Journal of Civil Structural Health Monitoring

https://doi.org/10.1007/s13349-020-00416-1

ORIGINAL PAPER

Performance‑based post‑earthquake decision making

for instrumented buildings
Milad Roohi1   · Eric M. Hernandez2

Received: 9 December 2019 / Revised: 26 May 2020 / Accepted: 8 June 2020

© Springer-Verlag GmbH Germany, part of Springer Nature 2020

Abstract
This paper develops a decision making framework for post-earthquake assessment of instrumented buildings in a manner
consistent with performance-based design criteria. This framework is achieved by simultaneously combining and advancing
existing knowledge from seismic structural health monitoring and performance-based earthquake engineering paradigms.
The framework consists of (1) measurement, (2) uncertainty modeling, (3) dynamic response reconstruction, (4) damage
estimation, and (5) performance-based assessment and decision making. In particular, the main objective is to reconstruct
inter-story drifts with a probabilistic measure of exceeding performance-based acceptance limits and determine the post-
earthquake re-occupancy classification of the instrumented building of interest. Since the proposed framework is probabilistic,
the outcome can be used to obtain the probability of losses based on the defined decision variables and be integrated into a
risk-based decision making process by city officials, building owners, and emergency managers. The framework is illustrated
using data from the Van Nuys hotel testbed, a seven-story reinforced concrete building instrumented by the California Strong
Motion Instrumentation Program (CSMIP Station 24386).

Keywords  Decision making · Performance-based earthquake engineering · Seismic structural health monitoring · Dynamic
response reconstruction · Instrumented buildings · Real-world validation
List of symbols G0 Constant power spectral density intensity
arg min Argument of the minimum hk Height of the kth story
𝐛1 Spatial distribution of excitation I(t) Non-negative envelope function
𝐛2 Spatial distribution of process noise 𝐊 Stiffness matrix
c2 Output location matrix 𝐌 Mass matrix
𝐂𝜉 Damping matrix n Number of degrees-of-freedom
e State error p[.] Probability
𝔼 Expected value 𝐏 State error covariance
E Viscous damping coefficient 𝐏ISD Inter-story drift error covariance
𝐄 Feedback matrix q(t) Displacement vector
𝐄opt Optimal feedback matrix q̂ (t) Displacement vector estimate
F(t) Corrective force ̇ Velocity vector
q(t)
fc′ Compressive strength of concrete ̈ Acceleration vector
q(t)
fR (.) Restoring force function Sü ∗ ü ∗ (𝜔) Kanai–Tajimi power spectral density
t Time
tr(.) Trace
* Milad Roohi ü g (t) Ground acceleration vector
mroohigh@colostate.edu
v(t) Measurement noise
Eric M. Hernandez w(t) Process noise
eric.hernandez@uvm.edu
z(t) Vector of auxiliary variables
1
Department of Civil and Environmental Engineering, y(t) Measured displacement vector
Colorado State University, Fort Collins, CO 80523, USA ̇ Measured velocity vector
y(t)
2
Department of Civil and Environmental Engineering, ÿ (t) Measured acceleration vector
University of Vermont, Burlington, VT 05405, USA 𝜉g Site dominant damping coefficient

123
Vol.:(0123456789)
 Journal of Civil Structural Health Monitoring

𝜎 Standard deviation earthquake sequence) have demonstrated that this is not a

𝜎 2 Variance hypothetical scenario, but a real possibility. Although most
𝜱(𝜔) Power spectral density engineered buildings are expected to survive a strong ground
𝜱ee (𝜔) Error spectral density matrix motion without collapse, it is not clear that these buildings
𝜱vv (𝜔) Power spectral density of measurement noise will be safe to re-occupy, especially if they can be subjected
𝜱ww (𝜔) Power spectral density of uncertain inputs to strong aftershock.
𝜔 Frequency In the United States, and other parts of the world, doc-
𝜔g Site dominant frequency uments such as ATC-20 [1, 2] offer guidelines for visual
N Normal distribution post-earthquake assessment and occupancy classification
of potentially damaged buildings as inspected (green tag),
Abbreviations
restricted use (yellow tag) and unsafe (red tag). Despite best
ASCE American Society of Civil Engineers
efforts by inspectors, visual inspections suffer from several
ATC​ Applied Technology Council
notable limitations, including but not limited to: (1) inspec-
C Collapse
tor bias and (or) experience-based variability, (2) lack of
CP Collapse prevention
access to damaged locations or members, (3) time consum-
CSMIP California Strong Motion Instrument Program
ing, and (4) qualitative in nature and not entirely quantitative
DM Damage measure
or physics-based. These limitations might lead inspectors
DoF Degree of freedom
to reach erroneous conclusions about which buildings are
DV Decision variable
safe to be re-occupied immediately and which ones are not,
EDP Engineering demand parameter
exacerbating earthquake losses.
EKF Extended Kalman filter
In the case of instrumented buildings, engineers can aug-
FE Finite element
ment the assessment by incorporating measurements during
FEMA Federal emergency management agency
earthquakes. However, despite the immediate appeal, there
IO Immediate occupancy
are technical, logistical, and economic challenges associ-
ISD Inter-story drift
ated with building instrumentation including but not lim-
KF Kalman filter
ited to: (a) building instrumentation and its maintenance are
LS Life safety
relatively expensive, and budget constraints may not allow
M Measurement
floor-by-floor or component-level instrumentation with only
NMBO Nonlinear model-based observer
a limited number of accelerometers (an average of 9 to 12)
PBA Performance-based assessment
installed per building, and (b) direct measurement of damage
PBEE Performance-Based Earthquake Engineering
is often not possible and only the consequences of damage
PBM Performance-based monitoring
can be measured, this requires the solution of an inverse
PEER Pacific Earthquake Engineering Research
problem that maps indirect measurements to physical struc-
PF Particle filters
tural damage. During the last few decades, researchers have
PL Performance level
proposed a broad range of methods for seismic structural
PSD Power spectral density
health monitoring and post-earthquake assessment of build-
RC Reinforced concrete
ings. Depending on the damage features, existing methods
RMS Root-mean-square
can be widely categorized into: (1) spectral methods (e.g.,
UKF Unscented Kalman filter
Bernal [3], Sadeghi Eshkevari et al. [4]), (2) state estima-
tion methods (e.g., Roohi et al. [5, 6], Erazo and Hernandez
[7], Ching et al [8], Hu et al. [9]), (3) wave propagation
1 Introduction methods (e.g., Safak [10], Todorovska and Trifunac [11]),
(4) energy methods (e.g., Roohi et al. [12], Hernandez and
After a potentially damaging earthquake, city officials must May [13]), (5) model updating methods (e.g., Simoen et al.
make decisions regarding the structural integrity of building [14], Astroza et al. [15]), (6) fragility analysis methods (e.g.,
structures under their jurisdiction. Furthermore, exposure Naeim et al. [16], Lenjani et al. [17]) and (7) machine-learn-
to sequential seismic events following a major earthquake ing methods (Azimi and Pekcan [18], Pan et al. [19], Man-
can exacerbate the problem resulting in cumulative physi- galathu et al. [20], Zhang et al. [21], Lenjani et al. [22]).
cal damage and, in some cases, even partial or complete For more information regarding these approaches as well as
collapse. Recent earthquakes in New Zealand (2010–2011 their advantages and disadvantages refer to Roohi et al. [12],
Canterbury sequence), Taiwan (1999 Chi–Chi mainshock- Roohi [23] and other references (e.g., Wu and Jahanshahi
large aftershocks sequence), Iran (2012 East Azerbaijan [24], Sohn et al. [25], Fan and Qiao [26], Azimi et al. [27]).
doublet), and Puerto Rico (2020 southwestern Puerto Rico

