Structural Safety and

Reliability Index
Structural Safety
• Safety is a relative term because it depends on the “purpose” of the structure.
• Generally, we define “failure” to be the state at which an intended purpose is
violated. The structure is “safe” if that purpose is satisfied.
• There are many types of “failures” :
• Ultimate (or strength ) failure: the member or structure loses structural capacity to
resist load effect.
• Serviceability failure: The member or structure loses architectural functionality
• Fatigue failure: The member suddenly loses capacity under small cyclic loads
• To indicate each of these failures, a mathematical statement (or engineering
model) must be written, e.g. :
• For strength: Resistance > Action (or R>S  R- S > 0).
• For serviceability : Max Deflection > Limit Deflection, or Max Vibration Frequency >
Permissible Frequency, or Max Crack Width > Permissible Crack Width
• For fatigue: Max cyclic stress > Endurance Limit, … etc
• Each one of these statements is called a “Limit State”
Performance Functions
• Each Limit State can be written into a function form, for instance for
Strength: R = S  g = R-S. This function is called a Performance
Function.
• Note that because the components of R and S are in fact random
{stochastic}, the function g is also random {stochastic}.
• The Limit State is satisfied at g = 0. This is called the Failure Surface.
• The failure state is defined to be all the cases that give g < 0 { or g > 0 if
the function is reversed.
• The Probability of Failure (P.o.F) is defined to be the ratio of all the
failure state cases to all possible cases of the structure during its life
span.
• The formal definition of probability of failure is P.o.F = P(g < 0 )
 Probability of Failure
• If g is a function of a single variable “X” then we can show a 1D representation of the PoF. The
calculations then are simple.
The area of Failure
where R < S

Frequency
Probabilistic Probabilistic
Probabilistic
Distribution of “R” Distribution of g
Distribution
Frequency

of “S”
P.o.F = P(g<0)
Increasing
Values of g

Increasing
μS = Mean Value μR = Mean Value of ”μg = Mean Value of “g
Values of X
of “S” S=R ”“R g=0

