Article

Structural Health Monitoring

11(3) 281–292

Damage detection for a frame Ó The Author(s) 2011

Reprints and permissions:
[Link]/[Link]
structure model using vibration DOI: 10.1177/1475921711430437
[Link]
displacement measurement

B Xu1,2, G Song3 and Sami F Masri4

Abstract
A structural parameter identification and damage detection approach using displacement measurement time series is
proposed, and the performance of the approach is validated experimentally with a frame structure model in a healthy
condition and with joint connection damages. The dynamic displacement response of the frame structure under base
excitation is measured by noncontact laser displacement sensors. The proposed approach is carried out using two
neural networks: one is called displacement-based neural network emulator (DNNE) and the other is called parametric
evaluation neural network (PENN). First, the theoretical basis and the selection of the input and output of the DNNE
and the PENN are explained, and second, an identification index called root mean square (RMS) of prediction displace-
ment difference vector (RMSPDDV) is defined. The performance of the proposed methodology for damage detection of
the frame structure model with different joint damage scenarios introduced by loosening the bolts connecting the beams
and columns is investigated with the direct use of displacement measurement under base excitations. The results from
the proposed time domain displacement-based identification approach are compared with them based on the extracted
frequencies and show that the proposed time domain methodology can identify the variation of interstory stiffness due
to the joint damage with acceptable accuracy without any modal shapes and frequencies extracted from a dynamic test.
The proposed approach provides an alternative way for damage detection of engineering structures by the direct use of
structural dynamic displacement measurements.

Keywords
parameter identification, damage detection, time series, displacement response, neural network, laser sensor, frame
structure, base excitation

Introduction and plays important roles in structural health monitor-

ing system. Because damage in an engineering structure
Civil infrastructures are continuously deteriorating due usually leads to a decrease in structural stiffness, dam-
to material aging, misuse, lack of proper maintenance, age to be identified is usually represented in the form of
and, in some cases, overstressing as a result of increas- a decrease in stiffness in a structural element level or
ing load demands and changing environments. Failure substructure level. Therefore, damage detection is
of these infrastructures often leads to a high social con-
sequence. It has been a crucial topic for developed and
developing countries to evaluate the current perfor- 1
College of Civil Engineering, Hunan University, Changsha, Hunan, China.
2
mance and the condition of engineering structures in Ministry of Education Key Laboratory of Building Safety and Energy
Efficiency (Hunan University), Changsha, Hunan, China.
service for the prevention of potential catastrophic 3
Department of Mechanical Engineering, University of Houston,
events. Houston, TX, USA.
Usually, parametric modeling and identification of 4
Department of Civil Engineering, University of Southern California, Los
civil engineering structures present a challenging prob- Angeles, CA, USA.
lem, and considerable efforts and progress have been
Corresponding author:
made in this field in the last two decades. Usually, dam- B. Xu, College of Civil Engineering, Hunan University, Changsha 410082,
age detection for civil engineering structures has been Hunan, China
carried out as a typical system identification problem Email: binxu@[Link]
282 Structural Health Monitoring 11(3)