123
Journal of Civil Structural Health Monitoring
Furthermore, efforts have been made to bridge the seis-
mic structural health monitoring and performance-based
earthquake engineering paradigms to develop probabilistic
frameworks that can assess the performance of buildings
and provide vital information to support decision mak-
ing immediately following an earthquake. In [28], Çelebi
et al. developed a seismic monitoring system for real-time
monitoring of inter-story drifts. This system requires meas-
uring the displacement of adjacent floors (i.e., inter-story
drifts) on multiple pairs of building floors and then relates
inter-story drifts to performance-based damage thresholds
specified in documents such as FEMA-356 or ASCE 41.
Miranda [29] and Naeim et al. [16] distinguished probabil-
istic measures based on fragility functions as the most prom-
ising tools for real-time damage detection and conducted
an exhaustive investigation on automated post-earthquake
damage assessment of instrumented buildings. Porter et al.
[30] developed a framework based on PEER PBEE meth-
odology for near-real-time loss estimation for instrumented
buildings. This framework performs Bayesian updating to
estimate the structural response based on measured base-
ment motion and a stochastic structural model and then
uses fragility functions to predict structural damage. Later,
Mitrani-Reiser et al. [31] proposed a decision-support sys-
tem for both immediate pre-event and post-event building
safety and loss assessment. This system integrated fragil-
ity curves and shaking intensity predictions from seismic
network data and employs the PEER PBEE methodology
to perform probabilistic damage and loss estimation and,
subsequently, determine post-earthquake occupancy class
of the building according to the ATC-20 procedure. Hwang
and Lignos [32] proposed a data-driven framework to esti-
mate story-based seismic engineering demand parameters
(EDPs) along the height of instrumented steel frame build-
ings with moment-resisting frames using a wavelet-based
damage sensitive feature. The estimated EDPs were used for
earthquake-induced loss assessment. In [33], Cremen and
Baker provided a methodology for quantifying the benefits
of building instrumentation, by measuring errors in damage
and loss consequence predictions calculated from the FEMA
P-58. Recently, Roohi et al. [5, 6] developed a nonlinear
model-based observer for response reconstruction (i.e., state
estimation) in nonlinear hysteretic structural systems and
employed the observed for performance-based assessment of
experimental and real-world large-scale instrumented steel
(Roohi et al. [34]), reinforced concrete (Roohi et al. [12])
and wood-frame (Roohi et al. [5]) buildings.
Despite the progress described, there still exists a lack
of a comprehensive performance-based rapid and reliable
decision making procedure to integrate all the available
information (such as measurements, structural drawings,
construction information) to assess the extent of the struc-
tural damage accurately and, subsequently, make informed
decisions. Such a procedure will lead to mitigate earth-
quake losses, reduce the decision making uncertainty, and
improve community resilience.
This paper develops a post-earthquake assessment
framework consistent with criteria from performance-
based design. The proposed framework consists of the fol-
lowing steps: (1) measurement, (2) uncertainty modeling,
(3) dynamic response reconstruction, (4) damage estima-
tion, and (5) performance-based assessment and decision
making. Since the proposed framework is developed on a
probabilistic basis, the outcome can be used to obtain the
probability of various losses based on the defined deci-
sion variable and be integrated into a risk-based decision
making process.
This paper is organized as follows. Section  3 pre-
sents the building model of interest. This is followed by
Sect. 4 that presents background on performance-based
earthquake engineering. Section 5 develops the proposed
framework for performance-based post-earthquake deci-
sion making. Finally, the paper ends with a case study of
the Van Nuys hotel testbed, a seven-story reinforced con-
crete (RC) building instrumented by the California Strong
Motion Instrumentation Program (CSMIP Station 24386),
to illustrate the effectiveness of the proposed framework.
2 Building model
The global response of typical multistory building
structures to seismic ground motions can be accurately
described for engineering purposes by
𝐌q(t)
̈ + 𝐂𝜉 q(t)
̇ + fR (q(t), q(t),
̇ z(t))
(1)
= −𝐌𝐛1 ü g (t) + 𝐛2 w(t)
where the vector q(t) ∈ ℝn contains the relative displace-
ment (with respect to the ground) of all stories. z(t) is a
vector of auxiliary variables dealing with material nonlinear-
ity and damage behavior. n denotes the number of geometric
DoF, 𝐌 = 𝐌T ∈ ℝn×n is the mass matrix, 𝐂𝜉 = 𝐂T𝜉 ∈ ℝn×n
is the damping matrix, fR (⋅) is the resultant global restoring
force vector. The matrix 𝐛1 ∈ ℝn×r is the influence matrix of
the r ground acceleration time histories defined by the vector
ü g (t) ∈ ℝr . The matrix 𝐛2 ∈ ℝn×p defines the spatial distribu-
tion the vector w(t) ∈ ℝp , which in the context of this paper
represents the process noise generated by unmeasured exci-
tations and (or) modeling errors.
123

3 Performance‑based earthquake
engineering
The aftermath of the 1994 Northridge and 1995 Kobe
earthquakes revealed significant vulnerability in the way
buildings and other structures were designed to resist
earthquakes. Following these, researchers and engineers
realized the need to develop seismic design and assess-
ment methods, which can improve the seismic vulner-
ability of structures and control earthquake losses. These
efforts resulted in the development of an important engi-
neering concept known as performance-based earthquake
engineering (PBEE). PBEE includes concepts and tech-
niques related to the design, construction, and mainte-
nance of structures aimed to ensure (as much as possible)
predictable performance objectives are met under earth-
quake demands [35]. In the first generation of the PBEE
documents in the United States (also called as PBEE-1),
the report of SEAOC Vision 2000 [35] made an impor-
tant step toward the realization of the PBD and PBA of
buildings. This report classified the system performance
levels as fully operational, operational, life safety, and near
collapse, and classified hazard levels as frequent, occa-
sional, rare, and very rare events. The stakeholders can
determine the desired performance objective of the system
based on the system performance levels corresponding to
different hazard levels. Subsequent documents of PBEE-1
such as ATC-40 [36], FEMA-273 [37], FEMA-356 [38]
and ASCE/SEI 41-13 [39] used a similar framework and
slightly modified the descriptions for system performance
and hazard levels. These documents established approxi-
mate relationships between seismic response parameters
(inter-story drifts, inelastic element deformations, and
element forces) and qualitative performance measures of
Immediate Occupancy (IO), Life Safety (LS), Collapse
Prevention (CP), and Collapse (C). They also proposed
component level acceptance criteria for structural and
Fig. 1  Summary of the PEER PBEE framework
123
Journal of Civil Structural Health Monitoring
non-structural elements for various static/dynamic linear/
nonlinear analysis.
The Pacific Earthquake Engineering Research Center
(PEER) considered the shortcomings of the PBEE-1 and
developed a second-generation of the PBEE framework
known as PEER PBEE (also called as PBEE-2). Figure 1
presents a summary of the PEER PBEE framework [40].
This framework provides a more robust and probabilistic
methodology based on four logical steps, including (1) haz-
ard analysis, (2) structural analysis, (3) damage analysis, and
(4) loss analysis. The outcome of every step is character-
ized by one of four generalized variables: intensity meas-
ure (IM), engineering demand parameter (EDP), damage
measure (DM), and decision variable (DV). These variables
are defined as follows, IM is a parametric representation of
ground motion intensity, such as peak ground acceleration,
EDP is a parametric representation of structural response
to ground motion, such as displacements, velocities, accel-
erations at all degrees of freedom, DM is a parametric
representation of a damage state such as cracks, failure in
connections or structural collapse, and DV is a parametric
expression of the decision variable, such as loss expressed
in terms of repair costs, casualties or lost occupancy time.
Using the Total Probability Theorem, the PEER PBEE
framework equation can be expressed
∭
p[DV] = p[DV|DM] p[DM|EDP] p[EDP|IM] …
(2)
p[IM|D] dIM . dEDP . dDM
where, the expression p[X|Y] refers to the probability den-
sity of X conditioned on knowledge of Y; D denotes facil-
ity location, structural, non-structural, and other features;
p[IM|D] is the conditional probability of experiencing a
given level of ground motion intensity given the location
D; p[EDP|IM] is the conditional probability of experienc-
ing a level of structural response, given a level of ground
motion intensity; p[DM|EDP] is the conditional probability
of experiencing the damage state, given a level of structural
Journal of Civil Structural Health Monitoring
response; p[DV|DM] is the conditional probability of expe-
riencing a loss of certain size, given a level of damage. The
expected loss or value of the decision variable p[DV] is
calculated as the sum of these quantities over all levels of
intensity, response, damage, and loss. The following section
aims to develop a framework for decision making regarding
the post-earthquake assessment of instrumented buildings in
a manner consistent with criteria from performance-based
design.
4 Proposed framework
for performance‑based post‑earthquake
decision making
The proposed decision making framework stems from the
performance-based design framework and consists of the fol-
lowing five steps: (1) measurement, (2) uncertainty modeling,
(3) dynamic response reconstruction, (4) damage estimation,
and (5) performance-based assessment and decision making.
Figure 2 presents a summary of the proposed framework char-
acterized by four variables consist of Response Measurement
(M), Engineering Demand Parameter (EDP), Damage Meas-
ure (DM), and Decision Variable (DV). Using the Total Prob-
ability Theorem, the decision variable can be expressed by
∭
p[DV] = p[DV|DM] p[DM|EDP] p[EDP|M] …
(3)
p[M] dM . dEDP . dDM
where p[M] is the probability density of measurement set,
and p[EDP|M] is the conditional probability of experiencing
a level of response given measurement set M . Except for a
few special cases, solving the multidimensional integrals
in Eq. 3 is very complex and challenging task as it requires
the complete probability distribution of the p[EDP|M] ,
p[DM|EDP] , and p[DV|EDP] to be estimated. For instance,
to estimate p[EDP|M] in the special case of linear structural
Fig. 2  Summary of the proposed performance-based post-earthquake decision making framework
systems (which can be described by linear models), the den-
sities p[EDP|M] are Gaussian. This means they can be char-
acterized by mean vectors and covariance matrices; thus, the
mathematical solution becomes trackable. This is important
because in real world application there are many cases that
can be addressed using this special case. However, in the
case of more complicated systems, where there is a need to
solve the nonlinear filtering problem, there does not exist a
finite set of parameters that can characterize the densities
p[EDP|M] . Instead, we seek algorithms that can provide
estimates based on approximations of the probability den-
sity functions using the first two statistical moments [23].
In the following, each step of solving Eq. 3 is discussed in
more details to obtain approximate solution of the p[DV]
and use the outcome for performance-based post-earthquake
decision making.
4.1 Measurement
The first step of the proposed framework is to perform seis-
mic instrumentation and measure dynamic response during
a seismic event. In practice, this process begins by determin-
ing the type, number, and locations of the sensors consider-
ing technical, logistical, and economic constraints. With-
out loss of generality, this paper focuses on accelerometers
as the sensor of choice due to their popularity, durability,
and reliability. In a typical setup, accelerations are meas-
ured horizontally in three independent and non-intersecting
directions and the vector of acceleration measurements,
ÿ (t) ∈ ℝm , is given by
ÿ (t) =
[ ] (4)
− 𝐜2 𝐌−1 𝐂𝜉 q(t)
̇ + fR (q(t), q(t),
̇ z(t)) − 𝐛𝟐 w(t) + v(t)
123

