[SESSION INDEX] [AUTHOR INDEX] [PRESENTER INDEX] 22nd ARRB Conference – Research into Practice, Canberra Australia, 2006

APPLICATION OF ARTIFICIAL NEURAL NETWORK (ANN) FOR PREDICTION OF PAVEMENT DETERIORATION FOR LOW VOLUME ROADS IN INDIA Mr. D.T.Thube and Dr. M. Parida, Indian Institute of Technology Roorkee, India Dr. S.S. Jain, Centre of Transportation Engg. (COTE) & Centre for Transportation Systems (CTRANS), Indian Institute of Technology Roorkee, India ABSTRACT The timely identification of undesirable cracks, raveling and rutting conditions is a critical step in pavement management at network level. To date many models have been developed for forecasting pavement condition. The most popular of them in developing countries is the World Bank developed model HDM-4. This paper summarizes the implementation of a pavement condition prediction methodology using Artificial Neural Network (ANN) for three individual ANN models to forecast cracking, raveling and rutting for low volume roads (LVR) in India. Road inventory data as well as five cycles of pavement performance data (pre-monsoon, postmonsoon and during winter season) including various pavement distresses, subgrade characterization and traffic data have been collected from 61 in-service LVR pavement sections in 2004, 2005 and 2006 and each individual ANN model was tested using this data. The modelling results suggest that the ANN models developed in the study satisfactorily forecast future cracking, ravelling and rutting. The performance of the ANN models is also compared with calibrated HDM-4 models using LVR validation sections in the study area.

INTRODUCTION Artificial neural networks employ mathematical simulation of biological nervous systems in order to process acquired information and derive predictive outputs after the network has been properly trained for pattern recognition. A neural network consists of numerous layers of parallel processing elements or neurons. One or more than one hidden layers may exist between an input and an output layer. The neurons in the hidden layers are connected to the neurons of a neighboring layer by weighting factors that can be adjusted during the model training process. The networks are organized according to training methods for specific applications. Figure1 illustrates a three-layer neural network consisting of four neurons in the input layer, four neurons in the hidden layer, and two neurons in the output layer, with interconnecting weighting factors, wij, between layers of neurons. “Training” of an ANN model is a procedure by which ANN repeatedly processes a set of test data (input-output data pairs), changing the values of its weights according to a predetermined algorithm in order to improve its performance. Backpropagation is the most popular algorithm for training ANNs (Lippman 1987). It is a supervised learning method in which an output error is fed backward through the network, altering connection weights so as to minimize the error between the network output and the targeted output. The following equation is used for correcting the weighting factor:

Δwij (n) = α Δwij (n − 1) − ε (∂E / ∂wij )

(1)

Δwij (n) and Δwij (n − 1) are weight increments between nodes i and j during the nth and (n−1)th steps. The momentum factor α is used to speed up the training in flat regions of

Where,

© ARRB Group Ltd and Authors 2006

1

22nd ARRB Conference – Research into Practice, Canberra Australia, 2006

the error surface and helps to prevent oscillations in the weights. A learning rate ε is used to increase the chance of avoiding the training process being trapped in local minima instead of global minima.

Input Layer

Hidden Layer

Output Layer

Weights, wij 1

1 1

2

2

3

3

4

4

Outputs 2

Figure 1: Typical Three-Layer Neural Network

RESEARCH REVIEW Artificial neural networks (ANNs) are valuable computational tools that are increasingly being used to solve resource-intensive complex problems as an alternative to using more traditional techniques. Ceylan et al. (2004) employed ANNs as pavement structural analysis tools for the rapid and accurate prediction of critical responses and deflection profiles of flexible pavements subjected to typical highway loadings. Meier et al. (1997) trained back propagation type ANNs as surrogates for ELP analysis in a computer program for back calculating pavement layer moduli and realized a 42 times increase in processing speed. Similar ANN applications were also reported by Meier and Rix (1995), Gucunski and Krstic (1996), Khazanovich and Roesler (1997), and Kim and Kim (1998). The research project team working on the development of the new mechanistic based AASHTO Pavement Design (NCHRP 1-37A) have also recognized ANN as nontraditional, yet very powerful computing techniques and took advantage of ANN models in preparing the 2002 Design Guide concrete pavement analysis package. In addition, artificial neural networks (Attoh-Okine 1994, 2000, 2001, 2002; Gunaratne and Lu, 2004; Choi, Adams, and Bahia, 2004; Sundin and Braban-Ledoux, 2001; Roberts and Attoh-Okine, 1998; Alsugair and Al-Qudrah, 1998; Huang and Moore, 1997; Eldin and Senouci, 1995; Fwa and Chan, 1993) have recently been used in pavement deterioration, pavement-performance prediction, flexible pavement cracking prediction, and condition ratings of jointed concrete pavements. As explained above, several neural network studies have been conducted to estimate current pavement condition, to predict pavement deterioration, and finally to assist engineers in selecting optimal maintenance and rehabilitation activities. Such applications help the pavement management engineers to choose the best available resource allocation strategies.

