You are on page 1of 12

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

net/publication/327078266

Prediction of Bearing Capacity of Stone Columns Placed in Soft Clay Using


SVR Model

Article · August 2018


DOI: 10.1007/s13369-018-3513-7

CITATIONS READS

0 11

2 authors:

Manita Das Ashim Kanti Dey


National Institute of Technology, Silchar National Institute of Technology, Silchar
5 PUBLICATIONS   0 CITATIONS    13 PUBLICATIONS   56 CITATIONS   

SEE PROFILE SEE PROFILE

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

PhD project View project

All content following this page was uploaded by Manita Das on 19 August 2018.

The user has requested enhancement of the downloaded file.


Arabian Journal for Science and Engineering
https://doi.org/10.1007/s13369-018-3513-7

RESEARCH ARTICLE - CIVIL ENGINEERING

Prediction of Bearing Capacity of Stone Columns Placed in Soft Clay


Using SVR Model
Manita Das1 · Ashim Kanti Dey1

Received: 26 November 2017 / Accepted: 9 August 2018


© King Fahd University of Petroleum & Minerals 2018

Abstract
It is well known that the construction on soft clayey soil is always a great challenge to the geotechnical engineers. The soft clay
poses high compressibility and low bearing capacity. It is a common practice to improve the properties of the soft clay prior to
any construction on it. In this respect, ground improvement by stone columns is a usual choice of the geotechnical engineers.
The stone columns increase the bearing capacity and reduce the settlement of the soft clay. Many theories are developed to
determine the bearing capacity of the soft soil reinforced with stone columns. However, most of the theories are site-specific
and do not show a very good match with the field observations. In this study, a large numbers of data were collected from
previously reported studies from various parts of the globe and an empirical formula based on support vector regression
(SVR) technique for the determination of the ultimate bearing capacity of the stone columns is achieved. Two different
techniques, namely tenfold cross-validation (qTFCV ) and non-cross-validation (qNCV ), are presented for the construction of
the SVR model. It is observed that the SVR method gives a better prediction than artificial neural network method. Laboratory
experiments were conducted to validate the SVR-ERBF empirical approach. The formula is also validated with two field
observations by two other investigators.

Keywords Support vector regression · Artificial neural network · Bearing capacity of stone columns · Soft clay

1 Introduction dite the consolidation of the soft clay. However, use of the
stone columns is mainly limited to embankment construc-
Construction on soft clay is always a great challenge to the tion, since the settlement cannot be completely eliminated.
geotechnical engineers. High compressibility and low shear Many researchers calculated the bearing capacity and the
strength are the two main characteristics of the soft clay to settlement of the soft soil reinforced with stone columns
pose this challenge. It is a usual practice to construct piles in experimentally [3–7]. Normally, the experimental results
this type of soil to transfer the load to the hard stratum below. have some inherent errors due to weather variation, installa-
For low- to medium-rise buildings, cost of construction of tion techniques, human factors, etc. So an appropriate analyt-
the piles is very high compared to the cost of the superstruc- ical or numerical study is also important. Some researchers
ture [1]. Hence, some ground improvement techniques may conducted analytical studies or used available software to
be adopted to solve the problem. Borthakur and Dey (2017) predict the behavior of the stone columns [8–14]. Various
improved the strength of soft soil by providing micro-piles as artificial intelligence (AI) approaches such as artificial neu-
reinforcement [2]. Construction of stone columns is another ral network (ANN), support vector regression (SVR), genetic
commonly used ground improvement technique for the soft programming, GEP, GPR, MARS, functional network can
clay. Stone columns increase the bearing capacity and expe- be used to predict an output from a number of input parame-
ters. Rentschler et al. (2017) obtained a better performance of
B Manita Das SVM than that of MARS in terms of coefficient of regression
manitta403@gmail.com R 2 [15]. Cheng and Wu (2008) obtained a better performance
Ashim Kanti Dey of SVM in learning and curve fitting and genetic algorithm
ashim_kanti@yahoo.co.in in optimization [16]. Olyaie et al. (2018) obtained a better
performance of SVM than GEP in their work with a lower
1 Civil Engineering Department, N.I.T. SILCHAR, Silchar, value of root-mean-square error (RMSE) and mean absolute
India

