See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/275816379

Artificial Neural Network Model for Prediction

of Bearing Capacity of Driven Pile

Article · April 2015

CITATION READS

1 64

3 authors:

Harnedi Maizir Nurly Gofar

Sekolah Tinggi Teknologi Pekanbaru Nanyang Technological University
13 PUBLICATIONS 71 CITATIONS 42 PUBLICATIONS 159 CITATIONS

SEE PROFILE SEE PROFILE

Khairul Anuar Kassim

Universiti Teknologi Malaysia
158 PUBLICATIONS 532 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Intelligent prediction for evaluation of single pile bearing cpaacity using Pile Dricving Analyzer (PDA)
datas View project

Intelligent prediction system on the structural response of multi-storey building based on Indonesian
seismic response spectrum loading View project

All content following this page was uploaded by Harnedi Maizir on 04 May 2015.

The user has requested enhancement of the downloaded file.

Maizir, et. al
ISSN 0853-2982

Jurnal Teoretis dan Terapan Bidang Rekayasa Sipil

Artificial Neural Network Model for Prediction of Bearing Capacity of

Driven Pile
Harnedi Maizir
Jurusan Teknik Sipil, Sekolah Tinggi Teknologi Pekanbaru, Jl. Dirgantara No. 4 Pekanbaru
E-mail: [email protected]
Nurly Gofar
Fakultas Teknik, Universitas Sriwijaya, Bukit Besar Palembang 30139
E-mail: [email protected]
Khairul Anuar Kassim
Faculty of Civil Engineering, Universiti Teknologi Malaysia, 81310 UTM Skudai, Johor, Malaysia
E-mail: [email protected]

Abstract

This paper presents the development of ANN model for prediction of axial capacity of a driven pile based on Pile
Driving Analyzer (PDA) test data. As many as 300 sets of high quality test data from dynamic load test performed at
several construction projects in Indonesia and Malaysia were selected for this study.Input considered in the model-
ing are pile characteristics (diameter, length as well as compression and tension capacity), pile set, and hammer
characteristics (ram weight, drop height, and energy transferred).An ANN model (named: ANN-HM) was developed
in this study using a computerized intelligent system for predicting the total pile capacity as well as shaft resistance
and end bearing capacity for various pile and hammer characteristics. The results show that the ANN-HM serves as
a reliable prediction tool to predict the resistance of the driven pile with coefficient of correlation (R) values close to
0.9 and mean squared error (MSE) less than 1% after 15,000 number of iteration process.

Keywords :Axial capacity, Shaft resistance, End bearing, PDA, ANN.

Abstrak

Makalah ini menyajikan pengembangan model ANN untuk prediksi kapasitas daya dukung axial tiang pancang
berdasarkan data uji Pile Driving Analyzer (PDA). Sebanyak 300 set data uji dari uji beban dinamis yang
dilakukan pada beberapa proyek konstruksi di Indonesia dan Malaysia dipilih untuk penelitian ini. Variabel bebas
yang digunakan adalah karakteristik tiang pancang (diameter, panjang serta kapasitas tekan dan tarik), set, dan
karakteristik palu penumbuk tiang (berat palu, tinggi jatuh dan energi yang ditransfer). Model ANN (yang
dinamakan: ANN-HM) dikembangkan dalam penelitian ini menggunakan intelligent system dalam ANN untuk
memprediksi daya dukung tiang total yang didistribusikan kepada tahanan ujung dan tahanan sisi untuk berbagai
jenis tiang dan palu penumbuk tiang. Hasil penelitian menunjukkan bahwa ANN-HM dapat diandalkan untuk
memprediksi daya dukung tiang pancang dengan koefisien korelasi (R) mendekati 0,9 dan rata-rata kesalahan
kuadrat (MSE) kurang dari 1 % setelah 15.000 kali proses iterasi.

