Prediction of National Airport Pavement Test Facility Pavement Layer ...

Prediction of National Airport Pavement Test Facility Pavement Layer Moduli from Heavy Weight Deflectometer Test Data Using Artificial Neural Networks Kasthurirangan Gopalakrishnan Department of Civil, Construction and Environmental Engineering Iowa State University 192 Town Engineering Building Ames, IA 50011 [email protected]

ABSTRACT The National Airport Pavement Test Facility (NAPTF) was constructed to generate full-scale testing data to investigate the performance of airport pavements subjected to complex gear loading configurations of new generation aircraft. During the first test program, the NAPTF test sections were simultaneously subjected to Boeing 777 trafficking in one lane and Boeing 747 trafficking in another lane using the National Airport Pavement Test Machine. To monitor the effect of time and traffic on pavement structural responses, heavy weight deflectometer (HWD) tests were conducted on the trafficked lanes and the untrafficked centerline of flexible test sections as trafficking progressed. The primary objective of this study was to develop a tool for backcalculating NAPTF non-linear flexible pavement layer moduli from HWD data using artificial neural networks (ANN). A multi-layer, feedforward network that uses an error-backpropagation algorithm was trained to approximate the HWD backcalculation function. The synthetic database generated using the non-linear pavement finite element program, ILLI-PAVE, was used to train the ANN. Using the ANN, the asphalt concrete moduli and subgrade moduli were successfully predicted. Further research is required to develop ANN models for predicting the granular layer moduli. These results could be used to compare the relative effect of Boeing 777 and Boeing 747 trafficking on the elastic moduli and characterize the seasonal variation in moduli values. The same concept could also be used for backcalculating non-linear pavement moduli of highway pavements for input into mechanistic-empirical analysis and design. Key words: artificial neural networks—ILLI-PAVE—National Airport Pavement Test Facility— pavement moduli

Proceedings of the 2005 Mid-Continent Transportation Research Symposium, Ames, Iowa, August 2005. © 2005 by Iowa State University. The contents of this paper reflect the views of the author(s), who are responsible for the facts and accuracy of the information presented herein.

INTRODUCTION The Federal Aviation Administration’s (FAA) National Airport Pavement Test Facility (NAPTF) is located at the Atlantic City International Airport, New Jersey. It was constructed to generate full-scale test data needed to develop pavement design procedures for the new generation of large civil transport aircraft, including the Boeing 777 and Boeing 747. During the first series of tests, two gear configurations, a six-wheel tridem landing gear (Boeing 777) in one lane and a four-wheel dual-tandem landing gear (Boeing 747) in the other lane, were tested simultaneously. Heavy weight deflectometer (HWD) tests were conducted at regular time intervals as trafficking continued. The primary objective of this study was to develop a tool for backcalculating NAPTF non-linear pavement layer moduli from HWD test data using artificial neural networks (ANN). The elastic layer moduli backcalculated from non-destructive test results are good indicators of pavement layer condition (Xu, Ranjinathan, and Kim 2001) as well as required inputs for the a priori mechanistic design of a flexible pavement. The backcalculation approach is particularly appealing for characterizing subgrade soils that display large variability in subgrade modulus (as large as 35%–50% over few miles of a pavement) (Thompson, Tutumluer, and Bejarano 1998). Conventional elastic layer program (ELP)-based backcalculation software assumes that pavement materials are linear-elastic, homogenous, and isotropic. The non-linearity or stress-dependency of resilient modulus for unbound granular materials and cohesive fine-grained subgrade soils is well documented in the literature (Hicks 1970; Thompson and Robnett 1979). Previous studies have observed the nonlinearity of underlying layers at the NAPTF. Gomez-Ramirez and Thompson (Garg and Marsey 2002) reported the presence of material non-linearity at NAPTF by separately analyzing the individual layer compression from multi-depth deflectometer (MDD) readings. Garg and Marsey (2002) have similarly observed the stress-dependent nature of the granular and subgrade layers in NAPTF flexible test sections. Therefore, it is more realistic to use non-linear layer moduli for conducting NAPTF pavement structural analysis and for studying the variation in moduli with trafficking. ILLI-PAVE is a two-dimensional axi-symmetric pavement finite-element software developed at the University of Illinois at Urbana-Champaign (UIUC) (Raad and Figueroa 1980). It incorporates stresssensitive material models and provides a more realistic representation of the pavement structure and its response to loading (NCHRP 1990). Based on extensive repeated laboratory testing data at UIUC, Thompson and Robnett (1979) indicated that the breakpoint resilient modulus (ERi), typically associated with a repeated deviator stress of about six psi, is a good indicator of the subgrade soil’s resilient modulus. The Asphalt Institute’s Thickness Design Manual MS-1 (Asphalt Institute 1982) recommends ERi (subgrade modulus at a deviator stress of six psi) as the subgrade modulus input for ELP analysis. The ERi is also one of the subgrade material property inputs to ILLI-PAVE. However, there is no commercial ILLI-PAVE-based backcalculation program currently available. Previous research at UIUC showed that the asphalt concrete moduli and non-linear subgrade moduli (ERi) could be successfully predicted using an ANN trained with the ILLI-PAVE database (Gopalakrishnan and Thompson 2004). Ceylan et al. (2004) demonstrated the use of ANNs trained with ILLI-PAVE results as pavement structural analysis tools for the rapid and accurate prediction of critical responses and deflection profiles of flexible pavements subjected to typical highway loadings. Ongoing research at Iowa State University on backcalculting pavement moduli using ANN is being performed under Dr. Halil Ceylan’s supervision. However, the current research specifically focuses on developing a tool for backcalculating NAPTF pavement nonlinear moduli using ANN trained with ILLI-PAVE results.

