knowledge extraction from aerodynamic design data and its application to 3d turbine blade geometries...

13
Knowledge Extraction from Aerodynamic Design Data and its Application to 3D Turbine Blade Geometries Lars Graening [email protected] http://www.honda-ri.de/ 03 July 2009 International Workshop on Machine Learning for Aerospace Marseille, France

Upload: griselda-miller

Post on 28-Dec-2015

229 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Knowledge Extraction from Aerodynamic Design Data and its Application to 3D Turbine Blade Geometries Lars Graening lars.graening@honda-ri.de

Knowledge Extraction from Aerodynamic Design Data and its Application to 3D Turbine

Blade Geometries

Lars Graening

[email protected]://www.honda-ri.de/

03 July 2009

International Workshop onMachine Learning for Aerospace

Marseille, France

Page 2: Knowledge Extraction from Aerodynamic Design Data and its Application to 3D Turbine Blade Geometries Lars Graening lars.graening@honda-ri.de

Outline

Validation of the extracted knowledge

Universal representation of design modifications

Knowledge extraction from unstructured surface meshes

Page 3: Knowledge Extraction from Aerodynamic Design Data and its Application to 3D Turbine Blade Geometries Lars Graening lars.graening@honda-ri.de

3D Turbine Blade

Knowledge extraction from a 3D turbine blade design data set

Ultra-low aspect ratio transonic turbine stator blade of a small Honda turbofan engine

Honda Business Jet

Part of the stator section 3D Blade Design

Pressure side

Suction side

Leading edge

Trailing edge

Page 4: Knowledge Extraction from Aerodynamic Design Data and its Application to 3D Turbine Blade Geometries Lars Graening lars.graening@honda-ri.de

Knowledge Extraction

DesignDatabase

KnowledgeExtraction

Design data

(Design,Performance)

Knowledge

(e.g. Design Rules)

Pre-Processing

Optimizer

Design Generation& Modification

Evaluation

How can we extract knowledge from the design data resulting from various optimization runs where different shape representations

are used?

• CAD software

• Rapid Prototyping

• Splines,FFD / DMFFD

• …

• Aerod. Engineer

• Computer:

• Evo. Opt.

• Response Surface

• …

• CFD simulation

• Wind tunnel

• Real physicalenvironment

• …

Usually huge amount of design data is generated during the design and optimization process.

Page 5: Knowledge Extraction from Aerodynamic Design Data and its Application to 3D Turbine Blade Geometries Lars Graening lars.graening@honda-ri.de

Universal Design Representation & Pre-Processing

ReferenceSurface Mesh

ModifiedSurface Mesh

Vertex AVertex A’

Displacement

Displacement:

,,, ri

ri

mj

mj

ri nvvvv

rmmr ff ,• Measure of the local deformation along the

normal vector of a reference design• Using performance difference instead of

performance values• An identification of corresponding vertices

is needed

The calculation of the displacement leads to a reduction in the number of parameters under consideration

Universal representation:

• Use unstructured surface mesh as a universal geometric representation

• Vertices sample the design surface, triangles define the neighborhood, normal vectors define local curvature

Page 6: Knowledge Extraction from Aerodynamic Design Data and its Application to 3D Turbine Blade Geometries Lars Graening lars.graening@honda-ri.de

Knowledge Extraction from Aerodynamic Design Data

Where are unchanged and mostfrequently changed design regions?

What are the mean differences ofmultiple designs relative to a base design?

Which design regions are sensitiveto performance changes?

How similar are two designs?

N

r

N

rm

mrjiNNi

1 1

2,,1

2

ri

rr

iN

R

N

rmm

rmrri

mrjir

i

1,1

,,,

N

rmm

mrji

riN ,1

,,1

1

log

Page 7: Knowledge Extraction from Aerodynamic Design Data and its Application to 3D Turbine Blade Geometries Lars Graening lars.graening@honda-ri.de

Knowledge Extraction from Aerodynamic Design Data

• A reduction of the parameters is needed• Assuming neighboring vertices with

similar sensitivity belong to the same design region

• Clustering vertices to sensitive design regions

• Cluster centers are used to analyze interrelations

How are distant design areasinterrelated?

Blade geometries

De

sig

n R

ule

s

Page 8: Knowledge Extraction from Aerodynamic Design Data and its Application to 3D Turbine Blade Geometries Lars Graening lars.graening@honda-ri.de

Design Rules

A: A reduction of the blade thickness is expected to increase the performance0000 ,

7,

10,

8 THENANDANDIF ArCC

ArCC

ArCC

B: Surprisingly, reducing the thickness at the suction side but increasing the thickness at the pressure side is expected to decrease the performance

000 ,7

,8 THENANDIF Ar

CCAr

CC

Pressure Side:

Suction Side:

Investigate the consequences of an interrelated displacement of distant vertices to the performance

How reliability is the extracted information?

Page 9: Knowledge Extraction from Aerodynamic Design Data and its Application to 3D Turbine Blade Geometries Lars Graening lars.graening@honda-ri.de

Knowledge Validation

Algorithm for Knowledge Validation:(using Direct Manipulation of Free Form Deformation)

Object Point

Control Point

Deform the design usingDirect Manipulation of Free-Form Deformation

Generate CFD mesh for thereference blade

Apply deformations to theCFD mesh

Simulate the flowusing CFD (HSTAR3D)

Dis

pla

cem

en

tP

erf

orm

an

ce

Kn

ow

led

ge

(e.

g. d

es

ign

ru

le)

Menzel, S., Olhofer, M., Sendhoff B.: Direct Manipulation of Free Form Deformation in Evolutionary Design Optimization, PPSN 2006

Page 10: Knowledge Extraction from Aerodynamic Design Data and its Application to 3D Turbine Blade Geometries Lars Graening lars.graening@honda-ri.de

Knowledge Validation

Reference blade

Surface Deformation

CFD Grid DeformationResulting Blade Performance after CFD(Pressure Loss)

worse

better

CC7

Sensitivity

The displacement of vertex CC7 is positive

correlated with the aerodynamic pressure loss of the turbine blade!1. The hypothesis of the direct relation between

performance and displacement of CC7 is true

2. Using DMFFD for validation and utilization works out well

Page 11: Knowledge Extraction from Aerodynamic Design Data and its Application to 3D Turbine Blade Geometries Lars Graening lars.graening@honda-ri.de

Knowledge Validation

A:

B:

738.9

499.11

Decrease in pressure loss of about -10.50 %

Increase in pressure lossof about 5.86 %

0000 ,7

,10

,8 THENANDANDIF Ar

CCAr

CCAr

CC

000 ,7

,8 THENANDIF Ar

CCAr

CC

The performed experiments confirm the expected outcome of the extracted hypothesis

Page 12: Knowledge Extraction from Aerodynamic Design Data and its Application to 3D Turbine Blade Geometries Lars Graening lars.graening@honda-ri.de

Summary & Outlook

• framework for extracting knowledge from aerodynamic design data

• technique based on DMFFD for validating the extracted knowledge

• validation experiments provide evidence for the reliability of the knowledge extraction techniques

• outlook measurements are needed that allow to extract rules which are potentially interesting for the aero engineer

Page 13: Knowledge Extraction from Aerodynamic Design Data and its Application to 3D Turbine Blade Geometries Lars Graening lars.graening@honda-ri.de

Thank You!