Systems Realization Laboratory
© 2008, Chris Paredis
Research Overview
Chris Paredis
Systems Realization LaboratoryProduct and Systems Lifecycle Management Center
G.W. Woodruff School of Mechanical EngineeringGeorgia Institute of Technology
www.srl.gatech.edu www.pslm.gatech.edu
Systems Realization Laboratory© 2008, Chris Paredis
Research Scope, Context, and Foundation Model-Based Systems Engineering
1. Reusable Analysis Models Defined in SysML
2. Predictive Trade-off Models
3. Variable Fidelity Analysis
4. Capturing Synthesis Knowledge as Graph Transformations
5. Risk Management in Systems Engineering
Summary
Presentation Overview
2
Systems Realization Laboratory© 2008, Chris Paredis
Research Area• Modeling and Simulation in Design
Application Focus: Systems engineering• Fluid power systems• Mechatronic systems
Research Focus• Decision theory• Modeling and Simulation• Information technology
Research Scope and Context
3
Systems Realization Laboratory© 2008, Chris Paredis
Product Development: A Decision-Based Perspective
4
Concept
Development Design
Production
& Testing
Sales &
Distribution
Maintenance
& Support
Portfolio
Planning
Decisions
Evaluate Alternatives
GenerateAlternatives
Select Alternative
KnowledgeInformation
GenericDecisionProcess
Modeling and Simulation Provides Informationin Support of Decisions
Systems Realization Laboratory© 2008, Chris Paredis
A Generic Design Decision
5
Decision
nA
11O
12O
1kO
21O
22O
2kO
1nO
2nO
nkO
11( )U O
12( )U O
1( )kU O
21( )U O
22( )U O
2( )kU O
1( )nU O
2( )nU O( )nkU O
Select iAwhere
is maximum
[ ]E U
Select iASelect iAwhere
is maximum
[ ]E U
1A
2A
Alternatives Outcomes(Attributes)
Preferences
Systems Realization Laboratory© 2008, Chris Paredis
Uncertainty
Why are Design Decisions Difficult?
Decision
nA
11O
12O
1kO
21O
22O
2kO
1nO
2nO
nkO
11( )U O
12( )U O
1( )kU O
21( )U O
22( )U O
2( )kU O
1( )nU O
2( )nU O( )nkU O
Select iAwhere
is maximum
[ ]E U
Select iASelect iAwhere
is maximum
[ ]E U
1A
2A
Uncertainty
Alternatives Outcomes(Attributes)
Preferences
Limited Resources
Infinite number of
alternatives
Systems Realization Laboratory© 2008, Chris Paredis
Integration of Design Product & Design Process
Goal: maximize overall utility• Must consider process in utility assessment
Value trade-offs between product and process• Decrease expected utility of product to reduce expected cost of
process• "Optimal" solution is guaranteed not to be optimal from a pure
product perspective
Information Economics• Meta-level decision:
How to frame your design problem?
Cost Value Cost Value
7
Systems Realization Laboratory© 2008, Chris Paredis
Systems Approach to Product Development
The systems approach provides a balanced trade-off between product and process objectives
Systematic decomposition• Large, flexible, yet manageable design space
Decoupling• Limit the risk due to unexpected interactions• Enable collaborative engineering
Helps us deal with complexity
8
Systems Realization Laboratory© 2008, Chris Paredis
Model-based Systems Engineering (MBSE)
Effective and Efficient Analysisof Alternatives• Model from different
perspectives• Model at different levels of
abstraction• Model reuse & modularity• Optimization with grid-computing
Effective Generationof Alternatives• Graph transformations for
generating plausible system architectures
• Automated generation of system models
9
Requirements & Objectives
Executable Simulations
System Behavior Models
Design Optimization
System Alternatives
ModelLibraries
MBSE: Model formally all aspects of a systems engineering problem
MBSE: Model formally all aspects of a systems engineering problem
Systems Realization Laboratory© 2008, Chris Paredis10
MBSE: Example Problem
Given:• Primary components• Decision objectives / preferences
Find:• Best system topology• Best component parameters
Given:• Primary components• Decision objectives / preferences
Find:• Best system topology• Best component parameters
Very large search and optimization problem• Many competing
objectives• Many topologies• Many component
types/sizes• Many control
strategiesExcavator
pump_vdisp
cylinder
accum
How to size and connect these?
engine
v_3way
10
Systems Realization Laboratory© 2008, Chris Paredis
Which decisions should be made at each point in the product development process?• How should decisions be ordered in a decision sequence?• How should the decisions be framed? Which decision alternatives
should one consider?