Gopalakrishnan

2

NATIONAL AIRPORT PAVEMENT TEST FACILITY The NAPTF test pavement area is 900 feet (274.3 meters) long and 60 feet (18.3 meters) wide. The first test series installation included a total of nine test sections (six flexible and three rigid) built on three different subgrade materials: low-strength (target CBR of 4), medium-strength (target CBR of 8), and high-strength (target CBR of 20). Two different base sections are used in flexible test sections: conventional (granular) and stabilized (asphalt concrete). The low-strength and the medium-strength flexible test sections alone are considered in this study. The naturally occurring sandy soil material at the NAPTF site underlies each subgrade layer. Pavement Sections Each NAPTF test section is identified using a three-character code, where the first character indicates the subgrade strength (L for low, M for medium, and H for high), the second character indicates the test pavement type (F for flexible and R for rigid), and the third character signifies whether the base material is conventional-aggregate (C) or asphalt-stabilized (S). Thus, test section MFC refers to a conventionalbase flexible pavement built over a medium strength subgrade, whereas test section MFS refers to a stabilized-base flexible pavement built over a medium-strength subgrade. Cross-sectional views of the asbuilt NAPTF flexible test items considered in this study are shown in Figure 1. Note that P-401 asphalt concrete was used in the surface layer and in the stabilized base layer as well.

LFS

LFC 5 in.

AC Surface (P-401)

Granular Base (P-209)

7.75 in.

Asphalt Stab. Base (P-401)

Granular Subbase (P-154)

36.4 in.


AC Surface (P-401)

LOW Strength Subgrade Subgrade=94.7 in.

MFS

MFC 5 in. 4.9 in.

29.6 in.

AC Surface (P-401) Granular Base (P-209) Granular Subbase (P-154)

5 in.

5.1 in.

AC Surface (P-401)

7.9 in.

Asphalt Stab. Base (P-401)

4.9 in.


8.5 in.

12.1 in.

MEDIUM Strength

LOW Strength Subgrade

MEDIUM Strength Subgrade

Subgrade

Subgrade=104.5 in.

Subgrade=94.8 in.

Subgrade=101.6 in.

Figure 1. Cross-sectional views of as-built NAPTF flexible test sections Traffic Testing A six-wheel dual-tridem gear configuration (Boeing 777) with 54-inch (1,372-mm) dual spacing and 57inch (1,448-mm) tandem spacing was loaded on the north wheel track, while the south side was loaded with a four-wheel dual-tandem gear configuration (Boeing 747) having 44-inch (1,118-mm) dual spacing and 58-inch (1,473-mm) tandem spacing. The wheel loads were set to 45,000 lbs (20.4 tons) each and the tire pressure (cold) was 188 psi (1,295 Kpa). In the LFC (a conventional aggregate-base pavement built over low-strength subgrade) and LFS (an asphalt-stabilized base pavement built over low-strength subgrade) test sections, the wheel loads were increased from 45,000 lbs (20.4 tons) to 65,000 lbs (29.4 tons) after 20,000 initial load repetitions. Throughout the traffic test program, the traffic speed was 5 mph (8 kmh).

