Documentos de Académico
Documentos de Profesional
Documentos de Cultura
ORIGINAL PAPER
Resumen Abstract
Este artculo presenta y discute el uso de las This paper presents and discusses the use
redes neuronales para determinar la duracin of neural networks to determine strong
de los movimientos fuertes del terreno. Para ground motion duration. Accelerometric data
tal efecto se desarroll un modelo neuronal, recorded in the Mexican cities of Puebla and
utilizando datos aceleromtricos registrados en Oaxaca are used to develop a neural model
las ciudades mexicanas de Puebla y Oaxaca, that predicts this duration in terms of the
que predice dicha duracin en trminos de la magnitude, epicenter distance, focal depth,
magnitud, distancia epicentral, profundidad soil characterization and azimuth. According
focal, caracterizacin del suelo y el azimut. Por to the above the neural model considers
lo que, el modelo considera los efectos tanto the effect of the seismogenic zone and the
de la zona sismognica como del tipo de suelo contribution of soil type to the duration of
en la duracin del movimiento. El esquema VWURQJ JURXQG PRWLRQ 7KH QDO VFKHPH
final permite una estimacin directa de la permits a direct estimation of the duration
duracin a partir de variables de fcil obtencin since it requires easy-to-obtain variables and
y no se basa en hiptesis restrictivas. Los does not have restrictive hypothesis. The
resultados presentados en este artculo indican results presented in this paper indicate that
que la alternativa del cmputo aproximado, the soft computing alternative, via the neural
particularmente las redes neuronales, es una model, is a reliable recording-based approach
poderosa aproximacin que se basa en los to explore and to quantify the effect of seismic
UHJLVWURVVtVPLFRVSDUDH[SORUDU\FXDQWLFDUORV and site conditions on duration estimation. An
efectos de las condiciones ssmicas y de sitio en HVVHQWLDO DQG VLJQLFDQW DVSHFW RI WKLV QHZ
la duracin del movimiento. Un aspecto esencial model is that, while being extremely simple,
\VLJQLFDQWHGHHVWHQXHYRPRGHORHVTXHD it also provides estimates of strong ground
pesar de ser extremadamente simple ofrece motions duration with remarkable accuracy.
HVWLPDFLRQHVGHGXUDFLyQFRQQRWDEOHHFLHQFLD $GGLWLRQDO EXW LPSRUWDQW VLGH EHQHWV DULVLQJ
$GLFLRQDO H LPSRUWDQWH VRQ ORV EHQHFLRV TXH from the models simplicity are the natural
arroja esta simplicidad sobre la separacin separation of source, path, and site effects and
natural de los efectos de la fuente, patrn o WKH DFFRPSDQ\LQJ FRPSXWDWLRQDO HIFLHQF\
GLUHFWLYLGDG\GHVLWLRDGHPiVGHODHFLHQFLD
computacional. Key words: strong ground motion duration,
JURXQGPRWLRQSDUDPHWHUVVLJQLFDQWGXUDWLRQ
Palabras clave: duracin del movimiento de ULDV,QWHQVLW\QHXUDOQHWZRUNVVRIWFRPSXWLQJ
terreno, parmetros de movimientos de terreno,
GXUDFLyQVLJQLFDWLYDLQWHQVLGDGGHULDVUHGHV
neuronales, cmputo aproximado.
L. Alcntara Nolasco
S. Garca
E. Ovando-Shelley
M. Antonio Macas Castillo
Instituto de Ingeniera
Universidad Nacional Autnoma de Mxico
Delegacin Coyoacn, 04510
Mxico D.F., Mxico
*
Corresponding author: leonardo@pumas.ii.unam.mx
221
L. Alcntara Nolasco, S. Garca, E. Ovando-Shelley and M.A. Macas Castillo
by the dendrites of other neurons. Figure 1 restricted in the same way and can in principle
VKRZV KRZ D VLPSOLHG QHXURQ ORRNV OLNH be between and , but they are very often
When a neuron receives enough electric pulses small values centered on zero (Broomhead and
WKURXJK LWV GHQGULWHV LW DFWLYDWHV DQG UHV D Lowe, 1988). Figure 2 provides a schematic
pulse through its axon, which is then received YLHZRIDQDUWLFLDOQHXURQ
by other neurons. In this way information
can propagate through the NN. The synaptic As mentioned earlier there are many
connections change throughout the lifetime of different activation functions, some of the most
a neuron and the amount of incoming pulses commonly used are threshold (Eq. 2), sigmoid
needed to activate a neuron (the threshold) (Eq.3) and hyperbolic tangent (Eq.4).
