Dynamic Fuzzy Wavelet Neural Network Model for Structural

System Identification
Hojjat Adeli, F.ASCE1; and Xiaomo Jiang2

Abstract: A new dynamic time-delay fuzzy wavelet neural network model is presented for nonparametric identification of structures
using the nonlinear autoregressive moving average with exogenous inputs approach. The model is based on the integration of four
different computing concepts: dynamic time delay neural network, wavelet, fuzzy logic, and the reconstructed state space concept from
the chaos theory. Noise in the signals is removed using the discrete wavelet packet transform method. In order to preserve the dynamics
of time series, the reconstructed state space concept from the chaos theory is employed to construct the input vector. In addition to
denoising, wavelets are employed in combination with two soft computing techniques, neural networks and fuzzy logic, to create a new
pattern recognition model to capture the characteristics of the time series sensor data accurately and efficiently. The model balances the
global and local influences of the training data and incorporates the imprecision existing in the sensor data effectively. Experimental
results on a five-story steel frame are employed to validate the computational model and demonstrate its accuracy and efficiency.
DOI: 10.1061/共ASCE兲0733-9445共2006兲132:1共102兲
CE Database subject headings: Fuzzy sets; Neural networks; Identification; Mathematical models; Structural analysis.

Introduction of the art up to 1995 is presented by Ghanem and Shinozuka

The goal of structural system identification research is to develop
a mathematical model for a structural system based on a set of
inputs and corresponding output measurements. When structures
are damaged during a strong ground motion changes occur to
their dynamic characteristics. Structural system identification
finds applications in: 共1兲 determination of the structural properties
such as the stiffness and natural periods and frequencies; 共2兲 non-
destructive damage evaluations, where input–output measure-
ments are used to assess nondestructively the damage severity and
location in an existing structure; 共3兲 health monitoring of the glo-
bal or local conditions of structures; and 共4兲 structural control
of smart structures which requires evaluation of dynamic re-
sponse of structures with various structural rigidities, masses, and
damping.
The complicated mechanics of earthquake and the significant
expense associated with detailed measurements of structural re-
sponse make accurate system identification of structures a chal-
lenging problem. This problem has attracted the attention of a
good number of researchers in recent years. A review of the state
1
Lichtenstein Professor, Dept. of Civil and Environmental
Engineering and Geodetic Science, The Ohio State Univ., 470 Hitchcock
Hall, 2070 Neil Ave., Columbus, OH 43210 共corresponding author兲.
E-mail: adeli.1@osu.edu
2
Graduate Research Associate and PhD Candidate, Dept. of Civil and
Environmental Engineering and Geodetic Science, The Ohio State Univ.,
470 Hitchcock Hall, 2070 Neil Ave., Columbus, OH 43210. E-mail:
jiang.98@osu.edu
Note. Associate Editor: Elisa D. Sotelino. Discussion open until June
1, 2006. Separate discussions must be submitted for individual papers. To
extend the closing date by one month, a written request must be filed with
the ASCE Managing Editor. The manuscript for this paper was submitted
for review and possible publication on April 26, 2004; approved on April
20, 2005. This paper is part of the Journal of Structural Engineering,
Vol. 132, No. 1, January 1, 2006. ©ASCE, ISSN 0733-9445/2006/1-102–
111/$25.00.
共1995兲. Examples of recent work are Chassiakos and Masri
共1996兲, Gurley and Kareem 共1999兲, Loh et al. 共2000兲, Masri et al.
共2000兲, Hung et al. 共2003兲, and Kijewski and Kareem 共2003兲.
There are two fundamentally different approaches for the
solution of the system identification problem: parametric method
and nonparametric method. The parametric method is a straight-
forward approach in which the parameters of an actual system
model are used directly to represent physical quantities such as
the structural stiffness and damping ratio. Estimation algorithms
such as the Kalman filter method, maximum likelihood method,
and recursive least square method are used to determine the
parameters of the system model. Parametric methods have been
applied successfully to identify the dynamic properties of linear-
ized and time-invariant equivalent structural systems 共Juang
1994; Ljung and Glad 1994; Ghanem and Shinozuka 1995兲.
Recently, wavelet-based approaches have been developed for
parametric identification of simple multidegree-of-freedom
共MDOF兲 systems 共Staszewski 1997 and 1998; Ghanem and
Romeo 2000; Hans et al. 2000; Lamarque et al. 2000; Chang et al.
2003; Kijewski and Kareem 2003兲. Staszewski 共1997 and 1998兲
uses a continuous Grossman–Morlet wavelet transform to decom-
pose the impulse response of MDOF systems into the time-scale
domain. Then, the structural parameters are estimated based on
the ridges 共where the energy of the signal is mainly concentrated
in the time-scale plane兲 and skeletons of the wavelet transform
共the values of the wavelet transform at its ridges兲. This method is
used for parametric estimation of the nonlinear and time variant
simple MDOF systems. Ghanem and Romeo 共2000兲 describe a
wavelet-based approach for parametric identification of single-
and simple MDOF linear time-varying dynamic systems. They
use the wavelet transform to decompose the governing differential
equations into discrete forms. Then, the structural parameters are
obtained from the solution of the transformed differential
equations using the least square minimization technique. The
authors report accurate estimation of the damping and stiffness
parameters of a simple linear time-invariant system. They also