carried out by determining structural stiffness values the results of structural testing. Additional applica-
and comparing them with previously determined refer- tions of ANNs in structural engineering include the
ence values or originally intended values determined by modeling of a nonlinear behavior of structures and/or
design drawings. Structural identification and damage structural components. For example, it is difficult to
detection methodologies can be carried out in fre- establish an accurate mathematical and parametric
quency or time domains, in parametric or nonpara- model for the widely used magnetorheological (MR)
metric forms, and in deterministic or in stochastic dampers for structural vibration control because there
ways, and the most widely employed mathematical are too many parameters to be involved and identified
model–based structural parameter identification algo- in the parametric model. The ANN shows its advan-
rithms developed in the last two decades are based on tages in solving these kinds of complicated problems.
modal parameters, namely, the natural frequencies, Chang and Roschke11 presented an alternative repre-
mode shapes, and their derivations.1 A comprehensive sentation of the dynamic performance of an MR
review of the vibration-based identification methodolo- damper in terms of a multilayer perceptron neural
gies can be found in several references.2–4 network. Zapateiro et al.12 also presented a modeling
Even the concept of the modal parameters–based approach for a shear-mode MR damper with ANN.
identification approaches is sound, it has been widely ANN also shows their ability in structural para-
recognized in the past years that natural frequencies meters identification and damage detection. Worden
that can be extracted relatively accurately from et al., Nakamura et al., Chen et al. and Yoshimura
dynamic measurements are usually insensitive to local et al. employed an ANN-based nonparametric
damages in engineering structures such as the cracks in approach for the detection of changes in the character-
reinforced structures and the loosening of connections istics of unknown structure systems.13–16 A study by
of composite members or structures, and mode shapes Masri et al.17 demonstrated that ANNs are powerful
extracted from traditional sensors may be too noise- tools for the identification of nonlinear systems typi-
corrupted to identify low to intermediate levels of local cally encountered in the structural dynamics field.
damage. Kosmatopoulos et al.18 employed ANNs to realize
Besides the modal parameters–based identification online tracking of nonlinear restoring forces in non-
algorithms, several time domain algorithms such as the linear structures. The basic idea of the above studies is
recursive least squares method and the sequential pre- to describe the nonlinear behavior of the system in a
diction error method have been developed for structural nonparametric way, and ANN shows its advantages in
identification. Moreover, considerable efforts and prog- handling the modeling problems for nonlinear complex
ress have been made in the modeling and identification systems. In fact, on the other hand, ANN has been
of engineering structures in a nonparametric way. As a employed to identify a dynamic system in a parametric
recognized comprehensive soft computation algorithm, way as an efficient inverse problem solver. In the past
artificial neural networks (ANNs) have drawn consider- years, great efforts have been made to enhance the
able attentions in the civil engineering and have been ability of ANNs to handle parametric identification
employed to model the behavior of both linear and problems. Most ANN-based structural identification
nonlinear materials, structural members or compo- approaches employ structural modal parameters
nents and even structural systems, due mainly to their including frequencies, modal shapes, and their deriva-
ability to approximate any arbitrary continuous func- tives to identify structural physical parameters by using
tion and mapping, and their parallel computation ANN as an optimization solver to describe the relation-
character. For example, ANN provides an alternative ship between modal parameters and structural physical
methodology for the modeling of the constitutive laws parameters. For example, Yun and Bahng19 presented
of engineering materials such as confined concrete a method for estimating the stiffness parameters of a
and composites. Lefik et al.5,6 proposed a modeling substructure in a complex structure by using an ANN
method for composites with ANNs. For reinforced with mode shapes and frequencies as input. Lee and
concrete (RC) members, Oreta and Kawashima7 Yun20 presented an effective method for damage esti-
established an ANN-based confined compressive mation of steel girder bridges using modal parameters
strength and strain model for circular concrete col- identified from the ambient vibration data using the
umns. Inel8 proposed an ANN model to describe the frequency-domain decomposition technique as input to
ultimate deformation capacity of RC columns. a back-propagation neural network (BPNN).
Mansour et al.9 proposed a shear strength prediction Considerably different from the above ANN-based
method for RC beams using ANNs. Yun et al.10 pre- frequency-domain identification and damage detection
sented an improved self-learning simulation approach approaches, Xu et al.21,22 presented a novel time
using a new ANN-based hysteretic material model to domain parametric identification methodology using
capture the cyclic behavior of the connections from both structural displacement and velocity measurement
Xu et al. 283

time series without any eigenvalues and eigenmodes Parametric identification method with
extraction processes. Moreover, the above methodol- displacement measurement time series
ogy has been developed and extended for large-scale
and complex structures by Xu and Wu23 and Wu et Identification methodology
al.24 to carry out decentralized and localized identifi- Consider an existing linear, viscously damped structure
cation using simulated dynamic response measure- whose parameters are to be identified. The information
ments of substructures with ANN. In recent years, employed to identify the structure in this study includes
with the great progress in sensing techniques, struc- the displacement response and the corresponding exter-
tural strain, stress, and even force under dynamic nal excitation.
excitation can be measured with acceptable accuracy. To facilitate the whole parametric identification pro-
For example, as one of the attractive advanced sen- cess, a reference structure and a number of associate
sing techniques for strain measurements, fiber optic structures that have the same overall dimension and
sensors are capable of providing distributed strain topology or numerical model as the object structure
measurements.25,26 The rapid development of these but different assumed structural parameters are con-
sensing techniques for different physical parameters structed. Two neural networks, which are constructed
necessitates the development of a new structural iden- and trained in advance, are employed to identify and to
tification methodology, which employs the new type detect damages of the object structure with the sole use
of measurements rather than the traditional accelera- of displacement measurement and excitation informa-
tion measurements. Xu et al.27 proposed a novel iden- tion. Because any online training for the two neural
tification strategy for an engineering truss structure networks and any optimization computations are unne-
model using vibration-induced strain time series. The cessary in the whole identification process, the identifi-
noncontact laser displacement sensors provide reli- cation and damage detection can be carried out
able measurement of vibration displacement response straightforwardly with the direct use of displacement
of engineering structures. measurements time series. This proposed approach is
More often than not, most of the current structural different from any of the current available eigenvalues-
identification approaches have been only tested by or eigenvectors-based identification and damage detec-
numerical simulation but have not been really tested in tion approaches where eigenvalues and/or eigenvectors
a more realistic way. Different from the previous stud- should be extracted first before identifying structural
ies by Xu et al. where the identification approaches parameters by solving an inverse problem. The two
were based on both velocity and displacement time neural networks play important roles in the whole iden-
series and the performance of the approaches was vali- tification procedure and their architectures and techni-
dated with numerical simulation only,21,22 in this study, cal backgrounds are described theoretically as follows.
a neural network-based structural parameter identifica-
tion and damage detection method with the direct and
sole use of displacement measurements from laser dis- Nonparametric modeling for reference structure with
placement sensors and excitation information is pro- displacement-based neural network emulator
posed and the performance of the approach is validated
(DNNE)
with dynamic test measurements of a frame structure
model. Because of different information used in the In most optimization algorithms for inverse analysis,
approach, the technical background of the construction the search procedure for the optimal solution starts at a
of the displacement-based emulator neural network certain point within the predetermined solution space
(DENN), which plays key roles in the proposed that may be determined by the preliminary estimation
method, is explained based on the discrete solution of on the region of the solutions of the problem. The refer-
the structural vibration equation. Then, the perfor- ence structure employed here plays a similar role as that
mance of the proposed methodology is examined with of the starting search point in any optimization algo-
the test measurements of a frame structure model on a rithms. In other words, the reference structure is just
shaking table. The interstory stiffness parameters of the like a baseline model in any traditional structural iden-
model structure in intact condition and damaged states tification and damage detection problems. Please note
with different levels of loosening damages in joints are that the reference structure does not necessarily have
identified using laser displacement measurements and the same structural parameters of the object structure
compared with the results from a traditional frequency- in a healthy condition. According to the design drawing
domain identification method. The results show that and the estimation of the material parameters of the
the proposed methodology can identify the interstory object structure, a reference structure can be assumed
stiffness of the frame structure in both healthy and and the corresponding finite element model (FEM) can
damaged states with acceptable accuracy. be determined. The structural parameters in terms of
284 Structural Health Monitoring 11(3)

