Sunteți pe pagina 1din 13

Engineering Structures 24 (2002) 10911103

www.elsevier.com/locate/engstruct

Theory of principal components analysis and applications to


multistory frame buildings responding to seismic excitation
Mark A. Aschheim a,, Edgar F. Black b, Isabel Cuesta c
a

Mid-America Earthquake Center, Department of Civil and Environmental Engineering, University of Illinois at Urbana-Champaign, 205 North
Mathews, Urbana, IL 61801, USA
b
Department of Civil Engineering, University of Notre Dame, 159 Fitzpatrick Hall of Engineering, University of Notre Dame, Notre Dame, IN
46556, USA
c
Los Alamos National Laboratory, P.O. Box 1663, MS C926, LANL ESA-WR, Los Alamos, NM 87545, USA
Received 9 February 2001; received in revised form 3 October 2001; accepted 29 October 2001

Abstract
Described herein is a technique of multivariate statistical analysis applied to the post-processing of dynamic response data. The
data may represent the linear or nonlinear response of structures, and may be obtained from computed simulations or from the
measured response of instrumented structures. When applied to displacement response data, an ordered set of orthonormal mode
shapes is obtained. The principal components analysis (PCA) mode shapes coincide with or are related to the elastic mode shapes
for linear elastic systems, and depart from these shapes as nonlinear response becomes more prominent. The PCA modes provide
an unambiguous and simple description of the predominant mode of structures responding to earthquake ground motions, and
thus improve the theoretical basis of nonlinear static procedures that use equivalent single-degree-of-freedom (SDOF) systems
for representing the response of structures subjected to damaging earthquake ground motions (e.g. the capacity spectrum and displacement coefficient methods). Where greater fidelity is desired, the most efficient representations are obtained by including as
few PCA modes as are needed for the degree of precision desired. This paper presents the theory of PCA and illustrates its
application to a 12-story frame building responding linearly and nonlinearly to earthquake ground motions. Equivalent SDOF
models of the structure are developed based on the PCA mode shapes, and these are applied to estimate the computed displacement
histories. 2002 Elsevier Science Ltd. All rights reserved.
Keywords: Dynamic response; Mode shapes; Nonlinear static procedures; Principal components analysis; Equivalent single-degree-of-freedom systems; Seismic design

1. Introduction
Linear and nonlinear dynamic response can generate
large amounts of time series data that may be difficult
to interpret, particularly when structures respond nonlinearly. Although various measures such as displacement
ductility and dissipated hysteretic energy are useful indicators of inelastic demand, only qualitative descriptions
of the displacement response have been available.
Researchers have referred to the predominant mode of
response of a nonlinear system [15], but have been

Corresponding author. Tel.: +1-217-333-7384; fax: +1-217-2658040.


E-mail address: aschheim@uiuc.edu (M.A. Aschheim).

unable to describe with specificity what this mode


shape is or the degree to which response is in this mode.
Principal components analysis (PCA) identifies these
mode shapes precisely and quantifies the degree to which
response is in each mode.
Principal components analysis determines an
orthonormal set of basis vectors that best represent the
response data. The ordering is such that the first basis
vector (hereafter referred to as a PCA mode) is the best
fit to the observed response history. The second PCA
mode is the best fit to the residuals, and so on, for each
mode. When applied to the displacement response of linear systems that have uniform values of nodal mass, the
mode shapes are observed to coincide with the elastic
mode shapes. For nonuniform mass distributions, the
PCA modes and elastic modes are related according to

0141-0296/02/$ - see front matter 2002 Elsevier Science Ltd. All rights reserved.
PII: S 0 1 4 1 - 0 2 9 6 ( 0 2 ) 0 0 0 3 6 - 6

1092

M.A. Aschheim et al. / Engineering Structures 24 (2002) 10911103

relationships that are described herein. The empiricallydetermined mode shapes may deviate from those
obtained for elastic response as response becomes nonlinear, and depend (indirectly) on properties of both the
structure and the excitation. The PCA mode shapes are
eigenvectors computed from the response data; the corresponding eigenvalues provide a quantitative measure
of the degree to which the response of the structure is
represented by each PCA mode shape. Thus, it is possible to identify the first mode of a structure responding
nonlinearly and to quantify the degree to which response
is represented by this mode. Explicit knowledge of the
properties of the structure (e.g. the mass and stiffness
matrices) is not needed for determination of the PCA
mode shapes; these properties are reflected in the displacement response history.
Since the PCA mode shapes are determined to maximize variances, they are natural choices to be used in
establishing parsimonious representations of the
response. Representations of the response that use the
first or first two PCA modal responses are preferable to
approximations in which responses based on the elastic
mode shapes are superposed, because a larger number
of elastic mode shapes may be necessary to represent
the displacement profile exhibited by systems
responding nonlinearly. The PCA mode shapes may be
useful for developing equivalentsingle-degree-of-freedom (SDOF) representations of the response of multidegree-of-freedom (MDOF) systems. The computed
PCA mode shapes portray the behavior of structures
being designed to resist earthquakes; thus, it is possible
in design to modify member strengths and stiffnesses
iteratively until the PCA mode shapes indicate that undesirable mechanisms do not develop in the structure. PCA
also is useful when modeling the response of instrumented structures, because model properties may be
constrained by requiring that PCA mode shapes obtained
from the computed data approximately match those
obtained from the recorded data. Examples of this application are given by Wissawapaisal and Aschheim [21]
and Inel [13], who studied the seismic response of instrumented bridges. Independently, Gutie rrez and Zaldivar
[10] recently applied PCA to data collected in pseudodynamic and shake table tests.
The idea of PCA originates in work by Pearson [18]
and was developed independently by Hotelling [11].
PCA finds modern use in statistical treatments of empirical data [16,17,20]. This paper presents the theoretical
basis for the technique using the nomenclature of structural engineering and applies the technique to the linear
and nonlinear responses of a 12-story moment-resistant
frame building. The effectiveness of the PCA representations of lateral forces and lateral displacements is compared, and equivalent SDOF (ESDOF) systems based
on the predominant PCA mode are used to estimate the
roof displacement history.

