Computers and Structures 77 (2000) 4152

www.elsevier.com/locate/compstruc

Substructural identication using neural networks

Chung-Bang Yun*, Eun Young Bahng
Department of Civil Engineering, Korea Advanced Institute of Science and Technology, Taejon, 305-701, South Korea
Received 18 February 1997; accepted 27 July 1999

Abstract
In relation to the problems of damage detection and safety assessment of existing structures, the estimation of the
element-level stiness parameters becomes an important issue. This study presents a method for estimating the
stiness parameters of a complex structural system by using a backpropagation neural network. Several techniques
are employed to overcome the issues associated with many unknown parameters in a large structural system. They
are the substructural identication and the submatrix scaling factor. The natural frequencies and mode shapes are
used as input patterns to the neural network for eective element-level identication particularly for the case with
incomplete measurements of the mode shapes. The Latin hypercube sampling and the component mode synthesis
methods are adapted for ecient generation of the patterns for training the neural network. Noise injection
technique is also employed during the learning process to reduce the deterioration of the estimation accuracy due to
measurement errors. Two numerical example analyses on a truss and a frame structures are presented to
demonstrate the eectiveness of the present method. 7 2000 Elsevier Science Ltd. All rights reserved.
Keywords: Substructuring identication; Neural networks; Modal data; Noise injection learning; Latin hypercube sampling; Component mode synthesis

1. Introduction
Structural identication has become increasingly an
important research topic in connection with damage
assessment and safety evaluation of existing structures
[14]. The stiness parameters such as axial and exural rigidities can be identied based on the measured
responses and excitations using various system identication techniques. Such inspection approaches provide
nondestructive means of safety assessment. Recent
development in this area is made possible by rapid
advances in computer technology for data acquisition,

* Corresponding author. Tel.: +82-42-869-3612; fax: +8242-869-3610.

E-mail address: [email protected] (C.B. Yun).

signal processing and analysis. In general, the system

identication methods inherently involve comprehensive search processes such as extended Kalman lter,
recursive least squares or modal perturbation. Most of
them are computationally expensive and tend to be numerically unstable for the case with a large-degrees-offreedom system. In recent years, instead of the comprehensive search processes, pattern-matching techniques
using neural networks have drawn considerable attention in the eld of damage assessment and structural
identication [512]. Neural networks have unique
capability to be trained to recognize given patterns and
to classify other untrained patterns. Hence, the neural
networks can be used for estimating the structural parameters with proper training.
Several researchers have dealt with neural network
approaches for damage estimation of small degrees-of-

0045-7949/00/$ - see front matter 7 2000 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 4 5 - 7 9 4 9 ( 9 9 ) 0 0 1 9 9 - 6

42

C.-B. Yun, E.Y. Bahng / Computers and Structures 77 (2000) 4152

freedom models. Wu et al. applied the neural networks

to recognize the locations and the extent of individual
member damage of a simple three story frame [5].
Szewczyk and Hajela proposed a modied counterpropagation neural network to estimate the stiness of individual structural elements based on observable static
displacements under prescribed loads [6]. Their pilot
studies showed that the neural-network-based
approach for damage assessment is promising. Tsou
and Shen identied damage of a spring-mass system
with eight degrees-of-freedom (DOF) that has closely
spaced eigenvalues [8]. They constructed three multilayer subnets which performed three tasks of input
pattern generation, damage location identication, and
damage severity determination. Their study showed
good performance for on-line damage identication.
Yagawa et al. applied neural network approach to estimate stable crack growth in welded specimens [9].
Yoshimura et al. proposed the Generalized-Space-Lattice transformation of training patterns to relax the illposedness of the problem. As the eect of transformation, a training process converged fast, and estimation
accuracy was improved signicantly [10].
In this study, the neural network-based approach is
extended to the estimation of stiness parameters of a
complex structural system with many unknowns. Several techniques, such as substructural identication
[1315] and submatrix scaling factor [16] are employed
to overcome the issues associated with many
unknowns. The concept of modal strain energy [17] is
utilized for selection of modes to be used as input to
the neural networks, and Latin hypercube sampling
technique [18] and the component mode synthesis
method [19] are employed for ecient generation of
training patterns. Two numerical example analyses on
a truss and a frame structures are presented to demonstrate the eectiveness of the proposed method.

Fig. 1. Architecture of backpropagation neural network.

