Beruflich Dokumente
Kultur Dokumente
2 3 , 4 3 3 4 5 (1994)
AND
JOSEPH PENZIEN'
Department of Civil Engineering, U.C. Berkeley, U.S.A. and International Civil Engineering Consultants. Inc.. Berkeley. C A 94704,
ti.S.A.
SUM MARY
An alternative method is presented for generating uniform-hazard response spectra making use of a seismic hazard curve
expressing annual probability of exceedance as a function of Peak Ground Acceleration (PGA), a set of normalized
(PGA = lg) mean response spectra, and a coefficient-of-variationfunction for the random normalized spectral values in
terms of period T. The practice of using mean + la normalized response spectra, scaled to specified PGA levels, is
discussed and associated mean return periods are evaluated.
INTRODUCTION
In the development of response spectra for use in designing seismic resistant structures, it has been a common
practice to establish first sets of spectra normalized to a Peak Ground Acceleration (PGA) equal to lg and
then to scale them down to specified PGA levels in their design applications.', The shapes of a single set of
normalized design spectra have usually been obtained by averaging statistically the spectra generated for
a family of real recorded accelerograms scaled to the lg PGA level. It is now known that each member of the
family should represent a common site condition and that all members should represent ground motions
produced by the critical larger-magnitude shorter-distance seismic events. Usually, the statistical averages
representing the mean-plus-one-standard-deviation (mean + la) levels have been used for selecting spectral
shapes. The recent emergence into engineering practice of Probability Risk Assessment (PRA) methodologies
makes it possible to use a consistent probabilistic approach. The final result of this approach is a seismic
hazard curve which is simply a plot expressing annual mean frequency of exceedance as a function of PGA
for the particular site under c ~ n s i d e r a t i o n . ~In, ~generating the seismic hazard curves, the effect of PGA
dispersion must be taken into consideration. Generally, the correction for error in the PGA attenuation
equation was assumed to be log-normal distribution.
While it is still a common practice to establish normalized site-dependent response spectra and to scale
them down to the design and maximum-probable PGA levels, it should be recognized that the resulting
spectra do not represent the same probability of exceedance over the full frequency (or period) range of
interest. Because of this deficiency, the present trend in engineering practice is to develop sets of design
response spectra which do represent uniform probabilities of exceedance over the entire frequency range of
interest. These so-called 'uniform-hazard response spectra' are now being established by generating first sets
*Professor.
'Graduate student.
:Professor Emeritus.
of seismic hazard curves, each of which expresses annual mean frequency of exceedance as a function of an
acceleration response spectral value for a specified discrete value of frequency (or period) and a specified
discrete value of damping. The procedure for generating these curves is the same as that described previously
for generating seismic hazard curves for PGA; however, the computational effort is much greater due to the
large number of discrete values of frequency and damping involved. Having these sets of hazard curves.
response spectra for a specified probability of exceedance over the entire frequency range of interest are
obtained directly therefrom.
P ( A ) = 1 - Q ( A ) and p(A)=
dP(A)
~
dA
Further, let S denote the random normalized acceleration response spectral value S,,(T, 4 ) for a specified
period T and damping ratio 5. Having obtained a set of mean normalized acceleration response spectrum
curves, denoted by p = p(T, <), and the corresponding set of variance curves, denoted by o2 = 02(T,<), the
probability density function for random variable S can be expressed using the log-normal form
in which [ and i are shape and scale parameters, respectively. These two modal parameters can be estimated
from data by using three different estimation methods: method of moment, maximum-likelihood method and
non-linear regression analysis.
It is reasonable to assume that random variables A and S are statistically independent, in which case their
joint probability density function is of the simple form
P(A9 S ) = P(A)P(S) (3)
This joint probability density function can now be used to obtain the desired probability distribution
function for random variable S,(T, <) = A x S representing the non-normalized acceleration response spec-
tral value for period T and damping ratio <. This action is taken using the relation
P(c) [
,S=O
m
p(S) [ll
A=O p ( A )dA] dS
Since C is used as a dummy variable for S,, the probability distribution function for S,(T, 4 ) is identically
equal to P ( C ) , i.e.
P(Sa) = P ( C ) 16)
UNIFORM-HAZARD RESPONSE SPECTRA 435
Having obtained the probability distribution functions P ( S , ) for the full range of discrete values of T and for
specified discrete values of 5, uniform-hazard acceleration response curves are obtained directly therefrom.
having unknown constants A, B and 01 which must be evaluated so as to satisfy the realistic boundary
conditions
g(w = p ) = g(w = - b ) = 0 (94
J g(w)dw= 1
-P
The numerical values of these constants will, of course, depend upon the cut-off limits & p. It is only the
upper cut-off limit which has a significant effect on the pertinent numerical results presented later; however,
the lower cut-off limit is applied to keep symmetry of the modified probability density function, thus allowing
a more efficient analytical solution. Figures 2(a) and 2(b) show the truncated function f(w) and modified
function g(w) for p = k2.5 and p = k3.0. Numerical values of A, B and a for these cut-off limits are also
shown in these figures.