123
Arabian Journal for Science and Engineering

relative error (MARE) [17]. Li and Miao (2015) used SVM to such as linear correlation coefficient (R), root-mean-squared
establish the prediction model for preprocessed data and used error (RMSE), and mean absolute percentage error (MAPE).
GA to select the optimum value of the global search [18]. Liu Since the data used in this study were collected from vari-
et al. [19] declared that GA gives satisfactory results in vari- ous places with various weather conditions, both from field
able selection, whereas SVM is a very promising tool for the and laboratory, it is necessary to normalize the data within
prediction approximation [19]. Hence, it was decided that a range. Hence, in the present study, the dataset was nor-
both ANN and SVR would be used in the present study and malized in the range [0, + 1] before constructing the SVR
the results would be compared to obtain a better solution. models [22,23]. The values were normalized by using the
Artificial neural network (ANN) is a computational tech- formula given below:
nique which has the capability of input–output mapping.
In geotechnical engineering, many researchers used ANN Normalized value
to solve various problems [20] although it has some major (Observed value − Minimum value)
= . (1)
drawbacks like poorer generalizing performance, over-fitting (Maximum value − Minimum value)
problems, and arriving at a local minimum and slow conver-
gence speed [21]. Chik et al. [23] and Chik and Aljanabi The normalization also improves the condition number of
[22] used ANN to find the settlement of the stone columns the Hessian matrix in the normalization problems [21]. The
[22,23]. Support vector regression (SVR) is a comparatively training set was used to construct the SVR model, and the
new artificial intelligence technique which does not have the testing set to describe the performance of already established
above drawbacks, but has many advantages such as sparse- models.
ness of its solution and minimization of a number of support
vectors. SVR is based on the theory proposed by Vapnik et
al. [24]. Samui [21] used SVR to predict the settlement value 3 Methodology
of shallow foundations [21].
In the present study, 102 numbers of data were collected With the collected data, an attempt was made to match
from previously published papers. It is observed that the the observed values with the calculated values of the bear-
existing theories [13,14,25] on bearing capacity and settle- ing capacity as per the existing theories [5,14,25], and the
ment of the stone columns do not match the calculated values comparison is shown in Fig. 1. It is clearly noted that the
with the observed values, i.e., target values. ANN and SVR calculated values do not at all match with the observed val-
techniques can be used to find a better prediction of the prob- ues. Thus, it can be said that the established theories are not
lem. Results show that SVR has a better prediction efficacy universal and are more site-specific. Hence, some better pre-
than the ANN. diction is required. Various intelligence methods can be used
An empirical formula of the bearing capacity was obtained for the prediction of bearing capacity of the stone columns.
using 80% of the results of SVR model and tested with the In this study, the artificial intelligence methods (AI) such
remaining 20% of the results. The formula was also validated as support vector regression (SVR) and artificial neural net-
with a set of experimental results undertaken in the present work (ANN) have been used to predict the bearing capacity
study. Two case histories on bearing capacity of the stone of the stone columns. To select the best model, two different
columns were also collected and validated with the empirical validation techniques (tenfold cross-validation (TFCV) and
formula. Thus, it is expected that the empirical formula may non-cross-validation technique (NCV)) were used for both
be useful for practicing engineers. ANN and SVR models. The results from the SVR model
were compared with the results from the ANN.
Tenfold cross-validation (qTFCV )
2 Normalization of Collected Data In the tenfold cross-validation (TFCV) model, the whole data
are divided into ten equal sections. Firstly, the first nine sets
As discussed above, 102 numbers of data were collected from
of data are used for training purpose and the remaining one
the published literature (described in Table 1) to study the
is used for the testing purpose. In the second iteration, out of
bearing capacity of the stone columns. The bearing capacity
these ten sets of data, other nine subsets are used for the train-
of the stone columns depends on undrained cohesion (cu ) of
ing purpose and remaining one is used for testing purpose.
the surrounding soft soil, angle of friction of stone column
By repeating this process, ten iterations were completed until
materials (ϕ), spacing-to-diameter ratio (s/d), and length (l)
all the subsets are used for training.
of the stone columns. Therefore, a model was tried to develop
with all these four parameters as the input parameters and the Non-cross-validation (qNCV )
bearing capacity as the output parameter. The performance In the non-cross-validation (NCV) model, 70% of the total
of this model was discussed with three statistical parameters collected data are used for training and the remaining 30%

123
Arabian Journal for Science and Engineering

Table 1 Description of collected data


Mean value of data No. of collected data Field test (F) or laboratory test (L) References
cu in kPa s/d φ in degree l in mm qu in kPa

17.00 3.00 43.00 450.00 1739.08 24 F Ambily, and Gandhi [4]


19.44 1.96 41.11 7888.89 274.44 9 F Poorooshasb and Meyerhof [32]
26.20 2.21 40.10 12.15 173.00 10 L Andreou and Papadopoulos [33]
13.13 2.13 40.00 127.50 69.25 4 L Etezad [34]
13.24 2.35 27.35 4250.00 298.82 17 F Ambily and Gandhi [3]
2.50 2.89 41.50 500.00 35.67 3 L Murugesan and Rajagopal [35]
6.00 2.00 48.00 217.50 29.00 10 L Malarvizhi and Ilampurthi [36]
15.00 2.00 42.00 500.00 550.00 1 L Mohanty and Samanta [7]
20.00 3.33 37.00 300.00 140.00 1 L Golait et. al. [37]
11.25 2.00 11.25 10, 000.00 188.13 8 F Zhang et. al. [38]
15.00 1.50 35.00 20, 000.00 1800.00 1 F Hassen et. al. [39]
11.18 3.00 39.86 3814.29 357.71 14 F Etezad et. al. [14]