2. Neural network-based structural identication

Studies on neural networks have been motivated to
imitate the way that the brain operates. It has recently
drawn considerable attention in various elds of
science and technology, such as character recognition,
electro-communication, image processing, and industrial control problems [20]. Many researchers have
developed various neural network models for dierent
purposes [21]. In this study, a model called backpropagation neural network (BPNN) or multilayer perceptron is used for structural identication. The BPNN
consists of an input layer, hidden layers, and an output
layer as in Fig. 1. The input and output relationship of
a neural network can be nonlinear as well as linear,
and its characteristics are determined by the weights
assigned to the connections between nodes in two adjacent layers. Changing these weights will change the
input/output behavior of the network. Systematic ways
of determining the weights of the network to achieve a
desired input/output relationship are referred to as
training or learning algorithm. In this study, the standard backpropagation algorithm is used. The basic
strategy for developing a neural network-based
approach to identify a structural system is to train the
BPNN to recognize the element-level structural parameters from the measurement data on the structural
behavior, such as natural frequencies and mode
shapes.
2.1. Substructural identication
For the identication of a structure with many
unknowns, it is not practical to identify all of the parameters in the structure at a same time, because most
of the identication techniques require expensive computation that would be even prohibitive, as the number
of unknown parameters increases. Several researches
were reported on identication of a part of a structure
so as to reduce the size of the system under consideration [1315]. Those works were based on the reasoning that the expected damages of a structure occur at
several critical locations, hence it is more reasonable to
concentrate the identication at critical locations of
the structure.
The local-identication method is generally based on
substructuring, in which the structure is subdivided

Fig. 2. Substructuring for localized identication.

C.-B. Yun, E.Y. Bahng / Computers and Structures 77 (2000) 4152

into several substructures and the identication is carried out on a substructure at a time. In the present
study, the substructure to be identied is called as the
internal substructure, while the others as the external
substructures as shown in Fig. 2. Since the parameters
to be estimated are limited to an internal substructure,
it is expected that the numerical problems such as
divergence or falling into local minima may be
avoided. Another advantage is that this approach
requires measurement data only on the substructure of
interest, instead of on the whole structure.
The two-stage identication is another idea to
reduce the number of variables. The two-stage identication was also studied by the present authors [22]. In
the study, Topole and Stubbs's damage index method
[23] was employed to select probable damaged members in the rst stage identication. Then the damage
severities of the selected members were estimated using
the neural network technique in the second stage
identication. However, there may be many cases
where we can estimate which region shall be investigated through visual inspection or knowledge of
experts without carrying out the rst stage identication. Then the substructural identication can be
performed on the interested part of the structure, while
the parameters of the other parts are still unknown. In
the worst case any information may not be obtained to
determine the substructure to be identied, then the
substructuring technique can be applied to each substructure in turn.
2.2. Submatrix scaling factors
The number of the parameters to be estimated shall
be kept reasonably small for successful identication
of a structure. Reduction of parameters will improve
the results of the identication and minimize the
required measurement data. For the purpose of ecient parameterization of the structure, the submatrix
scaling technique is adopted in this study [16]. Using
the technique, the stiness matrix of the system can be
described by introducing the submatrix scaling factors
(SSF) corresponding to the element level stiness
matrices as
K

q
q
X
X
Kj
sj K0j
j1

j1

where K is the stiness matrix of the structure in the

present condition, Kj and K0j are the present and the
reference values of the stiness submatrix for the jth
element transformed into the global coordinates. sj is
the j-th SSF, and q is the total number of submatrices.
Then, the identication of the stiness matrix can be
performed by estimating the submatrix scaling factors

43

instead of all of the stiness matrix coecients. Direct

identication of the system stiness matrix coecients
may face with problems which violate the basic properties of the stiness matrix related to the member connectivity, boundary condition, and those as
symmetricity and positive deniteness. However, the
use of the SSFs described in Eq. (1) can avoid such
problems.
2.3. Input to neural networks
A way of choosing the patterns representing the
characteristics of the structure, which are to be used as
the input to neural networks, is one of the most important subjects in this approach. Several researchers
have used the various input patterns (input vectors)
suitable for their purpose. For example, Wu et al. used
the frequency spectrum for each DOF of the example
structure for damage estimation [5]. This input has an
advantage that does not need the modal parameter
identication from the measurements. Nonetheless, to
characterize those spectral properties, many sampling
data are required. Accordingly, a large number of
input neurons in the neural networks are needed,
which may reduce the eciency and accuracy of the
training process. Tsou and Shen used the dynamic residual vector that can be obtained from the modal
data [8]. It provides a simple and eective way to
detect the damage and the length of the input pattern
can be reduced signicantly compared with the spectrum data. However, it still has the restriction that the
modes should be measured at every nite element
DOF. In reality, the modes obtained from a typical
modal survey are generally incomplete in the sense
that the mode vector coecients are available only at
the measured DOFs. Therefore, it is desirable to use
the input patterns which are more suitable for the case
with partial measurement data.
The natural frequencies and modes of a structure
are used as the input patterns in this study. The choice
has been made based on the following advantages: (1)
the length of the input pattern is limited to several
times of the number of the measured DOFs; (2) the
natural frequencies represent global behaviors, while
the mode vectors do local characteristics; and (3) they
can be obtained from the measurements of the structural behavior. The relationship between the modes
and the stiness matrix can be written as
KF MFL

