Advanced ANN Prediction of Moisture Damage in CNT Modified Asphalt Binder

Document Type: Regular Article

Author

Assistant Professor, Department of Civil Engineering, University of Bahrain, Bahrain

Abstract

Moisture penetration causes many direct and indirect distresses in flexible asphalt pavement. Due to damage in asphalt concrete and binder by moisture are the prime concern of failure for flexible pavement worldwide. The causes and prediction are investigated in this study. The asphalt binder was modified with carbon nanotubes (CNT) with very small percentages. The modified binder was simulated with moisture damage with AASHTO T-283 methods. In this study, polymer and carbon nanotubes (CNT) have been added to liquid asphalt binder to examine whether the resulting modified binder has improved moisture damage resistance. Using laboratory tested data, an artificial intelligence modeling technique has been utilized to determine the moisture damage behavior of the modified binder. Multi-Layer Perceptron (MLP) provides the best prediction for wet and dry samples AFM readings with R2 values respectively 0.6407 and 0.8371.

Highlights

Google Scholar

Keywords

Main Subjects


1. Introduction

Asphalt binder has been used in roadway pavements since the early 1900’s. The USA has nearly 3.2 million kilometers of pavement made of asphalt binder. Since the pavement is not impervious hence the water from various sources enters the pavement and consequently damage it. Nonetheless, the moisture damage occurs globally with the presence of water. In literature, asphalt pavement distresses are related to the moisture intrusion in asphalt roads. Four attributes, namely: the adhesive bond between the binder, cohesive resistance of binder, the frictional resistance between the aggregate particles and the aggregate interlock are related to pavement strength [1].

Damage due to moisture in pavement causes loss of integrity of a Hot Mix Asphalt (HMA) mix through the weakening of the adhesion bond between the asphalt binder and aggregate via the act of water. This occurrence is well known as the stripping which eventually leads to the breaking of the aggregate-asphalt adhesion bond and finally direct towards the failure of flexible pavement. Moreover, asphalt binder may experience changes in strength, stiffness, and viscosity which refer to the cohesion in asphalt as a result of the moisture-related sensitivity of asphalt pavements [2].

Although the research scientist and practicing engineers are thriving to prolong the life of the nation pavement till now due to lack of proper understanding of moisture interaction with asphalts, pavements are prone to early failure from the infiltration of moisture. Several distresses may occur individually and simultaneously inside pavement due to this moisture entrance. The water reduces the internal bonding strength which leads to fatigue cracking/rutting in pavements [3]. Since damage starts from nano to micro level, current research aims to conduct binder testing at nano level using Atomic Force Microscopy (AFM). The AFM produced laboratory output data were finally modeled with three advanced artificial neural intelligence tools.

2. Background of Research

Moisture damage in asphalt pavements is one of the main and primary distresses which is associated with the fracturing of the pavement surface, permanent deformation and excessive cracking [4]. Some literature related to pavement distresses, and ANN related works are investigated and described below:

Gandhi et al. constructed ANN models to predict tensile strength ratio (TSR) and the indirect tensile strength (ITS) of various asphalt pavement mixtures considering 5 no. of input variables such as aggregate source, asphalt binder source, anti-striping agent (ASA), conditioning duration, and asphalt binder content [5]. It is worth mentioning here that, ANN is a widely used computational tool that can identify the complex unknown relationship between the inputs and desired output to the system. The activity of an ANN was introduced following the activities of a biological brain. ANN-based pavement thickness determination in order to back-calculate the layer moduli was accomplished by Tarefder et al. [6]. Xiao and Amirkhanian also incorporated ANN for prediction of the stiffness behavior of asphalt concrete in order to reuse reclaimed asphalt pavement materials [7].

Although, artificial neural network (emphasized on data science and data mining) is not new in civil engineering application [8]. Mohammadhassani et al. applied ANFIS, ANN model to concrete deep beam experiment and analysis [9]. However, for moisture damage in asphalt binder and concrete in pavement construction, modeling with artificial network and data science is a relatively new area. Not too many attempts can be found to model and investigate artificial data for modeling the moisture damage and associated effect in the asphalt pavement area.

