You are on page 1of 11

Advances in Engineering Software 37 (2006) 421431 www.elsevier.

com/locate/advengsoft

Advances in backcalculating the mechanical properties of exible pavements


A. Burak Goktepe a, Emine Agar b, A. Hilmi Lav b,*
b a Department of Civil Engineering, Ege University, 35100 Izmir, Turkey Istanbul Technical University, Department of Civil Engineering, 34469 Ayazaga, Istanbul, Turkey

Received 22 March 2005; received in revised form 29 September 2005; accepted 5 October 2005 Available online 9 November 2005

Abstract In current practice, the evaluation of the performance of existing road pavements has become a priority issue for many highway engineers. To make appropriate rehabilitation and management decisions the engineer must rely on an efcient method for determining the structural conditions of pavements. Nondestructive testing (NDT) reveals the stressstrain properties of pavement layers at relatively low strain levels. Since the theoretical approaches used to determine the stressstrain relationships in pavement layers calculate the deections for given mechanical properties, it is necessary to make an inversion using a backcalculation tool. Several methods have been developed to backcalculate the mechanical properties of exible pavement; these methods vary in analysis type, material model, and optimization algorithm. This study is designed to explain these methods and to compare and contrast them in terms of modeling precision, computational expense, and calculation details. Consequently, innovations and advances in backcalculating exible pavements are considered in this paper. q 2005 Elsevier Ltd. All rights reserved.
Keywords: NDT; backcalculation; exible pavements; pavement analysis

1. Introduction The structural condition of a pavement structure generally provides the necessary information for: (a) an estimation of the pavements remaining life, and (b) the selection of a feasible rehabilitation and/or reconstruction strategy. Considering the length and the structure of a road network in a country, highway engineers need exible, fast, and reliable ways to determine the physical condition of the pavement section being examined. Nondestructive testing (NDT) methods are being currently used since they analyze the pavements structural condition in a nondestructive manner and are rapidly applied. Basically, NDT methods may be categorized as (a) deection basin methods and (b) surface wave methods. The former methods are fundamentally based on the measurement of surface deections emerging by the applied load as well, as on the correlation among these values and the stiffness of each layer. Obviously, the amount of surface deection depends on loading conditions (type, magnitude, contact area, and duration), measurement location, and layer properties (thickness, mass, and stiffness). Therefore, discrepancies
* Corresponding author. Tel.: C90 212 285 6535; fax: C90 212 286 5563. E-mail address: lav@itu.edu.tr (A. Hilmi Lav).

0965-9978/$ - see front matter q 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.advengsoft.2005.10.001

among NDT devices originate from the variations in loading conditions and measurement locations. Typical deection basin testing techniques widely used are the Dynaect road rater, falling weight deectometer (FWD), and rolling weight deectometer (RWD) tests [14]. Surface wave tests, on the other hand, record the Rayleigh waves induced by applied load and propagating through the pavement surface. These tests allow the travel time between successive receivers to be calculated for different excitation frequencies by collected wavelength data. Such procedures are also referred to as spectral analysis of surface waves (SASW) and depend on the phase velocities and the excitation frequencies [46]. A determination of material characteristics, namely material stiffness of each pavement layer, is essential in conducting pavement design using mechanistic approaches. Theoretical models (such as the layered elastic theory and nite element methods) use the stiffness properties of pavement layers to calculate resulting strains. In order to obtain the structural condition (in terms of elastic stiffness) of a pavement structure using measured surface deections, it is necessary to characterize the inverse mapping of theoretical pavement response model. Such numerical models involving parameter identication routines are generally referred to as pavement backcalculation techniques [7]. Backcalculation methods can generally be grouped into three basic categories: adaptive, static, and dynamic.

422

A. Burak Goktepe et al. / Advances in Engineering Software 37 (2006) 421431

In existing studies, adaptive processes are identied with neural networks and neuro-fuzzy systems. Adaptive methods do not directly utilize a pavement response model, but rather simulate inverse mapping by learning the target behavior via known input-output data patterns [813]. As their names imply, static and dynamic methods are classied by loading type, and utilize conventional pavement response models. Hence, they employ two calculation directions, namely forward and backward. In the forward direction of analysis, deections are computed for given trafc loads and pavement structure. In the backward pass of computation, the calculated values are compared with deections measured by NDT device, and new mechanical properties are determined by a parameter identication routine. These optimization steps are conducted until the differences between calculated and measured deections stay under a certain error criterion. It should be mentioned that the inverse process can be performed by several techniques, such as the least-squares, the gradient descent method, database search, and genetic algorithm [4,1418]. The objective of this study is to categorically explain exible pavement backcalculation methods as well as to compare and contrast them in terms of modeling precision, computational expense, calculation details, and data requirements. Innovations and advances in backcalculating the mechanical properties of exible pavements are considered in this state-of-the-art study (from 1983 to 2005). In this regard, developments made after the last state-of-the-art study [16] as well as the earlier studies on the backcalculation of exible pavements are presented here. In this context, nondestructive testing, pavement backcalculation procedures, and evaluations on backcalculation methods are given in the following sections. 2. Nondestructive testing methods Highway engineers seek the answers to two essential questions when performing the structural evaluation of a pavement system: (1) how much the pavement system can serve within limits under estimated trafc loads? (2) What is the optimum strategy for the pavement management? The answers of these questions are based on an optimization problem considered by the combination of material behaviors, mechanical modeling, statistical evaluation, and economical equilibriums. Therefore, the determination of mechanical properties is crucial for the problem statement description. In NDT methods, the mechanical properties of exible pavement layers are measured at low strain levels. The idea behinds the NDT methodology for the structural performance of the pavement system is inversely proportional with the amount of surface deections emerged by the applied load. Contradictions to this approach are geophysical methods, such as the spectral analysis of surface waves (SASW), which contains pavement shear modulus calculated by Rayleigh wave velocities [56]. Virtually, the discrepancies among deection based NDT methods are based on variations in loading details (type, duration, and magnitude) and deection measurement locations. Focusing on the type of loading provides us with