where M is the mass matrix, F is the modal matrix,

and L is the diagonal matrix containing the eigenvalues. The mass matrix is assumed to be known
exactly in this study.
Using the substructuring technique and referring to

44

C.-B. Yun, E.Y. Bahng / Computers and Structures 77 (2000) 4152

Fig. 2, the modal matrix F can be partitioned as

2

FL1
4
F FL2
FL3

3
FH
1
5
FH
2
FH
3

2.4. Generation of training and testing patterns

where the subscripts 1, 2, and 3 denote substructures,

while Substructure 2 is the internal substructure, and
the superscripts L and H represent lower and higher
modes, where only the lower modes are generally used
for the structural identication. Rewriting the ith
mode in the partitioned modal matrix FL2 as j~ i for the
simplicity, the input pattern can be dened as
Input pattern vector f fi , j~ 1i , . . . , j~ ni , i 1, . . . ,
mg

where fi is the ith natural frequency, j~ ji denotes the jth

component of j~ i which is normalized as j~ Ti j~ i 1, n is
the number of DOFs measured for Substructure 2, and
m is the number of modes to be included in the identication.
A useful relationship between the submatrix for the
element stiness and the modes may be obtained by
examining the strain energy of each submatrix with
respect to the modes [17]. The modal strain energy distribution among the stiness submatrices for the ith
mode can be evaluated as
bji

jTi K0j ji 1
jTi Mji o2i

where bji is the modal strain energy (MSE) coecient

of submatrix K0j in the ith mode, ji is the ith mode in
the modal matrix F, and oi is the ith circular natural
frequency. The submatrices containing large values of
the MSEs for a mode indicate that the corresponding
structural elements are major load carrying elements
for the particular mode. Stiness changes in those elements will cause signicant changes in the frequency
and mode shape. Hence, detection of the stiness
changes may be more eective by examining the modal
data for those elements. On the other hand, the structural elements having negligible MSEs reect that it
would be dicult to detect the stiness changes in
those elements using the corresponding modal data.
Topole and Stubbs utilized the changes in the MSEs
from the undamaged to the damaged structures to predict structural damage [23]. In the present study, however, the MSEs are used as indicators in selecting
important vibration modes for the input patterns. The
MSEs distribution gives a good information for selecting the modal data.

Since the neural network-based structural identication is highly dependent on the training patterns, it
is very important to prepare well-examined data sets.
The number of training patterns must be large enough
to represent the structural system properly. However,
for the computational eciency, it must be limited to
be reasonably small, because most of the computational time for this approach is required for preparing the training patterns and training the networks.
When the number of the unknown parameters is P
and each parameter has L sampling points, the size of
the whole population is L P. This size increases exponentially, as the number of the unknown parameters
does. To reduce the number of training patterns eectively, a sparse sampling algorithm such as Latin
hypercube technique [18] is employed in this study,
which makes the required number of samples reduced
from L P to L.
The training patterns for the proposed neural network-based method consist of the modal data as input
and the corresponding SSFs as output. To generate
training patterns, a series of eigenanalyses are to be
performed. However, repeated computations of eigenvalue problems for a complex structure can be very expensive. Thus, the xed-interface-component-modesynthesis (CMS) method [19] incorporating the substructuring technique is employed. In the component
mode synthesis method, each substructure is analyzed
independently for local mode shapes. Then, the component mode shapes are assembled to construct Ritz
vectors, which are subsequently used to compute the
mode shapes for the whole structure. When the number of DOFs is large, the computational eciency of
the CMS method becomes more signicant.
The neural-network approach is an approximation
technique that handles non-unique cases by either
returning one of the possible solutions or an average
taken over all possible solutions [6]. Although the
input information to the neural network is limited to
the components of the mode vectors for an internal
substructure by applying the substructuring technique
to the structural identication, most of the information
regarding the stiness parameters of the internal substructure is included in the components of the mode
vectors of the internal substructure, rather than those
of the external substructures. Thus, those stiness parameters may be estimated with a reasonable accuracy
from the components of mode vectors of the internal
substructure as the input to the neural network. On
the other hand, as the range of the submatrix scaling
factors in the external substructure reduces, the estimation accuracy increases. If possible, the uncertainties
of the unknowns in the external substructure shall be
reduced by using all the available information.