where 𝐜2 ∈ ℝm×n is a Boolean matrix that maps the DoFs
to the measurements, and v(t) ∈ ℝm×1 is the measurement
noise.
To determine the number and location of sensors (i.e.,
the measurement matrix 𝐜2 in Eq. 4) an optimality crite-
rion is needed to be defined. In this paper, the aim is to
place accelerometers in locations that contain maximum
information for dynamic response reconstruction, i.e.,
select the number and locations of sensors in a way that
minimizes the uncertainty of dynamic response recon-
struction. This minimization can be achieved by selecting
an optimality criterion based on the variance of a user-
defined objective function related to the state of the sys-
tem, such as displacement, internal forces, and stresses.
The proposed framework selects the objective function to
be the sum of the inter-story drift estimation variances.
Therefore, the optimal sensor placement can be achieved
by solving an optimization problem to select the optimal
measurement matrix, (𝐜2 )opt  , subject to maximum inter-
story drift (ISD) estimation variance being bounded by a
maximum allowable variance of 𝜎max 2
 , which can be speci-
fied based on the expected accuracy to determine perfor-
mance-based post-earthquake re-occupation category of
the building of interest. This optimization problem can be
formulated as follows
(𝐜2 )opt = arg min tr(𝐏ISD )
𝐜2
[ 2 ] (5)
s.t. max 𝜎ISD 2
(k, k) k=1∶n < 𝜎max
where tr(.) is trace of a square matrix (defined as the sum of
elements on the diagonal), 𝐏ISD (k, k) is the inter-story drift
estimation error covariance matrix, 𝜎ISD
2
(k, k) is the kth diag-
onal element of inter-story drift estimation error covariance
matrix, k is story number, and n is total number of stories.
Here, tr(𝐏ISD ) is given by
∑
n
tr(𝐏ISD ) = 𝐏ISD (k, k) (6)
k=1
where 𝐏ISD (k, k) is described as
𝐏ISD (k, k) =
{ Kanai–Tajimi PSD, Sü ∗ ü ∗ (𝜔) , is characterized by
𝐏(1, 1), if k = 1 (7)
𝐏(k, k) + 𝐏(k − 1, k − 1) − 2𝐏(k, k − 1), if k > 1
and 𝐏 , the displacement estimation error covariance matrix,
is given by
[[ ]T ]
𝐏 = 𝔼 q(t) − q̂ (t)][q(t) − q̂ (t) (8)
where q̂ (t) is the displacement vector estimate. Sec-
tion 5.3.1 will present an expression to determine 𝐏 (see
Eq. 19). After optimal sensor placement, noise-contaminated
123
Journal of Civil Structural Health Monitoring
measurements during seismic events, characterized by p[M] ,
are obtained.
4.2 Uncertainty modeling
The second step is to model the uncertainty in the unmeas-
ured ground motion excitations and the measurement noise.
The uncertainty modeling is explicitly performed using power
spectral density (PSD). The uncertainty models of this step
can be expressed by 𝜱vv (𝜔) and 𝜱ww (𝜔) , defined as the PSD
of the unmeasured excitations and the measurement noise,
respectively.
4.2.1 Measurement noise characterization
The measurement noise is modeled using a zero-mean Gauss-
ian sequence with a noise-to-signal RMS (root-mean-square)
selected based on the expected accuracy of measurements.
Using the white noise PSD, the PSD of measurement noise,
𝜱vv (𝜔) , is characterized as follows
𝜱vv (𝜔) = 𝜱0 (9)
which implies that the power of the measurement noise is
distributed uniformly over all frequency components. This
may or may not be valid, depending on the sensor. If a differ-
ent PSD is specified by the manufacturer, then it can simply
be used instead of the white assumption.
Additionally, in some modeling situations, one can use
the measurement noise to include the effects of unmodeled
dynamics (typically induced by modeling errors). Some
authors [6] have proposed updating the noise PSD in order to
include these effects. This is an interesting proposal; however,
it lies outside of the scope of this paper, and it will not be
pursued further.
4.2.2 Ground motion characterization
The seismic ground motions can be modeled using a
Kanai–Tajimi power spectral density corrected by an ampli-
tude function to obtain a non-stationary ground motion
acceleration. The PSD of ground motion, 𝜱ww (𝜔) , using the
𝜱ww (𝜔) = Sü ∗ ü ∗ (𝜔)
1 + 4𝜉g2 ( 𝜔𝜔 )2
g
(10)
= G0 [ ]2
1 − ( 𝜔𝜔 )2 + 4𝜉g2 ( 𝜔𝜔 )2
g g
where 𝜔 ∈ (−∞, ∞) ; 𝜉g and 𝜔g are the site dominant damp-
ing coefficient and frequency, respectively; G0 is the con-
stant power spectral intensity of the bed rock excitation. The
simulated ground motion acceleration u(t) ̈ is generated by
Journal of Civil Structural Health Monitoring
filtering a Gaussian white noise through a second order lin-
ear filter (single degree of freedom oscillator) with natural
frequency 𝜔g and viscous damping 𝜉g as follows
̇ + 𝜔2g u(t) = −w(t)
̈ + 2𝜉g 𝜔g u(t)
u(t)
where
√ response of the building at all DoF and (2) a mapping
K C
𝜔g = and 𝜉g =
M 2M𝜔g
M, C, and K are the mass, stiffness, and damping parameters
of the linear filter. w(t) is a Gaussian white noise process
with spectral density Sww (𝜔) = G0 . The Kanai–Tajimi PSD
models the earthquake-induced base motion as a stationary
stochastic process, under the premise that only the frequency
content is considered. To take into account the amplitude
variability of the motion the following time-dependent enve-
lope is used
ü gm (t) = I(t)u(t)
̈ In the literature, researchers have proposed four catego-
[ ]
= I(t) −𝜔2g u(t) − 2𝜉g 𝜔g u(t)
̇
where ü gm (t) is the simulated ground motion and I(t) is a
non-negative function representing the time-dependent enve-
lope. For the purpose of this study, the amplitude modulating
function I(t) is selected as
I(t) = te−𝛼t
The corrected realization of the Kanai–Tajimi PSD in Eq. 12
provides a filtered white noise stochastic time series with
appropriate frequency content and amplitude modulation
for ground acceleration during earthquakes. Therefore, the
model can be conveniently used for stochastic response anal-
ysis of structures to ground motion excitations. Figure 3 pre-
sents a schematic of the procedure used for seismic ground
motion simulation.
4.3 Dynamic response reconstruction
Once data becomes available from a seismic event and the
measurement noise and excitation have been characterized,
Fig. 3  Schematic of a linear filter with Gaussian white noise input to
model seismic ground motion
dynamic response reconstruction is the third step of the
proposed framework. Response reconstruction refers to the
estimation of unmeasured response quantities of interest
or engineering demand parameters (EDP) from a limited
(11) number of global dynamic response measurements, given
by p[EDP|M] . The information needed for reconstruct-
ing dynamic response are the following: (1) the dynamic
between the global and local DoF of every element. An
accurate response reconstruction in this step is vital to
prevent under-estimation or over-estimation of the actual
response of the building. Furthermore, the estimation
uncertainty bound of the reconstructed response helps to
develop a set of maximum, mean, and minimum seismic
demand to consider the best and worst-case scenarios in
assessing the performance of the instrumented building.
The estimated response parameters with their associated
uncertainties can form a demand set to perform damage
estimation.
(12) ries of state observers to perform response reconstruction
based on nonlinear filters including: (a) classical nonlinear
Bayesian filters (e.g., extended Kalman filter (EKF) [41]),
(b) modern nonlinear Bayesian (or statistically linearized)
filters (e.g., unscented Kalman filter (UKF) [42], (c) particle-
based nonlinear Bayesian filters (e.g. the particle filter (PF)
[43]), and (d) model-based state observers (e.g., nonlinear
(13) model-based observer (NMBO) [5]). From these response
reconstruction approaches, the proposed framework uses the
NMBO for response reconstruction in instrumented build-
ings. This is mainly because the NMBO has been formu-
lated in such a way that (1) it explicitly accounts for PSDs
of the excitations and measurement noise, and (2) it can
be implemented directly as a modified nonlinear structural
model of a system subjected to corrective forces. The latter
feature allows for the direct implementation of the nonlinear
state observer using high-fidelity finite element (FE) models.
Thus, the NMBO can take advantage of a wide range of
material and element models available in advanced nonlinear
simulation software packages, which not only results in the
ease of implementation but also improve the state estimation
accuracy due to better modeling capabilities. In contrast, the
existing state observers such as the EKF, UKF, and PF, are
associated with computational modeling limitations, because
the models that such state observers use for state estimation
are rather simple and cannot capture the complexity of the
nonlinear structural behavior. This issue is exacerbated in
problems that require modeling of highly nonlinear behavior.
The next subsection presents a summary of the NMBO and
its implementation.
123
 Journal of Civil Structural Health Monitoring

