Draft
DEVELOPMENT OF AN ARTIFICAL NEURAL NETWORK MODEL
TO PREDICT SUBGRADE RESILIENT MODULUS FROM
CONTINUOUS DEFLECTION TESTING
Journal: Canadian Journal of Civil Engineering
Manuscript ID cjce-2017-0132.R1
Manuscript Type: Article
Date Submitted by the Author: 27-Apr-2017
Complete List of Authors: Elbagalati, Omar; Louisiana State University, Civil and Environmental
Engineering Elseifi, Mostafa; LSU, CEE Gaspard, Kevin; Louisiana Transportation Res Zhang, Zhongjie; Louisiana Transportation Research Center
Is the invited manuscript for consideration in a Special
Issue? : N/A
Keyword: subgrade resilient modulus, pavement deflection, rolling wheel deflectometer, network level
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
DEVELOPMENT OF AN ARTIFICAL NEURAL NETWORK 1
MODEL TO PREDICT SUBGRADE RESILIENT MODULUS FROM 2
CONTINUOUS DEFLECTION TESTING 3
4
5
Omar Elbagalati
Graduate Research Assistant
Department of Civil and Environmental Engineering
Louisiana State University
3316s Patrick Taylor Hall, Baton Rouge, LA 70803
e-mail: [email protected]
Mostafa A. Elseifi (Corresponding Author)
Associate Professor
Department of Civil and Environmental Engineering
Louisiana State University
3316s Patrick Taylor Hall, Baton Rouge, LA 70803
e-mail: [email protected]
Kevin Gaspard
Senior Pavement Research Engineer
Louisiana Transportation Research Center
Louisiana State University
4101 Gourrier Ave., Baton Rouge, LA 70808
e-mail: [email protected]
Zhongjie Zhang
Pavement Geotechnical Research Administrator
Louisiana Transportation Research Center
Louisiana State University
4101 Gourrier Ave., Baton Rouge, LA 70808
e-mail: [email protected]
6
7
8
9
10
11
12
13
Page 1 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Elbagalati, et al. 2
ABSTRACT: The subgrade resilient modulus is an important parameter in pavement 14
analysis and design. However, available Non-Destructive Testing devices (NDT) such as 15
the Falling Weight Deflectometer (FWD) have limitations that prevent their widespread 16
use at the network-level. This study describes the development of a model that utilizes 17
the Rolling Wheel Deflectometer (RWD) measurements to predict the subgrade resilient 18
modulus at the network level for flexible pavements. RWD and FWD measurements 19
obtained from a testing program conducted in Louisiana were used to train an Artificial 20
Neural Network (ANN) based model. The ANN model was validated using data from a 21
testing program independently conducted in Minnesota. The ANN model showed 22
acceptable accuracy in both the development and validation phases with coefficients of 23
determination of 0.73 and 0.72, respectively. Furthermore, the limits of agreement 24
methodology showed that 95% of the differences between the subgrade resilient modulus 25
calculated based on FWD and RWD measurements will not exceed the range of ±21 MPa 26
(±3 ksi). 27
28
Keywords; Subgrade Resilient Modulus, Pavement Deflection, Rolling Wheel 29
Deflectometer, network level 30
31
32
33
34
35
36
Page 2 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Elbagalati, et al. 3
INTRODUCTION 37
The subgrade resilient modulus (Mr) is an essential design parameter for stress-strain 38
analysis of pavement structures (Ping et al. 2012). In addition, the subgrade resilient 39
modulus is a key input in the AASHTO 1993 pavement design methodology and in all 40
three hieratical levels of design in the new AASHTOWare Pavement-ME design 41
procedure (Rahim 2005; Hossain et al. 2011). Recent studies have indicated high 42
sensitivity of Pavement-ME predicted distresses to the subgrade resilient modulus 43
(Schwartz et al. 2013; Abd El-Hakim et al. 2016). Furthermore, the subgrade resilient 44
modulus was found to have a significant effect on the design thickness of asphalt 45
overlays (Wu and Gaspard 2009; Wu et al. 2013). 46
For new construction, it is possible to measure the subgrade resilient modulus in 47
the laboratory after collecting sufficient soil materials from the field (Ahmed et al. 2016). 48
However, collecting a sufficient amount of soil materials from in-service pavement 49
sections by extracting cores is a tedious procedure that disturbs traffic, is costly, and can 50
have a significant impact on the integrity of the pavement structure (Tarefder et al. 2015). 51
To overcome these difficulties, many Non-Destructive Testing (NDT) and semi-52
destructive devices have been utilized to assess subgrade material properties in-situ such 53
as the Falling Weight Deflectometer (FWD) and the Dynamic Cone Penetrometer (DCP). 54
Yet, the DCP test requires drilling holes in the pavement section (Mohammad et al. 55
2009). On the other hand, the stationary nature of the FWD has limited the device 56
production rates and reduced its applicability for network-level surveys (Rada et al. 57
2011). To address these limitations, a number of continuous deflection measuring NDT 58
Page 3 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Elbagalati, et al. 4
devices were developed in recent years such as the Traffic Speed Deflectometer (TSD) 59
and the Rolling Wheel Deflectometer (RWD) (Flintsch et al. 2013). 60