2. Theory of PCA
Consider a vector of displacement response data v representing the displacement relative to the ground at n
discrete locations of a MDOF system at an instant of
time. There are t observations of the n1 vector v over
time. It is obvious that the displacement response, v, at
any instant of time represents a linear combination of
the unit basis vectors that provide an orthonormal basis
for the n-dimensional space that contains v. The deviation of v from its mean over t observations,v, may be
expressed in terms of a new orthonormal basis, B:
vv Bu

(1)

where u represents the displacements relative to their


means in terms of the basis vectors contained in B. Since
B is orthonormal, BTB I. Therefore, BT B1, and
premultiplying (1) by BT gives
u BT(vv)

(2)

with the mean of u being the n1 vector 0.


Let the covariance of the n1 vector v be represented
by the nn covariance matrix Cv. The (i, j) element of
Cv is the covariance between the displacements at the
ith and jth degrees of freedom, vi and vj, over the t observations, given by

cov(vi,vj)

1
(v v )(v v )
t k 1 i,k i j,k j

(3)

where vi and vj are the means of vi and vj, respectively,


over the t observations. Because cov(vi, vj) is computed
using real-valued data, Cv is real, and because
cov(vi,vj) cov(vi,vj), Cv is symmetric.
Because Cv is real and symmetric, it can be diagonalized by an orthogonal matrix
TCv

(4)

where consists of the orthonormal eigenvectors of Cv


and is a diagonal matrix containing the eigenvalues
li of Cv. By convention, the eigenvectors of Cv are
arranged in sequence so that their corresponding eigenvalues are in descending order (l1l2ln).
By operating with standard identities on Eq. (2), the
covariance of u, Cu, can be expressed in terms of Cv:
Cu BTCvB.

(5)

The parallel structures of Eqs. (4) and (5) indicate that


Cu is a diagonal matrix consisting of the eigenvalues of
Cv if the new basis B is selected to be the set of
orthonormal eigenvectors of Cv.
As Cu is diagonal, cov(ui,uj) 0 for i j, leading to
the result that the displacements ui and uj (expressed in
terms of the orthonormal basis B ) are uncorrelated
for i j. Similarly, each displacement ui has variance
var(ui) cov(ui,ui) li.

M.A. Aschheim et al. / Engineering Structures 24 (2002) 10911103

In subsequent sections it will be shown that:

a2i li li.

(10)

Similarly, the principal components are uncorrelated


because all covariances (between different components)
are identically 0. Because is a basis for the space that
contains the deviation of v from its mean, eqs. (1) and
(2) provide that v can be expressed as a linear combination of the eigenvectors of Cv
v u v

(11)

where

2.1. Proof that the eigenvectors are principal


components

u T(vv ).

To see that the eigenvectors of Cv are the principal


components of the displacement history, consider the linear combination aTv that has maximum variance. The
weights a are selected to maximize var(aTv), under the
constraint that a has unit length, or aTa 1. Because
is orthonormal and is a basis for the displacement
space, a can be expressed as a linear combination of the
basis vectors
(6)

By applying a standard identity,

(12)

Eq. (11) represents the displacement response exactly,


and thus the entire set of n eigenvectors represents 100%
of the variance in v, which is given by tr(). Because
the bases are orthonormal, tr() tr(Cv), which indicates that the total variance is a constant for any given
data set. A subset of principal components may be used
to obtain efficient representations of v. Because the eigenvectors are sequenced according to their eigenvalues,
the most efficient representations of v using k eigenvectors will be obtained if the first k eigenvectors are selected:

var(a v) a Cva.
T

var(aTv) var(fTi v)

1. the eigenvectors of Cv are, in fact, the principal


components of the displacement response v (in other
words, f1 is so oriented that it maximizes the variance, f2 maximizes the variance with f1 removed, and
so on); and
2. the variance in the displacement response explained
by each principal component is proportional to the
eigenvalue associated with the principal component.

a a.

1093

(7)

fiui v .

(13)

Eqs. (4) and (5) result in

The proportion of total variance represented by the k


components used in eq. (13) is (l1 l2
lk) / tr(). Because the principal components are orthogonal, the coefficients yi are invariant with changes in k,
and because they have unit length, their values are
determined by the simple dot product given in (12). The
number of eigenvectors to be used depends on the
desired level of precision. Since each eigenvector
explains a proportion of the variance given by
(li) / tr() inspection of quickly reveals the value
added by additional components. The authors experience with civil engineering structures responding to
earthquake excitations is that the first PCA mode often
is sufficient to represent 90% or more of the variance in
the displacement response, while the first two PCA
modes often represent 9599% of the variance in the
displacement response.
Three unusual cases may occur:

2.2. The proportion of response explained by each


eigenvector

1. the number of observations, t, is less than the degrees


of freedom, n;
2. one or more of the displacement degrees of freedom
is a linear combination of the other displacement
responses; and
3. not all eigenvalues are distinct (that is, one or more
eigenvalues are repeated).