Which information should be used in support of these decisions?• How accurate should the information be?• How do we trade off accuracy versus cost?
Which tools allow us to obtain this information most cost effectively?• Which Information Technologies can be used to reduce the cost of
information?• Is it cost effective to use repositories of reusable models?
Research Questions
11
Systems Realization Laboratory© 2008, Chris Paredis
Research Scope, Context, and Foundation Model-Based Systems Engineering
1. Reusable Analysis Models Defined in SysML
2. Predictive Trade-off Models
3. Variable Fidelity Analysis
4. Capturing Synthesis Knowledge as Graph Transformations
5. Risk Management in Systems Engineering
Summary
Presentation Overview
12
Systems Realization Laboratory
© 2008, Chris Paredis
Formal Representation of Engineering Analysis Models in SysML
Student: Jonathan Jobe
Reusable Analysis Models
14
Multi-Aspect Component Models
• Components are the reusable elements of design
• MAsCoMs:– Group all models related to
single fluid power component– Multiple disciplines and levels
of abstraction– Modular– Formal & unambiguous
• Systems Modeling Language (OMG SysMLTM)
– Formally model systems design information, from requirements through testing
15
How to use MAsCoMs?
Design Concept Schematics
-Hydraulic System
ISO 1219Fluid Power Graphics
Log Splitter Design Example
16
Log Splitter—Model Composition
17
Automatic Translationfrom SysML to Modelica
Formal Graph Transformations Modelica
SysML
Systems Realization Laboratory© 2008, Chris Paredis
Research Scope, Context, and Foundation Model-Based Systems Engineering
1. Reusable Analysis Models Defined in SysML
2. Predictive Trade-off Models
3. Variable Fidelity Analysis
4. Capturing Synthesis Knowledge as Graph Transformations
5. Risk Management in Systems Engineering
Summary
Presentation Overview
18
Systems Realization Laboratory
© 2008, Chris Paredis
Reusable Models for Predicting the Performance of Design Concepts
Student: Rich Malak
Predictive Trade-off Models
Systems Realization Laboratory© 2008, Chris Paredis
Decision Chains: From Concept to Detail
DesignDecision
Alternative 1
Alternative 2
Alternative 3
Alternative n
...
One-Shot Decision A Chain of Decisions
Decision3.1
Decision1
Decision3.2
Alternative 1
Alternative 2
Alternative 3
Alternative 4
Decision3.3
Alternative…
Alternative…
Decision3.4
Alternative n-1
Alternative n
Decision2.1
Decision2.2
Concept 2.1
Concept 2.2
Concept 3.1
Concept 3.2
Concept 3.3
Concept 3.4
20
Systems Realization Laboratory© 2008, Chris Paredis
Decision Chains: From Concept to Detail
DesignDecision
Alternative 1
Alternative 2
Alternative 3
Alternative n
...
One-Shot Decision A Chain of Decisions
Decision3.1
Decision1
Decision3.2
Alternative 1
Alternative 2
Alternative 3
Alternative 4
Decision3.3
Alternative…
Alternative…
Decision3.4
Alternative n-1
Alternative n
Decision2.1
Decision2.2
Concept 2.1
Concept 2.2
Concept 3.1
Concept 3.2
Concept 3.3
Concept 3.4
Decision2.1
Alternative 1
Alternative 2
Alternative 3
Alternative 4
Each Concept = Set of Design Alternatives
21
Systems Realization Laboratory© 2008, Chris Paredis
Modeling Set-Based Design Concepts
What is the appropriate gear ratiofor the differential?
…it depends• On the other drivetrain
components• On manufacturing and
cost considerations• On the details of
differential itself
How do we best model a subsystem conceptto support decision making at the systems level?