C.-B. Yun, E.Y. Bahng / Computers and Structures 77 (2000) 4152

In the process of computing the mode vectors for

each pattern, the mode shapes may be obtained in
dierent order compared with the baseline case, particularly for the higher modes. In this study, the mode
shapes of each pattern are rearranged to match the
baseline shapes using the modal assurance criteria
(MAC) [24], in which the MAC is dened as
MACij

~ Tj j2
j j~ T
i j

~ i kj~ Tj j~ j j
j j~ T
i j

For instance, to select the mode matching to the ith

baseline shape j~ i , the mode j~ j corresponding to the
largest MACij shall be chosen. Additionally, the direction (sign) of the mode shape must be kept same as
the corresponding baseline shape. The condition can
be satised, if the inner product of the mode and its
baseline is always kept positive.
2.5. Noise injection learning
A crucial problem on the BPNN is its generalization
capability. Usually, training patterns used for learning
are only a limited number of samples selected from a
population of input and output patterns. Hence, a network successfully trained to a given set of samples
may not provide the desired input and output associations for untrained patterns, particularly in the cases
with measurement noises and errors in the mode estimation. Concerning this problem, several researchers
reported that adding some noise to input patterns
during backpropagation learning process can remarkably enhance the generalization capability of the resultant networks, if the mapping from the input space to
the output space is smooth [25,26]. Matsuoka found
that the noise injected into the input reduces the sensitivity of the network to the variation of the input [25]:
that is, it makes the mapping from the input space to
the output space be smooth.
In this study, noises are articially added to natural
frequencies and mode shapes in generating the training
and the testing patterns as
f

n
i

fi 1 vi

j~ nji j~ ji 1 vji

7
8

where f ni and j~ nji are the ith natural frequency with

noise and the j-th component of the corresponding
modes with noise, and vi and vji are Gaussian random
noises. The sequence of noise is generated by using a
subroutine which returns a series of random numbers
normally distributed with zero mean and a specied
standard deviation. Based on Eqs. (7) and (8), the
noise level can be dened as the value of the standard

45

deviation. For instance, if the standard deviation is

0.03, the noise level in the modal data can be referred
as 3% in the root-mean-square (RMS) level.

3. Illustrative numerical examples

Two dierent types of structures were selected for
example analyses to illustrate the applicability of the
proposed approach. They are a two-span planar truss
and a multi-storey frame structure.
3.1. Truss structure
A truss example is the two-span planar truss which
contains 55 truss elements, 24 nodes and 44 nodal
DOFs as shown in Fig. 3. It was assumed that the
baseline parameters were known. Values for the material and geometric properties are as follows: the elastic
modulus=200
GPa;
the
cross-sectional
area=0.03 m2; and the mass density=8000 kg/m3. As
the rst step of the identication procedure, the stiness matrix of the structure is represented as
K

55
X
sj K0j

j1

Then, the problem is to identify 55 unknown submatrix scaling factors, sj, j = 1, . . . , 55, based on the
measured modal data. Instead of identifying the whole
SSFs simultaneously, the structure was subdivided into
three substructures; an internal substructure and two
external substructures as shown in Fig. 3. It was
assumed that the unknown SSFs for the elements in
the internal substructure were between 0.3 and 1.7,
while those for the external substructures were between
0.5 and 1.5. The mode shapes were assumed to be
measured only at 10 test DOFs, which include the displacements in the x- and y-directions at ve nodes in
the internal-substructure.
3.1.1. Modal data for input to BPNN
For ecient training process, appropriate modal
data shall be selected as the input vector. It was
assumed that the rst ve modes of the internal-substructure were available at the test DOFs. The natural
frequencies and mode shapes of the reference structure,
in which all the SSFs are unity, are shown in Fig. 4.
The rst, second and fth modes are found to be vertical, i.e., exural modes, the third is a mixed mode with
the vertical and horizontal motions, and the fourth is a
horizontal mode. The MSE distribution in the rst ve
modes of the reference structure is shown in Fig. 5. It
shows that the Elements 1, 2, 4, 5, 7, and 10 are carrying large fractions of the strain energy in most of