Gopalakrishnan

3

NAPTF Flexible Pavement Failure Criterion The NAPTF failure criterion is based on the criterion utilized by the U.S. COE MWHGL Tests (Ahlvin et al. 1971). It is defined as 1-inch (25.4-mm) surface upheaval adjacent to the traffic lane. This is considered to reflect a structural or shearing failure in the subgrade. Data Availability All test data referenced in this paper are available for download on the FAA Airport Pavement Technology website: http://www.airporttech.tc.faa.gov/naptf/. HEAVY WEIGHT DEFLECTOMETER TESTS For HWD testing, the FAA HWD KUAB Model 240, configured with a 12-inch loading plate and a 27– 30 msec pulse width, was used. The deflections were measured with seven seismometers at offsets of 0inch (D0), 12-inch (305-mm) (D12), 24-inch (610-mm) (D24), 36-inch (914-mm) (D36), 48-inches (1,219mm) (D48), and 60-inch (1,524-mm) (D60) intervals from the center of the load. The HWD tests were performed at nominal force amplitudes of 12,000 lbs or 12 kip (53.4 kN), 24,000 lbs or 24 kip (106.8 kN), and 36,000 lbs or 36 kip (160.2 kN). These tests were performed on the centerline, Boeing 777 traffic lane (Lane 2) and Boeing 747 traffic lane (Lane 5). The HWD test sequences were repeated at 10-foot (3.1-meter) intervals along the test lanes. The location and orientation of HWD test lanes are illustrated in Figure 2.

N

-30 ft. -15 ft.

B777

LANE 2

0 ft. 15 ft.

C/L B747

LANE 5

30 ft. Figure 2. NAPTF HWD test lanes Pavement Temperature The temperature of the asphalt concrete layer at the time of FWD testing has a significant influence on the surface deflections. During the NAPTF construction, static temperature sensors were installed at different depths along the test sections to record the pavement temperatures at different times of the day. The HWD

Gopalakrishnan

4

tests were all conducted between January 11, 2000 and June 6, 2001. The asphalt concrete temperature varied between 40 0F to 75 0F during the entire duration of traffic testing. DATABASE GENERATION USING ILLI-PAVE To generate the synthetic database for training the ANN, each NAPTF flexible test section was modeled in ILLI-PAVE. The as-constructed layer thicknesses (see Figure 1) were used for each test section. The individual pavement layers were characterized as follows. The asphalt concrete layer and the sand layer were treated as linear elastic material. Stress-dependent elastic models along with Mohr-Coulomb failure criteria were applied for the base, subbase, and subgrade layers. The stress-hardening K-θ model was used for the base and subbase layers: MR =

σD = Kθ n εR

(1)

Where MR is resilient modulus (psi), θ is bulk stress (psi), and K and n are statistical parameters. The following relationship exists between K and n (R2 = 0.68, SEE = 0.22) (Rada and Witczak 1981): Log10 (K) =4.657 – 1.807n

(2)

The stress-softening bilinear model was used for the subgrade layer: M R = M Ri + K 1 .(σ d − σ di ) for σ d < σ di M R = M Ri + K 2 .(σ d − σ di ) for σ d > σ di

(3)

Where MR is resilient modulus (psi), σd is applied deviator stress (psi), and K1 and K2 are statistically determined coefficients from laboratory tests. A total of 5,000 input cases were generated for each test section by randomly varying the asphalt concrete and subgrade layer moduli and the ‘Kb’-‘nb’ and ‘Ks’-‘ns’ values (note that K and n are related) for the base and subbase layers, respectively. The effect of 36-kip HWD loading was simulated in ILLI-PAVE and the pavement surface deflections were computed. Initially, it was decided to use separate ANN models for each section. Of the total number of data sets for each test section, 3,750 data vectors were used in training the ANN and the remaining 1,250 data vectors were used to test the network after the training was completed. The range of layer properties used in training the ANN is summarized in Table 1. ARTIFICIAL NEURAL NETWORK ARCHITECTURE A generalized n-layer feedforward ANN that uses an error-backpropogation algorithm (Haykin 1994) was implemented in the Visual Basic (VB 6.0) programming language. The program can allow for a general number of inputs, hidden layers, hidden layer elements, and output layer elements. Two hidden layers were found to be sufficient for solving a problem of this size, and therefore the architecture was reduced to a four-layer feedforward network. A four-layer feedforward network consists of a set of sensory units (source nodes) that constitute the input layer, two hidden layers of computation nodes, and an output layer of computation nodes. The following notation is generally used to refer to a particular type of architecture that has two hidden layers: (# inputs)-(# hidden neurons)-(# hidden neurons)-(# outputs). For example, the notation 10-40-40-3 refers to an ANN architecture that takes in 10 inputs (features), has 2 hidden layers consisting of 40 neurons each, and produces 3 outputs.