Cv BCuB .
T

(8)

Substituting (6) and (8) into (7) leads to

var(aTv) aTTTa aTa

a2i li.

(9)

The constraint aTa 1 is equivalent to aTa 1, so


a2i 1. Then, var(aTv) is maximized by picking a to
be the unit vector that returns f1 to a (a1 1 and
aj 0 for all j 1). Thus, fT1v maximizes the variance,
and therefore f1 is the first principal component. Subsequent combinations to maximize the variance should
be uncorrelated with those that have already been taken.
Since the first k principal components (f1, f2,fk) are
uncorrelated, the (k+1)th principal component to maximize var(aTv) must select the largest remaining eigenvalue, lk+1, for 1k+1n. Thus, the (k+1)th principal
component is given by fk+1.

The variance explained by any eigenvector fi is its


eigenvalue li, since fi is selected by choosing an appropriate unit vector a in Eq. (6), for which Eq. (9) provides

In the first two cases, the rank of Cv will be reduced to

1094

M.A. Aschheim et al. / Engineering Structures 24 (2002) 10911103

a value less than n, and the number of nonzero eigenvalues will be reduced correspondingly. Meaningful characterization of dynamic response ordinarily will require t
to be substantially larger than n, so the first case usually
is not encountered. In the second case, a reduced basis
consisting of the eigenvectors associated with the nonzero eigenvalues will be sufficient to represent the
response using Eq. (12). In the third case, the eigenvectors associated with the repeated roots are not unique,
yet remain orthogonal to the other eigenvectors. Thus,
they may be chosen arbitrarily to be orthonormal to
themselves and will be orthonormal to the other eigenvectors.
All data presented in this paper are of the same type
(i.e. displacements, forces or accelerations). When mixed
data types are present (e.g. rotations and displacements),
either the data should be scaled to be of similar numeric
value, or formulations based on correlations may be
preferable to the present covariance formulation.

limit, coincide with the transformed elastic mode shapes


given by LTel.
Two special cases are of interest:

2.3. Relationship between the PCA and the elastic


mode shapes

3. Principal components in the response of a 12story building

A subsequent example illustrates that the PCA mode


shapes coincide with the elastic mode shapes for linear
elastic structures having uniform nodal masses, provided
that the data are obtained at small enough time
increments over a duration of sufficient length. This
observation is discussed for general distributions of mass
as follows.
The free vibration eigenproblem of elastic dynamics,
given by Kfel lMfel, may be transformed to a standard form [2] by introducing a Cholesky factorization of
the mass matrix given by M LLT. The entries of L,
Lij, are zero for i j. This factorization allows the elastic
lf
,
f
dynamics eigenproblem to be restated as K
L1K(LT)1. The transformed elastic mode
where K
LTfel, are orthogonal to one
shapes, given by f
another while satisfying the orthogonality relationships
required of the elastic mode shapes:

In the following, PCA is applied to the computed


dynamic response of a 12-story moment-resistant steel
frame building subjected to the 1940 NS El Centro earthquake ground motion. Linear elastic response is considered as well as nonlinear response involving either
beam-hinging or weak-story mechanisms. Estimates of
the displacement response are presented, based on
ESDOF models that are established using the predominant mode shapes determined by PCA.

j fTel, iLLTfel,j fTel, iMfel,j dij


Tif
f

(14)

where dij 1 for i j and dij 0 otherwise.


Because the elastic modal responses are independent
and, in general, are uncorrelated with one another, the
displacements u i and u j expressed in terms of the transformed mode shapes LTel will be uncorrelated, for
closely-spaced elastic response data of sufficient duration. Therefore, the off-diagonal terms in Cu will tend
to zero as t increases. If the principal components are
determined for transformed displacements v LTv,
then Cu is a diagonal matrix provided that the displacements u are expressed in terms of the set of orthonormal
basis vectors that are the eigenvectors of Cv . The PCA
mode shapes obtained for LTv exhibit the orthogonality
required of the elastic mode shapes [Eq. (14)], and, in the

1. If M is diagonal with uniform masses, M mI


LLT, requiring that L LT mI. Because this
reduces LTv and LTel to scalar operations on v and
el, respectively, the PCA mode shapes obtained on
v will coincide with the elastic mode shapes after normalization.
2. If M is diagonal with entries mi, M LLT requires
that L LT with the ith diagonal entry of L given
by mi. For this case, the PCA mode shapes obtained
on LTv will coincide with the transformed elastic
mode shapes given by LTel. Thus, the elastic mode
L1F
, proshapes are given by el (LT)1F
vided that sufficient response data v is available and
response is linear.