123
Arabian Journal for Science and Engineering

1000
IS code be noticed that ERBF kernel function shows a better per-
Etezad et.al.
900 formance than any other kernel functions mentioned here
Estimated bearing capacity in kPa

Afshar and Gazavi


800 in terms of RMSE value and MAPE value for training and
700
testing data.
600

500
3.2 Artificial Neural Network (ANN) Formation
400

300
Training and testing of ANN model
200 The collected data from the previous literature on the bearing
100 capacity of stone columns were divided into training, testing,
and validation data. With the training data, a network was
0
0 100 200 300 400 500 600 700 800 900 1000 developed and with the testing and validation data, the trained
Target bearing capacity value in kPa model was verified. In this case, Levenberg–Marquardt (LM)
algorithm was used for the prediction because it shows the
Fig. 1 Comparison of target bearing capacity value with calculated
bearing capacity value using some of the collected data
best result as compared to other algorithms [20]. The details
of all the results are listed in Table 4.
Figure 3a and b shows the mean standard error (MSE)
of the data are used for testing and validation of the ANN against epoch graph of TFCV and NCV model, respectively.
model. From the figures, it can be seen that with increase in number
of epoch, MSE value decreases which confirms that there is
no chance of over-fitting [20].
3.1 Support Vector Regression (SVR) Model
In the trial-and-error method, the number of hidden neu-
rons was chosen in such a way that it can show the best
The computation of SVR modeling was conducted in
performance. In the present study, back-propagation neural
MATLAB R2013a with SVR toolbox [26]. In the present
network (BPN) was used instead of radial basis function net-
study, for the prediction of the bearing capacity of the stone
work (RBFN) since BPN has a superior capability in pattern
columns, the following kernel functions were used: polyno-
classification system [23]. Therefore, by using the optimum
mial (POLY), radial basis function (RBF), and exponential
number of hidden neurons, the ANN was trained as shown in
radial basis function (ERBF). Trial-and-error method was
Fig. 4. The figure shows a very good correlation between the
used to get the best combination of the values of kernel
network output and the output of both TFCV (q90−10 ) and
parameters such as the degree of a polynomial function (de ),
NCV (q70−30 ) models. A comparison of R values, MAPE,
the width of RBF function (σ ), penalty parameter (C), and
and RMSE between TFCV and NCV is given in Table 4.
loss function parameter (ε). The best performance of the SVR
From the Table, it is seen that TFCV model gives a better
model was obtained by checking all these parameters in trial-
prediction since it shows a higher value of R and lower val-
and-error basis, which are listed in Table 2. The predicted
ues of MAPE, and RMSE as compared to NCV model. Chik
versus experimental, i.e., target values of the bearing capac-
et al. [23] and Chik and Aljanabi [22] also obtained a better
ity under SVR model, is shown in Fig. 2 for qTFCV and qNCV .
performance using TFCV model than that of NCV model
The optimal values of all the kernel parameters for both
[22,23].
the training and testing data are listed in Table 2. The number
of support vectors (nsv) is same for all the kernel functions;
nsv equals to 85 and 64 for qTFCV technique and qNCV tech-
nique, respectively. Among all the kernel functions, ERBF
shows the best correlation with high correlation coefficient 4 Results and Discussion
value as 0.99837 and 0.99709 for qTFCV and qNCV , respec-
tively. Mozumder et al. [27] also obtained the best result with From the study of both the artificial intelligent methods, i.e.,
ERBF-SVR of all the kernel functions. The performance of ANN and SVR, it is seen that for both the methods, TFCV
different kernel functions is presented in terms of root-mean- (qTFCV ) shows a better performance than NCV technique
square error (RMSE) and maximum absolute percentage (qNCV ). Again, SVR model shows a better performance than
error (MAPE) values. For both the validation techniques, that of ANN model, because a higher R-value is observed
qTFCV and qNCV , with different kernel functions, the RMSE for all the kernel functions in SVR model (Tables 3 and 4).
and MAPE values along with correlation coefficient values Moreover, the RMSE and MAPE values are less in SVR than
for training and testing data are given in Table 3. It can also in ANN.

