moisture damage modeling in lime and chemically modified

Research ArticleMoisture Damage Modeling in Lime andChemically Modified Asphalt at Nanolevel UsingEnsemble Computational Intelligence

M R Hassan 1 A Al Mamun 1 M I Hossain 1 andM Arifuzzaman 2

1King Fahd University of Petroleum ampMinerals Dhahran Saudi Arabia2University of Bahrain Zallaq Bahrain

Correspondence should be addressed to M R Hassan mrhassankfupmedusa

Received 17 August 2017 Revised 11 January 2018 Accepted 31 January 2018 Published 18 April 2018

Academic Editor Elio Masciari

Copyright copy 2018 M R Hassan et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

This paper measures the adhesioncohesion force among asphalt molecules at nanoscale level using an Atomic Force Microscopy(AFM) and models the moisture damage by applying state-of-the-art Computational Intelligence (CI) techniques (eg artificialneural network (ANN) support vector regression (SVR) and an Adaptive Neuro Fuzzy Inference System (ANFIS)) Variouscombinations of lime and chemicals as well as dry and wet environments are used to produce different asphalt samples Theparameters that were varied to generate different asphalt samples and measure the corresponding adhesioncohesion forces arepercentage of antistripping agents (eg Lime and Unichem) AFM tips 119870 values and AFM tip types The CI methods are trainedto model the adhesioncohesion forces given the variation in values of the above parameters To achieve enhanced performancethe statistical methods such as average weighted average and regression of the outputs generated by the CI techniques are usedThe experimental results show that of the three individual CI methods ANN can model moisture damage to lime- and chemicallymodified asphalt better than the other two CI techniques for both wet and dry conditions Moreover the ensemble of CI along withstatistical measurement provides better accuracy than any of the individual CI techniques

1 Introduction

Moisture plays a significant role in asphalt pavement failureMany developed countries have invested billions of dollars onroads and pavements to overcome all kinds of malfunction-ing An asphalt pavement may have different types of damagesuch as rutting fatigue creep revealing and moisture dam-ageOf these the exact nature ofmoisture damage phenomenaremains ambiguous as traditionalmacroscale andmicroscaletests cannot explain the damage thoroughly Although in thelast few decades researchers have made significant attemptsto explore moisture damage phenomena comprehensiveunderstanding of the reasons for moisture damage is stilllacking In this paper the authors attempt to characterizemoisture damage in lime- and chemically modified asphaltbinder at nanoscale level using Atomic Force Microscopy(AFM) and apply a number of Computational Intelligence(CI) methods to analyze and model the damage The authors

also propose an ensemble of CI along with some statisticalmethods to better model asphalt damage

Different types of methods have been proposed to iden-tify moisture damage in asphalt binder and pavement atmacrolevel For example for the AASHTO T-283 [1] spec-ification experiments are done in dry and wet conditionsStatic Immersion Tests (ASTMD 1664 AASHTOT-182) BoilTests (ASTM D 3625) and Indirect Tensile Tests (AASHTOT-283 and ASTM D 4867) are other examples where testsare done at macrolevel [2] The measurement of the surfaceforce of asphalt binder at nanoscale unit using AFM has beenlimited to asphalt morphology only The problems of designconstruction and use of functional structures have beenextensively analyzed at nanoscale unit in recent decade [3]Despite the fact that asphaltic materials such as asphalt aremainly used on a large scale as well as in large quantities forroad construction the mechanical (macroscopic) behavior

HindawiComputational Intelligence and NeuroscienceVolume 2018 Article ID 7525789 9 pageshttpsdoiorg10115520187525789

2 Computational Intelligence and Neuroscience

of these materials still depends to a large degree on thephysical properties in terms ofmicroscale and nanoscale level[4] Although engineers material producers and researchershave explored the potential for many years the analysis ofasphalt binder properties at nanoscale unit has been found tobe limited in utilizations [5] Since all the destruction (egadhesioncohesion failures) initiates from the nanoscale tothe microscale the current focus is now being placed onnanotechnological (the technology that analyzes the prop-erties in terms of nanoscale measurement) applications inasphalt The derived information can be useful in multiscalemodeling of moisture damage in asphalt To the knowledgeof the authors only a few studies by Berquand and Ohler [6](accessed January 1 2018) Tarefder and Zaman [7] Tarefderand Ahsan [8] Hassan [9] Arifuzzaman and Hassan [10]Arifuzzaman [11] and Arifuzzaman et al [12] have testednanoscale moisture damage using functionalized AFM tipsIn the study by Arifuzzaman andHassan [10] the researchersmodeled the moisture damage in polymer-modified asphaltwhile Hassan [9] modeled the moisture damage in CarbonNanotube (CNT) modified asphalt Tarefder and Zaman [7]studied the effect of polymer alteration due to moisturedamage in asphalt using AFM without applying any artifi-cial intelligence to detect the damaging behavior HoweverTarefder andAhsan [8] developed anANNmodel to quantifyadhesion from the AFM data Arifuzzaman [11] developedANN model to determine the moisture damage behavior ofthe CNT modified asphalt binder None of these previousstudies explained or predicted moisture damage in lime- andchemically modified asphalt binder In this paper the authorsdevelop an ensemble CI-statistical technique to model mois-ture damage in lime- and chemically modified asphalt Inthis process first a number of well-known CI techniques areapplied namely Artificial Neural Networks (ANNs) SupportVector Regression (SVR) and an Adaptive Neuro FuzzyInference System (ANFIS) Initially functionalized tips wereused to measure the weak intermolecular forces (eg adhe-sioncohesion) in polymer-modified asphalt binder systemsThe level of moisture damage was measured at nano-Newtonscale for varying measurements of lime antistripping agentsUnichem (UC) and spring constant values (119870 values) andvarying the functionalized AFM tips carboxyl (-COOH)ammin (-NH3) methyl (-CH3) among hydroxyl (-OH)and silicon nitride (Si3N4) Since no practical attempt hasbeen made to date to match the functional groups that existin asphalt binder this paper attempts to analyze the effectof these functional groups on moisture damage in asphaltusing the above-mentioned CI techniques To improve theperformance of the CI techniques further a number ofstatistical measurements are applied to aggregate the outputsof each of the CI techniquesTherefore the aggregated valuesare considered as predicted values

The rest of this paper is structured as follows ldquoBack-ground and Literature Reviewrdquo section provides an overviewof AFM ANN SVR and ANFIS and a literature review ofmoisture damage modeling using these techniques ldquoEnsem-ble of CI and statistical methodsrdquo section describes theensemble model ldquoMaterials Description and Experimenta-tionrdquo section presents the description of the dataset the

experimental setup the experimental results and the anal-ysis Finally ldquoConclusionrdquo section concludes the paper

2 Background and Literature Review

In this section AFM and the three CI techniques namelyANN SVR and ANFIS are briefly described At the endof each subsection a discussion about how and where thesemethods have been applied to model asphalt damage isprovided