102 / JOURNAL OF STRUCTURAL ENGINEERING © ASCE / JANUARY 2006

Downloaded 17 Aug 2010 to 130.54.110.32. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
demonstrate that the presence of noise affects the accuracy of the
estimated damping significantly but has little effect on stiffness.
Kijewski and Kareem 共2003兲 use Morlet wavelet decomposition
for identifying the frequency and damping parameters of simple
MDOF systems.
These parametric identification algorithms, however, depend
strongly on the accuracy of the measured data. They cannot pro-
vide the required accuracy and reliability needed for complex
system identifications of real life structures due to complicated
nonlinear nature of behavior of civil structures, and incomplete,
incoherent, and noise-contaminated measurements of structural
response to strong ground motions.
Nonparametric system identification approaches have been
widely used in health monitoring 共e.g., Nakamura et al. 1998;
Masri et al. 1996, 2000兲 and structural control of building struc-
tures 共e.g., Bani-Hani and Ghaboussi 1998; Brown and Yang
2001兲. They are particularly effective for large-scale structures
due to their complicated nonlinear behavior and the incomplete,
incoherent, and noise-contaminated measurements of structural
response under extreme loadings. In the nonparametric approach,
the input–output map is characterized and determined by a system
model that may not have any explicit physical meaning. In gen-
eral, the system model does not represent any physical quantity
directly, but it is trained to approximate a physical structure and
predict the structural responses. As such, the approach does not
require complete and coherent measurements of the structural re-
sponse to strong ground motions. It has better adaptability than
the parametric methods 共Billings and Tsang 1989; Billings et al.
1990; Sjöberg et al. 1995; Thomson et al. 1996; Aguirre et al.
2000兲.
In nonparametric system identifications the nonlinear autore-
gressive moving average with exogenous inputs 共NARMAX兲
approach is widely used for mapping the input–output relation-
ship 共Ljung and Glad 1994; Ljung 1999兲 due to the following
advantages: 共1兲 it does not impose any restriction on the nature of
system input excitation leading to its general applicability; 共2兲 it
uses constant parameters independent of the time variable, which
makes it relatively easy to incorporate a priori knowledge about
the system into the model; and 共3兲 it can model nonlinearity ef-
fectively in contrast to linear approaches such as the autoregres-
sive moving average method 共ARMA兲.
The key issue in nonparametric identification methods is an
effective approach for estimation of the coefficients of the
NARMAX. Sjöberg et al. 共1995兲 describe a number of soft com-
puting approaches for finding the NARMAX coefficients such as
the multilayer neural networks and fuzzy logic. Loh et al. 共2000兲
use the Levenberg–Marquardt back-propagation 共BP兲 neural net-
work for nonparametric identification of a five story test frame.
Masri et al. 共1993, 1996, 2000兲 investigate use of the BP neural
network method for structural system identification and health
monitoring. Their study shows that the neural network method is
a practical tool in detecting changes in nonlinear structures with
unknown constitutive properties and topologies. However, those
neural network methods suffer from some common drawbacks
such as lack of an efficient constructive model 共for example, re-
quiring arbitrary selection of the number of hidden nodes兲, slow
convergence rate, and entrapment in a local
minimum.
The writers found only one paper on the use of wavelets for
nonparametric system identification of structures. Hung et al.
共2003兲 used the wavelet neural network 共WNN兲 model, which
was proposed by Zhang and Benveniste 共1992兲 for signal process-
ing, for system identification of a five-story test frame subjected
JOURNAL OF STRUCTURAL ENGINEERING © ASCE / JANUARY 2006 / 103
Downloaded 17 Aug 2010 to 130.54.110.32. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
to simulated earthquake loadings on a shaking table. However,
the WNN method for structural system identification also suffers
from: 共1兲 lack of an efficient constructive model; 共2兲 the need to
find the model parameters such as the input vector dimension by
trial and error; 共3兲 slow convergence rate when there exists noise
in the measured data; and 共4兲 low identification accuracy when
there exists imprecision in the measured data.
In this paper, a new multiparadigm dynamic time-delay fuzzy
WNN model is presented for nonparametric identification of
structures using the NARMAX approach. The model is based on
the integration of four different computing concepts: dynamic
time delay neural network, wavelet, fuzzy logic, and the recon-
structed state space concept from the chaos theory. The goal is to
improve the accuracy and adaptability of nonparametric system
identification of structures under earthquake loadings. The results
of this research will help structural engineers better understand
the behavior of structures during earthquakes and design more
effective earthquake-resistant structures.
Constructing Dynamic Time-Delay Fuzzy-Wavelet
Neural Network Model
Creating State Space Vectors from Denoised Data
The NARMAX approach has proven to be a powerful tool for
mapping the nonlinear input–output relationship in system iden-
tification 共Juang 1994; Ljung and Glad 1994; Thomson et al.
1996; Aguirre et al. 2000兲. In this approach, the general discrete
dynamic input–output mapping is expressed as follows:
y共t兲 = f共xt−1,yt−1,et−1兲 + e共t兲 共1兲
where xt, yt, and et represent the input, output, and zero-mean
noise vectors at time t 共representing the discrete time series values
up to time t兲, respectively; f共.兲 = scalar nonlinear mapping or ap-
proximation function; and e共t兲 = error between the actual and es-
timated values of the future output y共t兲. The objective is to find
the relationship between the past observations 关xt−1 , yt−1兴 and the
future output y共t兲. We need to model the mapping function for
effective structural system identification.
In this work, first the noise in the signals is removed using the
discrete wavelet packet transform 共DWPT兲 method. The DWPT
provides more coefficients than the conventional discrete wavelet
transform 共DWT兲 representing additional subtle details of a sig-
nal. The DWPT-based denoising is presented in Jiang and Adeli
共2004兲. Next, the dynamic system is approximated as a functional
representation of the lagged past denoised inputs and outputs. The
reconstructed state space concept from the chaos theory is used to
simulate the multivariate properties of the nonlinear system effec-
tively. The input dimension or the size of the state space vector in
the NARMAX approach, D, is the summation of dimensions of
the structural inputs, Dx 共number of discrete time steps used for
any input time series兲, and the feedback outputs, Dy 共number of
discrete time steps used for the corresponding past output time
series兲.
A proper choice of the dimension for the NARMAX approach
plays an important role in reconstructing an appropriate multivari-
ate input. If D is too small, the model produces inaccurate
identification results. If D is too large, it can also lead to overes-
timation and inaccurate results in addition to increasing the com-
putational cost. System identification researchers usually use
a trial-and-error approach to find the most suitable value
for the input dimension in the NARMAX approach 共Zhang 1997;
Hung et al. 2003兲. However, the trial-and-error approach: 共1兲 does
not provide a rational basis for the selection; 共2兲 is cumbersome;