46

C.-B. Yun, E.Y. Bahng / Computers and Structures 77 (2000) 4152

Fig. 3. Substructuring for localized identication.

modes. Especially in the third and fourth modes which

contain the horizontal motions, the horizontal elements
such as the Elements 2, 5, 7, and 10 have relatively
large values of MSE. On the other hand, the MSE
coecients associated with the Elements 3, 6, 8, 9, and
11 are relatively small for all of the ve modes, which
indicates that the accuracy of the SSF estimation for
the elements using the rst ve modal data may be
poor. Based on the above observation, the rst four
modal data were selected as an input pattern for this
example. However, other cases with dierent combinations of modes were also tested for the purpose of
comparison.
3.1.2. Training and testing patterns
For the generation of training patterns using the
Latin hypercube sampling, 15 sampling points between
0.3 and 1.7 were selected with an equal interval of 0.1
for each SSF in the internal substructures, assuming
that each SSF is uniformly distributed. On the other

Fig. 4. First ve mode shapes of a truss structure.

hand, 11 sampling points between 0.5 and 1.5 were

taken with an interval of 0.1 for the SSFs in the external substructures. Consequently, the number of the
possible sample cases is 165 (=15  11), which was
not judged to be enough to represent the system adequately. Hence, 10 iterations of the Latin hypercube
sampling process were taken, and the 1650 training
patterns were sampled. For the cases of testing patterns, the sampling points for the internal substructure
were taken as the same with those for the training
data, however the number of sampling points for the
external substructures was taken to be 10 in the range
of 0.551.45 with an interval of 0.1. The dierences in
sampling points were imposed, because the real testing
patterns are not necessarily the same as the training
patterns. Thus, the number of possible sample cases
for testing was 150 (15  10).
Using the CMS method, natural frequencies and
mode shapes were computed for sampled values of the
SSFs. In the process of computing the mode vectors
for the sampled training and testing patterns and rearranging those in order of the baseline, the mode
shapes of some patterns may be very dierent with
those of the baseline. Accordingly, those patterns were
eliminated from the training and testing patterns.
Actually 1200 of the sampled training patterns were
used for training process and 100 of the sampled testing patterns were used for testing process. The elapse
time for generating whole training data set was about
53 sec by the Alpastation 2004/233 workstation. The
variations of the SSFs in the external substructures
were considered in preparation for the training data
set, though they were not included in the output vector. Therefore, the output vector of the BPNN consists
of the SSFs in the internal substructure, while the
input vector consists of the rst four natural frequencies and the mode shape components for the internal
structure. The training and testing data set were prepared for the cases with 0, 3, 5, and 7% noise in RMS
level. An example of training patterns was shown in
Table 1.

C.-B. Yun, E.Y. Bahng / Computers and Structures 77 (2000) 4152

47

Fig. 5. Modal strain energy distribution for internal substructure of a truss structure.

3.1.3. Training and testing processes

It is important to choose the proper network size. If
the network is too small, it may not be able to represent the system adequately. On the other hand, if the
network is too big, it becomes over-trained and may
provide erroneous results for untrained patterns. In
general, it is not straight forward to determine the best
size of the networks for a given system. It may be
found through trial and error process using knowledge
about the system. A four-layer neural network as

shown in Fig. 1 was selected for the present example.

The number of the neurons in the input layers varies
depending on the number of modes included: i.e. 11
neurons for each mode. However, the numbers of the
neurons in the rst hidden, the second hidden and the
output layers were taken to be xed to 17, 13, and 11,
respectively as shown in Fig. 1. Numerical investigations show that small changes of the numbers of the
neurons in the hidden layers have little eect on the
estimation results.

Table 1
One example of training patterns
(a) Input pattern without noise
5.352
7.594
10.415
12.913

0.362
0.045
0.347
0.032

0.340
0.080
0.345
0.242

0.050
0.076
0.321
0.358

0.268
0.064
0.418
0.637

0.113
0.071
0.468
0.591

0.002
0.187
0.121
0.208

0.299
0.386
0.175
0.086

0.290
0.374
0.196
0.022

0.495
0.572
0.283
0.036

0.497
0.571
0.317
0.071

0.050
0.074
0.318
0.334

0.268
0.064
0.430
0.615

0.105
0.069
0.485
0.622

0.003
0.207
0.122
0.212

0.299
0.357
0.159
0.075

0.279
0.371
0.209
0.021

0.496
0.576
0.305
0.034

0.511
0.581
0.295
0.067

0.3

0.9

1.1

0.6

1.6

1.1

1.1

1.6

(b) Input pattern with 3% noise in RMS

5.727
7.637
10.967
12.349

0.358
0.049
0.327
0.028

0.333
0.080
0.328
0.259

(c) Output pattern (SSFs, sIj )

