You are on page 1of 12

Code Calibration

and
Partial Safety Factors
Safety Levels of Design Codes
• Design codes provide the minimum requirements for safety. Based on how the
safety level is achieved in design codes, we can categorize the design codes into
4 levels:
• Level I codes, where only fixed deterministic values are used for design and
safety is reflected using central safety factors {i.e. using average values}. This
was common before 1960-1970s.
• Level II codes, where the partial safety factors are calibrated using basic
statistical information about design parameters {Generally Average Values and
Standard Deviations}. These represent most current design codes (AISC/ACI/IBC
Eurocode and other LRFD codes)
• Level III codes, where the probability of failure is explicitly calculated as design
parameter and the full PDF distributions of random variables are used in design.
These codes are still in research phase.
• Level IV codes, where coupling between cost/performance and probability of
failure is used for design of structures.
Calibration of Level II Codes
• In
  Level II, the safety factors used in the design are generally calibrated
using reliability analysis to achieve specific target Reliability Index { βo }
• The general format for LRFD design equation is:

where: Rn = Nominal Resistance, = Capacity Reduction Factor for Rn, =


Load Factor for load effect i, = The Nominal Load Effect resulting from
Load i.
• The factors and are called: Partial Safety Factors {PSF}.
• The PSF are calibrated (evaluated) such that the design equation above
achieve a specific target reliability index βo.
• Obviously, the resulting values of PSF will depend on the selected βo , and
on the statistical data for the material and dimensions and loads in the
country where it is to be applied.
Partial Safety Factors Based on FOSM
•  Suppose under some design scenario, the member is subjected to the
load effects Q1, Q2, …, etc. The performance function will be:
g = R – { Q1+Q2+ … } , here R and Qi are R.V.
• The LRFD design equation for this scenario (which uses deterministic
partial safety factors} will have the following format.

Here: Rn = Nominal Resistance, Qn=Nominal Load are Deterministic


values.
• Recall that the mean bias factor is defined as: and , therefore, we have:
and
• Also recall that based on FOSM, The Cornell Reliability Index for the “g”
above is given as:
PSF Based on First-Order Reliability
• For
  acceptable reliability-based design equation, the computed Reliability Index
must be bigger than a Target Reliability Index {i.e.: ≥ o }.
• Therefore, and doing the cross multiplication, we have:

• The following Algebraic manipulation can be made to simplify the terms:

• Substituting , where V = Coefficient of Variation:

• Now we define the factors The are called the “weight factors”.
• The reliability design equation then becomes:
PSF Based on FOSM
• The
  reliability-based equation can be re-arranged to become similar to the LRFD
code design equation:

• Compare to LRFD design equation: , the Partial Safety Factors {PSF} for this design
scenario are readily obtained:
 Where is the weight factor for the
,
R.V. X, defined as:
,
, … etc
• The weight factor operates as an influence factor which gives more weight to the load
that has higher variability. Therefore, we expect higher load factor for loads with
higher variability.
• Also the resistance that has higher variability {higher } will get a lower capacity
reduction factor
Example: FOSM PSF
•• Simply supported beam of length L = 5 m is subjected to uniform dead
 and live loads, the nominal capacity of the beam is given as M = F Z . item Mean CoV Bias
p y x Dead Load 10 kN/m 0.1 1.1
• Use the given information to determine the partial safety factors for the Live Load 25 kN/m 0.2 1.2
resistance and the dead and live loads assuming the target reliability Fy 2.45E+05 kPa 0.08 1.12
index = 3.5. Use FOSM theory and assume the deterministic design
Zx To Be Designed 0.05 1.1
equation has the LRFD form , assume independence.

 •First, we need to find an estimate for the variable Zx. We can estimate Zx using the
mean values:
Mp > Ma  FyZx > (wD+wL)*L2/8  Zx > (wD+wL)*L2/(8Fy) = (10+25)*52/(82.45e5)
We get Zx = 4.46e-4 m3.
• Also, we need to prepare the mean, COV and Bias for the resistance.
• Since Rn = Mp = FyZx  μR = μFy×μZx = 1.09e2 kN.m,
The same for the Bias: = Fy×Zx= 1.232
The CoV for R can be calculated as:

• Now we prepare the table for the calculation of PSF.