stiffness in this FEM model will be identified using the dividing the difference between the displacement
displacement measurements of the object structure to responses xk1 and xk2 by the sampling period of T.
describe damages. Therefore, according to Equation (6), it is easy to
Considering the reference structure that can be mod- understand that the displacement response xk is fully
eled as a linear viscously damped structure with N determined by xk1 , xk2 , and the excitation force
degrees of freedom (N-DOFs), the equation of motion €xg, k1 . Then, a neural network can be constructed and
under base excitation can be characterized by the fol- be employed to describe the mapping from xk1 , xk2 ,
lowing equation: and €xg, k1 to xk straightforwardly in a nonparametric
form. The employed DNNE uses xk1 , xk2 , and €xg, k1
M€x + C x_ + Kx =  MI€xg , ð1Þ as input and xk as output and acts as a nonparametric
model of the reference structure.
where the matrices M, C, and K 2 RN 3N , are the struc- The dynamic displacement response of the reference
ture mass, damping, and stiffness matrices, respectively; structure under the measured excitation is simulated
N is the number of DOFs; €x, x_ , and x 2 RN are the with any available numerical integration algorithms
acceleration, velocity, and displacement vectors, respec- and is used to train the DNNE to map the function
tively; I 2 RN is a unitary vector, and €xg is the excita- from xk1 , xk2 , and €xg, k1 to xk of the reference struc-
tion acceleration. ture. After being trained, the DENN is employed to
Equation (1) can be rewritten in the following first- forecast the displacement response of the reference
order vector differential equation: structure step by step according to the displacement
response at two previous consequent time steps and
Z_ = AZ + B€xg , ð2Þ external excitation acceleration information at the pre-
where the state vector Z and the system matrices A and vious time step as described in the following formula.
B are defined as follows:  
xfk = fDNNE xk2 , xk1 , €xg, k1 ðk = 2, . . . , K Þ, ð9Þ
     
x 0 I 0 where xfk is the forecasted displacement response of the
Z= , A= , B= :
x_ M 1 K M 1 C I reference structure at time kT, K is the sampling num-
ð3  5Þ ber of the dynamic response time series. Even though
ANN has been proposed to model and detect damage
The discrete time solution of the state equation can be
for structures in a nonparametric pattern and in a qua-
written as follows:
litative manner,28,29 the main motivation and objective
of this study is to develop an identification methodol-
Zk = PZk1 + Q€xg, k1 ðk = 1, . . . , K Þ ð6Þ
ogy with neural networks solely and directly using the
X
‘
(ADt)i displacement measurement time series, which can be
P = exp (ADt) = ð7Þ carried out in a quantitative way instead of in a qualita-
i=0
i!
tive way. For this purpose, another neural network
ðt
D called as parametric evaluation neural network
Q= exp (A  t)dt  B = A1 (P  I)B, ð8Þ (PENN) is introduced and plays a key role in the para-
metric identification and damage detection.
0

where T is time interval, Zk and Zk 21 are the state vari-

PENN for parameter identification and damage
ables at kT and (k21)T, respectively.
Equation (6) shows that the state variable at kT, Zk , detection using displacement measurements
is uniquely determined by the state vector Zk1 and The function of the PENN is to describe the mapping
excitation force at (k21)T, which is determined by the from a defined evaluation index, which is determined
base excitation €xg, k1 in this study. In other words, the by the displacement response measurement of the
mapping from the state variable Zk1 and excitation object structure with the help of the above-trained
acceleration €xg, k1 to state variable Zk is unique. DNNE, to the parameters of the object structure. In
Equation (6) is a parametric form describing the above order to train the PENN, a number of training data
mapping. Because only the displacement response of should be prepared.
the structure is employed for identification in this By changing the parameters of the FEM of the
study, the relationship between displacement responses above constructed reference structure, a number of
at consequent time steps should be clarified. From the associate structures that have structural physical para-
definition of velocity, it is easy to understand that the meters different from those of the reference structure
velocity response x_ k1 can be determined completely by are considered. The dynamic displacement response of
Xu et al. 285

