Fa gue behavior study of a dense graded HMA using the four point bending beam test aided by an in-situ instrumenta on at BR-116/RS, Brazil

Federal University of Rio Grande do Sul – Brazil, grabc@hotmail.com Federal University of Rio Grande do Sul – Brazil, lelio.brito@ufrgs.br State University of Rio Grande do Sul and Slac Na,onal Laboratory – Brazil, dionisio.doering@ufrgs.br North Carolina State University – United States, dmar,23@ncsu.edu State University of Rio Grande do Sul – Brazil, nplhilgert@gmail.com Federal University of Rio Grande do Sul – Brazil, marlova.johnston@ufrgs.br Federal University of Rio Grande do Sul – Brazil, jorge.cera1@ufrgs.br


INTRODUCTION
The fatigue cracking is recognized as one of the main distress types in asphalt concrete pavements (Pell et al., 1961;Rowe, 1996;Harvey et al., 1995;Di Benedetto et al., 2004). This phenomenon occurs due to the accumulation of damage under the effect of repeated traf+ic loading. The tensile strain developed in the bottom of the wearing course leads the layer to cracking and consequent loss of structural capacity.
Various laboratory testing methods have been developed to characterize the fatigue response of asphalt concrete mixtures. In Brazil, the test commonly employed to analyze the fatigue life of asphalt mixtures is the indirect tensile test by diametral compression in cylindrical samples (IDT). This test is relatively simple to conduct and low cost, but feature some acknowledged disadvantages, such as permanent deformation on the loading strips due to high shear stresses, impossibility of separating damage due to fatigue from damage due to excessive creep +low for testing in inelastic materials, such as asphalt, and underestimation of the fatigue life (Tangela et al., 1990;Di Benedetto et al., 2004). Nonetheless, four point bending beam tests are recognized to display a better ability to simulate +ield conditions (Tayebali et al., 1994). In Brazil, the four point bending beam test has being employed for asphalt mixtures fatigue characterization in various research projects, such as Colpo (2014), Klinsky et al. (2014), Melo (2014), Ceratti et al. (2017) and Bessa (2017).
The monitoring of stresses and strains developed in pavement structures is of interest to pavement researches and engineers for a proper understanding of their structural behavior, which is crucial to pavement design, performance prediction and road management systems (Zejiao, Yiqiu and Meili, 2012). Measuring of the mechanical and environmental response of an in-service segment through instrumentation allows the comparison with laboratory behavior of the same asphalt mixture.
Pavement instrumentation can be an important tool for monitoring in-service material performance, structural responses under actual environmental and loading conditions, and also assist in the calibration of models (Leiva-Villacorta and Timm, 2012). Various instrumentation efforts pursued such purpose, allowing the validation and calibration of models that are actually used in mechanistic-empirical pavement design methods (Timm, Priest and McEwen, 2004;Priest, 2005;Ullidtz et al., 2006).
In tolled roads the development of a signi+icant super+icial cracked area is limited by the control agencies and are subject to a penalty. Hence, pavement distress evolution depends on de-+lection control to help evaluating road structure as surfacing works are recurrent and the cracking assessment clouded. Because such procedure is affected by all layers' resilient behavior interdependence and is in+luenced by weather, testing protocol and loci as well as equipment, tools to aid engineers to better evaluate the pavement structural condition ought to be investigated. All considered, a proper fatigue performance model becomes vital to anticipate the right moment of an intervention or repair in the structure, thus allowing adequate network managing.
This research was divided into two stages: +ield instrumentation of a tolled road segment in a federal highway and laboratory fatigue tests by means of IDT and 4PBB tests. Since the relation between the +ield and laboratory response is considered a challenge for road engineers this work aims to contribute to a better outlook on the instrumentation support towards such effort.

