Lecture 5: Monte Carlo Simulation for Time-

Independent Reliability Analysis

 Lecture #5
Outline

• Two Numerical Methods for Time-Independent Reliability

Analysis
• (Direct) Monte Carlo Simulation
• Expansion Methods (Optional)

2
 Time-Independent Reliability Analysis

Recall from Lecture #4: Some Basic Math

 Definition: the probability that the system performance meets
the required design performance under various uncertainty
sources (e.g., material properties, geometric tolerances, and
loading conditions).
Performance function
R ( X) = P G ( X) ≤ 0 = 1 − P G ( X) > 0 (e.g., stress, strain, and
( ) ( ) power output)
where X = (X1, X2,…, XN)T models uncertainty sources.

 The uncertainty of X propagates and causes the uncertainty of G.

 Equating G to zero, i.e., G = 0, produces the limit-state function
(LSF), which separates the safe region ΩS = {x: G(x) ≤ 0} from the
failure region ΩF = {x: G(x) > 0}.

3
 Time-Independent Reliability Analysis

Recall from Lecture #4: Some Basic

Schematic of time-independent
Math reliability analysis in a 2-D case
 The basic idea is to compute

X2: operational
the probability that X is inside G≤0
the safe region. Safe region fX(x)

R ( X) = P ( G ( X) ≤ 0 )

factor
= ⋯ S f X ( x ) dx
Ω

where fX(x) denotes the joint G>0

Failure
PDF of X, and the safe region region X1: manufacturing
ΩS = {x: G(x) ≤ 0}. tolerance

4
 Monte Carlo Simulation for Reliability Analysis

Four Sequential Steps (Assuming Independence between X)

 Step 1: For each random variable Xi, generate M “pseudo random numbers”
vi1, vi2,…, viM from a standard uniform distribution (between 0 and 1). Then
obtain M random vectors v1, v2,…, vM with vj = [v1j, v2j,…, vNj]T.
 Step 2: Transform vij (i = 1, 2,…, N, j = 1, 2,…, M) to xij using CDF mapping
xij = FX−i1 ( vij )
For example, if we have two standard normal random variables X1 and X2, the
transformation for either can be expressed as xi = Φ‒1(vi).
 Step 3: Evaluate G(X) at the random samples xj. This step requires M function
evaluations and produces M random function values G(xj), for j = 1, 2,…, M.
 Step 4: Extract the probabilistic characteristics of G(X), including statistical
moments, reliability, and PDF. For example, reliability can be estimated as
Confidence interval ( 3 ) 1 M Safety indicator function
R = E ( IΩS ( x ) ) ≈  IΩS ( x j )
 Rˆ MCS − 3σ ˆ , Rˆ MCS + 3σ ˆ  M j =1 1, x ∈ΩS
 RMCS RMCS  IΩS ( x) = 
0, x ∈Ω \ ΩS
Uncertainty of mean estimate : Mean
σ ˆ = Rˆ
RMCS MCS (
1 − Rˆ /M
MCS ) estimate by MCS
5
 Monte Carlo Simulation for Reliability Analysis

Schematic of Reliability analysis Using Direct Monte Carlo Simulation

G(X) > 0 Failed sample

X2
Failure region Safe sample

G(X) ≤ 0 G(X) = 0
Safe region Failure surface (LSF)

I ΩS ( x ) = 1 I ΩS ( x ) = 0

fX(x)
Mean point

0 X1

6
 Monte Carlo Simulation for Reliability Analysis

How to Generate Normal (Gaussian) Random Numbers

 Transformation of a standard uniform random variable Xi to a
standard normal random variable Ui
1.0 1.0
0.8 CDF mapping 0.8
0.6 0.6
ΦU i(ui)

FXi(xi)
0.4 0.4
0.2 0.2
0 0
ui Ui xi Xi
0.9
1.0
0.6
fU i(ui)

fX i (xi)

0.3

0 00
0 ui Ui xi 1 Xi

7
 Monte Carlo Simulation for Reliability Analysis

An Example: Beam Design under Uncertainty

 The free end of a cantilever beam is subjected to two transverse
loads X and Y along the orthogonal directions.
L Y
X
t
w
 The LSF with respect to the maximum stress is expressed as
 600 600 
G = σ (w, t, X ,Y ) − S y =  2 Y + 2 X  − S y = 0
 wt wt 
where Sy is the yield strength, and w and t are the design
parameters.
 The cantilever beam is deemed to fail if G > 0 (i.e., the maximum
stress σ is higher than the yield strength Sy).

8
 Monte Carlo Simulation for Reliability Analysis

An Example: Beam Design under Uncertainty

 Suppose X, Y, Sy, w, and t are independent random variables.
Random
X [lb] Y [lb] Sy [psi] w [in] t [in]
variables
Distribution Normal Normal Normal Normal Normal
Mean 500 1,000 400,000 2 1
Standard
100 100 20,000 0.02 0.01
deviation

 Determine the empirical PDFs of σ(w, t, X, Y) and Sy using direct

Monte Carlo simulation (MCS) with 1 million random samples.
 Let us first look at a simpler case, where only X is a random
variable and the others are constants.

