You are on page 1of 10

ISSN 1064-2293, Eurasian Soil Science, 2009, Vol. 42, No. 13, pp. 1487–1496. © Pleiades Publishing, Ltd.

, 2009.

SOIL
PHYSICS

Comparison of Artificial Neural Networks Models


with Correlative Works on Undrained Shear Strength1
O. Sivrikaya
Associate Professor Doctor, Department of Civil Engineering, Ni ğde University, 51100 Ni ğde, Turkey
e-mail: osivrikaya@nigde.edu.tr

Abstract—In recent years, the Artificial Neural Network (ANN) modelling that has been used in the solution
of the complex problems has gained an increasing interest in soil science. The ANN modelling is also getting
more popular in soil mechanics applications. It is a preferable method among the other approaching methods
because of having quick results in test phase in short time. This paper describes the ANN models for estimating
undrained shear strength (Su) of cohesive soils from SPT (Standard Penetration Test) data with index properties
in Turkey. The performance of the ANN models is investigated using different input variables such as measured
N, corrected N (N60) value, natural water content (wn), liquid limit (wL), plasticity index (Ip). In this study the
ANN models are compared to empirical methods. The results indicate the superior performance of ANN models
over the empirical methods.
DOI: 10.1134/S1064229309130092

1 1. INTRODUCTION sible that employing different types of test leads to erro-


Correlations are important to estimate engineering neous results. Therefore, an engineer should be aware
properties of soils particularly for a project where there of those facts before using correlation equations.
is a financial limitation, lack of test equipment or lim- Unless those uncertainties are known, the used correla-
ited time. Therefore it is useful to estimate the engi- tion may lead to overestimated or underestimated
neering properties of soil by using other soil parameters design parameters. This study covers the use of the
which can be obtained easily. The relations between the ANN models to estimate undrained shear strength (Su)
values of various soil parameters in the field and/or lab- of cohesive soils using SPT-N value with index proper-
oratory conditions both assist the engineer for the pre- ties in Turkey.
liminary evaluation of the project and enable to check Standard Penetration Test (SPT) is one of the most
the consistency of the results determined by various common in-situ tests used for geotechnical design in
methods. Turkey and the world [2, 3]. Almost every soil explora-
Statistical regression modelling is a method in tion program also covers standard penetration test as
which the data corresponding to a dependent variable one of the principle parts [2, 4, 5]. The SPT is per-
are regressed against those of the independent variable. formed by driving a standard split-spoon sampling tube
In multiple regressions, a probabilistic model is of 30 cm into the ground at the bottom of a borehole
employed and it is assumed that it is normally distrib- with a 63.5 kg hammer falling from 76 cm. SPT-N
uted. All conventional regression methods have a limi- value, which is the number of hammer blows required
tation. It is also known that the empirical methods are to drive the sampler for last two 15 cm penetration, is
not applicable to the complex and highly non-linear called standard penetration resistance of soil. Many
problems [1]. factors and variables (the drilling methods, drill rods,
The correlations with SPT-N are commonly used in borehole sizes and stabilization, sampler, blow count
the preliminary stage of a project. However, many of rate, hammer configuration, energy corrections and test
those correlations, available in literature, are not clear procedure) affect the validity and usefulness of SPT
enough to apply to field data. It is also unknown results [6, 7]. Therefore, the SPT correction must be
whether the correlations include the corrections and made in order to get useable and useful SPT-N values
hence their use causes confusion. The other important in the geotechnical design and the determination of
issue is that which type of test is used for the determi- engineering properties. In recent years, various correc-
nation of the correlation equation. This is not clear tions have been developed for measured SPT-N values
either. The differences in the tests used in developing taking into account the effect of: rod length, borehole
these correlations play a very important role. It is pos- diameter, sampler type, type of hammer and its release
mechanism, blow count frequency, energy and the
1 The article is published in the original. effects of overburden pressure. While these corrections