22
Systems Realization Laboratory© 2008, Chris Paredis
Functional and Preference Attributes
What is important about the differential?• Gear ratio• Maximum torque• Maximum speed• Size, mass• Cost
What is NOT important about the differential?• Working principle• Detailed dimensions• Manufacturing processes
Functional Attributes
Preference Attributes
Only important to the extent that theyinfluence the preference attributes
Model Preference Attributes as Function of Key Functional Attributes
23
Systems Realization Laboratory© 2008, Chris Paredis
Predictive Tradeoff Model: Hydraulic Cylinder Key functional attributes
• Bore diameter
• Stroke length
Predicted preference attributes• Mass
• Cost
stroke
mass
borecost
stroke
mass
borecost utility
stroke
bore pumpdisp
Raw Cylinder Data Predictive Cylinder Model System Tradeoffs
stroke
mass
borecost
stroke
mass
borecost
stroke
mass
borecost utility
stroke
bore pumpdisp
Raw Cylinder Data Predictive Cylinder Model System Tradeoffs
Systems Realization Laboratory© 2008, Chris Paredis
Taking Designer Preferences into Account
Not all feasible designs are preferred: Pareto optimality
Non-dominated set (efficient frontier). Every point is optimal under some tradeoff
0.95 0.96 0.97 0.98 0.99 1
200
250
300
350
400
Reliability
Cos
t ($)
Valid Designs: Planetary Gear TrainFeasible Designs: Planetary Gear TrainFeasible, but dominated. No engineer will choose these
Care only about non-dominated set
25
Systems Realization Laboratory© 2008, Chris Paredis
Parameterized Efficient Sets
2.61222.9375
3.25003.5789
3.8750
0.95
0.96
0.97
0.98
0.99
1
180
200
220
240
260
280
300
320
340
360
Torque Ratio
Reliability
Co
st (
$)
No problem-independent preference for gear ratio.Solution: find efficient set as function of gear ratio.
Gear Ratio
26
Systems Realization Laboratory© 2008, Chris Paredis
Combining Multiple Concepts with the same Functionality
Blue reverted gear train
dominates at gear ratios < 5
Redplanetary gear train
dominates at gear ratios > 5
Best concept can be identified after consideringsystems-level tradeoff for gear ratio and other attributes.
33.5
44.5
55.5
6
0.940.95
0.960.97
0.980.99
1
150
200
250
300
350
400
450
500
Torque RatioReliability
Co
st (
$)
Gear Ratio
33.5
44.5
55.5
6
0.940.95
0.960.97
0.980.99
1
150
200
250
300
350
400
450
500
Torque RatioReliability
Co
st (
$)
Gear Ratio
27
Systems Realization Laboratory© 2008, Chris Paredis
Another Example: Small 4-stroke Engines
Some engines are dominated• E.g. same torque, mass,
power, but lower price
Not all combinationsof the functional attributes are feasible• Clustering algorithms
are used to define the feasible set(Support Vector Machines)
3D Projection of Small Engine Data (6D)(Clustering and Domination)
non-dominatedtradeoffs
max torque
dominatedtradeoffs
Cluster ID
mass
max power
3D Projection of Small Engine Data (6D)(Clustering and Domination)
non-dominatedtradeoffs
max torque
dominatedtradeoffs
Cluster ID
mass
max power
28
Project 2E Webcast - 23 July 2008 29
Predictive Tradeoff Modeling:Making System-Level Decisions
System Model
Utility Model
Other Components
…
Pump Attributes:• max displacement• max speed• efficiency• mass• price
Cylinder Attributes:• bore diam.• stroke length• mass• price
Engine Attributes:• max power• max speed• max torque• efficiency• mass• price
Typical Iterative Loop(optimization / search)
Often more than technical performance
• E.g., price/cost, mass, “-ilities” (reliability, usability,…)
Attribute values for a given component are not independent How to formalize the dependencies efficiently and effectively?
Attribute values for a given component are not independent How to formalize the dependencies efficiently and effectively?
Systems Realization Laboratory© 2008, Chris Paredis
Making System-Level Decisions with Predictive Models
stroke
mass
borecost
stroke
mass
borecost utility
stroke
bore pumpdisp
Raw Cylinder Data Predictive Cylinder Model System Tradeoffs
stroke
mass
borecost
stroke
mass
borecost
stroke
mass
borecost utility
stroke
bore pumpdisp
Raw Cylinder Data Predictive Cylinder Model System Tradeoffs
Engine & Pump
Hydraulic Cylinder & Ram
Directional Control Valve
Log Loading & Splitting Area
Engine & Pump
Hydraulic Cylinder & Ram
Directional Control Valve
Log Loading & Splitting Area
stroke
mass
borecost
stroke
mass
borecost utility
stroke
bore pumpdisp
Raw Cylinder Data Predictive Cylinder Model System Tradeoffs
stroke
mass
borecost
stroke
mass
borecost
stroke
mass
borecost utility
stroke
bore pumpdisp
Raw Cylinder Data Predictive Cylinder Model System Tradeoffs
Engine
Load
Engine
Load
Systems Realization Laboratory
© 2008, Chris Paredis
Adaptive Kriging Models
Students: Alek Kerzhner, Roxanne Moore
Efficient Design Optimization under Uncertainty
3232
Motivating Study:Backhoe Dig-Cycle
• Optimization under uncertainty
• LatinHyperCube sampler used to predict expected value
• Kriging model used in conjunction with sampler to generate response surface to reduce computational cost
optimizer
Latin Hypercube + Kriging response surface
Objectives:• Maximize Efficiency• Minimize Cost
Design variables:• bore diameters• pump max disp
Adaptive Kriging Models
• Why Surrogate Models?– Without Kriging: 50 optimizer iterations*
7 evaluations per iteration*1000 LHS samples = 350,000 runs
– With Kriging: 2000 seeding runs + 100 extra runs = 2,100 runs!