共3兲 is computationally time consuming; and 共4兲 does not guaran-
tee accurate system identification results.
An input time series is denoted by discrete quantities 兵xi其,
i = 1 , 2 , . . . , N, where N is the number of data points. The input
vector at time t, x共t兲, in the NARMAX approach 关Eq. 共1兲兴 is rep-
resented by the state space vector, xk, as follows:
xk = 关xk,xk+␶,xk+2␶, . . . ,xk+共Dx−1兲␶兴T k = 1,2, . . . . ,Na
where T denotes the transpose of a vector and the parameters Dx
and ␶ = dimension 共usually referred to as embedding dimension,
dE, in the chaos theory兲 共Jiang and Adeli 2003兲 and the lag time
index used to reconstruct the state space, respectively. The value
of ␶ is set equal to 1 in this study. The data point subscript, k,
satisfies the following condition:
k 艋 N − Dx
k苸Z
where Z = set of all integers. Based on Eq. 共3兲, the maximum
number of state space points, Na, is equal to
Na = N − Dx
Eq. 共2兲 represents the state space vector with one lag time
reconstructed from the sensor data. The feedback input vector y in
Eq. 共1兲 is defined similar to Eq. 共2兲 with Dy dimension. In the
theory of chaos, the set of points used to simulate the evolution
trajectory in the original state space is referred to as an attractor.
The purpose of the attractor is to unfold the time series back to a
multivariate state space representing the original physical system.
An attractor is the geometric invariance in the state space repre-
sentation of a time series. The invariants of the dynamic system
producing the time series are preserved if the time series is trans-
formed into a sufficiently large reconstructed state space 共defined
in terms of its dimension兲. Chaotic attractors often indicate chaos
in the physical system represented by time series. A proper choice
of the parameters Dx and Dy plays an important role in accurately
reconstructing the multivariate state space 共Jiang and Adeli 2003兲.
For the structural system identification problem, the false near-
est neighbor 共FNN兲 method was found to be more accurate than
the fill factor and average integral local deformation methods and
therefore is used in this research to find the optimum embedding
dimensions Dx and Dy. The FNN approach is based on the as-
sumption that a small embedding dimension results in state space
points that are far apart in the original state space to be considered
neighboring points in the reconstructed state space. Details of this
method can be found in Jiang and Adeli 共2003兲 and therefore will
not be described here.
Creating Fuzzy Wavelet Neural Network Model
The dynamic system function represented symbolically by f共.兲 in
Eq. 共1兲 is approximated using the wavelet transform functions and
wavelet coefficients expressed in a general form as follows
共Daubechies 1988 and 1992兲:
104 / JOURNAL OF STRUCTURAL ENGINEERING © ASCE / JANUARY 2006
Downloaded 17 Aug 2010 to 130.54.110.32. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
冉 冊
M D
Xkj − bij
f̄共Xk兲 = 兺 兺wi ␸
aij
k = 1, . . . ,Na
i=1 j=1
a,b 苸 Re 共5兲
␸共.兲 苸 L2共Re兲
where D = Dx + Dy = dimension of the input vector
Xk; Xkj represents the jth item in the kth input vector
Xk =
xk , xk+␶ , xk+2␶ , . . . , xk+共Dx−1兲␶xk , y k+␶ , y k+2␶ , . . . , y k+共Dy−1兲␶T; and
共2兲 M = number of wavelets obtained using the modified Gram–
Schmidt algorithm 共Zhang 1997; Jiang and Adeli 2005兲 and the
Akaike’s final prediction error method 共AFPE兲 共Ljung 1999;
Jiang and Adeli 2005兲. In the modified Gram–Schmidt algorithm,
first the wavelet which best approximates the measured data is
selected from the complete nonzero wavelet coefficient sets.
Next, this wavelet is combined with the remainder of wavelet sets
one at a time and the combination which best approximates the
measured data is determined. The procedure is repeated for all
nonzero wavelet coefficients. The symbol f̄共Xk兲 represents the
approximation of f共Xk兲, wi represents the discrete wavelet trans-
form coefficient, and ␸共.兲 = two-dimensional wavelet expansion
共3兲 functions with scaling and translation 共the multidimensional input
Xk results in a multidimensional wavelet sets兲. The Mexican hat
wavelet is used due to its analytical expression convenient for
decomposing multidimensional time series, its differentiability,
and its noncompactly supported but rapidly vanishing feature 共the
function is nonzero on an infinite interval but approaches zero
共4兲
rapidly兲 共Zhou and Adeli 2003兲 which is advantageous in this
work. The parameters aij ⫽ 0 and bij denote the frequency 共or
scale兲 and the time 共or space兲 location parameters corresponding
to the multidimensional input vector Xk, respectively, and Re is
the set of real numbers. The notation L2共Re兲 represents the square
summable constructed state space vectors.
Characteristics localization of time series in spatial 共or time兲
and frequency 共or scale兲 domains can be accomplished efficiently
through wavelet decomposition. The power of wavelets for time
series analysis stems from three features. First, wavelet analysis
can determine the sharp transitions simultaneously in both fre-
quency and time domains. Thus, wavelets can help identify non-
linear behavior displayed in any signal. Second, wavelet analysis
allows for an effective representation of discontinuities in the cha-
otic time series. The wavelet representation of information in the
time series allows for its hierarchical decomposition. In this way,
the information can be analyzed in components of desired char-
acteristics and at various levels of details. Third, when the infor-
mation in time series is transformed into the wavelet domain less
storage is required for its effective representation, resulting in
computational efficiency for large time series.
Eq. 共5兲 is used to construct a wavelet neural network for ef-
fective identification of general nonlinear systems. Unlike the
conventional multilayer neural network whose activation func-
tion, such as a sigmoid function, has infinite energy 共the function
is nonzero on an infinite interval and area under the function is
infinity兲, the proposed WNN uses a wavelet function with a
spatial-spectral zooming property, which influences the output of
the model only in the finite range of input time series. This prop-
erty: 共1兲 reduces the undesirable interaction effects among the
nodes of the neural network thus in general improving the accu-
racy of the system identification and 共2兲 accelerates the neural
network training process thus improving its computational effi-
ciency. The flip side of the WNN model is that local imprecision
in the training data may result in a large local output error because
of the same spatial-spectral zooming property.
The fuzzy logic theory is a powerful tool for effective repre-
sentation of imprecision existing in the time series data measured
by sensors 共Zadeh 1978兲. To overcome the aforementioned short-
coming of the WNN model, a fuzzy-WNN model is created based
on the following two steps. First, the fuzzy C-means 共FCM兲 clus-
tering algorithm presented in Jiang and Adeli 共2003兲 is used as a
data mining tool to divide the sensor data into clusters with com-
mon features. Next, a fuzzy-wavelet model is created by combin-
ing the fuzzy clusters of the reconstructed state space vectors with
the nonorthogonal wavelets defined by Eq. 共5兲.
The FCM algorithm performs a fuzzy partitioning of the data
set into classes, where the concept of fuzzy membership is used to
represent the degree to which a given data set belongs to some
cluster. The state space vectors reconstructed earlier are clustered
based on their similarities. The fuzzy clustering problem is de-
fined as a constrained optimization problem as follows: Fig. 1. Two clusters obtained using fuzzy C-means clustering
Minimize algorithm
Na M