BACKGROUND
Fatigue phenomenon is considered the main mechanism of distress in +lexible pavements (Pell et al., 1961;Preussler et al., 1981;Tangella et al., 1990;Matthews et al., 1993;Harvey et al., 1995;Kim et al., 1997;Di Benedetto et al., 2004;Tapsoba et al., 2013;Mateos et al., 2017), and is highly dependent not only on material characteristics, but also on traf+ic and environmental conditions, which makes its interpretation and modeling complex. The asphalt mixtures fatigue behavior is being studied throughout the years, with different considerations and approaches. Some approaches are purely empirical and observational, where the number of cycles at failure is related to the input of strain amplitude or stress level in a cyclic test. Other approaches include mechanistic elements and embodies a more phenomenological view, relating the number of cycles at failure with parameters such as initial stiffness, dissipated energy, fracture mechanics and continuum damage.
The classic approach to access the number of cycles at failure (Nf) for a given tensile strain amplitude or tensile stress level for asphalt mixture layers are so-called Wöhler curves (Equation 1 and 2).
where: Nf: Number of cycles at fatigue failure; εt: Tensile strain amplitude;

σt:
Tensile stress level; a, b, c, d: Regression coef+icient. Dissipated energy concept has been successfully used in the asphalt fatigue life prediction (Shen andCarpenter, 2005, 2007;Shen et al., 2006;Dondi et al., 2013;Subhy et al., 2017). Those methods are based on the assumption that the energy is absorbed by the material as damage, and the number of loading cycles at failure is related to the amount of energy dissipated during the test (Xiao, 2006).
Viscoelastic materials, when subjected to cyclic loading, present a hysteresis loop, once the loading and unloading path are not coincident due the viscoelasticity. The area inside this loop corresponds to the dissipated energy on that load cycle (Rowe, 1996;Ghuzlan and Carpenter, 2000), and can be calculated by Equation 3.
(3) where: DEi: Dissipated energy at load cycle i; εi: Strain amplitude at load cycle i; σi: Stress amplitude at load cycle i; φi: Phase shift between stress and strain at load cycle i. The classic fatigue models can be related to the initial dissipated energy as follows (Equation 4).

Fa gue Field Monitoring
For a proper construction or rehabilitation design of pavement structures, it is necessary to address the materials behavior under a wide range of loads and environmental conditions, such as temperature. To develop an adequate performance model for a given structure, it is mandatory to understand the differences between laboratory performance tests and actual +ield responses. In this sense, +ield instrumentation can be a powerful tool to better understand the actual response of the pavement under different conditions, which is not possible to simulate in laboratory (Al-Qadi and Nassar, 2003). Pavement instrumentation has being used in several countries with satisfactory results, that helps to calibrate and validate models and pavement design methods (Brown, 1977;Tabatabafe and Sebaaly, 1990;Van Deusen et al., 1992;Silva e Motta, 1999;Priest and Timm, 2006;. Most of the studies, nonetheless, are carried out in experimental sections and controlled loading environment. Garg et al. (2018) used +ield instrumentation data to develop a HMA fatigue model based on the ratio of dissipated energy change (RDEC). Various studies (Rowe, 1996;Ghuzlan and Carpenter, 2000;Subhy et al., 2017) have shown that RDEC, de+ined as the change of the dissipated energy between two cycles divided by the dissipated energy of previous cycle, provides a better indication of damage induced by the repeated load. This fatigue model for HMA is used in airport pavement thickness design software FAARFIELD. Leandri et al. (2013) observed that simulations of longitudinal and transversal strains provide a good +it with the registered wave pulse, in terms of both shape and peak values. The simulated vertical stress at the bottom of asphalt concrete layer well reproduce the peak values even if the shape is not exactly replicated, the tails of the registered pulse decrease more sharply than simulations.
Silva e Motta (1999) captured the Brazilian experience with pavement instrumentation and accredited that the +irst road instrumentation to an experiment in 1976, aiming to analyze pozzolanic mixtures in the highway BR-101, in Santa Catarina state. Gonçalves (2002) used pavement-embedded instrumentation to evaluate six experimental segments subjected to a heavy vehicle simulator in an accelerated pavement loading facility on the Pavement Laboratory at Federal University of Rio Grande do Sul (LAPAV-UFRGS). Bastos (2016) also instrumented a pavement segment subjected to a heavy vehicle simulator, using strain gages, pressure cells, temperature and moisture sensors at Ceará state.

EXPERIMENTAL INVESTIGATION
The instrumentation process of the road segment and laboratory tests are described below. A bespoken data collection system with embedded signal processing and instruments signal-conditioners are detailed elsewhere (Ceratti et al., 2017).

