You are on page 1of 20

Bull Earthquake Eng (2017) 15:5393–5412

DOI 10.1007/s10518-017-0190-y

ORIGINAL RESEARCH PAPER

Damage assessment by stiffness identification for a full-


scale three-story steel moment resisting frame building
subjected to a sequence of earthquake excitations

Cong Zhou1 • J. Geoffrey Chase1 • Geoffrey W. Rodgers1 •

Chikara Iihoshi2

Received: 16 February 2017 / Accepted: 12 July 2017 / Published online: 17 July 2017
 Springer Science+Business Media B.V. 2017

Abstract This research experimentally investigates a practical hysteresis loop analysis


method for damage assessment through stiffness identification using a full scale three-story
steel moment resistant frame structure subjected to a sequence of 6 shake table tests with
different magnitudes of 3D excitations. The effective elastic stiffness, ke, is calculated
based on the analysis of the experimentally measured and reconstructed hysteresis loops.
Structural degradation is then evaluated by tracking changes in identified stiffness over
time and across different earthquake events. Finally, the fundamental frequency of the
building is calculated using the identified stiffness to compare with the experimental
frequency from experimental transfer functions. Results show average differences between
the final stiffness of one event and the initial stiffness of the following event are less than
5% in both horizontal directions, indicating a good continuity and accuracy of the iden-
tification over events. The average difference between the experimental and calculated
fundamental frequency are less than 0.1 Hz in both directions over all events. The overall
results clearly delineate the capability of the method to accurately and consistently
quantify and localize structural damage that may not be detected by external visual
detection over several earthquake events of different magnitude, indicating its ability to be
used in long term monitoring.

Keywords Damage assessment  Structural health monitoring  Stiffness


identification  E-defence shake table test  Hysteresis loop analysis

& Cong Zhou


cong.zhou@canterbury.ac.nz
1
Department of Mechanical Engineering, University of Canterbury,
Private Bag 4800, Christchurch 8140, New Zealand
2
Asahi Kasei Homes Corp., Tokyo, Japan

123
5394 Bull Earthquake Eng (2017) 15:5393–5412

1 Introduction

The main goal of structural health monitoring (SHM) is to create practical, efficient
identification methods to clearly delineate changes in physical or modal parameters for
damage assessment and localization in long term and immediate post-event monitoring. To
date, system identification has a significant gap between theory and practice in achieving
this outcome, despite small scale laboratory and complex real structure structures being
tested (Belleri et al. 2014; Çatbaş et al. 2013; Ikeda 2016). In particular, method com-
plexity and robustness, and/or significant data requirements prevent achieving this goal in
civil SHM.
Commonly used SHM approaches include parametric model-based methods and non-
parametric methods. Many model-based methods, such as adaptive least mean squares
(LMS) method (Chase et al. 2005a, b; Nayyerloo et al. 2011), extended Kalman filters
(EKF) (Huang et al. 2010; Lei et al. 2015; Pan et al. 2016; Yang et al. 2006) and unscented
Kalman filters (UKF) (Al-Hussein and Haldar 2015; Wu and Smyth 2007, 2008; Xie and
Feng 2012), identify changes in structural stiffness of selected baseline model parameters
to reflect the severity of seismic damage. These model-based methods can successfully
detect and quantify damage severity when the chosen baseline model is accurate repre-
sentation of response. However, there is a significant, but unknown, risk of a poor iden-
tification result when the chosen model does not match the dynamics of the actual
measured system response since the actual outcome is not fully known (Yao and Pakzad
2014; Zhou et al. 2017). Finite element model (FEM) updating is another class of model-
based method for SHM in civil structures. The FEM updating determines damage as a
change of structural properties by updating the structural parameters, such as mass, stiff-
ness and damping, to match dynamic characteristics between the FEM and experimental
test extraction. However, the dynamic testing is prone to uncertainties and significant
frequency changes that may be identified in theses ultra-low magnitude input tests due to
changes in ambient conditions, rather than the desired identification of the existences of
damage (Huang et al. 2012, 2015; Im et al. 2013)
Nonparametric approaches have also been extensively applied in the field of health
monitoring and system identification. These methods are based on mathematic models,
such as neural networks (NN), fuzzy wavelet NN and nonlinear autoregressive moving
average with exogenous (NARMAX) models, which may not have any explicit physical
meaning (Londoño et al. 2015). The structural system is trained to approximate an arbitrary
continuous function that represent a physical structure and predict the structural response,
which is particularly effective for large structures due to their complicated nonlinear
behaviour and the incomplete, incoherent, and noise-contaminated measurements of
structural response under extreme loading (Adeli and Jiang 2006). However, a poorly
trained model may lead to in accurate results when the training data is incomplete or
corrupted, and converging towards global optimum and convergence speed is not guar-
anteed with a limited set of train samples (Sirca and Adeli 2012). Finally, these non-
parametric methods might not be able to provide assessment on damage severity, since the
model approximation lacks of direct of physical interpretation (Chang and Shi 2010).
The hysteresis loop analysis (HLA) method at the center of this work offers some
significant advantages compared to the traditional model-based and non-parametric SHM
algorithms. In particular, no differential equations need to be solved to evaluate modelling
errors or optimise parameters (Cifuentes and Iwan 1989; Xu et al. 2014, 2015; Zhou et al.
2017). Hence, the HLA method is based on fundamental mechanics, but not necessarily
limited to a model structure.

