Lecture 8: Reliability-Based Design Optimization

Lecture #8
Outline

• Example of Reliability-Based Design

• Design Methodology
• Design Formulation
• Optimization Results

• Handling Uncertainty in Design

• Formulation of Reliability-Based Design

2
 Reliability-Based Design

Design of Vehicle Control Arm: Design Methodology

Deterministic
Optimum
1st hotspot
[Life at 1st hotspot=Limit]
Failure Surface G1 = 0
X2: (thickness of

Initial
comp. 2)−1

Design 2nd hotspot

[Lives at both hotspots ≥ Limit]

Safe Region
0 X1: (thickness of
comp. 1)−1 Failure Surface G2 = 0
[Life at 2nd hotspot=Limit]

Hu C., and Youn B.D., “Adaptive-Sparse Polynomial Chaos Expansion for Reliability Analysis and Design of Complex Engineering
Systems,” Structural and Multidisciplinary Optimization, v43, n3, p419‒442, 2011.
3
 Reliability-Based Design

Design of Vehicle Control Arm: Design Methodology

1st hotspot
[Life at 1st hotspot=Limit]
Failure Surface G1 = 0
X2: (thickness of
comp. 2)−1

2nd hotspot

[Lives at both hotspots ≥ Limit]

Safe Region
Weight (cost)
0 X1: (thickness of
comp. 1)−1 Failure Surface G2 = 0 Reliability
[Life at 2nd hotspot=Limit]

Hu C., and Youn B.D., “Adaptive-Sparse Polynomial Chaos Expansion for Reliability Analysis and Design of Complex Engineering
Systems,” Structural and Multidisciplinary Optimization, v43, n3, p419‒442, 2011.
4
 Reliability-Based Design

How to Formulate this Design Problem?

Find design variables d = μ(X)

To minimize the design cost C(d)
Subject to
Reliability at the first hot spot R1(X; d) ≥ Rreq
Reliability at the second hot spot R2(X; d) ≥ Rreq
Other requirements (constraints)

5
 Reliability-Based Design

Design of Vehicle Control Arm: Optimization Results

Initial Optimum

Weight 31.473 32.717

Reliability

R6 0.3235 0.9987

R42 1.0000 1.0000

R43 1.0000 1.0000

Initial Stress
Contour
X3 X4 X2 R46 1.0000 1.0000
Initial design 91 hotspots
X7 (reliability
X1 R80 0.0050 0.9989
constraints)
X5 X6 Final Stress
Contour R87 1.0000 0.9991

Optimum design 0 99.87% 1

6
 Handling Uncertainty at Three Steps in Design

Model (PDF) Pf = Pr ( G > 0 )

Probability
Probability

Probability
Design 2

density
density

density
Data Design 1

Design variable 0 G 0 G

Step 1: Uncertainty Step 2: Uncertainty Step 3: Design under

modeling analysis uncertainty

• Probability theory • Reliability analysis • Reliability-based

• Statistics • Robustness analysis design
• Other theories • Sensitivity analysis • Robust design

Design process

PDF: Probability density function 7

 Reliability-Based Design

How to Formulate a General Reliability-Based Design Problem?

Formulation of reliability-based design optimization (RBDO)
Minimize f ( d )
Subject to R j = Pr ( G j ( X, Θ; d ) ≤ 0 ) ≥ Φ ( β t ) = R t , j = 1,..., N c
diL ≤ di ≤ diU , i = 1,..., N d

– f is the objective function, often the design cost (e.g., volume and mass)
– Random design vector X = (X1,…, XNd)T and random parameter vector Θ =
(θ1,…, θNr)T
– d = (d1,…, dNd)T = µ(X) is the mean design vector (or the design variable set)
– Rj is the reliability level of a given design d with respect to the jth design
constraint, for j = 1,…, Nc
– Gj is the performance function of the jth design constraint, for j = 1,…, Nc
– Rt is the target reliability level (i.e., a target reliability index of βt)
– diL and diU are respectively the lower and upper bounds on di, for i = 1,…, Nd.

8
 Reliability-Based Design

How to Formulate a General Reliability-Based Design Problem?