Field instrumenta on
The structure of test section selected consists of a +lexible pavement with a bituminous layer of a hot mix asphalt (HMA) with polymer-modi+ied binder (8 cm thickness), a dense-graded aggregate base (15 cm) and granular subbase (30 cm) over a clayey subgrade.
Sensors embedded in pavement sections have proved useful in providing measured responses as a basis of comparison. Various researches have instrumented pavement layers using strain gages and pressure cells to measure dynamic response to load in the form of strains and pressures (Timm, Priest and McEwen, 2004;Gonçalves, 2002;Leandri et al., 2013;Chaves et al., 2016). Following the same line of work, a road segment on BR-116/RS, in southern Brazil, was instrumented with these sensors in the bottom of asphalt wearing course.
Sensors were selected in accordance with literature and the researchers experience and included six strain gages (SG) (Figure 1a KM-120-120-H2-11-W1M) were oriented to measure longitudinal, transverse and 45° strain, and the thermistors (model LM35) have the temperature range from -55 ºC to +150 ºC with ± 0.5ºC accuracy.

Figure 1. Sensors used on the experimental segment
All the sensors were installed in the bottom of the pavement wearing course. The +irst set (PC2 + 3 SG) was spaced 5 m apart longitudinally from the second set (PC1 + 3 SG) and located under the external wheel path of the external road lane. Figure 2 shows the sensor layout along the external wheel path.

Figure 2. Sensors installation layout
A custom data acquisition system (DAS) was developed as part of this research. The DAS performs signal conditioning, analog signal +iltering and ampli+ication, digitization and data preprocessing before sending it to an embedded single board computer for storage. The pre-processing implements a high pass +ilter to remove base line level variations (caused mainly due to temperature variations) and a trigger algorithm that sends a +lag indicating the start of a dynamic event. The event +lag is set when the signal variation is higher than a user de+ined threshold. The system samples the analog signals at a rate of 450 points per second, per channel.

Laboratory tests
To evaluate the laboratory performance two test con+igurations were used, IDT and 4PBB. The specimens required for the tests were cored direct from the +ield, both cylindrical (for IDT tests) and prismatic (for the 4PBB tests). The procedure allowed evaluating the actual in situ density condition, as mixtures produced in an asphalt plant and in the laboratory may display differences in performance (Rahbar-Rastegar and Daniel, 2016). Binder content for the hot mix asphalt was 5.24% by weigh and featured apparent speci+ic gravity of 2.289.
The 4PBB test were carried out under controlled strain mode of 400 microstrains (με), 500 με and 600 με. In this study, a repeated sinusoidal loading at a frequency of 10 Hz was used. All tests were performed in a temperature-controlled chamber at 20 ± 0.5°C. The control and data acquisition software (UTS015) measured the de+lection of the beam specimen, computed the strain peak to peak in the specimen and adjusted the load applied by the loading device as per AASHTO T321 (2014).
The test apparatus also recorded load cycles, applied strain peak to peak, and beam de+lections. Failure was assumed to occur when the stiffness reached half of its initial value, which is determined from the load at approximately 50 repetitions; the test was terminated automatically when stiffness was reduced by 50%. The dissipated energy, tensile stress and strain and +lexural stiffness of fatigue beam are determined by Equations 3, 5, 6 and 7 with the software UTS015 (AASHTO T321, 2014). (5) where: σt: Tensile stress amplitude (kPa); P: Peak force (kN); L: Support span width (mm); b: Average beam width (mm); h: Average beam height (mm); εt: Tensile strain amplitude (με); δ: Peak de+lection at the center of beam (mm); a: Loading span width (mm); S: Flexural stiffness (MPa). Note the validity of the equations above assume homogeneity of a linear viscoelastic, isotropic material. These equations consider the most stressed +iber in the bottom of the specimen, which should be the more stressed +iber, and consider the stresses/strains to be linear across the center transversal section of the specimen. Furthermore, the assumptions involved in the equations above are good approximations in the +irst cycles whilst damage is reduced and material integrity is unchanged. Although such considerations limit a fatigue study, they are rather simple approaches for application in the present research. The hysteresis loop is calculated using these assumptions and so are calculated for the most stressed +iber.
The IDT was conducted under controlled stress mode of about 10%, 20%, 30%, 40% and 50% to the monotonic indirect tensile strength (ITS = 1.43 MPa). In this study, a cyclic haversine loading at 1Hz frequency, with pulse time of 0.1 s and 0.9 s of rest, was used. The failure criteria adopted was 50% reduction in the initial resilient modulus. All tests were performed in a temperature-controlled chamber at 20 ± 0.5°C. The tensile stress and resilient modulus (MR) are determined by Equations 8 and 9. (8) where: σt: Tensile stress level (MPa); P: Peak force (N); d: Average specimen diameter (mm); h: Average specimen height (mm); MR: Resilient modulus (MPa); ∆: Resilient displacement of the specimen (mm); µ: Coef+icient of Poisson.