123
Bull Earthquake Eng (2017) 15:5393–5412 5395

Reconstructed hysteresis loops were initially used as a rapid visual qualitative indicator
of system performance. Secant stiffness was deduced from the reconstructed hysteresis
loops to determine the occurrence of degradation and damage (Stephens and Yao 1987).
Based on new sensor technologies, a near real time SHM system using reconstructed
hysteresis loops called R-SHAPE was first applied at the Millikan Library Building (Iwan
2002; Iwan et al. 2013). The hysteresis loops for each floor were established in near real
time from data obtained during the earthquake, and the general shape of the reconstructed
hysteresis loops were examined quickly as an indicator of the type of structural behaviour.
However, they did not directly identify linear and nonlinear stiffness or behaviour, and thus
did not quantify or localize damage severity.
Therefore, a hysteresis loop analysis method based on a hypothesis test and an overall
multi-segment least square linear regression, was proposed for parameter identification of a
SDOF pinching system using reconstructed hysteresis loops (Zhou et al. 2015a, b). This
method showed the ability to track the variation of linear and nonlinear structural stiffness
values if degradation occurs during the earthquake and can be readily generalized to multi-
degree of freedom systems with different hysteretic behaviours, thus also localizing the
damage assessment.
SHM methods can be best validated numerically for their accuracy, where every value is
known. For this HLA method this validation has been undertaken for linear, and highly
nonlinear cases. In addition, they have also been tested numerically for accuracy and
robustness to different and variable hysteretic behaviours, as well as to noise in (Xu et al.
2014, 2015; Zhou et al. 2015a, 2017). This paper undertakes the necessary experimental
validation with comparison to design calculated values, and thus assessing robustness to
realistic conditions. Recently, a full-scale three storey steel moment resisting (SMRF)
structure was subjected to a sequence of earthquake excitations in the E-defence shake
table in Japan. It provides unique seismic data for a complex real-world structure with
realistic seismic response compared to smaller, scaled structures or simulated numerical
models. In this study, the HLA method is extended to experimentally examine its capability
to identify and track structural stiffness over time and different events for damage assess-
ment in long term monitoring. Only acceleration measurements, which can be easily
obtained with low cost accelerometers and distributed over the internet in real time if
necessary, were used as input information to this approach. Severity assessment and
localization of structural damage are achieved by tracking changes in elastic story stiffness
as a damage index over a sequence of six tests with different magnitude input events.

2 Test structure and E-defence shaking table tests

The test structure is a full-scale steel moment resisting frame (SMRF) building, as shown
in Fig. 1. The building on right hand has a couple of oil dampers connected to an upper
beam in the first story, and the building on the left hand is a normal building that is used in
this work. The structure has three stories and each story has a uniform height of 2870 mm.
The plan dimensions are 6405 mm in the x-direction (NS) and 7320 mm in the y-direction
(EW), respectively. Rigid connections are designed between the H-shaped steel beams and
cold-formed steel RHS columns. The seismic weights are 171.85, 171.85 and 90 kN for the
first, second and third story, respectively.
Shake table tests of the building were conducted at E-defence in Japan. Input ground
motions were applied in two horizontal (x, y) and vertical (z) directions. Three tri-axial

123
5396 Bull Earthquake Eng (2017) 15:5393–5412

Fig. 1 Photo of the E-defence test structure

accelerometers were installed on beams at each floor to record the structural response.
Figure 2 shows the locations and orientation of the accelerometers at each floor. Test
events include six 3D earthquake excitations applied sequentially with different magni-
tudes, as listed in Table 1, with #01 referring to the first test event.

3 Methods

3.1 Reconstruction of hysteresis loops

The equation of motion for a multi-degree-of-freedom inelastic structure subjected to


earthquake excitation can be expressed:
QðVðtÞÞ ¼ MI€ €  CðtÞVðtÞ
vg ðtÞ  MVðtÞ _ ð1Þ
_ and VðtÞ
where VðtÞ, VðtÞ € are the vectors of displacement, velocity and acceleration, M is
the constant mass matrix and C(t) is the time-varying Rayleigh damping matrix. QðVðtÞÞ is
Fig. 2 Accelerometers locations
at each floor, shown on a plan
view of the structure

123
Bull Earthquake Eng (2017) 15:5393–5412 5397

Table 1 Sequential shake


Test no. Input event PGA (g)
table tests in order given
y-direction x-direction Vertical
z-direction

#01 BSL2-18% 0.11 0.13 0.01


#02 Sannomal 0.22 0.16 0.01
#03 Uemachi 0.30 0.35 0.01
#04 Toshin-Seibu 0.62 0.63 0.06
#05 Sannomal 0.21 0.15 0.01
#06 Nankai- 0.87 0.74 0.03
Trough

the nonlinear time-varying restoring force vector determined by the time-varying structural
stiffness matrix K(t) and loading–unloading path.
In particular, the full scale E-defence test structure is a three-story SMRF building.
Thus, the Matrices of M, K(t), C(t), and QðVðtÞÞ are specifically defined:
2 3 2 3
m1 0 0 K1 ðtÞ þ K2 ðtÞ K2 ðtÞ 0
6 7 6 7
M ¼ 4 0 m2 0 5 KðtÞ ¼ 4 K 2 ðtÞ K2 ðtÞ þ K3 ðtÞ K3 ðtÞ 5
0 0 m3 0 K3 ðtÞ K3 ðtÞ
2 3 2 3 2 3
C11 C12 0 Q1 ðtÞ f1 ðtÞ  f2 ðtÞ
6 7 6 7 6 7
CðtÞ ¼ a0 M þ a1 KðtÞ ¼ 4 C21 C22 C23 5 Qs ðVðtÞÞ ¼ 4 Q2 ðtÞ 5 ¼ 4 f2 ðtÞ  f3 ðtÞ 5
0 C32 C33 Q3 ðtÞ f3 ðtÞ
ð2Þ
where m1, m2 and m3 are the mass for the first, second and third story, respectively; K1(t),
K2(t) and K3(t) are the time-varying story stiffnesses; a0 and a1 are mass-proportional and
stiffness-proportional damping coefficients, respectively; and f1(t), f2(t) and f3(t) are the
hysteretic restoring forces, which define Q1,2,3(t) the nonlinear restoring forces for each
story. The restoring forces, f1(t), f2(t) and f3(t), can be obtained:

