You are on page 1of 10
‘Advanced Testing and Characterization of Bituminous Materials — ‘Loizos, Pat, Scarpas & ALQadi (ods) (© 2009 Taylor & Francis Group, London, ISBN 978-0-415-55854-9 Enhanced algorithms for the derivation of material parameters from triaxial cyclic compression tests on asphalt specimen Karl Kappl & Ronald Blab Institute for Road Construction & Maintenance, Vienna University of Teenology, Vienna, Austria ABSTRACT: Material parameters, ie. complex modulus £* and phase lag gas a function of temperature and test frequency of asphalt mixtures, from cyclic material tests on asphalt specimens are fundamental input parameters for material characterization in mix design and ‘numerical calculations of pavement distress. Mostly, these material parameters are calculated by means of a “black box” software supplied by the machine manufacturer on the basis of | simple “peak finding” methods. However, it can be proven, that the used post-processing algorithm to interpret measurement data of cyclic tests is crucial on the accuracy of the calculated material parameters. Simple “peak finding methods”, where the machine data are used without diligent post processing, are not suitable for the calculation of complex ‘modulus or phase lag, because of potential errors. The paper gives a brief survey of various ‘mathematical methods for the determination of complex modulus £* and phase lag gcalcu- lated from various eyclic tests with sinusoidal loads. Furthermore, an enhanced approach is presented that allows the identification of differences in the evolution of material parameters ‘during loading and unloading phases of individual load eycles. 1 INTRODUCTION One of the most important and at the same time most sensitive subjects in materials research is the analysis of test data. Special caution is advisable in analyzing the outcome of eyclic tests, in which high frequencies are applied (sinusoidal, cosinusoidal or haversinusoidal loading at f> 1 Hz), as the material parameters derived from the test data analyzed (complex modu- lus £* and phase lag g in viscous materials) are extremely sensitive to the analysis method. used. These material parameters represent an important input in pavement design and finite element models used in projecting pavement performance (e.g. depth of rutting or fatigue resistance). They are usually determined by means of so-called “black-box” software sup- plied by machine manufacturers on the basis of simple “peak finding” methods (direct use of maximum/minimum points derived from raw test data). The use of “unfiltered raw data” includes the risk of producing inaccurate material parameters, as the “raw data”, because of the data capturing method used, come with a certain “filter noise”. Depending on the method. used for analysis, this noise' has a greater or lesser impact on complex modulus or phase lag 2. ANALYZING METHODS AND TECHNIQUES ‘The correct determination of material parameters of hot mixed asphalt HMA like ie. com- plex modulus £* or phase lag g obtained from dynamic or cyclic material tests became more and more important. The complex modulus £* is defined as the ratio of the maximum "Noise as a physical phenomenon, namely measurable Auetuations in current or voltage, was first described by Walter Schotthy in 1918, When these fluctuations are made audible through amplification, a typical noise is obtained after which the phenomenon was named. Meanwhile, the term “noise” is being used in a much more general sense (quoted from WIKIPEDIA) 357 applied stresses o divided by the maximum resulting strains ¢ of the tested material. The phase lag ¢ (phase angle) is obtained by determining the time lag between stress and strain signal and multipliying with the angular frequency of the conducted test. “The question as to whether itis necessary at all to use algorithms for determining dynamic ‘material parameters is easy to answer: any electrical signal, whether used for measuring or for machine control purposes, that is captured and recorded digitally by computer is subject, to a certain “noise” that is caused by various factors, One important factor in the produc- tion of filter noise is the way in which signals are transmitted by means of various physical media, currently primarily electric current. Before transmission, raw data are converted to ‘make them fit for transport via the medium selected. This may be done in analogue or digital ‘mode and involves sampling of data, their transformation into discrete values, and their final transport by means of electric pulses. The production and processing of a measuring signal can be described as follows: A sensor is used to pick up the signals that are often quite weak (in the range of a few millivolts). The signals are then amplified for further processing. The track/sample hold amplifier samples the amplified analogue signal at certain time intervals, A time-continuous signal is thus transformed into a time-diserete signal (see Figure 1), This signal can be processed by the digital signal processor. For a variety of reasons, the analogue- {o-digital transformation process also produces noise, which finally becomes part of the digi tised signal. Sources and causes of such noise