123
Arabian Journal for Science and Engineering

Table 2 Best combination of


Models Kernel functions de σ C ε
kernel parameters
(qTFCV ) Radial basis function (RBF) 1 10 0.001
polynomial function (POLY) 3 1 0.0001
Exponential radial basis function (ERBF) 1 12 0.0009
(qNCV ) Radial basis function (RBF) 1.2 10 0.001
polynomial function (POLY) 3 1 0.009
Exponential radial basis function (ERBF) 1 10 0.0001
SVR predicted bearing capacity value in kPa

1 1

SVR predicted bearing capacity value in kPa


ERBF, R=0.99837 ERBF, R=0.99709
0.9 RBF, R=0.99133
RBF, R=0.99146
0.8 POLY, R=0.98812 0.8 POLY. R=0.98525

0.7
0.6 0.6

0.5
0.4 0.4

0.3
0.2 0.2

0.1
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Target bearing capacity value in kPa Target bearing capacity value in kPa
(a) (b)

Fig. 2 SVR predicted bearing capacity value, qu against target bearing capacity value a tenfold cross-validation technique, b non-cross-validation
technique

Table 3 Error analysis of the


Models Kernel functions R RMSE MAPE
models
(qTFCV ) Radial basis function (RBF) Training 0.99146 0.9070 3.1970
Testing 0.99005 0.98554 5.0510
Polynomial function (POLY) Training 0.98812 1.2526 4.8573
Testing 0.9712 1.7122 6.5891
Exponential radial basis function (ERBF) Training 0.99837 0.1856 0.2672
Testing 0.99438 0.7570 3.9560
(qNCV ) Radial basis function (RBF) Training 0.99133 0.9737 4.2650
Testing 0.98952 1.1729 6.2726
Polynomial function (POLY) Training 0.98525 1.4543 5.3956
Testing 0.97042 1.9345 7.4377
Exponential radial basis function (ERBF) Training 0.99709 0.2320 0.3278
Testing 0.99193 0.8588 4.6433

Table 4 Optimum number of


ANN models Algorithm R RMSE MAPE No. of hidden neurons
hidden neurons
q90−10 Training 0.95472 1.612 5.0796 11
Testing 0.95235 1.892 5.235 11
q70−30 Training 0.94846 2.53 6.7131 7
Testing 0.932 2.65 6.754 7

123
Arabian Journal for Science and Engineering

Best Validation Performance is 0.095966 at epoch 1000 Best Validation Performance is 0.082873 at epoch 2
0 0
10 10 Train
Train
Validation Validation
Mean Squared Error (mse) Test Test

Mean Squared Error (mse)


Best Best

-1 -1
10 10

-2 -2
10 10
0 200 400 600 800 1000 0 0.5 1 1.5 2
1000 Epochs 2 Epochs
(a) (b)

Fig. 3 Mean standard error (MSE) versus epoch graph of a tenfold cross-validation (TFCV) model and bnon-cross-validation (NCV) model for
the algorithm LM

training: R=0.95472 training: R=0.94846


1 1
Data Data
0.9
Output ~= 0.88*Target + 0.0093

Fit Fit
Output ~= 0.88*Target + 0.038

0.8 Y =T Y =T
0.8

0.6 0.7

0.6
0.4
0.5

0.4
0.2
0.3
0 0.2

0.1
-0.2
0

-0.2 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1


Target Target
(a) (b)

Fig. 4 BPN network output versus training of a TFCV (tenfold cross-validation) model and b NCV (non-cross-validation) model for the algorithms
LM

5 Development of Empirical Formula for the SVR-ERBF simulated bearing capacity values and taking
Bearing Capacity Calculation friction angle (φ) and undrained cohesion of clay (cu ) as the
main factors keeping other parameters fixed at the reference
Based on the above discussions, SVR technique has been values. The reference values of all the parameters are selected
used for bearing capacity prediction of stone columns. The near to their mean values [29]. The mean values of the param-
empirical formula is obtained using 80% of the results of eters cu , s/d, φ, and l are obtained as 14.39 kPa, 2.47, 36.72◦ ,
SVR model for TFCV technique. Leung et al. [28] derived 2492.09 mm, respectively, and thus the reference values are
this type of empirical equation with several charts and cor- considered as 14 kPa, 2.5, 38◦ , and 2200 mm.
rection factors for his research [28]. As undrained cohesion For the prediction of bearing capacity of stone columns,
of the soft clay (cu ) and angle of friction of the stone column the following equation is proposed:
materials (φ) are the most sensitive parameters [22,23] for
the prediction of the bearing capacity of the stone columns, qemp = qmain × Fc , (2)
an equation of the bearing capacity, qmain , was formed using

123
Arabian Journal for Science and Engineering

where qemp is the corrected bearing capacity and Fc is the 1.3


s/d=1.5
correction factor for the other two parameters, namely s/d
1.2 s/d=1.6
and length. s/d=1.75
Using Statistical Package for the Social Sciences (SPSS), s/d=1.85
1.1
considering cu (in kPa) and φ (in degree) as the main factors, s/d=2
the equation of qmain (in kPa) is formed as given below (both 1 s/d=2.1

