Efficient Methodologies

for Reliability Based

Design Optimization

Harish Agarwal
Department of Aerospace and Mechanical Engineering
University of Notre Dame, IN - 46556
Email :- [email protected]

Presentation for Dr. Yoshimura and his research group

Overview

 Introduction.
 Deterministic design optimization.
 Motivation.
 Background on Reliability-based design optimization (RBDO).
 New Unilevel Method for RBDO.
 Example Problem.
 Summary.

2
Multidisciplinary Systems Design

 Linked simulation tools.

Elastic
 Highly coupled. Suspension
Structures
 Complex information exchange.
Aero- Crash-
 Computationally expensive. dynamics worthiness
 Highly nonlinear.
Occupant Fuel
Dynamics Economy

3
 Deterministic Design Optimization

Deterministic Optimum
90 Reliable Optimum
70
Note that almost 75%
50 of designs around the
deterministic optimum
fall in the failure domain
and hence fail.

 Optimized deterministic designs are driven to the limit of the design constraints
and can lead to catastrophic failure.

 A variety of different types of uncertainties are inherently present in engineering

design.

 Deterministic design optimization does not account for the uncertainties.

 Therefore, there is a strong need for optimization under uncertainty.

4
 Reliability Based Design Optimization (RBDO)

 In RBDO, the deterministic problem is reformulated and the failure driven

constraints are replaced with reliability constraints.

 The reliability constraints can be formulated by the reliability index approach (RIA)
or the performance measure approach (PMA).
 In RIA, are formulated as constraints on the probability of failure

 In PMA, are formulated as constraints on performances that satisfies a

probability requirement

 The computation of and requires solution to optimization problems.

5
 Reliability Analysis
 The probability of failure corresponding to a failure mode is given as

 It is almost impossible to compute the multi-dimensional integral. However,

approximations to the probability of failure can be obtained using the First Order
Reliability Method (FORM), which computes a Most Probable Point (MPP) of
failure.

(Failed)
(Safe)

(Failed)

(Safe)

Original Space Standard Space

6
 First Order Reliability Method (FORM)

 The MPP is computed by solving the following optimization problem.

 The first order approximation to the probability of failure is

 The optimizer may fail to provide a solution to the equality constrained FORM
problem (singularity).
 Limit state surface is far from the origin in U-space.
 The case never occurs at a given design setting ( the design
has a failure probability equal to zero or one).

 Padmanabhan and Batill [2002] addressed this problem by using a trust region
algorithm for equality constrained problems.
 Gives solution to
 The performance measure approach (PMA) avoids the singularities through an
inverse reliability analysis.

7
 Performance Measure Approach (PMA)

 The reliability constraints are formulated in

terms of the performance values that meets
a given probability requirement

 The following inverse reliability analysis

optimization problem is solved

 PMA formulation is robust compared to RIA.

 In current work, the PMA formulation has

been used.

8
FORM for RIA and PMA

Mean Value
Design Point

MPP

Locus of MPP

9
 Issues in Traditional RBDO Formulation

 It should be noted that irrespective of the Main Optimizer

formulation used to prescribe the reliability
constraints (RIA or PMA), the traditional RBDO
involves a nested optimization process.
 Each iteration of RBDO requires the evaluation Objective Function Reliability
and
of the reliability constraints which themselves Soft Constraints Constraints
require solution to optimization problems.
 Such a formulation is inherently
computationally intensive for problems where
the function evaluations are expensive (e.g.,
multidisciplinary systems). Deterministic
Analysis OR
 Moreover, the formulation becomes impractical
as the number of hard constraints increases
which is often the case in real-life design
Inner
problems. Optimization
Loops
 To reduce the computational cost, a variety of
RBDO techniques have been developed.
Engineering Simulation Model

10
Sequential Method for RBDO

 To alleviate the computational

cost associated with traditional
RBDO, an improved sequential
RBDO technique is developed Deterministic Optimization
in this investigation.

 The optimization and the

reliability assessment are
decoupled from each other.
Inverse Reliability Assessment
 The methodology requires the
gradient of the MPP with
respect to the decision variables
Yes Final
in order to update the MPP Converge
Design
during the optimization phase.
No

Calculate Optimal Sensitivities of MPPs

11
 Computation of optimal sensitivities of the MPP

 It should be noted that during the first iteration, the MPPs are set equal to the
mean values of the random variables. This is equivalent to solving a
deterministic optimization problem.

 In subsequent cycles, the MPP is updated based on reliability assessment at

the previous optimal design setting.

 A linear post optimality analysis is performed to compute the post-optimal

sensitivities of the MPP with respect to the design variables.
 This requires the Hessian of the limit state function at the MPP.
 Hessian update schemes can be employed (SR1, BFGS, etc.)

 If the Hessian is not available or is difficult to obtain, approximations to the limit

state function can be employed to estimate the sensitivities.

 The methodology is tested for analytical problems and simple engineering

design problems.

12
 New Unilevel Method – Deriving KKT conditions

The reliability constraints The Lagrangian for IRA is

in PMA are formulated as

The first order optimality

conditions require that the
The following inverse gradient of the Lagrangian
reliability analysis (IRA) should be zero
problem is solved

The corresponding KKT

conditions are

13
 Unilevel Method for RBDO
 A unilevel formulation for RBDO is
developed.
 The first order KKT conditions of the
inverse reliability analysis
optimization problem are imposed
at the system level directly as
equality constraints.
 Through algebraic manipulation, the
first order KKT conditions for the
inverse reliability optimization
problem can be reduced to

 It should be noted that the

optimization is performed at an
augmented space that consist of the
design variables and the MPPs.

14
Example Problem

CA1

CA2

RBDO
Inputs

15
 10
1000  The figure shows the
9
850 contours of the objective
8 and the constraints at
700 the mean values of the
7
55 0
random variables.
6
4 00  The infeasible region in
d2

5
is shaded.
4 0
25
 This problem has two
3 100
optimal solutions which
2 can be found by
0

choosing different
10

50
1
starting points.
0
-10 -8 -6 -4 -2 0 2 4 6 8 10
d1
Deterministic Optima
Reliable Optima

16
 Comparison of Different RBDO Methods
 Starting from the design [-5,3], an optimal solution of [-3.006,0.049] is obtained.
For this starting point, the number of system analysis required in different RBDO
methods is compared below.

 Starting from the design [5,3], an optimal solution of [2.9277,1.3426] is obtained.

For this starting point, the number of system analysis required in different RBDO
methods is compared below.

 Note that the traditional RBDO formulation that uses the RIA formulation to
prescribe the probabilistic constraint fail to converge. It is also observed that the
unilevel and the sequential RBDO methods developed in this research are
computationally efficient compared to the traditional approaches.

17
Closure

 Traditional reliability based design optimization (RBDO) involves the solution to

a nested optimization problem which is extremely computationally intensive.

 A sequential RBDO (work in progress) and a unilevel RBDO methodology are

developed in this investigation.

 These methodologies avoid the numerical instability associated with the

traditional RIA formulation for RBDO.

 Tests on an analytical problem show that these methodologies are significantly

computationally efficient.

 Current efforts are focused towards testing these approaches for large-scale
multidisciplinary problems.

18
Questions & Discussions

19