3.1. Building description and modeling


The 12-story steel moment-resistant frame building
(Fig. 1) was designed using the uniform building code
[12] distribution of lateral forces for uniform floor
masses equal to 551 kN per floor. The base shear
strength was established to limit drift response; the
design is described in more detail as the Flexible-12
frame in Black and Aschheim [3]. The frame was
designed only for lateral loads in order to validate a
design methodology. Lateral response was computed
using DRAIN-2DX [19]. Flexural response was modeled
using beam-column elements (Type 02) extending along
beam and column centerlines; the post-yield stiffness
was set equal to 10% of the initial stiffness.
3.2. Linear elastic response
Elastic response was obtained by subjecting the building to the 1940 NS El Centro ground motion with acceleration amplitudes scaled by 0.5. The computed displacement of the roof relative to the ground is shown in
Fig. 2; the peak of 0.1682 m occurred at 5.61 s. The first

M.A. Aschheim et al. / Engineering Structures 24 (2002) 10911103

1095

Fig. 3. Mode shapes obtained from modal and PCAs coincide for
large data sets, for structures responding elastically with uniform
mass distributions.

Table 1
Modal response data for elastic response to El Centro
Mode Viscous
Natural Spectral
Proportion Cumulative
damping period displacement of variance proportion
(percent of (s)
(m)
(%)
of variance
critical
damping)
1
2
3

Fig. 1. Framing configuration and steel member sizes.

Fig. 2. Computed roof displacement history for elastic response to


the El Centro ground motion.

three elastic mode shapes are shown by solid lines in


Fig. 3. The PCA modes were computed using 80 s of
displacement response data at 0.01-s intervals, and are
plotted in the same figure using circles. The PCA mode
shapes computed for this relatively large data set are
seen to coincide with the elastic mode shapes. Table 1

5.00
2.77
2.82

2.168
0.798
0.470

0.1081
0.0483
0.0295

96.85
2.87
0.26

96.85
99.72
99.98

summarizes elastic periods and spectral response data for


the first three elastic modes, along with the modal damping resulting from the use of mass and stiffness proportional damping in DRAIN-2DX. The USEE 2001
computer program [14] was used to compute the spectral
displacements of Table 1, for the scaled record and
damping present in the dynamic analyses. Because the
PCA mode shapes coincided with the elastic mode
shapes, the proportion of variance represented by these
modes, given by li /tr() is shown on the same lines
of the table. The elastic spectral data characterize peak
quantities while the variance data characterize the degree
to which each mode represents the displacement
response over the duration of response. It can be
observed that 97% of the variance in the displacement
response is represented by the first mode, while the first
two modes represent nearly 100% of the variance in the
displacement response history.
Based on the prominence of the first mode in the displacement response and the success that other investigators have had in using ESDOF models, one may
speculate that an ESDOF system may yield accurate estimates of the peak displacement and the displacement
response history. Using the ESDOF model of ATC-40
[1] with the principal component renormalized to unit
amplitude at the roof results in 1 1.372, where 1 is

1096

M.A. Aschheim et al. / Engineering Structures 24 (2002) 10911103

Fig. 4. Comparison of roof displacement response computed for the


MDOF system and estimated with the ESDOF system, for elastic
response.

a term that is often described at the participation factor


of the first mode [5]. Therefore, the roof displacement is
estimated to be 1.372 times the displacement of a SDOF
oscillator having a period of 2.168 s and viscous damping equal to 5% of critical damping. Fig. 4 compares the
estimated roof displacement history with values computed for the MDOF system. The estimated peak
response can be obtained from Table 1 as 1.372(0.1081
m)=0.148 m, a value that is 11.8% less than the computed peak of 0.1682 m. This difference is attributed to
the influence of higher modes.
To better understand the dynamic response of the system and the role of higher modes on response, PCA may
be applied to smaller intervals of the response history.
Advancing the interval over time results in a moving
window analysis. Such an analysis identifies how well
the various modes represent the response as a function
of time as well as possible changes in the mode shapes
with time. A 5-s window was advanced in 1-s increments
over the displacement response data. The proportion of
variance explained by each of the three components is
plotted in Fig. 5, with values plotted at the trailing edge

Fig. 5. The proportion of variance and cumulative proportion of variance in the displacement response represented by the first three principal components, obtained in a moving window analysis of elastic
response data.

of the 5-s window. The solid line indicates that response


after about 38 s is nearly entirely in the first mode. The
first two modes are sufficient to represent nearly all the
response for 5-s intervals that begin during the first 20
s. The third mode plays a significant role in the response
for 5-s intervals that begin between 20 and 35 s. Three
or fewer modes are seen to be sufficient to represent
nearly 100% of the variance in the displacement
response over the entire history. Since it is possible that
the principal components may rotate (while maintaining
orthonormality) as the moving window advances, the
mode shapes are plotted in Fig. 6 together with the elastic mode shapes. Fig. 6 indicates that the principal
components deviate little from the elastic mode shapes
during most intervals. Therefore, the changes in the predominance of the mode shapes over time in Fig. 5 relate
primarily to the excitation of higher modes rather than
substantial changes in the mode shapes themselves. This
can be expected for elastic response. The increased prominence of the second and third mode contributions at
about 28 s (Fig. 5) causes the roof to oscillate more rapidly from about 28 to 37 s (Fig. 2). Since the roof displacements during this interval are relatively small, it
is clear that the higher modes do not cause large roof
displacements in and of themselves, for this excitation.
The PCA technique also may be applied to the
dynamic story shears and to the lateral forces that equilibrate these story shears in a static analysis. Fig. 7 shows
the first three PCA modes of the lateral force distribution, computed in a moving window analysis over 80s of response data, along with the lateral force distributions that correspond to the elastic mode shapes. The
lateral force distribution represented by the first PCA
mode varies over time, sometimes resembling a first
mode distribution and other times resembling the second
or third elastic modal force distributions. The proportion
of variance in the lateral forces explained by each PCA
mode over 80 s of response data is presented in Table
2. It can be observed that all three PCA modes contribute

Fig. 6. Principal components of the displacement response obtained


in a moving window analysis of elastic response data.