Arifuzzaman and Hassan used SVR to predict the moisture damage in asphalt binder without adding CNT [10]. Adham and Arifuzzaman used CNT to resist moisture damage. They did not accomplish modeling work with ANN [11]. Amirkhanian et al. (2011) evaluated the rheological and engineering properties of different asphalt binder modified with various percentages of carbon nanoparticles (0.2%, 0.5%, 1.0%, and 1.5% by weight of the binder). They included in the experimental design of their study the utilization of three binder sources containing one type of nanoparticle (PG 64-22) [12]. Hassan used CNT and ANN to model moisture damage without application of Functional AFM tips (COOH, CH3, NH3 and OH groups) which are needed to predict the asphalt binder adhesion and cohesion forces [13]. To the knowledge, this research is the first time attempt to model the moisture damage behavior with CNT modified asphalt binder with functionalized AFM probes.

3. Aim of the study

To make a good relationship for moisture damage study in carbon nanotubes (CNT) mixed asphalt binder, the research-oriented specific objectives are proposed as:

1. Application of artificial neural network (i,e. SVM, ANFIS, and MLP) on CNT modified asphalt binder.

2. Model the large data set from moisture damage of asphalt binder using ANN, ANFIS, and SVM.

3. To compare the performance of all models and to suggest the best one.

4.1. Test matrix

In this experiment, very used small percentages of carbon nanotubes (CNT) were used in SB and SBS modified base asphalt binder. Both dry and wet samples were tested under AFM with functionalized tips (-CH3, NH3, -COOH, -OH and –Si3N4) to get 269 data point for force distance measurement. Three methods SVM, ANFIS, and ANN, were used several hundred times to model these 269 data points to achieve the perfect results. The description for CNT, SB and SBS additives are given below:

4.2. Materials and additives

Base Binder: The base binder was collected from an oil-asphalt refinery commercial factory. Then SB, SBS, and CNT were added to it according to the following discussions.

Description of polymer: Styrene-Butadiene (SB) polymer

This is a latex type polymer used to improve the elastic recovery as well as cohesive and adhesive properties of the pavement structures. In this experiment total, two percentages (4% and 5%) by weight of SB polymer was mixed with the base binder.

Styrene-Butadiene-Styrene (SBS) polymer: Becker et al. described SBS to be the most preferable and used polymer to modify the base binder. In this experiment, total two percentages (4% and 5%) by weight of SBS polymer was mixed with the base binder [14].

Carbon Nano Tubes (CNT): Carbon nanotubes are structural as rolled-up graphite sheets having one end as capped. These type of tiny tubes may have single walls or multiple walls. The single-wall carbon nanotubes (SWNT), unlike carbon/graphite black, possess highly desirable mechanical, electrical and thermal properties [15,16], with many potential applications in the computer, electronics as well as in aerospace industries. It is reported that the addition of a very small amount of carbon nanotube CNT (around 1%) by weight may increase both the compressive and the flexural strength as well as mechanical properties of civil engineering materials [17]. (Mann 2006). Our research found very few attempts to use CNT with asphalt binder for improvement [11,18]. In this study, Single Wall (0.5%, 1.0%, and 1.5%) CNT were used to resist moisture damage in asphalt binder.

4.3. Sample preparation

The moisture damage in the field was simulated in laboratory sample of the asphalt binder (AASHTO T-283 method). In this method, the asphalt samples were conditioned under freezing and thawing environment for several days. Thus the wet samples were ready for AFM testing. Dry samples were usual samples without any interference of water.

5. Modelling Moisture Damage Data with Artificial Intelligence (AI) Approaches

The input values for ANN were Condition (dry and wet), percentages of SB and SBS polymer mixed with the base binder, percentages of CNT, AFM tips type (COOH, CH3, NH3, OH, and Si3N4). The output result value was AFM experimental output as adhesion/cohesion force in terms of nano-newton in asphalt binder.

Statistical analysis of the data set is shown in Table 1.

The statistical analysis of the AFM laboratory generated data is shown in Table 1. The standard deviation value tells the necessity of artiļ¬cial intelligence modeling of the highly deviated values from the standard one. The Kurtosis values of the data indicate that all the AFM produced data are less outlier prone than that of the normal distribution of the data.

Table 1.

The statistical analysis of data used for the study.