Kata-kata Kunci: Kapasitasaksial, Tahanangeser, Tahananujung, PDA, ANN

1. Introduction Several researchers have attempted using ANN applica-

tion is to develop optimal models using simple data
Artificial neural networks (ANN) have been adopted obtained from PDA test. For example, Lee and Lee
used by researchers as a tool for the development of (1996) simulated data obtained from model pile using
predictive models on various geotechnical problems in-situ pile load test. The output of the model predic-
including bearing capacity of piles. The neural net- tion is the ultimate bearing capacity while the input
works in ANN model attempt to simulate the process includes penetration depth of pile, diameter of pile and
by which human brain learns to discern patterns in the mean normal stress. The study shows that the ANN
arrays of data. The ability of the neural network to model gives a maximum error not more than 25%.
learn complex association by examples makes it an Similarly, Abu Kiefa(1998) introduced a general
attractive method for estimating pile capacities based regression neural network for predicting the capacity of
on stress-wave measurements in pile driving analyzer
(PDA) test.
Vol. 22 No. 1 April 2015 49
 Artificial Neural Network Model for Prediction of Bearing Capacity of Driven Pile

a driven pile in cohesionless soil. Five variables were with the static load test, more dynamic test can be
selected as input data: angle of friction of soil, effective carried out as quality control to save cost and expedite
overburden pressure, pile length, and cross sectional construction in large project.
area of the pile. The results show that ANN model
gives a better prediction when compared with analytical Following Teh (1997), this model was developed based
and empirical methods such as Meyerhoff and Coyle & on the capacity obtained from CAPWAP analysis
Castello. Nawari et al. (2005) also studied a method for rather than the true bearing capacity of piles. As many
design and analysis of deep foundation using an artifi- as 300high quality load test records were acquired from
cial intelligence technique utilizing Back-propagation projects in Indonesia and Malaysia. These data was
Neural Network (BPNN) models. The inputs of the obtained from projects where at least one SLT was
network for training and testing are N-SPT value and performed, thus correlation was applied to the data
pile dimension. During the training phase, the measured before processing.
axial pile capacities were compared with the capacities
obtained by BPNN. The developed neural network An ANN system (named ANN-HM) was developed by
models were capable of reproducing the target output a computerized intelligent system based on these data
values with minimal error. Furthermore, Teh(1997) also for predicting the total axial bearing capacity of the
proposed a neural network for estimating the static pile single pile (Qt), as well as shaft resistance capacity (Qs)
capacity determined from dynamic stress-wave data and end bearing capacity (Qe) based on selected pile
(PDA) combined with Case Pile Wave Analysis and hammer characteristics. The performance of the
Program (CAPWAP) software for square precast rein- system was evaluated until the final model characteris-
forced concrete piles. The study was concerned with tics were achieved.
predicting CAPWAP output rather than true bearing
capacity of the pile. They used back-propagation neural 2. Artificial Neural Network (ANN)
networks (BPNN) for their analysis. The predicted
value gives a small error of about 10% when compared ANNs are mathematical inventions inspired by the
to actual data. Park and Cho (2010) applied the ANN function of the human brain and nervous system. The
model to predict the resistance of the driven pile in ANNs must be trained by inputting data repeatedly
dynamic load test. One hundred and sixty five(165) together with corresponding target outcomes. After
data from dynamic piles load test at various sites in sufficient numbers of training iterations, the ANNs
Korea were selected to develop the model. The results learn to recognize patterns of the data, hence; creating
indicate that the ANN model served as a reliable and an internal model of the governing data process. The
simple predictive tool to predict the resistance of the ANNs use this internal model to make predictions for
driven pile with R2 values close to 0.9. new input conditions. The input variables are called
neurons, and neurons send signals to other neurons.
This paper presents the application of the ANN for pre- Some input to the neuron may have greater importance
diction of axial capacity of driven pile by adopting than the other, and this is modeled by weighting the
PDA data collected from several construction projects input to the neuron. Thus, the neuron can be thought of
in Indonesia and Malaysia involving PDA test with as a small computing engine that takes in input,
CAPWAP analysis. The PDA combined with CAP- processes them, and then transmits an output.
WAP software is the most widely employed system for
High Strain Dynamic Pile Test (HSDPT) due to its sim- Literature study (Priddy and Keller, 2007) indicates
plicity and quick handling of the data. The procedure is that the most popular neural networks model is the
standardized in ASTM-D4945-08 (2008). Detail expla- Back-Propagation Neural Networks (BPNN). BPNN
nation of the analysis of pile driving formulas and wave consists of three inter-connected group of layers, i.e.
equation analysis adopted in PDA test and CAPWAP input layer, hidden layer and output layer. The purpose
analysis can be found elsewhere (e.g. Coduto (2001)). of BPNN training is to change iteratively the weights
of each neuron in a direction that minimizes the error.
The ability of CAPWAP procedure to accurately predict The number of hidden neurons is important in the
static capacity from dynamic pile testing has been BPNN. However, there is no exact method for deter-
demonstrated by Robert (1994) and Likins (2004) indi- mining the number for hidden layer neurons. As
cating a good correlation between CAPWAP prediction described by Nawari, et. al (2005), too many hidden
of pile capacity and Static Load Test (SLT) results. layers make the training time longer, but few hidden
On the contrary, Lok and Che (2004) indicated that the layers make learning algorithm be trapped in local
accuracy of HSDPT output relies mainly on the input minimum. Determining the suitable number of hidden
parameters such as hammer efficiency and damping layer requires trial and error process. Berke and Hajela
factor. Similar study was done by Gofar and Angelo (1991) suggested that the rough estimations of the best
(2006) on bored piles in Malaysia and they concluded number of nodes on the hidden layer should be
that the relationship between the results of SLT and between average and the sum of the nodes in the input
PDA test is site specific;hence only after correlation and output layers.