21 Theory of Atomic Force Microscopy (AFM) An AFM is atool to investigate the topographical and surface properties ofseveral engineering materials It is designed to evaluate vari-ous local properties such as height friction and magnetismIt scans and then calculates the local properties continuouslyto produce an asphalt binder bonding image All the AFMprobes has a very sharp tip with usual values 3 to 6120583m tallpyramid with a 15 to 40 nm end radius AFM scan measuresthe deflections of the cantilever which the lever performsthrough the reaction of a laser beam of the cantilever [13]The tip is positioned with high resolution by piezoelectricceramic that expands or contracts due to the voltage gradientand helps in positioning 3D devices with high accuracy[14]

211 AFM Testing AFM is used by the chemical engineeringresearch groups at Harvard University to evaluate adhesionand cohesion forces The use of AFM in the chemicalprocess is known as chemical force microscopy [15] AFMwas also used by Beach et al [16] to find drag off pro-cess between hexadecanethiol monolayers that automaticallypulled together the silicon nitride tip as cantilever and siliconwafer Okabe et al [17] used hydrophobic -CH3 and -COOHterminating alkane thiols functionally modified to measureadhesion forces and to image the samples Du et al [18]used AFM to find thin films of polymer-generated that yieldsstrength and modulus of elasticity Researchers have alsoused nonfunctionalized AFM tips to measure the rigidity ofasphalt binder [19] Pauli et al [20] measured the surfaceenergy of asphalt binder using AFM Masson et al [19] alsoused AFM to execute imaging of asphalt and connectedthe images with different chemically analyzed ingredients ofbitumen

22 Artificial Neural Network (ANN) Artificial Neural Net-works (ANNs) have shown their versatile application tooptimize and forecast immensely complex uncertain orpoorly understood situations by simulating the real fieldscenario for the last three decades in different disciplines ofcivil engineering Following the same trend the applicationof ANN in transportation engineering especially in pavementengineering has also received attention (eg Ceylan et al [21]Flood and Ian [22])

An ANN resembles the biological brain by being likea parallel distributed processor with high computationalpower Hecht-Nielsen and Robert [23] and interconnectedprocessing units or neuronsThe parallel processor is capableof solving nonlinear relations by storing and providing data

Computational Intelligence and Neuroscience 3

for use and the interconnection provides computationalpower for simultaneous data processing in the ANN ANNsare typically organized in layers and a typical ANN has threelayers an input layer hidden layer(s) and an output layerThe numbers of neurons at each layer are chosen based onthe system dynamics and the expected accuracyThe strengthof interconnection is measured by weighting the vectorsof the ANN [24] There is no strict regulation to emulatethe structure of an ANN However researchers have madedifferent conclusions addressing the maximum or minimumnumber of nodes in the hidden layers

23 Support Vector Regression (SVR) Support Vector Regres-sion (SVR) is an extension of well-known machine learningtool named Support Vector Machine (SVM) which wasintroduced by Vapnik [25] SVR selects the set of pointsin each class (support vectors) that are the nearest to theother class and through them computes a hyperplane whichis called the maximum-margin hyperplane and makes SVRrobust Using nonlinearmapping data aremapped on a high-dimensional space and linear regression is then applied to thetransformed feature space [26]

A study byChen and Su [27] has shown that the nonlinearmapping relationship between some fundamental variablesand the structural response was developed using SVM andMonte Carlo (MC) theory to calculate the failure probabilityand reliability of pavement The experimental results weresignificantly precise efficient and in good agreement withthe field conditions Yan et al [28] produced a forecast ofthe low number of dense asphalt-aggregate mixtures usingSVM and achieved good performance compared to multiple-least-squares-regression Ke-zhen et al [29] used SVM as thebasis of an insensitive loss function to measure pavementserviceability ratings of flexible pavement and found betterperformance compared to the AASHO model and the ANNmodel Soltani et al [30] found that a SVM algorithm wasthemost effective in forecasting the stiffness of a polyethyleneterephthalate-modified asphalt mixture with a better Corre-lation Coefficient (CC)

SVM was used to predict the mechanical behavior ofhot-mix asphalt using a comprehensive dataset and wasfound to be comparable to the ANN and better than amultivariate regression-basedmodel by [31] In another studyGopalakrishnan and Kim [32] showed the effectiveness ofSVM on skilled offline pavement backcalculation and foundcomparable performance toMultilayer Perceptrons (MLP) inmost cases and better in some special situations Maalouf etal [33] used SVR to predict the resilient modulus (MR) ofhot-mix asphalt and found superior performance to the leastsquares (LS) with reducedmean-squared error and improvedCC

24 Adaptive Neuro Fuzzy Inference System (ANFIS) Jang[34] developed the Adaptive Neuro Fuzzy Inference Sys-tem (ANFIS) which serves as a basis for constructing if-then rules and Fuzzy Inference Systems (FIS) ANFIS com-bines FIS with the concept of ANN The FIS displays theprevious information in a set of constraints by trimmingdown the input-output space that needs to be optimized A

backpropagation algorithm is applied to tune the parame-ters of the FIS in order to achieve a better performanceHere FIS reforms the structure of the adaptive networksDifferent methods like backward spreading least squaresestimation Kalman filter or hybrid learning algorithms incombination with numerous mathematical approaches canbe implemented to optimize the ANFIS parameters

A number of studies for example [35ndash41] have usedANFIS used to predict and control different engineering sys-tems Likewise there have been several successful implemen-tation of ANFIS in solving problems of pavement engineering[12 42] ANFIS and ANFIS-related hybrid learning algo-rithms have also been used to forecast the complex modulusof ethylene vinyl acetate-modified binder Yilmaz et al [43]Ozgan et al [44] predicted the stiffness modulus of asphaltcore samples using ANFIS Shafabakhsh and Tanakizadeh[45] scrutinized the changes in resalient modulus (MR)of asphalt mixtures under different loading patterns usingANFIS

3 Ensemble of CI and Statistical Methods

In an attempt to further enhance the prediction of bindersrsquoadhesioncohesion force an ensemble of ANN SVR andANFIS with a number of statistical methods is developedIn the ensemble method the predicted values producedby ANN SVR and ANFIS for each input sample in thetraining dataset are aggregated using the statistical methodsto generate a final prediction In this regard the followingthree statistical methods are applied to compute the finalpredicted values

(1) Average(2) Weighted average(3) Linear function

31 Aggregate by Average In this approach of the ensemblemethod the authors compute the average of the outputs ofthe above-mentioned CI techniques and consider the averagevalue as the ultimate prediction of adhesioncohesion forcein nano-Newton scale The final prediction is computedfollowing the equation