• If the R and S contain vectors of n-variables, then the

Joint PDF for the n-variables becomes an n-
dimensional surface.
• The calculation of probability of failure becomes the
matter of computing n-dimensional volumes under the
n-dimensional surface. It becomes quite complex.
Techniques for Computing PoF
• The topic of finding the PoF for multivariable performance function has
been (and still is) a hot research topic.
• Many methods exist for such goal. These can be grouped into 2 types
• Simulation Techniques
• Analytical Techniques
• In the simulation techniques, the PoF is directly computed by directly
simulating the performance function using random variables.
• In the analytical techniques, approximate methods are used and the
random distributions are approximated as normal distributions or even
using only mean and variance information.
• Most famous simulation techniques is a family of methods called Monte
Carlo Simulations, the Latin Hyper Cube Methods, Rosenbluth method
and many others.
Monte Carlo Simulations (MCS)
• In the MCS, we treat the performance function as a stochastic function.
• For each variable in the limit state function, a sufficiently large number of
random values are generated for the variable using its PDF.
• The RVs are substituted into the limit state function, and the failure is
assessed through checking if g < 0 or not.
• The probability of failure PoF is simply computed as the total number of
failure cases { n } (all cases with g < 0 ) divided by the total number of
trial made for the function { N } , i.e.: P.o.F = n / N.
• The MCS is used for cases of complex nonlinearities and complex
combinations of RVs that are difficult to simplify.
• However, MCS requires a lot of computational power because for good
accuracy, we generally need thousands of calculations to be done.
Sci-hub.com
Libgen.is
URL, or DOI.
 Example: MCS
• A simple steel beam is subjected to uniform dead and live Item PDF Mean CoV
loads. The bending nominal plastic capacity is given as Mn Dead Load -- 10 kN/m 0
= FyZx. The properties of Fy and Zx in addition to the live load
Live Load Gamma 20 kN/m 0.2
are given in the table. The length of the beam (L=5 m) and
the dead load will be considered deterministic. Calculate the Fy Normal 250 MPa 0.1
beam probability of failure using Monte Carlo Sim. Zx Normal 450E+03 mm3 0.15
• The design limit state for this case is R = S or Mn= Ma , FyZx = wL2/8 , and hence the
performance function becomes: g = FyZx – (wD+wL)*L2/8, here wD and L are deterministic.
• If we do a quick check using the mean values, we find that the design “seems” safe:
g = 250*450e3/1e6 – (10+20)*25/8 = 18.75 kN.m > 0
• We can even compute the Central Safety Factor SFc = RAve / SAve = 1.2
• However, we know that each variable is stochastic, thus, despite that the design is “safe” but
we don’t know “how much safe!!” it is. That is: we don’t know how “reliable” it is!
• To get the PoF by MCS, we need to use EXCEL for calculations
• First we need to get the parameters of the PDFs.
• For Live Load: Gamma Distribution ϴ = CoV2*μ=0.8 , k =( 1/CoV)2=25
• For Fy: Normal Distribution: μ = 250, σ = 25
• For Zx: Normal Distribution: μ = 450e3, σ = 67.5e3
 Results of the Example
• We need to repeat the MCS calculations hundreds or even thousands
of times. The higher the number of simulation points, the better the
estimate for the PoF
• The figure shows how the PoF changes as we increase the number of
MCS points.
• This drawing is called a “Sensitivity Drawing”, and it shows how
sensitive the PoF to the number of simulated points.
• The curve should converge to a constant value as Number of MCS
sim points approaches infinity.
We also need to plot the
• Here the PoF is nearly 0.23, Frequency Distributions
• i.e. there is 23% chance the beam will fail. for each variable, to
• A PoF = 0.23 is very large, and thus the design make sure we got them
is totally not reliable!! And we can surely say right!!
that this beam (With its cross-section and
material) is totally unsafe { From a Reliability
point of view }.
• Recall that the Central Factor of Safety is 1.2 ,
which gives a “false” sense of safety for the
given beam.
Acceptable PoF for Structural Safety
• Generally, the acceptable PoF depends on the type of structure, the risk of failure, the
consequences of failure and how uncertain we are about the mathematical model (or
limit state function g).
• The Acceptable PoF is called The Limit PoF = PoFLimit
• For Columns (R.C and Steel), PoFLimit can be as low as 0.0001 (for material strength
failure) to 0.00001 (for inelastic buckling failure)
• For Beams (R.C. and Steel), PoFLimit =0.0001
• For Masonry Walls and Steel Connections, PoFLimit =1e-6 to 1e-7
• A Reliable Design is the design that uses stochastic analysis such that PoF < PoFLimit
• For a structure, if all members are designed to have the same PoF, we say that the
structure has Homogenous (or Uniform) Failure State. This can be quite expensive
design (not economical).
• Generally, the structures are designed to have uniform “Risk” , or uniform “Hazard”.
This means that each member is given a PoFLimit that depends on how risky or
hazardous its failure.
Accuracy of Estimated PoF by MCS
• The
  PoF depends very much on the number of simulation points for MCS.
• Theoretically, we can only obtain the exact PoF when N  ∞, we call this exact PoF
the True PoF = (PoF)True
• Each time we do an MCS, we are producing a sample from the population of the
MCS. The values of PoF are called Sample PoF = (PoF)Sample
• Based on Estimation Theory in Probability the relation between the mean and
variance of a sample and the true value of PoF is:
E( PoFSample) = PoFTrue , and Var(PoFSample) = (PoFTrue)*(1-PoFTrue)/N
• Using these estimates, , inverting this relation gives the required number of MCS
points for a certain level of accuracy