f ␤共c兲 = 兺 兺
k=1 i=1
A␤ki储Xk − ci储2 共6兲 IF Amki:Xk

Subject to the constraints Na

M 兺 AmkiXk

兺
Aki = 1
i=1
共7兲
1 艋 k 艋 Na
Aki 艌 0
共8兲
1 艋 k 艋 N a, 1 艋 i 艋 M
where f ␤共c兲 = objective function; Xk = kth input state space vector;
Aki = membership function or the degree of membership of the kth
variable in the ith fuzzy implication rule; ci = 关c1i , c2i , . . . , cDi兴T
represents the D-dimensional center of the ith cluster; M = total
number of fuzzy implication rules; and Na = number of the train-
ing data sets defined by Eq. 共4兲. The parameter ␤ represents the
degree of fuzziness in the data. A value in the range 1 ⬍ ␤ 艋 2 is
often chosen 共larger values are selected for fuzzier data situa-
tions兲. A value of ␤ = 1.5 is chosen in this study. The parameter M
is also equal to the number of the clusters as well as the number
of the wavelets used in the WNN model.
The variables of the optimization problem represented by
Eqs. 共6兲–共8兲 are the membership grades, Aki. The total number of
the variables is equal to M times Na. The membership grade is
represented in terms of the Euclidean distance of the input vector
from the class center ci as follows 共Bezdek 1981; Jiang and Adeli
2003兲:
k=1
THEN ci = Na i = 1, . . . ,M
兺
Amki
k=1
k = 1, . . . ,Na 共10兲
where Amki:Xk
indicates the degree of membership of Xk in the ith
fuzzy implication rule at iteration m is Amki. Fig. 1 shows a con-
ceptual example of two clusters obtained from the data points in a
two-dimensional state space using the FCM clustering algorithm.
The partitioning of the data is mathematically expressed in
the fuzzy partition matrix A = 关Aki兴 whose elements are the
membership degrees of the data vectors Xk = 关x1 , x2兴 in the fuzzy
clusters ci.
The new fuzzy WNN model is formed using the fuzzy wave-
lets in the following three steps: 共1兲 construct the WNN model;
共2兲 find fuzzy clusters using the FCM algorithm; and 共3兲 replace
the conventional translation parameters, bij in Eq. 共5兲, by the
fuzzy clusters cij to form the fuzzy wavelet. Fig. 2 shows the
feedforward architecture of the fuzzy WNN model. It consists of
an input layer, a hidden fuzzy-wavelet layer, and an output layer.
There are D = Dx + Dy input nodes in the input layer. The optimum
number of fuzzy-wavelet nodes is found using the AFPE criterion
共Ljung 1999兲. This number is set equal to the number of fuzzy
clusters required in the FCM algorithm. The ellipses in the hidden
layer represent the fuzzy wavelet functions. The output in the

冋兺 冉 冊 册
M −1 output layer represents the approximated structural response
储Xk − cti储2 1/␤−1
关Eq. 共5兲兴.
Aki = 1 艋 k 艋 Na
j=1 储Xk − ctj储2
共9兲 Creating Dynamic Time-Delay Fuzzy Wavelet Neural
1艋i艋M Network Model
The iterative FCM algorithm presented in Jiang and Adeli Both earthquake records and structural responses are a sequence
共2003兲 is used to find the clusters and their centers. A fuzzy logic of ordered measurements in the form of time series produced by
inference mechanism is developed in the FCM algorithm using a single sensors. In order to extract useful information from the
set of IF–THEN fuzzy implication rules in the following form time series for accurate system identification of a structure using
共Sugeno and Kang 1988兲: the fuzzy WNN model, the input data from a single time series