0.5

1.5

1.7

48

C.-B. Yun, E.Y. Bahng / Computers and Structures 77 (2000) 4152

Table 2
One example of test outputs
NN Input
( fj, j~ ji )

NN output sIj )
Target output
Relative error (%)a
a

5.28
8.01
10.63
11.92

0.35
0.03
0.24
0.15

0.33
0.14
0.28
0.12

0.03
0.07
0.27
0.20

0.22
0.11
0.56
0.76

0.12
0.07
0.59
0.52

0.00
0.08
0.00
0.04

0.30
0.30
0.05
0.06

0.29
0.27
0.07
0.20

0.50
0.62
0.20
0.02

0.50
0.61
0.23
0.09

1.47
1.60
7.65

0.89
0.90
0.26

1.44
1.40
2.98

0.90
0.90
0.22

1.13
1.10
3.06

0.37
0.40
5.32

0.35
0.30
19.67

1.27
1.20
5.89

0.45
0.50
8.10

0.62
0.60
4.79

1.52
1.70
10.44

output
Relative error 100  j NN outputTarget
j.
Target output

The architecture of the neural network and the

required number of training patterns are inter-related.
However, there is no established method to determine
the relationship. According to the VapnikChervonenkis dimension theory [21], the required number of
training patterns may be roughly taken as two times of
the number of weights. In this study, however, the
number of the training patterns was taken as 1200
through trial and error, which is approximately the
number of weights including threshold (1153) as shown
in Fig. 1. It is found that the values of the weights
after the learning process are well-distributed without
severe saturation. It means that the present architecture is suitable and the number of the training patterns
is sucient.

The training process took about 1000 epochs to

learn the pattern representation using the standard BP
algorithm, in which the elapse time was about 220 s by
an Alpastation computer system. Generally the training process is a time consuming work. However, once
the training is completed, the neural networks can be
used very eciently and repeatedly for identifying the
system parameters. Hence, it can be a very useful tool
for structural health monitoring.
Each of the training patterns was used once for
training at an epoch, then the order of the training
patterns is randomly altered at the next epoch. An
example of the tested outputs was listed in the Table 2.
Fig. 6 shows the averages of the relative errors in the
estimated SSFs for various cases with dierent number

Fig. 6. Average estimation errors for dierent modes used in input patterns (for testing data set without noise).

C.-B. Yun, E.Y. Bahng / Computers and Structures 77 (2000) 4152

49

Fig. 8. Learning curves without and with noise injection in

training patterns (for testing data set with 7% noise).
Fig. 7. Estimation error for dierent noise injection learning.

of modes included in the input patterns. It can be

found that the accuracy of the estimation improves
with increasing number of the modes included until the
fourth mode. But adding the fth mode did not
improve the estimates. Hence, the rst four modes
were included in the input to the BPNN.
The eect of the measurement errors in the testing
data set and the eectiveness of the noise injection
learning were also investigated. Fig. 7 shows the estimation errors for the SSFs for dierent cases with various noise levels in the training and testing data sets.
The noises in the training patterns mean the noises
injected to the input patterns during the training process, and those in the testing patterns indicate the
measurement errors. Fig. 7 shows that the average estimation error for the noise-free testing data was
obtained as 8.9%, when the neural network has been
trained using the noise-free training data set. But, the
average error for the testing data with 5% noise using
the same neural network was found to be 18.3%,
which implies that the measurement error causes de-

terioration in the estimation. On the other hand, when

the network has been trained using the training data
set with 5% noise, the error for the testing data with
5% noise was obtained as 13.5%, which is much smaller than the error using the noise-free training data.
This indicates that it is desirable to use the 5% noise
data for training, if the measurement error is expected
to be 5%. As expected from the MSE distribution, the
estimation errors for the Elements 3, 6, 8, 9, and 11
are larger than those for the other elements. However,
the results generally lie in a satisfactory range. Table 3
shows the similar results for the cases with 3 and 7%
measurement noises. The results also indicate that the
estimation accuracy can be improved signicantly, if
noise with an intensity similar to the measurement

Table 3
Average estimation errors (%) for dierent noise injection
levels
Noise levels in RMS

In training data
0%
3%
5%
7%

In testing data
0%

3%

5%

7%

8.9
9.7
10.4
12.0

12.8
11.1
11.4
12.6

18.3
14.0
13.5
13.7

22.4
15.8
14.8
14.7

Fig. 9. Two-bay and ten-storey frame structure.