f1 ðtÞ ¼ Q1 ðtÞ þ Q2 ðtÞ þ Q3 ðtÞ


f2 ðtÞ ¼ Q2 ðtÞ þ Q3 ðtÞ ð3Þ
f3 ðtÞ ¼ Q3 ðtÞ

The ground acceleration v€g and structural accelerations (€


v1 ,€v2 ,€
v3 ) are measured at three
diagonal points for each story during the test, as shown in Fig. 2. Figure 3a, b show
recorded accelerations of each point of the third (top) story in the x-direction and
y-direction, respectively, during test #01 (BSL2-18%). It is noted that the accelerometer
installed at the middle point AX1Y1 in the x-direction (€vxX1Y1 ) did not work during the test.
x
Thus, no accelerations were recorded for v€X1Y1 at the third story, as shown in Fig. 3a.
However, it can be clearly seen from Fig. 3b that the measurement at the corner AX0Y0
vyX0Y0 ) is larger than the measurement at the middle point AX1Y1 (€
(€ vyX1Y1 ) in the y-
y
direction and the measurement at the other corner AX2Y2 (€ vX2Y2 ) is smaller than AX1Y1

123
5398 Bull Earthquake Eng (2017) 15:5393–5412

Fig. 3 Acceleration comparison of the three diagonal points in the a x-direction, and b y-direction of the
third (top) story during the test #01 (BSL2-18% event)

vyX1Y1 ) in the y-direction. In addition, Fig. 3a also shows the measurement at AX0Y0
(€
vxX0Y0 ) is larger than AX2Y2 (€
(€ vxX2Y2 ) in the x-direction. These results indicate a clear
contribution of torsional components to the building response during the test. Figure 4
shows how this torsional response component affects the measured linear acceleration
response of the building.
Assuming the floor is rigid, and denoting are v€xrigid and v€yrigid in the x-direction and y-
direction, respectively, the measured response at the three diagonal points (AX0Y0,
AX1Y1 and AX2Y2) in the x-direction and y-direction can then be defined:

Fig. 4 Illustration of the effect of torsion on the measured acceleration response

123
Bull Earthquake Eng (2017) 15:5393–5412 5399

v€xX0Y0 ¼ v€xrigid þ R  sin h1  sin h2


ð4Þ
v€xX2Y2 ¼ v€xrigid  R  sin h1  sin h2

v€yX0Y0 ¼ v€yrigid þ R  sin h1  cos h2


v€yX1Y1 ¼ v€yrigid ð5Þ
v€yX2Y2 ¼ v€yrigid  R  sin h1  cos h2

Thus, the torsional effects on the measured responses can be eliminated using the
average responses of ð€ vxX0Y0 þ v€xX2Y2 Þ=2 in the x-direction and ð€ vyX0Y0 þ v€yX1Y1 þ v€yX2Y2 Þ=3
in the y-direction, respectively. This approach enables evaluation of hysteresis loops in
each lateral direction using the model of Eqs. (1)–(3), without adding a torsional evalua-
tion equation.
The velocities (v_1 , v_2 , v_3 ) and displacements (v1,v2,v3) are obtained by direct and double
integration of measured accelerations (€ v1 , v€2 , v€3 ) after a band pass filtering with a window
width of 0.8–15 Hz based on the fact that no residual drift is observed from measured inter-
story displacement. Hence, the ability to obtain measurements if only acceleration is
measured ensures this approach is possible with minimum added investment in cost,
computation or complexity, presuming minimal or small residual displacements. However,
in cases unlike the one in this research, where there is significant residual or permanent
deflection, an accurate estimation of displacement can also be obtained using a GPS low
frequency method or Multi-rate Kalman method given required displacement measurement
with or without residual displacement. (Hann et al. 2009; Smyth and Wu 2007; Xu et al.
2015; Zhou et al. 2015a). A 5% damping ratio is typically assumed for the first and second
mode to calculate the Rayleigh damping matrix. The effect of different damping ratio on
the identification result is also investigated. Thus, the hysteresis loops of the sth story in
both x-direction and y-direction are reconstructed using the calculated restoring force fs and
relative deformation (vs - vs-1).
Figures 5 and 6 show the resulting hysteresis loops of the three stories subjected to the
six sequential earthquake events in Table 1 in the x-direction and y-direction, respectively.
It can be seen that the hysteresis loops for tests #04 (Toshin-Seibu) and #06 (Nankai-
Trough) clearly show the strongest response during earthquake excitation, which corre-
sponds their large input PGA compared to the other events in Table 1.

3.2 Hysteresis loop analysis

The restoring forces are path dependent and multivalued functions of displacements in the
hysteresis loops. For instance, Fig. 7a shows the hysteresis loop of the first story of the
#01(BSL2-18%) earthquake in the x-direction. However, the hysteresis loops can be
divided into a number of half cycles using turning points where the force is a local
maximum or minimum, as shown in Fig. 7b.
Thus, the restoring force for a given story is a single valued function of displacement
within each half cycle (j) of the hysteresis loop for that story (s), denoted fjs(t). For each
selected half cycle, the force–deformation (f–v) relationship can be defined using an r-
segment piecewise regression model (Zhou et al. 2015a):

123
5400 Bull Earthquake Eng (2017) 15:5393–5412

Fig. 5 Reconstructed hysteresis loops using experimental measurements in the x-direction

Fig. 6 Reconstructed hysteresis loops using experimental measurements in the y-direction

