Sie sind auf Seite 1von 6

Soil Dynamics and Earthquake Engineering 23 (2003) 631636

www.elsevier.com/locate/soildyn

Evaluation of liquefaction potential using neural-networks


and CPT results
M.H. Baziara,*, N. Nilipourb,1
a
College of Civil Engineering, Iran University of Science and Technology, Narmak, Tehran 16844, Iran
b
Laboratory of Hydraulic Constructions, Department of Civil Engineering, Swiss Federal Institute of Technology, CH-1015 Lausanne, Switzerland
Accepted 30 April 2003

Abstract
In this research, a reliable Cone Penetration Test data set was gathered with a wide range of parameters. This data was incorporated in a
Neural-Networks computer software called STATISTICA Neural-Networks. The back propagation algorithm with a multilayer perceptron
network is utilized to analyze the liquefaction occurrence in different sites. In this study, different sets of effective parameters for the neural-
network analyses are selected such that to reduce the noise and to obtain more accurate results.
Considering the relative importance of effective parameters in liquefaction assessment, it is indicated that s0 ; s00 together play a more
important role than what previously was assumed and hence the relative importance of the qc and seismic parameters are decreased compared
with the previous works. The results presented here have more accuracy than previous works while at the same time, the range of the
parameters used in this study is much wider than what was previously used. This range of parameters makes the proposed method applicable
for practical purposes.
q 2003 Elsevier Ltd. All rights reserved.
Keywords: Neural-network; Back propagation; Liquefaction; Cone penetration test

1. Introduction difficulties in obtaining undisturbed samples of loose


sandy soils, since 1980s many researchers have preferred
On many occasions, liquefaction occurs during earth- to use in situ tests to evaluate liquefaction and lateral
quakes. It is known as one of the most destructive spreading potential [2]. The Standard Penetration Test
phenomena caused by earthquake and has been widely because of its simplicity was one of the first in situ tests to be
seen in loose saturated sands soil deposit (Niigata 1964, widely used. Later Robertson and Companella [3], Shibata
Alaska 1964, Tangshan 1979, Loma Prieta 1989, Kobe and Teparaska [4] and Stark and Olson [5] adopted the Cone
1995, Turkey 1998, etc.). Increase in pore water pressure, Penetration Test (CPT) for evaluating liquefaction potential.
due to seismic shaking by the earthquake, causes the loose In order to evaluate liquefaction potential, it is necessary
saturated sandy soils to liquefy and lateral spreading, sand to find out what parameters control liquefaction occurrence,
boils, ground oscillations, settlements, etc. to appear at the including: earthquake parameters, soil properties, and stress
surface. conditions. In recent years, Artificial Intelligence has
Since 1970, scientists have conducted extensive research introduced a useful and powerful computational tool,
and have proposed many methods to predict the occurrence Artificial Neural-Networks (NN), for solving complicated
of liquefaction. In the beginning, undrained cyclic loading multifactor problems. By using this powerful tool, liquefac-
laboratory tests had been used to evaluate the liquefaction tion potential can be estimated more accurately than by
potential of a soil under a special cyclic loading, which conventional methods [6,7]. For this type of analysis, there
simulates an earthquake excitation [1]. But due to is no need to normalize and calibrate parameters since NN
can consider all effective parameters together and find the
* Corresponding author. Tel./fax: 98-21-7491031.
E-mail addresses: baziar@iust.ac.ir (M.H. Baziar); nima.nilipour@
relations between them. In the first attempt at using NN and
epfl.ch (N. Nilipour). CPT data for liquefaction assessment, researchers achieved
1
Tel.: 41-21-634-6359; fax: 41-21-693-2264. the same accuracy as the conventional method [8]. Using
0267-7261/03/$ - see front matter q 2003 Elsevier Ltd. All rights reserved.
doi:10.1016/S0267-7261(03)00068-X
632 M.H. Baziar, N. Nilipour / Soil Dynamics and Earthquake Engineering 23 (2003) 631636

a database with a wider range of parameters and larger data


set, and selecting the proper parameters as input, a NN
model can be improved with respect to accuracy, and the
relative importance of each parameter will be established
more reliably. It is also anticipated that utilizing powerful
NN software enables us to get more accurate results than the
previous works.

