EARTHQUAKE ENGINEERING A N D STRUCTURAL DYNAMICS, VOL.

14, 841-859 (1986)

MODAL ANALYSIS FOR RANDOM VIBRATION OF HYSTERETIC

FRAMES

T.T. BABER
Department of Ciuil Engineering, School of Engineering and Applied Science, University of Virginia, Charlottesuille, V A . 22901, U.S.A

SUMMARY
An algorithm is presented for the direct stiffness formulation of hysteretic frames which is suitable for random vibration
analysis under zero mean or non-zero mean random excitation. A modal transformation is used to reduce the number of
active degrees of freedom. First and second order response statistics are presented for two example frames using stochastic
equivalent linearization and Monte Carlo simulation. The modal transformation stabilizes the stationary response
calculation algorithm by eliminating superfluous higher modes. The iterations for transient response calculations under
non-zero mean excitations are destabilized by elimination of too many modes. The modal transformation allows
reasonable response estimates to be obtained with significant reductions in computing time. The expected lower modes
dominate the response, but inclusion of several higher modes is necessary to provide reasonable response estimates.

INTRODUCTION
Under extreme dynamic loadings, such as earthquake ground motion, extreme wind, or wave action, a
structure may undergo numerous cycles of inelastic response. Individual structural elements accumulate
damage which leads to significant reduction in element stiffness and strength. The response can be
characterized by degrading hysteresis at each of the damage sites, while response away from the damage
locations may remain largely elastic.
Excitations of the types mentioned above are frequently highly unpredictable, hence, stochastic process
models are useful in characterizing both the excitations and the structural response. A considerable body of
research exists, devoted to excitation process models, and response processes of linear and non-linear systems
under such excitations.'6.36'3833'
When the system under consideration is highly non-linear, an analytical solution for response probability
distributions becomes difficult. Even for single degree of freedom (SDOF) systems, a variety of analytical
approximations has been needed. These include equivalent linearization,'- 7 ~ 2 2 ~ 2 3 * 2 6 ~ 2 7 ~ 3-'3~73 3Gaussian
'
and non-Gaussian closure techniques,'. 5 , l 7 numerical solution of the Pontriagin-Witt equations,' and '
Wiener-Hermite series expansion^.'^
Perhaps the most extensively utilized technique in analysis of multidegree of freedom (MDOF) systems is
equivalent linearization. The theoretical basis for the method is well established, for both zero mean',''*23 and
non-zero mean3' responses. Kaul and Penzien" developed a linearization technique for the moment
equations derived from the F.P.K. equation of a multidegree of freedom structure with non-linearity confined
to the tridiagonal terms. Baber and Wen utilized equivalent linearization to analyze multidegree of freedom
shear-beam structures.536Baber and Wen used a variation upon the discrete hinge model concept formulated
previously for deterministic frame analysis". 13. ' '.and the method of equivalent linearization to obtain zero
mean response statistics of plane frames.' Baber4 extended this approach to the analysis of plane frames under
non-zero mean excitations, and considered gravity effects upon hysteretic frame response. Iwan and
KrousgrillZ2developed a finite element formulation of equivalent linearization for the dynamic analysis of
continuous systems under deterministic or random input. Other finite element models in non-linear analysis
have been de~eloped.~', 33*34

0098 -8847/86/060841-19$09.50 Received I5 July 1985

0 1986 by John Wiley & Sons, Ltd. Revised 2 January 1986
842 T. T.BABER

Modal analysis is well known as a means of decoupling equations of motion and reducing the number of
active degrees of freedom in the dynamic analysis of linear structures. Modal analysis has also been utilized in
the analysis of non-linear structures under deterministic, or random loadings. In the large deformation analysis
of elastic structures, the linear modes have been utilized as approximating functions t o reduce the problem
from partial to ordinary differential equations. Lin.,28.29Eastep and McIntosh" and Iwan' utilized
analytically derived modes with simple continua. Iwan and Krousgrill" proposed the use of modes derived
numerically from a finite element analysis. In most of the studies reported above, only a single mode was
considered.
Herein, research is reported on the random vibration of plane frames. After the stiffness formulation of the
model is summarized, a modal decomposition of the governing equations, based upon the modes of a related
linear system is presented with the objective of reducing the number of unknowns in random vibration
analysis. Numerical studies are presented, comparing the response obtained with a limited number of modes to
that obtained from the full set of equations, under zero and non-zero mean random excitation. The capabilities
and limitations of modal decomposition for random vibration of hysteretic frames are discussed in view of the
results of the numerical studies.