the mth associate structure, xm, k ðm = 1, . . . , M; usually treated as an inverse problem and is solved
k = 2, . . . , KÞ, under the same initial conditions that is with any of the available optimization approaches.
usually assumed to be zero and with the measured In contrast to the widely employed eigenvalues- and/or
external excitations, can also be determined by any eigenvectors-based traditional identification approaches,
numerical integration approach. in this study, the identification approach is implemented
In order to define an evaluation index, the following with the support of two neural networks.
operation is carried out. For a certain associate struc- In order to identify the structural parameters of the
ture such as the mth associate structure, treating its dis- object structure for which the displacement response
placement response vectors, xm, k2 and xm, k1 , and base measurements are known, the PENN is constructed
excitation, €xg, k1 ðk = 2, . . . , K Þ, as the input of the and trained to describe the mapping between the eva-
above-trained DNNE for the reference structure, a luation index and structural parameters as shown in
forecast displacement response vector the following equation:
xfm, k ðk = 2, . . . , K Þ at time kT can be determined
according to the following equation: ðMm , Km , Cm Þ = fPENN ðRMSPDDVm Þ: ð13Þ
  Corresponding to each associated structure, an eva-
xfm, k = fDNNE xm, k2 , xm, k1 , €xg, k1
luation index (vector) can be determined. Therefore,
ðm = 1, . . . , M; k = 2, . . . , K Þ ð10Þ the data composed of the structural parameters of the
M associate structures and the corresponding
As the structural parameters of the mth associate
RMSPDDVs can be determined. Based on the data,
structure differ from those of the reference structure,
the PENN is trained to describe the mapping from the
the dynamic displacement response of the associate
RMSPDDVs to the corresponding structural para-
structure determined by numerical integration does not
meters and is employed to identify the structural para-
correspond any more to the output of the DNNE.
meters of the object structure using the following
Therefore, a difference vector can be defined, in which
approach.
the component corresponding to the jth DOF of the
mth associate structure is determined by the following
equation: Parametric identification and damage detection using
displacement measurements
e(j) (j)f (j)
m, k = xm, k  xm, k
The proposed parameter identification and damage
ðj = 1, . . . N ; m = 1, . . . , M; k = 2, . . . , K Þ: ð11Þ detection approach is described in detail in Figure 1.
The DNNE and PENN are at the core of the proposed
The difference vector defined above should be fully
approach. The whole approach is implemented in three
determined by the parameters of the mth associate
phases. In Phase 1, a reference structure is assumed,
structure and be a function of the structural parameters
and the DNNE is constructed and trained with the
of the mth associate structure when the external excita-
simulated displacement response time series under the
tion is identical. Therefore, an evaluation index called
measured base excitation acceleration. In Phase 2, a
the root mean square of the prediction displacement
number of associate structures are assumed and the
difference vector (RMSPDDV) whose component cor-
corresponding displacement is determined by simula-
responding to the jth DOF of the mth associate struc-
tion. Then the simulated displacement is used as input
ture can be defined as follows:
to the above-trained DNNE, and the corresponding
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi output is compared with the simulated results and the
u
u X K
corresponding RMSPDDV is determined. Therefore,
(j) t 1
2
RMSPDDVm = e(j)
K  2 k =2 m, k the PENN is trained with the RMSPDDV of each asso-
ðm = 1, . . . , M; j = 1, . . . , N Þ: ð12Þ ciate structure and the corresponding structural para-
meters. The DNNE and PENN are constructed and
Furthermore, the evaluation index RMSPDDV cor- trained in advance in Phases 1 and 2, respectively, so
responding to the mth associate structure is dependent that the parameter identification in Phase 3 is a direct
on its structural parameters including mass, stiffness, process with the help of the trained DNNE and PENN
and damping coefficients of Mm, Km, and Cm. The using the displacement measurement and excitation as
objective of any traditional time domain structural input directly.
identification approaches is to identify structural para- In practice, in Phase 3, the RMSPDDV for the
meters in order to minimize the difference between the object structure can be determined by the output of the
dynamic measurement and the forecast values by the DNNE when the displacement measurements are used
mathematical model. This minimization problem is as input to the DNNE and the displacement
286 Structural Health Monitoring 11(3)

Construction of a reference structure

1-1
Displacementresponse of the reference
structureunder the sameinitial condition
and excitation as the object structure Known information: Displacement response of the
by numericalintegration object healthy structure or damaged structure under
1-2 measured base excitation
Generation of training patterns for
DNNE 3-1
1-3
1-4 Forecasted
Construction and training of DNNE 3-2 displacement
Trained DNNE response of the
object structure
3-3
2-2
2-3
Displacementresponse of the m-th Determination of the
Forecasted displacement Evaluation index RMSPDDV
associate structureunder the same
response of the m-th
initial condition and excitation as the
associate structure
object structure by numerical 3-4
integration(m=1,…,M) 2-4
2-1
Evaluation index 2-6 Construction and 2-7
Construction of M associate structures RMSPDDV
training of PENN
Trained PENN

2-5 Structural parameters of

3-5
the m-th associate
structure(m=1,…,M)
Structural parameters of
Training of NNE in advance the object structure or
damaged structure
Training of PENN in advance
Identification with the support of NNE and PENN

Figure 1. Displacement-based parametric identification and damage detection strategy with neural networks.

