Professional Documents
Culture Documents
Neural Network Estimating Compaction Quality of Hot Asphalt Mixes
Neural Network Estimating Compaction Quality of Hot Asphalt Mixes
Abstract: Continuous real-time estimating of compaction quality during the construction of a hot mix asphalt (HMA) pavement is addressed
in this paper. The densification of asphalt pavements during construction usually is accomplished by using vibratory compactors. During
compaction, the compactor and the asphalt mat form a coupled system whose dynamics are influenced by the changing stiffness of the mat.
The measured vibrations of the compactor along with process parameters such as lift thickness, mix type, mix temperature, and compaction
pressure can be used to predict the asphalt mat density. Contrary to existing techniques in the literature in which a model is developed to fit
experimental data and to predict mat density, a neural network-based approach is adopted that is model-free and uses pattern-recognition
techniques to estimate density. The neural network is designed to read the entire frequency spectrum of roller vibrations and to classify these
vibrations into different levels. The intelligent asphalt compaction analyzer (IACA) is then trained to convert these vibration levels into a
“number” indicative of the asphalt mat density at a given location. This two-step process eliminates the need for regression analysis and
produces more accurate density measurements than those reported elsewhere in the literature. Compaction studies of HMA mixes on a stiff
subgrade indicate that the changes in the vibration characteristics of the roller are attributable to an increased compaction of the HMA base.
The results also show that, with the neural network working as a classifier, the IACA can estimate the density continuously, and in real time,
with accuracy levels adequate for quality control in the field. DOI: 10.1061/(ASCE)CO.1943-7862.0000343. © 2011 American Society of
Civil Engineers.
CE Database subject headings: Asphalt pavements; Artificial intelligence; Compaction; Quality control; Vibration; Neural networks.
Author keywords: Asphalt pavements; Artificial intelligence; Compaction; Quality control; Vibration.
the asphalt community. modulus and depends on a number of design factors such as the
Intelligent compaction techniques gradually have been accepted viscosity of the asphalt, η, and the effective asphalt content %
by the construction industry; however, existing techniques are not by volume, V beff , and construction parameters such as the percent-
yet commercially available largely because of their inability to age air void. The dynamic modulus, E in 689,475.728 kPA
account for factors in the field such as the characteristics of the (105 psi), can be expressed by using the Witczak equation as fol-
compactor, subgrade characteristics, and mix properties that cause lows (Ayers and Witczak 1998):
inaccuracies in the estimated density. These unaccounted for
parameters make the practical application of the techniques diffi-
cult. In research conducted at the University of Oklahoma log E ¼ 1:249937 þ 0:02923ρ200 0:001767ðρ200 Þ2
(Commuri and Zaman 2008), a neural network-based strategy
was implemented to estimate the level of compaction. The intelli- 0:002841ρ4 0:005809V a 0:82208V beff =ðV beff þ V a Þ
gent asphalt compaction analyzer (IACA) developed in that study þ ½3:871977 0:0021ρ4 þ 0:003958ρ38 0:000017ðρ38 Þ2
was shown to be capable of estimating compaction density by using
an asphalt vibratory compactor in a laboratory setting. The neural þ 0:00547ρ34 =f1 þ e½0:6033130:313351 logðf Þ0:393532 logðηÞ g
network was shown to have the ability to classify the features ð2Þ
extracted from the vibration signals as those corresponding to
the densities of the asphalt specimen. Further, the generalization where f = loading frequency (Hz); η = binder viscosity at the
capabilities of the neural network enabled it to provide reasonable temperature of interest [100,000 kg=ðm · sÞ (106 poise)]; V a =
density estimates when presented with data different from the set air void content (% by volume); V beff = effective asphalt content
used to train the network. (% by volume); ρ34 = cumulative amount retained on the
In this paper, the validation of the IACA during compaction 19 mm sieve (% by mass); ρ38 = cumulative amount retained on
under controlled field conditions is presented. The hypothesis that the 9.5 mm sieve (% by mass); ρ4 = cumulative amount retained
a vibratory compactor and the HMA form a coupled system with on the 4.76 mm sieve (% by mass); and ρ200 = cumulative amount
characteristic vibrations during compaction formed the basis of the retained on the 0.075 mm sieve (% by mass).