MODEL FOR HYSTERETIC FRAMES

Baber and Wen' and Baber4 have developed an approach for the direct stiffness modelling of hysteretic plane
frames under zero mean, and non-zero mean excitation. The model consists of an assemblage of elastic
massless frame elements joined through discrete, hysteretic hinges as shown in Figures 1 and 6. Incorporation
of the hinge regions within the elements provides some significant assembly advantages4 Thus, the element
stiffness equation takes the symbolic form

In equation (l), the subscripts T and R indicate translational and rotational components respectively of force
f, and displacement v. The 4 vector contains the hysteretic subelement rotations. The element equations may
be transformed into global form in the usual manner.40
The hysteretic subelement restoring forces H iare given by rate type models of the form
Hi = gi (Hi,
&i, Hi (t),T d I, . . .) (2)
See the Appendix for examples. If a cycle independent post-yield restoring force is assumed, the Hiare related
to thef,, by
f,,= (1 - ~ l i H) i + ~ . i k i 4 ~ (3)
The f,, are given by equation (1). Hence, for element (m)

If the masses are lumped at the nodes, and have only translational inertia, then the usual direct stiffness
approach gives the translational equations of motion for the structure,
MV, + Cf, + K.rTVT + K,,v, + K,, 4 = F (t) (5)
and the rotational equations of equilibrium

where M is the diagonal mass matrix, and C is an optional classical damping matrix. The forcing function to be
considered is due to horizontal ground excitation and gravity
 MODAL ANALYSIS FOR RANDOM VIBRATION OF HYSTERETIC FRAMES 843

where g is the gravitational acceleration, t(t)

is the horizontal seismic acceleration, d,,,, and d,,, are direction
vectors to distinguish between horizontal and vertical degrees of freedom, respectively.
The equation set is completed by the rate equations

H = g(H, 6,.. . ) (8)

which are assembled from equation (2),and the assembled relationship between hysteretic element end forces
and element displacements
H = K l V T + KZVR K 3 4 + (9)
Equations (5b(9) which govern the response can be rewritten more conveniently4, as '

where Q1,Q2,Q3 and Q4 are constant matrices obtained as a consequence of the transformations. Equations
(lo), (12) and (13) are the governing equations for subsequent numerical analysis under deterministic or
random excitation.

MODAL DECOMPOSITION
If equations (1Ok(13)are utilized in a random vibration analysis, the variable set to be integrated is the vector
y = [vTVT 4'3'. Suppose that y is an N vector. If first and second order response statistics are to be computed,
e.g. by equivalent linearization, the number of unknowns is on the order of N ( N + 3)/2. Clearly, reducing the
size of N is a high priority.
Modal reduction of equations (lOF(13)can be achieved in a reasonably straightforward manner. The first
three terms of equation (10) describe the dynamics of a related linear system, which consists of the elastic
elements rigidly joined, without any hysteretic hinge elements. Hence, the translational variables v T can be
replaced by a modal set after solution of the generalized eigenvalue problem associated with these terms. Let
the eigenvalues be Ai and the associated eigenvectors Jli. The complete set of orthonormal eigenvectors
Y = [Jll Jlz . . . Jln] forms a basis for uT; this relationship may be expressed as
vT = Ya (14)
In many problems, a relatively good approximation to the actual solution vector is obtained using a subset of
only p eigenvectors, viz.
vT = y(P)alP)
P<n (15)
Substituting equation (15) into equations (lOF(13)and premultiplying equation (15) by Y@)Tgives
the desired
modal form
gi(P) + C y ) g ( P ) + A'P)a(P) +QY)d= f (iP. ) ( t ) (16)
(P) (P)
'R = KR:[KRTY a +KRH41 (17)
H = Qyla'pl+Q 2 4 (18)
Q'P'i'P'+ Q z 4 = g(H, 6, . . . ) (19)
v, = yI(PIalP) (20)
In these equations
844 T. T.BABER

The ClQ) matrix is diagonal, in view of the earlier assumption of classical damped modes. 1‘P) is the diagonal
matrix of (p) active eigenvalues, not necessarily consecutive.
If n is the original number of variables in v,, then equations (16t(20) represent an approximate set of
replacement equations for p < n.

RANDOM VIBRATION ANALYSIS

Except for the form of the Q3 and C matrices, which are replaced by the diagonal matrices 1 and Ci,
respectively, equations (16k(19) are identical in form to the set of equations (lOk(13) used in earlier random
vibration studies4 Hence, similar analytical approaches are appropriate.
(a) Excitation models
In the studies reported, a temporally modulated, white noise was employed. The white noise may be
multiplied by a normalized Shinozuka-Sato” or an Ang-Amin’ modulating function. Low and high pass
filter effects were excluded from the present study to reduce the number of variables, but present no added
difficulty.
(b) Monte Carlo simulation
Monte Carlo simulation (MCS) is the most straightforward means of estimating the system response, but
also perhaps the most expensive. The white noise may be easily simulated by a pseudo-random number
generator; techniques are well known.I4’30 Subsequent filtering is achieved by adding the linear filter equations
to the modal equation set.
At each time step the derivative vector a is obtained from equation (16), while H is obtained from equation
(18) and a is obtained by integrating a numerically. For many of the preferred forms of g, equation (19)may be
written as
g(H, 6, . . . ) = G(H, sgn($), . ..I6 (25)
Hence, provided sgn($,) is known, for all hysteretic subelements, 4is obtained by the expression
C$ = [Q2 - G(H, sgn(d), . . . )]-I QY)a(p) (26)
Integration in terms of the modal variables proceeds in a straightforward fashion, with vR and vT computed
only at time steps where output is desired.
Stationary response statistics may be estimated by Monte Carlo simulation with considerably less expense
than transient response statistics, by utilizing the ergodicity assumption. This approach may be suitable for
zero mean response estimates, but not for non-zero mean response estimates, since non-zero mean responses
may take a long time to reach stationarity, and the stationary response statistics may provide limited
information of the transient response characteristics under modulated excitation in the non-zero mean c ~ s e . ~ , ~
Moreover, there is a significant question, in view of the non-uniqueness of the stationary non-zero mean
response, of the appropriateness of the ergodicity assumption.
(c) Equivalent linearization
If f:“’(t) is a non-zero mean excitation, the mean response may be obtained by taking expected values of
equations (16k(19), giving
(PJ (P) (P) (PI
jibp’+Ci ji. + A pa + Q y ’ k = p y , (27)
PR = KR; [ K R T y (PI cdIP) + K R H p L g l (28)
PH = Q:“’pb“’+ Q ~ P $ (29)
QYJ&’’ + Q2& = E[g(H, 6,. . . )I (30)
(PI (P’
FT = pa (31)
 MODAL ANALYSIS FOR RANDOM VIBRATION OF HYSTERETIC FRAMES 845