50

C.-B. Yun, E.Y. Bahng / Computers and Structures 77 (2000) 4152

Table 4
Baseline parameters for frame examplea

Beams
Columns
a

Shape

Area (m2)

I (m4)

Density (kg/m3)

W24  15
W14  145

1.04  102
2.78  102

5.62  104
7.12  104

7850
7850

I is moment of inertia of the cross-sectional area; E is modulus of elasticity at 210 GPa.

noise level is imposed in the training process. Fig. 8 is

the learning curves to depict how the noise injection
learning works. The learning rate is relatively high at
the epochs from 1 to 100, and after that, it becomes
lower. It can be observed from the curves that the
noise injection improves the results of estimation
during the learning process. Generally, the back-propagation algorithm requires some reasonable criteria to
stop the weight adjustment. The number of learning
epochs was used as the termination criterion in this
paper, and it was taken as 1000 epochs, because the
learning curves showed that the decreasing rate of the
testing error became very at after the epoch as in
Fig. 8. The level of the testing error may be used as a
practical termination criterion, too.
3.2. Frame structure
The second example is a frame structure as shown in
Fig. 9. The planar frame structure contains 50 elements, 33 nodes and 90 nodal DOFs. The baseline
parameters were listed in Table 4. The SSFs of the internal substructure were estimated for twelve elements,
marked in Fig. 9. The natural frequencies and mode
shapes of the frame structure are shown in Fig. 10.
The four measurement cases in Fig. 11 were investigated on the eect of rotational and translational components of mode shapes. The rst four modes were
used as input patterns to the neural networks; 1200
training and 100 testing patterns were actually used
during the learning process, similarly with the truss
example.
Fig. 12 shows the average estimation errors for four

Fig. 10. First four mode shapes of a frame structure.

cases with dierent measured components of the mode

vectors. When all the DOFs at 9 nodes were used as
the input (Case 1), the estimation accuracy was good.
However, when only the x- and y-DOFs were included
in the input pattern (Case 2), the estimation accuracy
became poor. On the other hand, when the translational DOF at the mid point of each element in the internal substructure was added into the input pattern
(Case 3), the estimation accuracy was found to be
improved signicantly, which indicates that the translational data at the additional nodes may make up for
the information on the unmeasured rotational DOFs,
which are usually dicult to be measured in the eld
experiment. Additionally when the translational DOFs
at two quarter points of each element were added into
the input pattern (Case 4), the estimation accuracy was
found to be better than the case using the rotational
DOFs (Case 1). The average relative errors for the
training and testing data set with 5% noise were 8.8
and 9.8%, respectively. It means that the generalization performance of the present neural networks is
very good.

4. Conclusions
A neural network-based substructural identication
was presented for the estimation of the stiness parameters of a complex structural system, particularly
for the case with noisy and incomplete measurement of
the modal data. The proposed approach does not
require any complicated formulation for model reduction of the system, which is indispensable to the
conventional approaches for substructural identication. On the other hand, this approach requires generation of enough training patterns which can
adequately represent the relationship between the stiness parameters and the modal information. The Latin
hypercube sampling and the component-mode-synthesis methods were used for the ecient generation of
such training patterns.
The numerical example analyses were carried out
on a two-span truss and a multi-storey frame, and
the results were summarized as follows: (1) The substructuring technique and the concept of the subma-

C.-B. Yun, E.Y. Bahng / Computers and Structures 77 (2000) 4152

51

Fig. 11. Cases of measured DOFs (sensor locations).

Fig. 12. Average estimation error for dierent measured

DOFs with 5% measurement noise (w: Case 1, +: Case 2, Q:
Case 3, R: Case 4).

trix scaling factor were found to be very ecient to

reduce the number of unknown stiness parameters
to be estimated. (2) The modal strain energy coecients were eective indicators in selecting vibration
modes for the input to the neural networks. (3) In
the truss structure, the horizontal elements were
major load carrying members for lower modes,
hence the accuracy of the parameter estimation for
the elements was better than the others. (4) For the
case of the frame structure, the estimates became
deteriorated, when the rotational DOFs at the nodes
were not included in the input to the neural networks. However, the estimation accuracy could be
improved signicantly, when additional measurement
information on the translational DOFs within the elements was utilized. (5) The noise injection learning
using noise at similar level to the measurement error
was found to be very eective to improve the estimation accuracy. (6) The average relative estimation
errors for testing data set with various noise intensities were found to be in the range of 915%, which
shows the applicability of the present method for the
identication of large structural systems.