METHODOLOGY OF DEVELOPMENT OF ANN MODELS In this study, individual ANN models are developed to predict the value of total cracking, raveling and rut conditions for flexible LVR pavement sections and several ANN architectures are studied to obtain the best results. For the neural network trainings, the back propagation algorithm, which is the most commonly used type of artificial neural networks, is employed and

© ARRB Group Ltd and Authors 2006

2

22nd ARRB Conference – Research into Practice, Canberra Australia, 2006

identification of input parameters used for the trainings and the generation of the ANN training databases are discussed below. Identification of Input and Output Variables From the analysis of the performance model equations in HDM-4 models (Bennett, 2000; Odoki, 2000 and Morosiuk, 2001), it is evident that the pavement’s evolution over time fundamentally depends on four global variables–traffic, pavement age (calculated from the date of construction or most recent rehabilitation), dominant climatic conditions, and structural capacity–variables that help to define the initiation as well as the progression of the distress. They may exhibit together with the interaction among the different manifestations of damage and wear. Road inventory details as well as five cycles of pavement performance data (pre-monsoon, postmonsoon and during winter season) including various pavement distresses, sub grade characterization and traffic data have been collected from 61 in-service LVR pavement sections in year 2004, 2005 and 2006 and database developed from it, which are subsequently used for calibration of HDM-4 as well as developing ANN models. The category of data and variables used to generate the training data sets are selected primarily on the basis of HDM-4 models for related distresses and details of input/output variables for each individual ANN models are given in Table 1. Similiarly, by using the above periodic pavement performance data, HDM-4 pavement deterioration models have been calibrated for LVR in Indian conditions and the details of suggested HDM-4 calibration coefficients are in Table 2 (Thube et al. 2006). ANN Model Architecture The selection of ANN architecture is not a decision making process and most of the time, trial and error combined with engineering judgment is jointly employed to determine the appropriate for a particular problem. In the present study a number of input and output variables were kept constant as described above and variations are done in a number of hidden layers and neurons. The details of twelve ANN model architectures attempted for each cracking, raveling and rut depth progression models in present study are given in Table 3. Training and Testing Set Generation Given the various architecture of ANN models as above in Table 3, the weights of links among the neurons are resolved through the training process and sigmoidaxon function have been used for training. Neurosolution-5 (evaluation version) software has been used in the present study for analysis. The training process involves presenting all example pattern pairs in the training dataset to the network and adjusting the weights of the connections according to the weight adjustment rules. In the present study, training process has been carried out for 10000 epochs. After completion of the training procedure, the trained network is exposed to the testing dataset to check if the training is successful. The testing datasets are fed into the trained ANN and the testing error is calculated. If the testing error is still in an acceptable level, the ANN model is considered reasonable. The details of a number of data points selected for training and testing for various ANN models in the present study are given in Figure 2. The goodness of fit 2 (R ) details for various attempted ANN Architectures and distresses are given in Table 3. The ANN architecture model corresponding to minimum root mean square error (RMSE) and maximum goodness of fit (R2) are finally suggested among the different ANN architectures for each distress types.

© ARRB Group Ltd and Authors 2006

3

22nd ARRB Conference – Research into Practice, Canberra Australia, 2006

Table 1: Details of Input-output variable in different ANN Models Model Description

Cracking progression Model

Ravelling progression Model

Rut depth progression Model

ANN Input variable No.