three types of load applications, i.e. static, steady-state vibratory, and the time domain impulse. Static loading is the rst and the simplest case and cannot simulate the actual trafc loads; thus, in current deection basin tests, displacements are recorded along the pavement surface subjected to a steady state harmonic or a transient dynamic load [4,15,16]. The Benkelman beam and La Croix deectometer are NDT devices working with a static type of loading. Dynamic loading is precise and realistic to simulate the effect that is imposed by actual trafc loads. Dynamic tests are normally performed through the longitudinal direction of the pavement and measured deection basins characterize the structural integrity of the pavement system in accordance with the actual trafc loads that are dynamic in nature. The steady-state dynamic case is similar to the effect of the vehicle passing over the pavement section, and the loading period is related with the vehicles speed. Dynaect and Road Rater are two examples of popular steady-state dynamic loading devices. On the other hand, in the time domain impulse loading, an impulse load is applied on pavement surface and deection data is recorded in the time domain. Generally, several sensors are used to measure the deection values on different points of pavement surface. Falling Weight Deectometer (FWD) is a transient impulse loading device [3,1922]. In the FWD test, an impulse load within the range of [6.7156 kN] is impacted on the pavement surface, and related deection values are recorded in the time domain. The FWD load is induced by a circular plate as well as a rubber seal that is placed between the plate and the surface to avoid the instant impact effect. Generally, transient surface deections are measured at different locations (usually at six or seven locations) by geophones. Accordingly, peak values for each geophone are used to plot deection basin curve. Fig. 1 illustrates the typical result of an FWD test. As can be seen from the gure, peak deection is measured by the geophone directly below the load application point and deections are smaller for more distant geophones. Deection curves exhibit haversine behavior; therefore, they increase nonlinearly until the peak point and decrease after the peak by elapsing time. There are a number of FWDs commercially available, such as the Dynatest FWD, the KUAB FWD, and the Phoenix FWD. The FWDs simulate vehicular loading quite successfully, and have been preferred by several researchers [3,7,10,14,20,21,2325].
Deflection Maximum Deflection

Time

Radial distance

Fig. 1. Illustration of typical FWD deection graph.

A. Burak Goktepe et al. / Advances in Engineering Software 37 (2006) 421431

423

The primary disadvantage of FWD is that it has to stop at each test point for several minutes throughout the measuring process; thus, lowering the time efciency. The rolling weight deectometer (RWD), on the other hand, can measure deections at a certain speed and overcomes the major drawback of FWD. For this reason, among all the NDT methods, RWD is the fastest way to collect deection data. RWD utilizes a screening device to locate pavement sections with similar deections and consists of four noncontact optical load tires. Basically, RWD sensors perform two measurements at the same time: (a) The distance to the pavement and (b) vertical displacement of the sensor with respect to the laser beam. RWD can produce thousands of deection data measurements per hour, which greatly exceed the approximate thirty measurements per hour measured by the FWD [2]. In addition to the above explained large-scale deection basin NDT methods, there are also several small-scale (portable) dynamic plate test devices available; these measure the elastic stiffness modulus of considered foundation material. The most popular small-scale devices are: (a) German dynamic plate bearing test (GDPBT), (b) TRL foundation tester (TFT), and (c) Loadman. Basically, they are similar in terms of the mechanics of operation, but transducers utilized and analysis procedure lead to several variations in the measured outcomes [26]. The wave propagation tests record the stress waves propagating along the surface of the pavement due to applied load. By computing the travel time between adjacent receivers for different excitation frequencies, a dispersion curve is obtained through correlating phase velocities with wavelengths. As a common technique used in this kind of tests, waves are monitored and recorded by receivers, while time signals are transformed to the frequency domain. The thickness and stiffness of considered pavement layers are then obtained by an inversion process based on the propagation of generalized plane surface waves of the Rayleigh type. This technique is represented by the SASW method [27]. The most popular surface wave propagation test is the seismic pavement analyzer (SPA). The SPA method employs two types of pneumatic impact hammers to produce excitations possessing short and long wavelengths. SPA has a hammer with a contact load about 2.67 kN and with frequency up to 12,000 kHz. On the other hand, there are other geophysical testing devices, such as ground penetrating radar (GPR), used to evaluate the pavement quality. GPR is a pulse-echo method designed to locate structural objects and to evaluate layer properties. It functions like an ultrasound machine, but utilizes radio waves instead of sound waves [3]. Several researchers carried out extensive research to understand the different aspects of these techniques. However, the main disadvantage is that testing is performed rapidly and the obtained moduli are at low strain levels (in the range of strains where the moduli are not strain dependent). Accordingly, seismic wave methods are constrained in terms of the magnitudes of the loads used to generate seismic waves. Since pavement deformation is not the essential parameter in these

methods, low amplitude excitations are sufcient to generate these waves [3,6,2832]. 3. Backcalculation of exible pavements Backcalculation problems, also known as parameter identication problems, are prevalently used in many scientic disciplines. Basically, it is an optimization process performed to obtain inverse mapping of a known relation established by discrete or continuous data points. The backcalculation process in pavement system is the numerical analysis of measured surface deections performed for the estimation of layer stiffness parameters (namely, pavement moduli). In order to accomplish this, measured deections are matched with calculated deections. Deection calculations are conducted using an equivalent pavement response model with synthetic moduli. Usually, the matching process is performed by an iterative process, in which deection values are calculated using different set of moduli. Consequently, iterations are continued until a close match between measured and calculated deection values are satised [4,15,16,32,33]. Available sources reveal that numerous backcalculation techniques have been developed for the backcalculation of pavement layer moduli. The discrepancies among these models are related with the type of the pavement (forward) response model and the optimization procedure carried out for the determination of appropriate layer modulus values. On the basis of classications and implementations made by several researchers, the general classication of backcalculation methods is presented in Fig. 2 [4,1418,3436]. As can be seen from Fig. 2, pavement response analysis (calculation of deections) can be considered as either static or dynamic. Initial studies of this problem were based on static analysis, and attempts were made to establish the inverse correlation with functional, statistical, and empirical approaches. However, due to the inherent, sophisticated, and nonlinear nature of the problem, these attempts were not successful. In all static approaches, conventional methods are based on iterative optimization procedures, and forward pavement response is calculated using either layered elastic theory or nite element method (FEM) for linear or nonlinear elastic material behaviors. Since the nonlinearity is closer to the nature of pavement materials, it is obvious that the nonlinear analysis increases the precision of the forward model. In addition, optimization processes can be performed by using a parameter identication algorithm (PIA), such as nonlinear least squares, database search algorithm (DSA), and genetic algorithm (GA). It should be noted that GA is an AI (articial intelligence)-based model-free optimization technique, which mimics the theory of evolution [3,4,1620,36,37]. Schematic representation of static linear elastic conventional backcalculation methods with PIA, GA, and DSA optimization techniques are illustrated in Fig. 3. It should also be noted that DSA based backcalculation employs the pattern searching algorithm through optimization routines and utilizes a predetermined database instead of calculating deections in the each step of the optimization [15,16].