JOURNAL OF STRUCTURAL ENGINEERING © ASCE / JANUARY 2006 / 105

Downloaded 17 Aug 2010 to 130.54.110.32. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
 Fig. 2. Architecture of fuzzy-wavelet neural network
cannot be treated as independent data points. The data order in the
time series must be preserved in temporal processing to avoid
distorting the time signal and changing its frequency content.
Consequently, a dynamic time-delay fuzzy WNN model with
a recurrent feedback topology is created for structural system
identification.
In order to capture the linear characteristics of the nonlinear
system identification problems, a linear term is added to Eq. 共5兲 to
form the fuzzy WNN model expressed as
冉 冊
M D D
Xkj − cij
f̄共Xk兲 = 兺
i=1
兺
wi
j=1
␸
aij
+ 兺j b jXkj + d
where b j = weight of the link of the jth input to the output; and
d = bias term 共Fig. 3兲. The double summation in Eq. 共11兲 repre-
sents a linear combination of the weighted sum of the wavelets.
Eq. 共11兲 approximates the structural dynamics using a few wave-
lets in combination with the linear characteristics.
Fig. 3 shows the architecture of the multiparadigm dynamic
time-delay fuzzy WNN model for structural system identification.
It consists of six steps: 共1兲 de-noise the original time series sensor
data using the DWPT-based approach; 共2兲 construct the state
space vectors using the FNN approach; 共3兲 select wavelets using
the modified Gram–Schmidt algorithm; 共4兲 minimize the number
of the required wavelets using the Akaike’s final prediction error
criterion; 共5兲 find fuzzy clusters using the FCM algorithm; and 共6兲
construct the dynamic time-delay fuzzy-WNN model.
The topology of the dynamic fuzzy WNN model consists of an
input layer, a hidden 共fuzzy wavelet兲 layer, and an output layer.
The input layer has D nodes representing the constructed state
space vector of structural inputs x and feedback outputs y. In
addition to Dx structural inputs, Dy outputs, y n−1 , y n−2 , . . . , y n−Dy,
are fed back to the input layer at a time. The fuzzy wavelet layer
106 / JOURNAL OF STRUCTURAL ENGINEERING © ASCE / JANUARY 2006
Downloaded 17 Aug 2010 to 130.54.110.32. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
Fig. 3. Architecture of dynamic time-delay fuzzy wavelet neural
network model
consists of M fuzzy-wavelet nodes. The output layer has only one
node which is the predicted structural response 关Eq. 共11兲兴.
Unlike the conventional static neural networks that are based
on single-valued input nodes, the proposed model is a dynamic
neural network that preserves the time sequence of the input vec-
tors and memorizes the past of the time series sensor data. Each
共11兲 data set in the input layer is a state space vector, a delay input
with lag time of 1. The delay inputs are aggregated nonlinearly
through the fuzzy wavelet processing nodes in the fuzzy wavelet
layer to memorize the past.
Training Dynamic Time-Delay Fuzzy Wavelet Neural
Network Model
A hybrid learning algorithm, Levenberg–Marquardt 共LM兲-least
squares 共LS兲 algorithm, is developed for estimating the param-
eters of the fuzzy WNN model: parameters aij in the wavelet
basis, weights wi and b j, and the bias d 关Eq. 共11兲兴. The LM algo-
rithm, an approximate combination of the Gauss–Newton and
steepest descent algorithms, is used to estimate the parameters aij
of nonlinear wavelet functions. The approximation avoids the
second-order differentiation required in the Gauss–Newton algo-
rithm and overcomes the numerical instabilities encountered in
the steepest descent algorithm. The much simpler LS algorithm is
used to determine the parameters of the linear part of the fuzzy
WNN model. Next, a backtracking inexact linear search algo-
rithm is developed to automatically update the iteration step
 Fig. 5. Kobe1 earthquake ground motion record and acceleration
responses of five-story test frame

in Fig. 4. The new model is used to identify the response of the

second floor and compare the results with those of Hung et al.
Fig. 4. Model of five-story test steel frame for validating model
共2003兲.
Fig. 6共a兲 shows the measured data points from 7 s to 8 s of the
structural response at the first floor under the five earthquake
excitations mentioned earlier. It is found that at each floor all
length with the goal of accelerating the learning convergence rate structural acceleration response curves have the same general
of the model. shape at various earthquake excitations without outliers. The re-
sponse patterns of the structure subjected to different levels of
excitations are not changed significantly, but the peak ground val-
Validation of Fuzzy-Wavelet Neural Network Model ues of acceleration increase as the excitation level increases, as
expected.
Example Structure
Experimental results on a 1 / 2-scaled five-story steel frame re-
ported in recent literature were used to validate the new dynamic
time-delay fuzzy WNN model for system identification of struc-
tures 共Loh et al. 2000; Hung et al. 2003兲. The structure is a 3 m
long, 2 m wide, and 6.5 m high steel frame 共Fig. 4兲. Lead blocks
were nearly uniformly distributed on each floor such that the mass
of each floor was approximately equal to, 3,664 kg. This test
structure was subjected to the basic excitation of the original
Kobe earthquake with five different scales 共20, 32, 40, 52, and
60%兲 on a shaking table at the National Center for Research on
Earthquake Engineering in Taiwan 共Hung et al. 2003兲. These ex-
citations are denoted as Kobe1, Kobe2, Kobe3, Kobe4, and
Kobe5, respectively. Acceleration responses were measured at the
four corners of each floor in two horizontal directions over a
period of 25 s at increments of 0.001 s. Thus, around 25,000 out-
put data points are recorded in a single test. Fig. 5 shows the
Kobe1 earthquake ground motion record and the corresponding
acceleration responses of the five floors of the test frame. It
should be noted that in Fig. 5 the vertical scales for floor accel-
erations are different from that for the ground acceleration.
In order to improve computational efficiency, all acceleration
responses were normalized to the gravity acceleration 共g兲. For the
sake of comparison with the results of WNN presented by Hung
et al. 共2003兲, only the response data at the first, second, and third
floors are used for training and testing data sets in this study, Fig. 6. Data points from 7 s – 8 s of structural response at first floor
which is the same scheme used by Hung et al. 共2003兲, as depicted under five earthquake excitations: 共a兲 measured; and 共b兲 denoised