4.3.1 Nonlinear model‑based observer ( ( ) )−1

𝐇o = −𝐌𝜔2 + 𝐂𝜉 + 𝐜T2 𝐄𝐜2 i𝜔 + 𝐊0 (17)

The NMBO estimate of the displacement response, q̂ (t) , is the matrices 𝜱ww (𝜔) and 𝜱vv (𝜔) are, respectively, the PSDs
given by the solution of the following set of ordinary dif- of the uncertain excitation and measurement noise defined
ferential equations in the previous step.
The optimal feedback gain matrix, 𝐄opt , can be selected
𝐌q̂̈ (t) + (𝐂𝜉 + 𝐜T2 𝐄𝐜2 )q̂̇ (t)+fR (̂q(t), q̂̇ (t), z(t))
(14) by solving an optimization problem to minimize the inter-
= 𝐜T2 𝐄y(t)
̇ story drift estimation error covariance matrix, 𝐏ISD , given by

̇ is the measured velocity vector and 𝐄 ∈ ℝm×m

where y(t) 𝐄opt = arg min tr(𝐏ISD )
is the feedback gain. As can be seen, Eq. 14 is of the same
𝐄 (18)
form of the original nonlinear model of building in Eq. 1. s.t. 𝐄 ∈ ℝ+
A physical interpretation of the NMBO can be obtained by where 𝐏ISD is recalled from Eq. 7 as
viewing the right-hand side of Eq. 14 as a set of corrective
forces applied to a modified version of the original nonlin- 𝐏ISD (k, k)
{
ear model of interest in the left-hand side. The modifica- 𝐏(1, 1), if k = 1
=
tion consists of adding the damping term cT2 𝐄c2 , where the 𝐏(k, k) + 𝐏(k − 1, k − 1) − 2𝐏(k, k − 1), if k > 1
matrix 𝐄 is free to be selected. The diagonal terms of 𝐄 are (7)
equivalent to grounded dampers in the measurement loca- and 𝐏 is described using the PSD of estimation error,
tions, and the off-diagonal terms (typically set to zero) are 𝜱ee (𝜔) , as
equivalent to dampers connecting the respective DoF of the
[[ ][ ]T ]
measurement locations. To retain a physical interpretation, 𝐏 = 𝔼 q(t) − q̂ (t) q(t) − q̂ (t)
the constraints on 𝐄 are symmetry and positive definiteness.
+∞ (19)
The role of grounded dampers is to determine the relative
∫−∞
= 𝜱ee (𝜔)d𝜔
weight given to the measurements and the prediction of the
nonlinear model. In other words, selecting high damping
parameter values would put more weight on the measure- Any optimization algorithm (e.g., Matlab fminsearch) can
ments and selecting low damping parameter values would be used to solve the optimization in Eq. 18 by varying the
put more weight on the nonlinear model predictions for values of the diagonal elements of the 𝐄 matrix to deter-
response reconstruction. Also, the corrective forces 𝐜T2 𝐄y(t)
̇ mine the optimized feedback matrix, 𝐄opt . Figure 4 presents
are proportional to the velocity measurements and added a summary of the nonlinear model-data fusion using the
grounded dampers. The velocity measurements y(t) ̇ can be NMBO. Also, readers are kindly referred to [34, 44–46] for
obtained by integration of acceleration measurements ÿ (t) in implementation examples.
Eq. 4. The integration might add long period drifts in veloc-
ity measurements, and high-pass filtering can be performed
to remove these baseline shifts.
To determine 𝐄 , the objective function to be minimized is
the trace of the estimation error covariance matrix. Since for
a general nonlinear multi-variable case, a closed-form solu-
tion for the optimal matrix 𝐄 has not been found, a numerical
optimization algorithm is used. To derive the optimization
objective function, Eq. 14 is linearized as follows

𝐌q̂̈ (t) + (𝐂𝜉 + 𝐜T2 𝐄𝐜2 )q̂̇ (t) + 𝐊0 q̂ (t) = 𝐜T2 𝐄y(t)
̇ (15)

where 𝐊0 is the initial stiffness matrix. By defining the state

error as e(t) = q(t) − q̂ (t) , it was shown in [44] that the PSD
of estimation error, 𝜱ee (𝜔) , is given by

𝜱ee (𝜔) = 𝐇o 𝐛2 𝜱ww (𝜔) 𝐛T2 𝐇∗o

(16)
+ 𝐇o 𝐜T2 𝐄 𝜱vv (𝜔) 𝐄T 𝐜2 𝐇∗o

where 𝐇o is defined as Fig. 4  Summary of the nonlinear model-data fusion using the nonlin-
ear model-based observer (NMBO)

123
Journal of Civil Structural Health Monitoring

p[DVk ≥ PL] =
4.4 Damage estimation
�PL
p[DVk |DMk ] dDMk
The third step of the proposed framework is to estimate dam- (23)
�PL
age measure (DM) from the estimated response and compare = p(ISDk ) dISDk