OBJECTIVE 61
The objective of this study was to develop and validate a model to estimate subgrade 62
resilient modulus using RWD deflection measurements at the network-level for flexible 63
pavements. For model development, RWD and FWD measurements obtained from a 64
testing program conducted in Louisiana were used to train an Artificial Neural Network 65
(ANN) based model. After the learning process, the ANN model was validated using 66
RWD and FWD data from a testing program conducted independently at the MnROAD 67
facility in Minnesota. 68
BACKGROUND 69
The Rolling Wheel Deflectometer 70
The Rolling Wheel Deflectometer (RWD) is a pioneer test device for cost-effective 71
measurements of pavement deflection and surface properties at traffic speed, see Fig. 1. 72
On an Interstate, RWD can survey 400 lane-km (250 miles) per day compared to 64 lane-73
km (40 miles) per day for FWD. The most recent version of the RWD, which was 74
introduced in 2003, consists of a 16 m. (53-ft.) long semitrailer applying a standard 80 kN 75
(18,000-lbs.) load on the pavement structure by means of a regular dual-tire assembly 76
over the rear single axle (Briggs et al. 2000). The device operation speed can be as high 77
as 80 km/h (50 mph) causing no delays to the road users. Yet, the trailer is sufficiently 78
long to isolate the deflection basin produced by the RWD’s rear single axle from those 79
produced by the RWD tractor. 80
Page 4 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Elbagalati, et al. 5
FIGURE 1 81
82
Several modifications and upgrades were applied to the RWD since its first 83
introduction with respect to the laser sensors, data acquisition system, and software. The 84
laser collection system was moved between the tires, and a new procedure was 85
introduced for laser calibration. The laser sensors are set to collect a reading at a fixed 86
interval of 15.24 mm (0.6 in.) at all truck speeds. In 2009, a more accurate and stable 87
deflection measurement system customized for pavement applications was installed. The 88
upgraded system has a (101.1 mm) 4-in. measurement deflection range and has an 89
accuracy of ± 0.025 (0.001 in.). In the new system, four Selcom Model SLS 6000 laser 90
triangulation sensors are mounted at approximately 1.1 m. (3.6 ft.) above the roadway 91
surface with a 101.1 mm. measurement range. The laser sensors work simultaneously to 92
determine pavement deflections under the wheel load, with one sensor placed between 93
the dual tires to determine the maximum deflection. Two additional sensors are placed in 94
front of the wheels to measure a secondary pavement deflection at 457.2 mm (18 in.) 95
from the load. Prior to this research, no study has attempted to use the measurements 96
from the second sensor located 457.2 mm from the load in the analysis. 97
Typically, the RWD averages individual deflection readings over 160-meter (0.1 mile) 98
intervals, and reports the average deflection value along with its standard deviation. An 99
environmental chamber is utilized to maintain the measurement system at a constant 100
temperature. Further, the system includes a distance-measuring instrument (DMI) to 101
longitudinally reference collected data, an infrared thermometer to measure pavement 102
surface temperature, and a global positioning system (GPS) (ARA Inc. website). A 103
Page 5 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Elbagalati, et al. 6
Recent Strategic Highway Research Program 2 (SHRP2) study selected the RWD and 104
TSD as the most promising moving deflection measurement devices (Flintsch et al. 105
2013). 106
Calculating the Subgrade Resilient Modulus based on FWD data 107
In the 1993 AASHTO Guide, a detailed procedure is described to calculate the subgrade 108
resilient modulus based on FWD measurements as follows: 109
110 M� = C ∗ 0.24 ∗ Pd� ∗ r (1) 111
where Mr = subgrade resilient modulus (psi); C= Correction factor; P = FWD load (lb.); 112
dr = Deflection at distance r (in.); and r = Distance from the center of the FWD loading 113
plate (in.). 114
115
The relationship described in Equation (1) is valid at a distance r outside the pressure 116
bulb of the FWD load. Darter and co-authors recommended that the deflection used for 117
subgrade resilient modulus determination should be measured at a distance at least 0.7 118
times the radius of the stress bulb (ae), which can be calculated based on Equation 2 119
(Darter et al. 1991). For relatively thin pavements, the stress bulb was found to be 120
approximately 381 mm (15 in.), and for medium to thick pavements, the stress bulb 121
ranged from 660 to 840 mm (26 to 33 in.) (Darter et al. 1991): 122
123
a� =�a� + D� ∗ (E� M�� )�/�(2) 124
125
Page 6 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Elbagalati, et al. 7
where, a= FWD plate radius (in.); D= Total pavement thickness (in.); and Ep= effective 126
modulus of all pavement layers above the subgrade (psi), which can be calculated from 127
the following equation: 128
d� = 1.5pa� ! "1
M��1 + #$% &'()*+ ,�- + .1 − 0
&01234567 /E�8 9 : (3) 129
130