26

26

29

Description of variables Age in months; Initial cracking area ;sub grade properties-LL, PL, PI, Field moisture content, OMC,CBR(soaked),MDD;CRP; AADT detailsMT,NMT,% of Truck volume; composition of commercial vehicles; Environmental factors-% duration of dry season, mean monthly precipitation, mean temperature degrees, average temperature range, no of days having temperature> 32 degrees; Geometric properties - R+F (m/km),horizontal curvature (degree/km), speed limit, no. of R+F per km; other parameters-CDS,CDB and CRP ( as defined in HDM-4 Models) Age in months; Initial ravelling area ;sub grade properties-LL,PL,PI,Field moisture content, OMC,CBR(soaked),MDD;CRP; AADT detailsMT,NMT,% of Truck volume; composition of commercial vehicles; Environmental factors-% duration of dry season, mean monthly precipitation, mean temperature degrees, average temperature range, no of days having temperature > 32 degrees; Geometric properties - R+F (m/km),horizontal curvature (degree/km), speed limit, no of R+F per km ; other parameters-CDS,CDB and CRP (as defined in HDM-4 Models) Age in months, observed cracking, ravelling, edge break and pothole area during analysis period ,sub grade properties-LL, PL, PI, Field moisture content, OMC, CBR(soaked), MDD; CRP; AADT details-MT,NMT,% of Truck volume; composition of commercial vehicles; Environmental factors -% duration of dry season, mean monthly precipitation, mean temperature degrees, average temperature range, no of days having temperature > 32 degrees; Geometric properties-R+F (m/km), horizontal curvature(degree/km), speed limit, no of R+F per km ; Other parameters-CDS,CDB and CRP ( as defined in HDM-4 Models)

© ARRB Group Ltd and Authors 2006

ANN Output variable

Total cracking (% area)

Total ravelling (% area)

Total Rut depth (in mm)

4

22nd ARRB Conference – Research into Practice, Canberra Australia, 2006

Table 2: Details of suggested HDM-4 Pavement Deterioration Calibration Factors for Low Volume Roads in India Total Cracking

Raveling

Rut Depth

Progression

Progression

Progression

Kcpa

Kvp

Krst

Plain

0.23

0.34

2.7

Rolling

0.23

0.27

2.17

Mountainous

0.23

0.54

1.5

Average for Study area

0.227

0.381

2.122

Terrain Types

Table 3: Details of Different ANN Model Architectures Goodness of fit details at Testing Number Number ANN Model Cracking Ravelling Rut Depth of Architecture of Progression Progression Progression Hidden Number Neurons Layers RMSE R2 RMSE R2 RMSE R2 1

2

4

3.07

0.98

6.10

0.99

2.24

0.89

2

2

5

2.95

0.98

6.32

0.99

2.12

0.90

3

2

6

3.14

0.98

6.65

0.99

2.54

0.88

4

2

7

3.09

0.91

5.51

0.99

2.89

0.88

5

3

4

2.51

0.99

6.81

0.99

2.14

0.90

6

3

5

2.17

0.98

7.37

0.99

2.30

0.89

7

3

6

2.17

0.99

5.81

0.99

2.59

0.88

8

3

7

2.87

0.98

7.72

0.99

2.84

0.88

9

4

4

2.50

0.99

6.27

0.99

1.97

0.90

10

4

5

2.72

0.99

6.67

0.99

2.90

0.87

11

4

6

2.61

0.99

7.17

0.99

2.40

0.89

12

4

7

3.14

0.98

5.94

0.99

2.35

0.89

Suggested ANN Model Architecture Number

© ARRB Group Ltd and Authors 2006

7

4

9

5

22nd ARRB Conference – Research into Practice, Canberra Australia, 2006

300 Training

No of Data Points

250

Testing

224

Validation

248

224

200 150 100 50

94 61

61 20

30

20

0 Cracking Model

Ravelling Model

Rut depth Model

Pavement Deterioration Model Type Figure 2: Details of Data points for ANN Model Development

VALIDATION OF ANN MODELS After training and testing, the last and also the most critical step is to verify the model using a validation dataset. In the present study, LVR sections (different than of model developments) are selected and pavement performance data as per requirements for ANN and HDM-4 models are collected for these roads. Similarly, the details of observed cracking, raveling and rut depth for these roads are collected by carrying out visual distress survey. Predictions of cracking, raveling and rut depth are carried out by using the suggested and trained ANN models as given in Table 3, as well as using HDM-4 models and calibration coefficients as given in Table 2. Scatter plots have been plotted between observed vs. HDM-4 and ANN predicted distress and the details of scatter plots for cracking, raveling and rut depth distresses are given in Figs.3 to 5.The linear relationships and goodness of fit parameters like (R2) and root mean square error (RMSE) are calculated between observed vs. HDM-4 predicted and ANN predicted distresses for cracking, raveling and rut depth progressions and the details of the same are given in Table 4.