Gopalakrishnan

5

Table 1. Range of pavement layer properties used in generating the ANN training database Pavement layer

Thickness (inches)

Elastic layer modulus (ksi) Poisson’s ratio

Asphalt concrete

5 - MFC &LFC 10 - MFS & LFS 8 - MFC & LFC 8.5 - MFS 29.5 - LFS 12 - MFC 36.4 - LFC 95 - MFC & LFC 105 - MFS & LFS 120 - Medium 144 - Low

100 – 2,500

0.35

Kb: 1.6 – 20 nb: 0.2 – 0.8

0.35

Ks: 1.6 – 20 ns: 0.2 – 0.8 1.6 – 20

0.35 0.45

45

0.4

Base

Subbase Subgrade Sand

An ANN-based backcalculation procedure was developed to approximate the HWD backcalculation function. Using the ILLI-PAVE synthetic database, the ANN was trained to learn the relationship between the synthetic deflection basins (inputs) and the pavement layer moduli (outputs). Initialization of Weights The first step in back-propogation learning is to initialize the network. It is recommended that the initialization of the synaptic weights of the network be uniformly distributed inside a small range. A range of -0.2 to +0.2 was used for random initialization of all synaptic weight vectors in the network. Nonlinear Activation Function The model of each neuron in the hidden layer(s) and output layer of the network includes a nonlinearity at the output end. The presence of a nonlinear activation function, ϕ(.), is important because otherwise the input-output relation of the network could be reduced to that of a single-layer perceptron. The computation of the local gradient for each neuron of the multilayer perceptron requires that the function ϕ(.) be continuous. In other words, differentiability is the only requirement that an activation function would need to satisfy. For this problem, an asymmetric hyperbolic tangent function (tanh) was chosen for which the output amplitude lies inside the range -1 ≤ yj ≤ +1. Since we require the final outputs to be real values instead of binary outputs, a linear combiner model was used for neurons in the output layer, thus omitting the nonlinear activation function. Performance Measure (RMSE) In order to track the performance of the network, the root mean squared error (RMSE) at the end of each epoch was calculated. An epoch is defined as one full presentation of all the training vectors to the network. The RMSE at the end of each epoch is defined as the following:

∑ [d N

RMSE =

Gopalakrishnan

j =1

j

− Y (X

)]

2

j

N

6

Where dj is the desired response for the input training vector Xj, and N is the total number of input vectors presented to the network for training. For the network to learn the problem smoothly, a monotonic decrease in the RMSE is expected with an increase in the number of epochs. A smooth learning curve was achieved with a learning-rate parameter (η) of 0.001. ANN INPUTS AND OUTPUTS Deflection basin parameters (DBPs) derived from falling weight deflectometer (FWD) and/or HWD deflection measurements are shown to be good indicators of selected pavement properties and conditions (Hossain and Zanniewski 1991). Recently, Xu et al. (2001) used DBPs in developing new relationships between selected pavement layer condition indicators and FWD deflections by applying regression and ANN techniques. Apart from the six independent deflection measurements (D0 to D60), some of the commonly used DBPs were included as inputs for training the ANN. The DBPs considered in this study are shown in Table 2. Each DBP supposedly represents the condition of specific pavement layers. For example, AUPP is sensitive to the asphalt concrete layer properties, whereas BCI and AI4 are expected to reflect the condition of subgrade. The desired outputs from the ANN are asphalt concrete modulus (EAC), subgrade modulus (ERi), base modulus parameter (Kb or nb), and subbase modulus parameter (Ks or ns). Note that by predicting either K or n, the other parameter can be determined using the relation proposed by Rada and Witzcak (1981). Table 2. DBPs considered in this study Deflection basin parameter Formula (DBP) AREA Area under pavement profile (AUPP) Area index Base curvature index (BCI) Base damage index (BDI) Deflection ratio Shape factors

AREA = 6(D0 + 2D12 + 2D24 + D36)/D0 AUPP = (5D0 – 2D12 – 2D24 – D36)/2 AI4 = (D36 + D48)/2D0 BCI = D24 – D36 BCI2 = D60 – D48 BDI = D12 – D24 DR = D12/D0 F1 = (D0 – D24)/D12 F2 = (D12 – D36)/D24

SELECTION OF BEST-PERFORMANCE NETWORKS Separate ANN models were used for each desired output rather than using the same architecture to determine all the outputs together. The most effective set of input features for each ANN model were determined based on both engineering judgment and the experience gained through past research studies conducted at UIUC. Parametric analyses were performed by systematically varying the choice and number of inputs and number of hidden neurons to identify the best-performance networks. As it was found that the prediction accuracy of the network remained the same for hidden layers greater than or equal to two, the number of hidden layers was fixed at two for all runs. The learning curve (RMSE vs number of epochs) and the testing RMSE were studied in order to arrive at the best networks. A previous