2. Neural-networks

One of the more popular types of network is the


Multi-Layer Perceptron associated with the back
propagation algorithm [9], which is a supervised learning
algorithm.
Back propagation is widely used in prediction problems
such as weather forecasting, and stock market prices. In this
algorithm, weight of the synapses and connection between
neurons are modified in a way that the corresponding output
of each input data gets closer to the desired value. In other
words, the global error of the network gets minimized. The
most common error function is sum-squared function. In
this research, in order to utilize NN, powerful software,
STATISTICA Neural-Network has been used. The formu-
lation of Back Propagation algorithm in STATISTICA
Fig. 1. Schematic figures of an artificial neural-network.
Neural-Network is defined as [10]
Dwi;j t hdj oi aDwi;j t 2 1 1
momentum. Since the Back Propagation algorithm
where wi;j is synapses weight between neuron i and j; t is explores the error surface in order to find global minimum
number of epochs, a is momentum, h is learning rate, dj is error, repeating the training process and initializing the
local gradient error and oi is output of neuron i: values of synapses randomly each time, will increase the
The important characteristic of this software is its chance of reaching the global minimum error. As a matter
flexibility, which helps users to do the learning process of fact, one of the advantages of the NN is its ability for
successfully and minimize the error as much as possible. looking for minimum global error in an n-dimensional
The Multi-Layer Perceptron should have at least three layers space instead of just checking errors in different sets of
including: input, hidden, and output layers (Fig. 1). Multi- two- or three-dimensional spaces. In other words for the
Layer Perceptron can have more than one hidden layer, but case of liquefaction studied in this paper, instead of trying
usually one hidden layer is enough for modeling any to show boundaries in two-dimensional graphs, using
complicated problem. Actually the most important stage in calibrated parameters, or eventually three-dimensional
designing Multi-Layer Perceptron is to optimize the number graphs by presenting a liquefaction boundary for each
of neurons in the hidden layer. If the number of neurons in type of liquefiable soils, such as what has been done in
hidden layer is less than optimal, the training process cannot conventional methods, the NN model presents the
reach the global minimum error or the network is not able to boundary in a n-dimensional space.
learn properly. If the number of neurons in hidden layer is Also by making proper changes in the controlling
more than optimal, the possibility of over-learning or over- parameters, learning rate and momentum, or using especial
fitting becomes a problem. options such as noise and jug weight, different paths in the
Considering the number of neurons in input/output error surface can be explored to find the global minimum
layers, the complexity of problem and the size of the error. In other words, having constant values of learning
database, the network designer should determine the rate, momentum and noise during training process and using
optimal number of neurons in hidden layer [10]. same values for all synapses in the initializing stage, limits
After designing the network; with selection of the the chance of reaching the global minimum error. It has to
proper number of neurons in each layer, suitable be noted that the desired minimum error should be
activation/error and summation functions, and preparing considered for both training and testing data sets. In other
the training and testing data sets, the learning process words, over-learning causes high accuracy in training stage
should be initiated. The initialization process includes: but low accuracy in testing stage, which is not acceptable for
specifying initial values of synapses, learning rate and prediction purposes.
M.H. Baziar, N. Nilipour / Soil Dynamics and Earthquake Engineering 23 (2003) 631636 633