where, d0 = Deflection measured under the center of the loading plate (in). 131
Differences between FWD and RWD 132
Table 1 compares the general characteristics of FWD and RWD. Recent studies found 133
that the difference in deflection magnitude can be significant between a continuous 134
deflection measuring device such as RWD and FWD deflection measurements; yet, the 135
general trends were relatively the same when comparing pavements that were 136
structurally-sound or structurally-deficient (Katicha et al. 2013). Furthermore, both test 137
methods appear to properly reflect pavement conditions and structural integrity of the 138
road network by providing a greater average deflection and scattering for sites in poor 139
conditions. The effect of surface irregularities on the measurements of a moving device 140
(such as RWD), the difference in the load contact area between FWD and the RWD, and 141
the difference in loading mechanisms (rubber plate for the FWD and dual tire assembly 142
for the RWD), are all factors that may lead to deflection differences (Rada and Nazarian 143
2011). Therefore, the deflection basin characteristics for the RWD are not expected to 144
match with the ones from FWD. Accordingly, the approach presented in Equations (1) to 145
(3) cannot be directly applied to RWD measurements. To address this limitation, this 146
study made use of Artificial Neural Networks (ANNs). 147
Page 7 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Elbagalati, et al. 8
TABLE 1 148
149
Artificial Neural Networks 150
Artificial Neural Networks have commonly been used for solving complex engineering 151
problems in the last three decades (Ceylan et al. 2014). ANNs are parallel computing 152
schemes that imitate biological neural networks (Yet et al. 2014). They are effective and 153
accurate tools for solving complex nonlinear problems as they provide robust models that 154
can continuously be updated as new data become available. In addition, they can be used 155
in databases with either large or relatively small amount of data (Plati et al. 2015). Many 156
researchers have successfully used ANN-based models to backcalculate the layer 157
properties of in-service pavements, and they reported that ANNs are an effective tool for 158
backcalculation analysis (Kim et al. 2000; Gopalakrishnan et al. 2006; Bayrak et al. 2006; 159
Ceylan et al. 2005; genetic et al. 2006; Saltan et al. 2012). Therefore, ANN was selected 160
in the present study to backcalculate the subgrade resilient modulus from RWD 161
measurements. 162
DATA DESCRIPTION 163
Louisiana Testing Program 164
A two-phase comprehensive testing program was conducted in District 05 of Louisiana 165
(Abdel-Khalek et al. 2012). In the first phase, the complete asphalt road network (about 166
3,200 km [2,010 mi.]) was tested using the RWD deflection system based on the 167
manufacturer standard testing protocol. Researchers also selected 58 sections to be tested 168
using the FWD. In addition, 16 road-sections (2.4 km each [1.5 mi.]), referred to as 169
research sites, were selected and used for a detailed evaluation of RWD technology in the 170
Page 8 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Elbagalati, et al. 9
second phase (Elseifi et al. 2012). In addition to RWD testing, the field-testing plan in 171
Phase II conducted FWD testing on selected flexible and surface-treatment pavement test 172
sites. The testing plan specified that FWD testing should be conducted within 24 hours 173
following completion of RWD testing and at the same time of the day on the selected 174
sites, in order to maintain the same testing conditions. 175
Minnesota Testing Program 176
While the developed model was calibrated for the conditions pertinent to the Louisiana 177
road network, it was uncertain whether this model could be extended to different states 178
and varying road conditions and designs. To address this concern, FWD and RWD data 179
were obtained from a comprehensive testing program conducted at the MnROAD facility 180
in 2013. The testing program aimed at evaluating the accuracy of two continuous 181
deflection measuring devices (TSD and RWD) and to assess their use at the network level 182
(Sivaneswaran 2014). It is noted that the RWD prototype used in MnROAD was 183
measuring the secondary deflection (D1) at 381 mm (15 in.) instead of 457.2 mm (18 in.) 184
as it was the case in Louisiana. To assess the accuracy of these devices, 20 sensors were 185
installed in the MnROAD facility (strain gauges, pressure cells, geophones, 186
accelerometer, etc.). FWD was used to verify the performance of each sensor and to 187
evaluate the correlation between the continuous deflectometers and the FWD. Research 188
findings highlighted the benefits of using the RWD as well as the TSD for pavement 189
structural evaluation purposes at the network level. The flexible pavement test segments 190
at which both FWD and RWD measurements were conducted consisted of 16 road 191
sections that were used in the validation phase. 192
Page 9 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Elbagalati, et al. 10
METHODOLOGY 193