Subtracting equations (27b(31) from (16b(19) respectively leads to the non-symmetrical zero mean problem

where

describe the zero mean variables.

3
Except for equation (35) the set of reduced equations is linear. If it is assumed that and h are jointly
Gaussian at each time step, then equation (35) can be replaced by an 'equivalent' linear set, as shown by
K a z a k ~ and
v ~ ~Atalik and Utku,2 while the expected values in equation (30)can be computed in a reasonably
straightforward manner. The non-linear right hand side of equation (35) is replaced by the approximate
expression
g(h+p,,$+&,, . . . ) - E [ g ( h + p H , b + & . . . ) I = K,h+C,$ (37)
where K, and C, are the matrices of linearized system properties.
The linearized equation (35) can be solved explicitly for 3, giving,
where

bi
It is, of course, known that hi and are not jointly Gaussian, since hi is limited by the ultimate hysteretic
moment in most cases. Explicit forms for these limits have been given for several of the available rate type
hysteretic restoring forces. Hence, the linearization step is an approximation.
The linearized equations (32) and (38)for the response vector Y ' ~ )= [wb)' W@lT $TIT can be easily written in
the form

which leads, in the customary manner' to the equation

where

Equations (27) and (30) for the mean response, and equation (42) for the zero-time lag covariance response
constitute the set of differential equations to be integrated jointly with periodic updating of the matrices C, and
K,, provided the expected values can be computed.
In order to update C, and K , , and to evaluate the expected values in equation (30),it is necessary to compute
the mean responses pHi,b41i,and the joint second order response statistics c h i = J[Var (hi)],odi = J[Var
(&)I and pHiii = COV (f~,, 4i)/~Hzpdi.This presents some difficulty,since equation (30) is implicit in k ,which is
needed in order to evaluate the expected values. A simple fixed point iteration obtained by solving the linear
expression for does not converge. However, if the expected value computation in equation (30) is expanded
846 T. T. BABER

in a linear Taylor series expansion about a current point j&, in the vector be, the linear equations

are obtained which are easily solved for ir, at each iteration. The analysis proceeds as follows. At the beginning
of a time step an estimate Qei is computed by extrapolation from previous values. Those values, and similarly
obtained estimates for adi and pH,&,are used as a starting point on the iteration of equation (43), and the
equations described below as well. Substitution into equation (43) leads to an explicit form for an updated fi,
estimate, written compactly as

This procedure has reasonable convergence properties, without the additional complications of a complete
Newton-Raphson iteration involving the variables adi and pHi4?
The second order response statistics are obtained by using the linearized transformation

Iteration at each step is initiated by extrapolating the response from previous time steps. Equations (44) and
(45) applied to S ' p ) provide an updated estimate of the response statistics. The iteration is repeated until
convergence occurs at each time step.
The nodal response statistics of vT and vR are easily obtained, given the modal response, by using equations
(17) and (20).It is only necessary to compute the response statistics of vT and vRat time steps where output is to
be provided.

NUMERICAL STUDIES
In order to determine the performance of the nodal analysis approach in zero mean and non-zero mean
random vibration analysis, a series of numerical studies was conducted. Objectives of the studies were the
following.
1 . To determine, at least for the structures considered, the number of modes needed to achieve reasonable
approximations to the full (nodal) analysis.
2. To determine order of magnitude estimates of the computational savings attained by modal analysis in
linearization and simulation analysis of random vibration.