© ARRB Group Ltd and Authors 2006

6

22nd ARRB Conference – Research into Practice, Canberra Australia, 2006

Predicted Cracking ( % area)

40 Observed vs HDM-4 Predicted Observed vs ANN Predicted

30

20

10

0 0

10

20

30

40

Observed Cracking ( % area)

Figure 3: Scatter Plot Observed vs. ANN Model predicted and HDM-4 Model predicted Cracking

Predicted Ravelling (% area)

50 Observed vs HDM-4 Predicted Observed vs ANN-Predicted.

40

30

20

10

0 0

10

20

30

40

50

Observed Ravelling (% area)

Figure 4: Scatter Plot Observed vs. ANN Model predicted and HDM-4 Model predicted Ravelling

© ARRB Group Ltd and Authors 2006

7

22nd ARRB Conference – Research into Practice, Canberra Australia, 2006

30 Observed vs HDM-4 Predicted Observed vs ANN Predicted

Predicted Rut Depth (in mm)

25 20 15 10 5 0 0

5

10

15

20

25

30

Observed Rut Depth (in mm)

Figure 5: Scatter Plot Observed vs. ANN Model predicted and HDM-4 Model predicted Rut depth Table 4 : Details of Linear Relationships and statistical parameters between Observed vs HDM-4 and ANN Predicted distresses Sr. No.

Model Description

1

Cracking Progression Observed vs. HDM-4 Predict.

y = 1.1152x - 0.0768

0.97

1.73

Cracking Progression Observed vs. ANN Predict.

y = 1.1152x - 0.0768

0.98

0.93

Ravelling Progression Observed vs. HDM-4 Predict.

y = 1.1689x - 1.2155

0.97

4.06

Ravelling Progression Observed vs. ANN Predict.

y = 0.8694x + 1.1964

0.97

2.78

Rut depth Progression Observed vs. HDM-4 Predict.

y = 0.8906x + 1.2683

0.92

1.78

0.94

1.60

2

3

Linear relationship details

Rut depth Progression y = 0.9194x + 0.382 Observed vs. ANN Predict. (Note: y- predicted distress; x- observed distress)

R2

RMSE

CONCLUSIONS The following conclusions have been made based upon the study results: (i)

A unified ANN based model has been suggested for the prediction of cracking, ravelling and rutting progression for low volume roads in India.

(ii)

The suggested unified ANN model will be more useful than the HDM-4 models as the HDM-4 models require separate local calibrations for plain, rolling and mountainous terrains in the study area.

© ARRB Group Ltd and Authors 2006

8

22nd ARRB Conference – Research into Practice, Canberra Australia, 2006

(iii)

Different ANN model architectures were examined by the carrying out of various trials. The architecture corresponding to a minimum root mean square error at the testing stage has been suggested.

(iv)

The suggested ANN models showed a high goodness of fit (R2 value) of greater than 0.99 for cracking and ravelling distresses, and greater than 0.9 for rut depth progression at the testing stage.

(v)

The suggested ANN models also shows a high goodness of fit (R2 value) of greater than 0.94 for all cracking, ravelling and rut depth progression at the validation stage.

(vi)

The suggested ANN models shows high goodness of fit (R2) than that of HDM-4 predicted distresses and hence proves the efficacy of ANN models over HDM-4 models in prediction of distresses.

(vii)

The suggested ANN models shall be useful for the accurate prediction of cracking, ravelling and rut depth for low volume roads in India and assist in the development of maintenance strategies for low volume raods in India generally.

(viii)

The suggested models will also assist in the development of pavement management systems for low volume roads.

REFERENCES Ceylan, H. (2002). “Analysis and design of concrete pavement systems using artificial neural networks.” Ph.D. Dissertation, University of Illinois at Urbana-Champaign, December. Meier, R.W., Alexander, D.R., and Freeman, R. (1997). “Using artificial neural networks as a forward approach to backcalculation.” Transportation Research Record No. 1570, TRB National Research Council, Washington, D.C., pp. 126-133. Gucunski, N. and Krstic, V. (1996). “Backcalculation of pavement profiles from spectralanalysisof-surface waves test by neural networks using individual receiver spacing approach.” Transportation Research Record No. 1526, TRB, National Research Council, Washington, D.C., pp. 6-13. Khazanovich, L., and Roesler, J. (1997). “DIPLOBACK: Neural-network-based backcalculation program for composite pavements.” Transportation Research Record No. 1570, TRB, National Research Council, Washington, D.C., pp. 143-150. Long Term Pavement Performance Information Management System (2004): Pavement Performance Database User Reference Guide. Attoh-Okine, N. O. (1994). ‘‘Predicting roughness progression in flexible pavements using artificial neural networks.’’ Proc., 3rd Int. Conf. Managing Pavements, San Antonio Tex., 52–62. Attoh-Okine, N. O. (1997). ‘‘Rough set application to data mining principles in pavement management database.’’ J. Comput. Civ. Eng., 11(4), 231–237. Attoh-Okine, N.O. and Appea, A. (1998). “Predicting roughness progression models in flexible pavements - an evolutionary algorithm approach.” Intelligent Engineering Systems Through Artificial Neural Networks, 8, 845-853. Attoh-Okine, N. O. (1999). ‘‘Analysis of learning rate and momentum term in backpropagation neural network algorithm trained to predict pavement performance.’’ Adv. Eng. Software, 30(4), 293–302. Attoh-Okine, N.O. (2001).”Grouping pavement condition variables for performance modeling using self-organizing maps.” Computer Aided Civil and Infrastructure Engineering, 16(2),122-25.