To develop the proposed ANN model, the FWD sensor D7 (1500 mm from the center of 194
the plate) measurements were used in Equation (1) to calculate the subgrade resilient 195
modulus for the tested pavement sections, Step 1. No temperature correction was applied 196
for the D7 measurements as they are only affected by the subgrade properties (Lukanen et 197
al. 2000; Kim et al. 2002). Second, statistical correlations were investigated between 198
RWD measurements and the subgrade resilient modulus calculated from Step 1. The 199
RWD measurements were corrected to a reference temperature of 20oC using BELLS 200
equation and the AASHTO 1993 procedure (Kim et al. 2002; Elseifi et al. 2012). Finally, 201
the RWD measurements and the subgrade resilient modulus values calculated from FWD 202
measurements were used to develop and validate the ANN model. 203
Correlations between RWD Measurements and the Subgrade Resilient Modulus 204
As described earlier, the RWD reports average deflections for 160 m intervals along with 205
standard deviations. Thus, four readings can be obtained from the device; the average 206
deflection at the rear axle (D0) and its standard deviation (σD0), and the average deflection 207
at 457.2 mm (D1) and its standard deviation (σD1). The statistical correlations between 208
these four parameters and the subgrade resilient modulus were investigated for the 209
measurements obtained from the Louisiana testing program. 210
An analysis of variance (ANOVA) was conducted between the subgrade resilient 211
modulus and the four RWD measurements using the SAS 9.4 software. Table 2 212
summarizes the results of the statistical analyses. As shown in this table, Parameters D0, 213
σD0, and D1 were found to be significantly correlated to the subgrade resilient modulus. 214
On the other hand, σD1 was not statistically correlated to the subgrade resilient modulus. 215
Page 10 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Elbagalati, et al. 11
The coefficient of determination (R2) between each parameter and the subgrade resilient 216
modulus was also calculated. Fig. 2. presents the correlations between the D1 and D0 217
with the Mr. As shown in Fig. 2, there is a downward trend between the decrease in 218
subgrade resilient modulus and the measured RWD deflections; yet, considering only one 219
deflection measurement is not sufficient to accurately predict the subgrade resilient 220
modulus as evident from the low R2 shown in this figure. 221
Based on these findings, the three RWD measurements (D0, σD0, and D1) were 222
considered in the ANN model for prediction of the subgrade resilient modulus. As 223
previously noted, the RWD prototype used in MnROAD measured the secondary 224
deflection (D1) at 381 mm (15 in.) instead of 457.2 mm (18 in.). To develop a model that 225
is compatible with measurements of both prototypes, “D1/r” was used in the model 226
instead of D1; where, r is the radial distance from the RWD rear axle. A multi-linear 227
regression model was developed using SAS 9.4 and showed an R2 of 0.6 and an RMSE of 228
15%; therefore, ANN was utilized in the present study to develop a model with better 229
accuracy. Comparing linear regression with ANN was not in the scope of this study as 230
recent studies concluded that comparison between linear regression models and ANN-231
based models is not adequate (Karlaftis and Vlahogianni 2011). 232
233
Page 11 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Elbagalati, et al. 12
TABLE 2 234
FIGURE 2 235
ANN Model Development 236
A multilayered feed-forward ANN using a back-propagation error algorithm was 237
developed with a tan-sigmoid transfer function and a linear activation function. The 238
simplest network topology that produces acceptable prediction accuracy was selected to 239
avoid overfitting of the model (Kim et al. 2014; Lawrence et al. 1997). The network 240
topology consisted of three layers of neurons and two layers of weights; an input layer (i) 241
of 3 neurons; a hidden layer (j) of 2 neurons; a target output layer (k) of 1 neuron, layer 242
of weights between neuron layers i and j (ij), and layer of weights between neuron layers 243
j and k (jk). Weights in layers ij and jk were named “Wij” and “W‵
jk”, respectively. In 244
addition, bias values were added to the sums calculated at each neuron (except layer i). 245
Biases in layers j and k were named “bj” and “Bk”, respectively (Leverington 2012). To 246
train the network, such that the proper weights and biases are calculated, the input layer 247
was fed with the three selected RWD measurements, and the target layer was fed with the 248
subgrade resilient modulus values. The network structure is shown in Fig. 3. 249
FIGURE 3 250
251
Data from the Louisiana testing program were used in the model development phase (124 252
road segments). The data were divided into 70% for training, 15% for validation, and 253
15% for testing, so that more than 25 data points are used for validation and testing 254
purposes as recommended in literature (Beleites et al. 2013). To avoid overfitting and to 255