Gopalakrishnan

7

study that focused on the MFC section alone showed that the base and subbase moduli parameters were the hardest to predict (Gopalakrishnan and Thompson 2004). During the course of this study, the same conclusion was reached for other test sections. It was concluded that further research is needed to develop robust ANN models for predicting the base and subbase moduli parameters. RESEARCH RESULTS A summary of the sensitivity analyses performed to select the best-performance networks for predicting asphalt concrete modulus (EAC) and subgrade modulus (ERi) in NAPTF test sections are shown in Table 3. Note that the ANN inputs are similar for all four test sections. In Figure 3, the ANN-predicted moduli values and the target values are compared using the 1,250 test data vectors for each NAPTF section. Excellent agreement is found between the predicted and target values for both EAC and subgrade modulus ERi in all four test sections, except for ERi in LFS section, where an R2 value of 0.81 was obtained. Table 3. Summary of best-performance ANN pavement moduli prediction models NAPTF Training Testing Network Output Inputs section architecture RMSE RMSE MFC LFC MFS LFS

EAC

D0 ~ D60

6-40-40-1

71 ksi

69 ksi

ERi

D0 ~ D60, BCI, AI4

8-40-40-1

0.86 ksi

0.82 ksi

EAC

D0 ~ D60

6-40-40-1

100 ksi

97 ksi

ERi

D0 ~ D60, BCI, AI4

8-40-40-1

1.29 ksi

1.18 ksi

EAC

D0 ~ D60

6-40-40-1

69 ksi

67 ksi

ERi

D0 ~ D60, BCI, AI4

8-40-40-1

0.81 ksi

0.78 ksi

EAC

D0 ~ D60

6-40-40-1

90 ksi

90 ksi

ERi

D0 ~ D60, BCI, AI4

8-40-40-1

2.36 ksi

2.15 ksi

One of the major reasons for developing this ANN-based backcalculation procedure is to evaluate the structural integrity of the NAPTF pavement test sections reliably as they were subjected to traffic loading. The NAPTF test sections were subjected to trafficking until they exhibited failure. The MFC test section was the first one to fail at 12,952 load repetitions exhibiting 3 to 3.5 inches of rutting and severe cracking. In the MFS section, localized failure occurred in the Boeing 777 traffic lane toward the west end. At 19,900 passes, 3.5 inches of rut depth was observed on the Boeing 777 traffic lane, with upheaval outside the traffic path. Trafficking was terminated on the Boeing 777 traffic lane, but it continued on the Boeing 747 traffic lane. The Boeing 747 lane failed at 30,000 passes. HWD tests were not conducted on the MFS section beyond 19,900 passes. The low-strength test sections (LFC and LFS) showed few signs of genuine distress, even after 20,000 passes, and therefore the wheel loading was increased from 45,000 lbs to 65,000 lbs. The trafficking was terminated in the low-strength test sections after 28,000 passes of 65,000 lbs. While the MFC and MFS sections failed at the subgrade level, the LFC and LFS sections failed in the surface layers, signifying tire pressure or other upper layer failure effects, but not subgrade level failure (Gervais, Hayhoe, and Garg 2003).

Gopalakrishnan

8

ANN Predicted Subgrade Modulus, ERi (ksi)

ANN Predicted AC Modulus, EAC (ksi)

2500 MFC 2 R = 0.98

2000 1500 1000 500 0 0

500

1000

1500

2000

25 MFC 2 R = 0.97

20 15 10 5 0

2500

0


3000 LFC 2 R = 0.98

2500 2000 1500 1000 500 0 0

2000 1500 1000 500 0 0

5 0

2000 1500 1000 500 0 0

500 1000 1500 2000 2500 3000 Target AC Modulus, EAC (Ksi)

10

15

20

25

25 MFS 2 R = 0.98

20 15 10 5 0 0

5

10

15

20

25

Target Subgrade Modulus, ERi (ksi) ANN Predicted Subgrade Modulus, ERi (ksi)


2500

5

Target Subgrade Modulus, ERi (ksi)

Target AC Modulus, EAC (Ksi)

LFS 2 R = 0.98

25

10

500 1000 1500 2000 2500 3000

3000

20

15

0



2500

15

LFC R2 = 0.95

20


MFS 2 R = 0.99

10

25

500 1000 1500 2000 2500 3000

3000

5




25

LFS R2 = 0.81

20 15 10 5 0 0

5

10

15

20

25


Figure 3. ANN prediction of NAPTF asphalt concrete moduli (left) and subgrade moduli (right)