IACA. To minimize the effect of the subgrade on compactor vibra- Even when the design parameters are fixed, the dynamic modu-
tions, a test strip with a rigid subgrade was first constructed, and the lus is influenced by the amount of air voids in the HMA specimen
functioning of the IACA was studied. Calibration procedures were compacted, as shown in Eq. (2). Because the vibration of the as-
developed by using vibration data and measured density from phalt compactor during the pavement construction is a function of
roadway cores. The performance of the IACA was then verified the dynamic modulus of the pavement, these vibrations can be
against the density measured with a nonnuclear density gauge monitored to estimate the amount of air voids in the compacted
(NNDG) (TransTech Systems 2005; Troxler Electronic Laboratories HMA. However, such estimates of the density assume that the
2007a), and the densities measured from the extracted cores of underlying subgrade does not have any influence on the nature
compacted pavement. of the compactor vibrations. If this were not the case, changing
subgrade properties causes variations in roller vibrations and such
variations cannot be accounted for correctly during construction.
Theoretical Background for Intelligent Compaction The neural network is designed to read the entire frequency
spectrum of roller vibrations and to classify these vibrations into
The behavior on an HMA pavement under traffic and environmen-
different levels. The IACA is then trained to convert these vibration
tal conditions is dependent on the properties of the individual com-
levels into a “number” indicative of the asphalt mat density at a
ponents and those of the volumetric composition of the mix. In
given location. This two step process eliminates the need for regres-
mechanistic-empirical modeling of HMA pavements, the stress-
sion analysis (Minchin et al. 2001) and produces more accurate
strain relationship under a continuous sinusoidal loading is defined
density measurements than those reported elsewhere in the
by the complex dynamic modulus, E . The complex modulus is
literature.
defined as the ratio of the sinusoidal stresses amplitude, σ, and
In the research presented in this paper, a test site with a rigid
the sinusoidal strain amplitude, ε. Thus, the complex dynamic
subgrade was constructed. This rigid subgrade did not cause any
modulus is mathematically expressed by the following equation
appreciable change in roller vibrations during the compaction.
(Clyne et al. 2003):
Thus, any variations in roller vibrations would be the result of
E ¼ σ=ε ¼ σ0 sinðαtÞ=ε0 sinðαt ϕÞ ð1Þ the densification of the underlying HMA layer. The design and con-
struction of the reinforced concrete slab that functioned as a rigid
where σ0 = peak (i.e., maximum) stress; ε0 = peak (i.e., maximum) subgrade for the validation of the IACA hypothesis is presented in
strain; ϕ = phase angle (rad); α = angular velocity (rad=s); and the next section. The compaction of HMA mixes was then per-
t = time (s). formed on the top of this concrete pad in a controlled manner.
To minimize the effect of the subgrade on compactor vibrations, a Fast Fourier Transform
test pad consisting of a continuously reinforced concrete pavement
The frequency content of a continuous time signal xðtÞ can be an-
(CRCP) was designed to provide a stiff uniform subgrade over
alyzed by using the well-known Fourier transform. In the case of
which the HMA overlays could be constructed. It was anticipated
sampled signals, the discrete time Fourier transform (DTFT)
that the properties of such a subgrade would not alter during the
operates on a periodic discrete signal and relates it to a periodic
course of compaction. Thus, any changes observed in the vibration
continuous frequency spectrum. If x½n, n ¼ 0; …; N 1, is a col-
spectrum of the compactor during construction would be a result of
lection of “N” samples of xðtÞ obtained by sampling xðtÞ at a rate
the changing properties of the asphalt mat.
of f s Hz, then the DTFT Xðejω Þ decomposes the sequence x½n
The test site selected was a stretch of unused road on Mendel
into sine and cosine waves, with frequencies equally spaced
Downloaded from ascelibrary.org by IMPERIAL COLLEGE LONDON on 05/15/13. Copyright ASCE. For personal use only; all rights reserved.
was used in the design of the concrete slab for the test site. Xðejω Þ ¼ Re XðωÞ þ jImXðωÞ ¼ x½nejωn ð4Þ
n¼∞
The existing subgrade was stabilized with cement kiln dust and
compacted to 1;675 kg=m3 (104:4 lb=ft3 ). The subgrade moisture
The discrete Fourier transform (DFT) of the sequence x½n,
content was also determined to be within 2% of the optimum mois-
n ¼ 0; …; N 1 is a finite length sequence:
ture content (i.e., 19%). A concrete CRCP slab 4.27 m (14 ft) wide,
106.7 m (350 ft) long, and 15.25 cm (6 in.) thick was then con- X
N 1
structed on top of the compacted subgrade. The completed test strip X½k ¼ Xðejω Þjω¼2πk=N ¼ x½nej2πkn=N ; 0≤k ≤N1
is shown in Fig. 1. n¼0
ð5Þ
Experimental Setup for Use in Simulated Field Tests The fast Fourier transform (FFT) (Ingle and Proakis 2007) is a
practical approach to the numerical computation of the DTFT for a
The IACA functions on the hypothesis that the vibratory roller and finite length sequence and provided the power contained at each
the underlying HMA pavement form a coupled system whose re- frequency in the spectrum of the signal.