are, for example, resistance noise (“Nyquist noise”, shot noise or thermal noise), tube hiss, contact noise at interfaces between electric conductors/semi-conductors or quantisation noise occurring in the conversion of analogue signals to digital (Wendemuth 2005). Figure 2 illustrates this noise once again using the example of a signal from force measure- ‘ments made during dynamic materials testing. The force signal measured over a time period Figure 1. Idealized scheme of converting an analog, measuring signal (efi) into digital one (right) by ‘means of signal sumpling (middle) [Wendemuth 2005), force FIN} » Se eres [res a i) ‘inet 3) Figure 2. Schematic representation of a sinusoidal force signal allicted by signal (electronic) noise (blac line) and the mathematically fitted, sinusoidal regression function (gray line). 358 is represented by three sine waves. As can easily be seen from the diagram, the force signal measured includes a large amount of noise and thus substantial fluctuations in amplitude, ‘These fluctuations in amplitude caused by filter noise ultimately have significant effects on the calculation of the dynamic material parameters. Therefore, a simple method has to be found for eliminating or reducing this noise without affecting or changing the signals in their time ‘domains. The methods most frequently used are briefly discussed in the following chapter. 2A State of the art Pellinen summarized in (Pellinen et al. 2003) the most commonly used post processing analy~ sis techniques. These techniques comprise two dilferent steps: first of all test machine “raw” data have to be filtered (smoothing) before complex modulus and phase angel could be deter- mined (parameter estimation algorithms). Pellinen compared various analysis methods to determine dynamic modulus (amplitudes of waveforms) and phaselag (time shift), which are in detail + 2step methods applied over a certain number of cycles, in which raw data from the test are first smoothed by means of suitable algorithms. Then the wave amplitudes and phase lags are determined by means of linear regression (A and B in Figure 3) + Localized multiple linear regression analyses with peak peaking (taking only one point = the maximum of all localized points of a have sine wave) for the determination of wave ‘and phase lags from wave maxima and minima in the raw data (“peak finding Cand D in Figure 3). Use of wave maxima and minima without previous smoothing of raw data (E and F in Figure 3). + Sinusoidal regression over a certain number of cycles (G in Figure 3). Various combinations of these analysis methods. ‘The results of a statistical analysis of the material parameters E* and g determined by ‘means of these algorithms in a comparative test are represented graphically in Figure 3. In these Black diagrams, phase lag is plotted against the dynamic modulus £ In summary, (Pellinen et al. 2003) arrived at the following conclusions: * The analysis method used has a material influence on the outcome of the calculated dynamic material parameters, with the impact on the magnitude of dynamic modulus £* being less than on phase lag g. » ow ‘roam one ‘Balled pa sing = adr seam one ‘ssa dea siding EAF> NDcainemsting& vel nason cong Seu ges a ve 0hstbaaa 18 29 28 39 3 49 18-23-28 33 ae 43 eHenatots Log Pa Log EuPe Phase Angie. Oot FFigure 3. Schematic representation of a sinusoidal force signal aMicted by signal (electron) noise (black line) and the mathematically fitted, sinusoidal regression function (gry line) 359 Use of wave maxima and minima without previous smoothing of raw data results in the least fit between data measured and material parameters, ‘ Smoothing of the raw data obtained from the test apparatus followed by multiple linear regression analyses (e.g. using polynoms or power functions) over a certain range of cycles provides the best fit. + Sinusoidal or cosinusoidal regression over a certain number of cycles provides the best fit of all regression analyses compared, ‘ Incyclic tests in which plastic deformation occurs, the use of additional function terms in the regression functions is indispensable, however. 22. Enhanced regression approach As previously mentioned, sine and cosine functions are a suitable approach to regression analysis of data obtained from cyclic testing. It must be pointed out, however, that in cer~ tain tests, in which the material changes due to cracking or permanent deformation (such as fatigue testing or rutting tests), additional function terms are necessary in order to obtain good fit between the data measured and the mathematical regression functions, The most simple approach is the sine function with a constant term PMlaystz,45) =a, +0 Sine: f 1-403) a where (9 [stands forthe regression function ofthe analyzed signal, ya, and a [-]arethe regression coeiients, fH] stands forthe test frequency and [forthe time ‘The coe tient a represents the offset, a, the amplitude and a the phase lag from a sinusoidal signal Using this simple basis function (hereinafter ealled “simple sine function”), the data recorded (as shown, fr example, in Figure 4 for vil stesso, and axial strain, are t be “approximated” as closely as possible. The quality ofthis fits measured