 

Dry

Wet

Maximum

382.51

405.01

Minimum

41.95

106.45

Mean

168.4945

225.7373

St Dev

88.48392

80.15665

Kurtosis

0.360758

-0.996740

skewness

1.13887

0.50977

To model the 269-datasets from AFM run test a total of three advanced ANN was used in this research. The ANN modeling work was accomplished in this study be splitting data set into two subsets: training set (80%) and validation set (20%). In order to train the NN system for tracing the complex and the nonlinear relationships between output and input values the training data set was utilized. The stopping criterion (trials and error) helps prevent the overfitting in ANN. The forecasting ability of the network was tested with 20% ANN dataset. One of the key concern in the NN architecture selection are the hidden layers as well as the number of nodes in each one of them. It may depend on a complex relationship which need to be addressed further [19]. It also requires several trial-error as well as a good selective judgment for appropriate ANN architecture which is the best suited for the concerned data set.

The brief description of the three systems are described below:

5.1. Adaptive Network Fuzzy Inference System (ANFIS)

This special ANN was proposed by Jang [20]. It serves as a basic function to construct Fuzzy Interference System (FIS) and if-then formulated rules. More precisely, the ANFIS combines both neural network and fuzzy system from the best attributes and characteristics point of view. It can be successfully applied to many real-life problems which are complicated for modeling work such as highly nonlinear function, for identifying nonlinear components in control type of systems, foresee-disorder time series, etc. ANFIS description can be found more in details in [20].

5.2. Support Vector Machines (SVM)

Cores and Vapnik [21] proposed SVM. This simulation tool is best suitable outcome based geometric hyper-plane function for special separating data points. In this technique, the input data are mapped in highly dimensional characteristics space. The main and critical shortcomings of SVM are the duration for runtime and accuracy issues.

5.3. MLP based Artificial Neural Network (ANN)

This process can be described as the human brain performing computational programming which simulates the data learning and predicting tasks. One of the most popular used in this study is Multi-Layer Perceptron (MLP) which has vast applicability in pattern recognition and identification, speech recognition, control systems, petroleum engineering, etc. [22]. Hence the MLP based process can be used for novel and efficient system for computational efficiency.

The overall flow diagram for the above-mentioned methods is shown in Figure 1.

 

a) ANN

 

b) ANFIS

 

c) SVM

Fig. 1. a) Schematic of an ANN, b) ANFIS, c) SVM models.

6. Analyses of Experimental Data

The data Table 2 and Table 3 show the analysis summary of the dry and wet samples with the above three mentioned method concerning the six measured performances as:

Table 2.

Dry Samples Results.

 

SVM

ANFIS

MLP

 

Train (80%)

Test (20%)

Train (80%)

Test (20%)

Train (80%)

Test (20%)

RMSE

16.2117

42.9149

38.4550

36.8155

29.5374

34.8987

CC

0.9810

0.7567

0.8476

0.8161

0.9149

0.8386

MAPE

7.2659

31.2811

18.6932

24.0868

14.9529

20.8287

NRMSE

0.2227

0.6672

0.5282

0.5724

0.4057

0.5426

MAPE (/100)

0.0727

0.3128

0.1869

0.2409

0.1495

0.2083

MAE

10.3951

30.0103

26.6549

29.8941

22.5585

25.4221

SD Error

0.2714

6.3164

1.8757

3.6325

2.9787

1.1294

AOC

9.99340322

27.314005

25.8901

28.4880

22.1676

23.6528

Time (Sec)

0.780005

0.0156

0.6240

0.0156

5.6472

0.0000

MRE

7.2659

31.2811

18.6932

24.0868

14.9529

20.8287

 

Table 3.

Wet Samples Results.

 

SVM

ANFIS

MLP

 

Train (80%)

Test (20%)

Train (80%)

Test (20%)

Train (80%)

Test (20%)

RMSE

36.6028

48.7577

30.5458

54.7252

31.0656

46.2332

CC

0.9078

0.7723

0.9368

0.7563

0.9373

0.8004

MAPE

11.1859

17.4761

12.8757

18.7533

12.5359

19.0857

NRMSE

0.4175

0.6254

0.3484

0.7019

0.3543

0.5930

MAPE (/100)

0.1119

0.1748

0.1288

0.1875

0.1254

0.1909

MAE

22.7841

37.5612

24.0691

39.9564

24.0460

37.6852

SD Error

0.8901

2.2477

0.7049

1.2707

0.2676

2.0414

AOC

22.1674

35.4661

23.5699

37.5300

23.5184

35.7585

Time (Sec)

0.8736056

0.0000

0.6240

0.0156

2.8704

0.0000

MRE

11.1859

17.4761

12.8757

18.7533

12.5359

19.0857

 

The Correlation Coefficient (CC), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Mean Relative Error (MRE) and Standard Deviation (SD) were calculated to investigate the accuracy of our utilized models. To find how much strong co-relation exists between the trend and the observed data, CC value is used. The value of CC “1” represents very much ideal co-relation exists while “0” means no co-relation can be predicted for the tested results. MAE is the measurement of the size of error while RMSE utilizes squaring mechanism to enlarge bigger error as well as suppressing smaller error. MAPE co-relates the magnitude of error to the magnitude of the observed real data.