JOURNAL OF STRUCTURAL ENGINEERING © ASCE / JANUARY 2006 / 107

Downloaded 17 Aug 2010 to 130.54.110.32. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
Fig. 7. Root mean square error bar diagram of system identification
results
Data Denoising
The measured data contain noise. The DWPT-based denoising
approach and Daubechies wavelets of order 4 are employed for
denoising all measured data. As an example, Fig. 6共b兲 shows the
denoised results of the measured data points shown in Fig. 6共a兲. It
is observed that the DWPT-based de-noising approach smoothes
the contaminated data effectively while at the same time preserv-
ing the subtle features in the signal.
Constructing Dynamic Time-Delay Fuzzy Wavelet
Neural Network Model
The 14,000 denoised data points from 4 s to 18 s of the structural
response at the first, second, and third floors are extracted from
the 25,000 data points for constructing the fuzzy WNN model.
For reconstructing the state space vectors, the same optimum em-
bedding dimension of 6 is obtained using the FNN method for the
three floor acceleration responses of the structure under different
earthquake excitation levels. This is not unexpected because the
general shape of the structural response under various excitation
levels does not change significantly. The same dimension is there-
fore used for both measured inputs and feedback inputs of the
fuzzy WNN model 共Dx = 6 and Dy = 6兲. The number of input vec-
tors created is thus equal to Na = 14,000− 6 = 13,996 关Eq. 共4兲兴.
Next, a four-level wavelet decomposition is performed on the
13,994 training vectors using the Mexican hat wavelet. The empty
wavelets whose supports do not contain any data are eliminated,
resulting in 157 nonempty wavelets. Based on the AFPE criterion
and using the modified Gram–Schmidt algorithm, it was con-
cluded that 2 out of the 157 nonempty wavelets are sufficient to
construct an effective fuzzy WNN model. As such, there are 2
wavelet nodes in the hidden layer of the network 共Fig. 3兲.
Training Dynamic Fuzzy Wavelet Neural Network
Model
The acceleration responses of the first, second, and third floors
during the previous Dx = 6 time intervals are used as inputs, and
the current acceleration response of the second floor is used as the
output of the dynamic time-delay fuzzy WNN model 共Fig. 4兲. The
five sets of experimental data obtained at different excitation lev-
els 共Kobe1–Kobe5兲 are used to train the model, one set at a time,
resulting in five trained models. The dynamic time-delay fuzzy
WNN model converges very fast after only two training iterations
108 / JOURNAL OF STRUCTURAL ENGINEERING © ASCE / JANUARY 2006
Downloaded 17 Aug 2010 to 130.54.110.32. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
Fig. 8. Identified structural responses of second floor subjected to
Kobe1 earthquake using model trained by structural responses from
Kobe2 earthquake: 共a兲 comparison of acceleration responses, 共b兲
absolute error, and 共c兲 comparison of acceleration responses from
7 to 8 s
using the hybrid adaptive LM-LS learning algorithm. All absolute
errors between the computed and measured responses are less
than 0.002g 共less than 0.6%兲.
Testing Dynamic Fuzzy Wavelet Neural Network Model
The measured acceleration responses of the first and third floors
from the previous Dx = 6 time intervals and the computed accel-
eration response of the second floor from the previous Dy = 6 time
intervals are fed into the input layer to calculate the current ac-
celeration response of the second floor. Each trained model is
tested using all five different levels of excitations 共Kobe1–
Kobe5兲. As such, 25 different sets of system identification results
are produced. The root mean squared sum of the errors 共RMSE兲
between the measured and estimated output is computed for all 25
test cases. The results are shown as a bar diagram in Fig. 7. The
horizontal axis represents the five different excitation levels. As
an example, Fig. 8 shows the identified structural responses of the
second floor subjected to Kobe1 earthquake using the model
trained by the structural responses from Kobe2.
Several observations are made. First, all RMSE values for
identifying the structural responses under smaller Kobe1 and
Kobe2 excitations are less than 8.2%, which is slightly larger than
7% reported by Hung et al. 共2003兲. However, the maximum ab-
solute errors between the computed and measured responses are
around 0.04g, the same as that reported by Hung et al. In all
likelihood, the responses of the structure under earthquakes
Kobe1 and Kobe2 represent the response of the undamaged
structure.
Second, the RMSE values for identifying the structural re-
sponses under the stronger Kobe3, Kobe4, and Kobe5 excitations
are less than 11%, which is significantly lower than 28% reported
by Hung et al. 共2003兲. As an example, Fig. 9 shows the identified