measurements. Then, the parameters of the object Experimental verification for a frame
structure can be identified by inputting the correspond- structure model with displacement
ing RMSPDDV to the PENN trained in Phase 2. measurements from laser sensors
When damage occurs in the model structure and the
displacement response of the damaged structure under Description of frame structure and shaking table test
the identical base excitation are available, the A shaking table test for a two-story steel frame struc-
RMSPDDV of the damaged structure can also be ture model to be identified is conducted to verify the
determined in a similar way by the trained DNNE. performance of the proposed structural identification
Then, the parameters of the damaged structure can be methodology. The frame structure model on the shak-
forecasted from the well-trained PENN with the corre- ing table is illustrated in Figure 2. Because it is easy to
sponding RMSPDDV as input. determine the mass of a structure, which usually does
In order to help understand the whole procedure, the not change with the occurrence of damage, the mass of
calculation process in each phase is described in detail a structure is considered as a known constant here. The
with two numbers in the form of ‘a and b’, where ‘a’ is mass of the first and the second floors are 1.16 and
the phase number and ‘b’ is the number of the steps in 2.03 kg, respectively.30 The thickness of the two col-
the corresponding phase. It is clear that no eigenvalues umns is 1.78 mm. The dimension of the model struc-
or modal shapes and their deviations are needed to be ture and the corresponding two-DOF mass-spring-
extracted from the dynamic response measurements, dashpot numerical model are shown in Figure 3. The
and no direct optimization computation is employed in interstory height of each story is 483 mm. The damp-
the identification procedure. Therefore, the proposed ing ratio of both control frequencies determined by the
approach provides a different inverse analysis method free vibration test of the structure is about 0.4%, which
for identification and damage detection. The character represents very light damping. Because the damping
of the proposed approach is very helpful in online iden- ratio of the model structure is very small and has been
tification and damage detection because no eigenvalues determined by a free vibration test, in order to simplify
and eigenvector extraction and optimization calculation the problem here, the structural parameters to be iden-
for inverse problem are needed. tified in this study are focused on the interstory
Xu et al. 287

Power Spectrum Magnitude/dB

40
X: 1.419 X: 4.333
Y: 39.37 Y: 21.62
20

-20
Laser
displacement
sensors -40

-60
0 2 4 6 8 10
Frequency/Hz

Figure 4. Power spectrum of the displacement response of the

object structure.

Shaking table methodology, a sweep sinusoidal excitation test is also

Figure 2. Shaking table test. carried out in order to extract the natural frequencies
of the structure. Two natural frequencies, 1.419 and
4.333 Hz, of the object structure in a healthy condition
are acquired from the sweep sinusoid excitation test as
305 shown in Figure 4.
12.5

Architecture and training of DNNE

As described above, for the purpose of identification
470

m2 for the object structure, a reference structure with

assumed structural parameters should be constructed.
978

k2 ,c2 In practice, the structural parameters including mass,

stiffness, and damping coefficients can be determined
25

by the preliminary estimation on the structural material

m1
parameters as described in the design drawings. In this
470

study, according to the basic information of the materi-

k1, c1 als and the dimension of the structure, a reference
in: mm structure is constructed, whose interstory stiffness para-
meters for the first and second floors are assumed to be
Figure 3. Model dimension and computational model. 300 and 300 N/m, respectively. The displacement
response of the first and second floors under the mea-
sured base excitation is determined by Newmark-Beta
stiffness parameters, which are employed to describe integration method. The integration time step is
the condition of the frame structure. 0.001 s. Taking 3 s of displacement response of the
The shaking table is excited by a sinusoidal wave reference structure from time 6.0–9.0 s as the training
with a frequency of 2.5 Hz and an amplitude of 0.2 g. data sets, a typical three-layer neural network, as
The displacement measurement of the first and second shown in Figure 5, is constructed and trained as a non-
floors of the frame structure model and the shaking parametric model of the reference structure. For this
table are synchronously acquired by three laser displa- two-story frame structure model, the number of neu-
cement sensors with a sampling rate of 1000 Hz. rons in the input and output layers of the DNNE is set
Because of the space limitation, the displacement to be 5 and 2, respectively. The input information
response time series of the object structure under har- includes the displacement response of the reference
monic excitation employed for identification and dam- structure at (k 2 2)T and (k 2 1)T and the excitation
age detection are not shown here. In order to eliminate acceleration at (k 2 1)T, and the output represents its
the effect of measurement noise, the displacement mea- displacement response at kT. The number of neurons
surements are band-pass filtered before identification. in the hidden layer is 6, determined by trial and error.
In order to compare the results from the proposed The whole off-line training process takes 3000 epochs.
288 Structural Health Monitoring 11(3)