increase the network generalization ability, training was halted when the validation set 256
error stopped decreasing, as shown in Fig. 4. Since the testing data set had no effect on 257
Page 12 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Elbagalati, et al. 13
the training phase, it was used to provide an independent measure of the network 258
performance. 259
FIGURE 4 260
261
RESULTS AND ANALYSIS 262
The regression plots of the ANN model for the training, validating, testing, and overall 263
sets are shown in Fig. 5. All data processing was performed off-line using a commercial 264
software package (MATLAB R2013a, The MathWorks Inc.). As shown in this figure, 265
the model had acceptable prediction accuracy with an R2 of 0.73. In addition, the RMSE 266
(%) was calculated at 12%. The RMSE (%) was calculated as follows: 267
268
RMSE% = 100 ∗ &∑ [@��ABCD�A)*(EF$)GH%ICJI%D�A)*(KF$)]6MN O /∑ C%ICJI%D�A)*MN O (4) 269
270
FIGURE 5 271
272
Network Description 273
At the end of the learning phase, proper weights were assigned to every connection, and 274
proper biases were assigned to each neuron as follows: 275
276 PQR = S−37.4 −42.8 −0.03−0.2 −0.35 −0.17W 277
278 bY = S22.49−0.47W 279 W′Y] = [0.0271.80] 280 B] = 0.415
Page 13 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Elbagalati, et al. 14
Forward Calculations 281
Artificial neural network models are considered by many researchers as "black-boxes" 282
(Attoh-Okine et al. 2009; Hsie et al. 2012; Benítez et al. 1997; Prechelt 1998). With a 283
complex network structure, it is difficult to explicitly describe the learned relationship 284
between the input and the output variables. However, the simplicity of the model 285
presented in this study (only one hidden layer with only 2 neurons) offers a chance to 286
describe the network in a form of a simple equation. The general equation of a 287
backpropagation algorithm-based neural network with one hidden layer, one output 288
variable, and a tan-sigmoid (tansig) transfer function can be described as follows: 289
290 k = (B` + ∑ tansig(bY1∑ aBWBY)Of0Og0 WY]) (5) 291
where k= the model output at layer k; nj = number of neurons in the hidden layer; ni = 292
number of neurons in the input layer; ai = the input variables; and The tansig function can 293
be described as follows: 294
295 tansig(x) = �iG�ji�i1�ji (6) 296
297
The tansig transfer function forces the neurons in the hidden layer to produce outputs in 298
the range of -1 to +1, which accelerates the back-propagation algorithm (Vogl et al. 1988; 299
Dorofki et al. 2012). A linear activation function was then utilized to transfer the output 300
in layer k to the final output (Mr). The following expression describes the model 301
developed utilizing ANN to predict the subgrade resilient modulus based on RWD 302
measurements: 303
M� = 119.7 ∗ S0.415 + 0.027 ∗ tansig 222.49 − 37.4D� − 42.8σ$l − 0.03 $N� 5 +304 1.80 ∗ tansig 2−0.47 − 0.2D� − 0.35σ$l − 0.17 $N� 5W + 195.2 (7) 305
Page 14 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Elbagalati, et al. 15
Model Evaluation and Analysis 306
Limits of Agreement 307
The limits of agreement (LoA) methodology, developed by Bland and Altman, is a 308
simple powerful methodology for assessing agreements between two devices or 309
procedures (Bland and Altman 1986). The methodology was successful to the extent that 310
the reference that introduced this method has become one of the most cited statistical 311
papers (Ryan and woodall 2005). Bland and Altman concluded that using only regular 312
regression could be misleading when comparing two devices or methodologies for two 313
reasons. First, correlation depends on the range and distribution of the variables. 314
Second, correlation ignores any systematic bias between the two variables (Bland and 315
Altman 2003). A recent study concluded the usefulness of the LoA methodology for 316
comparing TSD and the FWD measurements (Katicha et al. 2013). 317
The procedure of the LoA methodology consists of the following steps: (1) plot a 318
chart with the differences between measurements by two methods on the Y-axis, and the 319
mean of the two measurements on the X-axis, (2) calculate the mean and the standard 320
deviation (σ) of the differences, and (3) calculate the mean ± 1.96 σ. One would then 321
expect 95% of differences between measurements by two methods to lie within these 322
limits. Fig. 6 shows the LoA between the subgrade resilient modulus values calculated 323
based on FWD and RWD measurements; the chart is also known as the Bland and 324
Altman chart. 325
As shown in Fig. 6, 95% of the differences between the Mr values calculated 326
based on the FWD and the RWD measurements did not exceed the range of ± 21 Mpa (± 327
3 ksi), which is reasonable especially at the network level. The figure provides a better 328
Page 15 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Elbagalati, et al. 16
understanding of the model accuracy in predicting Mr based on RWD data. The figure 329
also shows that the error in the predicted subgrade resilient modulus is independent of the 330
Mr value. 331
FIGURE 6 332
333
Model Validation 334