sponse is determined by the excitation frequency and the natural
vibration modes of the coupled system. To analyze the roller vibra- Experimental Setup
tions, triaxial accelerometers were fixed on the axle of the roller
drum, and roller vibrations were captured by using a data acquis- The experimental setup used to examine the changes in the fre-
ition system. The following discussion relates to the extraction of quency content of vibrations during the compaction process is
shown in Fig. 2. This experimental setup consisted of an Ingersoll-
Rand DD138HF dual-drum vibratory compactor (Volvo Construc-
tion Equipment 2011) instrumented with accelerometers, and a
real-time data acquisition system to analyze the vibration character-
istics and predict density. Vibrations of the roller during compac-
tion were translated into voltages with a triaxial accelerometer
capable of measuring accelerations along three orthogonal axes.
An accelerometer model CXL10HF3 from Crossbow (Moog
Crossbow 2007), capable of measuring 10 g acceleration up to
10 kHz, was mounted on the axle of the roller drum to measure
drum vibrations during compaction tests. The signal produced
by this accelerometer was then read by the data acquisition system.
The data acquisition system used, in this case, the xPC Target, a
rapid prototyping tool that can convert graphical models of data
acquisition circuitry into software that can be executed in real time.
The xPC Target is an Intel Pentium processor-based embedded
computer and is configured with Simulink. The Simulink software
is widely used for graphical programming and has capabilities that
allow for designing and testing systems with real data. Further-
Fig. 1. Completed test strip with CRCP subgrade (Image by S. Commuri,
more, models created in Simulink can be compiled to run in real
A. Mai, and M. Zaman)
time on different hardware platforms.
Fig. 2. Experimental setup: (a) compactor instrumentation; (b) functional schematic of analyzer
Fig. 3. Spectrograms of compaction vibrations for different process parameters: (a) Interstate I35—50.8 mm (2 in.) (S3–PG 64-22OK); (b) test strip
shown in Fig. 1—76.2 mm (3 in.) (S3–PG 64-22OK) on 152.4 mm (6 in.) concrete base; (c) Will Rogers Airport, Oklahoma City—76.2 mm (4 in.)
(S3–PG 64-22OK) on compacted subgrade; (d) Will Rogers Airport, Oklahoma City—50.8 mm (2 in.) (S4–PG70-28OK) on 203.2 mm (8 in.) S3
• The key features of the signals (i.e., frequencies and the corre- nal that could be reconstructed back by using the sampled signal.
sponding power) were extracted. To classify these vibrations, the 200 elements corresponding to
• The extracted features were compared with the features corre- the response above the excitation frequency of the compactor were
sponding to a set of known densities. used as input to the classifier. The NN classifier implemented was a
• The predicted density was calculated from the results of the pre- three-layer NN with 200 inputs, 10 nodes in the input layer, four
vious step and knowledge of the process parameters (e.g., mix nodes in the hidden layer, and one node in the output layer. The
type, mat temperature, and type of compactor). inputs of the NN corresponded to the outputs of the feature extrac-
The sensor module consisted of accelerometers for measuring tion module; that is, in this case, 200 features of the frequency
the compactor vibrations during operation, infrared temperature spectrum were considered. The output corresponded to a signal
sensors for measuring the temperature of the mix, a means for se- indicative of the level of compaction reached. The method to
lecting the amplitude and frequency of the vibration motors, and a extract the training data and validate the performance of the com-
means for recording the mix type and lift thickness. The vibration paction analyzer is discussed in the next section.