Gopalakrishnan

9

ANN vs FAABACKCAL Using the 36-kip HWD test data acquired at the NAPTF, the asphalt concrete moduli and subgrade moduli were backcalculated with the best-performance ANNs for all four sections. The results were then compared with those obtained using FAABACKCAL, an ELP-based backcalculation program. FAABACKCAL was developed under the sponsorship of the FAA Airport Technology Branch and is based on the LEAF layered elastic computation program. In this program, the pavement layer moduli and subgrade moduli are adjusted to minimize the root mean square (rms) of the differences between FWD/HWD sensor measurements and the LEAF-computed deflection basin for a specified pavement structure. A standard multidimensional simplex optimization routine is then used to adjust the moduli values (McQueen, Marsey, and Arze 2001). A stiff layer with a modulus of 1,000,000 psi and a Poisson’s ratio of 0.50 was used in backcalculation. Based on the as-constructed conditions, the stiff layer was set at 10 feet for the medium-strength test sections and at 12 feet for the low-strength sections. The most recent version of the FAA backcalculation software is called BAKFAA. The detailed backcalculation results for NAPTF sections using FAABACKCAL are reported elsewhere (Gopalakrishnan and Thompson 2004). The plots comparing the results of ANN-based backcalculation with those obtained using FAABACKCAL are shown in Figure 4 for asphalt concrete moduli and in Figure 5 for subgrade moduli. The gray arrow in the plots for the LFC and LFS sections indicate where (after 20,000 load repetitions) the wheel load was increased from 45,000 lbs to 65,000 lbs. In Figure 4, the variation in asphalt concrete temperature as a function of the number of load repetitions (N) in each test section is also included and is indicated by a gray line. In these plots, the changes in layer moduli in the Boeing 777 traffic lane and the Boeing 747 traffic lane are due to both traffic loading as well as variation in temperature and climate. The changes in pavement material properties in the untrafficked pavement centerline are only due to environmental effects. The trends for asphalt concrete moduli are very similar using both approaches. The asphalt concrete moduli values are significantly influenced by asphalt concrete temperature. In the MFC and LFC sections, the ANN approach seem to be more sensitive to trafficking and temperature, indicated by a sharp decrease in the moduli values towards the end of trafficking. In the LFC and LFS sections, the structural degradation resulting from increasing the wheel load from 45,000 lbs to 65,000 lbs after 20,000 repetitions is clearly reflected in the moduli values. Compared to the centerline moduli values, the traffic lane moduli values decrease further after 20,000 repetitions. In Figure 5, the Y-axis is magnified in the individual plots as the subgrade moduli values varied over a narrow range compared to the asphalt concrete moduli. Note that the subgrade modulus predicted by ANN is the stress-dependent breakpoint subgrade modulus (at a deviator stress of about 6 psi) and is indicated as ERi. The subgrade modulus backcalculated by FAABACKCAL is indicated as E. The initial magnitudes of subgrade modulus obtained using both the approaches are similar for the MFS and LFS sections. For the MFC and LFC sections, the initial subgrade moduli obtained using the ANN are about 5 psi higher than the those obtained using FAABACKCAL. The sharp decrease in the subgrade modulus in the Boeing 777 traffic lane of the MFS section towards the end of the trafficking, as captured by the ANN results, could be due to the localized failure, as mentioned earlier. The NAPTF rutting study results showed that the eest end of the Boeing 777 traffic lane in the MFS section exhibited a rapid increase in surface rutting after 12,000 load repetitions (Gopalakrishnan and Thompson 2003).

Gopalakrishnan

10

55

1500

60 65

ANN MFC

500

70

0 10

100

1000 N

10000

AC Modulus, EAC (ksi)

50

2000

55

1500

60

1000

65

AC Temperature (0F)

2500

70

100

1000 N

10000

0

75 100000

40

3500 3000

45

2500

50

2000

55

1500

60

1000

65

500

70

LFC

0

3000

40

3000

40

2500

45

2500

45

N

1000

50

2000

55

1500

60

1000

65

500

70

MFS

0 1

10

100

1000

1

75 10000 100000

10

100

55 1500

60

1000

65

500

70

MFS

0 1

10

100

2500

50

2000

55

1500

60

1000

65 70

LFS

0 100

1000

N

75 10000 100000

75 10000 100000

3500

40

3000

45

2500

50

2000

55

1500

60

1000

65

0

45


3000

AC Temperature (0F)

40

10

1000

N

3500

1

75 10000 100000

50

2000

N

500

N

1000

0

100


10

AC Temperature (0F)