1487
1488 SIVRIKAYA

are used, their use has led to confusion regarding the The undrained shear strength (Su) for cohesive soils
correlations used for corrected SPT-N values and for can be estimated indirectly by using the result of SPT
use uncorrected SPT-N values. [3, 11, 12]. Many studies indicate that the sensitivity of
Even though the SPT was originally developed for clays may give rise to lower SPT-N values for a given
coarse-grained soils, it has been applied to fine-grained undisturbed shear strength due to strength loss during
soils to estimate engineering properties. However, its sampler penetration [13, 14]. Many works have been
applicability for fine-grained soils is still argued. SPT- made on the relations between Su determined from UC
N value can be used to predict the bearing capacity and and UU test and SPT-N, and they are given in Table 1.
settlement of foundations on cohesionless soils, and Using the results of unconsolidated undrained (UU)
determination of engineering properties of cohesion- compression tests, Stroud [12] and Décourt [3] propose
less soils such as the relative density (Dr), the angle of the relationships between SPT-N value and Su accord-
shear strength (φ) and liquefaction potential, and engi- ing to the plasticity index, Ip, and corrected SPT-N
neering properties of cohesive soils such as undrained value respectively, in overconsolidated, insensitive
compressive strength (qu), undrained shear strength (Su) clays. The correlation equation proposed by Décourt
and coefficient of volume compressibility (mV) [8, 9]. [3] gives approximately twice the result of Stroud [11].
Measured blow counts (SPT-Nfield) can be normal- The remaining researchers given in Table 1 have devel-
ized to N60, which is the blow count corrected to 60% oped relations between Su and N by using the results of
of the theoretical free-fall hammer energy, and N1, 60, UC test.
which is the blow count corrected to 100 kPa of effec-
tive overburden pressure and 60% of the theoretical 1.2. Current Correlative Work between Su and SPT-N
free-fall hammer energy. The most general equations
for N60 and N1, 60 are as follows: In this study, laboratory test results and borehole
logs obtained from many parts of Turkey have been
N60 = CBCCCECRCBFCSCANfield, (1) provided by private companies, universities and one
public institution which carried out the work. Values for
N1, 60 = CNN60, (2)
Su, which are used for the correlations, were obtained
where CN is the overburden correction factor, CBF is the from depths close to the depth of the SPT. Also, a ques-
count frequency correction factor depending on soil tionnaire was prepared to collect information on the
type and underground water level, CE is the energy cor- SPT procedure and equipment used so that reliable cor-
rection factor, CR is the rod length correction factor, CB rections and correlations could be attempted.
is the borehole diameter correction factor, CS is the According to the results of the questionnaire, CE = 0.75
liner correction factor, and CA is the anvil correction due to the type of donut hammer used with release on two
factor, and CC is the hammer cushion correction factor turns of the rope [15, 16], CB = 1.00 due to the change
depending on test procedure and equipment [9]. of the borehole diameter from 65 mm to 115 mm [17],
Although all corrections are applied to non-cohesive CC = 1.00 due to no hammer cushion being used [3, 9],
soils, the overburden and blow count frequency correc- CS = 1.20 due to no liners being used [17], CA = 0.85
tion cannot be applied to cohesive soils in practice [9, due to using a donut hammer and small anvil [18] and
10]. As the overburden correction for clays is still sub- CR = 0.75, 0.85, 0.95 or 1.00 depending on the rod
ject to argument, and not normally made in practice3,, it length [17] are assumed. In the light of the results of the
is neglected in this study. Consequently, the general questionnaire, Eq. (3) becomes:
equation including corrections for cohesive soils can be N60 = 0.75CRNfield. (4)
rewritten as Eq. (3):
The correction factors are assumed according to the
N60 = CBCCCECRCSCANfield. (3) results of the questionnaire by taking the previous stud-
ies into consideration. How the correction factors are
determined is beyond the scope of the current paper.
1.1. Previous Correlative Works between Su and SPT-N To develop relations between Su and SPT-N, the
The undrained shear strength (Su) is a very impor- multi linear regression (MLR) analysis (Su = a · NorN60 +
tant parameter in order to make a geotechnical design b · wn + c · wL + d · Ip) by the method of least squares is
for structures and slopes founded on cohesive soils. Su performed by taking both test types and the SPT-N cor-
occurs as the cohesive soil is loaded rapidly and there is rections into consideration; a, b, c and d represent the
no time for pore water pressure to dissipate. The value regression coefficient. 100 data from UU test and
Su is determined by the laboratory/field vane (FV) tests 50 data from UC test are used for analysis respectively.
or Unconsolidated Undrained (UU) compression tests While SPT resistance (N, N60), natural water content
in the laboratory. In addition, Su can be obtained taking (wn), liquid limit (wL) and plasticity index (Ip) are taken
the half of unconfined compressive strength determined as the independent parameters, Su is taken as dependent
by the unconfined compression (UC) test (Su = qu/2). parameter. For these relations, the correlation coeffi-

EURASIAN SOIL SCIENCE Vol. 42 No. 13 2009


COMPARISON OF ARTIFICIAL NEURAL NETWORKS MODELS 1489

Table 1. Previous works on Correlations between SPT-N and Su


Author(s) Soil Type Su (kPa)
Sanglerat (1972) Clay 12.5N
Silty clay 10N
Terzaghiand Peck (1967) Fine-grained soil 6.25N
Sowers (1979) Highly plastic clay 12.5N
Medium plastic clay 7.5N
Low plastic clay and Silt 3.75N
Nixon (1982) Clay 12N
Sivrikaya and To ğrol (2002) Highly plastic clay (CH) 4.85Nfield, r = 0.83
n = 113 6.82N60, r = 0.80
Low plastic clay (CL) 3.35Nfield, r = 0.76
n = 72 4.93N60, r = 0.73
Fine-grained soil 4.32Nfield, r = 0.80
n = 226 6.18N60, r = 0.78
Stroud (1974)* Clay Medium plastic clay, (45)N
Su = f1N, f1 = f(Ip) Ip < 20 (67)N
Ip > 30 ≅4.2N
Décourt (1990)* Clay 12.5N
15N60
* Represents the correlations where Su is determined by UU test, the rest of them represent the correlations where Su is determined by UC test.