424

A. Burak Goktepe et al. / Advances in Engineering Software 37 (2006) 421431

Backcalculation Methods

Static

Adaptive

Dynamic

Closed Form

Graphical

Conventional

Empirical

Impulse Load

Vibratory Load

Linear

Nonlinear Frequency Domain Time Domain Steady State

Genetic Algorithm

Parameter Identification

Database Search

Linear

Nonlinear

Genetic Algorithm
Fig. 2. Overview of backcalculation methods.

Parameter Identification

3.1. Static backcalculation procedures The idea of calculating strains within the semi-innite layered elastic media developed by applied static loads goes back to the prior work of Burmister [38]. In this earlier attempt,
(a) Measured Peak Deflections R = max k R= max Load k Peak Applied I = max P Poisson ratio, Layer thickness , h

Burmister used Boussinesqs equations with several assumptions for two and three-layered elastic half-space. Due to layered elastic theory, for axisymmetric stress distribution and cylindrical coordinates, stresses and strains are characterized with a fourth-order differential equation involving four

Forward Analysis

k (E)

PIA, GA max k k(E) * max P Backward Analysis E<Changes> Error Criterion Satisfied

: Number of sensors k PIA : Parameter Identification Algorithm GA : Genetic Algorithm

E (b) Measured Peak Deflections R = max k Database R= max Load k Peak Applied I = max P Poisson ratio, Layer thickness , h Forward Pattern Searching Algorithm Backward Analysis max k k(E) * max P E<Changes> Error Criterion Satisfied

k: Number of sensors

E
Fig. 3. Schematics of static linear conventional backcalculation with (a) PIA-GA and (b)DSA.

A. Burak Goktepe et al. / Advances in Engineering Software 37 (2006) 421431

425

Measured Peak Deflections R = max k Peak Applied Load I = max Pj Poisson ratio, Layer thickness , h k : Number of sensors j : Number of load levels

G* = E(w) + iE(w)
Forward k (E)

E1, m, h1 E2, 2, h2 E3, 3, h3 E4, 4, h4

Asphalt Concrete Base Subbase Subgrade

G* = E(w) + (1 +ib)
PIA max kj k(E) * max Pj E<Changes> Backward Error Criterion Satisfied

G* = E(w) + (1 +ib) G* = E(w) + (1 +ib)

Fig. 5. Typical dynamic multilayer pavement model.

Fig. 4. Schematic representation of static nonlinear backcalculation with PIA.

integration constants that are determined by boundary and continuity conditions. Details of layered elastic theory and related forward calculation programs can be found elsewhere [22,39]. In the backcalculation process undertaken by static loading data, only the peak load application is utilized because of the linear elastic material assumption. Therefore, loading levels are not considered in static material behavior and in the related pavement response model, which is inherent to the nature of actual (in situ) load application. In order to compute more realistic deection values, the nonlinearity in the material behavior as well as using different load levels existing in the test method can be considered. Namely, in nonlinear analysis, measured deections are matched with the calculated deections for all load levels instead of a peak value [16]. In the light of this gure, Fig. 4 demonstrates the difference between linear and nonlinear static backcalculation techniques (for PIA routine). Analogous to static linear case, the stiffness of each layer (i.e. pavement moduli, E) is changed by a parameter identication algorithm (PIA) to match calculated (() and measured deections (d). Poissons ratio (n) and the thickness (h) of each layer are considered to be constant in this study. Nevertheless, Poissons ratio of each layer can be taken into account as an additional design variable, although constant assumption does not result in serious mistakes [4,16]. 3.2. Dynamic backcalculation procedures Dynamic pavement response models have also been adapted to the forward calculations of backcalculation procedures by several researchers [1,4,16,19,4054]. Basically, the dynamic response of a pavement depends on the elastic moduli, thickness, Poissons ratios, mass densities (r), and damping ratios (b) of each layer. Actually, the variations in Poissons ratios, mass densities, and damping ratios have only slight inuence on the dynamic response of the pavement; thus, their values are generally assumed to be known. Commonly, the unknown parameters in a dynamic backcalculation analysis are the complex moduli (G*) and the thickness of the pavements layers. The complex modulus is the function of angular frequency (u) and the three material properties, i.e. slope of creep compliance curve (m), internal damping ratio, and Youngs modulus (E). Creep compliance is a viscoelastic

property that is related with the asphalt concrete (AC) layer as well as internal damping is the function of inertia, which is considered for base, subbase, and subgrade layers in elastodynamic analyses [4,16,4244]. A typical illustration of four-layer linear dynamic pavement model is given in Fig. 5. Obviously, Youngs modulus is the fundamental property of elasticity and is accounted for all exible pavement layers. The viscoelastic properties of AC layer can be characterized by creep compliance dened in the time domain, and the dynamic complex modulus considered in the frequency domain [22,40,42,54]. For remaining layers (base and subgrade), complex modulus is generally assumed independent of the frequency, and several material models, such as Kelvin and Maxwell, can be utilized to characterize these layers [22,4245]. As emphasized before, loading is performed as either impulsive or vibratory in dynamic response analyses. Therefore, deection data can be obtained in the time domain for the time domain impulse loads, and in the frequency domain for steady state vibratory loads. In order to use the dynamic loading data, Fourier analyses are usually conducted for the transformation of the domain. Then, elastodynamic analysis, such as Green function solution and dynamic FEMs, are performed to calculate resulting surface deections. Fig. 6 depicts the schematical representation of dynamic backcalculation for steady-state vibratory loads [4,16]. Referring again to Fig. 6, since loading and deection data are in the frequency domain for steady-state vibratory tests, there is no need for a transformation in this type of backcalculation analysis. Nevertheless, for impulse loads as in FWD test, the time (t) domain data must be transformed to the frequency (u) domain data. In this context, there can be two
Deflection Functions R() = k e i t+k Load Functions I() = P e i t
Poissons ratio, Layer Thickness, Density, Elastic Modulus and Dynamic Parameters (, h, , E, m, )

Forward Analysis

k (G*)

R() k (G*) * I() Forward G*<Changes> Error Criterion Satisfied

G* f (m, E, , ) k: Number of sensors

Analysis

G*

Fig. 6. Schematic Representation of dynamic steady-state vibratory load backcalculation.