1


70

75 10000 100000

0


65

FAABACKCAL MFC

500 10

45

LFC

60

1000

0

3000

500

55

1500

75 100000

40

3500

50

2000

AC Temperature ( F)

1000

45

2500

AC Temperature ( F)

2000

40

AC Temperature (0F)

50

3000


45

2500

B777-FAABACKCAL B747-FAABACKCAL C/L-FAABACKCAL

AC Temperature ( F)

40 ACTemperature (0F)


3000


B777-ANN B747-ANN C/L-ANN

500

70

LFS

0 1

10

100

1000

75 10000 100000

N

Figure 4. Comparison of ANN-predicted asphalt concrete moduli (left) with FAABACKCAL asphalt concrete moduli (right) for NAPTF test sections

Gopalakrishnan

11

10

Subgrade Modulus, ERi (ksi)

Subgrade Modulus, E (ksi)

ANN MFC

20 19 18 17 16 15 14 13 12 11 10

100

1000 N

10000

MFS 1

10

100

1000

10000

B777-FAABACKCAL B747-FAABACKCAL C/L-FAABACKCAL

20 19 18 17 16 15 14 13 12 11 10

100000

FAABACKCAL MFC

10



B777-ANN B747-ANN C/L-ANN

20 19 18 17 16 15 14 13 12 11 10

20 19 18 17 16 15 14 13 12 11 10

100000

100

1000 N

1

10

100



14 12 10 8

LFC

4 100

N

1000

10000

16 14 12 10 8 6

LFS

4 1

10

100

1000 N

10000

100000

1000

10000

100000

10000

1000

10000

100000

16 14 12 10 8 6

100000

LFC

4

100000

1 Subgrade Modulus, E (ksi)


10

1000 N

16

1

100000

MFS

N

6

10000

10

100

10

100

N

16 14 12 10 8 6

LFS

4 1

N

Figure 5. Comparison of ANN-predicted subgrade moduli (left) with FAABACKCAL subgrade moduli (right) for NAPTF test sections

Gopalakrishnan

12

SUMMARY AND CONCLUSIONS The primary objective of this study was to develop a tool for backcalculating NAPTF non-linear pavement layer moduli from HWD data using ANN. The NAPTF sections were modeled in ILLI-PAVE and a synthetic database was generated for a range of moduli values. A multi-layer feedforward network that uses an error-backpropagation algorithm was successfully trained to approximate the HWD backcalculation function using the ILLI-PAVE database. ANN models were successfully developed for predicting asphalt concrete and non-linear subgrade moduli. However, the base/subbase moduli could not be predicted using ANN. The best-performance ANN models were used to predict asphalt concrete and subgrade moduli from NAPTF HWD deflection basins collected at regular intervals during trafficking. The results were compared with results obtained using FAABACKCAL, a conventional ELP-based backcalculation software. The results definitely indicate that the non-linear pavement layer moduli could be successfully backcalculated from FWD/HWD deflection basins using an ANN trained with ILLI-PAVE results. The backcalculated results could be used to establish realistic a priori moduli inputs to mechanistic pavement structural analysis. Although regression models could also be developed for predicting layer moduli using ILLI-PAVE results, a significant advantage of using the ANN-based approach over the regression approach is that the functional forms of the relationships are not needed a priori. Also, the robustness of the ANN can be improved by including the field data sets in the training process, as they implicitly incorporate noise and errors typically seen in field measurements. Further research is needed to develop robust ANN models for predicting base/subbase moduli parameters. It is proposed that by including the ANN-predicted EAC and ERi values as inputs to the ANN and by training the ANN with the ILLI-PAVE results for the 12-kip and 24-kip HWD loads, the chances of accurately predicting base/subbase moduli will increase. It is also noted that the ANN moduli backcalculation concept developed in this study could be used to backcalculate highway pavement non-linear moduli successfully.