Table 2. Correlations between SPT-N and Su in cohesive soils


UU Test Su = 3.33N – 0.75wn + 0.20wL + 1.67Ip r = 0.82 s = ±34 kPa
n = 100 Su = 4.43N60 – 1.29wn + 1.06wL + 1.02Ip r = 0.81 s = ±34 kPa
UC Test Su = 2.41N – 0.82wn + 0.14wL + 1.44Ip r = 0.86 s = ±28 kPa
n = 50 Su = 3.24N60 – 0.53wn – 0.43wL + 2.14Ip r = 0.87 s = ±27 kPa

cient (r) has been established, and the standard errors shown in Fig. 2. The strength of connections between
(s) are determined for each regression equation neurons is represented by numerical values, called
obtained. The obtained regression equations with n, r weights. Each neuron has an activation value that is a
and s are given in Table 2. function of the sum of inputs received from other neu-
rons through the weighted connections [20, 21]. ANNs
have ability in performing a good amount of generali-
1.3. Modelling with Artificial Neural Networks zation from the patterns on which they are trained.
Artificial Neural Networks (ANNs) are data pro- Training consists of exposing the neural network to a
cessing paradigms made up of highly interconnected set of known input-output patterns. The data are passed
nodes, called neurons. A neuron in ANN maps an input through the multi-layered (MLP) feed forward neural
pattern into an output pattern. Even though there are network in a forward direction only. As the data moves
various types of neural networks which differ in the forward, it is subjected to simple processing within the
architecture and the rules of learning it is indicated in neuron and along the links connecting neurons. The
the literature that a multilayer feedforward neural net- network performs successive iterations to adjust the
work model is the most commonly used network for its weights of each neuron in order to obtain the target out-
efficient generalization capabilities [19, 20, 21]. The puts according to a specific level of accuracy. The
typical multi-layer feed-forward neural networks used adjusting process of neuron weights is carried out to
in the current application are shown in Fig. 1. This type minimize the network error, which is defined as a
of neural network consists of an input layer, one or mean-squared error between the computed and target
more hidden layer(s) and an output layer. Layers con- output patterns. After the neural network is satisfacto-
sist of a number of processing units, called neurons as rily trained and tested, it is able to generalize rules and

EURASIAN SOIL SCIENCE Vol. 42 No. 13 2009


1490 SIVRIKAYA

I1
W1
W2 Single Neuron
I2
N
A = ΣIiWi + C f = 1/(1 + e–A) Output of Neuron
I3 W3
Wn
I4 W4
WL Su
Fig. 2. The Functional Representation of Single Neuron in
Ip the Hidden Layer.

2. ARTIFICIAL NEURAL NETWORK MODELS


FOR ESTIMATING SU
Input Hidden Output In this study, a feed-forward neural network which
Layer Layer Layer consists of multilayer perceptrons trained by back-
propagation algorithms is used. For this purpose a com-
Fig. 1. The Architecture of Feed-forward Artificial Neural puter program written in C language has been devel-
Networks.
oped. The training process is terminated when the num-
ber of iteration reaches its maximum value of 1000000
will be able to deal with unseen input data to predict or as the cross-validation indicates that there is a dete-
output within the domain covered by the training pat- rioration in learning.
terns [20]. Several methods exist to train a network. Cross-validation is performed as follows: Training
One of the most successful and widely used training set used for training artificial neural networks is divided
algorithms for MLP is the back propagation. In this into two parts as an estimation subset and a validation
method, the input is propagated from the input layer subset. Firstly, the network is trained using the estima-
through the hidden layers to the output layer. The cal- tion subset. Then the validation subset is given as input
culated network error is then back propagated from the to the trained neural network and finally the mean
output layer to the input layer. The aim of this process squared error is obtained. The error on the validation set
is to adjust weights so that the mean squared error is is monitored during the training process. If the error on
minimized. This process is repeated until the error is the validation set decreases in comparison with the
minimized to a preference level [20, 21]. error obtained in the previous epoch, the processes is
continued. However, if it increases, the program is
In the current study, the sigmoid function is used as stopped and then the obtained weights are used in test
an activation function, which is given as follows: process. The programme is stopped either at the end of
the cross-validation process or when the number of
log -sig ( x ) = 1/ ( 1 + e ).
– ax
(5) maximum iteration is exceed. The weight values deter-
mined from the outcome of the cross-validation process
The input layer feeds the network with input from are obtained as an optimum values. In this study, initial
outside. The hidden layers link the input layer to the values of weight matrices are taken randomly in inter-
output layer, extract and remember useful features from val of [–0.5, +0.5] in the program. As the momentum
the input data to estimate the output of the network. term is used, adaptive learning rate is not considered
Trial and error should be carried out in order to deter- and learning rate is kept constant. The momentum term
mine an adequate number of hidden layers and the and learning rate are taken as 0.4 and 0.5 respectively.
number of neurons in each hidden layer [19, 20, 21]. The proposed architecture of ANN models consist-
ANN modelling is getting more popular, and has ing of an input layer, a hidden layer and an output layer
been used commonly in engineering tasks [22]. The are shown in Fig. 1. Wi, Ii and C represent weight val-
success has been achieved on geotechnical engineering ues of input, input values and weight values of bias
applications such as the site characterization [23, 24], respectively.
the soil classification [25], the engineering properties of Measured SPT-N value (Nfield) or corrected SPT-N
clays [26], the stress-strain behaviours of soils [27], the value (N60), natural water content (wn), liquid limit (wL)
bearing capacity of piles [28, 29], the liquefaction and plasticity index (Ip), are used as input to the neural
potential of coarse-grained soils [30, 31], the compac- network models. Hence, the input layer has four neu-
tion of soils [32, 33] and the permeability of clayey rons. The only output is undrained shear strength (Su),
soils [32, 34] and permeability of fine-grained soils and therefore the output layer has only one neuron. The
[35] in geotechnical engineering. optimum numbers of neurons in the hidden layer are

