1 of 27

ADJECTIVE-BASED, CUSTOMER-ORIENTED SMART DESIGN AND APPLICATIONS IN AUTOMOTIVE AND SHIP BUILDING INDUSTRIES

1. Istanbul Technical University, Turkey

2. The University of Tokyo

2 of 27

2

Published in 2017

3 of 27

OBJECTIVE

An adjective based design concept for yacht hulls and automobiles.
Relationship between adjectives and geometric parameters.
Sampling technique for automatic search and generation of design variations for a CAD model.

3

Design based on adjectives

Sampled designs in design space

4 of 27

ADJECTIVE BASED DESIGN

Use adjectives to represent design

Customer oriented design
Easy communication between designers and customers

CAD based design

Use design parameters

4

design parameters

adjectives

Aggressive

Compact

Modern

Charismatic

Strong

5 of 27

ADJECTIVES

5

Strong

Charismatic

Aesthetic

Speedy

Modern

Compact

Comfortable

Cute

Aggressive

6 of 27

TSENG ET AL.’S WORK

Method to capture designers’ preference(s) for stylistic attributes.
Utilized vehicle design represented by line drawing silhouettes.
Neural network were used in conjunction with genetic algorithms.

6

Most Sportive

Least Sportive

Most Beautiful

Least Beautiful

7 of 27

TSENG ET AL.’S WORK

7

Parametric Design

Randomly generate Initial designs

Survey with 18 participants

Train four neural networks per participant

Adjectives

Sportive

Rugged

Beautiful

Fuel Efficient

Resulting neural networks inverted using a genetic algorithms to generate new designs

8 of 27

YACHT HULL DESIGN

10 Adjectives for hull expression
34 Design parameters
Survey set is applied

8

Hull adjectives

Speedy

Strong

Comfortable

Aesthetic

Usual

Aggressive

Compact

Cute

Charismatic

Modern

9 of 27

DESIGN FRAMEWORK

9

Entrance Section

Middle-body Section

Run Section

FP

TG: Top Guide

BG: Bottom Guide

FG: Feature Guide

SP: Station Profile

FP: Forward Profile

FP

10 of 27

DESIGN PARAMETERS

10

11 of 27

SURVEY-1 ~ HULL ADJECTIVE LEARNING

97 participants
Collected adjectives were filtered
Most used adjectives are selected

11

Strong
Speedy
Aesthetic
Usual (common)
Compact
Comfortable
Aggressive
Modern
Charismatic
Cute

In this step, hull adjectives expressing yacht hull shapes are explored.

Three main criteria are considered to determine the adjectives:

The adjectives should be frequently used by people.
They should be less relativistic and less related to the predefined geometric parameters.
The adjectives all must have different meaning from each other to prevent confusion for subjects.

By means of Survey 1, a variety of adjectives were collected from people. Fig. 6 (a) shows a question from Survey 1. Survey participants are asked to express 3D yacht hull models having a variety of

geometries that are collected from the net, by an adjective. This survey was applied to 97 participants with ten different hull models. Most of the adjectives were learned via Survey 1, some of which were eliminated later. We selected some of the adjectives to use in the next surveys based on the second above-mentioned criterion. For example, the adjectives bad, ugly, good are too much relativistic and the adjectives wide, long, thin are directly related to the geometric parameters. Additionally, some adjectives were grouped due to their close meanings. For instance, the adjectives strong, enduring, durable and solid have close meaning, and therefore they were grouped under the strong adjective. Fig. 6 (b) shows adjectives with the usage frequency, which have been determined after the filtration process. 10 most preferred adjectives were used in a study of Survey 3.

Finally, ten hull adjectives were selected that were investigated in the further steps, which are listed as follows: strong, speedy, comfortable, aesthetic, usual (common), aggressive, compact, modern, charismatic, cute.

12 of 27

SURVEY-2 PARAMETER ELIMINATION

75 participants
Base models (large, medium, small) are sampled
9 parameters were eliminated.

12

Some of the geometric parameters which do not have any impact on predetermined 15 hull adjectives are detected and eliminated in this step.

We divided the parameters into two groups as Basic parameters and Sub parameters (see Fig. 7 (a)). It is believed that some of the Sub parameters may not have any

impact on the predefined hull adjectives, and therefore they should be eliminated and not be used in the further steps to reduce the number of

geometric parameters. Such parameter elimination can facilitate the exploration process of finding the relationship between hull adjectives and geometric parameters.

First, a cubical design space (see Fig. 7 (b)) is generated whose parametric dimensions are selected as length (L), beam (B) and depth (D)

of the hull, since they are the most important Basic parameters. Three parametric levels are obtained for each parametric dimension, which

stand for small, medium and large values for L, B and D parametric dimensions.

Only three of these models are utilized, as it is impractical to use all these 27 models. These models reside on one diagonal

of the cubical design space, which can be regarded as small, medium and large size of hull models. We call these models Base models (see

Fig. 7 (b) models enclosed by blue rectangle). Parametric values of 16 Sub parameters are changed one by one on the Base models. Three

parametric levels are taken for each Sub parameter where the parameter values are small, medium and large.