M.A. Aschheim et al. / Engineering Structures 24 (2002) 10911103

1097

Table 2
Efficiency of modal representations of various quantities (elastic
response)

computed over 80-s of response data, is given in Table


2. The first PCA mode represents a greater proportion
of the modal shears than is the case for the equilibrating
lateral forces. Table 2 and Figs. 6 8 indicate that the
lateral forces and to a lesser extent the story shears vary
significantly relative to the first mode even though the
displacements are represented well by the first mode.
This may be explained by noting that the lateral forces
associated with mode i are proportional to w2i Sd,ii, where
wi is the modal circular frequency and Sd,i is the spectral
displacement associated with the modal frequency. The
frequency content of the El Centro motion is such that
the spectral displacements associated with the higher
modes are small, and thus the first PCA mode is sufficient to represent nearly all of the displacement
response. However, the squares of the higher mode frequencies elevate the lateral force contributions sufficiently that the first three PCA modal force distributions are needed to represent the variance in the lateral
forces with a similar degree of precision.

Quantity

3.3. Nonlinear response

Fig. 7. Principal components of the lateral forces required to statically equilibrate the story shears computed by dynamic analysis,
obtained in a moving window analysis of elastic response data.

Lateral forces
Story shears
Lateral
displacements

Proportion of variance
(%)

Cumulative proportion
of variance (%)

1st
mode

2nd
mode

3rd
mode

1st
mode

2nd
mode

3rd
mode

37.27
80.80
96.85

28.20
14.62
2.87

22.88
3.87
0.26

37.27
80.80
96.85

65.47
95.42
99.72

88.35
99.29
99.98

significantly to the lateral forces required to equilibrate


the story shears computed in the dynamic analysis.
A moving window analysis of the story shears resulted
in the modal shears shown in Fig. 8. These are seen to
vary somewhat from the elastic modal shears, particularly for some 5-s intervals. The proportion of variance in the story shears represented by each PCA mode,

Fig. 8. Principal components of the dynamic story shears, obtained


in a moving window analysis of elastic response data.

The original frame was provided with relatively strong


columns such that mechanisms develop that involve
beam plastic hinging. This section presents two cases of
nonlinear response to the El Centro ground motion. The
first case develops beam-hinging response, consistent
with the original design. In the second case, the column
flexural strengths in the first story were artificially weakened to induce a weak-story response to develop. For
both cases the El Centro ground motion was applied with
amplitude scaled by a factor of 2.
3.3.1. Beam-hinging response
The computed nonlinear response of the beam-hinging
frame resulted in the roof displacement response history
shown in Fig. 9. The peak roof displacement of 0.5966
m occurred at 5.64 s. Fig. 10 shows the first three elastic
and PCA mode shapes. The response was intermittently
inelastic, and the PCA mode shapes computed over the

Fig. 9. Roof displacement history of the beam-hinging frame.

1098

M.A. Aschheim et al. / Engineering Structures 24 (2002) 10911103

Fig. 10. The elastic (solid line) and PCA mode shapes (dashed line
with circles) are nearly the same even for intermittent inelastic
response of the beam-hinging frame, when computed over the 80-s
duration of response data.

time, and Fig. 12 shows the PCA mode shapes obtained


for each window, together with the elastic mode shapes.
The large departures from the third elastic mode shape
in Fig. 12 occurred at the end of the record at a time
that virtually all of the response was represented by the
first two principal components. Comparison of Figs. 5
and 11 indicates a small reduction in the variance
explained by the first principal component during the
early portion of the response when the largest roof displacements occurred. The variance explained by the first
two principal components is nearly the same for both
cases. Comparison of Figs. 6 and 12 reveals little difference in the PCA mode shapes, even though moderate
inelastic response developed in the beam-hinging case.
In the present case, the elastic mode shapes were considered adequate approximations of the predominant
mode shapes, based on Fig. 10. Nonlinear static
(pushover) analyses were done by applying lateral forces
in proportion to the first mode shape amplitude and mass
at each floor level. A bilinear curve was fitted to the
capacity curve (Fig. 13) to determine the yield strength
and displacement for response in the first mode. The displacement of the roof at yield is 0.353 m, or 0.72% of
the height of the building, and the base shear coefficient
at yield is 0.173. The post-yield stiffness is 17.5% of the
initial stiffness. The elastic mode shape and uniform
mass distribution result in 1 1.372, following the
ATC-40 [1] formulation. The yield strength coefficient

Fig. 11. The proportion of variance and cumulative proportion of


variance represented by the first three principal components, obtained
in a moving window analysis of the beam-hinging frame response data.

80-s duration of response data are nearly equal to the


elastic mode shapes. Because the elastic and PCA mode
shapes are nearly identical, results for each mode are
tabulated together in Table 3. The spectral displacements
of Table 3 are the peak elastic displacements computed
for the scaled record, using the viscous damping present
in the nonlinear dynamic analysis.
A moving window PCA analysis was done using a 5s window advanced in 1-s increments. Fig. 11 shows the
proportion of variance represented by each mode over

Fig. 12. The principal components obtained in a moving window


analysis of the displacement response data tend to follow the elastic
mode shapes, even for beam-hinging responses.