The generalization ability of the presented ANN model was tested using measurements 335
obtained from the testing program conducted at MnROAD. RWD data from 16 flexible 336
pavement testing cells were used as inputs in the ANN model to predict the subgrade 337
resilient modulus. The Mr predicted values were then compared with those calculated 338
based on FWD measurements, see Equation 1. The model showed acceptable accuracy 339
with an R2 of 0.72 and RMSE of 8% as shown in Fig. 7. 340
FIGURE 7 341
SUMMARY AND CONCLUSIONS 342
The objective of this study was to develop a model to utilize RWD measurements in 343
predicting the subgrade resilient modulus for flexible pavements. RWD and FWD 344
measurements were obtained from two comprehensive testing programs conducted in 345
Louisiana and Minnesota and were used to develop and validate an ANN model for 346
predicting the subgrade resilient modulus. The Louisiana testing program data were used 347
for developing the model, and data from the Minnesota testing program were used in the 348
model validation. The ANN model showed acceptable accuracy in both the development 349
and validation phases with coefficient of determination of 0.73 and 0.72, respectively. 350
The RMSE was found to be 12% and 8% in the development and the validation phases, 351
Page 16 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Elbagalati, et al. 17
respectively. Furthermore, the limits of agreement methodology showed that 95% of the 352
differences between the Mr values calculated based on FWD and RWD measurements 353
will not exceed the range of ± 21 MPa (± 3 ksi), which is acceptable especially at the 354
network level. 355
The availability of additional RWD and FWD testing will allow feeding and 356
refining the developed ANN model and will enhance its generalization ability. The 357
model presented in this study was developed and validated based on testing conducted on 358
flexible pavements. Conducting additional RWD and FWD testing on rigid pavement 359
sections will allow assessing the validity of the presented approach in predicting the 360
subgrade resilient modulus using RWD testing data conducted on different types of 361
pavements. 362
ACKNOWLEDGMENTS 363
The financial support of the Louisiana Transportation Research Center (LTRC) is greatly 364
appreciated. The authors also acknowledge the help of Nadarajah Sivaneswaran from the 365
FHWA Turner-Fairbank Highway Research Center in providing the MnROAD data. 366
REFERENCES 367
Abd El-Hakim, R., El-Badawy, S. M., Gabr, A. R., & Azam, A. M. 2016. Influence of
MEPDG Unbound Material Type and Material Characterization Input Level on
Pavement Performance. In Transportation Research Board 95th Annual
Meeting (No. 16-1668).
Abdel-Khalek, A., Elseifi, M., Gaspard, K., Zhang, Z., & Dasari, K. 2012. Model to
Estimate Pavement Structural Number at Network Level with Rolling Wheel
Page 17 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Elbagalati, et al. 18
Deflectometer Data. Transportation Research Record: Journal of the
Transportation Research Board, (2304), 142-149.
Ahmed, M. U., Hasan, M. M., & Tarefder, R. A. 2016. Study of Stress Dependency of
Unbound Layers Using Falling-Weight Deflectometer Tests. In Transportation
Research Board 95th Annual Meeting (No. 16-5706).
Attoh-Okine, N. O., Cooger, K., & Mensah, S. 2009. Multivariate adaptive regression
(MARS) and hinged hyperplanes (HHP) for doweled pavement performance
modeling. Construction and Building Materials, 23(9), 3020-3023.
Bayrak, M. B., & Ceylan, H. 2006. Backcalculation of Rigid Pavement Layer Parameters
Using Artificial Neural Networks. In Intelligent Engineering Systems through
Artificial Neural Networks, Volume 16. ASME Press.
Benítez, J. M., Castro, J. L., & Requena, I. 1997. Are artificial neural networks black
boxes? IEEE Transactions on neural networks, 8(5), 1156-1164.
Beleites, C., Neugebauer, U., Bocklitz, T., Krafft, C., & Popp, J. 2013. Sample size
planning for classification models. Analytica chimica acta, 760, 25-33.
Bland, J. M., & Altman, D. 1986. Statistical methods for assessing agreement between
two methods of clinical measurement. The lancet,327(8476), 307-310.
Bland, J. M., & Altman, D. G. 2003. Applying the right statistics: analyses of
measurement studies. Ultrasound in obstetrics & gynecology, 22(1), 85-93.
Briggs, R. C., Johnson, R. F., Stubstad, R. N., & Pierce, L. 2000. A comparison of the
rolling weight deflectometer with the falling weight deflectometer.
Page 18 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Elbagalati, et al. 19
In Nondestructive Testing of Pavements and Backcalculation of Moduli: Third
Volume. ASTM International.
Ceylan, H., Guclu, A., Tutumluer, E., & Thompson, M. R. 2005. Backcalculation of full-
depth asphalt pavement layer moduli considering nonlinear stress-dependent
subgrade behavior. International Journal of Pavement Engineering, 6(3), 171-182.
Ceylan, H., Bayrak, M. B., & Gopalakrishnan, K. 2014. Neural Networks Applications in
Pavement Engineering: A Recent Survey. International Journal of Pavement
Research and Technology, 7(6), 434-444.