by the coefficient of determination? Fe As clearly illustrated by Figure 4, however, the simple sine function often fails to provide a good repression to the points measured, pariularly when one of the signals hails to produce a constant line (asin the example of the iaxial cyclic compression test TCCT, see Figure 4, and the am is to analyze several cycles eg” 3 cycles sch a8 in Figure 4) The differences between the axial stain data measured and the line ited by means of the simple sine function are too big (see Figure 4), primarily because this function does not permit any “ting” to take into account the non-linear natu of the permanent strains Especially in triaxial cyclic compression test, n which cyclical load- Ing s performed to produce permanent deformation, the resulting axial or radial strains to longer oscillate around a horizontal baseline but along a non-linear deformation “To determine the regression parameters of sinusoidal deformation reactions not running in parallel with the ani a simple sine function with an additional linear term a, ) is there fore used asa next step Peay, ds.a4) =0) +0 Sina fas) at ° ‘The result of this new regression function is illustrated in Figure 5. In contrast to the simple sine function, the approximated sine waves of the axial deformations are no longer “The coefficient of determination R? ofthis non-linear regression is calculated in the following manner according to (Sachs 1989): VQ, Wey S{E x) wim arene saat squares of distance of the measured values y from the regression values with Q, being the covariance 300 [All ree [Ma Figure 4. Triaxaleycic compression test (TCT) data from axial loading ¢, and the resulting strain 4, and the mathematical approximation function—simple sine unction—plotied against time (left) and as stress-strain diagram (right) (Kappl 2007). Figure 5. Triaxial cyclic compression test (TCCT) data from axial loading o,, and the resulting strain 4, and the mathematical approximation funetion~sine function + linar term plotted against time (let y-axis with stresses, eft ais, and strain right axis) and as stress-strain diagram (right; y-axis with strains on right axis) (Kappl 2007) horizontal but slanting downwards linearly, reflecting more accurately the actual positions of | the points measured (see red lines in Figure 5). ‘This function appears more appropriate for TCT tests than the simple sine function, A closer look at the stress-strain diagram in Figure 5 (right half) reveals, however, that this sine function with a linear term does not fit sufficiently the points recorded, particularly in the vertex points of the ellipsis. The shape of the ellipsis in the stress-strain diagram that is, typical of an “undeformed” sine wave (= shape of ideal mathematical sine wave) does not ‘exactly fit the measuring points (see marking in Figure 5) FFor this reason, the sine function with the linear term was extended by adding a second sine term having double the frequency (47/) of the first sine portion (22). This “extended. sine function” is discussed in more detail below: This function reads: It...) = 04 +aysin ft Hay) tag + asin( Af +a) 6 offset (displacement of ordinate) of the base wave, a, = amplitude of base wave, 44, = phase angle, a,= linear term, a, = amplitude of the harmonic wave, a, = phase angle of harmonic wave, f= frequency in [Hz] and /, the time /at which a data point was recorded. ‘As Figure 6 shows, the data measured can be modelled with much greater accuracy when the extended sine function is used. Better fit is obvious in particular at the turning points of | the sine waves (see marking in right-hand diagram in Figure 6) 36 Teal a a on ac ater Oar ° Atal stain 6 Figure 6. Differences between the data measured and the values obtained from regression analysis using the extended sine function (double sine function and linear term) for cycles (Kappl 2007). z & i i g g i 5 4 i @ & Figure 7. Comparison of coefficient of determination of the three regression funetions for axial deformation after the First 10000 load eyeles on a simple logarithmic seale (left diagram = overview: right diagram zoomed in between 0.95 and 1.00 on y-axis) (Kapp 2007). 23 Comparative assessment of different regression functions ‘on the base of triaxial eyclic compression test data Figure 7 compares the “quality of fit” of the three regression functions for axial deformation using the TCCT test as an example. As can be seen, the coefficient of determination R° of the simple sine function according to equation (1) is below 0.95 during the first 100 load cycles During the first 10 load cycles, R? is even lower than 0.5, Its only later, al the beginning of the secondary creep phase, that R? exceeds the level of 0.95. The sine function with the linear term according to equation (2), on the other hand, has very good fit initially, during the first load cycles, but then loses some of its quality (R? around 0.98). The regression function with the dual sine according to (3) boasts a very high and constant R? throughout the entire test Here, R? was on the average 0.995 or higher 24 Differences in material parameters E* and phase lag in triaxial eyelie tests based on the enhanced regression function Finally, further special characteristics in material response are discussed by reference to the triaxial eyclic tests (TCCT). As Figure 4 shows, the sine function with the linear term (2) does rot produce a particularly good fit to the data measured (shown in left diagram of Figure 8 as black circles). Differences are still identifiable in the areas around the maximum and mini ‘mum (marked in Figure 8 by red arrows). 