also change. This process allows the NN to
learn (Tettamanzi and Tomassini, 2001). 1 if x + t > 0
g( x ) =
Mimicking the biological process the
0 if x + t 0
(2)
DUWLFLDO 11 DUH QRW LQWHOOLJHQW EXW WKH\ DUH
FDSDEOH IRU UHFRJQL]LQJ SDWWHUQV DQG QGLQJ 1
the rules behind complex data-problems. A g( x ) = 2 s ( x t )
VLQJOHDUWLFLDOQHXURQFDQEHLPSOHPHQWHGLQ 1+ e (3)
many different ways. The general mathematic
GHQLWLRQLVJLYHQE\HTXDWLRQ
sinh( s( x + t ))
g( x ) = tanh( s( x + t )) = =
n
cosh( s( x + t ))
y( x ) = g w x i
i=0 (1) es ( x +t ) es ( x +t ) e2 ( s ( x +t )) 1
=
where x is a neuron with n input dendrites (x0,, e s ( x +t ) + es ( x +t ) e2 ( s ( x +t )) + 1 (4)
xn) and one output axon y(x) and (w0,,wn)
are weights determining how much the inputs where t is the value that pushes the center of
should be weighted; g is an activation function the activation function away from zero and s is
that weights how powerful the output (if any) a steepness parameter. Sigmoid and hyperbolic
should be from the neuron, based on the sum tangent are both smooth differentiable
RI WKH LQSXW ,I WKH DUWLFLDO QHXURQ PLPLFV D functions, with very similar graphs. Note that
real neuron, the activation function g should the output range of the hyperbolic tangent
be a simple threshold function returning 0 or 1. goes from -1 to 1 and sigmoid has outputs
7KLVLVQRWWKHZD\DUWLFLDOQHXURQVDUHXVXDOO\ that range from 0 to 1. A graph of a sigmoid
implemented; it is better to have a smooth function is given in Figure 3 to illustrate
(preferably differentiable) activation function how the activation function looks like. The t
(Bishop, 1996). The output from the activation SDUDPHWHU LQ DQ DUWLFLDO QHXURQ FDQ EH VHHQ
function varies between 0 and 1, or between -1 as the amount of incoming pulses needed to
and 1, depending on which activation function activate a real neuron. A NN learns because
is used. The inputs and the weights are not this parameter and the weights are adjusted.
n
y( x ) = g wn +1 wi xi
i=0
(5)
ek = dk-yk (7) The 'wjk value is used to adjust the weight wjk by
wjk= wjk+'wjk and the BP algorithm moves on to
the next input and adjusts the weights according
where yk is the calculated output and dk is the to the output. This process goes on until a
desired output of neuron k. This error value certain stop criteria is reached. The stop criterion
is used to calculate a k value, which is again is typically determined by measuring the mean
used for adjusting the weights. The k value is square error of the training data while training
calculated by: with the data, when this mean square error
reaches a certain limit, the training is stopped.
k =ek g(yk ) (8)
In this section the mathematics of the BP
DOJRULWKP KDYH EHHQ EULH\ GLVFXVVHG EXW
where g is the derived activation function. since this report is mainly concerned with the
When the k value is calculated, the j values implementation of NN, the details necessary
can be calculated for preceding layers. The for implementing the algorithm has been left
j values of the previous layer are calculated out (for details see Hassoun, 1995 and Hertz
from the k values of this layer by the following et al., 1991).
equation:
InY = C1+C2 M+C3 M C4+C5 In R+C6 exp here a (t) is the acceleration time history, g is
(C7 M)+C8 R+ f (source) + f (site) the acceleration of gravity, and T represents the
complete duration of recording a (t). Figure 5
present the procedure followed to determine
VLnY= C9 (12) WKH VLJQLFDQW SDUDPHWHUV +XVLG 7KH
PRVW FRPPRQ PHDVXUH RI VLJQLFDQW GXUDWLRQ
where Y is the ground motion parameter of is a time interval between 5-95% of IA and is
interest, M the magnitude of the earthquake, denoted by Da5-95.