50 Jurnal Teknik Sipil

Maizir, et. al

Various transfer functions should be investigated to 3. Data Collection

achieve the best performance in training as well as in
testing steps. The optimal performance was obtained The data used for this study was collected from various
from tan-sigmoid-log sigmoid-log sigmoid-linear acti- projects involving dynamic pile tests in Indonesia and
vation functions in the first, second, third and output Malaysia. The tests were performed using PDA test
layers respectively. ANNs have the ability to model the based on ASTM D 4945-08 and CAPWAP software for
non-linear relationship between a set of input variables analysis of bearing capacity. Only high quality data
and the corresponding outputs without the need for from on concrete pile were used in the study. Three
predefined mathematical equations. hundred sets of high quality test data from dynamic
load test were selected for the subsequent study; as
Furthermore, ANNs do not need prior knowledge about many as 188 load test records were acquired from pro-
the nature of the relationship between the model inputs jects in Indonesia while the other 112 load test records
and corresponding outputs. ANNs use the data alone to were collected in Malaysia.
determine the structure of the model, as well as the
unknown model parameters. This enables ANNs to The spun piles used in this study are circular with
overcome the limitations of the existing methods. diameter between 300 and 500 mm and square with
sides between 200 and 400 mm. Equivalent diameter
Compared to traditional methods, ANNs tolerate rela- was calculated for the square piles based on the area of
tively imprecise, noisy or incomplete data. the pile cross section.
The approximate results are less vulnerable to outliers,
have better filtering capacity, and are more adaptive. As suggested by previous researchers, better selection
Moreover, ANNs also use massively parallel processes; of input data will significantly improve the accuracy of
therefore the numerous independent operations can be prediction; thus, in this research, the piles are grouped
executed simultaneously. into small, medium, and large diameter piles. There are
25 (12%) small (diameter less than 200mm) piles, 131
The most important process for artificial neural net- (63%) medium sized (diameter between 200mm and
works model is dividing the available data into two sets 400mm) piles and 53 (25%) large piles (diameter >
of data. As mentioned by Shahin (2010), the available 400mm) (Table 1).
data are randomly divided by two statistically
consistent sets: a training set for model calibration, and Table 1. Grouping of pile data based on diameter
an independent testing set for model verification. In Group Diameter Number of Percentage
this study, the training and testing procedure of the pile
Large pile > 400 mm 53 25%
ANN model to predict the ultimate bearing capacity of
Medium 200 – 400 mm 131 63%
the single driven pile is mentioned in Figure (1). pile
Small pile < 200 mm 25 12%