Darter, M. I., Elliott, R. P., & Hall, K. T. 1991. Revision of AASHTO Pavement Overlay
Design Procedures: Appendix: Documentation of Design Procedures. National
Cooperative Highway Research Program.
Dorofki, M., Elshafie, A. H., Jaafar, O., Karim, O. A., & Mastura, S. 2012. Comparison
of artificial neural network transfer functions abilities to simulate extreme runoff
data. International Proceedings of Chemical, Biological and Environmental
Engineering, 33, 39-44.
Elseifi, M. A., Abdel-Khalek, A. M., Gaspard, K., Zhang, Z., & Ismail, S. 2011.
Evaluation of continuous deflection testing using the rolling wheel deflectometer
in Louisiana. Journal of Transportation Engineering, 138(4), 414-422.
Flintsch, G., Katicha, S., Bryce, J., Ferne, B., Nell, S., & Diefenderfer, B.
2013. Assessment of continuous pavement deflection measuring
technologies (No. SHRP 2 Report S2-R06F-RW-1).
Page 19 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Elbagalati, et al. 20
Gopalakrishnan, K., Thompson, M. R., & Manik, A. 2006. Rapid finite-element based
airport pavement moduli solutions using neural networks.International Journal of
Computational Intelligence, 3(1), 63-71.
Hsie, M., Ho, Y. F., Lin, C. T., & Yeh, I. C. 2012. Modeling asphalt pavement overlay
transverse cracks using the genetic operation tree and Levenberg–Marquardt
Method. Expert Systems with Applications, 39(5), 4874-4881.
Hossain, Z., Zaman, M., Doiron, C., & Solanki, P. 2011. Characterization of subgrade
resilient modulus for pavement design. In Geo-Frontiers Congress 2011.
Karlaftis, M. G., & Vlahogianni, E. I. 2011. Statistical methods versus neural networks in
transportation research: Differences, similarities and some
insights. Transportation Research Part C: Emerging Technologies,19(3), 387-399.
Katicha, S. W., Flintsch, G. W., Ferne, B., & Bryce, J. 2014. Limits of agreement method
for comparing TSD and FWD measurements.International Journal of Pavement
Engineering, 15(6), 532-541.
Kim, M. Y., Burton, M., Prozzi, J. A., & Murphy, M. 2014. Maintenance and
rehabilitation project selection using artificial neural networks. InTransportation
Research Board 93rd Annual Meeting (No. 14-3620).
Kim, Y. R., & Park, H. G. 2002. Use of falling weight deflectometer multi-load data for
pavement strength estimation (No. FHWA/NC/2002-006,).
Kim, Y. R., Lee, Y. C., & Ranjithan, S. R. 2000. Flexible pavement condition evaluation
using deflection basin parameters and dynamic finite element analysis
Page 20 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Elbagalati, et al. 21
implemented by artificial neural networks. InNondestructive Testing of Pavements
and Backcalculation of Moduli: Third Volume. ASTM International.
Lawrence, S., Giles, C. L., & Tsoi, A. C. 1997. Lessons in neural network training:
Overfitting may be harder than expected. In AAAI/IAAI (pp. 540-545).
Leverington, D. 2012. A Basic Introduction to Feedforward Backpropagation Neural
Networks. Neural Network Basics.
Lukanen, E. O., Stubstad, R., & Briggs, R. 2000. Temperature predictions and adjustment
factors for asphalt pavement (No. FHWA-RD-98-085,).
Mohammad, L. N., Gaspard, K., Herath, A., & Nazzal, M. D. 2007. Comparative
evaluation of subgrade resilient modulus from non-destructive, in-situ, and
laboratory methods (No. FHWA/LA. 06/417).
Ping, W. C. V., Ling, C. C., & Sheng, B. 2012. Development of Resilient Modulus
Estimation Models for Florida Pavements. In Transportation Research Board 91st
Annual Meeting (No. 12-2177).
Plati, C., Georgiou, P., & Papavasiliou, V. 2016. Simulating pavement structural
condition using artificial neural networks. Structure and Infrastructure
Engineering, 12(9), 1127-1136.
Prechelt, L. 1998. Automatic early stopping using cross validation: quantifying the
criteria. Neural Networks, 11(4), 761-767.
Rada, G. R., & Nazarian, S. 2011. The state-of-the-technology of moving pavement
deflection testing. Final Report, FHWA-DTFH61-08-D-00025, US Department of
Transportation, Washington, DC.
Page 21 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Elbagalati, et al. 22
Rada, G., Nazarian, S., Daleiden, J., & Yu, T. 2011. Moving Pavement Deflection
Testing Devices: State of the Technology and Best Uses. InProc., 8th
International Conference on Managing Pavement Assets, Santiago, Chile.