3. Data set preprocessing includes scaling and shifting of all values of


data set in a way that all input parameters vary in the same
In order to solve a problem utilizing NN, having a proper range, e.g. [0,1], using mean value and difference of
data set is the most important requirement. For the data set, minimum and maximum values of each parameter.
three main characteristics are required: (1) reliability, which
includes being real and accurate, (2) having sufficient data
in relation to the dimensions and complexity of the problem 5. Data analysis
and (3) covering all aspects of the problem.
Stark in 1995 [5] compiled the 180 cases and proposed a Considering range of the output [0,1], the logistic
method to assess liquefaction potential. Robertson and function is the proper activation function. Other functions
Wride [11] later used this data set to propose new method such as the sigmoid function also were tested, but results
for liquefaction assessment. They also introduced a new confirmed the superiority of the logistic function as an
variable called Ic for their liquefaction assessment. Finally, activation function for modeling liquefaction using NN.
Youd et al. [12], used the data set compiled by Stark and Other parts of network such as the number of neurons in the
Suzuky and modified the Robertson boundaries for assess- hidden layer, learning rate, momentum and noise level have
ment of liquefaction potential. Due to the fact that there are to be chosen during the training process, depending on the
some questions regarding validity of Ic for soil classification circumstances and by monitoring error level.
especially for silty sand material, it was decided to focus on In this study, normalized and calibrated parameters such
Stark method. Furthermore, accuracy of two new methods is as qc1 ; and SSR unlike earlier studies have not been used as
not much different with the method proposed by Stark. an input parameter, due to making the data set noisy and not
In this research, 170 field case histories were selected suitable for the training process. In other words, the main
from a total of 180 case records compiled by Stark [5]. The reason for utilizing NN in complicated and multifactor
10 cases, discarded in this study, were those with some problems is its ability to find reliable relations between basic
effective parameters such as soil mean grain size D50 or qc parameters while analyzing all the effective parameters at
missing. The data set consists of 16 cases from Japan the same time.
(Niigata 1964 and Nihonkai Chubu 1983), 57 cases from At the beginning of the training process, random values
USA (San Fernando Valley 1971, Imperial Valley 1979 and were used as weights. Since during the training process, we
Loma Preiata 1989), 85 cases from China (Haicheng 1975 explore the error surface for a global minimum, the initial
and Tangshan 1976), 5 cases from Romania (Vrancea 1977) values play important role in this procedure. Therefore for
and 7 cases from Canada (Sanguenay 1988). Liquefaction each NN model, the training process has been conducted
occurred at 104 of these sites. The remaining 66 sites did not several times, each time with different random synapses
liquefy. Since the percentage of fines content, FC, was not values to reach results with least errors.
available for all cases, FC cannot be considered as an input In order to compare the performance of the software and
parameter in the process. methodology used in the training process, proposed model
In all cases, the appearance of sand boils, settlement or in Goh [8], B5, with the same data set is trained, utilizing
lateral ground spreading were as indications of liquefaction STATISTICA Neural-Network. As it is shown in Table 1,
occurrence. obtained relative importance using this software is compar-
The data set used in this work has more case records than able with values reported by Goh [8]. It can be seen that the
the data set used in the previous work [8]. The Imperial relative importance of qc in B5 Model is very significant
Valley 1979 data compiled by Shibata and Teparaska [4] compared with other parameters. The same input parameters
and used in Goh [8], have different maximum surface (D50 ; a; s00 ; qc ; M) were used in training model N5, using
accelerations, amax ; from those reported by Stark [5]. selected data set in this study. The relative importance of
Considering the reference research performed by Youd parameters was different and despite larger number of cases,
and Bennett [13] and other investigators [14 16], it seems the error was less. According to Table 1, relative importance
that the 0.6 g reported by Stark [5], is more reliable than the
0.8 g used by Goh [8] for amax in Heber Road site during Table 1
Imperial Valley 1979 earthquake. Comparison of obtained relative importance (%) of parameters

Relative importance (%)


4. Preprocessing D50 a s00 qc M

Before using a data set to train the NN, in order to avoid B5 in Goh paper [8] 12.9 17.2 16.2 36.6 17.3
any inaccuracy and obtain better training results, the data set B5, with STATISTICA 15.7 18.1 14.4 40.8 11.0
should be preprocessed. Replacing yes with 1 for and Goh data
N5 B5 with STATISTICA 17.2 15.3 27.6 26.9 13.0
liquefaction cases and no with 0 for non-liquefaction and Stark data
cases, the data set becomes completely numerical. Numerical
634 M.H. Baziar, N. Nilipour / Soil Dynamics and Earthquake Engineering 23 (2003) 631636