Three storey one bay frame, zero mean analysis

A zero mean model of a three storey one bay plane frame was considered. The frame and its idealization are
illustrated in Figure 1, with numerical properties given in Table I. The first four mode shapes and natural
frequencies of the associated linear structure are shown in Figure 2. The stationary zero mean response of the
structure to white noise excitation was estimated with 1, 2, 3 , 4 and 5 mode linearizations. It was originally
intended to compare the modal response estimates with a response estimate obtained from a non-modal
solution. However, the iteration did not converge for the full (non-modal) formulations, or for more than five
modes, as a consequence of numerical difficulties in solving equation (41) with Sfp)= 0, using the Barrels and
Stewart algorithm.8 This algorithm faih if the magnitudes of significant coefficients vary too widely.
Transforming into modal form, and eliminating the higher modes effectively reduced the range, and permitted
iterations to converge. In the studies conducted, little change was observed for more than three modes,
consequently the nodal model is considered to be unnecessary.
Figures 3-6 illustrate partial results of a series of stationary modal linearization studies at different power
spectral densities. The simulation data points were obtained using a single sample computed from a four mode
model with time averaging, under the assumption of ergodic behaviour. The system was allowed 5 seconds of
 MODAL ANALYSIS FOR RANDOM VIBRATION OF HYSTERETIC FRAMES 847

f
102" 0 LUMPED
MASSES

x DISCRETE
HINGES(H1)

I 8'-6" 102" m-
I,
29'-0"
(a) THREE STORY PLANE FRAME (b) DISCRETE HINGE MODEL FOR
ZERO MEAN RANDOM VIBRATION
ANALYSIS

Figure 1. Discrete hinge idealization for zero mean random vibration analysis, three storey frame

w z = 2144 wz = 2604
T =.43SEC. T i.12 SEC

P3l ~ o , 0 9 0
3rd MODE

T i.066SEC.
4th MODE

T t.044SEC.

Figure 2. First four natural modes of associated linear structure, three storey frame with symmetry considerations

response time to approach stationarity following which 25 second sample functions were computed and
averaged. Since the longest period as predicted by the associated linear system is well under one second these
results suggest at least 5-7 periods of starting transient discarded, and averaging over 25-30 periods of
oscillation.
Figure 3 shows estimates of the stationary RMS storey lateral translations for one mode, three mode and five
mode solutions. The two mode solution is not shown. A four mode solution coincides almost exactly with the
848 T. T. BABER

Table I. Three storey frame properties and model

(a) Original structure

Storey mass m = 0.333
Member A E I MY
Columns 28 in2 30,000 ksi 1,900 in4 10,297 in.k
Beams 25 in2 30,000 ksi 2,800 in4 8,783 in.k

(b) Model structure

Lumped masses Mi = 0.083
elastic subelement properties: same as original structure
Hysteretic hinge subelement properties
( 1 ) Type I (column) hinges.
A = 4,591,000 p = y = 0.02165 n=2
a = 0.0043
(2) Type 2 (Beam) Hinges
A = 6,839,000 p = y = 0.04432 n=2
a = 0.0043

.01
10 10 100
&iTq,

Figure 3. Modal estimates of RMS displacements as a function of power spectral density S o

five mode solution shown. Clearly, a significant change in response is caused by the introduction of the higher
modes. This is particularly significant since the normalized first mode response exceeded the largest response of
another mode (the second) by a factor of nearly 100. However, because of the system non-linearity, it was
necessary to include some higher modes, especially the second and third, to provide a reasonable
representation of the system stiffness.The limited number of simulation data points shown agree well with the
response estimates predicted by the linearization.
Figure 4 illustrates the RMS storey lateral velocities predicted by one and three mode linearization models,
and by four mode simulation with time averaging. In this case, a three mode solution provided essentially
 MODAL ANALYSIS FOR RANDOM VIBRATION OF HYSTERETIC FRAMES 849

100
U.
'i
0 MCS

-- 3 MODES
(5 MODES AGREES WELL WITH

10

01
10 10 100
mo
Figure 4. Modal estimates of RMS velocities as a function of power spectral density, So

1 10 100
J-f-5
Figure 5. Modal estimates of hinge element energy dissipation as a function of power spectral density, So. three mode--------,four mode
solution--------
850 T. T. BABER

b- /- -
30'-0"
il 12""

(a) ONE STORY FRAME

(b) DISCRETE HINGE IDEALIZATION

Figure 6 . Discrete hinge model for non-zero mean random vibration analysis, one storey frame

convergent response estimates, as the addition of more modes produced no visible change in the structure
response. The simulation data points agree closely with the linearization predictions. Figure 5 illustrates the
hysteretic element mean energy dissipation rates as predicted by one, two, three and four mode linearized
solutions and four mode simulations. A single mode analysis (not shown) provided a relatively poor estimate of
the response, in spite of the fact that the first mode is by far the dominant mode. By comparison a three mode
solution provides very good estimates. The five mode solution generally coincided with the four mode solution.
The fourth mode contributes non-negligible changes ta the energy dissipation rates, as seen in Figure 6 even
though the modal displacements of the fourth mode were about one ten thousandth of the first mode
displacements. This is caused by the interaction of modal responses in a non-linear system. The significance of
this interaction is shown by the first and second mode variance estimates shown in Table 11. The first mode
variances are particularly interesting. Even the presence of the fifth mode in the analysis causes some change in
the first mode response estimate. The second mode, on the other hand appears to be effected primarily by the
presence of the third mode.

One storey one buy frume, non-zero mean analysis

The second structure considered is a one storey one bay plane frame, shown in Figure 6, with properties as
given in Table 111. This structure was previously analysed4 using the original set of nodal variables, and two
different discrete hinge idealizations. In the present work, only the idealization shown in Figure 6 is considered.