EURASIAN SOIL SCIENCE Vol. 42 No. 13 2009


COMPARISON OF ARTIFICIAL NEURAL NETWORKS MODELS 1491

Table 3. Information on ANN models


Output Variable, Su
Model Number Input Varibles
Neuron Number of Hidden Layer Number of Epoch Mean-square error, (kPa)2
UU Test 1 N-wn-wL-Ip 4 77 0.006015
n = 100 2 N60-wn-wL-Ip 2 44 0.006711
UC Test 3 N-wn-wL-Ip 7 3055 0.002041
n = 50 4 N60-wn-wL-Ip 6 735 0.003746

determined by training several networks with different The normalized data are then used to train neural
numbers of hidden neurons. For Models 1–4, two, four, network. The data from the output neuron have to be
six and seven hidden neurons are obtained respectively postprocessed to convert the data back into unnormal-
and given in Table 3. The performances of ANN models ized units according to
are measured in terms of desired output and error crite-
rion. The output obtained at the end of each trail is com- o = o min + ( o n – o min ) ( o max – o min )/ ( o max – o min ). (8)
pared with the target output to calculate the mean-
square error. The backpropagation is performed to The values of imin, imax, omin, and omax for ANN mod-
adjust the weights and biases until a mean-square error els are given in Table 4.
is minimized. The mean-square errors, which vary from
0.0020 to 0.0067, are achieved with different number of
epochs for all the models (Table 3).
2.1. Procedure for Estimating Su
The data sets used for training and testing the neural using Artificial Neural Network
networks were obtained from the database covering
cohesive soils from distinct geotechnical test sites in The Artificial Neural Network models described in
Turkey. They are non-sensitive clays with wL in the this study can be used to predict undrained shear
range of 22 to 110 and wp in the range of 14 to 44. The strength (Su) of a new site and the procedure can be eas-
SPTs were performed to the depth of 30 m. The training ily programmed into a computer, or conducted with a
data sets included data from each site that were selected calculator that enables to perform simple matrix opera-
randomly. The data sets consist of 100 data from UU tions. The input data are first normalized using Eq. (6).
test and 50 data from UC tests respectively. While The normalized output on is then obtained by propagat-
70 data were used for training for UU tests, 30 data ing the normalized input vector through the network as
were used for testing. In addition, 35 data from UC test follows:
for training and 15 data for testing were used.
o n = log sig [ W 2 × log sig ( W 1 × i n + C ) + D ], (9)
Normalizing of the training data is carried out so
that the processed data is in the range between 0 and 1. where W1 is weight matrix representing connection
The training data sets (input and target outputs) are nor- weights between the input layer neurons and hidden
malized as shown below:
i n = i min + ( i – i min ) ( i max – i min )/ ( i max – i min ), (6) Table 4. Normalized and unormalized parameters for ANN
models
o n = o min + ( o – o min ) ( o max – o min )/ ( o max – o min ), (7)
Input and Output
imin imax omin omax
where i is matrix of input vectors; o is matrix of target Variables
output vectors; in is matrix of normalized input vectors; UU Test N (N60) 4.0 (2.55) 46.0 (27.0)
on is matrix of normalized target output vectors; imin is
wn 15.0 66.0
vector containing the minimum values of the original
input; imax is vector containing the maximum values of wL 23.0 118.0
the original input; omin is vector containing the mini- Ip 2.0 84.0
mum value of the target output; omax is vector contain- Su – – 17.0 260.0
ing the maximum value of the target output. While con- UC Test N (N60) 4.0 (3.375) 44.0 (33.0)
stant values of i min and o min were considered as 0.1, wn 29.0 59.0
i max and o max are considered as 0.9. The preprocessing wL 50.0 118.0
of the training data is performed so that the processed Ip 27.0 84.0
data is in the range of 0.1 to 0.9. qu/2 = Su – – 11.0 232.0