Cl
cu and φ cannot be zero): s/d=2.3

0.9 s/d=2.5

qmain = 0.234 ×2 −4.124 × +9.093 × cu + 46.539. (3) s/d=2.75

0.8 s/d=3
s/d=3.2
Fc can be written as
0.7 s/d=3.5
s  0.4 0.7 1 1.3 1.6 1.9 2.2 2.5 2.8
Fc , l = C ds × Cl , (4) l/2200
d
(a)
where C ds andCl are the correction factors corresponding to
1.3
the parameters ds and l, respectively. The correction factors l=150 mm

l=300 mm
are obtained considering that the bearing capacity depends 1.2 l=500 mm
only on one parameter at a time. l=1000 mm
To obtain the correction factor for s/d ratio (Cs/d ), first 1.1 l=1500 mm

of all, a master curve was drawn by varying the length l with l=2000 mm
1
the SVR-ERBF simulated bearing capacity values keeping Cs/d l=2500 mm

l=3000 mm
s/d as constant at its reference value. The SVR-ERBF sim- 0.9
l=3500 mm
ulated values were evaluated for each combination of s/d
l=4000 mm
and l. Now, for different values of l, the values of s/d were 0.8
l=4500 mm
varied within the entire range and the correction factors were l=5000 mm
0.7
determined by dividing these SVR-ERBF simulated bearing l=5500 mm

capacity values by the value obtained from the master curve. 0.6
l=6000 mm

Figure 5a shows the variation of Cs/d with SVR-ERBF bear- 0.6 0.8 1 1.2 1.4 1.6 1.8
ing capacity values for different values of l and 5(b) shows s/d/2.5
the variation of Cl with SVR-ERBF bearing capacity val- (b)
ues for different s/d ratios. In Fig. 5a, the increase in Cl
Fig. 5 Variation of correction factors a Cl with different values of s/d
can be observed with the increase in length of the stone and b Cs/d with different values of l
columns for all values of s/d ratio. At the reference point
l/2200 = 1, the value of correction factor is one because at  4  3
s s
this point, both the bearing capacity values are same. When d d
Cs/d = −0.762 + 4.172
magnitude of l is less than its reference value, the curve is 2.5 2.5
more spread, showing that Cl is more sensitive for lower  s 2  s 
d d
values ofl. Again, a decrease in correction factor (Cs/d ) is −8.045 + 6.02 − 0.384 (6)
2.5 2.5
observed in Fig. 5b with the increase in s/d for all values of
qemp = qmain × C d × Cl .
s (7)
length. The spread portion of the curve indicates that Cs/d
is more sensitive for s/d values lower than the reference
Therefore, after knowing qmain from Eq. (3) and correction
value.
factors from Eqs. 5 and 6, the bearing capacity of stone col-
To achieve an expression of the correction factor for both
umn can be found out by using the above-proposed equation
the parameters, the best-fitted line with the minimal least
(Eq. 7). In case the values of some parameters do not come
square error has been chosen. The expressions for the correc-
within the range assumed in this study, interpolation method
tion factors are as
may be used.
follows:
The proposed empirical equation has been tested by using
 2   remaining 20% of the results of ERBF-SVR model of the
l l
Cl = 0.0038 + 0.305 + 0.627 (5)
2200 2200

123
Arabian Journal for Science and Engineering

Empirically predicted bearing capacity value in kPa


1000
Training data
Testing data
900

800

700

600

500

400

300
Fig. 7 Plan of seven stone columns
200

100
0 100 200 300 400 500 600 700 800 900 1000
study the performance of the central stone column[25]. One
footing in the form of a steel plate of diameter 10 cm and
Target bearing capacity value in kPa
thickness 10 mm was placed over the central stone column
Fig. 6 Target bearing capacity against empirically predicted bearing as shown in Fig. 7. The vertical load was applied with a
capacity using 20% of the collected data hydraulic jack of capacity 5 ton through a load cell of same
capacity, and the settlements were measured with the LVDTs
of sensitivity 0.01 mm kept at diagonally opposite ends. Two
collected data. The calculated bearing capacity values from
dial gauges were also used for measuring the settlements and
the empirical equation are plotted against the target bearing
placed along the other diagonal. A twelve-channel portable
capacity values as shown in Fig. 6. The figure shows a very
data acquisition system was used for recording the data from
good correlation among the data with correlation coefficient
the load cell and the LVDTs. Figure 8 shows a photograph
value of R = 0.9887.
and a sectional elevation of the test setup.
The ultimate bearing capacity of the stone columns
was measured by double-tangent method from the load–
6 Performance Study of SVR Model
settlement curve. The method consists of drawing two
tangents, one at the initial portion of the load–settlement
The performance of the SVR model was studied by conduct-
curve and the other at the end portion of this curve, and
ing some laboratory experiments. The laboratory model tests
determining the load termed as ultimate bearing capacity
were conducted in a steel tank of size 1 m × 1 m × 1 m for the
corresponding to the intersection point of the two tangents.
determination of the bearing capacity of the stone columns.
Table 5 lists the results of the experimental investigation con-
A thick polythene sheet was laid on the inner side of the tank
ducted in this study. The last column of the table shows
to avoid side friction and loss of water. Clay lumps collected
ultimate bearing capacity, qu of the stone columns for dif-
from the paddy fields were pulverized, freed from grass, plant
ferent combinations of s/d and l.
roots, stone pebbles, and other foreign materials, and kept in
the sun for drying. Water was added to the clay powder and
placed in the tank in layers, each layer of 100 mm thick. The
6.1 Application of the SVR-ERBF Empirical Equation
total thickness of the soil bed was kept as 0.9 m. The tank
in Experimental Data
was covered with a thick polythene sheet, and the soil was
The bearing capacity values obtained from the proposed
kept idle for 7 days. The water content was tried to be main-
empirical equation are verified with those obtained from
tained near the liquid limit, and at this water content, the
experimental results. Figure 9 shows a comparison between
undrained cohesion was found to be around 20 kPa. To con-
the above two bearing capacity values in normalized form.
struct a stone column, a hole was first made using an auger
A very good correlation is noticed with a correlation value,
of 50 mm diameter and then was filled in layers with stone
R = 0.9734. It is distinctly noticed that almost all the data
chips of size 2–6 mm with proper compaction. It was tried to
obtained from the empirical equation lie within the region
maintain the density of the stone chips in the stone columns
bounded by an upper and lower bound at 95% confidence
around 15 kN/m3 . A direct shear test on the stone chips at
interval.
this density showed an angle of internal friction to be 37.27◦ .
Seven stone columns were arranged in a triangular pattern to