In seismic hazard analysis, the effect of PGA dispersion on the seismic risk, the actual ground intensity
Y, is defined:
Y, = N Y (10)
where N is the correction factor for error in the PGA attenuation equation. The required probability,
including the effect of this dispersion becomes
P [ Y a> y ] = P"Y > y]
. A=O. 0367
8=0.01065
0.75 - & = I 6024
. -0.75
-
0.50
. g(w)
,--. -0.50
100
- A=O. 01351
o - n nna??
u-".
I
i i-
\,".ILL
0.75 d=1.3467
-0.75
0.50
0 25
0 00
Figure 2. Truncated probability density functions f ( w ) and g ( w ) for (a) /3 = k2.5, (b) fi = +30
UNIFORM-HAZARD RESPONSE SPECTRA 431
where f N ( v )is the probability density function of N. Log-normal distribution was assumed to represent the
dispersions of the PGA values about their mean level. Set Z = lnN, then Z will follow the normal
distribution. Equation (11) can be rearranged in the following form:
in which p and o are defined as in equation (7), and a is a scaled factor which satisfies the unit area
requirement of normal distribution between the integration bounds. Zu and ZLspecify the upper and lower
bounds of integration. The distribution is truncated at some reasonable upper bound PGA value using
subjective judgement, as discussed before.
where MWR = icy=I (In PGAi - In PGAi).A normal probability plot comparing the observed distribution
of residuals for the ground motion model (using nearly 200 recorded strong motion accelerograms for rock
site in Taiwan to fit Campbells PGA attenuation model) with that based on a normal distribution and the
modified distribution g(w). Figures 3(a) and 3(b) show the comparison of plotted residuals with respect to two
different distributions: normal distribution f ( w ) and modified distribution g (w). As pointed out in these
figures, the cut-off limits of Taiwan data can be set at the f 2.50 levels, i.e. B = k 2.5. The dispersion on z,
z = In PGA, was totally confined to the region between cut-offs p = f2-5; thus, these B values have been
used in generating seismic hazard curves for PGA which consider the effect of dispersion.
- as I
1
- Campbell form
M 1 = 6.0 1 Campbell form
H>= 0.0 /
PEAK A C C E L E R A T I O N C g l
Figure 4. Seismic hazard curves on PGA
(or median) attenuation relations; Curve A was generated considering full dispersion of PGA values in the
log-normal form without applying cut-offs to the tails of the distribution; Curve B was generated considering
dispersion of PGA values in the log-normal form but applying the standard truncation procedure to the
function p(w), with cut-offs at w = f2.5,yielding the function f ( w ) ;and, Curve C was generated considering
dispersion of PGA values which follow a consistent probabilistic approach when generating the seismic
hazard curve; however, the log-normal form of distribution representing this dispersion should be truncated
following either the standard procedure or the modified procedure described previously. Of the two
procedures described, the modified procedure is preferred; thus, Curve C in Figure 4 has been used in
generating the seismic hazard curves for the normalized acceleration response spectral values S,, (T,5 ) from
which the corresponding uniform-hazard acceleration response spectra described subsequently were ob-
tained.
Estimated from
0 Taiwan Hard S i t e Data
m
04
V' ' ' ""'I 6' I
Taiwan Data: Hard S i t e Condition
0
0
0
0
oT '
0.00 0.50 1.00 1.50 2.00 2.50 3.00
T -Set.
Figure 6. Mean and mean + 1v normalized acceleration response spectra
spectrum; thus, showing that the coefficient-of-variationfunction for Taiwan data (Figure 5) can be applied
to corresponding data obtained elsewhere.
equally applicable to Seed's data, as demonstrated in Figure 6, it can be used to estimate the modal
parameters of uncertainty distribution.