362 In addition, as can be seen from the diagram, this regression function provides relatively good fit to the mathematical sine function when the test specimen is loaded. On loading, the axial deformation of the test specimen closely follows the projected loading curve of the hydraulic piston. On unloading, the fit is no longer satisfactory, however. Here, the test specimen has to provide the energy for “re-deformation” (recovery) alone. As the material of the test specimen (asphalt) is already highly viscous at the TCT test temperatures (as arule, T= 40 and 50°C), the recovery takes place at a slower pace. This is clearly apparent from Figure 8: the data measured (shown as black circles) show a certain time lag (graphi- cally displaced to the left) compared with the mathematical sine regression function (broken red line). ‘These differences between loading and unloading are even more visible in the left diagram of Figure 8. This diagram plots axial stresses against axial strains. The individual sine waves are shown as hystereses. It can be seen that when the test specimen is loaded and the sen- sors are exposed to responses from both the apparatus and the material, the hystereses are “dented” (ef arrow in the loading portion of the wave in Figure 8). All points measured are Within the “sine function hysteresis”. On unloading, on the other hand, when the sensors. measure only the response of the material, the hysteresis is “drawn out”. In addition, it ean be seen from Figure 8, that the points measured in the vortexes of the hysteresis are not located precisely on the y.,curve of the sine function with the linear term. In the top right of | the hystereses in Figure § (see marking), the data measured are scattered more widely than ‘on the opposite side (bottom left of diagram). This is also clearly noticeable (see markings “wide” and “small”) When using the extended sine function according to equation (3), approximation to the data measured is much better, both around the vertex and during loading and unloading phases. The changes in material response during loading and unloading phases described above ultimately also have an impact on the dynamic material parameters £* and g. Thus, for example, Figure 11 shows complex modulus £*, E,, E, (real and imaginary parts of dynami modulus £*) and the axial phase angle g,, of stone mastic asphalt SMA 11 PmB 45/80-65, SI, G1 containing diabase as aggregate. Strikingly, the dynamic material parameters change in these diagrams with increasing number of load cycles. This phenomenon can be illustrated. very clearly, especially for the phase angle g,. Figure 11 shows £* and g,, plotted against the ‘number of load cycles, with the phase angles determined, on the one hand, between the stress and strain peaks (named gy, ,) and between the lows (named 9, yy.) These key points are shown graphically in Figure 9 and Figure 10 for two different siress and strain waves, It is clearly evident that the phase angle ¢,, ,,., i8 produced by the imposition of loads—in the subject case in TCT tests by axial loading applied by the loading piston—while the phase angle 9, gy, Feflects the response of the material of the asphalt test specimen upon unloading. At this tiie, the sensor is not exposed to any impacts from the apparatus This ana oP oe a0 en go ae ta em rina 1 Figure 8. Simple sine function (sine term plus linear term: gray lie) versus extended sine function (second sine term with double frequency and linear term: blue line) derived from axial test specimen deformations (Kappl 2007). 363 angle thus represents the “unadulterated” response of the material during unloading of the test specimen, To illustrate this phenomenon, Figure 9 shows the phase lag angle between the “Max- Max" and “Min-Min” points at the beginning of the test (at the 40th load cycle). One ean clearly see that the axial phase angle @, is smaller than 9, ,,., between the “Min- Min’ points In Figure 10, the same material parameters are shown graphically at the end of ‘Axial Stress (N/mm) Time t (1/100 s) o 1 2 30 40 50 60 grace rx tres (regression) a i > ea eh ‘ Max + 7 LOADING UNLOADING — cru raow 195-79 785-78 175 Axial strain (%) 3 8.05 Figure9. Axial strains and stresses at maximum and minimum wave pointsin the loading and unload ing phases at the beginning of the test (40th load eyele)(Kappl 2007). Axial stress (N/mm®) 4 4 76 Time t (1/100 s) 3 6 300 7610 7620 7630 7640 7650 7660 six [| Pes Sis ar? Sigs (data) he / f [Tho | wdc aan f/ MIN TE ons as “1365 13.6 -1355 135 13.45 -134 13.35 Axial strain (%) Figure 10. Axial strains and stresses at maximum and minimum wave points in the loading and unloading phases at the end of the test (25000th load eyele) (Kappl 2007), 364 S Bg28Q ata RCO wi = =e ee Figure 11. Dynamic modulus Ete axial and radial phase angles