123
Arabian Journal for Science and Engineering

(a)

LO AD IN G FRA ME

LO AD C ELL
LVD T
DA TA AC QU IS ITION

PLA TE
S TON E
1.0 M C OLU MN S 0.9 M
HY DR AU LIC JA CK

SO FT CLA Y

1.0 M X 1.0 M

(b)
Fig. 8 Experimental setup. a Photograph of the test setup, b sectional elevation of the test setup

7 Applicability of the Empirical Approach • Calculation of empirical bearing capacity value qemp ,
from Eq. 7.
In this section, the application of different aspects of the
proposed empirical equation is briefly discussed. It is also Let us take two case studies from field data which were not
discussed that how it can be related to the field bearing capac- used to formulate the above empirical equation:
ity of the stone column.
Case 1:
Method: As per the reference Bhosle and Vaishampayan [30],
qemp is assumed as the bearing capacity of the stone columns The given data are the undrained cohesion of clay, cu =
in the field. Different steps are explained below which are 10 kPa; diameter of the stone column, d = 0.9 m; length,
involved in the process: l = 21m; spacing, s = 1.7m; friction angle, φ = 30◦ ; and
the observed bearing capacity value = 876.75 kPa.
• Determination of bearing capacity value (qmain ) from Determination of calculated bearing capacity value from
Eq. 3. the proposed empirical equation:
• Evaluation of all the input parameter values.
• Calculation of correction factors with the use of Eqs. 5 • From Eq. 3, qmain = 224 kPa.
and 6. • From Eqs. 5 and 6, Cl = 3.88, Cs/d = 1.12.

123
Arabian Journal for Science and Engineering

Table 5 Experimental test results Determination of calculated bearing capacity value from
cu (kPa) s/d ratio φ (degree) L (mm) qu (kPa) the proposed empirical equation:

20 1.5 37.27 300 382


• From Eq. 2, qmain = 753.66 kPa.
20 2 37.27 300 365
• From Eqs. 4 and 5, Cl = 1.6, Cs/d = 1.1.
20 2.5 37.27 300 352
• From Eq. 6, the estimated bearing capacity value is found
20 3 37.27 300 307
to be 1326.44 kPa.
20 3.5 37.27 300 265 • The percentage variation among the bearing capacity val-
20 1.5 37.27 400 394 ues is almost 3%.
20 2 37.27 400 382
20 2.5 37.27 400 367
20 3 37.27 400 304
8 Conclusions
20 3.5 37.27 400 282
20 1.5 37.27 500 403
In this study, the applicability of SVR model for the bearing
20 2 37.27 500 394
capacity prediction of the stone columns is discussed. The
20 2.5 37.27 500 379 data collected from the previous literature were used for this
20 3 37.27 500 316 purpose. To investigate the performance of the SVR model,
20 3.5 37.27 500 240 three kernel functions such as polynomial (POLY), radial
basis function (RBF), and exponential radial basis function
(ERBF) were used. To observe the efficacy of SVR model,
1
data
ANN model was also constructed. For both the cases, two
0.9 f itted curve techniques such as tenfold cross-validation technique and
Emperically predicted normalized bearing

non-cross-validation technique were used.