= FISD (ISDk ≥ PL)

the DMs with performance-based acceptance criteria. The
outcome of this step is given by p[DM|EDP] , which is the
probability of DM given EDP. Based on the selected dam-
age measure, the p[DM|EDP] is calculated at the element or
system level. Based on the observations from past earthquakes
that the main portion of the seismic damage and loss to the
structural and non-structural elements are associated with
excessive geometric deformations such as inter-story drifts.
Therefore, the maximum inter-story drift is considered as dam-
age measure, and the p[DM] ( = p[ISD] ) is reconstructed from
the estimated EDPs as discussed in the following.
where p(ISDk ≥ PL) is the probability of ISDk exceeding
specific performance levels (PL) at story k, and FISD is the
cumulative probability density (CDF) of the estimated ISD
at story k. Here, performance levels include IO, LS, CP and
C. Additionally, the probability of specific performance
level, p[ISDk = PL] , can be obtained for four classes of
performance levels including IO (i.e., ISDk < IO ), LS (i.e.,
IO ≤ ISDk < LS ), CP (i.e., LS ≤ ISDk < CP , and C (i.e.,
ISDk ≥ CP ) as follows

4.4.1 Inter‑story drift estimation p[DVk = PL] =

⎧ 1 − FISD (ISDk ≥ IO)
⎪ FISD (ISDk ≥ LS) − FISD (ISDk ≥ IO)
for PL = IO
The expected value of maximum ISD estimate at each story
⎨ F (ISD ≥ CP) − F (ISD ≥ LS)
for PL = LS
can be calculated using the NMBO displacement estimates,
⎩ FISD (ISDk ≥ CP)
⎪ ISD k ISD k for PL = CP
q̂ (t) , as follows for PL = C
(24)
[ ] max|̂qk (t) − q̂ k−1 (t)|
𝔼 ISDk = (20) Then, the post-earthquake building classification can be
hk determined based on the probabilities obtained from inter-
story performance assessment. Conservatively assuming
where hk is height of k-th story, and the uncertainty in ISD
that the ISD s are independent, the probability of exceeding
specific performance level, p[ISD ≥ PL] , for the building
estimation can be calculated as follows
2
𝜎ISD = 𝐏ISD(k,k) (21) can be calculated as follows
p[DV ≥ PL] = p[ISD ≥ PL]
k

where 𝜎ISDk is the uncertainty standard deviation of ISD esti-

(1 − p[ISDk ≥ PL])
mation for k-th story. The estimated ISDs and their uncer- ∏
n
=1−
tainties are subsequently used to reconstruct probability k=1 (25)
density function of ISD for each story, p(ISDk ) , assuming a
(1 − FISD [ISDk ≥ PL])
∏n
Gaussian (normal) distribution as follows =1−
( [ ] 2 ) k=1
p(ISDk ) ∼ N 𝔼 ISDk , 𝜎ISD (22)
k and similarly, building-level probability of specific perfor-
mance level, p[ISD = PL] , can be obtained for four classes
of performance levels as follows
4.5 Performance‑based assessment and decision
making p[DV = PL] =
⎧ 1 − p(ISD ≥ IO)
⎪ p(ISD ≥ LS) − p(ISD ≥ IO)
for PL = IO
The estimated p[DM] is subsequently used as input to perfor-
mance model, p[DV|DM] to estimate p[DV] or p[DV ≥ PL] ,
PL = LS (26)
⎨ p(ISD ≥ CP) − p(ISD ≥ LS)
for

defined as probability of DV exceeding specific performance
level (PL) based on performance-based acceptance criteria.
The acceptance criteria relate engineering demand parameters
(such as inter-story drifts, inelastic element deformations, and
element forces) to qualitative performance levels of Immediate
Occupancy (IO), Life Safety (LS), Collapse Prevention (CP),
and Collapse (C) [47]. Therefore, the probability of exceed-
ing specific performance level (PL), p[ISDk ≥ PL] , can be
calculated for each story as follows
⎩ p(ISD ≥ CP)
⎪ for PL = CP
for PL = C
5 Case‑study: Van Nuys hotel testbed
This section illustrates the proposed framework using seis-
mic response measurements from Van Nuys hotel. The
CSMIP instrumented this building as Station 24386, and
the recorded data of this building are available from several

123

earthquakes, including 1971 San Fernando, 1987 Whittier
Narrows, 1992 Big Bear, and 1994 Northridge earthquakes.
From these data, measurements during 1992 Big Bear and
1994 Northridge earthquakes are used in this study to dem-
onstrate the proposed framework.
Researchers have widely studied the Van Nuys building,
and the building was selected as a testbed for research stud-
ies by researchers in PEER [48]. The Van Nuys building
is a 7-story RC building located in San Fernando Valley
in California. The building plan is 18.9 m × 45.7 m in the
North-South and East-West directions, respectively. The
total height of the building is 19.88 m, with the first story
of 4.11 m tall, while the rest are 2.64 m approximately. The
structure was designed in 1965 and constructed in 1966. Its
vertical load transfer system consists of RC slabs supported
by concrete columns and spandrel beams at the perimeter.
The lateral resisting systems are made up of interior concrete
column-slab frames and exterior concrete column-spandrel
beam frames. The foundation consists of friction piles, and
the local soil conditions are classified as alluvium. The test-
bed building is described in more detail in [48, 49].
Since the Van Nuys building was instrumented and
inspected following earthquakes that affected the structure,
the history of damage suffered by this building is well-
documented. These documents show that the building has
experienced insignificant structural and mostly nonstructural
damage before the Northridge earthquake in 1994. However,
the Northridge earthquake extensively damaged the build-
ing. Post-earthquake inspection red-tagged the building and
revealed that the damage was severe in the south longitudi-
nal frame (Frame A). In Frame A, five of the nine columns
in the 4th story (between floors 4 and 5) were heavily dam-
aged due to inadequate transverse reinforcement, and shear
cracks ( ≥ 5 cm) and bending of longitudinal reinforcement
were easily visible [50].
5.1 Measurement
The CSMIP initially instrumented the building with nine
accelerometers at the 1st, 4th, and roof floors. Following the
San Fernando earthquake, CSMIP replaced the recording
layout by 16 remote accelerometer channels connected to a
central recording system. These channels are located at 1st,
2nd, 3rd, 6th, and roof floors. Five of these sensors measure
longitudinal accelerations, ten of them measure transverse
accelerations, and one of them measures the vertical accel-
eration. Figure 5 shows the location of accelerometers. Since
the Van Nuys building was already instrumented, the paper
will not pursue the optimal sensor placement step and read-
ers are referred to [5, 34] for implementation examples of
this step.
123
Journal of Civil Structural Health Monitoring
5.2 Uncertainty modeling
The PSD of measurement noise, 𝜱vv (𝜔) , in each meas-
ured channel was modeled as zero-mean white Gaussian
sequences with a noise-to-signal root-mean-square (RMS)
ratio of 0.02. Also, the PSD of ground motion were mod-
eled using Kanai–Tajimi PSD by defining 𝜉g = 0.35 for
both earthquakes, 𝜔g = 6𝜋 rad/s for Northridge earthquake
and 𝜔g = 2𝜋 rad/s for Big Bear earthquake. The underlying
white noise spectral density G0 for each direction of meas-
ured ground motion for each shake table test was found such
that about 95% of the Fourier transform of the measured
ground motion lies within two standard deviations of the
average from the Fourier transforms of an ensemble of 200
realizations of the Kanai–Tajimi stochastic process. 𝛼 was
selected as 0.12.
5.3 Dynamic response reconstruction
A nonlinear FE model and measurements of the Van Nuys
building was employed to implement the NMBO in Open-
SEES and perform response reconstruction. The follow-
ing subsections present the step-by-step formulation of the
OpenSEES NMBO.
5.3.1 Nonlinear modeling of the Van Nuys hotel testbed
in OpenSEES
The nonlinear FE model of the building was implemented
using a two-dimensional fixed-base model within the envi-
ronment of OpenSEES [51]. This model corresponds to
one of the longitudinal frames of the building (Frame A in
Fig. 5). In the FE model, beams and columns were mod-
eled based on distributed plasticity modeling approach, and
the force-based beam-column elements were used to accu-
rately determine yielding and plastic deformations at the
integration points along the element. Gauss-Lobatto inte-
gration approach was employed to evaluate the nonlinear
response of force-based elements. Each beam and column
element was discretized with four integration points, and the
cross-section of each element was subdivided into fibers.
The uniaxial Concrete01 material was selected to construct
a Kent-Scott-Park object with a degraded linear unload-
ing and reloading stiffness and zero tensile strength. The
uniaxial Steel01 material was used to model longitudinal
reinforcing steel as a bilinear model with kinematic harden-
ing. The elasticity modulus and strain hardening parameters
were assumed to be 200 GPa and 0.01, respectively. Due to
insufficient transverse reinforcement in beams and columns
[52], an unconfined concrete model was defined to model
concrete. The peak and post-peak strengths were defined at
Journal of Civil Structural Health Monitoring
Fig. 5  a Van Nuys hotel testbed (CSMIP Station 24386) and b Loca-
tion of building accelerometers on the West–East elevation and floor
plans
a strain of 0.002 and a compressive strain of 0.006, respec-
tively. The corresponding strength at ultimate strain was
defined as 0.05fc′ for fc� = 34.5 MPa and fc� = 27.6 MPa and
0.2fc′ for fc� = 20.7 MPa . Based on the recommendation of
[53], the expected yield strength of Grade 40 and Grade 60
steel were defined as 345 MPa (50 ksi) and 496 MPa (72
ksi), respectively, to account for inherent overstrength in the
original material and strength gained over time.
5.3.2 Numerical optimization
Numerical optimization was performed using Eq.  18.
Table 1 presents the optimized damper values for each seis-
mic event.
5.3.3 Formulation of the OpenSEES NMBO
The OpenSEES nonlinear FE model was modified by add-
ing grounded dampers in measurement locations and was
subjected to corrective forces. A nonlinear dynamic time
history analysis was performed to estimate the complete
dynamic response. Figure 6 presents a schematic of the
Van Nuys hotel testbed (with the location of accelerom-
eters) along with the OpenSEES nonlinear FE model under
uncertain measured strong ground motion and the Open-
SEES NMBO with corresponding added viscous dampers
and corrective forces in measurement locations.
5.3.4 Displacement response reconstruction results
Figure 7 compares the NMBO displacement estimates and
its estimation uncertainty with those obtained from (a)
response measurements and (b) open-loop analysis of the
FE model under measured ground motion at non-instru-
mented 4th and instrumented 7th stories. In the locations
where measurements are available, the comparison reveals
considerable errors of the FE model in the tracking of the
actual measurements. As seen, the estimation error of the FE
model under the Big Bear earthquake is higher than those
obtained using the Northridge earthquake. However, under
both earthquakes, the NMBO succeeds in accurately track-
ing the response in the measurement locations. A closer
look throughout the complete time history shows that the
NMBO estimates and their discrepancy with the actual value
lie within the confidence interval. The better performance
of the NMBO compared to open-loop FE model analysis
is mainly due to its capability to feed the measurements to
the FE model and improve its prediction capability. This is
one of the main advantages of performing response recon-
struction using the NMBO, which determines the effects
of the measurements in the feedback loop using the added
grounded dampers (i.e., feedback gain in control theory) to
formulate a sufficiently accurate (i.e., optimized) nonlin-
ear state observer. In the non-instrumented locations (e.g.,
4th story), there are no measurements available to compare
with the NMBO and the FE model estimates. However, there
exists a considerable difference between the NMBO and the
FE model estimates (similar to the estimates obtained for
the instrumented stories). It is important to note that the
capability of the NMBO in accurately estimating unmeas-
ured quantities of interest has been previously validated,
where a small number of measurements from an extensively
Table 1  Optimized damper values in kN s/m (kips.s/in) units
Story Big Bear earthquake Northridge earthquake
1 7283.11 (41.59) 5209.72 (29.75)
2 9357.25 (53.43) 6592.45 (37.64)
5 19,299.40 (110.20) 16,612.79 (94.86)
7 34,808.04 (198.76) 47,217.69 (269.62)
123