© ARRB Group Ltd and Authors 2006

9

22nd ARRB Conference – Research into Practice, Canberra Australia, 2006

Attoh-Okine, N. O. (2002). “Combining use of rough set and artificial neural networks in doweled-pavement-performance modeling-A hybrid approach” J. Transportation Engineering, ASCE, 128(3), pp 270-275. Mei, X. Gunaratne, M., Lu, J.J.,and Dietrich, B. (2004). “Neural network for rapid depth evaluation of shallow cracks in asphalt pavements.” Computer-Aided Civil and Infrastructure Engineering, 19 (3), pp. 223-230. Choi, J., Adams, T.M., Bahia, H. U. (2004).”Pavement roughness modeling using backpropagation neural networks” Computer-Aided Civil and Infrastructure Engineering, 19 (4), p 295-303 Sundin, S. and Braban-Ledoux, C. (2001). “Artificial intelligence-based decision support technologies in pavement management.” Computer-Aided Civil and Infrastructure Engineering, V 16, n 2, p 143-157. Roberts, C. A. and Attoh-Okine, N. O. (1998). ‘‘A comparative analysis of two artificial neural networks using pavement performance prediction.’’ Comput.-Aided Civil Infrastruct. Eng., 13, 339–348. Alsugair, A.M. and Al-Qudrah, A.A. (1998). “Artificial neural network approach for pavement maintenance.” J. Computing in Civil Engineering, ASCE, 12 (4), 249-55. Huang, Y., and Moore, R. K. (1997). ‘‘POS distress level probability prediction in PMS using ANNs’’. Paper No. 97-0419, presented at Transportation Research Board, Washington, D.C. Fwa, T. F.,and Chan, W. T. (1993). ‘‘Priority rating of highway maintenance need by neural networks’’. J. Transportation Engineering, ASCE, 119(3), 419–432. Lippman, R. 1987). “An introduction to computing with neural nets.” IEEE ASSP Mag., 4, 4–22. Thube, D.T., Jain,S.S. and Parida,M. (2006) “Development of Pavement Deterioration Models (Calibrating HDM-4) for Low Volume Roads in India”, 10th International Conference on Asphalt Pavements to held in August at Quebec City, Canada (accepted for publications)

ACKNOWLEDGEMENT The inputs received from Department of Science and Technology (DST) Sponsored Research Project Titled “Design Options Based on Performance Analysis of Rural Roads in Uttaranchal” for the preparation of this paper is thankfully acknowledged.

AUTHOR BIOGRAPHIES D.T. Thube is working as Executive Engineer, Public Works Department, Maharashtra State in India and presently a Research Scholar at IIT Roorkee, Roorkee. He has 15 years of field experience and 3 year of research. Professor S.S. Jain is presently heading the Centre of Transportation Systems of IIT Roorkee and obtained his B.E., M.E. & Ph.D. degree from University of Roorkee. He joined the faculty of University of Roorkee (now IIT, Roorkee) in 1979 and became Professor in 1996. He is member, Editorial Board, Journal of Institution of Civil Engineers (UK) & Member Technical Committee, World Roads Congress (PIARC), Paris, Member, Highway Standards & Specification Committee, Member, STAC, Ministry of Surface Transport & Highways, Govt.of India. Dr. M.Parida is an Associate Professor in Transportation Engineering in Department of Civil Engineering, IIT Roorkee. In recognition of his research contribution, he has been conferred

© ARRB Group Ltd and Authors 2006

10

22nd ARRB Conference – Research into Practice, Canberra Australia, 2006

with distinguished Award, medals & prizes such as Pt. Jawaharlal Nehru Birth Centenary Award 2002, Musaddilal Memorial Award, University of Roorkee, Khosla Research Commendation Award 2000, U.P. State Centre, Lucknow and IRC, Nawab Zain Yar Jung Bahadur Memorial Medal and Best Technical Paper Award by the Institution of Engineers (India)-2000. Dr. Parida is presently Council Member of Indian Roads Congress.

© ARRB Group Ltd and Authors 2006

11