Three methods of parameter estimation were used, i.e. method of moment, maximum-likelihood method
and non-linear regression analysis, to obtain 1 ( T )and [ ( T )of equation (2). Functions [ ( T )and 1 ( T )are then
introduced into equation (2) giving the log-normal probability density function for random variable
S,,(T, 005). Having this function p[S,,(T, 0*05)],it was truncated using the modified procedure for
/I= f3.0. The truncated version of this function and the function p ( A ) ,derived from Curve C in Figure 3, are
substituted into equation (5) and the double integration is carried out numerically for discrete values of T to
obtain the corresponding desired probability distribution functions P [S,(T,OOS)]. In carrying out this
double integration, the upper limit of the first integral can be changed from infinity to the upper-limit finite
value set by the truncation procedure applied to p ( S ) . Having the probability distribution functions
P [S,( T, 0.05)] for discrete values of T, the corresponding annual probability of exceedance functions
Q [S,( T, 0.05)] are obtained using the relation
Q [ S , ( T , O'OS)] = 1 - P [ S , ( T , 0'05)l (14)
Setting a fixed numerical value for Q [S,(T, 0-OS)], say, 0.010, over the entire period range of interest, the
corresponding acceleration response spectrum S,( T, 0.05) is obtained directly from Q [S,( T, O.OS)]. Since
mean return period measured in years is the reciprocal of annual probability of exceedance, the above
example probability of exceedance equal to 0.01 corresponds to a mean return period equal to 100 years.
Other uniform-hazard acceleration response spectra can similarly be obtained for other discrete values of
annual probability of exceedance, say, 0.00211 and 0.001, which represent mean return periods of 475 and
1000 years, respectively. The resulting uniform-hazard response spectra are shown in Figure 7.
.. 0'
PERIOD T CSECI
Figure 7. Uniform-hazard response spectra
UNIFORM-HAZARD RESPONSE SPECTRA 44 1
***.*Dab .....Data
Log-Normal, MM Log-Normal. MM
Log-Normal, ML Log-Normal. La
Log-normal, FIT -Log-normal. FIT
U N I F O R M H A Z A R D R.ESPONSE SPECTRA
W
e
ml
w
m
z
0
e
v)
W
a
Figure 9. Uniform-hazard response spectra developed by using three different methods of parameter estimation for the S,(T)
distribution. (Note: (. . . . .) maximum-likelihood method, (- - - - -) regressional analysis, (-) method of moment)
S, at periods of T = 0-5 and 1.0 sec, respectively, by using the three methods of parameter estimation
[Maximum Likelihood (ML) method, Method of Moment [MM] and non-linear regressional analysis
(FIT)]. The influence of the shapes of distribution function on the calculation of uniform-hazard response
spectra is shown in Figure 9 for a return period of lo00 years. It is found that the uniform-hazard response
spectra developed by using the parameters estimated by method of moment gave a lower spectral values than
others.
442 C.-H. LOH, W.-Y. JEAN AND J. PENZIEN
U N I F O R M HAZARD RESPONSE S P E C T R A
Figure 10. Sensitivity study on uniform-hazard response spectra for 4 soil types (SI: Rock site, S2: Stiff soil condition, S3: Deep
cohesionless soil, S4: Soft to medium clay and sand)
0
Modified 2.0 S e
Modified 9.0 Sigma 3 0.01
4
2
32 om1
04
\
\
\
\
\
\
\
\
1 o.oooo1
Modified 2.0 Sigma
\
\
-PGA S,, -
Modified 9.0 Sigma
(Mean + 1 6 )
Row001 0.000001
0.0 et RI OJ ae 1.0 13 1.4 0.0 01 0.4 0.0 0.8 1.0 ld 1.4
Damign Reaponre Spectra ( 0 ) Design Remponse Spectra (g)
Figure 11. Comparison among the annual probability of exceedance function of Q(S,(T,005)) and the design response spectral value
D,;(a) for period = 1.0 sec; (b) for period = 0.5 sec
-
spectral value C d [ = A S,,(T, 0.05)] is also plotted in this figure for comparison. For example, as shown in
Figure 1 l(a) for period = 1.0 sec, at an annual probability of exceedance equal to 0.0021 (about 475 years
return period) the annual probability of exceedance function of Q [S,( T, 0.05)]for truncation level /3 = f2 0
has a lower risk than the risk curve of design response spectral value, while the curve with truncation level of
/3 = f3.0 gives a higher risk than the design curve. It means that if we use the hazard curve of
{ 1 - P [ S , ( T, 4)] } with truncation level of /3 = k 2.0 for our design curve, then the present design criteria by
+
scaling the normalized acceleration mean 1o response spectra to the PGA level of 475 years return period
will be very conservative. On the other hand, if the hazard curve of (1 - P[S,(T, ()I} with truncation level of
/3 = f3.0 was used for design, then the present design procedure will not be on the conservative side.