R a measure of the distance from the source to
the site being considered. C1-C9 are constants Predictive relationships have also been
to be determined. The lnY term describes developed for frequency-dependent duration
the uncertainty in the value of the ground parameters evaluated from bandpassed
motion parameter given by the predicative accelerograms (e.g., Bolt, 1973, Trifunac
relationship. and Westermo, 1982; Mohraz and Peng,
1989; and Novikova and Trifunac, 1994).
Regarding duration parameters many types These relationships have several limitations
of predictive relationships have been proposed WKDW DUH EDVLFDOO\ DVVRFLDWHG ZLWK D GHFLHQW
(Bommer and Martinez-Pereira, 1999), but representation of magnitude or site effects.
EUDFNHWHG GXUDWLRQ DQG VLJQLFDQW GXUDWLRQ Additionally, none of these have been
relationships are the most commonly used. derived from the energy integral. Some other
7KH IRUPHU LV GHQHG DV WKH WLPH HODSVHG restrictions are related to measured distan-
EHWZHHQWKHUVWDQGODVWH[FXUVLRQVEH\RQGD ce (normally the epicentral distance, not the
VSHFLHGWKUHVKROGDFFHOHUDWLRQ7KDWGHQLWLRQ FORVHVW VLWHVRXUFH GLVWDQFH DQG QDOO\ WKHUH
has shown to be sensitive to the threshold are still others having to do with the regression
acceleration considered and to small events method used to derive the relationships
WKDW RFFXU DW WKH QDO SDUW RI D UHFRUGLQJ (Kempton and Stewart, 2006).
6LJQLFDQWGXUDWLRQLVEDVHGRQWKHGLVVLSDWLRQ
of energy, within a time interval, and this In what follows we develop a predictive
energy is represented by the integral of the QHXURQDO PRGHO IRU VLJQLFDQW GXUDWLRQ WKDW
Figure 5. 6LJQLFDQW
duration parameters
(acceleration 5-95%).
Geo-Coordinates
Station Station name Soil
Lat N Long W
Code
The earthquakes in the database have A set of 26 events was used (Table 2) to
magnitudes ranging from 4.1 to 8.1. Most of design the topology of the NNs. These events
WKHHYHQWVRULJLQDWHGDORQJFRDVWRIWKH3DFLF were selected on the basis of the quality and
Ocean in the states of Michoacan, Guerrero resolution of the records. Accelerograms with
and Oaxaca. The epicenters of the remaining low signal to noise ratios were deleted from
three events, those of October 24, 1980, April the database. Both horizontal components and
3, 1997 and June15, 1999 were located in the vertical direction of each seismic event were
Puebla-Oaxaca border. Epicentral distances to considered.
stations in the city of Puebla range from 300
to 500 km and in only one case it reached 800 It is clear that the inputs and output spaces
km. That is why accelerations produced by the DUH QRW FRPSOHWHO\ GHQHG WKH SKHQRPHQD
earthquakes considered in this research did knowledge and monitoring process contain
not exceed 10 gal in Puebla. fuzzy stages and noisy sources. Many authors
have highlighted the danger of inferring a
In a second stage the database was process law using a model constructed from
expanded with accelerograms from the noisy data (Jones et al., 2007). It is imperative
Instituto de Ingeniera UNAM Accelerographic we draw a distinction between the subject of
Network (Alcntara et al., 2000). The added this investigation and that of discovering a
acceleration histories were recorded in stations process from records. The main characteristic
on rock located the coastal region of the of NN model is unrevealed functional forms.