We collect data from 75 participants in two sets with one-week interval in order to make participants focus well on the

surveys. Survey data is analyzed for each Base model separately as seen in Fig. 8 (b) and percentages for the Changed, Unchanged and Neutral selections

are calculated. If percentages higher than 80% is obtained for the Unchanged option for all three Base models, the Sub parameter is eliminated.

Otherwise, it is not eliminated.

13 of 27

SURVEY-3~DATA SET PREPARATION

L54 orthogonal array to sample designs
10 adjectives are matched with 54 hull using box-plot

13

Twenty-five geometric parameters left after eliminating some of the geometric parameters. When a naive approach with three parametric levels for each

geometric parameter is used, 847,288,609,443 (325) possible yacht hull design variations can be obtained, which is impractical to be used in Survey 3 for participant scoring.

SAMPLING: Therefore, the Taguchi experimental method is utilized to obtain sampled yacht hull models in the design space. The Taguchi experimental method suggests using orthogonal arrays

(Bush, 1952) which can provide uniform sampling from the design space by taking hull samples having more geometric variations as much as possible with a less number of hulls. L54 orthogonal array is chosen for 25 geometric parameters with three parametric levels (1, 2, 3) for each parameter.

SURVEY: “1” stands for the hull that has the adjective characteristics (“well” and “very well”) and “0” is for the hull

without the adjective property (“poorly” and “very poorly”). Likert-type scale can be thought as a pair of scales, which has a balance point

(Neither “3”) and two sides (poorly (1–2) and well (4–5)). The side, having more weight wins. In other words, if the box is placed above the

score “3”, it represents %75 of the data (sum of the box and the upper whisker) and belongs to the score “well”. Therefore, it is assigned to the

class “1”. If it is below the score “3”, this model is poorly described by this adjective so as to be assigned to the class “0”.

14 of 27

LEARNING RELATIONS-1

Input: Standardized parameter dimension values
Output: adjective classes (1,0)
54 observations
GMDH-type neural network to find relations

14

15 of 27

LEARNING RELATIONS-2

Mathematical models were obtained for each adjective

15

16 of 27

RESULTS OF ATTRIBUTE BASED DESIGN

16

17 of 27

DESIGN SAMPLING

A design sampling technique called Sampling-TLBO (S-TLBO) is developed.
S-TLBO - based on teaching learning based optimization (TLBO).

17

S-TLBO

Design Parameters

Design Constraints

Parametric Bounds

Number of Designs (N)

18 of 27

TEACHING LEARNING BASED OPTIMIZATION (TLBO)

Inspired from typical class room teaching and learning process
This algorithm simulates two fundamental modes of learning.
Through the teacher (Teacher Phase)
Interacting with other learners (Learner Phase)

18

Group of Students

Different Subjects

Scores

Teacher

Population

Different Design Variables

Fitness Value of the Problem

Best Solution

Analogies

19 of 27

LEARNING PHASES OF TLBO

19

Teacher Phase

During this phase teacher gives knowledge to students.
Students modify themselves using following equation:

Learner Phase

Each learner learn form other learners by interacting.
Knowledge is transferred if interacted learner is better then interactor learner.

20 of 27

S-TLBO

20

Properties of S-TLBO

21 of 27

S-TLBO

21

Space-filling Designs

Uniformly distributed designs over the design space.
Space-filling is achieved using the Audze and Eglais‘ (2011) technique.

With Space-filling

Without Space-filling

Non-collapsingness

Designs should not share the same design parameter values.
This helps to maintain diversity in the selected designs.

22 of 27

S-TLBO

22

Large portion of designs is at boundaries

Space-filling designs

Non-collapsing

Uncovered regions

Combination of both

✔

🗶

Sampling in Constrained Spaces

Static constraint handling mechanism is utilized.
Design are panelized if they violate any constraint.

23 of 27

RESULTS OF S-TLBO

23

To validate the performance of S-TLBO, two different CAD models are utilized: a yacht hull and a wine glass (without base).

Yacht Hull

CAD models are parameterized
Range of each parameter is defined

24 of 27

RESULTS OF S-TLBO

24

25 of 27

CONCLUSIONS

Adjectives for yacht models were explored.
Novel design framework with geometric parameters was proposed.
Three different survey types were introduced to capture human aesthetic shape understandings.
Mathematical models were obtained, which represent hulls expressed by adjectives.
A new sampling approach S-TLBO is proposed based on teaching-learning-based optimization for CAD model generation.
To obtain distinct designs S-TLBO favors designs with space-filling and non-collapsing properties.

25

26 of 27

FUTURE WORKS

The number of adjectives can be increased.
An extra survey will be performed to check the reliability of this study’s results.
S-TLBO can be implemented to sample designs based on the performance criterion.
S-TLBO will be utilized for generative design formation of industrial products.

26

27 of 27

ACKNOWLEDGMENT

This work was supported by The Scientific and Technological Research Council of Turkey (Project Number: 214M333)

27