Table 2
Results and details of designed neural-networks models

Model Input parameters No. of neurons in Error and accuracy (%)


hidden layer
Training Testing

N5 D50 ; a; s00 ; qc ; M 5 1 99 1 97
N6-s0 D50 ; a; s00 ; qc ; M; s0 5 0 100 1 97
N6-FC D50 ; a; s00 ; qc ; M; FC 5 0 100 2 80
N7-Z D50 ; a; s00 ; qc ; M; s0 ; Z 6 0 100 2 94

of s00 and qc in N5 model are significantly changed (about the multidimensional space, five-dimensional in case of
13%) compared with B5 models and it shows almost the Model N5 and six-dimensional in case of Model N6,
same relative importance value for s00 as for qc : This finding liquefaction boundaries could be presented by projecting the
indicates that in liquefaction potential assessment, effective obtained multidimensional boundary on two- or three-
stress plays a more important role than what model B5 dimensional sub-spaces. Therefore, in any sets of two- or
represented. Due to a higher level of accuracy and a wider three-dimensional sub-space, the error will not be more than
range of parameters in model N5, it can be concluded that the global error in multidimensional reference space.
the proposed relative importance values are more realistic However, these projected boundaries are only valid for those
than those suggested by model B5 in Ref. [8]. fixed value of the effective parameters based on which the
The methodology for calculating the relative importance, boundaries are projected and derived. Hence for a hypothetical
used in this study, was suggested by Garson [17]. earthquake with constant value of M and, amax ; liquefaction
Although adding more effective parameters as input boundaries can be obtained for specific stress condition, using
parameter to NN may logically lead to more accurate results, Model N6-s0 : The results show that corresponding liquefac-
but the unavailability and difficulty of achieving more useful tion boundaries of different values of effective stress in
parameters in liquefaction assessment, such as relative function of D50 ; and qc are almost parallel. This reveals the
density, void ratio, etc. considered as the main limitations. direct relation between effective stress and liquefaction
The effective parameters in liquefaction assessment, in occurrence, while the total stress assumed to be constant.
general, can be classified into three main groups: (1) stress As it is seen in Table 2, the other three models have zero
conditions, (2) soil parameters and (3) seismic parameters. error in the training process and one or two errors in the
The most usual and available effective parameters are testing process. It can be concluded that the N6-s0 model
total stress, effective stress and depth in Group (1), median gives higher accuracy and lower relative importance for qc
grain diameter and fines content in Group (2), and with respect to the other presented models. Although
magnitude and peak acceleration of earthquake in Group
(3). Since stress conditions and soil parameters affect CPT
results, qc can be placed in both Groups (1) and (2). As qc is
affected by several parameters, models with additional
available input parameters such as s0 and Z (depth of soil
layer) from Group (1) and FC from Group (2) are designed
in order to decrease the relative importance of qc : The
lowering of relative importance of qc as well as the exertion
of the direct effect of additional parameters decrease the
noise of the data set and lead to obtain models with better
training and generalization results.
N6- s0 ; N7-Z and N6-FC models with additional
parameters with respect to N5 model were studied. Table 2
shows the input parameters, number of neurons in hidden
layer and accuracy in the training and testing phases for
different NN models studied in this paper. As it is seen for
the N5 model, there is one error in the training and one error
in the testing processes.
Fig. 2 depicts the projection of five dimensional space (D50 ;
a; s00 ; qc ; M) in two-dimensional spaces of D50 ; and qc for
the N5 model. As it is seen for the Tangshan Earthquake and
with s00 100 kPa, there is only one error for training stage. It Fig. 2. Liquefaction boundary for 1976 Tangshan Earthquake (M 7:8;
should be mentioned that after minimizing the global error in amax 0:2 g and s00 100 kPa) using model N5.
M.H. Baziar, N. Nilipour / Soil Dynamics and Earthquake Engineering 23 (2003) 631636 635

Table 3 Table 5
Obtained relative importance (%) of input parameters Range of parameters in the reference data set