Figure 1. The procedure of training ANNs model (Zhao (2010))

Vol. 22 No. 1 April 2015 51

 Artificial Neural Network Model for Prediction of Bearing Capacity of Driven Pile

Eight parameters were selected as input parameters for Database is prepared for training and testing procedure
the subsequent study i.e. pile equivalent diameter (D), of the proposed neural network. The input and target
embedment length (L), compression stress (CS), output variables were normalized in values ranged
tension stress, (TS), vertical displacement or set between 0 and 1 before training, then data are ready to
(DFN), ram weight, (WH), drop height, (DH) and be used for the training and testing with ANN model.
energy transferred (EMX). The target output variables Rescaling is often accomplished by using a linear inter-
for this study are the total axial bearing capacity of the polation formula, as mentioned by Priddy and Keller
single pile (Qt) which distributed to shaft resistance (2007) given by Equation (1) :
capacity (Qs), and end bearing capacity (Qe). These
( − )
target values were obtained from the output of CAP- ′
= (1)
WAP analysis which was adjusted to the results of SLT ( − )
test on the related projects. where:
xi’ : value of data normalization
4. Modeling Process maxvalue : maximum data value
minvalue : minimum data value
Based on the collected data, the neural network model
developed in this study uses 8 (eight) nodes in the The training was performed in this study using back-
input layer (D, L, CS, TS, DFN, WH, DH, and EMX), propagation algorithm because BPNN can provide
8 (eight) nodes in the hidden layer (q1, q2,…..q8), and accurate predictions to any continuous function with
three nodes in the target output (Qt, Qe, and Qs). sufficient neurons. The target of ANN model is not the
Figure 2 shows the architecture of the complete best training data, but the best responds of the training
network for this study. data to be used for testing data. The selection of hidden
nodes size should be done carefully to get the good
reliable model. Hecht-Nelson (1980) used Kolmogo-
rov’stheorem in which any function of input (i) varia-
bles may represent by the required number of hidden
nodes (h) or h = 2i + 1. The value of mean squared
error (MSE) is a goal of the analysis and this value
varies with the coefficient of determination R2 of the
testing result (Park and Cho (2010). The R2 value de-
scribed the contribution of input value in predicting the
target output value, which means the lowest MSE value
gives the better prediction of axial bearing capacity
of driven pile.
Once the ANN model is evaluated in training phase,
the measured axial pile bearing capacities are com-
pared with the capacities obtained by BPNN. The de-
veloped ANN model should be capable of reproducing
the target output values with minimum error. After that,
the result of the trained model is used for a new set of
data and being examined again. The iterative process
was performed until the training, testing and validation
phases are successfully accomplished. Finally, the de-
veloped ANN model can be used as a tool for predict-
ing axial bearing capacity of single driven pile.
The subsequent step in the development of ANN model
is to determine the number of hidden layers and the
number of hidden nodes in each hidden layer. In this
study, the optimal ANN model is obtained by a trial
and error in which the ANN model is trained using one
hidden layer with 2, 3, 4,…8 hidden layer nodes, re-
spectively. The procedure of using the ANN model for
predicting the ultimate pile bearing capacity of the sin-
gle driven pile is explained in Figure (3). The ANN
output data, should be turned back to the actual value
data (un-normalized step), and then, the final output
Figure 2. The architecture of the developed ANNs
Model predicted by the ANN result can be compared with the
target output from PDA test result. This is called the
testing and validation process.
52 Jurnal Teknik Sipil
Maizir, et. al