426

A. Burak Goktepe et al. / Advances in Engineering Software 37 (2006) 421431

Deflection Histories R(t) = k (t) Load Time History I(t) = P (t) Poissons ratio, Layer Thickness, Density, Elastic Modulus and Dynamic Parameters (, h, , E, m, )

Fast Fourier Transformation (FFT) Fast Fourier Transformation (FFT)

k (j)

k (j) P (j)

P (j)

Backward Analysis

k (G*) PIA k0 (j) k (G*) G* f (m, E, , ) k: Number of sensors

Error Criterion Satisfied G*

G*<Changes>

Fig. 7. Schematic representation of dynamic the frequency domain tting for impulse load backcalculation.

possibilities, i.e. (a) the frequency domain tting and (b) the time domain tting. In the time domain tting, since computed deections are in the frequency domain, inverse Fourier transformation should be carried out to compare calculated and measured deections [16]. Details of impulse load dynamic backcalculation for both cases are explained in Figs. 7 and 8. It should be underlined for dynamic backcalculation analyses that the consideration of material behaviors in nonlinear manner drastically increases the computational complexity. Therefore, for the majority of dynamic backcalculation analyses, related

algorithms take linear material behavior into account, instead of nonlinear. Furthermore, similar to the static backcalculation methods, optimization process of dynamic backcalculation can be performed by parameter identication or genetic algorithm routine [4,16,17,37]. Obviously, the fundamental discrepancies among dynamic backcalculation procedures come from the philosophies behind the utilized forward response models. In essence, dynamic pavement (forward) response models fall into two categories depending on the solution methodologies, i.e., analytical and numerical. Analytical methods are successful in simulating

Deflection Histories R(t) = k (t) Load Time History I(t) = P (t) Poissons ratio, Layer Thickness, Density, Elastic Modulus and Dynamic Parameters (, h, , E, m, ) Backward Analysis Fast Fourier Transformation (FFT) Fast Fourier Transformation (FFT) P (j) = I (j) Forward Analysis k (G*) RC () = k (G*) *I (j) G* f (m, E, , ) k: Number of sensors

RC (t)

Error Criterion Satisfied

G*

RC (t) R PIA

G*
Fig. 8. Schematic representation of dynamic the time domain tting for impulse load backcalculation.

Forward Analysis

k0 (j)

A. Burak Goktepe et al. / Advances in Engineering Software 37 (2006) 421431

427

the wave propagation in elastic media; nevertheless, they need to be applied under certain geometric and boundary conditions [10,47]. The rst efcient analytical solution implemented for the characterization of the wave propagation in layered elastic media was developed by Kausel and Roesset [55]. In this pioneering study, they developed layer stiffness matrices by improving the transfer matrix concept introduced by Haskell [56] and generalized the lumped mass method of Lysmer [57]. Therefore, Kausel and Roesset presented both semi-discrete and continuous analyses to solve the equations of motion. Continuous solutions are based on the functional description of the wave motion in the time domain and the numerical integration of obtained dynamic equation. Semi-discrete solutions depend on the interpolation of wave propagation in the horizontal direction by considering sublayers with respect to the wavelengths of Rayleigh waves. However, their solutions were found to be instable with respect to several complications in numerical integrations for continuous analyses as well as robust due to computational complexity for semi-discrete analyses [47,58]. In essence, the solutions of Kausel and Roesset [55] are based on stiffness matrices establishing the relationship between stresses and strains at the top and bottom of each layer. Then, Fourier superposition is applied to extend the plane stresses and strains to axisymmetric conditions. Consequently, pavement response to a unit vibratory steadystate load is computed by Green function solutions characterized by means of transcendental functions integrals [54]. In this context, Roesset and Shao [59], Chang et al. [60], Magnuson [61], Kang [33] implemented several dynamic pavement analyzing programs, namely PUNCH, UTFWD, SCALPOT, and GREEN respectively, while some other researchers employed the above outlined dynamic analysis of Kausel and Roesset [33,62,63]. Apart from these, Hardy and Cebon [64] employed the theory of convolution in numerical computations to characterize the strains developed under moving dynamic loads for varying speeds and frequencies. Another analytical method was developed by Nilsson et al. [44] involving viscoelastic parameters of AC layer. In this model, they adopted Burgers method to characterize the shear modulus and considered linear elastic (frequency independent) bulk modulus. Authors also implemented the viscoelastic pavement analysis program named VEROAD [44]. In another notable study, Kang [33] developed a multi-frequency pavement analyzing tool focusing on dynamic Green exibility inuence functions. Additionally, Kang calculated surface deections considering viscoelastic aspects. Apart from these, Liang and Zhu [45] applied the modied Vlasov model under dynamic conditions comprising an AC layer over Vlasov subgrade. They considered creep compliance, damping ratio, depth to rigid bottom, and fatigue cracking issues in their notable study [45]. Eventually, Liang and Zeng [52] presented an analytical solution technique for the simulation of transient wave propagation based on axisymmetric dynamic loads. Basically, in these closed-form solutions, they used both LaplaceHankel transformation and transfer matrix methodology to calculate strains in the time domain [52].