119910avg = 119910ann + 119910svr + 119910anfis3 (1)

where 119910ann 119910svr and 119910anfis are the values predicted by ANNSVR and ANFIS for a given input sample (ie input sample= ⟨tiptype lime unichem⟩)32 Aggregate by Weighted Average A weight is assignedto each of the predicted values produced by individualCI techniques The Normalized Root Mean Square Error(NRMSE) value for the training dataset for each individualCI method is considered as the weight for the respective CImethod Therefore there are three weights for the three CImethods Once the weights are obtained the following equa-tion is used to generate the prediction using the predicted


outputs of each of the CI methods for a given input datasample

119910119908avg = 119908ann times 119910ann + 119908svr times 119910svr + 119908anfis times 119910anfis119908ann + 119908svr + 119908anfis (2)

where 119910ann 119910svr and 119910anfis refer to (1) and 119908ann 119908svr and119908anfis are the NRMSE values for the training dataset of therespective CI methods The NRMSE is computed followingthe equation below

NRMSE = 1 minus10038171003817100381710038171003817119910act minus 119910pred1003817100381710038171003817100381710038171003817100381710038171003817119910act minus 119910pred10038171003817100381710038171003817

(3)

where 119910act is the desired value of binder force that needs tobe predicted and 119910pred is the predicted value produced bythe respective CI method given an input sample ⟨tiptypelime unichem⟩33 Aggregate by Linear Function In this approach of ensem-ble method the set of predicted values using the trainingdataset are mapped to the desired output values using a linearfunction119891(sdot)The linear function that is used in the ensemblemethod is as follows

997888997888rarr119884act =3sum119894=1

119887119894 times997888rarr119884119894 (4)

In the above equation997888rarr119884119894 represents the list of predicted

values using the training dataset and the 119894th CI method 119887119894represents the coefficients for the 119894th CI method (here the 1stCI method is considered as ANN the 2nd CI method is SVRand the 3rd is ANFIS)

To compute the values of 119887119894 a least square estimation isapplied In this regard if 119884pred refers to the all the predictedvalues of each of the CI methods for the training dataset (iecolumn one of 119884pred would consists of the predicted valuesusing ANN for each of the data samples in training datasetcolumn two would consists of the same using SVR and thethird column would contain those using ANFIS) and 997888997888rarr119884act isthe list of actual values which are required to be predicted forthe training data samples the value of

997888rarr119887 (where 997888rarr119887 = ⟨1198871 11988721198873⟩) is computed using the following equation997888rarr119887 = (119884119879pred119884pred)minus1 997888997888rarr119884act (5)

As soon as the value of 119887119894 is available for any test datasample the prediction is computed as follows

119910linear = 1198871 times 119910ann + 1198872 times 119910svr + 1198873 times 119910anfis (6)

where 119910ann 119910svr and 119910anfis refer to (1)4 Materials Description and Experimentation

41 Materials In this study the aim is to model the moisturedamage in lime- and chemically modified asphalt Thereforethe authors have used the following materials to mix with theasphalt binder lime andUnichemDetails of the additives areprovided below

411 Lime Lime has been used in the hot-mix asphaltbitumen pavement industry for more than 25 years Itcontributes to both the rheological andmechanical propertiesin asphalt mixtures It also improves moisture sensitivityresistance as well as fracture toughness strength and reducesthe rate of aging due to oxidization in asphalt binders In thisstudy the authors added 05 10 and 15 lime by weightof asphalt to prepare the AFM samples for testing

412 Unichem A dark brown colored antistripping agentUnichem is liquid and has a viscosity of 236mPasdots at 378∘Ca density of 831 lbgal and an ash point of 100∘C Heatstable ingredients in Unichem help it to perform efficientlyin verifying heating condition Unichem has a good capacityto be absorbed to increase the wettability of the aggregatesurface and this plays a significant role in coating the aggre-gate easily As an antistripping agent unlike lime Unichemdoes not modify asphalt chemically It may be effective inconcentrations ranging from 025 to 15 by weight of theasphalt binder Therefore in the experiments the authorsused Unichem concentrations within the range of 025 to15 with respect to the weight of the asphalt binder

413 AFM Testing and Tips Functionalization Functional-ization of AFM tips was carried out to simplify the asphaltbinder chemistry through various groupings like -COOH -NH3 -CH3 and -OH Asphalt binder has hydrogen atom-saturated long carbon chains It has single C-C and C-Hbonds with evenly distributed electrons which are slightlyinclined to move around According to Little and Jones[46] these long carbon chains saturated with hydrogen areconsidered nonpolar in nature The Van der Waals forceswhich play a significant role due to their additive naturein these large molecules facilitate the interaction of thesenonpolar molecules Asphalt binder consists of some heteroatoms such as sulfur (S) nitrogen (N) and oxygen (O)which form different functional groups in the asphalt systemA group of atoms with a particular arrangement is called afunctional group This kind of group governs a specific typeof characteristics in the entire molecule The composition ofa specific group dictates the names of the functional groupsFor example -COOH is a carboxyl functional group in theasphalt molecule Typically the most common functionalgroups in asphalt binder are -COOH -NH3 -CH3 and -OHTherefore these functional groups were implemented tothe AFM tips in this investigation A total of four differentAFM tips are used in the current research They are differentfrom the different chemical view point and are modifiedwith asphalt functionals such as COOH -NH3 CH3 andOH All the types were classified as either hydrophilice orhydrophobic The COOH and OH are hydrophilic tips andthe CH3 and -NH3 are the hydrophobic tips Probing thefilm of asphalt binder surface with the help of functionalizedtips facilitates and measures cohesion forces between the twoasphalt molecules The functionalization task was done withthe help of a professional company (Novascan TechnologiesAmes IA USA)The -COOH -NH3 -CH3 and -OH groupswhich are themajor part of asphalt are used to functionalizedthe tips


Figure 1 The AFMmachine used in the experiments

In AFM testing the sharp tip of the free end of itscantilever scrutinizes the surface of asphalt binder This tiphas a beam bounce cantilever with a length of 125120583m with90 kHz frequency and a spring constant (119896) value of about3Nm Here silicon nitride (Si3N4) AFM tips were func-tionalized with the above-mentioned functional groups fromNovascan Technologies (USA) The tips were adjusted withthe self-accumulation of a thin monolayer film onto AFMtips followed by immersion of the AFM tips in a solution ofthiol Vaidya and Chaudhury [47]Thiol is connected with theAFM tip surface and the competent functional group by itstwo ends The attraction or repulsion generated between thetips and the binder surface results in bending or deflection ofthe cantileverThis deflection of the cantilever is measured bythe built-in laser beam reflection facilities in AFM Once thisdeflection is measured it is converted to acting attraction orrepulsion on the cantilever tips through multiplying by thespring constant It is a consequence of the space linking theAFM tip and the sample surface Figure 1 shows the AFMthat has been used to collect data in this paper The mainchallenge in asphalt testing with AFM tips is the stickinessof asphalt binders The sample is usually very soft at elevatedtemperature which causes the tips to be contaminated andeven some times damaged by the asphalt Thereby the wholetesting was conducted under noncontact mode in order tobe keep the AFM tips safer from the asphalt contaminationMoreover all the samples were put in freezer before testingso that it may not have a softer surface which is undesirableduring AFM testing