Formulation of reliability-based design optimization (RBDO)
Minimize f ( d )
Subject to R j = Pr ( G j ( X, Θ; d ) ≤ 0 ) ≥ Φ ( β t ) = R t , j = 1,..., N c
diL ≤ di ≤ diU , i = 1,..., N d

– f is the objective function, often the design cost (e.g., volume and mass)

9
 Reliability-Based Design

How to Formulate a General Reliability-Based Design Problem?

Formulation of reliability-based design optimization (RBDO)
Minimize f ( d )
Subject to R j = Pr ( G j ( X, Θ; d ) ≤ 0 ) ≥ Φ ( β t ) = R t , j = 1,..., N c
diL ≤ di ≤ diU , i = 1,..., N d

– f is the objective function, often the design cost (e.g., volume and mass)
– Random design vector X = (X1,…, XNd)T and random parameter vector Θ =
(θ1,…, θNr)T

10
 Reliability-Based Design

How to Formulate a General Reliability-Based Design Problem?

Formulation of reliability-based design optimization (RBDO)
Minimize f ( d )
Subject to R j = Pr ( G j ( X, Θ; d ) ≤ 0 ) ≥ Φ ( β t ) = R t , j = 1,..., N c
diL ≤ di ≤ diU , i = 1,..., N d

– f is the objective function, often the design cost (e.g., volume and mass)
– Random design vector X = (X1,…, XNd)T and random parameter vector Θ =
(θ1,…, θNr)T
– d = (d1,…, dNd)T = µ(X) is the mean design vector (or the design variable set)

11
 Reliability-Based Design

How to Formulate a General Reliability-Based Design Problem?

Formulation of reliability-based design optimization (RBDO)
Minimize f ( d )
Subject to R j = Pr ( G j ( X, Θ; d ) ≤ 0 ) ≥ Φ ( β t ) = R t , j = 1,..., N c
diL ≤ di ≤ diU , i = 1,..., N d

– f is the objective function, often the design cost (e.g., volume and mass)
– Random design vector X = (X1,…, XNd)T and random parameter vector Θ =
(θ1,…, θNr)T
– d = (d1,…, dNd)T = µ(X) is the mean design vector (or the design variable set)
– Rj is the reliability level of a given design d with respect to the jth design
constraint, for j = 1,…, Nc

12
 Reliability-Based Design

How to Formulate a General Reliability-Based Design Problem?

Formulation of reliability-based design optimization (RBDO)
Minimize f ( d )
Subject to R j = Pr ( G j ( X, Θ; d ) ≤ 0 ) ≥ Φ ( β t ) = R t , j = 1,..., N c
diL ≤ di ≤ diU , i = 1,..., N d

– f is the objective function, often the design cost (e.g., volume and mass)
– Random design vector X = (X1,…, XNd)T and random parameter vector Θ =
(θ1,…, θNr)T
– d = (d1,…, dNd)T = µ(X) is the mean design vector (or the design variable set)
– Rj is the reliability level of a given design d with respect to the jth design
constraint, for j = 1,…, Nc
– Gj is the performance function of the jth design constraint, for j = 1,…, Nc

13
 Reliability-Based Design

How to Formulate a General Reliability-Based Design Problem?

Formulation of reliability-based design optimization (RBDO)
Minimize f ( d )
Subject to R j = Pr ( G j ( X, Θ; d ) ≤ 0 ) ≥ Φ ( β t ) = R t , j = 1,..., N c
diL ≤ di ≤ diU , i = 1,..., N d

– f is the objective function, often the design cost (e.g., volume and mass)
– Random design vector X = (X1,…, XNd)T and random parameter vector Θ =
(θ1,…, θNr)T
– d = (d1,…, dNd)T = µ(X) is the mean design vector (or the design variable set)
– Rj is the reliability level of a given design d with respect to the jth design
constraint, for j = 1,…, Nc
– Gj is the performance function of the jth design constraint, for j = 1,…, Nc
– Rt is the target reliability level (i.e., a target reliability index of βt)

14
 Reliability-Based Design

How to Formulate a General Reliability-Based Design Problem?