For all cases, the performance of models was better for training datasets (80% samples) as compared to test datasets (20%) which are understandable. However, the variation between model performance for train and test datasets is more in case of dry samples as compared to wet samples. This could be because dry samples’ data contains more variations/noise. Moreover, the performance measures of all models for error terms (RMSE, MAE, etc.) were higher for wet samples’ test dataset while the relative error terms (MAPE) were lower for them. This could be because the magnitude of the dependent variable for wet samples was higher, so the error term magnitude was higher corresponding to that. However, the performance of wet samples’ models was better, which resulted in better relative error terms.

 

Fig. 2. MLP result on Predicted vs. Measured Data (Wet Samples).

 

Fig. 3. MLP result on Predicted vs. Measured Data (Dry Samples).

This experimental test analysis run is made with different values of dry and wet samples with SVM, MLP, and ANFIS-FCM method. For dry samples, the plot of predicted versus measured is developed in these three methods. Generated R2 values and equations are shown through linear trend line in the corresponding graph plots for each method in Figures 2 and 3.

The best two selected graphs out of sixteen are shown below based on the R2 values. Plotting of wet material analysis through SVM, MLP & ANFIS has been created. For wet materials, generated R2 values and equations are also shown through linear trend line in Figure 2 and 3.

The training and prediction time was calculated in terms of seconds in this study. The prediction time was very low for all the models because it has a comparatively small dataset and does not include any optimization algorithm as the model is already developed. Among the three methods for overall computation, MLP took the highest amount of time for training of both the dry and wet samples which was done for 80% of the total dataset. Hence, use of MLP on a large dataset cannot be recommended based upon these observations.

7. Discussion and Conclusion

The investigations provide a detail description of the test methods taking place for dry and wet samples. The measurement techniques & procedures are developed by a generation of linear trend line through the plotted points and comparing the R2 value to find the efficient technique or procedures for the test of dry and wet samples.

Artificial neural network and related modeling (ANFIS, MLP, and SVM) tools are used for the first time on AFM produced moisture damage data of CNT modified asphalt binder.

The conclusions drawn from the experimental and modeling work are listed below:

1. The neural network can address the complex relationship between adhesion and test variables incorporated in AFM testing. The developed NN model shows good prediction ability.

2. All models employed in this study gave CC value of more than 70% for all cases which is evident of their applicability for the prediction problem of this study. Moreover, the standard deviation of error was also relatively small for all models for training as well as prediction datasets which shows the robustness of these models.

3. The study shows that MLP test for dry and wet sample produces convincing error terms with the RMSE improvement of 2–8 units (approximately 3 – 5%) over SVM and ANFIS method.

4. So, it can be concluded with the promising fact that MLP procedure for dry (R2 = 0.8371) and wet (R2 = 0.6407) sample is more effective to predict real-life moisture damage data.

5. Despite the above reasoning, it was also observed that MLP took longer training time than SVM and ANFIS. This could be due to optimization algorithm of MLP models; hence their applicability for large datasets should be further investigated.

8. Recommendation

The economic aspect, health consequences, and availability of CNT utilization in road and highways are recommended for further investigation.

Acknowledgment

The author would like to thank University of Bahrain (UOB) for computational facilities, library usage and to support this research. Special appreciation to Engr. G. Baig from KFUPM, Saudi Arabia for preparing AFM testing samples.

References

[1]       DiVito JA, Morris GR. Silane pretreatment of mineral aggregate to prevent stripping in flexible pavements, Transportation Research Board; 1982, p. 104–11.

[2]       Martin AE, Rand D, Weitzel D, Sebaaly P, Lane L, Bressette T, et al. Topic 7: Field experience. Natl Semin Moisture Sensit Asph Pavements, Transp Res Board (TRB), Washington, DC, 2003, p. 229–58.

[3]       Sebaaly PE. Comparison of lime and liquid additives on the moisture damage of hot mix asphalt mixtures. National Lime Association; 2007.