The neural network model was developed by using the the training, testing and validation phase and the
software Matlabver 7.10.0.499 (R2010a) licensed un- measured bearing capacity using CAPWAP is
der UniversitiTeknologi Malaysia (UTM). During presented in Figure 3. It should be noted that the
analysis, the 300 test data acquired from construction coefficient of correlation (R) between the prediction
projects were divided into three categories:training set and measured value of total bearing capacity is quite
data (70% or 208); testing set data (15% or 41), and high i.e. 0.9377 after the validation process which gives
validation set data (15% or 41). The training, testing a coefficient of determination (R2) value of 0.88.
and validation processes were performed for 20,000 Figure 4 and 5respectively present the same
iterations with interval of 5,000 iterations. Table (1) comparison for haft resistance and end bearing
represents the performance of the model in terms of capacity. The figures suggest that the prediction using
coefficient of correlation (R) and mean squared error ANN-HM model is remarkably in good agreement with
(MSE) for different iteration number during the train- actual data especially for total bearing capacity.The
ing, testing and validation process. Analysis of output good combination of the experimentation indicated the
indicates that the best results were obtained for the highest coefficient of correlation (R) and lowest error
conditions given in Table (2). rate (MSE) for the training, testing and validation
phase.
5. Performance of the ANN-HM Model
The ANN model developed in this study to predict the
axial bearing capacity of driven pile for conditions
mentiond in Table 2 is subsequently referred as ANN-
HM. Comparisonbetween thepredicted values of total
bearing capacity of pile by ANN-HM model during

a. Training (R=0.91855) b. Testing (R=0.89361)

a. Validation (R=0.82333)
Figure 3. Comparison of predicted and measured total bearing capacity

Vol. 22 No. 1 April 2015 53

 Artificial Neural Network Model for Prediction of Bearing Capacity of Driven Pile

Table 2. Summary of ANN-HM model performance for different iteration number

Iteration Coefficient of correlation (R) Mean squared error (MSE)
Number Training Testing Validation Training Testing Validation
Total capacity 5,000 0.8766 0.7746 0.9507
Shaft resistance 5,000 0.7292 0.5688 0.7937 0.0104 0.0111 0.0072
End bearing 5,000 0.7634 0.7433 0.8791
Total capacity 10,000 0.9107 0.8862 0.8124
Shaft resistance 10,000 0.8268 0.8199 0.6399 0.0081 0.0096 0.0079
End bearing 10,000 0.8013 0.8013 0.7903
Total capacity 15,000 0.8894 0.9539 0.9377
Shaft resistance 15,000 0.8615 0.8367 0.8852 0.0075 0.0059 0.0078
End bearing 15,000 0.7815 0.9308 0.9098
Total capacity 20,000 0.8907 0.8429 0.9327
Shaft resistance 20,000 0.8655 0.7429 0.8589 0.009 0.0076 0.0064
End bearing 20,000 0.7862 0.8129 0.8111
Table 3. Characteristics of the developed ANN model
Number of iteration 15,000
Learning rate 0.15
Target error 0.005
Number of hidden layer 1
Number of hidden node 8

a. Training (R=0.83796) b. Testing (R=0.84201)

c. Validation (R=0.64696)
Figure 4. Comparison of predicted and measured Shaft Resistance

54 Jurnal Teknik Sipil

Maizir, et. al

a. Training (R=0.82459) b. Testing (R=0.86301)

c. Validation (R=0.80555)

Figure 5. Comparison of predicted and measured end bearing capacity

Vol. 22 No. 1 April 2015 55

 Artificial Neural Network Model for Prediction of Bearing Capacity of Driven Pile

6. Conclusions Likins, G., 2004, Correlation of CAPWAP with Static

Load Tests.Kuala Lumpur: Proc. 7th Interna-
Based on the ANN model (HM) results, the following tional Conference on the Application of Stress
conclusion can be drawn. wave Theory to Piles.