52

C.-B. Yun, E.Y. Bahng / Computers and Structures 77 (2000) 4152

Acknowledgements
The authors would like to thank to the Korea
Earthquake Engineering Research Center and the
Korea Science and Engineering Foundation for their
nancial support for this study.

[12]
[13]
[14]

References
[1] Natke HG, Yao JT-P. System identication approach in
structural damage evaluation. In: Proceedings of ASCE
Structural Congress '86, Preprint 17-1, 1986.
[2] Hong K-S, Yun C-B. Improved method for frequency
domain identication of structures. Engng Struct
1993;15(3):17988.
[3] Ghanem R, Shinozuka M. Structural-system identication I: theory. J Engng Mech ASCE 1995;121(2):255
64.
[4] Yun C-B, Lee H-J, Lee C-G. On sequential prediction
error method for structural identication. J Engng Mech
ASCE 1997;123(2):115122.
[5] Wu X, Ghaboussi J, Garrett Jr JH. Use of neural networks in detection of structural damage. Comput Struct
1992;42(4):64959.
[6] Szewczyk Z, Hajela P. Neural network based damage
detection in structures. ASCE Journal of Computing and
Civil Engineering 1994;8(2):16378.
[7] Manning RA. Structural damage detection using active
members and neural networks. AIAA Journal
1994;32(6):13313.
[8] Tsou P, Shen M-HH. Structural damage detection and
identication using neural networks. AIAA Journal
1994;32(1):17683.
[9] Yagawa G, Matsuda A, Kawate H, Yoshimura S.
Neural network appoach to estimate stable crack growth
in welded specimens. International Journal of Pressure
Vessels and Piping 1995;63:30318.
[10] Yoshimura S, Matsuda A, Yagawa A. New regularization by transformation for neural network based inverse
analyses and its application to structure identication.
International Journal of Numerical Methods in
Engineering 1996;39:395368.
[11] Okuda H, Miyazaki H, Yagawa G. A neural network
approach for modeling of viscoplastic material behaviors.

[15]

[16]
[17]
[18]

[19]
[20]
[21]
[22]

[23]
[24]
[25]
[26]

Advanced
Computer
Applications,
ASME/PVP
1994;274:1415.
Pandey PC, Barai SV. Multilayer perceptron in damage
detection of bridge structures. Comput Struct
1995;54(4):597608.
Koh CG, See LM, Balendra T. Estimation of structural
parameters in the time domain: a substructure approach.
Earthquake Engng Struct Dyn 1991;20:787801.
Oreta AWC, Tanabe T-A. Element identication of
member properties of framed structures. J Struct Engng
ASCE 1994;120(7):196176.
Yun C-B, Lee H-J. Damage estimation of structures
incorporating
substructural
identication.
In:
Proceedings of Asian-Pacic Symposium on Structural
Reliability and its Applications (APSSRA), Tokyo
Japan, 1995. pp. 220228.
Lim TW. Submatrix approach to stiness matrix correction using modal test data. AIAA Journal
1990;28(6):112330.
Lim TW. Structural damage detection using modal test
data. AIAA Journal 1991;227174.
Press W, Teukolsky SA, Vetterling WT, Flannery BP. In:
Numerical recipes in C The art of scientic computing. New York: Cambridge University Press, 1992. pp.
315.
Craig Jr RR, Bampton MCC. Coupling of substructures
for dynamic analysis. AIAA Journal 1968;6(7):13139.
Haykin S. Neural networks A comprehensive foundation. New York: Macmillan, 1994.
Hush DR, Horne BG. Progress in supervised neural networks. IEEE Signal Processing Magazine, 1993. p. 839.
Yun C-B, Bahng EY. Neural network approach to
damage assessment of civil structure. In: Proceedings of
the Structural Engineers World Congress (SEWC '98),
San Francisco, California, USA, 1923 July 1998, Paper
reference: T1312, 1998.
Topole KG, Stubbs N. Non-destructive damage evaluation of a structure from limited modal parameters.
Earthquake Engng Struct Dyn 1995;24:142736.
Ewins DJ. In: Modal testing: theory and practice.
England: Research Studies Press, 1986. pp. 2245.
Matsuoka K. Noise injection into inputs in back-propagation learning. IEEE Trans Systems, Man and
Cybernetics 1992;22(3):43640.
Holmstrom L, Koistinen P. Using additive noise in backpropagation training. IEEE Trans Neural Networks
1992;3(1):2438.