Table 11. Modal responses at S o = 422.4, first and

second modes

Number of First mode Second mode

modes response 02m2 response u2m2
1 2.2179 -
2 3.4793 0.05612
3 3.8084 0.07444
4 4.6031 0.07298
5 4.9861 0.07582
 MODAL ANALYSIS FOR RANDOM VIBRATION OF HYSTERETIC FRAMES 85 1

Table 111. One storey plane frame properties

(a) Structure properties

Storey mass = 0.35 kip.sec2/in-
Element properties
Element A I E MY
Beam 30 in2 1,700 in4 30,000 k / K 2 6,661 in.k
Column 37 in2 1,200 in4 30,000 k / K 2 7,204 in.k

(b) Discrete hinge model properties

(i) Lumped masses
Corner masses = 0.058 kip.sec’/in ~ ’
Third point masses = 0.1 17 kipsec2/in- ’
(ii) Elastic subelement properties-see (a)
(iii) Hysteretic beam subelement properties
A = 4,250,000 n= 1
= y = 319 CI = 0.005

As discussed in earlier work? a unique stationary solution for the non-zero mean case apparently does not
exist, since the final mean response depends upon the transient excitation characteristics in the period
preceding stationarity. Moreover, the earlier studies on this, and other non-zero mean hysteretic systems3
suggest that if stationarity is approached, the displacements involved may be so excessive as to reflect a
collapsed state due to shakedown which exceeds the limits of the model. With these considerations in mind, the
discrete hinge model was analysed for six seconds of response under a white noise with power spectral density
So = 422.4. The white noise was stationary, except for a 0.05 second parabolic starting transient which was
included to enhance convergence in early studies, and was retained after convergence difficulties were
eliminated.
The first six modes of the associated linear structure are shown in Figure 7. Considering the nature of the
excitation and the presence of the gravity loads, one would expect the first and second modes to dominate the

FIRST MODE
we= 109.0 w 2 = 712
w = 10.44 w = 26.7
T = 0.602 SEC. T 10.235 SEC.

THIRD MODE FOURTH MODE

w2 = 5679 w2 = 64,740
w = 75.36 w =254
T = 0834SEC T =.0247 SEC.

FIFTH MODE
w2 = 134,400
w = 366.6
T = .017 SEC.

Figure 7. First six modes of associated linear structure, one storey frame
852 T. T. BABER

variance and means of the responses, respectively. Thus a two mode solution seems to be the logical minimum
for this structure. It was not possible to obtain convergence of the linearized system properties using the
modified Newton-Raphson method described above for fewer than four modes. It seems likely that, in the
presence of non-zero mean effects, the first two modes provide a sufficiently poor representation of the non-
linear system that divergence of the iterations occurs. Hence, results are presented herein for four and six mode
linearization analyses (4ML and 6ML) and for four mode and full nodal simulations (4MS, FS). A six mode
simulation was in general agreement with the full simulation. The simulations were obtained from ensemble
averaging 100 samples of response obtained for five seconds of response time at a time step of O-OOo5 seconds.
Results are designated in the figures by the abbreviations given above. In some instances, because of the limited
sample sizes, the expected response symmetry at the nodal locations did not occur in the simulation. Where this
happened simulation response statistics from nodes on either -side-of the structure centreline were averaged.
Statistics so computed are designated by an overbar, e.g. (4MS, FS).
Figure 8 shows the RMS lateral displacement of node three of the frame under So = 422.4, with a previously
applied gravity load a = 3864 in/sec2, taken as the initial state. The four mode solution predicts slightly
smaller transverse displacements than do the six mode or full (nodal) solutions. The ensemble averaged
simulation predicts about 3WO per cent larger RMS displacements than the linearization model at this
excitation level. Figure 9 shows the corresponding RMS velocity estimates. The four mode solutions are in
almost complete agreement with the six mode and nodal solutions.
Figure 10 shows the vertical translation of the third points of the beam under combined vertical and gravity
loads. Again, the four mode solution indicates somewhat smaller displacements than the six mode and nodal
solutions. Moreover, the linearization predicts displacements 2 5 3 0 per cent smaller than the simulation
solutions. Both the linearization and simulation solutions predict that continuing vertical translations occur
during the lateral excitation, a kind of stochastic dynamic shakedown behaviour.
Figures 11 and 12 provide further information upon the type of behaviour leading to the increasing vertical
displacements. Figure 11 shows the mean hysteretic restoring moments of third point and corner hinge
elements as a function of time during the random excitation. The corner hinge elements restoring moments
drop to about one third of their original value within about the first second of response, then decrease
somewhat more slowly, apparently toward zero, although zero is not reached within the response time
calculated. The third point hinge mean moments nearly double during the first second of response, then
decrease gradually, with no apparent tendency to level off. In both cases, the four mode analysis predicts
somewhat smaller hysteretic restoring moments than do the six mode and nodal analyses, and the simulations
generally predict somewhat smaller mean hysteretic restoring forces than the linearizations.
Figure 12 illustrates the RMS hysteretic restoring moments of the hinge subelements. Four mode and six
mode linearizations agree quite well with the nodal linearization solution, and four mode simulation agrees
with the nodal simulation. The linearization studies predict slightly lower RMS hinge moments than do the