• Adaptive Kriging model– Kriging: computational complexity is O(n4)– Adaptive approach: keep n small
• Many localized (small) response surfaces• Automatically requests additional simulations • Automatically regenerates response surface
3333
Systems Realization Laboratory© 2008, Chris Paredis
The Information Economic Challenge
Level of Exploration / Optimization
Levelof
FidelityLevel of Effort
Required
Systems Realization Laboratory© 2008, Chris Paredis
The Information Economic Challenge
Level of Exploration / Optimization
Levelof
FidelityLevel of Effort
Required
Systems Realization Laboratory© 2008, Chris Paredis
Variable Fidelity Modeling
Low fidelity for broad explorationidentify promising regions of design space
High fidelity for detailed optimization
Feasible solutions
x1
x2
A
B
C
x1
Obj
Systems Realization Laboratory© 2008, Chris Paredis
Research Scope, Context, and Foundation Model-Based Systems Engineering
1. Reusable Analysis Models Defined in SysML
2. Predictive Trade-off Models
3. Variable Fidelity Analysis
4. Capturing Synthesis Knowledge as Graph Transformations
5. Risk Management in Systems Engineering
Summary
Presentation Overview
37
Systems Realization Laboratory
© 2008, Chris Paredis
Formal Representation of Domain-Specific Knowledge
Student: Alek Kerzhner
Capturing Synthesis Knowledge as Graph Transformations
Systems Realization Laboratory© 2008, Chris Paredis
Most Knowledge Can Be Representedas Graphs or Graph Transformations
Requirements & Objectives
SysML
Executable Simulations
Dymola
System Behavior Models SysML
Topology Generation using Graph Transf
Model Composition using Graph Transf
Model Translation using Graph Transf
Design Optimization ModelCenter
System Alternatives
MAsCoMs SysML
Simulation Configuration using Graph Transf
Hydraulic_Subsystem Schematic[Block] ibd [ ]
actuator : Double-ActingCylinder
a : FlowPort
b : FlowPorthousing : FlowPort
rod : FlowPort
valve : 4port3wayServoValve
cylA : FlowPort
cylB : FlowPort
portP : FlowPort
portT : FlowPort
pump : FDpump
discharge : FlowPort
suction : FlowPort
housing : FlowPort
inputShaft : FlowPort
tank-to-pump : Line
a : FlowPort
b : FlowPort
pump-to-valve : Line
a : FlowPort
b : FlowPort
valve-to-cylP1 : Line
a : FlowPort
b : FlowPort
valve-to-cylP2 : Line
a : FlowPort
b : FlowPortfilter : Filter
in : FlowPort
out : FlowPort
valve-to-filter : Line
a : FlowPort
b : FlowPort
filter-to-tank : Line
a : FlowPort
b : FlowPort
tank : Tank
return : FlowPort
sump : FlowPort
hydraulics
world
x
y
Dig Cycle
Arm
Boom
Sw ing
BucketTraj
systemalternative
behaviormodel
simulationconfiguration
Systems Realization Laboratory© 2008, Chris Paredis
MOFLON: Meta-Modeling & Graph Xform Tool
Define graph transformationsin Storyboards:activity diagrams + graphs
Easily integrated with SysMLtools(through standardized JMI interface)
Capture domain-specific knowledge
40
Systems Realization Laboratory© 2008, Chris Paredis
Generative Grammar for Design Synthesis
Graph Transformation rules to generate systems
Generate random system alternatives by applying rules in randomized order
Improve system alternatives through evolutionary search algorithms
41
Systems Realization Laboratory© 2008, Chris Paredis
Composition of Multi-Aspect Models
System ModelsMACM
Repository
ArchitectureOptimization
Perspective A
Perspective CPerspective B
AutomatedComposition
42
Systems Realization Laboratory© 2008, Chris Paredis
Research Scope, Context, and Foundation Model-Based Systems Engineering
1. Reusable Analysis Models Defined in SysML
2. Predictive Trade-off Models
3. Variable Fidelity Analysis
4. Capturing Synthesis Knowledge as Graph Transformations
5. Risk Management in Systems Engineering
Summary
Presentation Overview
43
Systems Realization Laboratory
© 2008, Chris Paredis
Trading off Product Quality and Risk versus Design Process Costs
Student: Stephanie Thompson
Risk Management in Systems Engineering
Systems Realization Laboratory© 2008, Chris Paredis
Maximize expected utility! Of what?• Product (artifact) utility• Product + Design Process utility
Instead of “What configuration of the product gives us the best product
utility?” we should ask
“What can we do now to give us the best balance of product and process utility later?”