signals were sampled at 1,000 samples per second with xPC Target
running on an Intel Pentium 4 processor and an IO301 embedded
data acquisition system. The sampled input was presented to the Table 1. Gradation for Mix S3 (PG64-22OK)
feature extractor (FE) module. The FE module implemented an Sieve number Job formula (% passing sieve)
FFT of the input signal to extract the features corresponding to
vibrations at different salient frequencies. Preprocessing the data 25.4 mm (1 in.) 100
to extract the features reduced the amount of data to be considered 19 mm (0.75 in.) 98
in the classification process and therefore, reduced the algorithmic 12.7 mm (0.5 in.) 88
complexity of the classifier. The neural network classifier was a 9.52 mm (0.375 in.) 72
multilayer neural network (NN) trained to classify the extracted Number 4; 4.75 mm (0.187 in.) 40
features into different classes. The compaction analyzer then Number 8; 2.36 mm (0.0929 in.) 30
postprocessesd the output of the NN and predicted the degree of Number 16; 1.18 mm (0.464 in.) 21
compaction in real time. Number 30; 0.6 mm (0.0236 in.) 16
Number 50; 0.3 mm (0.0118 in.) 11
Number 100; 0.15 mm (0.0059 in.) 8
Number 200; 0.075 mm (0.0029 in.) 4.2
Fig. 4. Output of compaction analyzer and actual density measure- Fig. 5. Output prediction error of neural network after each training
ments from NDG cycle
excluding air voids, and “D” represents the width of the roller The performance of the IACA prototype was analyzed during the
drum. The target compaction in the field is specified as a percentage compaction of asphalt mixes on the controlled test strip described
Fig. 7. Spectrogram of effect of changes in density between first and second pass
Fig. 8. Comparison of predicted and measured density: (a) density predicted by the IACA; (b) density measured with NNDG
in Fig. 1. The pavement was 91.44 m (300 ft) long, 3.6576 m The measured core densities were used to calibrate the NNDG. The
(12 ft) wide, 7.63 cm (3 in.) thick and was constructed from an calibration data was used to adjust the raw density readings obtained
S3 (PG64-22OK) mix (see Table 1). Initially, several overlays were by the NNDG to compare the readings with those of the cores.
constructed with the S3 mix, the vibrations of the machine were The vibration data from the spectrogram was correlated with the
collected, and the corresponding spectrograms were computed. density measurements to extract the data for training the neural net-
Several readings with an NNDG were also recorded during each work. Locations on the mat with densities of 90, 92, and 94% were
roller pass. On the completion of the overlay, several cores were identified, and the FFT output corresponding to these locations
extracted from the compacted pavement and their density was mea- were identified by using the GPS measurements. Eight columns
of FFT data, corresponding to a linear travel of 0.3048 m (1 ft),
sured in the laboratory in accordance with the AASHTO T 166
were selected at each of these locations to constitute the training
(AASHTO 2003) and OHD L-45 (Oklahoma DOT 2008) standards.
data for the neural network. The training error for each epoch
of the training is shown in Fig. 5. The training was stopped once
the required precision (106 , corresponding to 1 prediction error in
106 trials when using the training data) was obtained.
The performance of the trained IACA was verified during the
construction of asphalt pavement on the test strip. The output of
the accelerometer and the GPS measurements of the compactor lo-
cation were collected, and the spectrogram was plotted against the
distance traveled by the compactor for each roller pass. After each
roller pass, the density was measured with the NNDG at specific
points on the asphalt mat (Table 2). The densities measured after
each pass are shown in Fig. 6. In the first three passes, the density
mostly increases after each pass. However, “roll over” occurs after
the third pass, and subsequent roller passes cause a reduction in the
density of the compacted pavement. Such observed roll over just
describes the compaction process and is independent of the IACA
readings. The spectrogram of the compactor vibrations over the
first two passes is shown in Fig. 7, for which the effect of increased
density on the compactor vibration is observed easily.
The data from Table 2 were used to train the IACA to extract the
relevant features from the vibration signal and to estimate the level
of compaction. The estimated density during the final pass of the
first stretch is shown in Fig. 8. The predicted density correlates very
well with the densities measured with the NNDG. Fig. 9 shows the
final compacted density of the entire test strip, as predicted by the
IACA. The systematic scatter shown in Fig. 9 can be attributed to
the “double bounce” typically observed when compaction is per-
formed on a highly compacted or rigid subgrade. Usually, this con-
dition is detected by the roller operator, and the amplitude setting of
the eccentrics is adjusted to minimize the double bounce. The effect
Fig. 10. Comparison of measured density and extracted core density:
of double bounce on the crushing of aggregates or the compaction
(a) NNDG; (b) IACA
quality was not investigated in this research. A comparison of the
Fig. 12. Spectrogram of roller vibration and predicted density of roller pass during construction
Acknowledgments
References
AASHTO. (2003). “Bulk specific gravity of compacted bituminous mix-
tures using saturated surface dry specimens.” T 166-05, Washington,
DC.
AASHTO. (2004). “Standard method of test for the moisture-density rela-
tions of soils using a 2.5-kg (5.5-lb) rammer and a 305-mm (12-in.)
drop.” T 99, Washington, DC.
Ammann Schweiz AG. (2007). “Intelligent compaction.” 〈http://www
.ammann-group.com/en/home/technology/intelligent-ground-compaction/〉
(Oct. 1, 2007).