: l0
0 1 2 3 4 5 6
t (sec)

Figure 8. RMS lateral displacement of node three of single storey frame

 MODAL ANALYSIS FOR RANDOM VIBRATION OF HYSTERETIC FRAMES 853

25
=x

20

15

10

0
0 1 +. 2 3 4 5 6
t (sec)

Figure 9. RMS lateral velocity of node three of single storey frame under white noise, S o = 422.4

O0

-6 -

-5 -

-31 gA /'

01 1 I
0 1 2 3 4 5 6
t (sec)

Figure 10. Mean vertical translation of beam third point hinges under white noise, So = 422.4
854 T.T. BABER

4000

3000
A A

4MS

1000

I 4
0 ; 2 3 4 5 6
t (sec)

Figure 1 1 . Mean hinge hysteretic restoring forces of single storey frame, So = 4224

W O L 0-( FS, 4MS

4000
CORNER HINGES FL, 6ML. 4ML.ZM

3000tf

THIRD POINT HINGES

//

t (sec)

Figure 12. RMS hinge hysteretic restoring forces of single storey frame, So = 422.4
 MODAL ANALYSIS FOR RANDOM VIBRATION OF HYSTERETIC FRAMES 855

simulation analyses, but all analyses predict a rapid increase of RMS hinge force to stationarity in about one
second.
From Figures 11 and 12, it is apparent that the initial transient period of about one second represents a
transfer of gravity load moment from the corner hinges to the third point hinges. The corner hinges must
dissipate the majority of the energy due to hysteretic behaviour. Hence the mean moments are relieved through
inelastic action. After about one second, the gradual decrease of the third point moments, and the levelling off
of the corner moments during the remainder of the response, represent the attempt by the structure to find a
way to carry the gravity load. The presence of significant RMS third point hinge moment indicates that some
inelastic rotations are accumulating there as well. Eventually, the transfer will be complete,@ this model, when
the gravity loads have transferred to the post-yield 'elastic' hinges. If a cycle independent post-yield restoring
force does not exist, which may well be the case in a more complete model, this eventual levelling off may not
occur.
Figure 13 shows the energy dissipation predicted by the hinge elements at the corners and third points.
Again, the four mode analysis predicts slightly different values than does the nodal simulation. The simulation
also predicts somewhat higher energy dissipation by the third point hinges, than d o the linearizations. Note
that for the corner hinges, the four mode simulation and the nodal simulations gave quite similar results, so
only a representative 4MS point is plotted to show the degree of agreement. Figures 14 and 15 illustrate the
modal means and RMS displacements, respectively. The first mode dominates the RMS response while the
second mode controls the mean response. The order of magnitudes of the modal responses of modes one and
two are considerably larger than the next significant modes. Even so, some changes in response statistics were
caused by adding modes five and six to the analysis. As in the zero mean case, the higher modes contribute only
trace responses, but interact with the lower mode responses because of the system nonlinearity.

COMPUTATIONAL ADVANTAGES OF THE MODAL ANALYSIS

The use of modal coordinates, and a reduced equation set, has the potential to reduce the computational effort
of obtaining random vibration statistics for hysteretic multidegree of freedom structures. Table IV illustrates
this point further by comparing solution times for a five second response calculation of the single storey frame
using four mode, six mode and nodal analysis. In equivalent linearization, reducing the number of translational

PLci

'""-1 FL, 6ML, 4ML 1

0 1 2 3 4 5 6
t (sec)

Figure 13. Hysteretic subelement energy dissipation

856 T.T.BABER

t (sec)

Figure 14. Mean modal displacements, one storey frame

10 1 I I I I
0-
Mi r4MS MODE1

FL.6ML 4ML
1

.01

001

.0001
0 1 2 3 4 5 6
t (sec)

Figure 15. RMS modal displacements, one storey frame

 MODAL ANALYSIS FOR RANDOM VIBRATION OF HYSTERETIC FRAMES 857

Table IV. Execution times for single storey frame,

non-zero mean model*

Structural Linearization Simulation

model analysis analysis

4 modes 217 sec 1753 sec

6 modes 496 sec -
8 nodes 748 sec 2187 sec

* On CDC Cyber 855 operating under nos. 2.2.

(nodal or modal) variables to be integrated results in a substantial saving ofcomputational labour, even though
the same number of hysteretic hinge elements are present in the iteration at each time step. The computational
saving results from the significant reduction in the number of simultaneous differential equations to be
integrated at each time step.