Fig. 6  a Schematic of the Van Nuys hotel testbed with location of
accelerometers, b the OpenSEES nonlinear FE model subject to
uncertain measured strong ground motion, and c the OpenSEES
NMBO with corresponding added viscous dampers and corrective
forces in measurement locations
instrumented large-scale building was used to reconstruct
complete dynamic response (see Roohi et al. [5]).
123
Journal of Civil Structural Health Monitoring
Fig. 7  Comparison of displacement estimates using NMBO with esti-
mates obtained from open-loop FE analysis and actual measurements
in 4th and 7th stories during Big Bear and Northridge earthquakes.
Measured represents measured response, FE Model represents the
open-loop FE analysis of OpenSEES model under measured ground
motion and NMBO represents the estimated response using the Open-
SEES NMBO with sensor measurements from measured locations
along with 1 𝜎 estimation uncertainty bound
5.4 Damage estimation
This section reconstructs the ISD at each story to perform
damage estimation. As seen in Fig. 5, the displacements of
Journal of Civil Structural Health Monitoring

adjacent stories are only measured for the first two stories, 7
and thus, the actual inter-story drifts can be obtained for
6
these stories and used for validation of the estimates using
the NMBO and the FE model. Figure 8 presents the maxi- 5
mum inter-story drift ratios obtained from response meas- 4
urements for the first two stories, along with those estimated
3
for all the stories using the NMBO and open-loop FE model
analysis during the 1992 Big Bear and the 1994 Northridge 2
earthquakes. As seen, the NMBO outperforms the FE model
1
in the estimation of inter-story drifts, and therefore, the
NMBO ISD estimates and their uncertainties are used then 0
to reconstruct PDF of ISD for each story, p(ISDk ) . Figure 9 0 0.2 0.4 0.6 0.8 1

depicts the estimated PDF of maximum ISD obtained by fit-

ting normal distribution based on first and second moment (a) 1992 Big Bear earthquake
estimates of maximum ISD using the OpenSEES NMBO 7
along with performance-based acceptance criteria of IO,
LS, and CP. 6

Figure 10 presents the reconstructed CDF of maximum 5

ISD for each story during the 1992 Big Bear and 1994 the
Northridge earthquakes. Additionally, performance-based
acceptance criteria of IO, LS, and CP, along with the esti-
mated probability of exceeding various performance levels,
are presented. Table 2 presents the story-by-story probabil-
ity of exceeding a specific performance level, and Fig. 11
depicts the estimated probability of story-level post-earth-
quake performance levels of the Van Nuys building. As seen
in the case of the Big Bear earthquake, the p[DVk ] results
show that all the stories remain in the IO performance-level
with 1.00 probability of exceedance. Also, in the case of the
Northridge earthquake, the p[DVk ] results show that the 6th
and 7th stories can be classified as IO with 0.57 and 0.99
probabilities of exceedance; the 1st, 2nd, 4th, and 5th stories
can be classified as LS with 0.65, 0.86, 0.96, and 0.56 prob-
abilities of exceedance; the 3rd story can be as CP with 0.65
probability of exceedance.
5.5 Performance‑based assessment and decision
making
Table 3 presents the building level probability of exceed-
ing and classifying specific performance levels for the Van
Nuys building, and Fig. 12 depicts the estimated probability
of post-earthquake building classification for the Van Nuys
building. The p[DV] results show that there is 1.00 and 0.81
probabilities of exceedance, respectively, to classify the
building as IO and CP during the 1992 Big Bear and 1994
Northridge earthquakes.
5.6 Discussion on post‑earthquake assessment
and decision making results
The preceding sections illustrated that the proposed deci-
sion making framework is capable of combining a refined
4
3
2
1
0
0 0.5 1 1.5 2 2.5 3
(b) 1994 Northridge earthquakes
Fig. 8  Comparison of maximum inter-story drift ratios obtained from
response measurements with those estimated using NMBO and open-
loop FE model analysis during 1992 Big Bear and 1994 Northridge
earthquakes
distributed plasticity FE model and a limited number of
response measurements to reconstruct the seismic response
of instrumented buildings accurately. Subsequently, the esti-
mated response quantities and their associated uncertainty
can be used to reconstruct inter-story drifts with a probabil-
istic measure of exceeding performance-based acceptance
limits and determining the post-earthquake re-occupancy
classification of the instrumented building of interest. The
estimates during the Big Bear earthquake demonstrate that
all the stories remained in the IO performance-level, and
the building can be classified as IO by a 1.00 probability of
exceedance. Additionally, the estimates during the North-
ridge earthquake indicate that the higher floors remained
IO, and middle and lower floors passed the LS and CP per-
formance levels. In particular, stories 3 and 4 are classified
as CP and LS, respectively, with the 0.65 and 0.96 prob-
abilities of exceedance. It is important to note the classifica-
tions for these two stories are not consistent with the severe

