV (%)
9.3
32.6
C3
0.2157
0.1608
0.1730
8.4
(b) Dynamic backcalculation
FIGURE 9 Frequency distributions of static and dynamic backcalculation results (Texas LTPP data)
(CV : a coefficient of variation)
TRB 2006 Annual Meeting CD-ROM
Paper revised from original submittal.
21
80
0.8
Load
60
0.6
D0-measured
40
0.4
20
0.2
0
0.0
-20
-0.2
-40
-0.4
-60
-0.6
D0-computed
Deflection (mm)
Load (kN)
Kunihito MATSUI et al.
D30-measured
D30-computed
D60-measured
D60-computed
D90-measured
D90-computed
-80
D120-measured
-0.8
0
10
20
30
40
D120-computed
50
D150-measured
Time (ms)
D150-computed
(a) Surface deflections
Vertical displacement (mm)
0.1
0.0
A.C. Measured
-0.1
A.C. Computed
B.C. Measured
-0.2
B.C. Computed
SG Measured
-0.3
SG Computed
-0.4
-0.5
0
10
20
30
40
50
Time (ms)
(b) Vertical displacements at MDD sensor locations
FIGURE 10 Surface deflections and vertical displacements at MDD sensor locations (Texas LTPP
data)
TRB 2006 Annual Meeting CD-ROM
Paper revised from original submittal.
Kunihito MATSUI et al.
C
L
22
C
L r =15 cm
r =15 cm
P =49 kN
P =49 kN
Surface course
ν1=0.35
ρ1=0.0023 kg/cm3
h1=5.1 cm
ν1=0.35
Surface course ρ =0.0023 kg/cm3
1
h1=24.6 cm
Base course
ν2=0.35
ρ2=0.0019 kg/cm3
h2=9.3 cm
Base course
ν2=0.35
ρ2=0.0019 kg/cm3
h2=15.3 cm
Subbase
ν3=0.35
ρ3=0.0019 kg/cm3
h3=24.4 cm
Subbase
ν3=0.35
ρ3=0.0019 kg/cm3
h3=18.2 cm
Subgrade
ν4=0.40
ρ4=0.0018 kg/cm3
Subgrade
ν4=0.40
ρ4=0.0018 kg/cm3
(a) Pavement section C
(b) Pavement section D
FIGURE 11 Road test sections in Japan.
TRB 2006 Annual Meeting CD-ROM
Paper revised from original submittal.
23
1200
1000
800
600
400
200
0
1200
1000
800
600
400
200
0
Frequency
1200
1000
800
600
400
200
0
Layer modulus (MPa)
(a) Pavement section C
Frequency
E4
-8
80 0
-8
85 5
-9
90 0
95 95
10 100
0
10 105
5
-1
10
Mean: 101 MPa
CV: 0.8 %
75
75
-8
80 0
-8
85 5
-9
90 0
9 5 95
10 100
0
10 105
5
-1
10
Mean: 83.0 MPa
CV: 0.3 %
1200
1000
800
600
400
200
0
Frequecy
Frequency
E4
Mean: 177 MPa
CV: 12.5 %
Layer modulus (MPa)
Layer modulus (MPa)
1200
1000
800
600
400
200
0
E3
13
0
15 150
0
17 170
0
19 190
0
21 210
0
23 230
0
25 250
0
-2
70
Mean: 175 MPa
CV: 3.1 %
13
0
15 150
0
17 170
0
19 190
0
21 210
0
23 230
0
25 250
0
-2
70
Frequency
E3
Mean: 259 MPa
CV: 3.9 %
Layer modulus (MPa)
Layer modulus (MPa)
1200
1000
800
600
400
200
0
E2
20
0
24 240
0
28 280
0
32 320
0
36 360
0
40 400
0
44 440
0
-4
80
Mean: 318MPa
CV: 3.9 %
20
0
24 240
0
28 280
0
32 320
0
36 360
0
40 400
0
44 440
0
-4
80
Frequency
E2
Mean: 5,762 MPa
CV: 2.5 %
Layer modulus (MPa)
Layer modulus (MPa)
1200
1000
800
600
400
200
0
E1
30
00
40 - 40
00 00
50 - 50
00 00
60 - 60
00 00
70 - 70
00 00
80 - 80
0
0
90 0 - 9 0
00 00
-1 0
00
00
30
00
4 0 - 40
00 00
5 0 - 50
00 00
6 0 - 60
00 00
7 0 - 70
00 00
8 0 - 80
00 0
90 - 9 0
00 00
-1 0
00
00
Frequency
1200
E1
1000
800
600
400 Mean: 7842 MPa
200 CV.9.7%
0
Frequency
Kunihito MATSUI et al.
Layer modulus (MPa)
(b) Pavement section D
FIGURE 12 Frequency distributions of static backcalculation (test site in Japan).
CV : Coefficient of Variation
TRB 2006 Annual Meeting CD-ROM
Paper revised from original submittal.
Kunihito MATSUI et al.
E1
Frequency
Mean: 4888 MPa
CV: 6.9 %
1200
1000
800
600
400
200
0
Mean: 4807 MPa
CV: 9.3 %
25
00
3 0 - 30
00 00
3 5 - 35
00 00
4 0 - 40
00 00
4 5 - 45
00 00
5 0 - 50
00 00
5 5 - 55
00 00
-6
00
0
E1
25
00
3 0 - 30
00 00
3 5 - 35
00 00
4 0 - 40
00 00
4 5 - 45
00 00
5 0 - 50
00 00
5 5 - 55
00 00
-6
00
0
Frequency
1200
1000
800
600
400
200
0
24
Layer modulus (MPa)
Layer modulus (MPa)
E2
Frequency
Mean: 196 MPa
CV: 16.9 %
1200
1000
800
600
400
200
0
Mean: 236 MPa
CV: 21.6 %
15
0
20 200
0
25 250
0
30 300
0
35 350
0
40 400
0
45 450
0
-5
00
E2
15
0
20 200
0
25 250
0
30 300
0
35 350
0
40 400
0
45 450
0
-5
00
Frequency
1200
1000
800
600
400
200
0
Layer modulus (MPa)
Layer modulus (MPa)
E3
Mean: 160 MPa
CV: 31.5 %
10 100
0
15 150
0
20 200
0
25 250
0
30 300
0
35 350
0
-4
00
Mean: 167 MPa
CV: 7.7 %
1200
1000
800
600
400
200
0
Frequency
E3
50
50
10 100
0
15 150
0
20 200
0
25 250
0
30 300
0
35 350
0
-4
00
Frequency
1200
1000
800
600
400
200
0
Layer modulus (MPa)
Layer modulus (MPa)
(a) Pavement section C
E4
-5
55 5
-6
60 0
-6
65 5
-7
70 0
-7
75 5
-8
0
-5
0
Mean: 58.6 MPa
CV: 2.3 %
50
45
Layer modulus (MPa)
45
Mean: 61.3 MPa
CV: 1.7 %
1200
1000
800
600
400
200
0
Frequency
E4
-5
50 0
-5
55 5
-6
60 0
-6
65 5
-7
70 0
-7
75 5
-8
0
Frequency
1200
1000
800
600
400
200
0
Layer modulus (MPa)
(b) Pavement section D
FIGURE 13 Frequency distributions of dynamic backcalculation (test site in Japan).
(CV : Coefficient of Variation)
TRB 2006 Annual Meeting CD-ROM
Paper revised from original submittal.