SUMMARY AND CONCLUSIONS

A model has been developed for the zero or non-zero mean random vibration analysis of hysteretic plane
frames. The plane frame is idealized as a series of elastic subelements connected by discrete hysteretic hinge
subelements. Lumped masses are located at the nodal interfaces. This model may be subjected to Gaussian
white noise or filtered modulated white noise as desired.
Under random vibration analysis for second order statistics the number of response variables increases as
the square of the number of displacement variables, leading to significant computational expenses. Moreover,
since it has been shown that a unique stationary non-zero mean solution may not exist in all cases, transient
analysis is frequently necessary.
To reduce the number of displacement variables in a non-linear random vibration analysis a modal
transformation has been developed. The transformation uses the mode shapes of an ‘associated’ linear
structure, which coincides with the non-linear discrete hinge model, except that the hinge elements are replaced
by rigid nodes.
Two structures were analysed using the modal analysis technique. A three storey one bay frame was analysed
under zero mean base excitation. A one storey one bay frame was analysed under gravity load plus zero mean
horizontal base excitation.
The following conclusions may be drawn.
1. The use of modal coordinates provides the opportunity for significant reduction of computational effort
by decreasing the number of variables in the problem, at the expense of some additional approximation.
Greater computational savings are achieved in equivalent linearization analysis than in Monte Carlo
simulation.
2. Under zero mean base excitation, the first lateral mode dominates the response (three storey frame).
3. Under gravity load plus base excitation the first lateral mode and the first vertical mode are the dominant
modes in the response.
4. It may be necessary to include several modes in addition to the dominant modes to obtain reasonable
response estimates, because of the interaction between modal responses caused by the system non-
linearity. That is, responses of the dominant modes generally increased when additional modes were
included in the analyses.
5. In stationary zero mean random vibration analysis by equivalent linearization the elimination of
superfluous higher modes improves the convergence of the iterative solution by the Bartels and Stewart’
algorithm.
6. In transient non-zero mean random vibration, elimination of too many modes caused the iteration at
each time step to diverge.
858 T. T. BABER

7. The equivalent linearization analysis tends to overestimate system stiffness as compared with Monte
Carlo simulation, but does correctly predict the significant response trends.
8. Under combined gravity and lateral base excitation, a shakedown response of beams can occur unless the
system can find a way to transfer the gravity loads to regions not affected by the lateral e ~ c i t a t i o n . ~
9. Additional computational savings may be achievable by increasing the time step size when higher modes
are eliminated, but this aspect of the computational problem was not considered in detail here.

APPENDIX
Some rate type hysteresis models
A variety of hysteresis models have been described using rate type models. In the present studies, two
primary models were utilized. These are summarized herein.
(a) Smooth system (SS) model
The smooth system rate model, modified by Baber and to incorporate deterioration, is given by

In equation (46), Ai establishes the initial slope, pi and yi establish the ultimate hysteresis restoring force

ni contributes to the ultimate hysteresis restoring force equation, and controls the sharpness of yielding. qi and
vi together with Ai control the rate and type of deterioration.
(b) Single element pinching (SEP) model
Noori and Baber3 conducted a number of studies aimed at generalizing the deterioration of the smooth
system model to include hysteresis pinching. A simple rate type model which incorporates the desired pinching,
and which has performed well in both zero mean and non-zero mean random vibration studies, is given by
1 .
Qi = - {Ai$,-vi [ ~ i l H i l " ~ - ' H i l ~ i l + y i I H i l ["l ~-6,exp(-H,?/6:)]
~i]}
rli

The parameters 6 , and Zj2 are response history dependent functions which control the amount and spread of
pinching, respectively.
REFERENCES
1. J. C. Anderson, and V. V. Bertero, 'Seismic behavior of multistorey frames designed by different philosophies', Report No.
EERC-69-1I, Earthquake Engineering Research Center, University of California, Berkeley, CA, 1969.
2. T. S. Atalik, and S. Utku, 'Stochastic linearization of multi-degree of freedom nonlinear systems', Earthquake eng. struct. dyn. 4,
41 1-420 (1976).
3. T. T. Baber, 'Nonzero mean random vibration of hysteretic systems', J. eng. mech. ASCE 110, 10361049 (1984).
4. T. T. Baber, 'Nonzero mean random vibration of hysteretic frames', Comput. Struct. 23, 265277 (1986).
5. T. T. Baber and Y. K. Wen, 'Stochastic equivalent linearization for hysteretic, degrading multistorey structures', Civil Engineering
Studies, Structural Research Series No. 471, University of Illinois at Urbana-Champaign, 1979.
6. T. T. Baber and Y.K. Wen, 'Random vibration of hysteretic, degrading systems', J. eng. mech. div. ASCE 107, 1069-1089 (1981).
7. T. T. Baber and Y. K. Wen, 'Stochastic response of multistorey yielding frames', Earthquake eng. struct. dyn. 10, 4 0 3 4 1 6 (1982).
8. R. H. Bartels, and G. W. Stewart, 'Algorithm 432: solution of the matrix equation AX + XB = C (F4)'. Commun. ACM 15, 82G-826
(1972).
9. J. J. Beaman and J. K. Hedrick, 'Improved statistical linearization for analysis and control of non-linear stochastic systems: Part I: An
extended statistical linearization technique'. J. dyn. syst. meas. control A S M E 103, 1 4 2 1 (198 1).
10. G. V. Berg, 'Response of multi-storey structures to earthquakes', J. eng. mech. dio. ASCE 87, 1-16 (1961).
11. L. A. Bergman and B. F. Spencer, Jr., 'Solution of the first passage problem for simple linear and nonlinear oscillators by the finite
element method', T.and A. M . Report No. 461, (LIILU-ENG 834007). University of Illinois, 1983.
12. T. K. Caughey and F. Ma, 'The steady-state response of aclass of dynamical systems to stochastic excitation', J. appl. mech. A S M E 49,
62-32 (1982).
13. R. W. Clough, K. L. Benuska and E. L. Wilson, 'Inelastic earthquake response of tall buildings', Proc., 3rd world conf. earthquake eng.
Auckland, New Zealand 2, 6 8 6 9 (1965).
14. R. W. Clough and J. Penzien, Dynamics of Structures, McGraw-Hill, New York, 1975.
 MODAL ANALYSIS FOR RANDOM VIBRATION OF HYSTERETIC FRAMES 859