123

(a) 1992 Big Bear earthquake
Fig. 9  Reconstructed probability density function for maximum inter-story drift ratios of the Van Nuys building during the a 1992 Big Bear and
b 1994 Northridge earthquakes
failure of 5 of 9 columns in the fourth story. However, the
building-level estimate showed that the building could be
classified as CP by a 0.80 probability of exceedance. There-
fore, the building-level post-earthquake assessment results
are consistent with the building’s actual performance and
post-earthquake inspection reports following the Big Bear
123
Journal of Civil Structural Health Monitoring
(b) 1994 Northridge earthquake
and Northridge earthquakes. In [12], Roohi et al. have shown
that if the objective is the high-resolution story- or element-
level damage detection and localization, other damage sen-
sitive response parameters such as element-level dissipated
energy, demand-to-capacity ratios, and ductility demand
can provide more accurate assessment results compared to
Journal of Civil Structural Health Monitoring

(a) 1992 Big Bear Earthquake (b) 1994 Northridge Earthquake

Fig. 10  Reconstructed cumulative density function for maximum inter-story drift ratios and probability of exceeding IO, LS, and CP perfor-
mance levels of Van Nuys building during a 1992 Big Bear and b 1994 Northridge earthquakes

inter-story drifts. However, for the application of interest in 6 Conclusions

this paper, the applicability of the proposed framework is
validated for rapid performance-based post-earthquake re-
occupancy classification and decision making in the context
of a real-world building that experienced severe structural
damage during sequential seismic events.
This paper develops a performance-based post-earthquake
decision making framework. This framework consists of
the following five steps: (1) measurement, (2) uncertainty

123
 Journal of Civil Structural Health Monitoring

modeling, (3) dynamic response reconstruction, (4) dam-

age estimation, and (5) performance-based assessment and
decision making. The first step involves the optimal sensor
placement of accelerometers. The objective is to select the
number and locations of sensors in a way that minimizes
the uncertainty in the estimation of displacement response
at all stories. The second step is to characterize the uncer-
tainty in the unmeasured ground motion excitations and the
measurement noise. The uncertainty modeling is explicitly
performed using power spectral density. The third step is
to implement nonlinear model-data fusion by incorporat-
ing a nonlinear structural model, measurements, uncertainty
models, and reconstruct probabilistic engineering demand
parameters (EDP) in all structural members given the meas-
urements. The fourth step is to use the estimated EDPs as
(a) 1992 Big Bear earthquake input to damage models and reconstruct the probability den-
sity of damage measures (DM). The DMs are then employed
to estimate the probability of decision-variable (DV) exceed-
ing the acceptance criteria from the PBEE concept to deter-
mine the post-earthquake re-occupancy category of the
instrumented building. Since this concept is developed on a
probabilistic basis, the results can also be used to obtain the
probability of various losses based on the defined decision
variable and loss model. The outcome of this framework can
be integrated into a decision making process by city officials,
building owners, and emergency managers.
The framework was successfully implemented using
measured data from the seven-story Van Nuys hotel testbed
instrumented by CSMIP (Station 24386) during the 1992
Big Bear and the 1994 Northridge earthquakes. A nonlin-
ear model-based observer of the building was implemented
(b) 1994 Northridge earthquake using a distributed plasticity finite element model and
measured data to reconstruct seismic response during each
Fig. 11  Story-by-story estimated probability of post-earthquake per- earthquake. The estimated seismic response was then used to
formance levels of Van Nuys building during b 1992 Big Bear and b reconstruct probability and cumulative density of inter-story
1994 Northridge earthquakes drifts and determine the performance-based post-earthquake
re-occupation category of the building following each earth-
quake. The performance categories were estimated as IO
(immediate occupancy) for the Big Bear earthquake and CP

Table 2  Story-by-story k (story) 1 2 3 4 5 6 7
probability of exceeding specific
performance level for the Van 1992 Big Bear earthquake
Nuys building during 1992
  p[ISDk < IO] 1.00 1.00 1.00 1.00 1.00 1.00 1.00
  p[ISDk ≥ IO]
Big Bear and 1994 Northridge
earthquakes 0.00 0.00 0.00 0.00 0.00 0.00 0.00
  p[ISDk ≥ LS] 0.00 0.00 0.00 0.00 0.00 0.00 0.00
  p[ISDk ≥ CP] 0.00 0.00 0.00 0.00 0.00 0.00 0.00
1994 Northridge earthquake
  p[ISDk < IO] 0.01 0.00 0.00 0.00 0.44 0.57 0.99
  p[ISDk ≥ IO] 0.99 1.00 1.00 1.00 0.56 0.43 0.01
  p[ISDk ≥ LS] 0.33 0.13 0.66 0.04 0.00 0.00 0.00
  p[ISDk ≥ CP] 0.00 0.00 0.01 0.00 0.00 0.00 0.00

123
Journal of Civil Structural Health Monitoring

Table 3  Building-level probability of exceeding and classifying spe- 3. Bernal D (2006) Flexibility-based damage localization from sto-
cific performance level for the Van Nuys building during 1992 Big chastic realization results. J Eng Mech 132(6):651–658
Bear and 1994 Northridge earthquakes 4. Sadeghi Eshkevari S, Heydari N, Nathan Kutz J, Pakzad SN, Dip-
las P, Sadeghi Eshkevari S (2019) Operational vision-based modal
1992 Big Bear earthquake identification of structures: a novel framework. Struct Health
  p[ISD < IO] 1.00 p[ISD = IO] 1.00 Monit. https​://doi.org/10.12783​/shm20​19/32502​
  p[ISD ≥ IO]
5. Roohi M, Hernandez EM, Rosowsky D (2019) Nonlinear seismic
0.00 p[ISD = LS] 0.00
  p[ISD ≥ LS]
response reconstruction and performance assessment of instru-
0.00 p[ISD = CP] 0.00 mented wood-frame buildings—validation using NEESWood
  p[ISD ≥ CP] 0.00 p[ISD = C] 0.00 capstone full-scale tests. Struct Control Health Monit 26(9):e2373
1994 Northridge earthquake 6. Roohi M, Erazo K, Rosowsky D,  Hernandez EM (2020) An
extended model-based observer for state estimation in nonlinear
  p[ISD < IO] 0.00 p[ISD = IO] 0.00
  p[ISD ≥ IO]
hysteretic structural systems. Mech Syst Signal Process. https​://
1.00 p[ISD = LS] 0.80 doi.org/10.1016/j.ymssp​.2020.10701​5
  p[ISD ≥ LS] 0.81 p[ISD = CP] 0.19 7. Erazo K, Hernandez EM (2016) Uncertainty quantification of
  p[ISD ≥ CP] 0.01 p[ISD = C] 0.01 state estimation in nonlinear structural systems with application
to seismic response in buildings. ASCE-ASME J Risk Uncertain
Eng Syst Part A: Civ Eng 2(3):B5015001
8. Ching J, Beck JL, Porter KA, Shaikhutdinov R (2006) Bayesian
state estimation method for nonlinear systems and its application
to recorded seismic response. J Eng Mech 132(4):396–410
9. Hu RP, Xu YL (2019) Shm-based seismic performance assess-
ment of high-rise buildings under long-period ground motion. J
Struct Eng 145(6):04019038
(a) 1992 Big Bear earthquake 10. Şafak E (1999) Wave-propagation formulation of seismic response
of multistory buildings. J Struct Eng 125(4):426–437
11. Todorovska MI, Trifunac MD (2010) Earthquake damage detection
in the imperial county services building II: analysis of novelties via
wavelets. Struct Control Health Monit 17(8):895–917
12. Roohi M, Hernandez EM, Rosowsky D (2020) Reconstructing
element-by-element dissipated hysteretic energy in instrumented
(b) 1994 Northridge earthquake buildings: application to the Van Nuys Hotel testbed. ASCE J Eng
Mech. arXiv preprint arXiv​:2002.12426​ (under review)
13. Hernandez EM, May G (2012) Dissipated energy ratio as a feature
Fig. 12  Building-level estimated probability of post-earthquake per- for earthquake-induced damage detection of instrumented structures.
formance levels of Van Nuys building during a 1992 Big Bear and b J Eng Mech 139(11):1521–1529
1994 Northridge earthquakes 14. Simoen E, De Roeck G, Lombaert G (2015) Dealing with uncer-
tainty in model updating for damage assessment: a review. Mech
Syst Signal Process 56:123–149
(collapse prevention) for the Northridge earthquake. The 15. Astroza R, Ebrahimian H, Conte JP (2019) Performance comparison
post-earthquake assessment results were consistent with the of kalman-based filters for nonlinear structural finite element model
building’s actual performance and visual inspection reports. updating. J Sound Vib 438:520–542
16. Naeim F, Hagie S, Alimoradi A, Miranda E (2006) Automated post-
The study illustrates the capability of the proposed perfor- earthquake damage assessment of instrumented buildings. In: Wasti
mance-based framework to help structural engineers make ST, Ozcebe G (eds) Advances in earthquake engineering for urban
informed and swift decisions regarding post-earthquake risk reduction. Nato science series: IV: Earth and environmental
assessment of critical instrumented building structures and sciences, vol 66. Springer, Dordrecht
17. Lenjani A, Bilionis I, Dyke SJ, Yeum CM, Monteiro R (2020) A
to improve earthquake resiliency of communities. resilience-based method for prioritizing post-event building inspec-
tions. Nat Hazards 100(2):877–896. https​://doi.org/10.1007/s1106​
Acknowledgements  Support for this research provided, in part, by 9-019-03849​-0
award No. 1453502 from the National Science Foundation is grate- 18. Mohsen A, Pekcan G (2020) Structural health monitoring using
fully acknowledged. extremely compressed data through deep learning. Comput-Aided
Civ Infrastruct Eng 35:597–614
19. Pan H, Azimi M, Gui G, Yan F, Lin Z (2017) Vibration-based sup-
port vector machine for structural health monitoring. In: Interna-
tional conference on experimental vibration analysis for civil engi-
References neering structures, pp 167–178. Springer, Dordrecht
20. Mangalathu S, Sun H, Nweke CC, Yi Z, Burton HV (2020) Classify-
1. ATC (1989) Procedures for postearthquake safety evaluations of ing earthquake damage to buildings using machine learning. Earthq
buildings, report ATC-20. Technical report, Applied Technology Spectra 36(1):183–208
Council (ATC), Redwood City, CA 21. Zhang Y, Burton HV, Sun H, Shokrabadi M (2018) A machine learn-
2. ATC (1995) Addendum to the ATC-20 postearthquake building ing framework for assessing post-earthquake structural safety. Struct
safety evaluation procedures. Technical report, Applied Technol- Saf 72:1–16
ogy Council (ATC), Redwood City, CA 22. Lenjani A, Dyke SJ, Bilionis I, Yeum CM, Kamiya K, Choi J, Liu
X, Chowdhury AG (2020) Towards fully automated post-event