• For the previous example, if we want to detect a PoF of 0.2, with a CoV for the
estimates of PoF equals 0.1, then Nreq = 400 points. If CoV of the sampled points is
to taken 0.01, then Nreq = 40,000 points!!
 MCS for Correlated Variables
•• In
  the previous example, we treated the variables to be totally independent, and thus no correlation existed
between the variables. Thus we generated the random data assuming total independence.
• Many times the variables are correlated, and the correlation must be preserved during the MCS.
• A method to preserve correlation is to convert the correlated variables into un-correlated through an eigen
transformation.
• The Covariance Matrix , represents the covariances of the original (correlated) variables, Var( i, j) =
ρij×σi×σj where ρij is the correlation coefficient between variables i, j. For i=j=1, then ρij = 1 and Var(1,1) = (σ1)2.
Notice for correlated variables, the off-diagonal terms may not be zero. The non-correlated variables must
have all the off-diagonal values to be zero. Cx is symmetric matrix.
• The eigen transformation gives a transformation matrix T such that Cy= T-1CxT , where Cy is the variances of
the un-correlated variables. Note that Cy is always diagonal matrix (i.e. uncorrelated)
• The mean values of the un-correlated variables Y is then computed using ΩY = T-1 ΩX where ΩX is the
vector of mean values of the correlated variables X, and ΩY is the new mean values of the un-correlated
variables Y.
• Now using the Cy and ΩY matrices, we can generate uncorrelated random variables Y for the MCS. Then a
transformation is done backward to bring back the values of X ( X = T·Y ) for computation of the
performance function
The Rosenbluth Point Estimate Method (RPE)
• In
  the RPE, the performance function is approximated using point estimates forits
mean and variance. So, RPE does not require the PDF for the variables. Thus, it is a
Level II reliability method.
• Assume the limit-state function depends on N variables { g(X1,..XN) }
• Compute the value of g at the mean values of all Xi and call this Yo
Yo = g(X1 = μ1 , X2 = μ2, … , Xn = μn )
• For each variable Xi, compute the performance function at two points (μi+σi) and at
(μi-σi) while keeping the other variables at their mean values. i.e.:
Yi+ = g ( X1 = μ1 , …, Xi = μi+ σi , … , Xn = μn )
Yi– = g ( X1 = μ1 , …, Xi = μi – σi , … , Xn = μn )
• Compute the factors Yi =( Yi+ + Yi– )/2 , Vi =( Yi+ – Yi– )/( Yi+ + Yi– )
• The mean and C.o.V of g is then and
• Once you have and you can estimate the PoF using these information assuming
Normal Distribution for P(g < 0 )
 Example: RPE
Item PDF Mean CoV σ
• Take the previous example, estimate the PoF Dead Load -- 10 kN/m 0 0
Live Load Gamma 20 kN/m 0.2 4
using the Rosenbluth Method
Fy Normal 250 MPa 0.1 25
g = R – S = FyZx – (wd+wL)*L^2/8 450E+03
Zx Normal mm3 0.15 67.5e3
• The calculations are straightforward
and easy to put to EXCEL.
• The resulting PoF is 0.265 , which is
Item Means Live+s Live - s Fy + s Fy - s Zx + s Zx -s
close to the PoF estimated using MCS
Live 20 24 16 20 20 20 20
(0.23)
Fy 250 250 250 275 225 250 250
• The advantage is the very low number Zx 4.50E+05 4.50E+05 4.50E+05 4.50E+05 4.50E+05 5.18E+05 3.83E+05
of calculations, compared to MCS g 18.75 6.25 31.25 30 7.5 35.625 1.875
• Because of its efficiency, this method Yi 18.75 18.75 18.75
can be used for cases when the Vi -0.66667 0.6 0.9
performance function is not analytic, 1+Vi^2 1.444444 1.36 1.81
such as a finite element model that mu g 18.75
gives the failure states. CoVg 1.598638309
Sigma g 29.9744683
• The method was originally developed PoF 0.265810857
for variables of normal distributions, but
can be approximately used for other
cases.
 Reliability Index { R.I. }
  is a fundamental concept in reliability, and it is used interchangeably with the
• R.I.
probability of failure.
• Assume the performance function is g = R – S. Assume also that R and S are random
variables, each with known means (μR and μS) and known standard deviations (σR and
σS )
• Now assuming that R and S are independent, then g is also a random variable
whose mean value and standard deviation are given as:

• Further assuming that R, and S are normally distributed variables, then g is also
normally distributed variable, then the CDF for g can be used to estimate P.o.F as
PoF = P(g<0) = F(0) = Φ( (0 – μg )/σg ) = Φ(– μg /σg ).
Where F(x) is the CDF of g, and Φ is the CDF of the standard normal distribution.
• Cornell (1968) made the definition that – μg /σg = β = The Reliability Index. It is
called Cornell Reliability Index {βCornell}. Thus, PoF = Φ(-β), or β = - Φ-1(PoF)
Reliability Index { R.I. } Probabilistic
P.o.F = P(g < 0) Distribution of g
• The definition of the R.I. by Cornell can be
visualized as the number of standard
β × σg
deviations below the mean of (the assumed) ”μg = Mean Value of “g

normally distributed “g” to the Standard

Normal
• It gives an indication to the PoF and it is more P.o.F = P(g < 0) μ =0
Distribution
commonly used than PoF in reliability and safety Origin of Standard
β
analysis. Normal PDF