Relative importance (%) s0 s00 s0 2 s00 qc D50 M a


(kPa) (kPa) (kPa) (kPa) (mm) (Richter) (g)
D50 a s00 qc M s0 FC Z
Min 16.7 13.9 0 0.38 0.016 5.9 0.1
N5 17.2 15.3 27.6 26.9 13.0 Max 296.3 227.5 123.8 26 0.48 7.8 0.6
N6-s0 14.4 14.0 14.2 23.1 11.3 23.0 Average 114.48 78.69 35.79 6.47 0.17 7.29 0.28
N6-FC 13.4 8.1 29.9 26.9 14.0 7.7
N7-Z 15.4 12.5 17.6 14.7 14.3 10.9 14.5
one error in the testing phase. It should be noted that 10
cases from Loma Prieta 1989, reported by Stark [5], due to
relative importance of s00 decreased, the sum of relative
uncertainty in D50 and qc were discarded in this research,
importance values of stress condition parameters (s00 and
and one of the cases which showed an error in the proposed
s0 ) increased. Since s00 and s0 are somehow dependent, it is
boundary for clean sand by Stark [5] was among these 10
better to consider their relative importance together as stress
discarded cases.
condition importance. Also having both s00 and s0 as input
Table 4 summarizes the accuracy of the different studies
parameters implicitly takes into account the effect of ground
mentioned in this paper. It can be seen that the conventional
water level and pore water pressure which are important
method of Stark [5] with the same data set as used in this
parameters in liquefaction occurrence.
research has the accuracy of almost 90%. It should be
Due to unavailability of FC for all cases, the data set used in
mentioned that other conventional methods, such as NCEER
model N6-FC has in total 80 cases, where 70 cases are selected
method [12], have the same accuracy as the Stark method.
for training and 10 cases for testing. Since both FC and D50 are
The previous NN study [8] also has the total accuracy of
soil parameters, by adding FC as an input parameter, a
97.2% but for 109 cases out of 180 cases reported by Stark,
decrease in relative importance of D50 was predictable.
while the presented study with model N6-s0 shows higher
In model N7-Z another stress condition parameter, Z; is
accuracy, 99% in total, despite larger number of cases.
added as an input parameter. In conventional methods [3,5]
The range of parameters of reference data used in this
a function of depth, rd ; was used to calibrate SSR and
study is presented in Table 5. As it is shown in this table, the
consider the effect of depth in order to determine a boundary
results of this work can be applicable to the most of practical
for liquefaction occurrence with qc1 and SSR. In this model
purposes. It should be mentioned that using data set with a
the effect of Z is directly exerted to calculation and caused
wider range or with the same range but with a larger number
considerable decrease of relative importance of qc : Table 3
of cases improves the generalization and applicability of
presents the relative importance of parameters for each
this method.
model obtained in this study. Due to the change in effective
parameters, relative importance of parameters changed.
In comparison with the conventional method using the
same data set [5], the results in this study are more 6. Summary and conclusions
accurate. Proposed liquefaction-potential boundaries given
by Stark [5] were not able to predict correctly 17 cases NN is a powerful tool in predicting the liquefaction
among a total of 180 cases (two errors for Clean Sand, potential and more accurate results than the conventional
seven errors for Silty Sand and eight errors for Silty Sand methods are obtained. No calibration and normalization
to Sandy Silt) while the NN results for the proposed model with respect to the other parameters is needed. Also the
of N6-s0 showed zero error in the training process and just relative importance of the effective parameters can be
compared.
Table 4 Using normalized and calibrated parameters make the
Accuracy of each method data set noisy and are not suitable for the training process. In
Total Total Total Number of cases and other words, NN has the ability to find the relations between
number number accuracy% accuracy basic parameters of a multifactor problem, helping to
of cases of errors achieve better results.
Training Testing Adding s0 as an input parameter causes lowering of
relative importance of qc and decreasing of the global error
Stark 180 17 90.5
Stark (with 170 16 90.6 in the network. s00 and s0 together play a more important
same cases role than qc in liquefaction assessment. s00 directly affects
as this study) the liquefaction potential, since obtained boundaries with
Goh [8] 109 3 97.2 74, 98.6% 35, 94.2% different values of s00 are almost parallel.
This study 170 1 99.4 134, 100% 36, 97.2%
In order to apply the results of proposed NN model for
N6-s0
practical problems, the range of the parameters in
636 M.H. Baziar, N. Nilipour / Soil Dynamics and Earthquake Engineering 23 (2003) 631636