Frame of Reference
45
Systems Realization Laboratory© 2008, Chris Paredis
When is it valuable to perform additional analyses to reduce risk?
When is it valuable to perform detailed uncertainty analysis? In which order should the analyses be performed? How many design concepts should one carry forward in
parallel? …
Approach: • Investigate simplified design scenario for which a theoretically
optimal design process can be determined• Generalize these theoretical insights toward practical guidelines
Questions To Answer
46
Systems Realization Laboratory© 2008, Chris Paredis
Simple Illustration Problem – Problem Definition
Two Concepts• Designer must choose concept
A or B• The utilities of concepts A and
B are uncertain
Two Analyses• Low quality (LQ) and high
quality (HQ) analyses to refine the utility estimates
• LQ is cheaper, but the HQ is more precise.
• Cost of analyses reduces expected utility.
47
2~ ,iuu N
Lu u
Hu u
2~ 0,N
2~ 0,N
Systems Realization Laboratory© 2008, Chris Paredis
Normal distributions:• The utilities of A and B, and uncertainties of analyses are all
distributed normally.• All derived conditional distributions are also Normal
Limited scope of decision tree• The low quality analyses have already been performed for both
concepts, reducing the size of the decision tree. • In further work, the decision tree will include branches to perform the
low quality analyses.
Simplifying Assumptions
48
Systems Realization Laboratory© 2008, Chris Paredis
Decision Tree of Design Process
49
Select A
Select B
HQ A
HQ B HQ A
HQ B
Select A
Select A
Select B
Select B
Select B
Select B
Select A
Select A
ALp u
BLp u
|A AL
A
u up u
|B BL
B
u up u
|A AH L
AHu u
p u
|B BH L
BHu u
p u
|B BH L
BHu u
p u
|A AH L
AHu u
p u
|A AL
A
u up u
|B BL
B
u up u
| ,A A AH L
A
u u up u
| ,B B BH L
B
u u up u
| ,A A AH L
A
u u up u
| ,B B BH L
B
u u up u
uA
uB
uA-cHQA
uB-cHQA
uA-cHQB
uB-cHQB
uA-cHQA-cHQB
uA-cHQA-cHQB
uB-cHQA-cHQB
uB-cHQA-cHQB
| ,A A AH L
A
u u up u
| ,B B BH L
B
u u up u
Systems Realization Laboratory© 2008, Chris Paredis
Apply Bayes’ Theoremto derive conditional probabilities
All distributions end up being Normal:
Probability Distributions for Chance Events
50
|
,x Y y
Y
p x yp x
p y
2 2 2 2
| 2 2 2 2~ ,
L
u L uu u
u u
up u N
2 2 2 2 2 2 2 2
| 2 2 2 2~ ,
H L
u L u uu u H
u u
up u N
2 2 2 2
| 2 2 2 2~ ,
H
u H uu u
u u
up u N
2 2 2 2 2 2 2 2
| 2 2 2 2~ ,
L H
u H u u du u L
u u
up u N
2 2 2 2 2 2 2 2 2
| , 2 2 2 2 2 2 2 2 2 2 2 2~ ,
L H
u L u H uu u u
u u u u
u up u N
Systems Realization Laboratory© 2008, Chris Paredis
Decision Map
51
Nominal Parameters:
μA = μB = 100
σ2A = σ2
B = 502
HQ costs = 10
σ2L = 452
σ2H = 52
Systems Realization Laboratory© 2008, Chris Paredis
Research Scope, Context, and Foundation Model-Based Systems Engineering
1. Reusable Analysis Models Defined in SysML
2. Predictive Trade-off Models
3. Variable Fidelity Analysis
4. Capturing Synthesis Knowledge as Graph Transformations
5. Risk Management in Systems Engineering
Summary
Presentation Overview
52
Systems Realization Laboratory© 2008, Chris Paredis
Summary
53
Overall Research ThemeHow can one design systems better at a lower cost?
Guiding PrincipleMaximize utility of both product and process
Increase the value – Decrease the cost
Solution PrinciplesUtility theory
Modularity in components and modelsFormal knowledge capture and reuseGraph manipulation & transformation