123
Bull Earthquake Eng (2017) 15:5393–5412 5401

Fig. 7 Turing points of dividing half cycles for, a the hysteresis loop of the first story and b the time history
of restoring force of the first story in the x-direction during the BSL2-18% event. The significant half cycles
are denoted

fs1j ðiÞ ¼ k1 vðiÞ þ fb1 þ e1 ðiÞ vð1Þ  vðiÞ  vðt1 Þ


fs2j ðiÞ ¼ k2 vðiÞ þ fb2 þ e2 ðiÞ vðt1 Þ  vðiÞ  vðt2 Þ
ð6Þ

fsrj ðiÞ ¼ kr vðiÞ þ fbr þ er ðiÞ vðtr1 Þ  vðiÞ  vðnÞ

where n is the number of observations in this half cycle; e(i) is the disturbance caused by
model uncertainty and measurement noise, which allows different variance of errors and
different distribution of the data across segment (Bai and Perron 2003); (t1, t2,…, tr-1) are
the breakpoints that separate the half cycle into r-segments; (fb1, fb2…, fbr) are the intercept
force for each segment; and (k1, k2,…, kr) are the stiffness values or hysteresis loop slope
within each segment of the half cycle. If the optimal number of segments, r, for each half
cycle is obtained, the breakpoints, stiffness and intercept force can then be determined
using regression analysis by minimising overall sum squared residuals (SSR) (Zhou 2016).
In particular, a sup F-type hypothesis test between the null hypothesis of a linear half
cycle (one-segment) and the alternative hypothesis of a nonlinear half cycle (2, 3 or 4
segments) is defined (Bai and Perron 1998):
 
1 N  2r d0 S0 ðSðZZ 0 Þ1 S0 Þ1 Sd
Fðr ¼ 2; 3; 4j1Þ ¼ ð7Þ
N 2ðr  1Þ Rr
where N is the number of observation in the selected half cycle; Rr is the overall SSR under
the alternative hypothesis; Z = diag(Z1,…,Zr) with Zi ¼ ð ðvti1 þ1 1Þ ; . . .; ðvti 1 ÞÞ0 ;
d = (d1,…, dr)0 is the regression coefficient vector with di = (ki, fbi), and S is the con-
ventional matrix so that Sd = (d1 - d2, …, dr-1 - dr). If the maximum value of F(r = 2,
3, 4|1) is smaller than the predefined critical value of 16.79, the half cycle is optimized as a
linear regression model (one-segment) without breakpoint. Otherwise, a further F(r ? 1|r),
where r = 2 and 3, test is conducted to determine the number of breakpoints in the
nonlinear half cycle, which is defined:

123
5402 Bull Earthquake Eng (2017) 15:5393–5412

P 
r 0
n j¼1 R 1 ðT j Þ  R rþ1
Fðr þ 1jrÞ ¼ ð8Þ
Rr
( )
Xt1
2
X
t2
2
X
n
2
Rr ¼ min ½fs1 ðiÞ  k1 vðiÞ  fb1  þ ½fs2 ðiÞ  k2 vðiÞ  fb2  þ    þ ½fsr ðiÞ  kr vðiÞ  fbr 
i¼1 i¼t1 þ1 i¼tr1 þ1

ð9Þ

R0rþ1 ¼ min inf Rrþ1 ðvð1Þ; vðt1 Þ; . . .; vðtj1 Þ; vðsÞ; vðtj Þ; . . .; vðtr1 Þ; vðnÞÞ ð10Þ
1  j  r vðsÞ2Ki;g

 
Ki; g ¼ vðsÞ; vðtj1 Þ þ 0:05ðvðtj Þ  vðtj1 ÞÞ  vðsÞ  vðtj1 Þ  0:05ðvðtj Þ  vðtj1 ÞÞ
ð11Þ
where Rr is the overall SSR under the null hypothesis, R0rþ1 is the overall SSR under the
alternative hypothesis and R1(Tj) is the SSR at the jth segment.
In particular, if the F value of F(3|2) is smaller than the predefined critical value of
17.89, the half cycle is optimized as a 2-segment regression model. Otherwise, the
assumption is updated with the hypothesis test between 3-segment and 4-segment models
to obtain F(4|3). Again, the half cycle is optimized as a 3-segment regression model if
F(4|3) \ 18.86, or a 4-segment regression model if F(4|3) C 18.86.

3.3 Stiffness identification

Although the identified slope value for each segment of the regression model represents a
structural stiffness, this identified stiffness can be distorted due to measurement noise,
stiffness discontinuity of loading–unloading direction, and the effect of the smooth
response transition from elastic to plastic behaviour and thus between these segments
(Zhou 2016; Zhou et al. 2015a). A simple moving average (SMA) method is the most
common and easiest digital filter to smooth the identification result. However, the per-
formance of the SMA highly depends on the window size, and is less robust if high
frequency data is not removed (Smith 1999). Thus, a wavelet multiresolution analysis
(MRA) is applied to decompose a more accurate stiffness value from noisy identified slope
values. In particular, the time-varying vector of the identified slopes S(t) from each half
cycle are transformed into a series of time-invariant scale and wavelet coefficients using
multiresolution expansions (Mallat 1989):
X
1 X
J X
1
SðtÞ ¼ KðtÞ þ EðtÞ ¼ cJ;k /l;k ðtÞ þ dJ;k wj;k ðtÞ ð12Þ
k¼1 j¼1 k¼1

where cj,k and dj,k are the scaling and wavelet coefficients, respectively, / is an
orthonormal basis for the scaling function of the approximation space, and w is an
orthonormal basis of wavelet space related to the scaling function, /. Thus, the approxi-
mation terms interpret the inherent stiffness K(t) at a coarse scale, J, and the detail terms
characterize the discontinuity and uncertainty information E(t) at a finer scale j = 1,…, J.
The resolution scale, J, thus determines the occurrence or absence of any specific
phenomenon is determined based on statistical laws of the loglog type (Coca and Billings
2001; Ovanesova and Suarez 2004; Sjöberg et al. 1995; Wei and Billings 2002):