ðk1 , . . . , kn , . . . , kN Þm = fPENN ðRMSPDDVm Þ

Displacement ðn = 1, . . . , N Þ, ð14Þ
response at
time (k-2)T
Displacement where kn is the interstory stiffness of the nth story of
response at time
Displacement kT
mth associate structure.
response at To generate training patterns for the PENN, a num-
time (k-1)T
ber of associate structures with different interstory stiff-
Excitation ness from the reference structure are considered. In this
acceleration at
time (k-1)T study, suppose the variation of the interstory stiffness
of the object structure is within 680% of the values of
Figure 5. Architecture of DNNE. the reference structure, then it is a relatively wide
region. In fact, there are infinite combinations of the
interstory stiffness within this space. In general, the
Figure 6 gives the comparison between the displace- higher the accuracy of the mapping (or the curve)
ment responses determined by Newmark-Beta method between the interstory stiffness with the corresponding
and those predicted by the trained DNNE. The relative RMSPDDV is the more combinations should be con-
RMS errors between the two curves corresponding to sidered. For simplicity, let the interstory stiffness of the
the first and second floors are 1.03% and 0.85%, first and second floors take some certain discrete values
respectively. It can be seen that the maximum relative within the interested space. Here, let the interstory stiff-
RMS error can reach a very small value and the trained ness take the discrete values from 120 to 480 N/m with
DNNE can be treated as a nonparametric model of the an increment of 30 N/m and then a total of 169 combi-
reference structure to forecast the displacement nations can be constructed. In order to improve the
response for the reference structure according to displa- training efficiency, not all of the combinations are
cement response at two previous consequent time steps employed to prepare the training data for PENN. Here,
and the base excitation acceleration at the previous 121 combinations among them are selected randomly
time step with high accuracy. from the above 169 combinations as the structural
parameters of the associate structures. The displace-
ment response of each associate structure under the
Architecture and training of the PENN measured base excitation is determined by Newmark-
The mass distribution and the damping ratio of the Beta integration method. Using the displacement
model structure assumed to be known in this study response of each associate structure from 6.0 to 9.0 s as
and the structural parameters to be identified are input to the above well-trained DNNE, totally 121
the interstory stiffness; therefore, the evaluation index RMSPDDVs can be obtained and employed as training
completely depends on the stiffness parameters. patterns for the PENN.
So Equation (13) can be rewritten in the following Figure 7 shows the architecture of the PENN. It con-
form: tains four layers. The number of neurons in the input

6
Forecasted Target 6
Forecasted Target
4
Displacement / mm

4
Displacement / mm

2
2
0
0
-2
-2
-4
-4
-6
6 6.5 7 7.5 8 8.5 9 6 7 8 9
Time / s Time / s

(a) 1st floor (b) 2nd floor

Figure 6. Comparisons between the displacement responses of the reference structure determined by Newmark-Beta method and
those forecasted by the trained DNNE.
Xu et al. 289

Table 1. Comparisons of the identified interstory stiffness of

healthy object structure

Identified by Eigenvalues Relative

RMSPDDV Interstorey PENN (N/m) based (N/m) error (%)
Corresponding stiffness of
to each each floor 1st floor 391.9 383.6 2.16
DOF 2nd floor 373.3 361.7 3.21

8
Figure 7. Architecture of PENN. Measured Calculated
6

Displacement / mm
2
600
Forecasted Target 0
1st floor /N/m

400 -2

-4
200 -6

0 20 40 60 80 100 120 -8
0 5 10 15 20 25 30 35
Time / s
600
Forecasted Target (a) 1st floor
2nd floor /N/m

8
400 Measured Calculated
6

200 4
Displacement / mm

0 20 40 60 80 100 120 2

0
Figure 8. Comparisons between the target stiffness and the
-2
output of the trained PENN.
-4

-6
and output layers is set to be 2 and 2, respectively, and
-8
8 for the two hidden layers determined by trial and 0 5 10 15 20 25 30 35
Time / s
error. The input represents the RMSPDDV compo-
nents corresponding to each DOF and the output (b) 2nd floor

means the identified interstory stiffness of the structure. Figure 9. Comparison of displacement measurement with
The entire off-line training process for the PENN takes those determined by numerical integration with the identified
5000 epochs. Figure 8 gives the comparisons between parameters.
the target stiffness and the output of the trained PENN.
The relative RMS errors corresponding to the inters- object structure are identified. Table 1 shows the iden-
tory stiffness of the first and second floors are 4.93% tified interstory stiffness of the healthy frame struc-
and 3.78%, respectively, for all the 121 training pat- ture and the comparison with the results from an
terns. The results show that the trained PENN can eigenvalue-based identification method. The relative
describe the function between the RMSPDDV and the error is defined as the ratio of the difference between
corresponding interstory stiffness with acceptable the identified interstory stiffness by the proposed
accuracy. approach and them by the frequency-based identifica-
tion method divided by the latter. It demonstrates
Parameter identification of the healthy object that the interstory stiffness of the healthy frame struc-
ture model can be identified with acceptable accuracy
structure using displacement measurements using the proposed methodology. Comparison of the
Using 3 s (from 6.0 to 9.0 s) of the displacement test displacement measurements with those recon-
measurement from laser displacement sensors of the structed by numerical integration with the identified
object structure in a healthy condition as input to the parameters is shown in Figure 9. It is clear that the
above-trained DNNE, the corresponding RMSPDDV reconstructed displacement responses of the object
is determined. Then, inputting the RMSPDDV to the structure agree well with the measurements from the
trained PENN, the stiffness parameters of the healthy laser displacement sensors.
290 Structural Health Monitoring 11(3)

3
2

Loosened
1 screws

(a) Joints location and their numbers (b) Damage case No.1

Location of 1st floor

loosened
Loosened screws
Screws
Shaking table

1st floor

(c) Damage case No.2 (d) Damage case No.3

Figure 10. Damaged structures.