42 Asphalt Sample Preparation Using Binders The asphaltsamples were prepared in the asphalt concrete laboratoryThesamples were 10 times 10mm in length by width and thicknessof about 1mm Such dimensions are ideal for AFM testingThe idea behind the samples thickness came from the actualasphalt coating when the aggregates are coated by the binderduring the mixing time The samples were deposited in theglass substrate which later onwere inserted on the scanner forimaging as well as finding adhesion and cohesion forces Allthe fresh samples went through AASHTO T-283 [1] standardprocedures for the moisture damage simulation

43 Dataset As discussed in Section 41 the asphalt binderrsquosadhesioncohesion force varies with the variation of lime and

Table 1 Adhesioncohesion force summary for varying lime andchemical binders in asphalts

Parameter Dry samples Wet samplesNano-Newton Nano-Newton

Maximum 44437 67917Minimum 3102 11527Average 23747 35535Standard deviation 11624 13082Kurtosis minus109 minus080Skewness minus033 052

Unichembinders In this study datawere collected by varyingthe lime and Unichem as asphalt binder and then used AFMto measure the adhesioncohesion force at nanoscale levelSince this study aims to model the moisture damage in lime-and chemically modified asphalt binder the inputs to theCI techniques used in this study are AFM tip types andthe percentages of lime and Unichem while the predictedoutputs are the adhesioncohesion forces Table 1 summarizesthe adhesioncohesion forcesmeasured for the varying valuesof lime and Unichem as well as the AFM tip types for twodifferent conditions wet and dry

44 Moisture Damage Modeling Results In this study oneof the aims is to apply CI techniques in modeling asphaltdamage due to moisture As stated in ldquoBackground andLiterature Reviewrdquo section three well-known CI methodsANN SVR andANFIS have been used Each of themethodswas used for both wet and dry asphalt samples Manyresearchers (Mas and Ahlfeld [48] Hota et al [49]) have used80 of the data for training their proposed computationalmodel and the remaining 20 of data for testing the efficacyof the model that is these 20 data are kept unknown to thetrained model The authors of this study followed the sameapproach in the experiment and hence used 80 of the data(selected randomly) for training the respective CI methodsand the remaining 20 data for testing the trained methodsTables 2ndash5 report the experimental results in terms of thefollowing performance metrics

(1) Normalized Root Mean Square Error (NRMSE)(refers to (3))

(2) Correlation Coefficient (CC)

CC = 119899sum119910act119910pred minus sum119910actsum119910predradic119899sum1199102act minus (sum119910act)2radic119899sum1199102pred minus (sum119910pred)2

(7)

(3) Mean Absolute Percentage Error (MAPE)

MAPE = 100119899119899sum119894=1

10038161003816100381610038161003816100381610038161003816119910act minus 119910pred119910act10038161003816100381610038161003816100381610038161003816 (8)

In the above performance metrics 119910act and 119910pred refer to(3) and 119899 represents total number of data samples that areused to evaluate the respective CI technique The CC value


Table 2 Performance measurement of the ANN SVR and ANFIS for modeling asphalts (wet samples)

SVR ANN ANFISTrain Test Train Test Train Test

NRMSE 04589 06170 04356 06017 03545 06905CC 08891 07986 08988 07996 09342 07741MAPE () 01579 01696 01488 01532 01763 02558

Table 3 Performance measurement of the ANN SVR and ANFIS for modeling asphalts (dry samples)


NRMSE 03596 07062 02667 05859 03464 06958CC 09323 07847 09649 08196 09372 07749MAPE () 01952 02612 01623 02125 01763 02358

Table 4 Performances of the ensemble of CI-statistical methods for modeling asphalts (wet samples)

Ensemble of CI-average Ensemble of CI-weighted average Ensemble of CI-linear functionTrain Test Train Test Train Test

NRMSE 03750 05976 03705 06003 03449 06490CC 09269 08099 09288 08087 09383 07903MAPE () 01352 01495 01387 01505 01205 01834

Table 5 Performances of the ensemble of CI-statistical measurements for modeling asphalts (dry samples)


NRMSE 03057 06526 03034 06497 02656 05813CC 09520 07959 09528 07969 09651 08219MAPE () 01703 02222 01694 02195 01645 02112

ldquoONErdquo indicates perfect correlation where ldquoZEROrdquo indicatesno correlation between the prediction and actualmeasuredoutcome The MAPE is the indication of the error as apercentage to depict it explicitly and therefore a smaller valueofMAPE indicates better performance Similar to theMAPEa smaller NRMSE value indicates better performance

45 Analysis of Experimental Results Tables 2 and 3 showthe moisture damage modeling performances of the three CImethods ANN SVR and ANFIS From these tables overallit is noticed that for wet samples ANN canmodel the asphaltdamage (in terms of adhesioncohesion force prediction)with high accuracy For dry samples the prediction resultsof the ANN model are also very satisfactory compared withthe performances of the other CI methods used in this study

In evaluating the performance of wet samples the exper-imental results in Table 2 indicate that the NRMSE deviatesfrom 06905 (by ANFIS for the test data) followed by SVR(NRMSE = 06170 for test data) and then the ANN producesan NRMSE value of 06017 These results prove that ANNperforms better than the other two methods in modelingmoisture damage to lime- and chemically modified asphaltsA similar performance trend is noticed for the performance

metrics CC andMAPE for wet samples From these findingsa conclusion can be drawn that ANN is the most preferredCI method for moisture damage modeling at nanoscale levelAll these analyses indicate that the performance of ANNis convincing and its performance is very consistent formodeling moisture damage in wet samples of asphalts

Table 3 shows the prediction performances (in terms ofNRMSE) of CI techniques for dry samples As shown in thetable ANFIS produced an NRMSE value of 06958 (for testdata) SVR produced 07062 (for test data) and ANN 06049(for test data) Although ANFIS performed better than SVRthe performance of ANN was the best of the three Howeverthe large difference between the prediction performance oftraining and test data for all the CImethods (eg SVR 07062minus 03596 = 03466 ANN 05859 minus 02667 = 03192 andANFIS 06958 minus 03464 = 03494) indicates that there maybe a possibility of further improving the performances of eachof the methods Similar to the performance metric NRMSEa large difference is seen between the CC values for trainingand testing data respectively Nonetheless a CC value 08196is still a good prediction performance due to being close tothe exact performance (ie 10) An analysis of MAPE valuesalso identified a similar performance trend to those of CC


600

550

500

450

400

350

300

250

200

150

0 2 4 6 8 10 12 14 16 18 20

Observations

Adhe

sion

cohe

sion

forc

e

Predicted versus measured output for wet samples

ANNSVRANFISEnsemble-CI-Avg

Ensemble-CI-Wt-AvgEnsemble-CI-LRegMeasured

Figure 2 Predicted andmeasured adhesioncohesion forces for wetsamples

and NRMSE Interestingly for MAPE the performances gapbetween training and test data is very small (for ANN 02099minus 01864 = 00235) and therefore these performances can beconsidered consistent