123
Bull Earthquake Eng (2017) 15:5393–5412 5403

N 2N
 2J  ð13Þ
ln N ln N
where N is the number of observations in the full space of the slopes and thus the number
of half cycle identified.
The application of MRA is not unique and depends on the choice of wavelet. In this
case, the biorthogonal wavelet, bior6.8, which is a compactly supported spline wavelet for
symmetry and exact reconstruction (Ovanesova and Suarez 2004), is chosen. The analysis
is done using Matlab Toolbox (The MathWorks, Inc., R2016a), and compared to the
identification result using the SMA method, as further validation.
Hence, the stiffnesses (k1, k2,…, kr) for each segment of each selected half cycle fitted
by the optimized regression models can be tracked as damage indices over time for damage
identification. For this steel structure, the effective pre-yielding stiffness, ke, is represented
by the slope of the first-segment (k1). The degradation of the pre-yielding stiffness, ke, is
used to identify the evolution of damage that leads to a decrease of natural frequency, w,
which is particular appropriate for largely linear steel structure responses as seen in Figs. 5
and 6 (Londoño et al. 2015; Rao and Gupta 1999). The flowchart of the identification
procedure is shown in Fig. 8.

3.4 Impact of changing damping force

The damping ratio changes over time and different types of structures. These variations in
damping makes many such problems less identifiable, especially since damping in real
structures is not necessarily purely viscous and thus difficult to identify to a specific model
dynamic. Thus, in most hysteresis methods, the damping force is included in the total
restoring force to calculate an effective stiffness using a simpler mathematical model
(Cifuentes and Iwan 1989; Iwan 2002; Iwan and Cifuentes 1986; Iwan et al. 2013; Toussi
and Yao 1983):
f ðv; vÞ vg þ v€Þ
_ ¼ m  ð€ ð14Þ
where f ðv; vÞ
_ is the total restoring force including both spring force and damping force. An
effective stiffness is then calculated from f ðv; vÞ=v
_ to indicate structural damage by
comparing with the initial effective stiffness.
In the proposed HLA approach of this research, the restoring force for the hysteresis
loop is calculated by assuming 5% Rayleigh damping ratio, as typically used by design
codes and standards for the calculation of classical damping matrices if similar material
and damping mechanisms are distributed over the height (Atkinson and Pierre 2004;
Chopra 1995; Pekcan et al. 1999). The effect of damping ratio on the identification
results is further investigated by assuming different damping ratios. In particular,
damping ratios of 3, 5, 7 and 10%, which cover a typical range of damping ratios for
common structures (Belleri et al. 2014; Chopra 1995), are adopted for the first two
modes to recalculate the evolution of the elastic stiffness, ke, for the second story in the
x-direction and the first story in the y-direction during test #04 where the greatest
stiffness degradation occurred.

123
5404 Bull Earthquake Eng (2017) 15:5393–5412

Fig. 8 Flowchart of the HLA identification procedure

4 Results and discussions

4.1 Identified effective pre-yielding stiffness ke

To evaluate the accuracy of the fitness between the identified regression model and
measurement data, the residual errors Ereg related of all the half cycles over 6 test events
are calculated (Draper and Smith 2014):
Pn 2
i¼1 ðyi  f ðiÞÞ
Ereg ¼ P n
ð15Þ
Þ2
i¼1 ðyi  y

where yi is the observation data, f(i) is the regression data and y is the mean of observed
data.
Figure 9 shows the cumulative distribution function (CDF) of residual errors between
the identified regression model and measurement data for all three stories (DOF1-DOF3)
under the 6 earthquake events in both x and y directions. Over 95% half cycles are fitted
with residual errors less than 5% for all the cases, which indicates the identified piecewise
regression models accurately capture the measured force–deformation relationship (hys-
teresis loop) of each story of the test structure under different earthquake excitations.

123
Bull Earthquake Eng (2017) 15:5393–5412 5405

Fig. 9 CDF of residual errors between the identified regression model and measurements data for DOF1-
DOF3 under 6 earthquake events in both the a x-direction and b y-direction. N is the number of cycles

Therefore, the slope of the first segment of the selected half cycle can be used to accurately
represent the effective elastic stiffness of the structure for damage assessment.
Figures 10 and 11 show the evolution of the identified elastic stiffness ke for each story
over the six different earthquake events of Table 1 in the x-direction and y-direction,
respectively. The solid line represents the results from SMA moving average of the slope
values, and the dashed line shows the results from MRA analysis. The MRA results show
smoother changes over time than the SMA method. In addition, the MRA analysis is
automatic without requiring human input. However, both methods yield very similar
overall drops of stiffness over different events.
No significant stiffness drops were identified for all three stories in both directions
during tests #01-#03, but much stronger tests #04 (Toshin-Seibu) and #06 (Nankai-Trough)
led to significant stiffness changes in DOF1-3. The stiffness drops of test #04 (Toshin-
Seibu) are clearly larger than test #06 (Nankai-Trough), which indicates a more severe
structural degradation and damage during test #04, even though the structural response and