Gopalakrishnan

13

ACKNOWLEDGMENTS This paper was prepared from a study conducted in the Center of Excellence for Airport Technology. Funding for the Center is provided in part by the Federal Aviation Administration under research grant 95-C-001. The Center is maintained at the University of Illinois at Urbana-Champaign, which works in partnership with Northwestern University and the Federal Aviation Administration. The author gratefully acknowledges the financial assistance and support by Professor Marshall R. Thompson for conducting this research. The author is also grateful to Dr. Franco Gomez-Ramirez for his valuable help and suggestions in developing the training database. REFERENCES Ahlvin, R. G., H. H. Ulery, R. L. Hutchinson, and J. L. Rice. 1971. Multiple-Wheel Heavy Gear Load Pavement Tests, Vol. 1: Basic Report. Technical Report No. AFWL-TR-70-113. Vicksburg, MS: U.S. Army Engineer Waterways Experiment Station. Asphalt Institute. 1982. Research and Development of The Asphalt Institute’s Thickness Design Manual (MS-1) Ninth Edition. Research Report No. 82-2. College Park, MD: The Asphalt Institute. Ceylan, H., Tutumluer, E., Thompson, M. R., and F. Gomez-Ramirez. 2004. Neural Network-Based Structural Models for Rapid Analysis of Flexible Pavements with Unbound Aggregate Layers. Proceedings of the Sixth International Symposium on Pavements Unbound, Nottingham, England, UK. Garg, N. and W. H. Marsey. 2002. Comparison Between Falling Weight Deflectometer and Static Deflection Measurements on Flexible Pavement at the National Airport Pavement Facility (NAPTF). Paper presented at the 2002 Federal Aviation Administration Airport Technology Conference, Chicago, IL. Gervais, E. L., G. F. Hayhoe, and N. Garg. 2003. Towards a Permanent Solution for 6-Wheel Landing Gear Aircraft. Proceedings of the 2003 ASCE Airfield Specialty Conference, Las Vegas, NV. Gopalakrishnan, K. and M. R. Thompson. 2003. Rutting Study of NAPTF Flexible Pavement Test Sections. In Proceedings of the 2003 ASCE Airfield Specialty Conference, Las Vegas, NV. Gopalakrishnan, K. and M. R. Thompson. 2004. Backcalculation of Airport Flexible Pavement NonLinear Moduli Using Artificial Neural Networks. Proceedings of the 17th International Florida Artificial Intelligence Research Symposium Conference, FLAIRS-2004, Miami Beach, FL. Gopalakrishnan, K. and M. R. Thompson. 2004. Comparative Effect of B777 and B747 Trafficking on Elastic Layer Moduli of NAPTF Flexible Pavements. In Proceedings of the 2004 FAA Worldwide Airport Technology Transfer Conference, Atlantic City, NJ. Gomez-Ramirez, F. M., and M. R. Thompson. 2002. Characterizing Aircraft Multiple Wheel Load Interaction for Airport Flexible Pavement Design. Civil Engineering Studies, COE Report. University of Illinois at Urbana-Champaign. Haykin, S. 1994. Neural Networks: A Comprehensive Foundation. New York: Macmillan College Publishing Company, Inc. Hossain, S. M. and J. P. Zaniewski. 1991. Characterization of Falling Weight Deflectometer Deflection Basin. Transportation Research Record 1293. Washington, D.C.: TRB, National Research Council, pp. 1–11. Hicks, R. G. 1970. Factors Influencing the Resilient Properties of Granular Materials. Ph.D. dissertation, University of California, Berkeley. McQueen, R. D., W. Marsey, and J. M. Arze. 2001. Analysis of Nondestructive Data on Flexible Pavement Acquired at the National Airport Pavement Test Facility. In Proceedings of the 2001 ASCE Airfield Pavement Specialty Conference, Chicago, IL.

Gopalakrishnan

14

NCHRP 1-26. 1990. Calibrated Mechanistic Structural Analysis Procedures for Pavements. Final Report. National Cooperative Highway Research Program Project 1-26. Washington, D.C.: TRB, National Research Council. Raad, L. and J. L. Figueroa. 1980. Load Response of Transportation Support Systems. Transportation Engineering Journal 16 (TE1). Rada, G. and M. W. Witczak. 1981. Comprehensive Evaluation of Laboratory Resilient Moduli Results for Granular Material. Transportation Research Record: Journal of the Transportation Research Board, No.810. Washington, D.C.: TRB, National Research Council. Thompson, M. R. and Q. L. Robnett. 1979. Resilient Properties of Subgrade Soils. Transportation Engineering Journal 105.1. Thompson, M. R., E. Tutumluer, and M. Bejarano. 1998. Granular Material and Soil Moduli Review of the Literature. Final Report. Center of Excellence for Airport Pavement Research COE Report No. 1. University of Illinois at Urbana-Champaign. U.S. COE. 2001. O & M: PAVER, Asphalt Surfaced Airfields Pavement Condition Index (PCI). Unified Facilities Criteria 3-270-06. U.S. Army Corps of Engineers (preparing activity), Naval Facilities Engineering Command, Air Force Civil Engineering Support Agency. Xu, B., S. R. Ranjithan and Y. R. Kim. 2001. Development of Relationships between FWD Deflections and Asphalt Pavement Layer Condition Indicators. Paper presented at the 81st Annual Meeting of the Transportation Research Board, Washington, D.C.

Gopalakrishnan

15