Table 3
Response data for the beam-hinging frame
Mode

Viscous damping (percent of Period (s)


critical damping)

Spectral
displacement (m)

Peak displacement Proportion of


(m)
variance (%)

Cumulative proportion
of variance (%)

1
2
3

5.00
2.77
2.82

0.4323
0.1930
0.1181

0.4207
0.1930
0.1181

96.75
99.58
99.96

2.168
0.798
0.470

96.75
2.84
0.37

M.A. Aschheim et al. / Engineering Structures 24 (2002) 10911103

Fig. 13. Capacity curve obtained by applying lateral forces proportional to the product of the elastic modal amplitude and mass at
each floor in a nonlinear static analysis.

of the ESDOF system, Cy, determined by the ATC-40


method is 0.220. This value may be determined more
directly by:
Cy

2p 2xy
T1 g

(15)

where xy is the yield displacement of the ESDOF system


and T1 is the fundamental period of vibration of the
MDOF system. When the ESDOF system is determined
with a mode shape that differs from the elastic mode
shape, Eq. (15) ensures the initial period of the ESDOF
system matches the fundamental period of the MDOF
system. (The periods obtained with the ATC-40 [1] and
FEMA [7,8] formulations do not, in general, match the
fundamental period of the MDOF system.) Roof displacement histories were estimated based on the potentially nonlinear response of the analogous SDOF oscillator. The estimated roof displacement history is
compared with the roof displacement history computed
in the nonlinear response of the MDOF system in Fig.
14. It can be observed that the response history estimated

Fig. 14. Comparison of roof displacement histories computed for the


beam-hinging MDOF system and estimated using an ESDOF system
based on the first mode.

1099

using the ESDOF system captures the essence of the roof


displacement response computed for the MDOF system.
The peak displacements of the ESDOF systems corresponding to the first three modes are reported in Table
3. The 1st mode ESDOF system reached a peak ductility
of m 1.64 and peak displacement of 0.4207 m; the
estimate of peak roof displacement due to the first mode
is 1.372(0.4207 m)=0.5772 m. The contribution of the
second mode to the peak roof displacement is estimated
as 0.5682(0.1930)=0.1097 m, based on 2 0.5682,
although its timing may differ from the timing of the
first mode peak. An SRSS combination results in a peak
roof displacement estimate of (0.57722 0.10972)0.5
0.5875m, nearly equal to the first mode estimate. Relative to the computed peak displacement of 0.5966 m, the
first mode estimate is 3.3% low and the SRSS estimate
for the first two modes is 1.5% low.
3.3.2. Weak-story response
To introduce a weak-story response, the frame was
artificially modified by reducing the strengths of the lowest story columns to 25% of their original values. The
elastic properties were not changed; hence all building
frames described in this paper have the same elastic
mode shapes and natural periods. The weak-story model
was subjected to the same ground motion record used
for the beam-hinging frame. The computed roof displacement history is shown in Fig. 15. Peak roof displacements of 0.4011 and 0.3998 m occurred at 4.69
and 5.69 s, respectively, and a small permanent displacement remained at the end of the response.
Fig. 16 plots the PCA mode shapes computed for 80
s of displacement response data. Although a weak-story
mechanism was not apparent in the displaced shape at
the end of the response, the PCA mode shapes clearly
identify the intermittent weak story responses. The variance in the displacement response represented by the
first three modes is given in Table 4. While the mode
shapes are clearly different from the elastic and beamhinging cases, the first and second modes were able to
represent similar proportions of the variance. The pro-

Fig. 15.

Roof displacement history of the weak-story frame.

1100

M.A. Aschheim et al. / Engineering Structures 24 (2002) 10911103

Fig. 16. PCA mode shapes computed over the 80-s response clearly
indicate weak story behavior.

Table 4
Response data for the weak-story frame
PCA mode

1
2
3

Proportion of
variance (%)
96.57
2.75
0.56

Cumulative proportion of
variance
96.57
99.33
99.89

portion of variance explained by each mode is plotted


cumulatively as a function of time in Fig. 17, based on
a moving window analysis (a 5-s window advanced in
1-s increments). Comparison to Fig. 11 reveals overall
similarities. The corresponding mode shapes obtained in
the moving window analysis are plotted in Fig. 18. Weak
story behavior is evident in the first three modes, particularly for the 5-s intervals that begin during the first 12
s. Peak roof displacements occurred at 4.69 and 5.69 s,
which are approximately centered in the 5-s windows
that begin between 2 and 3 s. The first, second, and third
mode shapes appear not to change significantly in this

Fig. 17. The proportion of variance and cumulative proportion of


variance represented by the first three principal components in a moving window analysis of the weak-story frame response data.

Fig. 18. The first three PCA mode shapes, obtained in a moving window analysis of the weak-story frame displacement response data.
Weak story behavior is clearly evident for 5-s intervals that begin in
the first 12 s of response.

interval. The mode shapes at 3 s are plotted in Fig. 19,


and these mode shapes are used subsequently to estimate
response based on an ESDOF model.
The data of Fig. 19 indicate that response in the first
mode changes from the elastic mode shape at low amplitudes to become dominated by a weak-story mechanism
as displacements increase. To represent such a response,
the ideal ESDOF system would reflect this change in
predominant mode shape as displacements increase. Ideally, the capacity curve would be generated using an
adaptive load pattern [4,6,9] such that the deflected
shape of the system corresponds to the predominant
mode of response. The change in deflected shape affects
the mass that participates in the response; the lateral
resistance provided by the structure relative to the participating mass ideally would be used to define the loaddeformation response of the ESDOF system. These
theoretical considerations provide a context in which
decisions are made about how an ESDOF model should
be developed. Because ESDOF systems necessarily omit
(or oversimplify) interactions with higher modes, they

Fig. 19.
3 s.