123
5406 Bull Earthquake Eng (2017) 15:5393–5412

Fig. 10 Identified evolution of effective elastic stiffness (ke) in the x-direction over events. The solid line
km represents the moving average stiffness and the dash line kw represents the wavelet stiffness

input PGA of test #06 appears to be larger than test #04 as shown in Figs. 5 and 6. This
difference may be a result of test order. However, most importantly, the identification
results show the method is able to quantify degradation and damage that may not be
observed or easily estimated from its dynamic response and external input parameters.
Equally importantly, while the structural response of the first 3 tests #01 (BSL2-18%),
#02 (Sannomal), and #03 (Uemachi) appear to be within linear behaviour in Figs. 5 and 6.
An approximate 13% reduction of overall stiffness was identified for both directions prior
to the stronger test #04. Thus, damage leading to measurable stiffness degradation may not
be observed with external visual appearance and/or by examination of the general shape of
hysteresis loops if structural elements only experience micro cracking or reinforcement slip
and de-bonding during small shaking events (Maeck et al. 2000; Zhou 2016; Zhou et al.
2017). Therefore, this result shows the ability of this method to robustly detect damage that
may not be observed by external visual inspection due to smaller and more common
events, but which significantly alters future performance from that expected. Hence, small
events can aggregate to measureable damage that could affect performance in large events,
but may go undetected in otherwise apparently linear responses.
More importantly, the identified stiffness, ke, matched well between different earth-
quake events for all the stories in both x and y directions. Figure 12 compares the identified
final stiffness ke of one earthquake event and the initial stiffness ke of the following
earthquake event in both directions. Ideally, these values should be the same. The stiffness
values in the x-direction are slightly higher than stiffness in the y-direction. Thus, the
overall error in the x-direction is relatively smaller than in the y-direction on average as a
percentage because part of the error is due to fixed sensor errors. However, it is important

123
Bull Earthquake Eng (2017) 15:5393–5412 5407

Fig. 11 Identified evolution of effective elastic stiffness (ke) in the y-direction over events. The solid line
km represents the moving average stiffness and the dash line kw represents the wavelet stiffness

to note the average differences between the final and initial stiffness over all comparisons
is less than 5%, which validates the continuity and accuracy of the identification approach
across different earthquake events, indicating its ability to be used in long term monitoring.
Finally, changes of stiffness are tracked over time at each floor, which thus provides
significant advantages and insight, particularly in damage level and localization, over
traditional frequency based methods assessing natural frequency and mode shape.
To further validate the accuracy of the identification results and relate them to typical
methods, the fundamental resonant frequency, w, of the test structure during each event is
calculated using the identified final stiffness, ke, of each test and compared with the
experimental frequency from the experimental transfer function. Table 2 lists the com-
parison of the calculated and experimental transfer function frequency of all tests in the x-
direction and y-direction, respectively. The calculated frequency shows very similar trends
to the experimental transfer function frequency over events with differences smaller than
5% or 0.1 Hz. This result thus further demonstrates the capability of the method to track
the changes of structural effective stiffness, ke, accurately and consistently over several
earthquake events of different magnitudes, as well as validating its result against other
forms of assessment (Doebling et al. 1996; Fan and Qiao 2011; Salawu 1997; Yan et al.
2007).
Hence, the overall result indicates the HLA method offers potentially significant
advantages over parametric model-based methods because it is model-free based on fun-
damental underlying structural mechanics without requiring prior definition of baseline
mode as seen in prior work (Zhou et al. 2017). Thus, it is fully generalizable to different
type of structural behaviours. In addition, the method uses a far simpler computational

123
5408 Bull Earthquake Eng (2017) 15:5393–5412

Fig. 12 Comparison between the identified final and initial stiffness ke over different earthquake events in
the a x-direction and b y-direction

Table 2 Comparison of the experimental and calculated fundamental frequency w


Test x-direction y-direction

Experimental Calculated Difference (%) Experimental Calculated Difference (%)


(Hz) (Hz) (Hz) (Hz)

#01 2.759 2.805 –1.7 2.686 2.784 –3.6


#02 2.618 2.698 –3.1 2.579 2.678 –3.8
#03 2.661 2.529 5.0 2.588 2.626 –1.5
#04 2.100 2.184 –4.0 2.051 2.142 –4.4
#05 2.234 2.268 –1.5 2.145 2.222 –3.6
#06 1.862 1.883 –1.1 1.941 1.950 –0.5

algorithm than genetic algorithms and many other nonparametric algorithms whose con-
vergence towards global optimum and speed of convergence are not always guaranteed.
Finally, the method is tested in this work using a real full-scale complex structure that has
more uncertainty, nonlinearity and complexity, which thus provides a significant real-

123
Bull Earthquake Eng (2017) 15:5393–5412 5409

Fig. 13 Effect of damping ratio on the identification results for a the second story in the x-direction; b the
first story in the y-direction, during the Toshin-Seibu earthquake (test #04)

Fig. 14 Effect of damping ratio on the restoring force for the second story in x-direction during the Toshin-
Seibu earthquake (test #04)

world implementation validation, well beyond numerical analysis and carefully controlled
laboratory small scale structure tests.

4.2 Effect of damping force on identification results

Figure 13 shows the identified changes of ke with different damping ratios. The identified
stiffness evolution shows a very similar trajectory over different damping ratios. In
addition, the identified stiffness values for different damping ratios vary only slightly over
time with the greatest difference of 8.1%. This result indicates the damping force is a
relatively small part of the total damping force in the elastic segments, as further illustrated
in Fig. 14. It can be seen the change of restoring force is very small with damping force
changing from 3 to 10%. Thus, the identification method can still capture the changes of
stiffness over time and different earthquake events to indicate the occurrence of structural
damage and degradation, even when the actual damping ratio is changing over time and/or
is different than assumed.