• Generally structures are designed for a specific

target reliability index
• If the calculated reliability index is smaller than
the target then the structure is not reliable.
• Note the definition of R.I. depends on two values
for each variable (Mean and Variance), for this,
it is called “Second Moment Reliability Index”
Target Reliability Index
• Target R.I. (βo) depends on the type of failure, and its consequences and its
risk or impact, and the uncertainty associated with computing the failure.
 Example:
Steel Connection may fail in fracture when the tensile force P exceeds the fracture
  Mean CoV
t
Fu (Mpa) 400 0.01
capacity Pr = U*Anet*Fu. Find the Cornell reliability index for a member connection whose
properties are shown in the table. Assume all variables to be normally distributed and Anet (mm2) 3000 0.12
independent. Assume the shear-lag factor U = 0.6. Assume the model uncertainty for the Pt (kN) 200 0.15
resistance is 0.8 and its CoV is 0.15.
•• The
  limit state function is g = R-S = UAnet*Fu – Pt .It is true that all variables are normal, however, the
function g may not be normal, but we will assume it is normal.
• First get the statistics of the resistance. R = UAnetFu is the nominal resistance, and this must be
multiplied by the model, material and fabrication variations. Here only the model variation is given.
Thus R = UAnet*Fu*P. Also, despite the parameters of the Resistance are normal, however, the
resistance itself maynot be normal, but we will assume it to be normal.
• The mean resistance is μR = U×μAnet×μFu×μP ×μM ×μF =0.6*3000*400*0.8*1*1 = 576 kN.
• The CoV of the resistance
• The standard dev of R is σR = μR×VR = 109.4 kN.
 PoF = CDF Stand Norm P(g<0) = CDF(-3.3)
• This value of R.I. is small for a steel connection (should be at least 4.0!).
• How to increase ? This is a design question. We can increase Anet! This is equivalent to increasing the
mean value of Anet . Or.. We can reduce the CoV of the resistance (by requiring better quality control)
First-Order, Second-Moment Mean-Value Reliability {FOSM }

•• In
  case the limit state function is linear combination of variables
i.e., g(X1,..Xn) = ao + a1*X1 + … + an*Xn, then we can generalize the definition of β into:

• In case the function “g” is nonlinear, then we can linearize the function “g” by Taylor
Expanding it around the mean values. Then β becomes:

Here , and the partial derivative is evaluated at the mean values of all the variables Xi
• This definition is First-Order (because we took only the linear terms in Taylor expansion) and
Second-moment, because we used statistical information up to the variance (second
moment) and Mean-Value because we evaluate the derivatives at mean values.
Non-invariance of Cornell Reliability Index
• The definition of the R.I. as given by Cornell has a severe problem. It is not invariant!
• This means that the answer (for β) depends on the shape of “g” function.
• Consider the example of a simply supported beam   Mean Variance
with a mid-span force P. The properties of the beam Force P kN 10 4
Length L m 8 1.00E-02
are given in the table.
Plastic Mod. Z m 1.00E-04 4.00E-10
3
• There are many ways to define the performance
function “g”, but lets consider two ways: Fy Mpa 600 1.00E+03

g1 = ZFy - PL/4 and g2 = Fy – PL/(4Z)

• If we do the calculation, then we get two answers: β1= 2.48, and β2 = 3.48
• Obviously, this is a big problem and the cause for this discrepancy is the nonlinear
term in the performance function g2. The nonlinearity causes a significant shift from
the Normal Distribution of g.
Hasofer-Lind Reliability Index
• The non-invariance problem of Cornell Reliability index can be solved by
redefining the Reliability Index to be the shortest distance between the curve
of “g = 0” to the origin of the normalized variables
• This definition is called Hasofer-Lind Reliability Index.
• Consider random variables X1, … , Xn. The standardized (or normalized or
reduced) random variables are Z1, … , Zn , where Zi = (Xi-μi)/σi
• The limit state function g(X1, … , Xn ) is re-written in terms of Zi by substituting
Xi= μi +σi Zi thus becoming g(Z1, … , Zn )
• The Hasofer-Lind Reliability Index (βH-L ) is
defined to be shortest distance between the curve
g(Z1, … , Zn )=0 and the point Z1=Z2=…=Zn=0 in
the standardized space of variables.
• The point on the failure surface that gives βH-L is
called the “Design Point”.
 Computation of Hasofer-Lind Reliability Index.
• Finding
  the shortest distance between g=0 and the origin of the PDF for the g is
not simple and requires iterations for the case of nonlinear g.
• Many methods can be used for calculating βH-L.These methods are all
minimization techniques for finding the shortest distance.
• We can use EXCEL to do the calculations through its SOLVER.
• First, we need to assume initial values for the random variables but make sure
they all satisfy g = 0.
• This can be done by assuming n-1 { X1 , … , Xn-1 } variables, then solve for the
last {Xn} variable through imposing the condition g = 0.
• The assumed variables are transformed into the reduced (or standardized) space
using Zi = (Xi-μi)/σi
• The Hasofer-Lind Reliability Index βH-L is then computed using the SOLVER by
minimizing the distance between the point of all Zi to the origin of the reduced
space, i.e. βH-L = Minimize
 Example: H-L Reliability
Calculate the Hasofer-Lind reliability index for the 3-span continuous
beam shown. Take the serviceability limit state Max Def > L/360.
For 3-span beams, the max deflection = 0.0069wL4/EI. Use the given Variable Mean Std Dev
properties for the random variables. Assume them all to be Normally L 5m 0
Distributed. w 10 kN/m 0.4
E 2e7 kN/m2 5.00E+06
I 8.00E-04 1.50E-04
• The Performance Function is written as
• Here we have 3 random variables. We will assume values for two of them and find the third using
the relation g = 0.
 • So, Assume initial values (Use mean values); E = μ =
E