0.8
From the present study, following conclusions are drawn:
0.7
capacity value

0.6 (a) For both the cases of SVR and ANN, tenfold cross-
0.5
validation technique shows better results than non-cross-
validation technique.
0.4 (b) SVR model shows better results than ANN model.
0.3 (c) From the statistical performance of the SVR models, i.e.,
from R, RMSE, MAPE, and prediction error variability,
0.2
the best performance of ERBF kernel function can be
0.1 noticed compared to remaining two kernel functions such
as RBF and POLY.
0
0 0.2 0.4 0.6 0.8 1 (d) By using the SVR-ERBF simulated results, an empir-
Experimental normalized bearing capacity value ical equation of the ultimate bearing capacity of stone
columns is obtained.
Fig. 9 Experimental versus empirically predicted normalized bearing
capacity value (e) The empirical equation of the bearing capacity is vali-
dated with laboratory results as well as two field obser-
vations obtained from the published literature.
• From Eq. 7, the estimated bearing capacity value is found (f) The empirical equation obtained in this study is not site-
to be 974 kPa. specific and can be used in any site condition with known
• The percentage variation among the results is almost values of cu , φ, s/d, and length of the stone columns.
10%.

References
Case 2:
As per Report No. FHWA/RD-83/026 (1983) [31], 1. Gucuyen, E.; Erdem, R.T.; Gokku, U.: Irregular wave effects on
The given data are the undrained cohesion of clay, cu = dynamic behavior of piles. Arab J. Sci. Eng. 38, 1047–1057 (2013).
https://doi.org/10.1007/s13369-012-0428-6
47.9 kPa; diameter of the stone column, d = 0.915 m;
2. Borthakur, N.; Dey, A.K.: Experimental investigation on load car-
length, l = 7 m; s/d = 2; friction angle, φ = 44◦ ; and rying capacity of micropiles in soft clay. Arab J. Sci. Eng. (2017).
the observed bearing capacity value = 1368.7 kPa. https://doi.org/10.1007/s13369-017-2894-3