the reference data set has to be considered. Having more [8] Goh ATC. Neural-network modeling of CPT seismic liquefaction
CPT data, the proposed NN model can be improved in data. J Geotech Engng, ASCE 1996;122(1):70 3.
[9] Rumelhart DE, Hinton GE, Williams RJ. Learning internal represen-
generalization and applicability.
tations by error propagation. In: Rumelhart DE, McClelland JL,
editors. Parallel distributed processing, vol. 1. Cambridge, MA: MIT
Press; 1986.
Acknowledgements [10] Electronic Manual for STATISTICA Neural Networks, Release 3.0
D; 1998.
We are very grateful to Professor Caro Lucas for his [11] Robertson PK, Wride CE. Evaluating cyclic liquefaction potential
using the cone penetration test. Can J Geotech, Ottawa 1998;35(3):
Neural-Network advisory and providing technical support
442 59.
from School of Intelligent Systems, Institute for Studies in [12] Youd TL, Idriss IM, Andrus RD, Arango I, Castro G, Christian JT,
Theoretical Physics and Mathematics (IPM), Tehran, Iran. Dobry R, Liam Finn WD, Harder Jr. LF, Hynes ME, Ishihara K,
Koester JP, Liao SSC, Marcuson III WF, Martin GR, Mitchell JK,
Moriwaki Y, Power MS, Robertson PK, Seed RB, Stokoe II KH.
References Liquefaction resistance of soils: summary report from the 1996
NCEER and 1998 NCEER/NSF workshops on evaluation of
liquefaction resistance of soils. J Geotech Geoenviron Engng,
[1] Castro G, Poulos SJ, France JW, Enos JL. Liquefaction induced by
ASCE 2001;127(10):81733.
cyclic loading. Winchester, Mass: Geotechnical Engineers Inc.; 1982.
[13] Youd TL, Bennett MJ. Liquefaction sites, imperial valley, California.
[2] Seed HB, Idriss IM, Arango I. Evaluation of liquefaction potential
J Geotech Engng, ASCE 1983;109(3):44057.
using field performance data. J Geotech Engng, ASCE 1983;109(3):
45882. [14] Youd TL, Bartlett SF. US case histories of liquefaction-induced
[3] Robertson PK, Campanella RG. Liquefaction potential of sands using ground displacement. Proceedings of the first Japan US Workshop
the CPT. J Geotech Engng, ASCE 1985;111(3):384 403. on Liquefaction, Large Ground Deformation and their Effects on
[4] Shibata T, Teparaska W. Evaluation of liquefaction potentials of soils Lifeline Facilities; 1988. p. 2231.
using cone penetration tests. Soils Found, Tokyo, Jpn 1988;28(2): [15] Joyner WB, Boore DM. Peak horizontal accelerations and velocity
4960. from strong-motion records from the 1979 Imperial Valley,
[5] Stark TD, Olson M. Liquefaction resistance using CPT and field case California, earthquake. Bull Seismol Soc Am 1981;71:2011 38.
histories. J Geotech Engng, ASCE 1995;121(12):856 69. [16] Bartlett SF, Youd TL. Empirical analysis of horizontal ground
[6] Goh ATC. Seismic liquefaction potential assessed by neural networks. displacement generated by liquefaction-induced lateral spread.
J Geotech Engng, ASCE 1994;120(9):146780. Technical Report NCEER-92-0021, Buffalo, NY: National Center
[7] Wang J, Rahman MS. A neural network model for liquefaction- for Earthquake Engineering Research; 1992.
induced horizontal ground displacement. Soil Dyn Earthquake Engng [17] Garson GD. Interpreting neural-networks connection weight. AI
1999;18:555 68. Expert 1991;6(7):47 51.

Das könnte Ihnen auch gefallen