Professional Documents
Culture Documents
PROBABILITY-BASED CRITERIA FOR STRUCTURAL DESIGN Ellingwood1982
PROBABILITY-BASED CRITERIA FOR STRUCTURAL DESIGN Ellingwood1982
Center for Building Technology, National Bureau of Standards, Washington, DC 20234 (U.S.A.)
and
Theodore V. Galambos
Department of Civil and Mineral Engineering, University of Minnesota, Minneapolis, MN 55455 (U. S.A.)
Keywords: Buildings (codes); design (buildings); limit states; loads (forces); masonry; probability
theory; reinforced concrete; reliability; safety; specifications; standards; steel; structural
engineering.
ABSTRACT
In working stress design [1,2,6], the design desire to quantify performance of structures
loads are usually unfactored and are close to and to treat uncertainties in resistance, loads
the maximum probable loads to occur during and analysis in a more ration~i way. Limit
approximately a 50-year period of time [7]. states design, in contrast to working stress
The elastically computed stress arising from design, is a behavior-oriented design method
these loads is limited to the allowable stress, that requires specification writers and design-
which is a specified fraction of the failure ers to consider explicitly the ~tructural re-
stress. The ratio of the failure stress to ulti- quirements for function and safet5 at service
mate stress is the overall factor of safety. and extreme load levels [8]. The probabilistic
Most checks of serviceability also utilize un- approach is suggested by the observation that
factored loads. In strength design [4,6], the many of the design variables exhibit statistical
loads are multiplied by load factors that are regularity. Some nominal loads in a number
generally greater than unity and the design of standards (e.g. [3] and [5]) already have a
resistance is obtained by multiplying the statistical basis. This article describes how
calculated strength by a resistance factor less safety-related performance criteria for struc-
than unity. tural design can be selected using reliability
The factors of safety or load and resistance analysis to integrate available statistical data
factors are provided to take into account the on resistances and loads.
uncertain nature of the loads, resistances and
models used in structural analysis. These un-
certainties may lead to unfavorable deviations BASIC APPROACH
in the loads and in the load-carrying capacity
of the structure from what was assumed to The development of practical probability-
proportion and detail its members. Over a based loading and resistance criteria that
period of time, standard-writing committees would be acceptable to the professional de-
have selected or adjusted these factors on the sign community at large has five essential
basis of experience with existing structures. components:
perceptions regarding the accuracy of struc- Ill Develop statistical data to describe the
tural analysis methods, and intuition. Al- basic load and resistance variables in eqn. 1
though this process usually has served the [7.10-121.
profession well, its intuitive basis makes it (2) Establish procedures for calculating retia-
difficult to quantify acceptable performance bilities of structural members and systems
and to achieve anything close to uniform reli- [13-15]. Ideally, the performance criteria
ability in all structures designed by any par- should be based on a system reliability re-
ticular specification. Moreover, there is no qmrement. However. current practice usually
assurance that the performance criteria will be is to check performance on the basis of indi-
adequate if they are applied in nonroutine vidual member behavior.
design situations for which there may be little (3) Establish target reliabilities for design by
experience. Pressures to reduce design costs analyzing reliabilities associated with struc-
are expected to lead to the rapid introduction tural members designed according to existing
of new construction materials and systems criteria [7.10.16]. T h i s e n a b l e s the
without substantial experience with their per- probability-based criteria to be related to ex-
formance, making this last issue particularly isting acceptable practice and provides the
serious. continuity that is necessary from one design
The development of probability-based limit specification to the next.
states design [7 9] has been motivated by a (4) Select a deterministic format for eqn. I
17
that balances theoretical appeal with the need strengths and dimensions may be more ap-
for simple safety and serviceability checking propriate [7,9].
procedures in professional practice [7-9,16].
Determine general load criteria suitable for all Structural loads
construction materials. Most structural loads may be thought of as
(5) Develop resistance criteria that are con- consisting of a basic load parameter which is
sistent with the load criteria selected in step 4 essentially independent of the structure; a
such that reliabilities are close to the target modeling parameter that transforms the spa-
values selected in step 3 [7,16]. tially and temporally varying load into an
Although steps 4 and 5 are closely related, equivalent uniform (and usually static) load
they have been identified separately. Divi- for analysis and design purposes, and finally,
sions of responsibility among the groups pre- an influence coefficient or analysis factor that
paring standards and specifications are such transforms the uniform load into a structural
that the load and resistance criteria usually action such as a beam moment or column
are developed by different groups at different thrust. Uncertainties in the basic environ-
times. These five steps are elaborated upon in
ment, the load modeling, and structural anal-
the following sections.
ysis all contribute to the variability in the
load effects used to calculate structural relia-
bility.