between vane maxim and minima (let) and the real and imaginary parts £, and E, for SMA II Pm 45/80-64, SL, GI diabase (simple logarthiie seale on abscissa) (Kapp 2007, the test, after 25000 load eycles. Here one can clearly see that the axial phase angle @,, hhas become much smaller after exposure to the loading piston, ‘These effects may be explained by changes in the structure or displacements of particles in the asphalt test specimen, which may lead to the formation of “mineral skeleton pillars” in the cylindrical test specimen. As is known from soil engineering, aggregates (rock), when laterally constrained as in the case of triaxial tests, react elastically to impacts from outside (also see (Huschek 1977). This is reflected by the reduction in the phase angle towards “more clastics values” during TCCT tests, With dynamic modulus £*, on the other hand, as shown in Figure 11, there is no big dif ference between the dynamic modulus £,,..,,, (unloading) derived from the peak values and E*,,, y,, (loading) derived from the minimum points of the waves. The two material parameters are almost the same. It isstriking that as the loading time (number of load cycles) increases, the dynamic modu- lu rises up to a limit value. This means that a constant modulus £” is reached after an initial phase of stresses and strains (about 100 load cycles). This would be in line with reports by Francken (eg. (Francken 1977)), who uses a constant “plasticity modulus” for his permanent deformation model (BRRC model. 3 CONCLUSIONS: In summary, the following phenomena were observed during triaxial cyclic testing in the ‘dynamic material parameters of the test specimen calculated by means of the extended sine function: «In the primary creep phase (initial phase) of the TCCT tests, the dynamic material param- eters change significantly: During the first 100 to 200 load cycles, the dynamic modulus £* and its elastic part £, rise degressively as loading time increases, up to an almost constant limit value ‘The viscous part £, increases slightly during the first loading cycles but subsequently. decreases again to a constant limit value. Here, it must be diflerentiated between the vis- cous modulus determined between the maximum (unloading: reaction of material) and ‘minimum (during the loading phase) data points: The values found in the unloading phase Egon, A Almost constant and linear, while the E, Values produced by loading ini- tially increase in a slightly degressive manner and then (alter the initial phase) fall degres- sively toa limit value. ‘The axial phase angles decrease disproportionately in the primary creep phase and then remain almost constant. A difference exists also in the phase angles between stresses and strains between @, (during unloading) and 9, yg. (during loading), as both 365 parameters develop differently: The @,,.yu yy quickly fall from very high levels (high vis- cosity) to low values (high elasticity), while’... (in the unloading phases) remain constant over the loading time, with smaller phase angles ‘These phenomena had not been detected and analysed previously by the analysis software used, which did not contain the regression algorithms required for this purpose. Its, how- ever, specifically this “evolution” of the dynamic material parameters that may have a eriti- cal impact on the outcome of projections or model calculations. It should be studied in more detail in the future, ACKNOWLEDGEMENTS ‘The authors’ sincere thanks go to those who have made the evaluation of this mathemati- cal algorithm possible: Mr. Walter Fusseis, who formulated mathematical functions for this algorithm. Cordial thanks are also due to Mr. Michael Wistuba and Mr. Markus Spiezl, who in numerous talks contributed their ideas and constructive criticism and were significantly involved in the development of this analytical method, REFERENCES [EN 12697-25 (2004), Asphalt—Test methods for hot mix asphalt Part 25: Cyelie compression tes Comité européen de normalisation CEN, Brussels, 2004-12-01 EN 12697-26 2005). Asphalt-—Test methods for hot mix asphalt--Part 26: Stiffness: Comité européen ‘de normalisation CEN, Brussels, 2004-10-01 Francken Louis (1977): Permanent deformation law of bitwninous road mixes in repeated triaxial com= pression. Proceedings of the 4th international conference on structural design of asphalt pavements, Vol. 1, Ann Arbor, 1977, Huschek’S. (1979): Imerpretation der Ergebnisse aus Kriechversuchen mit Asphaltprobekirpern. Collo- ‘quium 77 zur platischen Verformbarket von Asphaltmischungen, Mittelung 37 der ETH Zirich, November 1977 Kappl Karl (2007): Bewertung und Modellierung des Verformungsverhaltens von Asphalten mit Hilfe von Erklischen Triaxialprifingen. Dissertation am Insitt fr StraBenbau und Straeneshaltung, TU Wien, Wien, 2007 Pellinen Terri & Crockford B, (2003): Comparison of analysis tchnigues 10 obtain modulus and phase ‘angle from sinusoidal test data. 6th RILEM Symposium PTEBM'03, Zurich, Conference Proceed- ings RILEM. 2008, Sachs Ludwig (1989): Angewandte Staitk. Berlin (a... Springer, 1989, Wendemuth A, Andelice E & Barth S. (2005): Grundlagen der digitalen Signalverarbeitung. Berlin (ua), Springer, 2008. 366,

You might also like