Rahim, A. M. 2005. Subgrade soil index properties to estimate resilient modulus for
pavement design. International Journal of Pavement Engineering, 6(3), 163-169.
Rakesh, N., Jain, A. K., Reddy, M. A., & Reddy, K. S. 2006. Artificial neural networks—
genetic algorithm based model for backcalculation of pavement layer
moduli. International Journal of Pavement Engineering, 7(3), 221-230.
Rolling Wheel Deflectometer. https://www.ara.com/projects/rolling-wheel-deflectometer.
Accessed Jun. 20, 2016.
Ryan, T. P., & Woodall, W. H. 2005. The most-cited statistical papers.Journal of Applied
Statistics, 32(5), 461-474.
Saltan, M., Uz, V. E., & Aktas, B. 2013. Artificial neural networks–based
backcalculation of the structural properties of a typical flexible pavement.Neural
Computing and Applications, 23(6), 1703-1710.
Schwartz, C., Li, R., Ceylan, H., Kim, S., & Gopalakrishnan, K. 2013. Global sensitivity
analysis of mechanistic-empirical performance predictions for flexible
pavements. Transportation Research Record: Journal of the Transportation
Research Board, (2368), 12-23.
Sivaneswaran, N. 2014. Network Level Pavement Structural Evaluations-A Way
Forward. In Pavement Evaluation Conference 2014.
Page 22 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Elbagalati, et al. 23
Tarefder, R. A., Ahsan, S., & Ahmed, M. U. 2014. Neural network–based thickness
determination model to improve backcalculation of layer moduli without
coring. International Journal of Geomechanics, 15(3), 04014058.
Vogl, T. P., Mangis, J. K., Rigler, A. K., Zink, W. T., & Alkon, D. L. 1988. Accelerating
the convergence of the back-propagation method. Biological cybernetics, 59(4-5),
257-263.
Wu, Z., & Gaspard, K. 2009. Mechanistic flexible pavement overlay design
program (No. FHWA/LA. 08/454).
Wu, Z., Zhang, Z., & Abadie, C. 2013. Determining structural strength of existing asphalt
layer using condition survey data. International Journal of Pavement
Engineering, 14(7), 603-611.
Ye, Z., Xu, Y., Veneziano, D., & Shi, X. 2014. Evaluation of winter maintenance
chemicals and crashes with an artificial neural network.Transportation Research
Record: Journal of the Transportation Research Board, (2440), 43-50.
Page 23 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Elbagalati, et al. 24
List of Tables
Table 1. Comparison between RWD and FWD Characteristics
Table 2. Correlation between RWD measurements and the FWD Subgrade Resilient
Modulus
List of Figures
Fig. 1. General Layout of the Rolling Wheel Deflectometer
Fig. 2. Correlation between the Mr and (a) D1 and (b) D0
Fig. 3. Structure of the ANN Network
Fig. 4. ANN Model Performance
Fig. 5. Regression plots of the developed ANN model for (a) the training data set (b) the
validation data set (c) the testing data set and (d) All data
Fig. 6. Bland and Altman Chart for Mr calculated based on FWD and RWD
measurements
Fig. 7. Model Validation Using Data from the MnROAD Testing Program
368
369
370
371
372
373
374
375
376
377
Page 24 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Elbagalati, et al. 25
378
Table 1. Comparison between RWD and FWD Characteristics 379
Factor RWD FWD
Operational Speed Traffic Speed Stationary
Deflection Sensor
Accuracy
6.25 microns 0.254 microns
Number of Operators 2 1
Productivity (km/day) 160 to 320 4 to 40
Number of Sensors 1 to 2 7 to 9
Applied Load (kN) 80 26 to 80
Load Type Transient wheel load Impact circular plate
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
Page 25 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Elbagalati, et al. 26
402
Table 2. Correlation between RWD measurements and the FWD Subgrade Resilient 403
Modulus 404
Parameter P-Value R2
D0 <0.0001 0.2950
σD0 <0.0001 0.4023
D1 <0.0005 0.1933
σD1 0.9087 0.1679
405
406
407
408
409
410 411
Page 26 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Page 27 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
(a)
(b)
R² = 0.1933
0
20
40
60
80
100
120
140
160
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35
MrMPa
D1 mm
R² = 0.2954
0
20
40
60
80
100
120
140
160
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Mr MPa
D0 mm
Page 28 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Page 29 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Page 30 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
Page 31 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
-30
-20
-10
0
10
20
30
40 50 60 70 80 90 100 110 120 130
Mr difference
Mr mean
Mean difference Upper limit of agreement Lower limit of agreement
Page 32 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering
Draft
R² = 0.7175
60
70
80
90
100
110
120
130
60 70 80 90 100 110 120 130
Mr (RWD) MPa
Mr (FWD) MPa
Page 33 of 33
https://mc06.manuscriptcentral.com/cjce-pubs
Canadian Journal of Civil Engineering