123

data collection and analysis: pre-event and post-event informa-
tion fusion. Eng Struct 208:109884. https​://doi.org/10.1016/j.engst​
ruct.2019.10988​4
23. Roohi M (2019) Performance-based seismic monitoring of instru-
mented buildings. PhD thesis, Graduate College Dissertations and
Theses. 1140. University of Vermont
24. Wu RT, Jahanshahi MR (2020) Data fusion approaches for struc-
tural health monitoring and system identification: past, present, and
future. Struct Health Monit 19(2):552–586
25. Sohn H, Farrar CR, Hemez FM, Shunk DD, Stinemates DW, Nadler
BR, Czarnecki JJ (2003) A review of structural health monitoring
literature: 1996–2001. Los Alamos National Laboratory, USA, pp
1–7
26. Fan W, Qiao P (2011) Vibration-based damage identification
methods: a review and comparative study. Struct Health Monit
10(1):83–111
27. Azimi M, Eslamlou AD (2020) Data-driven structural health moni-
toring and damage detection through deep learning: state-of-the-art
review. Sensors 20(10):2778
28. Celebi M, Sanli A, Sinclair M, Gallant S, Radulescu D (2004) Real-
time seismic monitoring needs of a building owner—and the solu-
tion: a cooperative effort. Earthq Spectra 20(2):333–346
29. Miranda E (2006) Use of probability-based measures for automated
damage assessment. Struct Des Tall Spec Build 15(1):35–50
30. Porter K, Mitrani-Reiser J, Beck JL (2006) Near-real-time loss
estimation for instrumented buildings. Struct Des Tall Spec Build
15(1):3–20
31. Mitrani-Resier J, Wu S, Beck JL (2016) Virtual inspector and its
application to immediate pre-event and post-event earthquake loss
and safety assessment of buildings. Nat Hazards 81(3):1861–1878
32. Hwang SH, Lignos DG (2018) Assessment of structural damage
detection methods for steel structures using full-scale experimental
data and nonlinear analysis. Bull Earthq Eng 16(7):2971–2999
33. Cremen G, Baker JW (2018) Quantifying the benefits of building
instruments to FEMA p-58 rapid post-earthquake damage and loss
predictions. Eng Struct 176:243–253
34. Hernandez E, Roohi M, Rosowsky D (2018) Estimation of element-
by-element demand-to-capacity ratios in instrumented SMRF
buildings using measured seismic response. Earthq Eng Struct Dyn
47(12):2561–2578
35. SEAOC (1995) Vision 2000: performance based seismic engineer-
ing of buildings. Structural Engineers Association of California,
Sacramento
36. ATC (1996) Seismic evaluation and retrofit of concrete buildings.
2. Appendices. Applied Technology Council (ATC), Redwood City,
CA
37. FEMA (1997) NEHRP guidelines for the seismic rehabilitation of
buildings. FEMA-273, Federal Emergency Management Agency,
Washington, DC
38. FEMA (2000) Commentary for the seismic rehabilitation of build-
ings. FEMA-356, Federal Emergency Management Agency, Wash-
ington, DC
123
Journal of Civil Structural Health Monitoring
39. ASCE (2013) ASCE/SEI 41-13: seismic evaluation and retrofit
of existing buildings. Technical report, American Society of Civil
Engineers, ASCE/SEI 41–13, Reston, VA
40. Porter KA (2003) An overview of PEER’s performance-based earth-
quake engineering methodology. In: Proceedings of ninth interna-
tional conference on applications of statistics and probability in civil
engineering. Citeseer
41. Gelb A (1974) Applied optimal estimation. MIT press, Cambridge
42. Julier S, Uhlmann J, Durrant-Whyte HF (2000) A new method for
the nonlinear transformation of means and covariances in filters and
estimators. IEEE Trans Autom control 45(3):477–482
43. Doucet A, Godsill S, Andrieu C (2000) On sequential monte carlo
sampling methods for bayesian filtering. Stat Comput 10(3):197–208
44. Hernandez EM (2011) A natural observer for optimal state estima-
tion in second order linear structural systems. Mech Syst Signal
Process 25(8):2938–2947
45. Hernandez EM (2013) Optimal model-based state estimation in
mechanical and structural systems. Struct Control Health Monit
20(4):532–543
46. Roohi M, Hernandez EM, Rosowsky D (2019) Nonlinear seismic
response reconstruction in minimally instrumented buildings—vali-
dation Using Neeswood capstone full-scale tests. Structural health
monitoring 2019. Presented at the structural health monitoring 2019.
https​://doi.org/10.12783​/shm20​19/32390​
47. FEMA-356 (2000) Prestandard and commentary for the seismic
rehabilitation of buildings. American Society of Civil Engineers
(ASCE), Reston
48. Krawinkler H (2005) Van Nuys hotel building testbed report: exer-
cising seismic performance assessment. Pacific Earthquake Engi-
neering Research Center, College of Engineering of California
49. Trifunac MD, Ivanovic SS, Todorovska MI (1999) Instrumented
7-storey reinforced concrete building in Van Nuys, California:
description of the damage from the 1994 Northridge earthquake
and strong motion data. Report CE 99:2
50. Trifunac MD, Ivanovic SS (2003) Analysis of drifts in a seven-story
reinforced concrete structure. University of Southern California
Report CE, pp 3–10
51. Frank M, Fenves GL, Scott MH et al (2000) Open system for earth-
quake engineering simulation. University of California, Berkeley
52. Jalayer F, Ebrahimian H, Miano A, Manfredi G, Sezen H (2017)
Analytical fragility assessment using unscaled ground motion
records. Earthq Eng Struct Dyn 46(15):2639–2663
53. Saiful Islam M (1996) Analysis of the northridge earthquake
response of a damaged non-ductile concrete frame building. Struct
Des Tall Build 5(3):151–182
Publisher’s Note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.