123
Arabian Journal for Science and Engineering

3. Ambily., A.P.; Gandhi, S.R.: Experimental and theoretical evalua- 23. Chik, Z.; Aljanabi, Q.A.; Kasa, A.; Taha, M.R.: Tenfold cross vali-
tion of stone column in soft clay (2004) ICGGE. dation artificial neural network modeling of the settlement behavior
4. Ambily, A.P.; Gandhi, S.R.: Behavior of stone columns based of a stone column under a highway embankment. Arab J. Geosci.
on experimental and fem analysis. J. Geotech. Geoenvi- 7, 4877–4887 (2014). https://doi.org/10.1007/s12517-013-1128-
ron. Eng. ASCE 133, 405–415 (2007). https://doi.org/10.1061/ 6
(ASCE)1090-0241(2007)133:4(405) 24. Vapnik, V.N.; Golowich, S.E.; Smola, A.: Support Vector Method
5. Afshar, J.N.; Ghazavi, M.: Experimental studies on bearing capac- for Function Approximation, Regression Estimation, and Signal
ity of geosynthetic reinforced stone columns. Arab J. Sci. Eng. 39, Processing, Advances in Neural Information Processing Systems.
1559–1571 (2014). https://doi.org/10.1007/s13369-013-0709-8 Morgan Kaufmann, San Mateo (1996)
6. Castro, J.: An analytical solution for the settlement of stone 25. IS 15284-1: Design and construction for ground improvement-
columns beneath rigid footings. Acta Geotech. (2014). https://doi. Guidelines. Part-1: Stone columns [CED 43: Soil and foundation
org/10.1007/s11440-014-0358-4 engineering], ICS 93.020. (2003)
7. Mohanty, P.; Samanta, M.: Experimental and numerical studies 26. Gunn, S.R.: Support vector machines for classification and regres-
on response of the stone column in layered soil. Int. J. Geosynth. sion. Tech Rep. University of Southampton. (2001) http://www.
Ground Eng. 1, 27 (2015). https://doi.org/10.1007/s40891-015- isis.ecs.soton.ac.uk/isystems/kernel/svm.zip,
0029-z 27. Mozumder, et al.: Support vector regression approach to predict
8. Goh, A.T.C.; Kulhawy, F.H.; Chua, C.G.: Bayesian neural network the strength of FRP confined concrete. Arab J. Sci. Eng. (2016)
analysis of undrained side resistance of drilled shafts. J. Geotech. https://doi.org/10.1007/s13369-016-2340-y
Geoenviron. Eng. ASCE 131(1), 84–93 (2005) 28. Leung, : Empirical approach for determining ultimate FRP strain in
9. Lee, J.S.; Pande, G.N.: Analysis of stone-column reinforced Foun- FRP-strengthened concrete beams. J. Compos. Constr. 10(2), 125–
dations. Int. J. Numer. Anal. Meth. Geomech. 22, 1001–1020 138 (2006). https://doi.org/10.1061/(ASCE)1090-0268(2006)10:
(1998) 2(125)
10. Shahu, J.T.; Reddy, Y.R.: Clayey soil reinforced with stone column 29. Mozumder, R.A.; Laskar, A.I.; Hussain, M.: Empirical approach
group: model tests and analyses. J. Geotech. Geoenviron. Eng., for strength prediction of geopolymer stabilized clayey soil using
ASCE 137(12) (2011). https://doi.org/10.1061/(ASCE)GT.1943- support vector machines. Constr. Build. Mater. 132, 412–424
5606.0000552 (2017)
11. Niroumand, H.; Kassim, K.A.; Yah, C.S.: Soil improvement by 30. Bhosle, S.; Vaishampayan, V.V.: Ground Improvement Using Vibro
reinforced stone columns based on experimental work. EJGE16 Stone Columns Capacity Of Stone Column. Guntur, INDIA (2009).
(2011) IGC 2009
12. Nassaji, F.; Asakereh, A.: Effect of granular bed on behaviour of 31. Report no. FHWA/RD-83/026 (1983) www.dot.ca.gov/hq/esc/
stone column improved ground. Int. J. Sci. Eng. Investig. 2(23) geotech/references/Ground.../4-Stone_Columns.pdf.
(2013) 32. Poorooshasb, H.B.; Meyerhof, G.: Analysis of behavior of stone
13. Afshar J.N.; Ghazavi, M.: A simple analytical method for calcula- columns and lime columns. Comput. Geotech. 20(1), 47–70 (1997)
tion of bearing capacity of stone column. Int. J. Civil Eng. 12(1) 33. Andreou, P.; Papadopoulos, V.: Factors affecting the settlement
(2014) estimation of stone column reinforced soils. Geotech. Geol. Eng.
14. Etezad, M.; Hanna, A.M.; Ayadat, T.: Bearing capacity of a group 32(23), 1175–1185 (2014). https://doi.org/10.1007/s10706-014-
of stone columns in soft soil. Int. J. Geomech. ASCE, (2015) 9788-x
15. Rentschler, T.; Schmidt, K.; Kuhn, P.; Scholten, T.: Three- 34. Etezad, M.: Geotechnical performance of group of stone columns.
dimensional mapping of soil organic carbon (SOC) based on Phd Thesis (2007)
multi-scale digital terrain analysis and data mining in Jiangxi 35. Murugesan, S.; Rajagopal, K.: Performance of encased stone
Province, PR China. (202), Wageningen, 26 June–1 (2017) columns and design guidelines for construction on soft clay soils.
16. Cheng, M.Y.; Wu, Y.W.: Dynamic prediction of project success In: Proceedings of the 4th Asian Regional Conference on Geosyn-
using evolutionary support vector machine inference model. In: thetics, June 17–20, Shanghai, China (2008)
Proceedings of the 25th International Symposium on Automation 36. Malarvizhi, S.N.; Ilampurthi, K.: Comparative study on the behav-
and Robotics in Construction, vol. 203, p. 2005 (2008) ior of encased stone column and conventional stone column. Soils
17. Olyaie, E.; Heydari, M.; Banejad, H.; Chau, K.W.: A labora- found. Jpn. Geotechn. Soc. 47(5), 873–885 (2007)
tory investigation on the potential of computational intelligence 37. Golait, Y.S.; Satyanarayana, V.; Raju, S.S.V.: Concept of Under
approaches to estimate the discharge coefficient of piano key weir. Reamed Cemented Stone Columns for Soft Clay Ground Improve-
18. Li, X.; Miao, J.: Life Prediction Method of Lithium-Ion Bat- ment. Guntur, INDIA (2009). IGC 2009
tery Based on Grey Support Vector Machines. World Academy 38. Zhang, L.; Zhao, M.; Shi, C.; Zhao, H.: Settlement calcula-
of Science, Engineering and Technology, International Journal of tion of composite foundation reinforced with stone columns.
Medical and Health Sciences, 9(5), (2015) Int. J. Geomech. 13(3), 248–256 (2013). https://doi.org/10.1061/
19. Liu, H.X.; Zhang, R.S.; Yao, X.J.; Liu, M.C.; Hu, Z.D.; Fan, B.T.: (ASCE)GM.1943-5622.0000212
Prediction of the isoelectric point of an amino acid based on GA- 39. Hassen, G.; Buhan, P.D.; Abdelkrim, M.: Finite element imple-
PLS and SVMs. J. Chem. Inf. Comput. Sci. 44(1), 161–167 (2004) mentation of a homogenized constitutive law for stone column-
20. Kostic, S.; Vasovic, D.: Prediction model for compressive strength reinforced foundation soils, with application to the design of
of basic concrete mixture using artificial neural networks. Neu- structures. Comput. Geotech. 2010(37), 40–49 (2010)
ral Comput. Appl. 26, 1005–1024 (2015). https://doi.org/10.1007/
s00521-014-1763-1
21. Samui, P.: Support vector machine applied to settlement of shallow
foundations on cohesionless soils. Comput. Geotechn. 35, 419–427
(2008)
22. Chik, Z.; Aljanabi, Q.A.: Intelligent prediction of settlement ratio
for soft clay with stone columns using embankment improvement
techniques. Neural Comput. Appl. 25, 73–82 (2014). https://doi.
org/10.1007/s00521-013-1449-0

123

View publication stats

You might also like