Having achieved a very good prediction of the adhe-sioncohesion forces of lime- and chemically modifiedasphalts an attempt is made to enhance the predictionperformances of the CI methods an ensemble of the CImethods with a number of statistical methods is developedas discussed in ldquoEnsemble of CI and Statistical Methodsrdquo sec-tion Tables 4 and 5 show the prediction performances of thethree different ensemble methods As shown in Table 4 theperformance of the ensemble ofCI-averagemethodproduceda very good prediction performance in terms of NRMSE CCand MAPE for the wet asphalt samples It is encouraging toachieve better accuracy than any of the individual CImethodsfor the wet samples (both training and test datasets) of lime-and chemically modified asphalts The prediction accuraciesof the ensemble methods of CI with average and weightedaverage respectively did not produce a better predictionaccuracy than that of the ANN However the ensemble ofCI-linear function produced an accurate prediction and wasbetter than any prediction accuracies of ANN SVR andANFIS Once again the better prediction performances ofthe ensemble methods by combining the predicted values ofANN SVR and ANFIS using average weighted average andlinear function techniques respectively would encourageprofessionals to use these methods for modeling moisturedamage at nanoscale

To summarize Figure 2 plots the measured versus pre-dicted adhesion force data (for wet samples) produced byeach of the three CImethods and the three different ensemblemethods As the plot shows the values predicted by theensemble CI-average technique are very close to the mea-sured values especially for the data samples (observations)from 2 to 5 and 9 to 13 This further suggests the superiorityof the ensemble techniques to the individual CI methodsin modeling moisture damage in asphalt samples under wet

450

400

350

300

250

200

150

100

50

0

0 2 4 6 8 10 12 14 16 18 20

Observations

Adhe

sion

cohe

sion

forc

e

Predicted versus measured output for dry samples



Figure 3 Predicted andmeasured adhesioncohesion forces for drysamples

conditions Figure 3 plots the measured versus predictedvalues for dry samples This plot reveals that the predictedvalues produced by the ensemble of CI-linear function arevery close to the measured values of adhesioncohesionforces This further confirms that the ensemble techniqueshave merit for modeling moisture damage of lime- andchemically modified asphalts under dry conditions

5 Conclusions

In this study an analysis of moisture damage in lime-and chemically modified asphalts was accomplished using anumber of CI techniques namely ANN SVR and ANFISEnsembles of CI and other statistical methods (averageweighted average and linear function) are also developedto better model the moisture damage of asphalts underboth dry and wet conditions It is well known that binderscan improve moisture damage in asphalt In this study theauthors used lime and chemical as asphalt binders and theadhesioncohesion force was then computed at nanoscaleusing AFM Since it is quite difficult and expensive todo the experiments for a range of different amounts ofasphalt binders and measure the adhesioncohesion forcesaccordingly computational modeling of the damage with thevariations of asphalt binders was generatedThe experimentalresults reveal that an ensemble of CI techniques following theaveraging of each of the outputs of CI techniques can modelasphalt damage under wet conditions while an ensembleof CI techniques along with a linear function can producevery accurate predictions of adhesioncohesion forces underdry conditions In conclusion an ensemble of CI-statisticalmeasurements can be applied to capture the complex systemof relationships among various chemical functional groupsthat might influence the intermolecular adhesioncohesionforces of lime- and chemically modified asphalt due tomoisture-induced damage Our future plan is to extendthis research to validate the performance of asphalt bindersthough comparing the adhesioncohesion forces at nanoscale


and macroscale units respectively that would follow the CItechniques developed in this study

Conflicts of Interest

The authors declare that they have no conflicts of interest

Acknowledgments

The authors would like to acknowledge King Fahd Universityof Petroleum amp Minerals and KACST-NSTIP (Grant no 13-NAN525-04) to provide the facilities and finance in order tosupport this research

References

[1] A AASHTO T283 Standard method of test for resistance ofcompacted asphalt mixtures to moisture-induced damage Amer-icanAssociation of StateHighwayAndTransportationOfficials2002

[2] M Arifuzzaman Nano-scale evaluation of moisture damage inasphalt [PhD dissertation] 2010 httpssearchproquestcomdocview853118198accountid=27795

[3] R Kelsall I W Hamley and M Geoghegan Nanoscale Scienceand Technology John Wiley amp Sons 2005

[4] M Partl R Gubler and M Hugener ldquoNano-Science and-Technology for Asphalt Pavementsrdquo in Nanotechnology inConstruction Special Publication pp 343ndash355 Royal Society ofChemistry Cambridge UK 2004

[5] J Yang and S Tighe ldquoA review of advances of nanotechnology inasphalt mixturesrdquo Procedia - Social and Behavioral Sciences vol96 pp 1269ndash1276 2013

[6] A Berquand and B Ohler Common approaches to tipfunctionalization for afm-based molecular recognition measure-ments 2018 2018 httpswwwnanowerkcomnanocatalogproductimages1493739121AN130-RevA0-Common Approachesto Tip Functionalization for AFM Based Molecular Recogni-tion-AppNotepdf

[7] R A Tarefder and A M Zaman ldquoNanoscale evaluation ofmoisture damage in polymer modified asphaltsrdquo Journal ofMaterials in Civil Engineering vol 22 no 7 pp 714ndash725 2010

[8] R A Tarefder and S Ahsan ldquoNeural network modelling ofasphalt adhesion determined by AFMrdquo Journal of Microscopyvol 254 no 1 pp 31ndash41 2014

[9] M R Hassan ldquoModeling of moisture damage in carbon nanotube modified asphalt using hybrid of artificial neural networkand other computational intelligence approachesrdquo Journal ofComputational and Theoretical Nanoscience vol 12 no 11 pp1ndash8 2015

[10] M Arifuzzaman and M R Hassan ldquoMoisture damage pre-diction of polymer modified asphalt binder using SupportVector Regressionrdquo Journal of Computational and TheoreticalNanoscience vol 11 no 10 pp 2221ndash2227 2014

[11] MArifuzzaman ldquoAdvanced annprediction ofmoisture damagein cnt modified asphalt binderrdquo Soft Computing in Civil Engi-neering vol 1 no 1 pp 1ndash11 2017

[12] M Arifuzzaman M S Islam and M I Hossain ldquoMoisturedamage evaluation in SBS and limemodified asphalt usingAFMand artificial intelligencerdquo Neural Computing and Applicationsvol 28 no 1 pp 125ndash134 2017

[13] G Binnig C F Quate and C Gerber ldquoAtomic force micro-scoperdquo Physical Review Letters vol 56 no 9 pp 930ndash933 1986

[14] D Croft G Shed and S Devasia ldquoCreep hysteresis and vibra-tion compensation for piezoactuators atomic force microscopyapplicationrdquo Journal of Dynamic Systems Measurement andControl vol 123 no 1 pp 35ndash43 2001

[15] A Noy D V Vezenov and C M Lieber ldquoChemical forcemicroscopyrdquo Annual Review of Materials Research vol 27 no1 pp 381ndash421 1997