Fa gue life
The results obtained for 4PBB fatigue tests are shown on Table 1. The results presented are the average of three specimens and the coef+icients of variation (COV) are also shown Figure 3 presents the fatigue plot of 4PBB fatigue tests considering tensile strain and initial dissipated energy.  The 4PBB fatigue tests performed on the mixture used on the experimental segment show a low variability with small coef+icient of variation. Both fatigue models, considering tensile strain and dissipated energy, show a good +itting to the data, with coef+icients of determination (R 2 ) of 0.94 and 0.93, respectively.
The hysteresis loops obtained for the 4PBB test with peak to peak strain of 500 µε is pictured on Figure 4. The loop area represents the dissipated energy due to dissipative phenomena as plastic dissipation and, mainly, damage growth.
The dissipated energy changes continuously along the fatigue process as the number of cycles increase, and the cracks initiate and propagate. In Figure 4 it can be observed a decrease in the loop area (i. e. energy) as the number of cycles increase, it happens as a consequence of the modulus reduction in the material throughout the test, in which the stress decreases to keep the displacement amplitude constant. The dissipated energy vs number of cycles can be +itted by a power function, as shown in Figure 5. The IDT fatigue tests results are shown on Table 2. The Figure 6 presents the fatigue curve of IDT fatigue tests.  The results for IDT fatigue tests shown high coef+icients of variation, which can be accredited to the instability promoted by the coring bit used to extract the specimens from the road segment, and also by the severe loads applied, once the coef+icient is higher for higher loads. Despite the variation, the coef+icient of determination for the +itting is high (R 2 = 0.98).
The hysteresis loops obtained for an IDT test under controlled stress mode (stress level equivalent to 20% of ITS) is pictured on Figure 7. The energy increases with growing number of cycles due to a combined effect of strain increase, due to material's micro cracking yielding to stiffness reduction, while stress is kept constant (Brito et al., 2008). It can be observed that in this test con+iguration there is an important strain accumulation on the material from cycle to cycle, this creep also contributes to damage. In IDT tests for viscoelastic materials is not possible to separate the creep damage from the cyclic loading.
The different nature on how stresses and temperatures are combined in a scattered fashion in a real pavement create an unorderly fatigue development, as opposed to the laboratory. In IDT, strains increase constantly along the test, under controlled stress level, whereas in 4PBB, stresses decrease along the test, under a constant amplitude of controlled strain, in both scenarios resulting in stiffness reduction due to materials' integrity degradation. Fatigue results are typically adjusted by a function which relates the effort imposed in the test to the number of cycles at fatigue failure. However, knowing the actual effort induced in the +ield is important to determine in which level the fatigue behavior should be analyzed, since the laboratory tests are carried out in different levels.