states of Michoacan, Guerrero and Oaxaca, The NN data-driven system is a black-box
and forcefully had to be generated by one of representation that has been found exceedingly
the earthquakes we had already catalogued in useful in seismic issues but the natural principle
Table 1. The seismic stations we considered are that explains the underlying processes remains
VKRZQLQ)LJXUHOOHGVTXDUHVDVZHOODVWKH cryptic. Many efforts have been developed to
locations of the epicenters (inverted triangles). examine the input/output relationships in a
They were 88 three-component accelerograms numerical data-set in order to improve the NN
LQWKHQDOGDWDEDVH modeling capabilities, for example Gamma test
(Kemp et al., 2005; Jones et al., 2007; Evans
and Jones, 2002), but as far as the authors trying 25 different topologies. Thus, this can
experience, none of these attempts are be considered as the model having the best
applicable to the high dimension of the seismic generalization capabilities using the selected
phenomena or the extremely complex neural learning algorithm, architecture, and nodal
models for predicting seismic attributes. hidden structure. In Figure 9 the estimated
values obtained for a second set of unseen
Neural approximation patterns (validation set) are compared with
the numerical predictions obtained using the
7KH UVW VWHS LQ GHYHORSLQJ D 11 LV WKH relationship proposed by Reinoso and Ordaz
representation of the set of input and output (2001). The neuronal relationship follows
cells. There are no clear-cut procedures more narrowly the overall trend but fails in
WR GHQH WKLV FRQVWUXFWLRQ VWHS :KLOH WKH VRPHFDVHVFRHIFLHQWVRIFRUUHODWLRQDURXQG
optimum architecture --hidden nodes and R2=0.75). It should be stressed that the NN
associated weights-- is obtained when the has better interpolation and extrapolation
error function is minimized (i.e., the sum of the capabilities than the traditional functional
patterns of the squared differences between DSSURDFKHV )XUWKHUPRUH WKH LQXHQFH RI
the actual and desired outputs is minimum) directivity and fault mechanism on duration
the numerical or categorical representation FDQ EH LGHQWLHG ZLWK WKH 11 EDVHG RQ D
of inputs and outputs also depends on the multidimensional environment (Figure 10)
modelers experience and knowledge and a
trial-and-error procedure must be followed in NN based on information compiled from
order to achieve a suitable design. Oaxaca
Table 4. Continue.
as the architecture. Again DH1, DH2, and DV are as a system with particular stiffness and
included as outputs for neural mapping and damping characteristics that determine how
EHWZHHQWKHYHLQSXWVIRXUDUHQXPHULFDO0 will the soil column vibrate and for how long,
R, FD, and AZ) and one is a class node (soil as seismic waves traverse it and after their
type ST). The best model during the training passage through the deposit.
and testing stages has two hidden layers of
150 nodes each and was found through an Through the {M, R, FD, AZ, ST} o {DH1,
exhaustive trial and error process. DH2, DV} mapping, the neuronal approach we
The results of the RACO NN are summarized SUHVHQWHGRIIHUVWKHH[LELOLW\WRWDUELWUDULO\
in Figure 13. These graphs show the predicting complex trends in magnitude and distance
capabilities of the neural system comparing the dependence and to recognize and select
task-D values with those obtained during the DPRQJWKHWUDGHRIIVWKDWDUHSUHVHQWLQWWLQJ
NN training phase. It can be observed that the observed parameters within the range of
the durations estimated with the NN match variables present in data.
quite well calculated values throughout the
full distance and magnitude ranges for the Conclusions
seismogenic zones considered in this study.
Duration times from events separated to be $UWLFLDO QHXUDO QHWZRUNV ZHUH XVHG WR
used as testing patterns are presented and estimate strong ground motion duration.
compared with the neuronal blind evaluations These networks were developed using a back
in Figure 14. The results are very consistent propagation algorithm and multi-layer feed-
and remarkably better than those obtained forward architecture in the training stage.
when analyzing RACP database. The linguistic In developing the networks it was assumed
expression of soil type is obviously a superior that the parameters that have the greatest
representation of the soil effect on D prediction. LQXHQFH RQ VWURQJ PRWLRQ GXUDWLRQ DUH
magnitude, epicentral distance, focal depth,
A sensitivity study for the input variables soil characterization and azimuth. These
was conducted for the three neuronal modules. parameters include the effects of seismic
The results are given in Figure 15 and are valid source, distance, materials and directivity.
only for the data base utilized. Nevertheless, The many topologies tested and the input
after conducting several sensitivity analyses sensitivity developed drive to the conclusion
changing the database composition, it was WKDW D EURDG VRLOW\SH FODVVLFDWLRQ LQ WKHVH
found that the RACO trend prevails: ST (soil investigation three soil types) provides a better
type) is the most relevant parameter (has correlation with seismic phenomena than the
the larger relevance), followed by azimuth more commonly used natural period Ts.
Az, whereas M, FD and R turned out to be less
LQXHQWLDO11VIRUWKHKRUL]RQWDODQGYHUWLFDO Overall, the results presented here show
components are complex topologies that WKDWDUWLFLDOQHXUDOQHWZRUNVSURYLGHJRRGDQG
assign nearly the same weights to the three reasonable estimates of strong ground motion
input variables that describe the event, but an duration in each one of the three orthogonal
important conclusion is that the material type components of the accelerograms recorded in
in the deposit and the seismogenic zone are the cities of Puebla and Oaxaca using easy-to-
YHU\UHOHYDQWWRGHQH'7KLVQGLQJFDQEH obtain input parameters: S M, R, FD and Az.
explained if we conceptualize the soil deposit
Figure 13. NN results for RACO, training stage. Figure 14. NN results for RACO, testing stage.