Example \ Cont.
 • The calculations are summarized in the following EXCEL table.
• The resulting PSF are: = 0.87, γDead = 1.13, and γLive = 1.56
• So, the deterministic LRFD design equation will be: 0.87 FyZx > (1.13 wD + 1.56 wL) L2/8,
where the {Fy ,Zx , wD , wL } are based on the NOMINAL values.
• Note these are very close to those used by the ASCE/IBC codes for residential buildings.
{0.9,1.2,1.6} item Mean CoV Bias Sigma ai
• This is because the data on the Dead Load 10 0.1 1.1 1 0.086884
statistical variations are very Live Load 25 0.2 1.2 5 0.434422
close to those used for Fyy 2.45E+05 0.08 1.12 --- ---
Zxx 4.46E-04 0.05 1.1 --- ---
residential buildings in USA.
R 1.09E+02 0.09434 1.232 10.3184169 0.896509
• If the statistical variation data
change, then the PSF will BetaO 3.5 Σ{σ22} 132.469727
become different.
• Note also that selecting another Phi 0.867306095
Phi 0.867306095
mean value for Zx will also γDead 1.133450484
γDead 1.133450484
change the PSF!! γLive 1.564914374
γLive 1.564914374
PSF Based on Hasofer-Lind R.I.
•• Because
  it is invariant, the Hasofer-Lind R.I. is used for nonlinear performance functions and can be
easily generalized to non-normal and/or correlated variables.
• The definition of H-L R.I. can be used to determine more accurate Partial Safety Factors.
• The idea is to do “inverse” analysis for the βH-L and keep searching for the design and mean values
that produce the target value of R.I. , i.e. βH-L = βo and at the same time the design values should
satisfy the limit state (g=0)
• This solution requires iterations because it is a nonlinear problem. The procedure is:
1) Suppose we have the performance function g(X1, … , Xn)=0, where g < 0 is considered failure.
2) Assume initial values for X1 to Xn-1, and then find Xn by solving the g = 0. Call these assumed values as “design
values” Xd = { X1d, …, Xnd}
3) Calculate the derivatives of g w.r.t Xi, Gi = -(Vi μi )dg/dXid. Put the derivatives Gi into a vector G = [ G1 , … , Gn ]T.
Note that Gi is effectively the contribution of the standard deviation of variable i in from g.
4) Compute the direction cosines (or weight factors) = . Note for Linear g :
5) Compute the new design values using Xid = μi (1+ ZiVi ), where Zi = βo for i = 1 … n-1, and solve for Xnd with the
new Xid,…Xnd-1 by imposing that g = 0.
6) Update the mean values, which you need for computing Gi using μi = Xid / (1+ ZiVi) , Zi = βo
7) Go back to step 3, and repeat until step 6: recalculate a new Gi, then a new , then find the new design values
Xid, and then the new mean values μi , then loop back to step 3…etc…Until the values of (or the values of Xid)
converge.
8) Compute the PSF as γi = Xid / Xnominal = λiXid / μi
PSF Based on H-L R.I. g(X1, … Xn) = 0

Assume initial values for : X1 … Xn-1


• The
  flow-chart shows the steps for the iterative Solve g = 0 for Xn
procedure to finding the PSF. Do the same for the Mean Values
• This can be programmed in any software,
 Compute gradients: Gi = -V μ dg/dXid
including EXCEL!. i i
Compute the weights:
• The H-L approach is more accurate in the cases
where “g” is a nonlinear function of the variables   Compute the new values (X1 … Xn-1)
and the variables are Normally-Distributed. Xid = μi (1+ Zi Vi ), Zi = βo
Solve g = 0 for Xn
• In case there is a correlation between
variables, the only change is when computing , it   Update the new mean values
μi = Xid / (1+ Zi Vi) , Zi = βo
becomes where is the Covariance Matrix.
• In case the variables are not normal, we use  Get new gradients: Gi = -Vi μi dg/dXid
the Rackwitz-Fiesseler normalization Compute new weights:

procedure, where the mean values are   Check: Σ (new – old )2 < Tolerance
computed using equivalent normal distributions. NO
YES
Compute PSF: γi = λiXid / μi
Example: H-L PSF item Mean CoV Bias
• The previous example will be solved this time using the H-L Dead Load 10 kN/m 0.1 1.1
iterative procedure. The data will be from the previous example. Live Load 25 kN/m 0.2 1.2
Fy 2.45E+05 kPa 0.08 1.12
• The performance function is g = Mp - Ma  g = FyZx – (wD+wL)*L2/8 Zx To Be Designed 0.05 1.1
• Note that the partial derivatives of g are:
GwD = - ( – L2/8)*VwD* μwD = ( L2/8)*VwD* μwD WD WL Fy Zx
Bias 1.1 1.2 1.12 1.1
GwL = - ( – L2/8)*VwL* μwL = ( L2/8)*VwL* μwL CoV 0.1 0.2 0.08 0.05
GFy = - ( Zx)*VFy* μFy and GZx= - ( Fy)*VZx* μZx Desing values 10 25 2.45E+05 4.46E-04
 • We begin by assuming values for wD, wL, Fy, and then compute Zx Mean Values 10 25 2.45E+05 4.46E-04