The PCA mode shapes for the 5-s window beginning at

M.A. Aschheim et al. / Engineering Structures 24 (2002) 10911103

will always be approximate. The benefit of an ESDOF


analysis is the ease with which approximate information
can be generated, and to preserve this utility, analyses
reported here make use of an invariant lateral load pattern throughout the linear and nonlinear portions of
response. The lateral load patterns examined here are
based on the PCA mode shape, a displaced shape associated with the development of the weak-story mechanism, and the elastic mode shape. The yield strength
coefficient is determined using Eq. (15) to ensure that
the period of the ESDOF system matches the fundamental period of vibration of the MDOF system.
Because a nonlinear dynamic analysis had been done,
it was possible to determine the capacity curve (Fig. 20)
by applying lateral forces proportional to the first PCA
mode shape amplitude (Fig. 19) and mass at each floor.
A bilinear curve fit to the capacity curve had a roof displacement at yield of 0.140 m, base shear coefficient at
yield of 7.55%, and post-yield stiffness equal to 52% of
the initial stiffness. Using the ATC-40 [1] formulation
with the PCA mode shape results in 1 1.307. Therefore, the yield displacement of the ESDOF system is
0.140 / 1.307 0.107m. Eq. (15) requires the yield
strength coefficient of the ESDOF system, Cy, to be
9.14%. The viscous damping, as a percent of critical
damping, and post-yield stiffness, as a percent of initial
stiffness, associated with response of the MDOF system
in the first mode were preserved in the ESDOF system.
The peak displacement of the ESDOF system is 0.343
m; thus, the estimate of the peak roof displacement is
1.307(0.343) 0.448m. This value exceeds the computed peak response of 0.4011 m by 12%. Even so, the
estimated roof displacement history compares well to the
roof displacement history computed for the MDOF system, as shown in Fig. 21.
Alternative load patterns may be applied to generate
the capacity curve. If the weak-story mechanism has
been identified, a blind estimate of the displacement
response may be made by:

1101

Fig. 21. Estimated roof displacement history using the PCA mode
shape in conjunction with an analogous SDOF oscillator, and the
response computed by nonlinear dynamic analysis of the MDOF system.

1. applying lateral forces proportional to the floor


masses in the pushover analysis; and
2. using elastic modal properties (1 1.372 and
T1 2.17s) to convert between the ESDOF system
and the MDOF system.
This results in a roof displacement at yield of 0.118 m,
base shear coefficient at yield of 7.58%, and a peak roof
displacement of 1.372(0.322)=0.442 m. This estimate is
about as good as the one computed using the first PCA
mode but was made without first identifying the PCA
mode shape.
If lateral forces are applied proportional to the first
elastic mode, the roof displacement at yield is 0.153 m
and the base shear coefficient at yield is 7.47%. Using
the elastic mode shape and period to obtain the ESDOF
system results in a first mode estimate of peak roof displacement of 0.494 m, 23% higher than the computed
peak.

4. Conclusions
The theory of PCA was presented using the nomenclature of structural engineering. Examples illustrated the
application of the theory for a 12-story moment-resistant
frame building having uniformly-distributed lumped
masses, subjected to earthquake excitations. Linear
response and nonlinear responses characterized by
beam-hinging or weak-story mechanisms were considered. Displacement estimates were made using equivalent SDOF models, based on the principal components identified in the response data. The following
conclusions can be made.

Fig. 20. Capacity curve obtained by applying forces in proportion to


the amplitude of the first PCA mode (at 3 s) and mass at each floor.

1. The displacement response of MDOF structures subjected to base excitation may be characterized by a
predominant mode. The precise mode shape and the

1102

2.

3.

4.

5.

6.

M.A. Aschheim et al. / Engineering Structures 24 (2002) 10911103

degree to which the response is characterized by this


mode shape may be determined for a given excitation
by the analytical procedure known as PCA. This procedure is applicable to data obtained from structures
responding linearly or nonlinearly. The PCA mode
shapes identified in different intervals of time may
differ from one another.
In the cases investigated, nearly all of the variance in
the displacement response can be represented by one
or two principal components. The first mode contribution to the total variance in the displacement
response was observed to approach 100% towards the
end of the response. Lateral forces and story shears
vary more rapidly than do the displacements and often
require inclusion of a larger number of principal
components to achieve similar precision.
The PCA mode shapes coincide with the elastic mode
shapes, when they are determined for the displacement response of linear systems having uniform mass
distributions using data sets of sufficient size.
Relationships between the PCA and elastic mode
shapes were provided for nonuniform mass distributions. The sequencing of the PCA and elastic mode
shapes may differ, depending on the degree to which
response in a given mode is excited by the ground
motion.
The first PCA mode represented 95% of the variance of the displacement response for each of the 12story buildings and excitations studied, when the
analysis was done over an 80-s window. This was the
case despite the potential for significant higher modes
that might be expected for a frame building having a
fundamental period of vibration of 2.17 s. During
smaller intervals of the response, the relative contribution of the second and third modes was larger,
although this occurred when the roof displacements
were substantially below their peak values.
The beam-hinging frame resulted in predominant
mode shapes that were nearly equal to the elastic
mode shapes. An equivalent SDOF model based on
the first mode was useful for estimating the peak displacement response and roof displacement response
history. The predominant mode shapes of the weakstory frame were very different from the elastic mode
shapes and clearly indicated the development of a
weak-story mechanism. An equivalent SDOF model
based on the PCA mode shape determined in the 5s interval in which the peak roof displacement
occurred was useful for estimating the peak roof displacement of the weak-story frame. For these cases,
estimates of peak roof displacement based on the first
PCA mode were within 12% of the peak quantities
computed in nonlinear dynamic analyses of the
MDOF structures.
The PCA technique may be used to identify the development of weak-story mechanisms even when such