Identification for damaged structures with Damage Case 2: Screws on the top end of the columns
joint connection loosening of the first floor and the bottom end of the columns of
the second floor (Locations 2 and 3) are loosened, as
In the frame structure, the two ends of each column are shown in Figure 10(c).
connected to the floors or base with three screws, as Damage Case 3: Screws on the top and bottom ends of
shown in Figure 10. In this study, local damage was the column of the first floor and the bottom end of the
introduced by loosening connection screws of some columns of the second floor (Locations 1, 2, and 3) are
joints. In order to indicate the location of joint connec- loosened, as shown in Figure 10(d).
tion damage, the numbers 1, 2, 3, and 4 are employed
as shown in Figure 10(a). The following three damage
scenarios are considered here. Shaking table tests on the above three damaged frame
structures are carried out under an identical base exci-
tation. The displacement measurements of the first and
Damage Case 1: Screws on the top end of the column second floors are synchronously acquired by laser dis-
of the first floor (Location 2) are loosened, as shown in placement sensors. With the above-mentioned identifi-
Figure 10(b). cation approach, the interstory stiffness parameters of
Xu et al. 291

Table 2. Natural frequencies of the three damaged structures measurements time series with neural networks is pro-
posed. The technical background and the implementa-
Damaged structures Natural frequency (Hz) tion of the proposed methodology are explained, and the
No. 1 1st order 1.144 architecture of the proposed DNNE and PENN are
2nd order 4.086 described. The performance of the proposed strategy is
No. 2 1st order 1.068 evaluated experimentally using displacement measure-
2nd order 3.326 ments from laser displacement sensors with a frame
No. 3 1st order 0.992 structure in a healthy and damaged condition under base
2nd order 3.204
excitation.
Interstory stiffness identification results for the
model structure show that the proposed methodology
Table 3. Comparisons of the identified interstory stiffness of can identify the interstory stiffness of the frame struc-
damaged structures ture with an acceptable accuracy when displacement
response time series are available. Identified results for
Identified Eigenvalues Relative
different damage scenarios also indicate that the
by PENN based error (%)
(N/m) identification method will not only identify the location well but also
(N/m) the extent of the damage. The identified stiffness results
agree well with them determined by the traditional
No. 1 1st floor 210.2 202.6 3.75 frequency-based identification method.
2nd floor 387.3 395.7 22.12
No. 2 1st floor 210.9 206.5 2.13
In contrast to the traditional modal parameters–
2nd floor 220.7 224.2 21.56 based identification methods, the strategy proposed in
No. 3 1st floor 175.3 167.4 4.72 this study does not require the extraction of structural
2nd floor 212.5 221.3 23.98 dynamic characteristics such as frequencies and mode
shapes from measurements. Thus, the damage detection
approach under discussion possesses some unique merits
the damaged structures are identified using 3 s of the
that other traditional algorithms using eigenvalues and
displacement measurements of the damaged structures
mode shapes do not have. Most of the vibration-based
from 6.0 to 9.0 s. Table 2 shows the frequencies of the
identification approaches require optimization for
three damaged structures. It is clear that the two fre-
inverse analysis, and it is hard to implement near-real-
quencies of the damaged structures have a decrease.
time or online identification. Because the DENN and
Table 3 gives the comparisons of the interstory stiffness
PENN with parallel computation characters have been
of the three damaged structures between those identi-
trained in advance, the identification process is carried
fied by the proposed approach and those determined
out with the direct use of several seconds of structural
with eigenvalues-based method. The results show that
dynamic response and thus provides a near-real-time
the interstory stiffness of the damaged frame structure
approach that can be used in online structural health
decreases with the loosening of the connection screws,
monitoring, as well as postearthquake damage detection.
and the proposed method gives acceptable damage
In this study, a different structural damage detection
detection results for the damaged frame structure. The
approach, namely, using noncontact laser sensors, is
relative errors are also listed in Table 3. Because the proposed. One of the limitations of the proposed
frame structure model has a strong floor and the stiff- method is that the excitation information is needed.
ness of the thin columns is relatively smaller, the struc- The approach should be extended when external excita-
ture model is a typical shear frame. The loosening of tion is unknown. Moreover, the model structure studied
screws leads to the decrease in the interstory stiffness of is relatively simple; further study on the performance of
the corresponding story. The identified results are rea- the proposed approach for large-scale structures with
sonable. The results show that, with the help of two noisy measurements will be carried out in the future.
neural networks of DNNE and PENN, the proposed Because a fixed reference structure is needed for the dis-
strategy can identify the structural stiffness, which can placement measurement by laser sensors (unlike accel-
represent the severity of the damage in the structure erometers) and the measurement range of a traditional
induced by the loosening of connections, using displa- laser sensor is usually limited, some advanced remote
cement response measurement time series. sensing techniques with enough measurement region,
sampling frequency, and accuracy such as digital image
processing techniques should be employed instead of
Concluding remarks
the laser sensors for the application of the proposed
In this article, a direct parametric identification and dam- approach in large-scale engineering structures in real-
age detection methodology solely using displacement life practice.
292 Structural Health Monitoring 11(3)