[16] E R Beach G W Tormoen and J Drelich ldquoPull-off forcesmeasured between hexadecanethiol self-assembled monolayersin air using an atomic forcemicroscope Analysis of surface freeenergyrdquo Journal of Adhesion Science and Technology vol 16 no7 pp 845ndash868 2002

[17] Y Okabe U Akiba and M Fujihira ldquoChemical forcemicroscopy of -CH3 and -COOH terminal groups in mixedself-assembled monolayers by pulsed-force-mode atomic forcemicroscopyrdquo Applied Surface Science vol 157 no 4 pp 398ndash404 2000

[18] B Du M R VanLandingham Q Zhang and T He ldquoDirectmeasurement of plowing friction and wear of a polymer thinfilm using the atomic force microscoperdquo Journal of MaterialsResearch vol 16 no 5 pp 1487ndash1492 2001

[19] J-FMasson V Leblond and JMargeson ldquoBitumenmorpholo-gies by phase-detection atomic force microscopyrdquo Journal ofMicroscopy vol 221 no 1 pp 17ndash29 2006

[20] A T Pauli R W Grimes A G Beemer T F Turner and J FBranthaver ldquoMorphology of asphalts asphalt fractions andmodel wax-doped asphalts studied by atomic forcemicroscopyrdquoInternational Journal of Pavement Engineering vol 12 no 4 pp291ndash309 2011

[21] H Ceylan M B Bayrak and K Gopalakrishnan ldquoNeural net-works applications in pavement engineering A recent surveyrdquoInternational Journal of Pavement Research and Technology vol7 no 6 pp 434ndash444 2014

[22] I Flood and Ian ldquoTowards the next generation of artificialneural networks for civil engineeringrdquo Advanced EngineeringInformatics vol 22 no 1 pp 4ndash14 2008

[23] R Hecht-Nielsen and Robert Neurocomputing Addison-Wesley Pub Co 1990

[24] D E Rumelhart J LMcClelland and SD P R GUniversity ofCalifornia Parallel distributed processing explorations in themicrostructure of cognition MIT Press 1986

[25] V N Vapnik The Nature of Statistical Learning Theory vol 8Springer 2000

[26] MMohammadhassani H Nezamabadi-PourM Z JumaatMJameel andAM S Arumugam ldquoApplication of artificial neuralnetworks (ANNs) and linear regressions (LR) to predict thedeflection of concrete deep beamsrdquo Computers and Concretevol 11 no 3 pp 237ndash252 2013

[27] G Chen and G Su ldquoStudy on pavement design based on MC-SVM method of reliabilityrdquo in Proceedings of the 2010 Inter-national Conference on Mechanic Automation and ControlEngineering MACE2010 pp 4877ndash4880 June 2010

[28] K-Z Yan D-D Ge and Z Zhang ldquoSupport Vector MachineModels for Prediction of Flow Number of Asphalt MixturesrdquoInternational Journal of Pavement Research and Technology vol7 no 1 pp 31ndash39 2014


[29] Y Ke-zhen H Liao H Yin and L Huang ldquoPredicting thePavement Serviceability Ratio of Flexible Pavement with Sup-port Vector Machinesrdquo in Road Pavement and Material Char-acterization Modeling and Maintenance pp 24ndash32 AmericanSociety of Civil Engineers Reston VA USA 2011

[30] M Soltani T B Moghaddam M R Karim S Shamshirbandand C Sudheer ldquoStiffness performance of polyethylene tereph-thalate modified asphalt mixtures estimation using supportvector machine-firefly algorithmrdquo Measurement vol 63 pp232ndash239 2015

[31] K Gopalakrishnan and S Kim ldquoSupport vector machines fornonlinear pavement backanalysisrdquo Journal of Civil Engineeringvol 38 no 2 pp 173ndash190 2010

[32] K Gopalakrishnan and S Kim ldquoSupport vector machinesapproach to HMA stiffness predictionrdquo Journal of EngineeringMechanics vol 137 no 2 pp 138ndash146 2010

[33] M Maalouf N Khoury and T B Trafalis ldquoSupport vectorregression to predict asphalt mix performancerdquo InternationalJournal for Numerical and Analytical Methods in Geomechanicsvol 32 no 16 pp 1989ndash1996 2008

[34] J S R Jang ldquoANFIS adaptive-network-based fuzzy inferencesystemrdquo IEEE Transactions on Systems Man and Cyberneticsvol 23 no 3 pp 665ndash685 1993

[35] A Talei L H C Chua and C Quek ldquoA novel application of aneuro-fuzzy computational technique in event-based rainfall-runoff modelingrdquo Expert Systems with Applications vol 37 no12 pp 7456ndash7468 2010

[36] D Petkovic M Issa N D Pavlovic N T Pavlovic and LZentner ldquoAdaptive neuro-fuzzy estimation of conductive sili-cone rubber mechanical propertiesrdquo Expert Systems with Appli-cations vol 39 no 10 pp 9477ndash9482 2012

[37] D Petkovic and Z Cojbasic ldquoAdaptive neuro-fuzzy estimationof autonomic nervous system parameters effect on heart ratevariabilityrdquo Neural Computing and Applications vol 21 no 8pp 2065ndash2070 2012

[38] D Petkovic N D Pavlovic Z Cojbasic and N T PavlovicldquoAdaptive neuro fuzzy estimation of underactuated roboticgripper contact forcesrdquo Expert Systems with Applications vol40 no 1 pp 281ndash286 2013

[39] D Petkovic Z Cojbasic and S Lukic ldquoAdaptive neuro fuzzyselection of heart rate variability parameters affected by auto-nomic nervous systemrdquo Expert Systems with Applications vol40 no 11 pp 4490ndash4495 2013

[40] C L Wu and K W Chau ldquoPrediction of rainfall time seriesusing modular soft computing methodsrdquo Engineering Applica-tions of Artificial Intelligence vol 26 no 3 pp 997ndash1007 2013

[41] S Shamshirband D Petkovic R Hashim S Motamedi and NB Anuar ldquoAn appraisal of wind turbine wake models byadaptive neuro-fuzzy methodologyrdquo International Journal ofElectrical Power amp Energy Systems vol 63 pp 618ndash624 2014

[42] M Mohammadhassani H Nezamabadi-Pour M Jumaat MJameel S J S Hakim and M Zargar ldquoApplication of theANFIS model in deflection prediction of concrete deep beamrdquoStructural Engineering andMechanics vol 45 no 3 pp 319ndash3322013

[43] M Yilmaz B V Kok B Sengoz A Sengur and EAvci ldquoInvesti-gation of complex modulus of base and EVAmodified bitumenwith Adaptive-Network-Based Fuzzy Inference Systemrdquo ExpertSystems with Applications vol 38 no 1 pp 969ndash974 2011

[44] E Ozgan I Korkmaz andM Emiroglu ldquoAdaptive neuro-fuzzyinference approach for prediction the stiffness modulus on

asphalt concreterdquo Advances in Engineering Software vol 45 no1 pp 100ndash104 2012