using the g-function: Zx = (wD+wL)*L2/(8*Fy) PSF 1.1 1.2 1.12 1.1

• We assume the initial values of the design variables to be equal to


the mean values. WD WL Fy Zx
• In EXCEL this will look like as shown table. Note at this initial step, Bias 1.1 1.2 1.12 1.1

the PSF are simply = Bias* Design Values / Mean Values = Bias CoV 0.1 0.2 0.08 0.05
Desing values 10 25 2.45E+05 4.46E-04
Mean Values 10 25 2.45E+05 4.46E-04
• Then compute Gi , and Zi using the equations Gi = -Vi μi dg/dXid PSF 1.1 1.2 1.12 1.1
and and Zi = βo. Use the values of Xid and μi as assumed above. It Gi 3.125 15.625 -8.75E+00 -5.47E+00

will look like the shown table. alpha 0.164616023 0.82308 -4.61E-01 -2.88E-01
Z-val 0.576156081 2.88078 -1.61E+00 -1.01E+00
Example \ Cont. •  Then update the PSF = Bias*Design Values/Mean Values.
• Using the new design and mean values, recalculate Gi ,
and Zi and redo the updating of Xid and μi
• Then use the new Zi to estimate new design
• Keep repeating the calculations until the PSF converge to
values ( using Xid = μi (1+ Zi Vi ), ) and update fixed values. The WL
final result isZx shown
the mean values ( using μi = Xid / (1+ Zi Vi) ) . Bias
WD Fy
1.1 1.2 1.12 1.1

Remember that the last Design value { Here it CoV


Desing values
0.1
10
0.2
25
0.08
2.45E+05
0.05
4.46E-04
is Zx } must be computed using the function g = Mean Values 10 25 2.45E+05 4.46E-04

0, { i.e.; Zx = (wD+wL)*L2/(8*Fy) } Gi
PSF 1.1
3.125
1.2 1.12 1.1
15.625 -8.75E+00 -5.47E+00

WD WL Fy Zx alpha 0.164616023 0.82308 -4.61E-01 -2.88E-01

 • From the results:


Z-val 0.576156081 2.88078 -1.61E+00 -1.01E+00
Bias 1.1 1.2 1.12 1.1 Design values 10.57615608 39.4039 2.13E+05 7.32E-04
Mean values 10 25 2.45E+05 7.71E-04
CoV 0.1 0.2 0.08 0.05 PSF 1.163377169 1.891387 0.975454 1.044545 γDead = 1.15
Gi 3.125 15.625 -1.43E+01 -8.22E+00
Desing values 10 25 2.45E+05 4.46E-04 alpha 0.136080127 0.680401 -6.25E-01 -3.58E-01 γLive = 1.77
Z-val 0.476280443 2.381402 -2.19E+00 -1.25E+00

Mean Values 10 25 2.45E+05 4.46E-04


Design values
Mean values
10.47628044
10
36.90701 2.02E+05 7.33E-04
25 2.45E+05 7.81E-04
= 0.92*1.03= 0.95
PSF 1.152390849 1.771537 0.9240845 1.031062
PSF 1.1 1.2 1.12 1.1 Gi 3.125 15.625 -1.44E+01 -7.90E+00
alpha 0.13672177 0.683609 -6.28E-01 -3.46E-01
Gi 3.125 15.625 -8.75E+00 -5.47E+00 Z-val
Design values
0.478526196
10.4785262
2.392631 -2.20E+00 -1.21E+00
36.96315 2.02E+05 7.34E-04
Mean values 10 25 2.45E+05 7.82E-04
alpha 0.164616023 0.82308 -4.61E-01 -2.88E-01 PSF 1.152637882 1.774231 0.923014 1.033477
Z-val 0.576156081 2.88078 -1.61E+00 -1.01E+00 Gi 3.125 15.625 -1.44E+01 -7.89E+00
alpha 0.136610703 0.683054 -6.29E-01 -3.45E-01
Design values 10.57615608 39.4039 2.13E+05 7.32E-04 Z-val 0.47813746 2.390687 -2.20E+00 -1.21E+00
Design values 10.47813746 36.95344 2.02E+05 7.34E-04
Mean values 10 25 2.45E+05 7.71E-04 Mean values 10 25 2.45E+05 7.82E-04

PSF 1.163377169 1.891387 0.975454 1.044545 Gi


PSF 1.152595121
3.125
1.773765 0.922703 1.033605
15.625 -1.44E+01 -7.89E+00
alpha 0.136609464 0.683047 -6.29E-01 -3.45E-01
Z-val 0.478133125 2.390666 -2.20E+00 -1.21E+00
Design values 10.47813313 36.95333 2.02E+05 7.34E-04
Mean values 10 25 2.45E+05 7.82E-04
PSF 1.152594644 1.77376 0.9226803 1.033628

You might also like