[4]       Kutay ME, Aydilek AH, Masad E, Harman T. Computational and experimental evaluation of hydraulic conductivity anisotropy in hot-mix asphalt. Int J Pavement Eng 2007;8:29–43. doi:10.1080/10298430600819147.

[5]       Gandhi T, Xiao F-P, Amirkhanian SN. Estimating indirect tensile strength of mixtures containing anti-stripping agents using an artificial neural network approach. Int J Pavement Res Technol 2009;2:1–12. doi:10.6135/ijprt.org.tw/2009.2(1).1.

[6]       Tarefder RA, Ahsan S, Ahmed MU. Neural Network–Based Thickness Determination Model to Improve Backcalculation of Layer Moduli without Coring. Int J Geomech 2015;15:04014058. doi:10.1061/(ASCE)GM.1943-5622.0000407.

[7]       Xiao F, Amirkhanian SN. Artificial Neural Network Approach to Estimating Stiffness Behavior of Rubberized Asphalt Concrete Containing Reclaimed Asphalt Pavement. J Transp Eng 2009;135:580–9. doi:10.1061/(ASCE)TE.1943-5436.0000014.

[8]       Flood I, Kartam N. Neural Networks in Civil Engineering. I: Principles and Understanding. J Comput Civ Eng 1994;8:131–48. doi:10.1061/(ASCE)0887-3801(1994)8:2(131).

[9]       Mohammadhassani M, Nezamabadi-Pour H, Jumaat M, Jameel M, Hakim SJS, Zargar M. Application of the ANFIS model in deflection prediction of concrete deep beam. Struct Eng Mech 2013;45:323–36. doi:10.12989/sem.2013.45.3.323.

[10]     Arifuzzaman M, Hassan MR. Moisture Damage Prediction of Polymer Modified Asphalt Binder Using Support Vector Regression. J Comput Theor Nanosci 2014;11:2221–7. doi:10.1166/jctn.2014.3630.

[11]     Al-Adham K, Arifuzzaman M. Moisture damage evaluation in carbon nanotubes reinforced asphalts. Sustain Eco-efficiency, Conserv Transp Infrastruct Asset Manag, CRC Press; 2014, p. 103–9. doi:10.1201/b16730-17.

[12]     Amirkhanian AN, Xiao F-P, Amirkhanian SN. Characterization of unaged asphalt binder modified with carbon nano particles. Int J Pavement Res Technol 2011;4:281–6. doi:10.6135/ijprt.org.tw/2011.4(5).281.

[13]     Hassan MR. Modeling of Moisture Damage in Carbon Nano Tube Modified Asphalt Using Hybrid of Artificial Neural Network and Other Computational Intelligence Approaches. J Comput Theor Nanosci 2015;12:4927–34. doi:10.1166/jctn.2015.4461.

[14]     Becker Y, Mendez MP, Rodriguez Y. Polymer modified asphalt. Vis Tecnol 2001;9:39–50.

[15]     Ball P. Roll up for the revolution. Nature 2001;414:142–4. doi:10.1038/35102721.

[16]     Arepalli S, Nikolaev P, Holmes W, Files BS. Production and measurements of individual single-wall nanotubes and small ropes of carbon. Appl Phys Lett 2001;78:1610–2. doi:10.1063/1.1352659.

[17]     Mann S. Nanotechnology and Construction. Nanoforum Report European Nanotechnology Gateway. 2008.

[18]     Tarefder RA, Zaman A. Carbon Nanotube Modified Asphalt Binders for Sustainable Roadways, 2017, p. 623–33. doi:10.1007/978-3-319-41682-3_52.

[19]     Tsoukalas LH, Uhrig RE. Fuzzy and neural approaches in engineering. John Wiley & Sons, Inc.; 1997.

[20]     Jang J-SR. ANFIS: adaptive-network-based fuzzy inference system. IEEE Trans Syst Man Cybern 1993;23:665–85. doi:10.1109/21.256541.

[21]     Cortes C, Vapnik V. Support-vector networks. Mach Learn 1995;20:273–97. doi:10.1007/BF00994018.

[22]     Agatonovic-Kustrin S, Beresford R. Basic concepts of artificial neural network (ANN) modeling and its application in pharmaceutical research. J Pharm Biomed Anal 2000;22:717–27. doi:10.1016/S0731-7085(99)00272-1.