Luna M.,(CICESE), Espinosa J.M., Cuellar Hertz J., Krogh A., Palmer R.G., 1991, Introduction
A., Camarillo L., Ramos S., Snchez M., to the Theory of Neural Computing. Addison-
(CIRES), Quaas R., Guevara E., Flores J.A., Wesley, California, 327 pp.
Lpez B., Ruiz R., (CENAPRED), Guevara E.,
Pacheco J., (IG-UNAM), Ramrez M., Aguilar Jones A.J., Evans D., Kemp S.E., 2007, A note
J., Jurez J., Vera R., Gama A., Cruz R., on the Gamma test analysis of noisy input/
Hurtado F., M. del Campo R., Vera F., (RIIS), output data and noisy time series. Physica
Alcntara L. (SMIS), 2000, Mexican Strong D, 229:1, 1-8.
Motion Data Base, CD Rom Vol 2. Mexican
Society for Earthquake Engineering. Tettamanzi A., Tomassini M., 2001, Soft
Computing: integrating evolutionary,
Anderson J.G., 2004, Quantitative measure of neural, and fuzzy systems. Springer-Verlag,
JRRGQHVVRIW RI V\QWKHWLF VHLVPRJUDPV Berlin, 328 pp.
in 13th World Conference on Earthquake
Engineering, 243, Vancouver, Canada, 1-6 Husid L.R., 1969, Caractersticas de terremotos -
August. anlisis general. Revista del IDIEM, 8, 2142.
ULDV $ $ PHDVXUH RI HDUWKTXDNH Kemp S.E., Wilson I.D., Ware J.A., 2005, A
intensity, Seismic Design of Nuclear Power tutorial on the Gamma test. International
Plants. R. Hansen, Editor, MIT Press, Journal of Simulation: Systems, Science
Cambridge, Mass., 438-483. and Technology, 6:1-2, 67-75.
Bishop C.M., 1996, Neural networks for pattern Kempton J., Stewart J., 2006, Prediction
recognition. Oxford University Press, New HTXDWLRQV IRU VLJQLFDQW GXUDWLRQ RI
York, 482 pp. earthquake ground motions considering
site and near-source effects. Earthquake
Bolt B.A., 1973, Duration of strong ground Spectra, 22:4, 9851013.
motion, in 5th World Conference on
Earthquake Engineering, 292, Rome, Italy, Mohraz B., Peng M.H., 1989, The use of low-
25-29 June. SDVVOWHULQJLQGHWHUPLQLQJWKHGXUDWLRQRI
strong ground motion. Publication PVP-182,
Bommer J.J., Martinez-Pereira A., 1996, Pressure Vessels and Piping Division, ASME,
The prediction of strong-motion duration 197200.
for engineering design, in 11th World
Conference on Earthquake Engineering, 84, Novikova E.I., Trifunac, M.D., 1994, Duration of
Acapulco, Mexico, 23-28 June. strong ground motion in terms of earthquake
magnitude, epicentral distance, site
Bommer J.J., Martinez-Pereira A., 1999, The conditions and site geometry. Earthquake
effective duration of earthquake strong Eng. Struct. Dyn., 23, 10231043.
motion. Journal of Earthquake Engineering,
3:2, 127172. Reinoso E., Ordaz M., 2001, Duration of strong
ground motion during Mexican earthquakes
Broomhead D.S., Lowe D., 1988, Multivariable in terms of magnitude, distance to the
functional interpolation and adaptive rupture area and dominant site period.
networks, Complex Systems, 2, 321-355. Earthquake Eng. Struct. Dyn, 30, 653-673.
Evans D., Jones A.J., 2002, A proof of the 6DUPD 6. (QHUJ\ X[ RI VWURQJ
Gamma test. Proc. R. Soc. Lond., 458, earthquakes. Tectonophysics, 11, 159173.
2759-2799.
Trifunac M.D., Westermo B.D., 1982, Duration
+DVVRXQ0+)XQGDPHQWDOVRI$UWLFLDO of strong earthquake shaking. Soil Dynamics
Neural Networks. MIT Press, Massachusetts, and Earthquake Engineering, 1:3, 117-121.
511 pp.