5 Conclusions

This research provides unique full-scale experimental validation of a model-free,


mechanics-relevant stiffness identification method for damage assessment and structural
health monitoring. The effective stiffness, ke, of a full-scale SMRF building are calculated
using the HLA method to detect and quantify damage in both x and y shaking direction
over a sequence of 6 earthquake events. The results from experimental transfer functions

123
5410 Bull Earthquake Eng (2017) 15:5393–5412

for each event are employed to further validate the accuracy of the identified stiffness. The
main conclusion drawn from this study are:
• CDFs of residual errors clearly show 95% of experimentally measured half cycles for
all cases are fitted with residual errors less than 5% using the identified regression
models, indicating a good accuracy of hypothesis test and regression analysis.
• The identified final stiffness of each individual event matched well with the identified
initial stiffness of the next event with the average difference smaller than 5%. In
addition, the differences between the experimental fundamental frequency and the
calculated frequency from the identified stiffness for all the 6 sequential tests are less
than 0.1 Hz. These results validate the capability of the method to quantify and localize
damage consistently and accurately in a long term monitoring.
• Measureable stiffness degradation is identified during the small test events, while a
largely linear response is apparently observed in the hysteresis loops. Hence, it is able
to detect small, unobserved damage that aggregates to significant changes over several
small events.
• The evolution of stiffness is identified with a similar trend and small deviation over a
typical range of damping ratios, indicating the robustness of this method to different
structures with different damping effects.
• This method avoids constraint to a highly specific complex model and uses a far
simpler computational algorithm than many current methods, which ensures rapid or
real-time assessment to offer significant information of structural damage immediately
after an event.

Acknowledgements Deserved acknowledgement is to be given to Dr. M. Koyama and Dr. S. Nakata of


Asahi Kasei Homes Corp., Japan, and Dr. T. Sasaki of E-Defense, Japan for their contribution to the shake
table test.

References
Adeli H, Jiang X (2006) Dynamic fuzzy wavelet neural network model for structural system identification.
J Struct Eng 132:102–111. doi:10.1061/(asce)0733-9445(2006)132:1(102)
Al-Hussein A, Haldar A (2015) Structural health assessment at a local level using minimum information.
Eng Struct 88:100–110. doi:10.1016/j.engstruct.2015.01.026
Atkinson GM, Pierre J-R (2004) Ground-motion response spectra in eastern North America for different
critical damping values. Seismol Res Lett 75:541–545. doi:10.1785/gssrl.75.4.541
Bai J, Perron P (1998) Estimating and testing linear models with multiple structural changes. Econometrica
66:47–78. doi:10.2307/2998540
Bai J, Perron P (2003) Computation and analysis of multiple structural change models. J Appl Econom
18:1–22. doi:10.1002/jae.659
Belleri A, Moaveni B, Restrepo JI (2014) Damage assessment through structural identification of a three-
story large-scale precast concrete structure. Earthq Eng Struct Dyn 43:61–76. doi:10.1002/eqe.2332
Çatbaş FN, Kijewski-Correa T, Aktan AE (2013) Structural identification of constructed systems:
approaches, methods, and technologies for effective practice of St-Id. Am Soc Civ Eng. doi:10.1061/
9780784411971
Chang C-C, Shi Y (2010) Identification of time-varying hysteretic structures using wavelet multiresolution
analysis. Int J Non Linear Mech 45:21–34. doi:10.1016/j.ijnonlinmec.2009.08.009
Chase J, Leo Hwang K, Barroso L, Mander J (2005a) A simple LMS-based approach to the structural health
monitoring benchmark problem. Earthq Eng Struct Dyn 34:575–594. doi:10.1002/eqe.433
Chase JG, Spieth HA, Blome CF, Mander J (2005b) LMS-based structural health monitoring of a non-linear
rocking structure. Earthq Eng Struct Dyn 34:909–930. doi:10.1002/eqe.460

123
Bull Earthquake Eng (2017) 15:5393–5412 5411

Chopra AK (1995) Dynamics of structures: theory and applications to earthquake engineering, vol 3.
Prentice Hall, New Jersey
Cifuentes AO, Iwan WD (1989) Nonlinear system identification based on modelling of restoring force
behaviour. Soil Dyn Earthq Eng 8:2–8. doi:10.1016/s0267-7261(89)80002-8
Coca D, Billings S (2001) Non-linear system identification using wavelet multiresolution models. Int J
Control 74:1718–1736
Doebling SW, Farrar CR, Prime MB, Shevitz DW (1996) Damage identification and health monitoring of
structural and mechanical systems from changes in their vibration characteristics: a literature review.
Los Alamos National Lab., NM (United States), Los Alamos National Lab., NM (United States)
Draper NR, Smith H (2014) Applied regression analysis. Wiley, New Jersey
Fan W, Qiao P (2011) Vibration-based damage identification methods: a review and comparative study.
Structural Health Monitoring 10:83–111
Hann CE, Singh-Levett I, Deam BL, Mander JB, Chase JG (2009) Real-time system identification of a
nonlinear four-story steel frame structure—application to structural health monitoring. IEEE Sens J
9:1339–1346. doi:10.1109/JSEN.2009.2022434
Huang H, Yang JN, Zhou L (2010) Comparison of various structural damage tracking techniques based on
experimental data. Smart Struct Syst 6:1057–1077. doi:10.12989/sss.2010.6.9.1057
Huang Q, Gardoni P, Hurlebaus S (2012) A probabilistic damage detection approach using vibration-based
nondestructive testing. Struct Saf 38:11–21. doi:10.1016/j.strusafe.2012.01.004
Huang Q, Gardoni P, Hurlebaus S (2015) Assessment of modal parameters considering measurement and
modeling errors. Smart Struct Syst 15:717–733. doi:10.12989/sss.2015.15.3.717
Ikeda Y (2016) Verification of system identification utilizing shaking table tests of a full-scale 4-story steel
building. Earthq Eng Struct Dyn 45:543–562. doi:10.1002/eqe.2670
Im SB, Cloudt HC, Fogle JA, Hurlebaus S (2013) Enhanced damage index method using torsion modes of
structures. Smart Struct Syst 12:427–440. doi:10.12989/sss.2013.12.3_4.427
Iwan WD (2002) R-SHAPE: a real-time structural health and performance evaluation system. In: Pro-
ceedings of the US europe workshop on sensors and smart structures technology, Como and Somma
Lombardo, Italy, April 12–13. pp 33–38
Iwan WD, Cifuentes AO (1986) A model for system identification of degrading structures. Earthq Eng
Struct Dyn 14:877–890
Iwan WD, Radulescu DC, Radulescu C (2013) Extreme event performance evaluation using real-time
hysteresis monitoring. US Patent 8,538,734
Lei Y, Chen F, Zhou H (2015) An algorithm based on two-step Kalman filter for intelligent structural
damage detection. Struct Control Health Monit 22:694–706. doi:10.1002/stc.1712
Londoño JM, Neild SA, Cooper JE (2015) Identification of backbone curves of nonlinear systems from
resonance decay responses. J Sound Vib 348:224–238. doi:10.1016/j.jsv.2015.03.015
Maeck J et al (2000) Damage identification in reinforced concrete structures by dynamic stiffness deter-
mination. Eng Struct 22:1339–1349. doi:10.1016/s0141-0296(99)00074-7
Mallat SG (1989) A theory for multiresolution signal decomposition: the wavelet representation. IEEE
Trans Pattern Anal Mach Intell 11:674–693. doi:10.1109/34.192463
Nayyerloo M, Chase J, MacRae G, Chen X (2011) LMS-based approach to structural health monitoring of
nonlinear hysteretic structures. Struct Health Monit 10:429–444. doi:10.1177/1475921710379519
Ovanesova A, Suarez L (2004) Applications of wavelet transforms to damage detection in frame structures.
Eng Struct 26:39–49. doi:10.1016/j.engstruct.2003.08.009
Pan S, Xiao D, Xing S, Law S, Du P, Li Y (2016) A general extended Kalman filter for simultaneous
estimation of system and unknown inputs. Eng Struct 109:85–98. doi:10.1016/j.engstruct.2015.11.014
Pekcan G, Mander JB, Chen SS (1999) Fundamental considerations for the design of non-linear viscous
dampers. Earthq Eng Struct Dyn 28:1405–1425. doi:10.1002/(SICI)1096-9845(199911)28:11\1405:
AID-EQE875[3.0.CO;2-A
Rao J, Gupta K (1999) Introductory course on theory and practice of mechanical vibrations. New Age
International, New Delhi
Salawu O (1997) Detection of structural damage through changes in frequency: a review. Eng Struct
19:718–723
Sirca G, Adeli H (2012) System identification in structural engineering. Sci Iran 19:1355–1364. doi:10.
1016/j.scient.2012.09.002
Sjöberg J et al (1995) Nonlinear black-box modeling in system identification: a unified overview. Auto-
matica 31:1691–1724
Smith SW (1999) The scientist and engineer’s guide to digital signal processing. California Technical
Publishing, San Diego