STATISTICAL DATA BASE
When loads are combined, the probability
distribution of the maximum of their sum
The probability distribution of each load
during some time period T may be required
and resistance variable used to calculate limit
for the reliability analysis. If only one time-
state probabilities and to develop perfor-
varying load is combined with permanent
mance criteria should characterize the uncer-
load, this calculation is relatively straightfor-
tainty in the variable that would be expected
ward. However, if more than one time-varying
in structures in service. Thus, measurements
load acts, the probability distribution of the
of variability obtained, e.g., from laboratory
maximum of a sum of random load processes,
tests of material strengths or from local or
limited load measurements may not be suffi- Z=max[X,(t)+X2(t)+...+Xk(t) ] (2)
T
cient for determining the probability distribu-
tions. Sources of uncertainty due to modeling is required. The computation of the probabil-
and limited data must also be considered [13]. ity distribution Fz(x ) is complex. Thus, ap-
Many common limit states such as fracture proximations have been developed that re-
of tension members or flexural yielding of duce the load combination analysis to a prob-
beams can be formulated as a linear combina- lem of combining random variables rather
tion of resistance and load effects. Because of than random processes.
the emphasis that these linear limit states A simple model for this purpose presumes
have received, the basic resistance variable that Z occurs when one of the loads attains its
described in the following has been taken to maximum value during T while the other loads
be the strength of the structural member in are at their point-in-time values [17]. The
question, while the basic load variable is the point-in-time load is simply the load Xj that
load effect that is dimensionally consistent would be measured if the random load pro-
with the resistance. In some cases where the cess were to be sampled at an arbitrary time
limit state is highly nonlinear, however, a instant. Thus, eqn. 2 is replaced by
formulation in terms of the basic material
Z=max max Xi(t ) + •. (3)
i T i~ l • j
18
Equation 3 appears to be a good approxima- the distributions best fitting the tipper per-
tion to eqn. 2 in many practical cases [18]. centiles of the actual probability distributions
Moreover, it is consistent with the observa- for the loads, which frequently were obtained
tion that those structural failures that are not by numerical integration [7]. While the sta-
caused by negligence or accidents usually oc- tistical estimates presented a~e considered
cur when one load attains an extreme value typical, some variation with geographic loca-
rather than under a combination of several tion may be observed with the environmental
time-varying loads [7]. Equation 3 implies that loads.
a load process can be described with suffi-
cient accuracy by a point-in-time distribution Structural resistance
and a maximum distribution, each of which
may be obtained rather easily. The resistance of structural members and
Table 1 summarizes statistical data ob- systems to applied loads is normally ex-
tained in recent studies to characterize loads pressed by a formula that has been deriVed:
for probability-based structural design (e.g. from theories of structural mechanics and has
[7]). The mean values are presented with ref- been verified by laboratory testing. This mod-
erence to the nominal design loads in Ameri- eled resistance may be denoted R,,, or
can National Standard A58 [3]. The coeffi- R,,, = R,,,( X,. X2 ..... ). (4)
cients of variation reflect sources of uncer-
tainty due to basic or inherent variability, the in which X~ are the basic (and usually ran-
load modeling, and structural analysis. In dom) variables used to determine load-carry-
simple frame structures and others for which ing capacity, such as material strengths or
accurate analysis techniques exist, the uncer- structural member sizes. The idealizations of
tainty due to structural analysis is negligible structural behavior that underlie the formula-
in comparison to the other two. The assumed tion of R,, contribute additional uncertainty
probability distributions and associated means in the in-situ resistance of structural mem-
and coefficients of variation were selected as bers. This modeling uncertainty is manifested
TABLE 1
Typical statistical data on loads
TABLE 2
Typical statistical data on resistance
Structural steel
Tension members, yield 1.05 0.11 Lognormal
Compact beam, uniform moment 1.07 0.13 Lognormal
Concentrically loaded column, intermediate slenderness 1.08 0.15 Lognormal
Beam-column 1.07 0.15 Lognormal
Cold-formed steel, braced beams 1.17 0.17 Lognormal
Reinforced concrete
Beam in flexure 1.12 0.13 Lognormal
Beam in shear 1.10 0.20 Lognormal
Short tied columns, compressive failure 0.98 0.16 Lognormal
Plant precast and pretensioned beams 1.06 0.09 Lognormal
Two-way slabs, flexure 1.12 0.14 Lognormal
Unreinforced masonry
Walls in compression, uninspected workmanship 5.30 0.18 Lognormal
/
assessment must focus on those designs for
fXl-X 2 which there is professional consensus that
performance currently is satisfactory and
which are not unduly conservative. The fol-
lowing reliability assessments of structural
members are referenced to a 50-year period of
time.