[45] G Shafabakhsh and A Tanakizadeh ldquoInvestigation of loadingfeatures effects on resilient modulus of asphalt mixtures usingAdaptive Neuro-Fuzzy Inference Systemrdquo Construction andBuilding Materials vol 76 pp 256ndash263 2015

[46] D N Little and D Jones ldquoChemical and mechanical processesof moisture damage in hot-mix asphalt pavementsrdquo in Proceed-ings of the National Seminar on Moisture Sensitivity of AsphaltPavements pp 37ndash70 2003

[47] A Vaidya and M K Chaudhury ldquoSynthesis and surface prop-erties of environmentally responsive segmented polyurethanesrdquoJournal of Colloid and Interface Science vol 249 no 1 pp 235ndash245 2002

[48] D M L Mas and D P Ahlfeld ldquoComparing artificial neuralnetworks and regression models for predicting faecal coliformconcentrationsrdquoHydrological Sciences Journal vol 52 no 4 pp713ndash731 2007

[49] H S Hota A K Shrivas and S K Singhai ldquoArtificial NeuralNetwork Decision Tree and Statistical Techniques Applied forDesigning and Developing E-mail Classifierrdquo InternationalJournal of Recent Technology and Engineering pp 2277ndash38782013

Computer Games Technology

International Journal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Journal ofEngineeringVolume 2018

Advances in

FuzzySystems


Volume 2018


ReconfigurableComputing



Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

thinspArtificial Intelligence

Hindawiwwwhindawicom Volumethinsp2018


Civil EngineeringAdvances in


Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications


Hindawi

wwwhindawicom Volume 2018

Advances in

Multimedia


Biomedical Imaging



Engineering Mathematics


RoboticsJournal of



Computational Intelligence and Neuroscience


Mathematical Problems in Engineering

Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018


Human-ComputerInteraction

Advances in


Scientic Programming

Submit your manuscripts atwwwhindawicom


of these materials still depends to a large degree on thephysical properties in terms ofmicroscale and nanoscale level[4] Although engineers material producers and researchershave explored the potential for many years the analysis ofasphalt binder properties at nanoscale unit has been found tobe limited in utilizations [5] Since all the destruction (egadhesioncohesion failures) initiates from the nanoscale tothe microscale the current focus is now being placed onnanotechnological (the technology that analyzes the prop-erties in terms of nanoscale measurement) applications inasphalt The derived information can be useful in multiscalemodeling of moisture damage in asphalt To the knowledgeof the authors only a few studies by Berquand and Ohler [6](accessed January 1 2018) Tarefder and Zaman [7] Tarefderand Ahsan [8] Hassan [9] Arifuzzaman and Hassan [10]Arifuzzaman [11] and Arifuzzaman et al [12] have testednanoscale moisture damage using functionalized AFM tipsIn the study by Arifuzzaman andHassan [10] the researchersmodeled the moisture damage in polymer-modified asphaltwhile Hassan [9] modeled the moisture damage in CarbonNanotube (CNT) modified asphalt Tarefder and Zaman [7]studied the effect of polymer alteration due to moisturedamage in asphalt using AFM without applying any artifi-cial intelligence to detect the damaging behavior HoweverTarefder andAhsan [8] developed anANNmodel to quantifyadhesion from the AFM data Arifuzzaman [11] developedANN model to determine the moisture damage behavior ofthe CNT modified asphalt binder None of these previousstudies explained or predicted moisture damage in lime- andchemically modified asphalt binder In this paper the authorsdevelop an ensemble CI-statistical technique to model mois-ture damage in lime- and chemically modified asphalt Inthis process first a number of well-known CI techniques areapplied namely Artificial Neural Networks (ANNs) SupportVector Regression (SVR) and an Adaptive Neuro FuzzyInference System (ANFIS) Initially functionalized tips wereused to measure the weak intermolecular forces (eg adhe-sioncohesion) in polymer-modified asphalt binder systemsThe level of moisture damage was measured at nano-Newtonscale for varying measurements of lime antistripping agentsUnichem (UC) and spring constant values (119870 values) andvarying the functionalized AFM tips carboxyl (-COOH)ammin (-NH3) methyl (-CH3) among hydroxyl (-OH)and silicon nitride (Si3N4) Since no practical attempt hasbeen made to date to match the functional groups that existin asphalt binder this paper attempts to analyze the effectof these functional groups on moisture damage in asphaltusing the above-mentioned CI techniques To improve theperformance of the CI techniques further a number ofstatistical measurements are applied to aggregate the outputsof each of the CI techniquesTherefore the aggregated valuesare considered as predicted values

The rest of this paper is structured as follows ldquoBack-ground and Literature Reviewrdquo section provides an overviewof AFM ANN SVR and ANFIS and a literature review ofmoisture damage modeling using these techniques ldquoEnsem-ble of CI and statistical methodsrdquo section describes theensemble model ldquoMaterials Description and Experimenta-tionrdquo section presents the description of the dataset the

experimental setup the experimental results and the anal-ysis Finally ldquoConclusionrdquo section concludes the paper

2 Background and Literature Review

In this section AFM and the three CI techniques namelyANN SVR and ANFIS are briefly described At the endof each subsection a discussion about how and where thesemethods have been applied to model asphalt damage isprovided

21 Theory of Atomic Force Microscopy (AFM) An AFM is atool to investigate the topographical and surface properties ofseveral engineering materials It is designed to evaluate vari-ous local properties such as height friction and magnetismIt scans and then calculates the local properties continuouslyto produce an asphalt binder bonding image All the AFMprobes has a very sharp tip with usual values 3 to 6120583m tallpyramid with a 15 to 40 nm end radius AFM scan measuresthe deflections of the cantilever which the lever performsthrough the reaction of a laser beam of the cantilever [13]The tip is positioned with high resolution by piezoelectricceramic that expands or contracts due to the voltage gradientand helps in positioning 3D devices with high accuracy[14]

211 AFM Testing AFM is used by the chemical engineeringresearch groups at Harvard University to evaluate adhesionand cohesion forces The use of AFM in the chemicalprocess is known as chemical force microscopy [15] AFMwas also used by Beach et al [16] to find drag off pro-cess between hexadecanethiol monolayers that automaticallypulled together the silicon nitride tip as cantilever and siliconwafer Okabe et al [17] used hydrophobic -CH3 and -COOHterminating alkane thiols functionally modified to measureadhesion forces and to image the samples Du et al [18]used AFM to find thin films of polymer-generated that yieldsstrength and modulus of elasticity Researchers have alsoused nonfunctionalized AFM tips to measure the rigidity ofasphalt binder [19] Pauli et al [20] measured the surfaceenergy of asphalt binder using AFM Masson et al [19] alsoused AFM to execute imaging of asphalt and connectedthe images with different chemically analyzed ingredients ofbitumen