9
 Monte Carlo Simulation for Reliability Analysis

An Example: Beam Design under Uncertainty

 Step 1: Define mean and standard deviation of X [lb]

>> u_X = 500;

>> s_X = 100;

 Step 2: Define values of the other parameters

>> % d1: Y [lb]; d2: Sy [psi]; d3: w [in]; d4: t [in]

>> d = [1000 400000 2 1];

 Step 3: Generate MCS sample points for X

>> ns = 1000000;
>> xx = normrnd(u_X,s_X,1,ns);

10
 Monte Carlo Simulation for Reliability Analysis

An Example: Beam Design under Uncertainty

 Step 4: Evaluate G function at MCS sample points

>> zz = 600/(d(3)*d(4)^2)*d(1) +....

600/(d(3)^2*d(4))*xx - d(2);

 Step 5: Estimate reliability based on G values at MCS sample

points

>> Rel_MCS = length(find(zz ≤ 0))/ns;

11
 Monte Carlo Simulation for Reliability Analysis

An Example: Beam Design under Uncertainty

 Step 6: Plot empirical PDF of σ [psi]

>> nbar = 80;

>> D = zz + d(2);
>> [heights,locations] = hist(D,nbar);
>> width = locations(2)-locations(1);
>> heights = (heights/ns)/width;
>> bar(locations,heights,'w'); hold on;
>> xlabel('Stress (psi)');
>> ylabel('Probability density');

12
 Expansion Methods for Reliability Analysis

Some Basic Math (Optional)

 Recall: Consider a performance function G as the difference
between the strength S of and load L on an engineered system
G = L−S
where S and L are random variables and follow normal
distributions with probability density functions fS and fL.
 By assuming statistical independence between S and L, we can
compute the reliability based on the standard normal CDF
 µ −µ  Probability distribution of G
R = Φ S L
 G ~ N ( µG ,σ G2 )
 σ 2 +σ 2 
 L S 
µG = µL − µS , σ G = σ L2 + σ S2
 Reliability is estimated based on the first two statistical moments
(mean and standard deviation).

13
 Expansion Methods for Reliability Analysis

Some Basic Math (Optional)

 The first-order Taylor series expansion of G(X) at the mean value

G ( X) ≈ G (μX ) + 
(
N ∂G μ
X)

i =1 ∂X i
(
X i − µ Xi )
= a1 X 1 + ⋯ + aN X N + b
This can be rewritten in a vector form as G(X) = aTX + b, where a =
[a1, a2,…, aN]T contains the first-order partial derivatives of G with
respect to X and is called a sensitivity vector of G.
 The first-order approximate mean and variance of G
µG = E [G ] ≈ E aT X + b  = aTμ X + b σ G2 = aT Σ Xa
 Assuming G is normal, the reliability can be estimated as
 µG 
R = Φ − 
 Gσ
14
 Expansion Methods for Reliability Analysis

An Example: Reliability Analysis of A Cantilever Beam (Optional)

 Group discussion (4 min): Assume that the vertical loads P1 and
P2 follow normal distributions with means 1000 lb and 500 lb,
and standard deviations 100 lb and 50 lb, respectively. In
addition, assume that the bending moment is deterministic (fixed
at 10,000 lb-ft).
P1 P2

m
10 ft 10 ft

 Compute the following using the first-order expansion method:

(1) the mean and standard deviation of the maximum moment at
the fixed end; and (2) the reliability with an allowable moment
ma = 33,000 lb-ft.

15
 Expansion Methods for Reliability Analysis

An Example: Reliability Analysis of A Cantilever Beam (Optional)

 Solution
 At the fixed end, the maximum moment can be expressed as
mmax = 10P1 + 20P2 + 10000 = aT X + b
where a = [10, 20]T, b = 10000, and X = [P1, P2]T. Note that the mean vector μX
= [1000, 50]T, and the covariance matrix ΣX = [1002, 0; 0, 502].
 We first compute the mean of mmax
1000
µm = aTμX + b = [10 20]   + 10000
 500 
max

= 30,000 lb-ft
 We then compute the variance and standard deviation of mmax

1002 0  10 
σ 2
mmax = a ΣXa = [10 20] 
T
2 
 0 50  20
2
= 2 ×106 [ lb-ft ]
σ m = 2 ×106 ≈ 1414.2 lb-ft
max
16
 Expansion Methods for Reliability Analysis

An Example: Reliability Analysis of A Cantilever Beam (Optional)

 Solution
 Finally, the reliability can be computed as
Performance function
R = P ( mmax ≤ ma )
G = mmax − ma
 mmax − µmmax ma − µmmax 
= P ≤ 
 σm σ
 max mmax 
 33,000 − 30,000 
= P z ≤ = 2.12 
 1414 
= Φ ( 2.12) = 98.3%

 Note that the performance function is a simple linear function of normally

distributed random variables P1 and P2. Therefore, the reliability estimate
obtained from the first-order expansion method is exact.

17