On the other hand, numerical methods, which are commonly identied with FEM, overcome the obstacles related with the complexities in geometric shapes and boundary conditions. Several studies used FEM to analyze exible pavement structure [46,47,52]. FEM is theoretically a powerful tool for the dynamic analysis of a layered stratum; nevertheless, those using it must avoid the errors that arise from the reection of waves at the boundaries of nite elements. In order to overcome this obstacle, large domains and model sizes should be considered in the model [53]. Additionally, discretization of the domain and a methodology based on the calculation of internal forces element by element is obviously computationally expensive. Therefore, despite the success and the general analysis ability of FEM in the solution of complex problems, it usually requires extensive computational effort [10,46,47,53]. In order to perform an efcient and precise backcalculation analysis, Al-Khoury et al. utilized axisymmetric layer and halfspace spectral elements to characterize the dynamic behavior of exible pavements by [47]. Essentially, the spectral element technique is the combination of FEM with wave propagation basics; however, each layer is represented by one element instead of subdivisions. It is obvious that this overcomes the computational drawback of FEM for use with inverse problems that requires iterative computations [47,48]. In addition, this same group also considered viscoelastic aspects for AC layer to realistically model the exible pavement system [49]. Therefore, they adopted Burgers model into spectral element technique to characterize the viscoelastic material behavior with a differential operator. Eventually, Al-Khoury et al. [50] completed their spectral element based forward-backward calculation procedure by considering poroelastic behavior for the base layer. In this context they utilized the Biots theory of wave propagation in a saturated porous stratum. As a result, spectral element method is quite successful in simulating dynamic pavement response as well as making required calculations quickly involving viscoelastic and poroelastic aspects [4750]. 3.3. Adaptive backcalculation procedures The last backcalculation technique, adaptive backcalculation, is fundamentally different from traditional techniques. In the adaptive technique, the two-phased (forward and backward) structure of traditional backcalculation approaches is combined into one step utilizing a supervised learning algorithm. Namely, an adaptive system is taught by known input-output patterns, and it simulates the nonlinear mapping between input and output spaces as a functional approximater. A basic illustration of an adaptive system is given in Fig. 9.
Adaptive System Output Error Calculation

Input

Target

Training Parameter Identification Algorithm

Error

Fig. 9. Basics of adaptive system.

428

A. Burak Goktepe et al. / Advances in Engineering Software 37 (2006) 421431

(a) R = max k E I = max Pj , h

Weights <Changes> k j EM ES (b) moduli R = max k I = max Pj , h : Number of sensors : Number of load levels : In-situ (measured) moduli : Synthetic (calculated) moduli )

Training Algorithm

Error

EM ,ES Error Criterion Satisfied E

N N N

f f f E

Parameters <Changes> k j EM ES : Number of sensors : Number of load levels : In-situ (measured) moduli : Synthetic (calculated) moduli

Training Algorithm

Error

EM ,ES Error Criterion Satisfied E

Fig. 10. Illustration of (a) ANN- and (b) ANFIS-based static nonlinear backcalculation.

This idea was rst introduced by Meier and Rix [8], who applied articial neural networks (ANN) for the SASW test data inversion and the backcalculation of exible pavement layer properties. Successively, Meier and Rix [9] veried the susceptibility of ANN methodology for pavement moduli backcalculation utilizing FWD data. Later, they published the complementary article comprising the dynamic aspects and the rigid bottom depth concepts [10]. Apart from these, several other studies focusing on ANN-based pavement backcalculation models were carried out [1113,36,6567]. An example ANN-based backcalculation model for nonlinear elastic material behavior and static loading is indicated in Fig. 10a. It should be noted that ANN can solely learn the mapping characterized by inputoutput patterns; therefore, underlying material model and mechanical analysis do not exist in ANNbased backcalculation. The performance, then, of ANN-based backcalculation is based on the quality and quantity of training data [65]. There are other adaptive learning methodologies that can be used for the backcalculation problem. In this context, neurofuzzy inference can also be employed to backcalculate pavement moduli with existing input-output data expressing the target behavior [13]. Similarly, adaptive neuro-fuzzy

inference system (ANFIS) can be utilized to backcalculate pavement moduli. A typical ANFIS-based backcalculation procedure is schematically depicted in Fig. 10b. Nevertheless, fuzzy inference and the adaptation of fuzzy system are not appropriate for large numbers of input-output patterns and detailed input space partitionings. However, it can be a good choice for a small amount of training data involving considerable amount of uncertainty [68]. It should be added for adaptive backcalculation that training data may involve either in situ (measured) or synthetic (calculated) pavement moduli. The training process can be performed by either experimental data to characterize specic test section or with synthetically collected data to inversely simulate the pavement response model. The fundamental advantages of adaptive backcalculation methodologies are that they present real-time backcalculation ability and precise analysis results. 4. Evaluations on backcalculation methods It is obvious that dynamic backcalculation analysis is more advantageous than the static approach. First of all, it is appropriate for the inherent nature of the problem; thus, precise

A. Burak Goktepe et al. / Advances in Engineering Software 37 (2006) 421431

429

results could be obtained using dynamic approach. In addition, not only peak deformation values, but also the entire deection record is used in the calculations of dynamic pavement response analysis. Furthermore, viscoelastic properties of AC layer can efciently be considered in dynamic analysis. Eventually, the thickness of subgrade, which is assumed semi-innite in static analyses, can be considered in dynamic model [4,15,16,19,33,6063,69,70]. Besides the existing advantages of the dynamic approach, it has also several obstacles originating from the complexity and the computational expense of dynamic analyses. Nevertheless, in many problems it is tedious to get all necessary data required for a dynamic analysis. Consequently, static approaches are generally preferred in the majority of pavement backcalculation studies, because of their simplicity and acceptable error ranges. Particularly in cases when the depth to the rigid bottom is greater than 1215 m, the discrepancy between static and dynamic approaches may be assumed to be acceptable [19,60, 63,7175]. In essence, the precision of a backcalculation analysis is sensitive to type and coverage of pavement response analysis; nevertheless, the efciency is affected by the optimization methodology employed in the backward analysis performed to match calculated and measured deections. Therefore, in order to increase the performance of the backcalculation analysis, it is necessary to use an optimization technique that is capable of overcoming local minima problems as well as derivative-based errors, while also being fast and precise. In this context, several parameter identication techniques (linear least squares, gradient descent, NewtonRaphson, and GaussNewton methods), pattern searching method (Hooke Jeeves algorithm), and evolutionary computation (GA) were utilized in the developed backcalculation algorithms so far. Apart from the advantage of GA on preventing the local minima problem and derivative-based errors, GA-based optimization produces almost identical outcomes in comparison with least-squares methods. However, GA-based backcalculation is ineffective in terms of computing speed [3,4,15 17,20,3537,7678]. There is no underlying structure of adaptive backcalculation methods to calculate mechanical pavement responses; therefore, they are limited in terms of size, reliability, range, and distribution of training data in terms of characterizing the considered behavior. This is why they must be applied carefully and cannot be shifted with mechanical approaches such as elastodynamic analysis. Together with these; adaptive methods are capable of supplying real-time backcalculation and providing accurate outcomes with the help of their outstanding modeling abilities. It is obvious that these advantages of adaptive backcalculation methods provide signicant earnings for pavement engineers in terms of time and money. The spectral element method is probably the most promising forward analysis method for use in backcalculation analyses. Moreover, it is possible to adopt viscoelastic and poroelastic material behaviors into spectral element analysis with small modications. Therefore, precise and efcient backcalculation