Fig. 13. Region showing different compaction levels on north and Ayres, M. Jr., and Witczak, M. W. (1998). “AYMA: Mechanistic probabi-
south lanes of shoulder listic system to evaluate flexible pavement performance.” Transporta-
tion Research Record 1629, Transportation Research Board,
Washington, DC., 137–148.
be between Level 2 and Level 3, that is, between 92 and 94% BOMAG. (2007). “VarioControl.” 〈http://www.bomag.com/world/EN/
variocontrol.htm〉 (Oct. 1, 2007).
compaction. The complete as-built compaction map and corre-
Clyne, T. R., Li, X., Marasteanu, M. O., and Skok, E. L. (2003). “Dynamic
sponding density measurements from the use of the NNDG of and resilient modulus of MN/DOT asphalt mixes.” MN/RC-2003-09,
the north and south lanes are shown in Fig. 13. The estimated den- Minnesota Dept. of Transportation, St. Paul, MN, 4–14.
sity of 90, 92, and 94% are indicated by the gray intensity with the Commuri, S., and Zaman, M. (2008). “A novel neural network-based
dark gray area corresponding to the highest density (i.e., 94%). The asphalt compaction analyzer.” Int. J. Pavement Eng., 9(3), 177–188.
density estimated by the IACA correlates well with the density GEODYNAMIK. (2007). “Compactometer, compaction meter for vibra-
measured by an NNDG. In the northern lane, the compaction qual- tory rollers ALFA-030.” 〈http://www.geodynamik.com/languages/pdf/
ity is good (∼93%) and the variations in density are significant ma30gb203.pdf〉, (Oct. 1, 2007).
(91–94%), which match the change in the gray intensity of the Hawaii Asphalt Paving Industry (HAPI). (2011). “Compaction.” 〈http://
hawaiiasphalt.org/index.php/compaction.html〉 (Aug. 18, 2011).
as-built map. In the southern lane, the NNDG readings indicate
Ingle, V. K., and Proakis, J. G. (2007). “Digital signal processing using
uniformly good compaction (∼93%), which corresponds to more Matlab.” Bookware companion series, 2nd Ed., Thomson Engineering,
regions represented in dark gray. Toronto, 119–174.
These results indicate that the IACA can predict the density con- Jaselskis, J. E., Han, H., Tan, L., and Grigas, J. (1998). “Roller mountable
tinuously in real time during pavement compaction. The prediction asphalt pavement quality indicator.” Proc., 1998 Transportation Conf.,
accuracy is comparable to the accuracy of the NNDG. Transportation Research Board, Washington, DC, 192–194.
Minchin, R. E. (1999). “An asphalt compaction quality control model using
vibration signature analysis.” Ph.D. thesis, The Pennsylvania State
Conclusions Univ., Philadelphia.
Minchin, R. E., and Thomas, H. R. (2003). “Validation of vibration-based
In this paper, the design of a neural network-based IACA was onboard asphalt density measuring system.” J. Constr. Eng. Manage.,
129(1), 1–7.
presented. A procedure to calibrate the IACA by using field com-
Minchin, R. E., Thomas, H. R., and Swanson, D. C. (2001). “Theory
paction data was propsed, and the performance of the IACA was behind vibration-based, onboard asphalt, density measuring system.”
validated during construction in controlled field settings. The ex- Transportation Research Record 1761, Transportation Research Board,
perimental results showed that the IACA can be trained to extract Washington, DC, 70–78.
the features from the compactor vibrations and to use these features Moog Crossbow. (2007) “Introduction to accelerometer applications.” 〈http://
to estimate the level of compaction. The calibration procedure is www.xbow.com/defense-solutions/support/appnotes-accelerometer
straightforward and can easily account for the different parameters .html〉 (Oct. 1, 2007).
such as base/intermediate/surface course, lift thickness, mix type, Mooney, M. A., Gorman, P. B., and Gonzalez, J. N. (2005). “Vibration
and construction equipment encountered at each site. The IACA based health monitoring during earthwork construction.” Struct. Health
Monit., 4(2), 137–152.
was calibrated by using the maximum theoretical density, which Mooney, M. A., Tawfik, E. F., Chan, G. B., and Pan, J. L. (2002). “Health
accounts for the difference in the type of mixes. The estimated den- monitoring during vibratory compaction of soil.” Proc. SPIE 4696,
sity correlated well with the density measured from compacted 17(2002), 112–123.
cores, and the measurement error was comparable to the errors ob- Oklahoma DOT. (2008). “Method of test for determining the specific
served when using tools that measure density at discrete points. gravity and unit weight of compacted bituminous mixtures using the