EURASIAN SOIL SCIENCE Vol. 42 No. 13 2009


1492 SIVRIKAYA

5 5
ANN/ measured ANN S.D. = 0.299 Avg = 0.908
ANN S.D. = 0.316 Avg = 0.950
Normalized Su,

4 MLR S.D. = 0.352 Avg. = 0.965 4

ANN/ measured
ANN ANN

Normalized Su,
MLR S.D. = 0.341 Avg. = 0.978
3 MLR MLR
3
2
2
1
1
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
Data points 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
400 Data points
350 400
MLR
Predicted Su, kPa

300 350

Predicted Su, kPa


ANN MLR
250 300
ANN
200 250
150 200
100 150
50 100
50
0 50 100 150 200 250 300 350 400
Observed Su, kPa 0 50 100 150 200 250 300 350 400
Observed Su, kPa

Fig. 3. Comparison of results from Regression and ANN for Fig. 4. Comparison of results from Regression and ANN for
Model 1. Model 2.

layer; W2 is weight matrix representing connection regression analysis it is seen that the regression coeffi-
weights between the hidden layer neurons and the out- cients are at the order of a for N/N60, d for Ip, b for wn
put neuron; C is bias vector for the hidden layer neu- and c for wL from high to low.
rons; and D is bias for output neuron. The logical sig- The undrained shear strengths (Su) predicted by the
moid function is defined in Eq. (5). The normalized out-
ANN in test phase are shown in Figs. 3–6. The ANN
put on is then unnormalized using Eq. (8) to obtain the
models are found to be able to learn the relation
actual undrained shear strength values. between undreaned shear strength and SPT-N value
with index properties of cohesive soils. They show a
3. COMPARISON OF ANN PREDICTIONS relatively good agreement between the ANN predic-
WITH RESULTS OF CORRELATION tions and the actual data. The undreaned shear
EQUATIONS strengths predicted by the ANN models for SPT-Nfieid
and SPT-N60 are compared with the measured Su values
In performing the analyses of MLR and ANN, SPT- in Figs. 3b–6b.
N value (N/N60) and consistency indexes (wn, wL and Ip)
are used to predict Su by taking SPT corrections and test In general the undreaned shear strengths predicted
type (UU and UC tets) into consideration. The predic- by ANN models give reasonable results. In order to see
tions of the ANN models for testing are also compared how accurate the results of the ANN models are, the
with those calculated using the empirical methods. data estimated by the ANN models are normalized with
The multi linear regression (MLR) analyses are per- the measured Su and the results are plotted in a normal-
formed by using the data sets which consist of 100 data ized form (Su(ANN)/Su(measured)) as illustrated in
of Su from UU test, N/N60, wn, wL and Ip and 50 data of Figs. 3a–6a. The average values and the standard devi-
Su from UC tets, N/N60, wn, wL and Ip. The results of the ations of the normalized undreaned shear strengths for
regression analyses are given in Table 2. In general, the SPT-Nfield and SPT-N60 are given in Table 5.
correlations obtained from regression analysis give The average values and the standard deviations of
quite satisfactory results in terms of correlation coeffi- normalized undreaned shear strengths for SPT-Nfield are
cient and standard error. It is found that the best corre- in the range of 0.908 – 1.046 and 0.287 – 0.299, respec-
lation equation is obtained from data of Su from UC tively. However, they are in the range of 0.950 – 1.127
tets, N, wn, wL and Ip in terms of the correlation coeffi- and 0.316 – 0.331 respectively for SPT-N60. It is found
cient (r = 0.86) and standard error (s = ±28 kPa). In all that Model 4 including SPT-N60, wn, wL and Ip data give

EURASIAN SOIL SCIENCE Vol. 42 No. 13 2009


COMPARISON OF ARTIFICIAL NEURAL NETWORKS MODELS 1493

5 5
ANN S.D. = 0.287 Avg = 1.046 ANN S.D. = 0.331 Avg = 1.127
ANN/ measured
4 MLR S.D. = 0.608 Avg. = 1.157 4
Normalized Su,

ANN/ measured
MLR S.D. = 0.414 Avg. = 1.102

Normalized Su,
ANN ANN
3 MLR 3 MLR
2 2
1 1

0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
Data points 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
Data points
400
400
350
Predicted Su, kPa