2e7, I = μI = 8e-4, then

• Transform these values into the standard form: ZE = 0,
ZI = 0 , and Zw = 104
• The distance is
• This process is easy to setup in EXCEL, as shown
• Use SOLVER, to minimize by changing values of E and
I. Note that w must always be computed by g = 0.
Solving βH-L Using EXCEL SOLVER
• The SOLVER can be added to EXCEL through going to Files > Options, then click
on Add-ins, then go to Manage: Excel Add-ins, click on Go.. Then check the Solver
Add-in
 Solving βH-L Using EXCEL SOLVER
• The SOLVER is then found under “Data” Tab,

• Doing the SOLVER!

Gives you βH-L = 3.18
• The design point is:
E = 4.37e6
I = 7.14e-4
W = 0.109
R.I. for Non-Normal Random Variables
• The Cornell and Hasofer-Lind Reliability Indices are all based on normalizing
(standardizing) the random variables.
• The main assumption there is that the R.V are all Normally Distributed.
• Most of the times, the random variables are not normally distributed, and thus
a modification must be made to account for this
• Only Simulation Techniques (like MCS or RPE) can consider “exactly” the non-
random variables.
• If we will use the Hasofer-Lind Procedure, then an approximation is made to
standardized the non-normal variables
• The procedure is called Rackwitz-Fiessler {R-F} Normalization Procedure
• In the R-F: the R.V. is transformed to become equivalent to a normal R.V.
• The basic idea is to have the CDF of the non-normal RV equal to CDF of a
standard normal variable with equivalent mean and variance at the given value
of R.V.
Rackwitz-Fiessler {R-F} Normalization
•  Specifically, if X is not normally distributed. The PDF and CDF for X are fx
and FX. Then the equivalent mean and variance for X is found by
imposing the condition that:
where FX is the CDF for the non-normal X, and is the CDF for the
standard normal distribution. The and are the “equivalent” mean and std
dev for the “equivalent” normal distribution at point X.
• Taking derivative dFX/dx of the previous equation we get , where fx is the
PDF for X and is the PDF for the standard normal variable.
• The two equations can be solved together and we get:

• Notice that the equivalent and keep changing as X itself change.

• The standard values are then computed as usual but using and
Example: H-L R.I. with non-Normal R.V.
• The
  previous 3-span cantilever beam will be solved but this time ALL the
random variables will all be assumed to follow Gamma Distribution.
• The only change here is the computation of the new standardized values.
With R-F procedure,
• For the Gamma Distribution, the parameters are computed as:
ϴ = (σx )2/ μx and k = (μx)2/(σx )2 in EXCEL alpha=k, and Beta = ϴ
• Using EXCEL, the value of Psi is computed as:
Psi = NORM.INV( GAMMA.DIST( X , (μx/σx )2 , (σx )2/ μx , TRUE ) ,0 , 1)
• Doing the optimization using EXCEL
SOLVER, we find Beta = 4.427
• Note that using Normal Distributions ,
the value of Beta = 3.18, which is
“generally” conservative.
 Modifications for Correlated Variables
• In
  some cases, the R.V. are correlated. For instance, variations in span lengths
in frames, or wind and snow loads, dead loads and EQ loads..etc.
• The way to deal with these variables is to do an eigen transformation of the
correlated variables X to a space of non-correlated variables Y.
• Define the the Covariance Matrix , which represents the covariances of the
correlated “X” variables, Var( i, j) = ρij×σi×σj where ρij is the correlation
coefficient between X variables i, j. For i=j=1, then ρij = 1 and Var(1,1) = (σ1)2.
• The eigen transformation matrix T is defined as: Cy= TTCxT , where Cy is the
variances of the un-correlated variables. Note that Cy is always diagonal matrix
(i.e. uncorrelated)
• The “new” mean values ΩY of the un-correlated Y’s are computed ΩY = TT ΩX
where ΩX is the vector of mean values of the correlated variables X.
• Now we can use the Cy and ΩY to do the standardization of the variables in H-L