TABLE 3
Summary of reliabilities associated with current criteria for dead and live loads
Structural steel
Tension members, yielding ( L o / D , = 2) 2.5
Tension members, fracture ( L o / D n = 2) 3.4
Compact beam, uniform moment ( L 0 / D n = 2) 2.9
Column, intermediate slenderness, instability 3.0
Bolted or welded connections ( L o / D n = 2) 4.0
Reinforced Concrete
Beam, flexure 2.9
Beam, shear 2.3
Plant precast, pretensioned beam, flexure 3.6
Short tied column, compression failure 3.4
Unreinforced masonry
Wall in compression, uninspected workmanship, story-height 7.0
/3 for nonreinforced masonry walls loaded justment in stress or load apparently accounts
under conditions approaching pure compres- for more than the small probability that ex-
sion appears to be excessively conservative. treme loads coincide. Wind and earthquake
Similar analyses have been performed for loads affect the entire structural system.
cold-formed steel, aluminum and wood whereas the live and snow loads tend to bc
structural members [7]. more localized. Thus, it is possible that tile
one-third increase is an implicit allowance for
Wind and earthquake loads tile contribution of load-sharing and ductility
to system reliability. On the other hand, it
Combinations including both live and wind also is possible that current structures are
loads would need to consider both the combi- overdesigned with respect to gravit> loads and
nations D + Lapt q-- 1,~ and D + L + I'Vapt (see underdesigned with regard to wind and earth-
eqn. 3). The variation in /7 with L o / D ~ and quake effects. Additional research should ad-
W~/D, is shown in Fig. 3 for the case where dress this question.
gravity and wind loads are additive. The value
of /3 is in the range 2.0 2.5 when the wind
load effect is a major component of the total PROBABILITY-BASED LOADING AND RE-
load effect. Similar analyses of load combina- SISTANCE CRITERIA
tions involving earthquake loads lead to /3-
values of 2.0 or less [7]. Load and resistance criteria must be
The lower apparent reliability for load selected to be consistent with the performance
combinations involving wind or earthquake and reliability requirements prescribed by
loads is due, in part, to the 1/3-increase in specification committees. Equations 8 through
allowable stress or reduction of 0.25 in total 11 show that partial safety factors on the
factored loads permitted by all current stan- nominal load and resistance variables could
dards used in this evaluation. The small likeli- be derived that would lead to designs with
hood that two or more time-varying loads reliability index /3o in all cases. Figure 4 il-
attain their 50-year m a x i m u m values simulta- lustrates the factors YD, "/L, and Ys on dead,
neously is taken into account by the load
combination analysis, eqn. 3. Thus, the ad-
2,0 f
/
iI 2 / TL
Lo/D
0
-0
/ / - 0.5 Steel beQms
/
4
1.5
, ' , 1,0
4 ' / .,
%
1.0
.t ~ 0.5
o
D+
D+S
- - -
live and snow loads and resistance factor q~ equation format and partial safety factors can
for checking the flexural capacity of steel be chosen so as to minimize this deviation, in
beams that would be obtained by this analy- an average sense, over those design situations
sis. It may be observed that the required to which the checking equations are likely to
values of ,~, ~'D, "fL, and 3's vary with L,/Do apply.
and S,/Dn. This variation is significant for 7c Numerous checking equation formats are
and ~'s- Variations in the partial safety factors possible, and one must be selected that strikes
of this magnitude would not be acceptable in a balance between theoretical appeal and sim-
standards and specifications used for routine plicity. Strong preferences have been ex-
design because the factors would change dur- presses by professional committees [4,6] for
ing each iteration of structural dimensioning selecting the resistance criterion as ~ R n, in
and detailing. which the resistance factor ~ is applied to
Practically speaking, load factors must be strength. A load format can be devised [7]
selected to be constant and independent of that is consistent with the treatment of com-
Ln/Dn, Sn/D~, etc. The fact that there are binations of loads in Eqn. 3, and yet retains
certain general performance requirements that the nominal loads that currently are specified
are the same for all structures suggests that it in the A58 Standard [3]. The checking equa-
would be desirable for the load requirements tion for the strength limit states becomes
to be the same for all construction materials
¢bRn>yDDn+Y,Qo,+ ~ yjQn,, (12)
[7,9]. Uniform load requirements would have j4=i
significant practical advantages over the di-
verse treatment of loads in existing standards in which Q. are nominal loads, YiQn, is the
[1,2,4,6[. Moreover, the resistance criteria factored principal variable load and terms
should depend only on the limit state and
"YjQn,are equivalent to the factored arbitrary-
consequences of a failure and not on the point-in-time loads in Eqn. 3.