Signal processing -field data
The longitudinal strain gage shows alternate signal of contraction and extension, due to dynamic nature of the loads, which impose different efforts depending on the position of gage in relation to the axle, while the transversal strain is always extension and the vertical stress is always compression (Perret, 2003).
To validate the +ield signals, some modelling were carried out using the software 3D-MOVE. The pavement parameters used in the analysis are shown in Table 3. The longitudinal and transversal strain, and vertical pressure obtained from the instrumentation as well as from 3D-MOVE software are presented in Figure 7, for a single axle dual tires load with tire pressure of 560 kPa, speed of 80 km/h and temperature of 32ºC, for the asphalt layer. This temperature is based on the thermocouples reading embedded on the layer for the average temperature on which most of the readings were registered.
The simulated pavement responses are consistent with the +ield measurements, which validates the data acquisition system. Regarding the shape of the signal, the vertical stress, transversal strain and longitudinal strain shows a better match for the set of data collected from the instrumentation than the signal magnitude, which can be attributed to the axle position in relation to the sensors. According with Leandri et al. (2013), for little errors in the measurement of the relative position, the difference in the simulation can be high. The shape of stress load measured by the pressure cell (PC) shows a difference on the unloading path compared to the simulated data (Figure 8a). The shape of unloading path is related to the relaxation behavior of the asphalt materials and the observed difference can be either a data acquisition problem or an inability of the software to describe well this phenomenon of viscoelastic materials. Further investigation should be addressed to account for this difference. The simulated pavement responses are consistent with the +ield measurements, which validates the data acquisition system. Regarding the shape of the signal, the vertical stress, transversal strain and longitudinal strain shows a better match for the set of data collected from the instrumentation than the signal magnitude, which can be attributed to the axle position in relation to the sensors. According with Leandri et al. (2013), for little errors in the measurement of the relative position, the difference in the simulation can be high. The shape of stress load measured by the pressure cell (PC) shows a difference on the unloading path compared to the simulated data (Figure 8a). The shape of unloading path is related to the relaxation behavior of the asphalt materials and the observed difference can be either a data acquisition problem or an inability of the software to describe well this phenomenon of viscoelastic materials. Further investigation should be addressed to account for this difference

BR-116 Field Results
A series of 3000 events were collected to determine the level of stress and strains developed in the bottom of asphalt layer. The results (Figure 9) demonstrates that the longitudinal strains and vertical stresses rarely exceed 250 microstrains and 400 kPa, respectively. A "zero" vertical stress and "zero" longitudinal strain in Fgiure 9, means 0 to 50 stress or strain reading; it is likely these readings are produced by light vehicles or high stress contact loads further away from sensors. In laboratory, the strain level is usually higher than those obtained from +ield data, to make the tests faster. In the strain level observed in +ield data (less than 250 µɛ), the 4PBB shows a number of cycles to fatigue failure of 1.81*10 6 , while IDT shows a number of cycles to fatigue failure of 1.62*10 3 (for a stress of 400 kPa), according to the fatigue models calibrated for the mixtures in this study.
The number of repetitions of a standard axle (N) in the road segment during the trial period, based on the American Association of State Highway and Transportation Of+icials (AASHTO) equivalent factors, was of 1.05*10 6 ESALs. This value is higher than the Nf obtained from fatigue tests, even considering the same tensile strain observed in +ield; nonetheless, the pavement section did not develop any visible super+icial cracks during the trial period. Such difference corroborates to the importance of a correlation factor, or transfer function, from the laboratory to +ield result (Priest, 2005;Nascimento, 2015;Fritzen, 2016).
Considering a shift-factor of "4" to 4PBB tests  and "10 4 " for IDT tests (Pinto, 1991), from literature, the number of cycles at failure would be 7.24*10 6 and 6.41*10 7 , respectively, for 250 microstrains and 400 kPa. A 4PBB approximation would suggest a pavement outlasting 6,9 times more traf+ic than that from the trial period, while an IDT fatigue life would suggest 61 times more traf+ic until pavement failed due to fatigue cracking. Although results are not shown in this research, the continuous monitoring of the section assured a much better correlation to the 4PBB estimate.

CONCLUSIONS
The primary goal of this research was to draw a comparison between 4PBB and IDT fatigue tests results supported by an in situ instrumented test section. In order to accomplish the task, conceiving a reliable instrumentation capable of coping with real traf+ic condition was prime. The data acquisition system developed has proved to be effective and the readings were reliable, highlighting a good adherence between modelled and collected results. The +iltering and pre-processing of signals allowed a loading wave shape without distortions and interferences, therefore becoming easily understandable and interpretable. The algorithm for identi+ication of a new event (vehicle hit) is also adequate, showing a good sensibility for this application and optimizing data collection size.
The fatigue tests protocols used for the research demonstrated to produce good quality results. Regardless the aforementioned issues with IDT test con+iguration, the strain measurement allows for further analysis, such as dissipated energy calculation, which traditionally is not possible. The availability of a fatigue model for the 4PBB test using the dissipated energy concept turns available further discussion of the better correlation of this test to +ield data. Finally, although the pavement structure displayed a longer fatigue life span than laboratory tests, the 4PBB had a closer relation to +ield performance in comparison to IDT fatigue tests.