analyses can be performed with the spectral element method. The hybrid Powel optimization method can be preferred in the optimization process of backcalculation procedures using spectral element analysis [4750]. FEM enables the general solution of exible pavement system subjected to dynamic loads. Nevertheless, the computational expense of this technique makes it tedious to use in backcalculation problems. In order, thus, to develop a complex and specic pavement model involving three-dimensional variations and nonlinear material properties, the FEM approach should be preferred [33,7981]. 5. Conclusions This paper addresses the advances in backcalculation of the mechanical properties of exible pavements. Within this content, the following conclusions were drawn in this study: With the consideration of the actual trafc loads as well as the material behavior of asphalt concrete, dynamic and viscoelastic pavement response analyses give precise outcomes. When the rigid bottom depth is higher than approximately 15m and practical considerations are valid for the problem at hand, the static approach can be applied to forward pavement analysis. Otherwise, static pavement response analysis and resulting backcalculation process may be misleading. Adaptive backcalculation methods perform real-time backcalculation analyses as a functional mapper. Nevertheless, there is no underlying mechanical background in these techniques; therefore, they must be applied carefully. FEM presents generalized problem solving ability, where complex geometries and boundary conditions are involved. In addition, nonlinear material properties and threedimensional analyzing are possible with FEM. However, the computational expense of FEM makes it tedious to use in backcalculation problems. The spectral element method is capable of performing both precise as well as fast pavement response analyses. Moreover, viscoelasticty and poroelasticity can also be considered in spectral element method to obtain more realistic outcomes. The optimization technique is also the complementary issue for a backcalculation analysis. For this reason, the preference of correct optimization algorithm inuences speed and precision of the backcalculation analysis.

References
[1] Callop AC, Cebon D. Stiffness reductions of exible pavements due to cumulative fatigue damage. J Transp Eng ASCE 1996;122:1319. [2] Briggs RC, Lukanen EO. Variations in backcalculated pavement layer moduli in LTTP seasonal monitoring sites. In: Tayabji SD, Lukanen EO, editors. NDT of pavements and backcalculation of moduli: 3. Special technical publication, STP 1375. Pennsylvania: ASTM Publication; 2000. p. 11328.

430

A. Burak Goktepe et al. / Advances in Engineering Software 37 (2006) 421431 [23] Tholen O, Sharma J, Terrel RL. Comparison of FWD with other deection testing devices. Transportation research record 1007, TRB. Washington, DC: National Research Council; 1985 p. 2026. [24] Newcomb, DE. Development and evaluation of regression method to interpret dynamic pavement deections, PhD Dissertation, department of civil engineering. Seattle: University of Washington 1986. [25] Bentsen RA, Bush A, Harrison JA. Evaluation of NDT equipment for aireld pavements, technical report, GL-89-3. Vicksburg: US Army Engineer Waterways Experiment Station; 1989. [26] Fleming PR. Small-scale dynamic devices for the measurement of elastic stiffness modulus on pavement foundations. In: Tayabji SD, Lukanen EO, editors. Nondestructive testing of pavements and backcalculation of moduli, 3. Special technical publication, STP 1375. Pennsylvania: ASTM Publication; 2000. p. 4158. [27] Nazarian S, Stokoe KH. Use of surface waves in pavement evaluation. Transportation research record, 1070, TRB. Washington, DC: National Res. Council; 1986 p. 132144. [28] Nazarian S, Stokoe KH, Hudson WR. Use of spectral analysis of surface waves method for determination of moduli and thicknesses of pavement systems. Transportation research record, 921, TRB. Washington, DC: National Research Council; 1983 p. 3845. [29] Shao KY. Dynamic interpretation of dynaect, falling weight deectometer and spectral analysis of surface waves tests on pavement system, PhD Dissertation. Texas: University of Texas at Austin 1985. [30] Sanchez-Salinero I. Analytical investigation of seismic methods used for engineering applications, PhD Dissertation, department of civil engineering, Texas: University of Texas at Austin 1987. [31] Sheu JC, Stokoe KH, Roesset JM. Effects of reected waves on SASW testing of pavements. Transportation research record, TRB, 1196. Washington, DC: National Research Council; 1988 p. 5161. [32] Kang YV. Effect of nite width on dynamic deections of pavements, PhD Dissertation, department of civil engineering. Texas: Univ. of Texas at Austin 1990. [33] Kang YV. Multi-frequency backcalculation of pavement layer moduli. J Transp Eng ASCE 1998;124(1):7381. [34] Chang, DW. Nonlinear effects on dynamic response of pavements using the NDT technique, PhD Dissertation, department of civil engineering, Austin: University of Texas at Austin 1991. [35] Harichandran RS, Yeh MS, Baladi GY. MICH-PAVE: a nonlinear nite element program for analysis of exible pavements. Transportation research record 1286, TRB. Washington, DC: National Research Council; 1990 p. 123131. [36] Terzi S, Saltan M, Yildirim T. Optimization of deection basin by genetic algorithm and neural network approach. International Conference on Neural Networks, ICANN/ICONIP 2003, LNCS 2714, 2003: 662669. [37] Reddy MA, Reddy KS, Pandey BB. Selection of genetic algorithm parameters for backcalculation of pavement moduli. Int J Pavement Eng 2004;5(2):8190. [38] Burmister DM. The general theory of stresses and displacements in layered soil systems. J Appl Phys 1945;23:8494. [39] Monismith CL. Analytically based asphalt pavement design and rehabilitation: theory to practice. Transportation research record, 1345, TRB. Washington, DC: National Research Council; 1992 p. 526. [40] Sousa JB, Monismith CL. Dynamic response of paving materials Transportation research record, 1136, TRB. Washington, DC: National Research Council; 1987 p. 5768. [41] Magnuson AH, Lytton RL, Briggs RC. Comparison of computer predictions and eld data for dynamic analysis of falling weight deectometer data. Transportation research record, 1293, TRB. Washington, DC: National Research Council; 1991 p. 124135. [42] Stubbs N, Torpunuri VS, Lytton RL, Magnuson AH. A methodology to identify material properties in pavements modeled as layered viscoelastic half-spaces. In: Von Quintus HL, Bush AJ, Baladi GY, editors. NDT of pavements and backcalculation of moduli, 2. Special technical publication, STP 1198. Pennsylvania: ASTM Publication; 1994. p. 159 69. [43] Maina JW, Yokota H, Mnanga DA, Masuda S. Prediction of pavement deterioration based on FWD results. In: Tayabji SD, Lukanen EO, editors.