Fig. 9. Identified structural responses of second floor subjected to
Kobe5 earthquake using model trained by structural responses from
Kobe3: 共a兲 comparison of acceleration responses; 共b兲 absolute error,
and 共c兲 comparison of acceleration responses from 7 to 8 s
structural responses of the second floor subjected to Kobe5 earth-
quake using the model trained by the structural responses from
Kobe3. The maximum absolute errors between the computed and
measured response are less than 0.14g compared with 0.2g re-
ported by Hung et al. 共2003兲. It is concluded that the model
trained using the structural response data under a lower excitation
can be used to identify the structural responses under a higher
excitation accurately.
Third, the maximum RMSE value 共11%兲 for identifying the
structural responses under the stronger Kobe3, Kobe4, and Kobe5
excitations is larger than the corresponding value 共8.2%兲 for iden-
tifying the structural responses under lower excitations 共Kobe1
and Kobe2兲. These results imply changes in the structural prop-
erties during larger excitations Kobe3, Kobe4, and Kobe5 because
of damage or inelastic behavior. However, the changes do not
appear to be significant.
Fig. 10 shows the identified structural responses of the second
floor subjected to a lower Kobe2 excitation using the model
trained by the structural responses from the stronger Kobe5 exci-
tation. In this figure the maximum absolute error is around 0.05g,
which is significantly lower than the value of 0.2g reported in
Hung et al. 共2003兲. It is concluded that the model trained using
the structural response data under a larger excitation can still be
used to identify the structural responses under a lower excitation
accurately.
Conclusions
Accurate system identification requires capturing the dynamic
characteristics of time series response data. In this research, a
JOURNAL OF STRUCTURAL ENGINEERING © ASCE / JANUARY 2006 / 109
Downloaded 17 Aug 2010 to 130.54.110.32. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
Fig. 10. Identified structural responses of second floor subjected to
lower Kobe2 excitation using model trained by structural responses
from stronger Kobe5 excitation: 共a兲 comparison of acceleration
responses; 共b兲 absolute error; and 共c兲 comparison of acceleration
responses from 7 to 8 s
dynamic time-delay fuzzy wavelet neural network model is cre-
ated for nonparametric identification of structures using the
NARMAX approach. The following conclusions are made.
1. The wavelet packet denoising technique employed in the
model speeds up the training convergence and improves the
system identification accuracy. Fig. 11 shows comparisons of
the identification results using the denoised and original
structural response data under Kobe4 excitation using the
model trained by structural responses from Kobe1. The
model is implemented in a combination of C⫹⫹ program-
ming language and MATLAB 6.1 共2001兲 on a Windows XP
Professional platform and a 1.5 GHz Intel Pentium 4 proces-
sor. The CPU time for training the model is 54 s using the
denoised data and 129 s using the original data. More sig-
nificantly, the maximum identification error increases to 15%
when the original data are used to train the model. 共The
maximum identification error is 7% when data are denoised.兲
2. In order to preserve the dynamics of time series, the recon-
structed state space concept from the chaos theory is em-
ployed to construct the input vector. Rather than choosing the
dimension of the input space arbitrarily or by trial and error
which is the norm in the structural system identification field,
an optimum value for the input space dimension is found
using the FNN method.
3. Wavelets are employed in two different contexts. First, for
denoising the data as explained under 共1兲. Second, it is used
in combination with two soft computing techniques, neural
networks and fuzzy logic, to create a new pattern recognition
model to capture the characteristics of the time series sensor
data accurately and efficiently. The result is a new fuzzy