MLR 350
300 MLR

Predicted Su, kPa


ANN 300
250 ANN
250
200
200
150
150
100
100
50
50
0 50 100 150 200 250 300 350 400
Observed Su, kPa 0 50 100 150 200 250 300 350 400
Observed Su, kPa
Fig. 5. Comparison of results from Regression and ANN for Fig. 6. Comparison of results from Regression and ANN for
Model 3. Model 4.

poor predictions in comparisons with the other ANN for using UU and UC test results. The comparisons are
models. The minimum value of 0.002041 (kPa)2 of plotted and illustrated in Figs. 3–6. It is seen that the
mean-square error is found in Model 3. From the stan- undrained shear strengths predicted by the ANN mod-
dard deviation point of view, compared with the MLR, els and the correlation equations are compatible with
the ANN models have given more reliable predictions each other and their trends are similar. However, all
in terms of estimating Su, and also Model 3 has a mini- ANN models give closer results than the correlation
mum standard deviation (0.287) of normalized Su. equations (Figs. 3a–6a). It is shown that the ANN mod-
Average value of normalized Su is found to be 1.046 in els developed in the current study give more reliable
Model 3. The ANN models including input data of N, predictions than the correlation equations. Model 3 and
wn, wL, Ip (Model 1 and 3) for the different test types are Model 4 have given the best prediction in terms of both
observed to give the best results, and are shown in MLR and ANN.
Figs. 3 and 5, respectively.
The trained weights (W1, W2) and biases (C, D)
The undrained shear strengths predicted by the obtained from the ANN models in this study are given
ANN models with SPT-Nfield and SPT-N60 are com- in Appendix. Thus, using Eqs. (6)–(9) with W1, W2, C,
pared with those predicted by the regression analyses D and Table 4 it will be possible to obtain the actual

Table 5. Comparison of ANN models with measured test results


ANN Regression
Normalized Su, Normalized Su,
Model
Su(ANN)/Su(measured) Su(regression)/Su(measured)
number Mean-square
error (kPa)2 Standard deviation, Corelation Standard Average, Standard
Average, Avg.
S.D. coefficient, r error, s (kPa) Avg. deviation, S.D.
1 0.006015 0.908 0.299 0.82 34 0.965 0.352
2 0.006711 0.950 0.316 0.81 34 0.978 0.341
3 0.002041 1.046 0.287 0.86 28 1.157 0.608
4 0.003746 1.127 0.331 0.87 27 1.102 0.414

EURASIAN SOIL SCIENCE Vol. 42 No. 13 2009


1494 SIVRIKAYA

undrained shear strength values for any input values in r : Correlation coefficient
each model. s : Standard error value for Su
S.D. : Standard deviation
4. CONCLUSIONS σv0 : Effective overburden pressure
This study deals with the comparison between the φ : Angle of shear strength
multi linear regression (MLR) and Artificial Neural
Network (ANN) analyses. The ANN models with a
feed forward, back-propapagation are developed to pre- ACKNOWLEDGMENTS
dict undrained shear strength (Su) of cohesive soils The authors are grateful to the companies of Yüksel,
from SPT data with index properties in Turkey. For this Toker, STFA, Istanbul Technical University and Gen-
purpose, a computer program is written in C language eral Directorate of Highways of Turkey for providing
using feed-forward back-propagation algorithm. In the borehole logs and laboratory test data.
terms of SPT correction and test type, four different
ANN models are constructed by using the input param-
eters of N/N60, wn, wL, Ip. The ideal network architec- APPENDIX
tures, for which network response compared to desired
ones is best, are determined for each ANN models. The
performance of the developed ANN models is investi- Weight Matrices and Biases Vectors for ANN Models
gated. The undrained shear strengths estimated by the Trained weigh matrices and biases vectors for
ANN models are also compared with empirical correla- Model 1.
tions. It is found that the ANN models give quite reli-
able predictions of the undrained shear strength in com-
parison with the correlation equations. They indicate – 1.641337 0.020699 0.114314 – 1.151760
that the ANN models have the superior performance
over the empirical methods. W 1 = 1.186435 – 0.967583 0.715441 1.484889
– 0.767256 – 0.743987 – 0.870100 – 0.409642
– 2.116101 0.361025 – 1.066565 – 1.384910
Nomenclature
SPT-N, N : SPT blow count 0.046065
SPT-N60, N60 : Blow count corrected to 60% of the
theoretical free-fall hammer energy; C = – 0.727639
– 0.137237
SPT-N60, N1, 60 : Blow count corrected to 100 kPa of
effective overburden pressure and 60% of the theoreti- 1.215493
cal free-fall hammer energy.
CN : Overburden correction factor
CE : Energy correction factor W 2 = – 1.830777 2.491047 – 0.565619 – 2.754919
CR : Rod length correction factor
D = – 0.226426
CB : Borehole diameter correction factor
CS : Liner correction factor Trained weight matrices and biases vectors for
CA : Anvil correction factor Model 2.
CBF : Blow count frequency correction factor
CC : Hammer cushion correction factor
ER : Hammer energy ratio W1 = 0.681193 – 1.590745 1.206406 0.926811
vs : Shear wave velocity – 2.156853 0.508218 – 2.013736 – 0.692355
mv : Coefficient of volume change
qu : Unconfined compressive strength C = 0.009636
Su : Undrained shear strength 1.424131
Dr : Relative density
wn : Water content W 2 = 2.012356 – 3.628955
wL : Liquid limit
D = – 0.332096
Ip : Plasticity index
a, b, c, d : Regression coefficients Trained weight matrices and biases vectors for
n : Data number Model 3.