The load and resistance factors must be
loads.
Figure 4 indicates that if the partial safety apportioned to the individual variables so that
factors on resistance and loads are specified the reliabilities are nearly uniform for all load
as constant in the checking equation, the re- situations and are consistent with the conse-
sulting designs will deviate unavoidably from quence to the structure of entering the various
the target reliability 130. However, a checking limit states. The following procedure may be
TABLE 4
Assumed weights Pi for gravity load combinations
Steel D + L
D + S 0.00 0.10 0.20 0.25 0.35 0.07 0.03
RC D + L 0.10 0.45 0.30 0.10 0.05 0.00 0.00
D+ S 0.30 0.40 0.20 0.05 0.05 0.00 0.00
Masonry D+ L
D + S 0.36 0.36 0.20 0.06 0.02 0.00 0.00
24
used [7]. A required nominal resistance R* load factor, 7D. Figure 4 suggests that this
may be computed from Eqn. 8 for a target factor could be about 1.10 (this factor takes
reliability/3o and a given set of nominal loads. into account the fact that dead loads tend to
Similarly, when the load and resistance fac- be underestimated). However, this low a value
tors are fixed, a nominal resistance may be of "/i) runs counter to professional experience
calculated from Eqn. 12, which usually is and would not be accepted. Thus, fD was set
somewhat different than R*. A set of load equal to 1.2, which appeared to be the mini-
and resistance factors ,~ and 7 may then be m u m acceptable value. The second relates to
selected that minimizes the weighted squared ~-factors. According to Fig. 4, ,) for flexure
difference between R* and R n, defined by might be as high as 0.95 in some cases. How-
ever, experience has shown that if q~ is set at
I ( ~ , ~,) = Epj(R*., - Ro~) ~ (13) 0.9 or higher for ordinary construction, it may
.i
be difficult to adjust the resistance criteria to
over a predefined set of dead, live, wind, account for improvements in fabrication or
snow, and earthquake loads for which the quality control, both of which tend to reduce
criteria are to apply. The relative weights pj l;~, without increasing ,~ above unity. Thus,
assigned to the D + L and D + S load com- the loading criteria were derived with the
binations are shown in Table 4 for steel, rein- additional constraint that the ,~ for flexure in
forced concrete, and masonry construction [7]. steel and reinforced concrete should not ex-
Note that the effect of dead load tends to be a ceed about 0.85-0.9 for average conditions.
major component of the total load effect in Third, in current practice the same load factor
concrete and masonry construction. is assigned to live and snow loads; thus, while
Fig. 4 indicates that these factors should be
Loading criteria somewhat different, 7c was set equal to "/s-
The reliability analysis of designs conform-
Load criteria must be set first if the goal of ing to existing criteria indicated that values of
uniform load requirements for all construc- /3 for many flexural and compression mem-
tion materials is to be attained. In fact, it is bers tend to fall within the range 2.5-3.0.
quite likely that if the methodology were to be Thus, the new toad criteria are chosen so that
applied separately to different construction specification committees can achieve retiabili-
materials, different loading criteria for each ties within this range for most ordinary design
might be obtained. Because this would be situations if they so desire. The calculation of
undesirable, and because of the nature of the suitable load (and resistance) factors accord-
standard development process, it is necessary ing to Eqn. 13 is a constrained optimization
to separate the resistance and load criteria, process and is described in [7]. The load
even if such a separation might appear to be criteria obtained are
artificial. However, useful load criteria cannot 1.4D,1
be attained without some consideration of the
corresponding resistance criteria likely to be 1.2 D n + 1.6L n
used. Moreover, current inclinations and ex- 1.2D n + 1.6Sn + (0.5L, or 0.SW,) (14)
periences of the engineering profession in- 1.2D. + 1.3Wn + 0.5L n
fluence the selection of both loading and re-
sistance criteria and place constraints on the 1.2D n + 1.5En + (0.5Lo or 0.2S,)
optimization of Eqn. 13. 0.9D, -- (1.3Wn or 1.5E.)
Three examples illustrate this last point.
The first relates to the selection of the dead Some load factors are less than unity in
25
5
PROPOSED CRITERIA
4 Rnew
Proposed: CR : 1 . 2 D + 1.6L
n
Rold
n
- "~- .~,. 0.75
0.80
0.85 1,05
/, -- 0.90
Current ,
practice
1.0 I
1,0/ 1.5 210 2,5
Lo / On
0 I I I I I 0.95
0.5 1.0 1.5 2.0 2 5 3.0
Ln/Dn