Fig. 11. Comparisons of identification results using denoised and
original structural response data under Kobe4 excitation using model
trained by structural responses from Kobe1: 共a兲 comparison of
acceleration responses using denoised sensor data and 共b兲 comparison
of acceleration responses using original sensor data
WNN model. The fuzzy WNN model: 共1兲 balances the global
and local influences of the training data due to the comple-
mentary properties of two soft computing techniques, neural
network and wavelet transform; 共2兲 incorporates the impre-
cision existing in the sensor data effectively due to the use
of fuzzy clustering technique; 共3兲 provides more accurate
system identifications; and 共4兲 results in fast training conver-
gence thus significantly reducing the computational
requirements.
4. The number of fuzzy WNN nodes in the hidden layer is
selected by the Akaike’s final prediction error criterion, thus
avoiding the arbitrary selection of this important neural net-
work parameter.
5. The model provides more accurate nonlinear approximation
because it is based on the integration of chaos theory 共based
on nonlinear dynamics theory兲, wavelets 共a signal processing
method兲, and two complementary soft computing methods,
i.e., fuzzy logic and neural network.
Acknowledgment
The data used to train and validate the new computational model
was provided by Professor Shih-Lin Hung of National Chiao
Tung University and National Center Research on Earthquake
Engineering 共NCREE兲, Republic of China, which is gratefully
acknowledged.
References
Aguirre, L. A., Donoso-Garcia, P. F., and Santos-Filho, R. 共2000兲. “Use
of a priori information in the identification of global nonlinear
110 / JOURNAL OF STRUCTURAL ENGINEERING © ASCE / JANUARY 2006
Downloaded 17 Aug 2010 to 130.54.110.32. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
models—A case study using a buck converter.” IEEE Trans. Circuits
Syst., I: Fundam. Theory Appl., 47共7兲, 1081–1085.
Bani-Hani, K., and Ghaboussi, J. 共1998兲. “Nonlinear structural control
using neural networks.” J. Eng. Mech., 124共3兲, 319–327.
Bezdek, J. C. 共1981兲. Pattern recognition with fuzzy objective function
algorithms, Plenum, New York.
Billings, S. A., and Tsang, K. M. 共1989兲. “Spectral analysis for nonlinear
system, Part I: Parametric nonlinear spectral analysis.” Mech. Syst.
Signal Process., 3共4兲, 319–339.
Billings, S. A., Tsang, K. M., and Tomlinson, G. R. 共1990兲. “Spectral
analysis for nonlinear system, Part III: Case study examples.” Mech.
Syst. Signal Process., 4共1兲, 3–21.
Brown, A. S., and Yang, H. T. Y. 共2001兲. “Neural networks for multiob-
jective adaptive structural control.” J. Struct. Eng., 127共2兲, 203–210.
Chang, C. C., Sun, Z., and Li, N. 共2003兲. “Identification of structural
dynamic properties using wavelet transform.” Structural health moni-
toring and intelligent infrastructure, Z. Wu and M. Abe, eds., Vol. 2,
Balkema, Lisse, The Netherlands, 1243–1248.
Chassiakos, A. G., and Masri, S. F. 共1996兲. “Modeling unknown struc-
tural system through the use of neural networks.” Earthquake Eng.
Struct. Dyn., 25共2兲, 117–128.
Daubechies, I. 共1988兲. “Orthonormal bases of compactly supported wave-
lets.” Commun. Pure Appl. Math., 41, 909–996.
Daubechies, I. 共1992兲. Ten lectures on wavelets, Society for Industrial and
Applied Mathematics, Phildelphia.
Ghanem, R., and Romeo, F. 共2000兲. “A wavelet-based approach for the
identification of linear time-varying dynamical systems.” J. Sound
Vib., 234共4兲, 555–576.
Ghanem, R., and Shinozuka, M. 共1995兲. “Structural-system identification,
I and II.” J. Eng. Mech., 121共2兲, 255–273.
Gurley, K., and Kareem, A. 共1999兲. “Applications of wavelet transforms
in earthquake, wind and ocean engineering.” Eng. Struct., 21共2兲,
149–167.
Hans, S., Ibraim, E., Pernot, S., Boutin, C., and Lamarque, C. H. 共2000兲.
“Damping identification in multi-degree-of-freedom systems via a
wavelet-logarithmic decrement—Part 2: Study of a civil engineering
building.” J. Sound Vib., 235共3兲, 375–403.
Hung, S. L., Huang, C. S., Wen, C. M., and Hsu, Y. C. 共2003兲. “Nonpara-
metric identification of a building structure from experimental data
using wavelet neural network.” Comput. Aided Civ. Infrastruct. Eng.,
18, 358–370.
Jiang, X., and Adeli, H. 共2003兲. “Fuzzy clustering approach for accurate
embedding dimension identification in chaotic time series.” Integrated
Computer-Aided Engineering, 10共3兲, 287–302.
Jiang, X., and Adeli, H. 共2004兲. “Wavelet packet-autocorrelation function
method for traffic flow pattern analysis.” Comput. Aided Civ. Infra-
struct. Eng., 19共5兲, 324–337.
Jiang, X., and Adeli, H. 共2005兲. “Dynamic wavelet neural network model
for traffic flow forecasting.” J. Transp. Eng., 131共10兲, 771–779.
Juang, J. N. 共1994兲. Applied system identification, Prentice Hall, Engle-
wood Cliffs, N.J.
Kijewski, T., and Kareem, A. 共2003兲. “Wavelet transforms for system
identification in civil engineering.” Comput. Aided Civ. Infrastruct.
Eng., 18, 339–355.
Lamarque, C. H., Pernot, S., and Cuer, A. 共2000兲. “Damping identifica-
tion in multi-degree-of-freedom systems via a wavelet-logarithmic
decrement—Part 1: Theory.” J. Sound Vib., 235共3兲, 361–374.
Ljung, L. 共1999兲. System identification—Theory for the user, 2nd Ed.,
Prentice Hall, Upper Saddle River, N.J.
Ljung, L., and Glad, T. 共1994兲. Modeling of dynamic systems, PTR Pren-
tice Hall, Englewood Cliffs, N.J.
Loh, C. H., Lin, C. Y., and Huang, C. C. 共2000兲. “Time domain identifi-
cation of frames under earthquake loading.” J. Eng. Mech., 126共7兲,
693–703.
Masri, S. F., Chassiakos, A. G., and Caughey, T. K. 共1993兲. “Identifica-
tion of nonlinear dynamic systems using neural networks.” J. Appl.
Mech., 60, 123–133.
Masri, S. F., Nakamura, M., Chassiakos, A. G., and Caughey, T. K.
 共1996兲. “Neural network approach to the detection of changes in
structural parameters.” J. Eng. Mech., 122共4兲, 350–360.
Masri, S. F., Smyth, A. W., Chassiakos, A. G., Caughey, T. K., and
Hunter, N. F. 共2000兲. “Application of neural networks for detection of
changes in nonlinear systems.” J. Eng. Mech., 126共7兲, 666–676.
Nakamura, M., Masri, S. F., Chassiakos, A. G., and Caughey, T. K.
共1998兲. “A method for nonparametric damage detection through the
use of neural networks.” Earthquake Eng. Struct. Dyn., 27共9兲,
997–1010.
Sjöberg, J., Zhang, Q., Ljung, L., Benveniste, A., Delyon, B., Glorennec,
P., Hjalmarsson, H., and Juditsky, A. 共1995兲. “Nonlinear black-box
modeling in system identification: A unified overview.” Automatica,
31共12兲, 1691–1724.
Staszewski, W. J. 共1997兲. “Identification of damping in MDOF systems
using time-scale decomposition.” J. Sound Vib., 203共2兲, 283–305.
Staszewski, W. J. 共1998兲. “Identification of nonlinear systems using
multi-scale ridges and skeletons of the wavelet transform.” J. Sound
Vib., 214共4兲, 639–658.
Sugeno, M., and Kang, G. T. 共1988兲. “Structure identification of fuzzy
model.” Fuzzy Sets Syst., 28, 15–34.
Thomson, M., Schooling, S. P., and Soufian, M. 共1996兲. “The practical
application of a nonlinear identification methodology.” Control Eng.
Pract., 4共3兲, 295–306.
Zadeh, L. A. 共1978兲. “Fuzzy set as a basis for a theory of possibility.”
Fuzzy Sets Syst., 1共1兲, 3–28.
Zhang, Q. 共1997兲. “Using wavelet network in nonparametric estimate.”
IEEE Trans. Neural Netw., 8共2兲, 227–236.
Zhang, Q., and Benveniste, A. 共1992兲. “Wavelet networks.” IEEE Trans.
Neural Netw., 3共6兲, 889–898.
Zhou, Z., and Adeli, H. 共2003兲. “Time-frequency signal analysis of earth-
quake records using mexican hat wavelets.” Comput. Aided Civ. In-
frastruct. Eng., 18共5兲, 379–389.

JOURNAL OF STRUCTURAL ENGINEERING © ASCE / JANUARY 2006 / 111

Downloaded 17 Aug 2010 to 130.54.110.32. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org