EURASIAN SOIL SCIENCE Vol. 42 No. 13 2009


COMPARISON OF ARTIFICIAL NEURAL NETWORKS MODELS 1495

– 0.028020 – 0.668899 – 0.784396 1.426922 – 0.475710


– 8.365220 – 0.923563 0.286794 2.301130 – 0.353806
– 1.118238 – 0.522389 – 0.471086 – 1.041396 0.734285
W 1 = – 0.134636 – 1.649379 – 0.470940 2.931011 C = – 1.528042
– 1.334828 – 0.222841 0.803687 – 1.144850 0.031970
2.058268 1.983643 – 1.062060 – 4.009289 2.039184
– 0.512270 0.593175 – 1.149971 – 0.293065 – 0.198409

W 2 = 1.294322 – 5.417977 – 0.596980 2.614359 – 1.367820 – 3.389836 0.278718 ; D = 2.398733

Trained weight matrices and biases vectors for Model 4.

0.305404 – 2.625917 1.482484 3.136755 – 3.063223


– 2.890808 0.044873 0.261528 – 0.197483 – 0.294324
W 1 = – 1.146895 – 0.080578 – 0.594889 0.535454 C = – 1.049339
– 2.080694 1.159951 – 0.045858 – 0.865012 0.241180
0.229003 – 1.094909 0.607204 0.583468 – 1.396263
– 0.502698 – 0.583852 0.120863 – 0.581868 – 0.951017

W 2 = 3.662750 – 1.877834 – 1.013901 – 1.559210 0.936132 0.008299 ; D = 0.011605

REFERENCES 9. J. McGregor and J. M. Duncan, Performance and Use of


the Standard Penetration Test in Geotechnical Engineer-
1. M. W. Gardner and S. R. Dorling, “Artificial Neural Net- ing Practice Report of CGPR (Virginia Polytechnic
works (the Multilayer Perception)—A Review of Appli- Institute and State University, Virginia, 1998).
cations in the Atmospheric Sciences,” Atm. Environ. 32
(14/15), 2627–2636 (1998). 10. S. Saran, Analysis and Design of Substructure (Balkema,
Rotterdam, 1996).
2. H. T. Durgunoglu and E. Togrol, “Penetration Testing in
Turkey: State-of-the-Art Report,” Proc. of the 1st Euro- 11. M. A. Stroud, “The Standard Penetration Test in Insensi-
pean Symp. on Penetration Testing (Stockholm, Sweden, tive Clays and Soft Rock,” Proc. of the 1st European
1974) p. 137. Symp. on Penetration Testing (Stockholm, Sweden,
1974).
3. L. Décourt, “The Standard Penetration Test: State-of-
the-Art-Report,” Norwegian Geotechnical Institute Pub- 12. O. Sivrikaya and E. Toğrol, “Relations between SPT-N
lication No. 179, (Oslo, Norway, 1990), p. 1–12. and qu,” 5th Intern. Congress on Advances in Civil Engi-
neering (Istanbul, Turkey, 2002), pp. 943–952.
4. H. M. Horn, “North American Experience in Sampling
and Laboratory Dynamic Testing,” ASTM Geotechn. 13. V. De Mello, “The Standard Penetration Test - A State-
Testing J. 2 (2), 84–97 (1979). of-the-Art Report,” Proc. of the 4th Pan-American Conf.
on Soil Mechanics and Foundation Engineering (San
5. H. Mori, “Review of Japanese Subsurface Investigation Juan, Puerto Rico, 1971).
Techniques,” J. South East Asian Soc. Geotechn. Engin.
10, 219–242 (1979). 14. J. K. Mitchell, F. Guzikowski, and W. C. B. Villet, “The
Measurement of Soil Properties in Situ: Present Meth-
6. I. K. Nixon, “Standard Penetration Test: State of the Art ods, Their Applicability and Potential,” Lawrence Berke-
Report,” Proc. of the 2nd Eur. Symp. on Penetration Test- ley Lab. Rep. No. 6363 (Univ. of California, Berkeley,
ing (Amsterdam, the Netherlands, 1982), pp. 3–21. 1978).
7. D. P. Coduto, Foundation Design: Principles and Prac- 15. C. R. I. Clayton, “SPT Energy Transmission: Theory,
tice (Prentice-Hall, USA, 1994). Measurement and Significance,” Ground Engin. 23 (10),
35–43 (1990).
8. B. B. Broms and N. Flodin, “History of Soil Penetration
Testing,” Proc. of 1st Intern. Symp. on Penetration Test- 16. H. B. Seed, K. Tokimatsu, L. F. Harder, and R. M. Chung,
ing (Orlando, Norway, 1988), pp. 157–220. “Influence of SPT Procedures in Soil Liquefaction