Formulation of reliability-based design optimization (RBDO)
Minimize f ( d )
Subject to R j = Pr ( G j ( X, Θ; d ) ≤ 0 ) ≥ Φ ( β t ) = R t , j = 1,..., N c
diL ≤ di ≤ diU , i = 1,..., N d

– f is the objective function, often the design cost (e.g., volume and mass)
– Random design vector X = (X1,…, XNd)T and random parameter vector Θ =
(θ1,…, θNr)T
– d = (d1,…, dNd)T = µ(X) is the mean design vector (or the design variable set)
– Rj is the reliability level of a given design d with respect to the jth design
constraint, for j = 1,…, Nc
– Gj is the performance function of the jth design constraint, for j = 1,…, Nc
– Rt is the target reliability level (i.e., a target reliability index of βt)
– diL and diU are respectively the lower and upper bounds on di, for i = 1,…, Nd.

15
 Reliability-Based Design

How to Solve an RBDO Problem?

Reliability index approach (RIA)
Minimize f ( d )
Subject to G rj = β t − β js ≤ 0, j = 1,..., N c
diL ≤ di ≤ d iU , i = 1,..., N d

Each probabilistic constraint is represented by using a reliability index.

The reliability index in RIA can be evaluated using FORM.
The first-order estimate of the reliability index is defined as the distance
between the MPP and origin, βis,FORM = ||uj*||.
Because of its simplicity and efficiency, the HL–RF method is often employed
to search for the MPP in RIA.

16
 Reliability-Based Design

How to Solve an RBDO Problem?

Design point located by RIA at the kth iteration (both X1 and X2
follow normal distributions with same mean and variance)

Safe region
X2
G1 ≤ 0 & G2 ≤ 0

β1(X;dk) < βt β2(X;dk) > βt

dk

G2 = 0
x1*
G2 > 0
G1 > 0 Failure
Failure x2*
G1 = 0 region
region
X1

17
 Reliability-Based Design

How to Solve an RBDO Problem?

Optimum design located by RIA (both X1 and X2 follow normal
distributions with same mean and variance)

Safe region
X2
G1 ≤ 0 & G2 ≤ 0

β2(X;dFinal) ≈ βt
β1(X;dFinal) ≈ βt

dFinal G2 = 0

G2 > 0
G1 > 0 x2 * Failure
Failure x1 * G1 = 0 region
region
X1

18
 Reliability-Based Design

Example on RBDO with RIA

Consider a mathematical problem for design optimization. The
problem has two design variables X = [X1, X2]T and involves three
performance functions that are defined as follows:
G1 ( X) = X12 X 2 / 20 − 1
2 2
G2 ( X) = ( X1 + X 2 − 5) / 30 + ( X1 − X 2 − 12) /120 − 1
G3 ( X) = 80 / ( X12 + 8 X 2 + 5) − 1
The two variables follow independent normal distributions with
means d = [d1, d2]T and standard deviations 0.3 and 0.3. A
deterministic design optimization problem is formulated as
follows.
Minimize f ( d ) = d1 + d2
Subject to G j ( d ) ≤ 0, j = 1,2,3
1 ≤ d1 ≤ 10, 1 ≤ d2 ≤ 10

19
 Reliability-Based Design

Example on RBDO with RIA

The RBDO problem with a target reliability of 99.87% (i.e., a
target reliability index of βt = 3.0) is formulated as follows
Minimize f ( d ) = d1 + d2
Subject to Pr ( G j ( X; d ) ≤ 0) ≥ Φ ( β t ) , j = 1,2,3
0 ≤ d1 ≤ 10, ?0 ≤ d2 ≤ 0

The initial design is d(0) = [5.0, 5.0]T.

1) Solve the deterministic design optimization problem using the
‘fmincon’ function in MATLAB.
2) Solve the RBDO problem using RIA with the HL-RF method by
modifying the 99-line RBDO code (downloadable under the
“Homework 3” assignment). Start the design optimization from
both the initial design d(0) = [5.0, 5.0]T and deterministic optimum
design obtained from 1).

20