[3] Tawq K, Armaghani J, Sobanjo J. Seismic pavement analyzer vs. FWD for pavement evaluation: comparative study. In: Tayabji SD, Lukanen EO, editors. NDT of pavements and backcalculation of moduli, 3. Special technical publication, STP 1375. Pennsylvania: ASTM Publication; 2000. p. 32745. [4] Ullidtz P. Will nonlinear backcalculation help. In: Tayabji SD, Lukanen EO, editors. NDT of pavements and backcalculation of moduli, 3. Special technical publication, STP 1375. Pennsylvania: ASTM Publication; 2000. p. 1422. [5] Nazarian S, Stokoe KH. Nondestructive evaluation of pavements by surface wave method. In: Bush AJ, Baladi GY, editors. NDT of pavements and backcalculation of moduli, 1. Special technical publication, STP 1026. Pennsylvania: ASTM Publication; 1989. p. 119 37. [6] Seng CR, Stokoe KH, Roesset JM. Effect of depth to bedrock on the accuracy of backcalculated moduli obtained with dynaect and FWD tests. Special report, SP 1175-5. Cnter for transportation research. Texas: University of Texas at Austin; 1993. [7] Zhou H. Comparison of backcalculated and laboratory measured moduli on AC and granular base layer materials. In: Tayabji SD, Lukanen EO, editors. NDT of pavements and backcalculation of moduli, 3. Special technical publication, STP 1375, vol. 337. Pennsylvania: ASTM Publication; 2000. p. 16172. [8] Meier RW, Rix GJ. An initial study of surface wave inversion using articial neural networks. ASTM Geotechn Test J 1993;16:42531. [9] Meier RW, Rix GJ. Backcalculation of exible pavement moduli using articial neural networks. Transportation research record, 1448, TRB. Washington, DC: National Research Council; 1994 p. 7582. [10] Meier RW, Rix GJ. Backcalculation of exible pavement moduli from dynamic deection basins using articial neural networks. Transportation research record, 1473, TRB. Washington, DC: National Research Council; 1995 p. 7281. [11] Tutumluer E, Seyhan U. Neural network modeling of anisotropic aggregate behavior from repeated load triaxial tests. Transportation research record, 1615. Washington, DC: National Research Council; 1998 p. 8693. [12] Kim YR, Xu B, Kim Y. A new backcalculation procedure based on dispersion analysis of FWD time history deections and surface wave measurements using ANNs. In: Tayabji SD, Lukanen EO, editors. NDT of pavements and backcalculation of moduli, 3, STP 1375. Pennsylvania: ASTM Publication; 2000. p. 297312. [13] Saltan M. Modeling deection basin using neurofuzzy in backcalculating exible pavement layer moduli. Pak J Inf Tech 2002;1:1807. [14] Hoffman MS, Thompson MR. Backcalculating nonlinear resilient moduli from deection data. Transportation research record, 852, TRB. Washington DC: National Research Council; 1982 p. 4251. [15] Lytton RL. Backcalculation of layer moduli, state of the art. In: Bush AJ, Baladi GY, editors. NDT of pavements and backcalculation of moduli, 1. Special technical publication, STP 1026. Pennsylvania: ASTM Publication; 1989. p. 738. [16] Uzan J. Advanced backcalculation techniques. In: Von Quintus HL, Bush AJ, Baladi GY, editors. NDT of pavements and backcalculation of moduli, 2. Special technical publication, STP 1198. Pennsylvania: ASTM Publication; 1994. p. 337. [17] Fwa TF, Tan CY, Chan WT. Backcalculation analysis of pavement-layer moduli using genetic algorithms. Transportation research record 1570, TRB. Washington DC: National Research Council; 1997 p. 134142. [18] Reddy MA, Reddy KS, Pandey BB. Backcalculation of pavement moduli using genetic algorithms. J Highw Res Board 2002;66:110. [19] Mamlouk MS. Use of dynamic analysis in predicting eld multilayer pavement moduli. Transportation research record, 1043, TRB. Washington, DC: National Research Council; 1985 p. 113119. [20] Sebaaly B, Davies TG, Mamlouk MS. Dynamics of falling weight deectometer. J Transp Eng ASCE 1985;111:61832. [21] Stolle DFE. Modeling of dynamic response of pavements to impact loading. Comput Geotech 1991;11(1):8394. [22] Huang YH. Pavement analysis and design. New Jersey: Prentice Hall Inc; 1993.