1. The ANN-HM model gives a good prediction of Lee, I.M. and Lee, J.H.,1996,Prediction of Pile
axial bearing capacity of driven pile which is Bearing Capacity Using Artificial Neural
implied by the high coefficients of correlation (R) Networks. Computers and Geotechnics, 18(3):
during the training, testing and validation phases. 189-200.

2. The result of mean squared error (MSE) gives more Lok, T.M.H. and Che, W.F., 2004, Axial Capacity
conservative value prediction of axial bearing Prediction for Driven Piles Using ANN: Model
capacity compare with the result from CAPWAP Comparison., ASCE.
analysis.
Nawari, N.O., Liang, R., and Nusairat, J., 2005,
3. The ANN-HM serve as a reliable prediction tools to Artificial Intelligence Techniques for the
predict the resistance of the driven pile (including Design and Analysis of Deep Foundations.
shaft resistance and end bearing capacity) with Engineering Journal of Geotechnical Engi-
coefficient of correlation (R) values close to 0.9 and neering.
mean squared error (MSE) less than 1% after 15,000
number of iteration process. Note this model Park, H.I. and Cho, C.W., 2010, Neural Network
considers only pile and hammer characteristic while Model for Predicting the Resistance of Driven
soil parameters are considered similar. Pile. Marine Georesources and Geotechnology,
28: 324-344.
Acknowledgement
Priddy, K.L. and Keller, P.E.,2007, Artificial Neural
The PDA test data used in this study are obtained with Network, An Introduction. New Delhi:
permission from PT. Delta Systech Indonesia and Prentice-Hall
Transwaja Technology Company Malaysia.
Robert, Y., 1994, A New Approach to the Analysis of
High Strength Dynamic Pile Loat Test Data.
References
Canadian Geotecnical Engineering Journal.
Abu Kiefa, M.A.,1998, General Regression Neural Net- 31(2) 246-253
works for Driven Piles in Cohessionless Soils.
Shahin, M.A., 2010, Intelligent Computing for
Journal Geotechnicaland Geoenvironmental.
Modeling Axial Capacity of Pile Foundation.
Engineering, ASCE.124(12): 1177 - 1185.
Canadian Geotech Engineering Journal.47:
ASTM-D4945-08, 2008, Standard Test Method for 230 - 243.
High-Strain Dynamic Testing of Piles, ASTM
Teh, C.I., Wong, K., Goh, A and Jaritngam, S., 1997,
International, West Conshohocken, PA.
Prediction of Pile Capacity Using Neural Net-
Berke, L. and Hajela, P., 1991, Application of Neural works. Computing in Civil Engineering,
Nets in Structural Optimisation. NATO/ April:129-138
AGARD Avanced Study Institute, Germany, 23 Zhao, J., Jiewen, T., and Yongqiang, S., 2010,
(I-II): 731-745. An ANN Model for Predicting Level Ultimate
Bearing Capacity of PHC Pipe Pile. Earth and
Coduto, D.P., 2001, Foundation Design : Priciples and Space : Engineering, Science, Construction
Practice. New Jersey: Prentice Hall Inc. and Operations in Challenging Environments
Gofar, N. and Angelo, M., 2006, Evaluation of Design
Capacity of Bored Piles based on High Strain
Dynamic Test. Jakarta: Proceedings PIT HATTI
X.

Hecht-Nelson, R., 1987, Kolmogorov's Mapping

Neural Network Existem Theorem. In Proc. of
1st IEEE Annual Int'l Conf. on Neural
Networks. San Diego, NJ, USA.

56 Jurnal Teknik Sipil

View publication stats