15. S. H. Crandall, “on-Gaussian closure for random vibration of nonlinear oscillator’, lnt. j . nonlinear mech. 15, 303-313 (1980).
16. S. H. Crandall, and W. Q. Zhu, ‘Random vibration: a survey of recent developments’, J. appl. mech. ASME 50, 953-962 (1983).
17. P. K. Dash and R. N. Iyengar, ‘Analysis of randomly time varying systems by Gaussian closure technique’, J. sound uib. 83,241-251
(1982).
18. F. E. Eastep and S. C. McIntosh, Jr., ‘Analysis of nonlinear panel flutter and response under random excitation of nonlinear
aerodynamic loading’, A I A A j . 10, 2 7 6 2 8 1 (1972).
19. S. C. Goel and G. V. Berg, ‘Inelastic earthquake response of tall steel frames’, J. struct. diu. ASCE 94, 1907-1914 (1968).
20. C. Hwang and W. S. Pi, ‘Nonlinear acoustic response anaiysis of plates using the finite element method, A I A A j . 10,276281 (1972).
21. W. D. Iwan, ‘On the steady-state response of a one dimensional yielding continuum’, J. appl. mech. ASME 37, 72CL727 (1970).
22. W. D. Iwan and C. M. Krousgrill, Jr., ‘Equivalent linearization for continuous dynamical systems’, J. appl. mech. ASME 50,41-20
(1983).
23. W. D. Iwan, and A. B. Mason, Jr., ‘Equivalent linearization for systems subjected to non-stationary random excitation’, Inr. j .
nonlinear mech. 15, 71-82 (1980).
24. A. Jahedi and G. Ahmadi, ‘Application of Wiener-Hermite expansion to non-stationary random vibration of a Duffing oscillator’,J.
appl. mech. ASME 5 0 , 4 3 W 2 (1983).
25. K. Kanai, ‘Semi-empirical formula for the seismic characteristics of the ground’, Bull. earthquake res. inst. Tokyo Uniu. 35, 308-325,
(1967).
26. M. P. Kaul and J. Penzien, ‘Stochastic seismic analysis of yielding offshore towers’, J. eng. mech. diu. ASCE 100, 10251038 (1974).
27. 1. E. Kazakov, ‘Statistical analysis of systems with multi-dimensional nonlinearities’, Automat. remote contrd 26, 458464 (1965).
28. Y. K. Lin, ‘Response of a nonlinear flat panel to periodic and randomly-varying loading’, J. aerospace sci. 29. 102%1033 (1962).
29. Y. K. Lin, Probabilistic Theory of Structural Dynamics, McGraw-Hill, New York, 1967.
30. G. A. Mihram, Simulution Statistical Foundations and Methodology, Academic Press, New York, 1972.
31. M. N. Noori and T. T. Baber, ‘Random vibration of degrading systems with general hysteresis’, Report No. VVA/526378/CE85/104,
Department of Civil Engineering, University of Virginia, 1984.
32. M. Shinozuka, and Y.Sato, ‘Simulation of nonstationary random processes’, J. eng. mech. diu., ASCE 93, No. EMI, 1 1 4 0 (1967).
33. M. P. Singh and M. Ghafary-Ashtiany, ‘Seismically induced stresses and stability of soil media’, Soildyn. earthquake eng. 1, 167-177
(1978).
34. M. P., Singh and T. P., Khatua, ‘Stochastic seismic stability prediction of earth dams’, Proc. ASCE specialty conf earthquake eng. soil
dyn. Pasadena, CA (1978).
35. P-T. D. Spanos, ‘Formulation of stochastic linearization for symmetric or asymmetric MDOF nonlinear systems’, J. appl. mech.
ASME 47, 209 (1980).
36. P-T. D. Spanos, ‘Stochastic linearization in structural dynamics’, Appl. mech. rev. 34, 1-8 (1981).
37. P-T. D Spanos, ‘Stochastic linearization method for dynamic systems with asymmetric nonlinearities’, Report E M R L No. II26,
Engineering Mechanics Research Laboratory, University of Texas at Austin, Texas, 1978.
38. C. W. S. To, The response of nonlinear structures to random excitation’, Shock uib. dig 16, No. 4, 13-33 (1984).
39. E. H. Vanmarcke ‘Some recent developments in random vibration’, Appl. mech. rev. 32, 1197-1202 (1979).
40. W. Weaver, and J. M. Gere, Matrix Analysis of Framed Structures, 2nd edn., Van Nostrand Reinhold, New York, 1980.