A. Burak Goktepe et al. / Advances in Engineering Software 37 (2006) 421431 NDT of pavements and backcalculation of moduli, 3. Special technical publication, STP 1375. Pennsylvania: ASTM Publication; 2000. p. 95 109. Nilsson RN, Oost I, Hopman PC. Visco-elastic analysis of full scale pavements: validation of VEROAD. Transportation research record, 1539, TRB. Washington, DC: National Research Council; 1996 p. 8187. Liang RY, Zhu JX. Efcient computational algorithms for forward and backward analysis of a dynamic pavement system. Comput Struct 1998; 69:25563. Zaghloul SM, White TD. Use of a three dimensional dynamic nite element program for the analysis of exible pavement. Transportation research record, 1388, TRB, Washington, DC 1993 p. 6069. Al-Khoury R, Scarpas A, Kasbergen C, Blaauwendraad J. Spectral element technique for efcient parameter identication of layered media. I. Forward calculation. Int J Solids Struct 2001;38:160523. Al-Khoury R, Scarpas A, Kasbergen C, Blaauwendraad J. Spectral element technique for efcient parameter identication of layered media. II. Inverse calculation. Int J Solids Struct 2001;38:875372. Al-Khoury R, Scarpas A, Kasbergen C, Blaauwendraad J. Spectral element technique for efcient parameter identication of layered media. III. Viscoelastic aspects. Int J Solids Struct 2002;39:2189201. Al-Khoury R, Kasbergen C, Scarpas A, Blaauwendraad J. Poroelastic spectral element for wave propagation and parameter identication in multi-layer systems. Int J Solids Struct 2002;39:407391. Dong QX, Hachiya Y, Takahashi O, Tsubokawa Y, Matsui K. An efcient backcalculation algorithm of time domain for large-scale pavement structures using Ritz vectors. Finite Elem Anal Des 2002;38:113150. Liang R, Zeng S. Efcient dynamic analysis of multilayered system during falling weight deectometer experiments. J Transp Eng ASCE 2002;128(4):36674. Shoukry SN, William GW. Performance evaluation of backcalculation algorithms through three dimensional nite-element modeling of pavement structures,1655, TRB. Washington, DC: National Research Council; 2000 p. 152160. Roesset JM. Stiffness and damping coefcients of foundations. In: ONeal MW, Dobry K, editors. Proceedings of ASCE national convention on dynamic response of pile foundations, analytical aspects, New York; 1980, 1980. p. 130. Kausel E, Roesset JM. Stiffness matrices for layered soils. Bull Seismol Soc Am 1981;71:174361. Haskell NA. The dispersion of surface waves on multilayered media. Bull Seismol Soc Am 1953;71:1734. Lysmer J. Lumped mass method for rayleigh waves. Bull Seismol Soc Am 1970;60:89104. Founquinos R, Roesset JM, Stokoe II KH. Response of pavement systems to dynamic loads imposed by nondestructive tests. Transportation research record, 1504, TRB, Washington, DC 1995 p. 5767. Roesset JM, Shao KY. Dynamic interpretation of Dynaect and falling weight deectometer tests. Transportation research record, 1022, TRB, Washington, DC 1985 p. 716. Chang DW, Kang YV, Roesset JM, Stokoe II KH. Effect of depth to bedrock on deection basins obtained with dynaect and falling weight deectometer tests. Transportation research record, 1355, TRB, Washington, DC 1992 p. 816. Magnuson AH. Computer analysis of falling weight deectometer data, Part I. Vertical displacement computations on the surface of a uniform surface pressure distribution, research report no: 1215-1F, Texas Transportation Institute, Collage Station; 1988. Mamlouk MS, Davies GD. Elasto-dynamic analysis of pavement deections. J Transp Eng ASCE 1984;110(6):53667. Uzan J. Dynamic linear backcalculation of pavement material parameters. J Transp Eng ASCE 1994;120(1):10925.

431

[44]

[45]

[46]

[47]

[48]

[49]

[50]

[51]

[52]

[53]

[54]

[55] [56] [57] [58]

[59]

[60]

[61]

[62] [63]

[64] Hardy MSA, Cebon D. Response of continued pavements to moving dynamic loads. J Eng Mech ASCE 1993;119(9):176280. [65] Tutumluer E, Meier RW. Attempt at resilient modulus modeling using articial neural networks. Transportation research record, 1560, TRB, Washington, DC; 1996 p. 16. [66] Abdallah I, Ferregut C, Nazarian S, Lucero MO. Prediction of remaining life of exible pavements with articial neural network models. In: Tayabji SD, Lukanen EO, editors. NDT of pavements and backcalculation of moduli, 3. Special technical publication, STP 1375. Pennsylvania: ASTM Publication; 2000. p. 48498. [67] Saltan M, Tigdemir M, Karasahin M. Articial neural network application for exible pavement thickness modeling. Turkish J Eng Environ Sci 2002;26:2438. [68] Goktepe AB, Agar E, Lav AH. Comparison of multilayer perceptron and adaptive neuro-fuzzy system on backcalculating the mechanical properties of exible pavements. ARI Bulletin of Istanbul Technical University; in press. [69] Uzan J, Scullion T. Verication of backcalculation procedures. Proceedings of third International Conference on Bearing Capacity of Roads and Airelds, Trondheim, Norway, June 812, 1990: 447458. [70] Ong CL, Newcomb ED, Siddharthan R. Comparison of dynamic and static backcalculation moduli for three-layer pavements. Transportation research record, 1293, TRB. Washington, DC: National Research Council; 2002 p. 8992. [71] Mamlouk MS. Use of dynamic analysis in predicting eld multilayer pavement moduli. Transportation research record, 1043, TRB. Washington, DC: National Research Council; 1986 p. 113119. [72] Hossain ASM, Zaniewski JP. Detection and determination of depth of rigid bottom in backcalculation of layer moduli from FWD data. Transportation research record, 1293, TRB. Washington, DC: National Research Council; 1991 p. 124135. [73] Mustaque ASM, Zaniewski JP. Detection and determination of depth of rigid bottom in backcalculation of layer moduli from falling weight deectometer. Transportation research record, 1293, TRB. Washington, DC: National Research Council; 1991 p. 124135. [74] Mera RF. Dynamic nondestructive testing of pavements, PhD Dissertation, department of civil engineering. Texas: University of Texas at Austin 1995. [75] Chatti K, Kim T. Effect of frequency-dependent asphalt concrete Layer moduli on pavement response. In: Tayabji SD, Lukanen EO, editors. NDT of pavements and backcalculation of moduli, 3. Special technical publication, STP 1375. Pennsylvania: ASTM Publication; 2000. p. 365 82. [76] Uzan J, Lytton RL, Germann FP. General procedure for backcalculating layer moduli. In: Bush AJ, Baladi GY, editors. NDT of pavements and backcalculation of moduli, 1. Special technical publication, STP 1026. Pennsylvania: ASTM Publication; 1989. p. 738. [77] Sivaneswaran N, Kramer SL, Mahoney JP. Advanced backcalculation using a nonlinear least squares optimization technique. Transportation research record, 1293, TRB. Washington, DC: National Research Council; 1991 p. 93102. [78] Harichandran RS, Ramon CM, Mahmood T, Baladi GY. MICHBACK users manual. Michigan: Michigan Department of Transportation; 1994. [79] Sidderharthan RV, Yao J, Sebaaly PE. Pavement strain from moving dynamic 3D load distribution. J Transport Eng ASCE 1998;124(6): 55766. [80] Huhtala M, Pihlajamaki J. New concepts on load equivalency measurements Proceeding of seventh international conference on asphalt pavements 1992 p. 194208. [81] Sousa JB, Weissman SL, Sackman JL, Monismith CL. Nonlinear elastic viscous with damage model to predict permanent deformation of asphalt concrete mixes Transportation research record, 1136, TRB. Washington, DC: National Research Council; 1993 p. 5768.

You might also like