22 Artificial Neural Network (ANN) Artificial Neural Net-works (ANNs) have shown their versatile application tooptimize and forecast immensely complex uncertain orpoorly understood situations by simulating the real fieldscenario for the last three decades in different disciplines ofcivil engineering Following the same trend the applicationof ANN in transportation engineering especially in pavementengineering has also received attention (eg Ceylan et al [21]Flood and Ian [22])

An ANN resembles the biological brain by being likea parallel distributed processor with high computationalpower Hecht-Nielsen and Robert [23] and interconnectedprocessing units or neuronsThe parallel processor is capableof solving nonlinear relations by storing and providing data













119910avg = 119910ann + 119910svr + 119910anfis3 (1)







(3)


997888997888rarr119884act =3sum119894=1

119887119894 times997888rarr119884119894 (4)

























(7)


MAPE = 100119899119899sum119894=1






NRMSE 04589 06170 04356 06017 03545 06905CC 08891 07986 08988 07996 09342 07741MAPE () 01579 01696 01488 01532 01763 02558



NRMSE 03596 07062 02667 05859 03464 06958CC 09323 07847 09649 08196 09372 07749MAPE () 01952 02612 01623 02125 01763 02358



NRMSE 03750 05976 03705 06003 03449 06490CC 09269 08099 09288 08087 09383 07903MAPE () 01352 01495 01387 01505 01205 01834



NRMSE 03057 06526 03034 06497 02656 05813CC 09520 07959 09528 07969 09651 08219MAPE () 01703 02222 01694 02195 01645 02112







600

550

500

450

400

350

300

250

200

150

0 2 4 6 8 10 12 14 16 18 20

Observations

Adhe

sion

cohe

sion

forc

e








450

400

350

300

250

200

150

100

50

0

0 2 4 6 8 10 12 14 16 18 20

Observations

Adhe

sion

cohe

sion

forc

e






5 Conclusions






Acknowledgments


References

























































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in
















119910avg = 119910ann + 119910svr + 119910anfis3 (1)







(3)


997888997888rarr119884act =3sum119894=1

119887119894 times997888rarr119884119894 (4)

























(7)


MAPE = 100119899119899sum119894=1






NRMSE 04589 06170 04356 06017 03545 06905CC 08891 07986 08988 07996 09342 07741MAPE () 01579 01696 01488 01532 01763 02558



NRMSE 03596 07062 02667 05859 03464 06958CC 09323 07847 09649 08196 09372 07749MAPE () 01952 02612 01623 02125 01763 02358



NRMSE 03750 05976 03705 06003 03449 06490CC 09269 08099 09288 08087 09383 07903MAPE () 01352 01495 01387 01505 01205 01834



NRMSE 03057 06526 03034 06497 02656 05813CC 09520 07959 09528 07969 09651 08219MAPE () 01703 02222 01694 02195 01645 02112







600

550

500

450

400

350

300

250

200

150

0 2 4 6 8 10 12 14 16 18 20

Observations

Adhe

sion

cohe

sion

forc

e








450

400

350

300

250

200

150

100

50

0

0 2 4 6 8 10 12 14 16 18 20

Observations

Adhe

sion

cohe

sion

forc

e






5 Conclusions






Acknowledgments


References

























































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in









(3)


997888997888rarr119884act =3sum119894=1

119887119894 times997888rarr119884119894 (4)

























(7)


MAPE = 100119899119899sum119894=1






NRMSE 04589 06170 04356 06017 03545 06905CC 08891 07986 08988 07996 09342 07741MAPE () 01579 01696 01488 01532 01763 02558



NRMSE 03596 07062 02667 05859 03464 06958CC 09323 07847 09649 08196 09372 07749MAPE () 01952 02612 01623 02125 01763 02358



NRMSE 03750 05976 03705 06003 03449 06490CC 09269 08099 09288 08087 09383 07903MAPE () 01352 01495 01387 01505 01205 01834



NRMSE 03057 06526 03034 06497 02656 05813CC 09520 07959 09528 07969 09651 08219MAPE () 01703 02222 01694 02195 01645 02112







600

550

500

450

400

350

300

250

200

150

0 2 4 6 8 10 12 14 16 18 20

Observations

Adhe

sion

cohe

sion

forc

e








450

400

350

300

250

200

150

100

50

0

0 2 4 6 8 10 12 14 16 18 20

Observations

Adhe

sion

cohe

sion

forc

e






5 Conclusions






Acknowledgments


References

























































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in

















(7)


MAPE = 100119899119899sum119894=1






NRMSE 04589 06170 04356 06017 03545 06905CC 08891 07986 08988 07996 09342 07741MAPE () 01579 01696 01488 01532 01763 02558



NRMSE 03596 07062 02667 05859 03464 06958CC 09323 07847 09649 08196 09372 07749MAPE () 01952 02612 01623 02125 01763 02358



NRMSE 03750 05976 03705 06003 03449 06490CC 09269 08099 09288 08087 09383 07903MAPE () 01352 01495 01387 01505 01205 01834



NRMSE 03057 06526 03034 06497 02656 05813CC 09520 07959 09528 07969 09651 08219MAPE () 01703 02222 01694 02195 01645 02112







600

550

500

450

400

350

300

250

200

150

0 2 4 6 8 10 12 14 16 18 20

Observations

Adhe

sion

cohe

sion

forc

e








450

400

350

300

250

200

150

100

50

0

0 2 4 6 8 10 12 14 16 18 20

Observations

Adhe

sion

cohe

sion

forc

e






5 Conclusions






Acknowledgments


References

























































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in







NRMSE 04589 06170 04356 06017 03545 06905CC 08891 07986 08988 07996 09342 07741MAPE () 01579 01696 01488 01532 01763 02558



NRMSE 03596 07062 02667 05859 03464 06958CC 09323 07847 09649 08196 09372 07749MAPE () 01952 02612 01623 02125 01763 02358



NRMSE 03750 05976 03705 06003 03449 06490CC 09269 08099 09288 08087 09383 07903MAPE () 01352 01495 01387 01505 01205 01834



NRMSE 03057 06526 03034 06497 02656 05813CC 09520 07959 09528 07969 09651 08219MAPE () 01703 02222 01694 02195 01645 02112







600

550

500

450

400

350

300

250

200

150

0 2 4 6 8 10 12 14 16 18 20

Observations

Adhe

sion

cohe

sion

forc

e








450

400

350

300

250

200

150

100

50

0

0 2 4 6 8 10 12 14 16 18 20

Observations

Adhe

sion

cohe

sion

forc

e






5 Conclusions






Acknowledgments


References

























































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in





600

550

500

450

400

350

300

250

200

150

0 2 4 6 8 10 12 14 16 18 20

Observations

Adhe

sion

cohe

sion

forc

e








450

400

350

300

250

200

150

100

50

0

0 2 4 6 8 10 12 14 16 18 20

Observations

Adhe

sion

cohe

sion

forc

e






5 Conclusions






Acknowledgments


References

























































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in








Acknowledgments


References

























































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in
































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in




moisture damage modeling in lime and chemically modified

Documents