EURASIAN SOIL SCIENCE Vol. 42 No. 13 2009


1496 SIVRIKAYA

Resistance Evaluations,” ASCE J. Geotech. Engin. 111 26. P. U. Kurup, N. K. Dudani, “Neural Networks for Profil-
(12), 1425–1445 (1985). ing Stress History of Clays from PCPT Data,” ASCE J.
Geotechn. Geoenviron. Engin. 128 (7), 569–579 (2002).
17. A. W. Skempton, “Standard Penetration Test Procedures
and the Effects in Sands of Overburden Pressure, Rela- 27. D. Penumadu, L. Jin-Nan, J. L. Chameau, and S. Aru-
tive Density, Particle Size, Aging and Overconsolida- mugam, “Rate Dependant Behavior of Clays Using Neu-
tion,” Geotechnique 36 (3), 425–447 (1986). ral Networks,” Proc. of the 13th Intern. Conf. on Soil
Mechanics and Foundation Engineering (New Delhi,
18. K. Tokimatsu, “Penetration Tests for Dynamic Prob- India, 1994).
lems,” Proc. of the 1st Intern. Symp. on Penetration Test-
ing (Orlando, Norway, 1988). 28. W. T. Chan, Y. K. Chow, and L. F. Liu, “Neural Net-
work—Alternative to Pile Driving Formulas,” Comp.
19. N. Kartam, I. Flood, and J. H. Garrett, Artificial Neural Geotechn. 17 (2), 135–156 (1995).
Networks for Civil Engineers: Fundamentals and Appli- 29. I. M. Lee and J. H. Lee, “Prediction of Pile Bearing
cations (ASCE, New York, 1997). Capacity Using Artificial Neural Networks,” Comp.
20. I. Flood and N. Kartam, “Neural Network in Civil Engi- Geotechn. 18 (3), 189–200 (1996).
neering I: Principles and Understandings,” ASCE J. 30. A. T. C. Goh, “Neural Network Modeling of CPT Seis-
Comput. Civil Eng. 8 (2), 131—148 (1994). mic Liquefaction Data,” ASCE J. Geotechn. Engin. 122
21. I. Flood and N. Kartam, “Neural Network in Civil engi- (1), 70–73 (1996).
neering II: Systems and Applications,” ASCE J. Comput. 31. D. Ural and H. Saka, “Liquefaction Prediction by Neural
Civil Eng. 8 (2), 149—162 (1994). Networks,” Electr. J. Geotechn. Engin., 1089–3032
(1998).
22. M. A. Shahin, M. B. Jaksa, and H. R. Maier, “Artificial
Neural Network Applications in Geotechnical Engineer- 32. Y. M. Najjar, I. A. Basheer, and W. A. Naouss, “On the
ing,” Austr. Geomech., 49–62 (2001). Identification of Compaction Characteristics by Neuro-
nets,” Comp. Geotechn. 18 (3), 167–187 (1996).
23. Y. Zhou and X. Wu, “Use of Neural Networks in the 33. I. A. Basheer, “Empirical Modeling of the Compaction
Analysis and Interpretation of Site Investigation Data,” Curve of Cohesive Soils,” Can. Geotechn. J. 38 29–45
Computer and Geotechnics 16, 105–122 (1994). (2001).
24. D. M. Rizzo and D. E. Dougherty, “Application of Arti- 34. M. M. Gribb and G. W. Gribb, “Use of Neural Networks
ficial Neural Network for Site Characterization Using for Hydraulic Conductivity Determination in Unsatur-
Hard and Soft Information,” Proc. of 10th Intern. Conf. ated Soil,” Proc. 2nd Intern. Conf. on Ground Water
on Computational Methods in Water Resources (Kluwer Ecology (Atlanta, USA, 1994), pp. 155–163.
Academic, Dordrecht, 1994), Part 12, pp. 793–799.
35. S. Akbulut, “Artificial Neural Networks for Predicting
25. V. Cal, “Neural Network—Alternative to Pile Driving the Hydraulic Conductivity of Coarse-Grained Soils,”
Formulas,” Adv. Engin. Software 22 (2), 95–97 (1995). Eur. Soil Sci. 38 (4), 392–398 (2005).

EURASIAN SOIL SCIENCE Vol. 42 No. 13 2009

You might also like