Acknowledgements 15. Chen HM, Qi GZ, Yang JCS and Amini F. Neural net-
The authors gratefully acknowledge the support provided by work for structural dynamic model identification. J Eng
the Program for New Century Excellent Talents in University Mech 1995; 121: 1377–1381.
(NCET-08-0178) and the Hunan Provincial National Natural 16. Yoshimura S, Yagawa G, Oishi A and Yamada K.
Science Foundation under grant no. 08JJ1009. Partial sup- Neural network based inverse analysis for defect identifi-
port from the NSFC grant no. 50978092 in conducting this cation with laser ultrasonics. Key Eng Mater 1998; 145–
research is also greatly appreciated. 149: 443–452.
17. Masri SF, Chassiakos AG and Caughey TK. Identifica-
tion of nonlinear dynamic systems using neural networks,
References ASME J Appl Mech 1993; 60: 123–133.
1. Doebling SW, Farrar CR and Prime MB. A summary 18. Kosmatopoulos E, Smyth A, Masri S and Chassiakos A.
review of vibration-based damage identification methods. Robust adaptive neural estimation of restoring forces in
Shock Vib Dig 1998; 30(2): 91–105. nonlinear structures. ASME J Appl Mech 2001; 68(6):
2. Chang FK. Structural health monitoring: Demands and 880–893.
challenges. New York, NY: CRC Press, 2001. 19. Yun CB and Bahng EY. Substructural identification
3. Inman DJ. Vibration with control. Chichester: Wiley, using neural networks. Comput Struct 2000; 77: 41–52.
2006. 20. Lee JJ and Yun CB. Damage diagnosis of steel girder
4. Wu ZS, Xu B and Harada T. Review on structural health bridges using ambient vibration data. Eng Struct 2006;
monitoring for infrastructure. JSCE J Appl Mech 2003; 6: 28: 912–925.
1043–1054. 21. Xu B, Wu ZS, Chen G and Yokoyama K. Direct identifi-
5. Lefik M, Boso DP and Schrefler BA. Artificial neural cation of structural parameters from dynamic responses
networks in numerical modelling of composites. Comput with neural networks. Eng Appl Artif Intell 2004; 17(8):
Meth Appl Mech Eng 2009; 198: 1785–1804. 931–943.
6. Lefik M and Wojciechowski M. Artificial neural network 22. Xu B, Wu ZS, Yokoyama K, Harada T and Chen G. A
as a numerical form of effective constitutive law for soft post-earthquake damage identification methodology
composites with parametrized and hierarchical micro- using vibration time series. Smart Mater Struct 2005;
structure. Comput Assist Mech Eng Sci 2005; 12: 183– 14(3): s116–s124.
194. 23. Xu B and Wu ZS. Localized and decentralized identifica-
7. Oreta AWC and Kawashima K. Neural network model- tion for large-scale structures. Comput Assist Mech Eng
ing of confined compressive strength and strain of circu- Sci 2007; 14: 361–378.
lar concrete columns. ASCE J Struct Eng 2003; 129(4): 24. Wu ZS, Xu B and Yokoyama K. Decentralized para-
554–561. metric damage based on neural networks. Comput Aided
8. Inel M. Modeling ultimate deformation capacity of RC Civ Infrastruct Eng 2002; 17: 175–184.
columns using artificial neural networks. Eng Struct 2007; 25. Ansari F. Fiber optic sensors for construction materials
29: 329–335. and bridges. In: Proceedings of the International Work-
9. Mansour MY, Dicleli M, Lee JY and Zhang J. Predicting shop on Fiber Optic Sensors for Construction Materials
the shear strength of reinforced concrete beams using arti- and Bridges, Lancaster, PA: Technomic Publishing, 1998.
ficial neural networks. Eng Struct 2004; 26: 781–799. 26. Conte JP and Liu M. Use of long-gage fiber optic sensors
10. Yun GJ, Ghaboussi J and Elnashai AS. Self-learning for earthquake response monitoring and nondestructive eva-
simulation method for inverse nonlinear modeling of cyc- luation of structures. CUREE publication no. CKIII-04,
lic behavior of connections. Comput Meth Appl Mech CUREE-Kajima Joint Research Program-Phase III, 2001.
Eng 2008; 197: 2836–2857. 27. Xu B, Chen G and Wu ZS. Parametric identification for
11. Chang CC and Roschke P. Neural network modeling of a truss structure using axial strain. Comput Aided Civ
a magnetorheological damper. J Intell Mater Syst Struct Infrastruct Eng 2007; 22: 34–46.
1998; 9: 755–764. 28. Chen HM, Qi GZ, Yang JCS and Amini F. Neural net-
12. Zapateiro M, Luo N, Taylor E and Dyke SJ. Modeling work for structural dynamic model identification. J Eng
and identification of a class of MR fluid foam dampers. Mech 1995; 121: 1377–1381.
Smart Structures and Systems 2010, 6(2): 101–113 29. Barai SV and Pandey PC. Time-delay neural networks in
13. Worden K. Structural fault detection using a novelty damage detection of railway bridges. Adv Eng Software
measure. J Sound Vib 1997; 201(1): 85–101. 1997; 28: 1–10.
14. Nakamura M, Masri SF, Chassiakos AG and Caughey 30. Huo LS, Song GB, Li HN and Grigoriadis K. HN
TK. A method for non-parametric damage detection robust control design of active structural vibration sup-
through the use of neural networks. Earthq Eng Struct pression using an active mass damper. Smart Mater
Dyn 1998; 27(6): 997–1010. Struct 2008; 17: 015021.