mechanisms are not apparent in the final displaced


shape of the structure. The PCA technique may be
useful for evaluating when the distribution of member
strengths and stiffnesses is adequate to prevent the
development of undesirable mechanisms.

Acknowledgments
The development of this paper involved a number of
people. Early applications of the method were by graduate students Andrew Jackson, Chatdanai Wissawapaisal,
and Mehmet Inel. Keith Hjelmstad and Ping Gu provided helpful comments to an early manuscript. The
careful review and comments provided by anonymous
reviewers are greatly appreciated. The support of the
National Science Foundation via a CAREER award
(CMS-9984830) to the first author is gratefully acknowledged. This work was supported in part by the Earthquake Engineering Research Centers Program of the
National Science Foundation under Award No. EEC9701785.

References
[1] ATC-40, Seismic evaluation and retrofit of concrete buildings.
Report No. ATC-40. Redwood City (CA): Applied Technology
Council; 1996.
[2] Bathe KJ. Finite Element Procedures in Engineering Analysis,
Englewood Cliffs (NJ): Prentice-Hall; 1982, p. 5739.
[3] Black EF, Aschheim M. Seismic design and evaluation of multistory buildings using yield point spectra. CD Release 00-04. University of Illinois (Urbana): Mid-America Earthquake Center;
2000.
[4] Bracci JB, Kunnath SK, Reinhorn AM. Seismic performance and
retrofit evaluation of reinforced concrete structures. ASCE Journal of Structural Engineering, American Society of Civil Engineers 1997;123:310.
[5] Chopra AK, Goel RK. Capacity-demand-diagram methods based
on
inelastic
design
spectrum.
Earthquake
Spectra
1999;15(4):63756.
[6] Elnashai AS. Advanced inelastic static (pushover) analysis for
seismic design and assessment. In: G. Penelis International Symposium on Concrete and Masonry Structures; 2000, p. 2334.
[7] FEMA-273. NERHP Guidelines for the Seismic Rehabilitation of
Buildings. Report FEMA-273. Federal Emergency Management
Agency; 1997.
[8] FEMA-274. NERHP Commentary on the Guidelines for the Seismic Rehabilitation of Buildings. Report FEMA-274. Federal
Emergency Management Agency; 1997.
[9] Gupta B, Kunnath SK. Spectra-based pushover procedure for
seismic evaluation. Earthquake Spectra 2000;16(2):36791.
[10] Gutie rrez E, Zaldivar JM. The application of Karhunen-Loe ve,
or principal component analysis method, to study the non-linear
seismic response of structures. Earthquake Engineering and
Structural Dynamics 2000;29:126186.
[11] Hotelling H. Analysis of a complex of statistical variables into
principal components. Journal of Educational Psychology
1933;24:417520.
[12] ICBO. Uniform building code. International Conference of Building Officials, Whittier, CA; 1997.

M.A. Aschheim et al. / Engineering Structures 24 (2002) 10911103

[13] Inel M. Displacement-based strategies for the performance-based


seismic design of short bridges considering embankment flexibility. Ph.D. thesis, University of Illinois, Urbana (in preparation).
[14] Inel M, Bretz E, Black E, Aschheim M, Abrams D. USEE 2001
utility software for earthquake engineering: program, report, and
users manual. CD Release 01-05. University of Illinois (Urbana):
Mid-America Earthquake Center. Available for download from
http://mae.ce.uiuc.edu/.
[15] Krawinkler H. Pushover analysis: why, how, when, and when not
to use it. In: Proceedings of the 65th Annual Convention, Structural Engineers Association of California, Oct. 16, Maui
(Hawaii), p. 1736; 1996.
[16] Mardia V K, Kent JT, Bibby JM. Multivariate analysis. London:
Academic Press, 1979.
[17] Dunteman GH. Principal components analysis. Quantitative

[18]
[19]

[20]
[21]

1103

applications in the social sciences, Series No. 07-069. Newbury


Park (CA): Sage, 1989.
Pearson K. On lines and planes of closest fit to systems of points
in space. Phil. 1901;2:55972.
Prakash V, Powell GH, Campbell S. Drain-2DX base program
description and user guide, v. 1.10, Report No. UCB/SEMM93/17. Berkeley: Department of Civil and Environmental Engineering, University of California; 1993.
Venables WN, Ripley BD. Modern Applied Statistics with SPlus. 2nd ed. Berlin: Springer, 1997
Wissawapaisal C, Aschheim M. Modeling the transverse response
of short bridges subjected to earthquakes. CD Release 00-05.
University of Illinois (Urbana): Mid-America Earthquake
Center; 2000.

S-ar putea să vă placă și