123
5412 Bull Earthquake Eng (2017) 15:5393–5412

Smyth A, Wu M (2007) Multi-rate Kalman filtering for the data fusion of displacement and acceleration
response measurements in dynamic system monitoring. Mech Syst Signal Process 21:706–723. doi:10.
1016/j.ymssp.2006.03.005
Stephens JE, Yao JT (1987) Damage assessment using response measurements. J Struct Eng 113:787–801.
doi:10.1061/(asce)0733-9445(1987)113:4(787)
Toussi S, Yao JT (1983) Hysteresis identification of existing structures. J Eng Mech 109:1189–1202
Wei H-L, Billings S (2002) Identification of time-varying systems using multiresolution wavelet models. Int
J Syst Sci 33:1217–1228
Wu M, Smyth AW (2007) Application of the unscented Kalman filter for real-time nonlinear structural
system identification. Struct Control Health Monit 14:971–990. doi:10.1002/stc.186
Wu M, Smyth A (2008) Real-time parameter estimation for degrading and pinching hysteretic models. Int J
Non Linear Mech 43:822–833. doi:10.1016/j.ijnonlinmec.2008.05.010
Xie Z, Feng J (2012) Real-time nonlinear structural system identification via iterated unscented Kalman
filter. Mech Syst Signal Process 28:309–322. doi:10.1016/j.ymssp.2011.02.005
Xu C, Chase JG, Rodgers GW (2014) Physical parameter identification of nonlinear base-isolated buildings
using seismic response data. Comput Struct 145:47–57. doi:10.1016/j.compstruc.2014.08.006
Xu C, Chase JG, Rodgers GW (2015) Nonlinear regression based health monitoring of hysteretic structures
under seismic excitation. Shock Vib 2015:1–12. doi:10.1155/2015/193136
Yan Y, Cheng L, Wu Z, Yam L (2007) Development in vibration-based structural damage detection
technique. Mech Syst Signal Process 21:2198–2211
Yang JN, Lin S, Huang H, Zhou L (2006) An adaptive extended Kalman filter for structural damage
identification. Struct Control Health Monit 13:849–867. doi:10.1002/stc.84
Yao R, Pakzad SN (2014) Time and frequency domain regression-based stiffness estimation and damage
identification. Struct Control Health Monit 21:356–380. doi:10.1002/stc.1570
Zhou C (2016) Efficient hysteresis loop analysis based structural health monitoring of civil structures.
University of Canterbury
Zhou C, Chase JG, Rodgers GW, Tomlinson H, Xu C (2015a) Physical parameter identification of structural
systems with hysteretic pinching. Comput Aided Civ Infrastruct Eng 30:247–262. doi:10.1111/mice.
12108
Zhou C, Chase JG, Rodgers GW, Xu C, Tomlinson H (2015b) Overall damage identification of flag-shaped
hysteresis systems under seismic excitation. Smart Struct Syst 16:163–181. doi:10.12989/sss.2015.16.
1.163
Zhou C, Chase JG, Rodgers GW, Xu C (2017) Comparing model-based adaptive LMS filters and a model-
free hysteresis loop analysis method for structural health monitoring. Mech Syst Signal Process
84:384–398. doi:10.1016/j.ymssp.2016.07.030

123

You might also like