As a numerical example, let Curve C in Figure 4 represent the site-specific seismic hazard curve on PGA
+
and let Seed's 5 per cent-damped mean 1o normalized acceleration response spectrum shown in Figure 6
represent Sd(T,0.05). Let RA(A) denote the mean return period of PGA = A as given by the seismic hazard
curve on PGA, recognizing that R , ( A ) = l/Q(A), and let Rd(A, T, denote the mean return period e)
r)
measured in years of S,(T, = A x &(TI <) as obtained by
Rd(T, = l/{ - p [ s a ( T , ()I} (15)
in which P[S,(T, r)]
is given by equation (5). Having specified a particular value of A, the corresponding
values of R , ( A ) and Rd(A, T, 5 ) are obtained. By the procedure described above, Rd versus RA relations can
be obtained for discrete values of T as shown in Figure 12. The irregular form of the function Rd versus T is
certainly due to the difference between hazard curves of Q [ S , ( T ,0.05)] and P ( A ) * S , , . To demonstrate the
use of Rd versus RA curves in Figure 12, choose R , = 475 years which, as seen in Figure 4, corresponds to
A = 049. Specifying arbitrarily T = 1.0 sec, it is shown in Figures 12(a) and 12(b) that Rd = lo00 years (if
/3 = f2.0 is used in the truncation of spectral acceleration distribution) and Rd = 740 years (if fi = f3.0 is
used in the truncation of spectral acceleration distribution), respectively. If T = 0.8 sec is chosen, it is also
shown in Figures 12(a)and 12(b),Rd is 470 years and 370 years, respectively. This example shows that when
scaling the mean + 16 normalized response spectra by some value A for design purposes, the return periods
444 C.-H. LOH, W.-Y. JEAN AND J. PENZIEN
PG A : 2 . 5 CJ, Sa: 2 u
10000
Q? 1000
4
0
.rl
k
0)
a
g
4
100
a,
ffi
10
10 100 1000 101 00
Return Period, RA
PGA:2.5u,Sa:3 CJ
1000 &.-.-. -
2.0 sec
100 1.0 sec
0.1 sec
0.0 sec
0.8 sec
00
Figure 12. Return period R , vs. R A for discrete values of period T. (a) for p = k2.0 as truncation bound in S, distribution, (b) for
B = k 3.0 as truncation bound in S, distribution
on the scaled spectral values are generally greater than the return period on A (except T = 0.5 sec for
/$ = & 3.0). Comparing return periods as described above it is concluded that if the uniform-hazard response
spectra developed by using p = 2.0 as truncation was used for design purposes, then the present practical
design procedure (normalized response spectral by their associated PGA level) is very conservative. However,
if the uniform-hazard response spectra developed by using fl = 3.0 as truncation was used for design, then
thc present practical design procedure will be a little non-conservative (especially at T = 0.5 sec).
CONCLUSION
The alternative method presented, herein for generating uniform-hazard response spectra is efficient from
B computational point of view. It makes possible the conversion of existing site-specific seismic hazard curves
on PGA and their corresponding site-specific normalized acceleration response spectra to uniform-hazard
response spectra, thus making it unnecessary to generate a large set of seismic hazard curves on normalized
response spectral values for discrete values of period T and damping ratio t. Discussion on the PGA
UNIFORM-HAZARD RESPONSE SPECTRA 445
dispersion as well as the truncation bound of spectral acceleration probability distribution to the influence of
uniform hazard response is presented. For Taiwan data, it is recommended that truncation at 1 = f2.5 for
PGA dispersion is suitable for practical design purposes. The present design procedure, normalized response
spectra by their associated PGA level with given return period, was discussed and compared with the
proposed uniform-hazard response spectra.
REFERENCES
1. H. B. Seed, J. Ugas and J. Lysmer, Site-dependent spectra for earthquake resistant design, Report No. EERC74-12, Earthquake
Engineering Research Center, University of California, Berkeley, 1974.
2. N. M. Newmark, J. A. Blume and K. K. Kapur, Seismic design spectra for nuclear power plants, 1.power div. ASCE 99,287-302
(1973).
3. C. A. Cornell, Engineering seismic risk analysis, Bull. seism. SOC. Am. 58, 1583-1606 (1968).
4. A. Der Kiureghian and A. H.-S. Ang, A fault-rupture model for seismic risk analysis, Bull. seism. SOC. Am. 67, 1173-1194 (1977).
5. EERI, The basics of seismic risk analysis, Earthquake spectra 5, 675-702 (1989).
6. A. H.-S. Ang and W. H.Tang, Probability concepts in engineering planning and design, in Basic Principles, Vol. I, Wiley, New York,
1975.
7. K. W. Campbell, Near-source attenuation of peak horizontal acceleration, Bull. seism. soc. Am. 71, 2039-2070 (1981).