Computational Intelligence In Manufacturing Handbook P7

Chang, C. Alec et al " Intelligent Design Retrieving Systems Using Neural Networks"
Computational Intelligence in Manufacturing Handbook
Edited by Jun Wang et al
Boca Raton: CRC Press LLC,2001

©2001 CRC Press LLC

7

Intelligent Design
Retrieving Systems

Using Neural Networks

7.1 Introduction

7.2 Characteristics of Intelligent Design Retrieval

7.3 Structure of an Intelligent System

7.4 Performing Fuzzy Association

7.5 Implementation Example


7.1 Introduction

Design is a process of generating a description of a set of methods that satisfy all requirements. Generally
speaking, a design process model consists of the following four major activities: analysis of a problem,
conceptual design, embodiment design, and detailing design. Among these, the conceptual design stage
is considered a higher level design phase, which requires more creativity, imagination, intuition, and

knowledge than detail design stages. Conceptual design is also the phase where the most important
decisions are made, and where engineering science, practical knowledge, production methods, and
commercial aspects are brought together. During conceptual design, designers must be aware of the
component structures, such as important geometric features and technical attributes that match a par-
ticular set of functions with new design tasks.
Several disciplines, such as variant design, analogical design, and case-based design, have been explored
to computerize the procedure of conceptual design in CAD systems. These techniques follow similar
problem-solving paradigms that support retrieval of an existing design specification for the purpose of
adaptation. In order to identify similar existing designs, the development of an efficient design retrieval
mechanism is of major concern. Design retrieval mechanisms may range from manual search to com-
puterized identification systems based on tailored criteria such as targeted features. Once a similar design
is identified, a number of techniques may be employed to adapt this design based upon current design
goals and constraints. After adapting the retrieved design, a new but similar artifact can be created.

7.1.1 Information Retrieval Systems vs. Design Retrieving Systems

An information retrieval system is a system that is capable of storage, retrieval, and maintenance of
information. Major problems have been found in employing traditional information retrieval methods
for component design retrieval. First, these systems focus on the processing of textual sources. This type
of design information would be hard to describe using traditional textual data.
Another major problem with using traditional information retrieving methods is the use of search
algorithms such as Boolean logic. In a typical Boolean retrieval process, all matched items are returned,

C. Alec Chang

University of Missouri – Columbia

Chieh-Yuan Tsai

Yuan-Ze University

©2001 CRC Press LLC

and all nonmatched documents are rejected. The component design process is an associative activity
through which “designers retrieve previous designs with similar attributes in memory,” not designs with
identical features for a target component.

7.1.2 Group Technology-Based Indexing

Group technology (GT) related systems such as Optiz codes, MICLASS, DCLASS, KK-3, etc., and other
tailored approaches are the most widely used indexing methods for components in industry. While these
methods are suitable as a general search mechanism for an existing component in a database, they suffer
critical drawbacks when they are used as retrieval indexes in the conceptual design task for new components.
Lately, several methods have been developed to fulfill the needs for component design such as indexing
by skeleton, by material, by operation, or by manufacturing process. However, indexing numbers chosen
for these design retrieving systems must be redefined again and again due to fixed GT codes for part
description, and many similar reference designs are still missed. In the context of GT, items to be
associated through similarity are not properly defined.

7.1.3 Other Design Indexing

Several researchers also experiment with image-bitmap-based indexing methods. Back-propagation neu-
ral networks have been used as an associative memory to search corresponding bitmaps for conceptual
designs. Adaptive resonance theory (ART) networks are also explored for the creation of part families in
design tasks (Kumara and Kamarthi, 1992). Other researchers also propose the use of neural networks
with bitmaps for the retrieval of engineering designs (Smith et al., 1997). However, these approaches are
not proper tools for conceptual design tasks because bitmaps are not available without a prototype design,
and a prototype design is the result of a conceptual design. The limitations in hidden line representation
as well as internal features also make them difficult to use in practice.

7.1.4 Feature-Based Modeling

A

part feature

is a parameter set that has specified meanings to manufacturing and design engineers.
Using proper classification schemes, part features can represent form features, tolerance features, assembly
features, functional features, or material features. Comprehensive reviews on feature-based modeling and
feature recognition methods can be found in recent papers (Allada, 1995). There are important works
related to feature mapping processes that transform initial feature models into a product model (Chen,
1989; Case et al., 1994; Lim et al., 1995; Perng and Chang, 1997; Lee and Kim, 1998; and Tseng, 1999).

7.2 Characteristics of Intelligent Design Retrieval

There is no doubt that design is one of the most interesting, complicated, and challenging problem-
solving activities that human beings could ever encounter. Design is a highly knowledge-intensive area.
Most of the practical problems we face in design are either too complex or too ill defined to analyze
using conventional approaches. For the conceptual design stage of industrial components, we urgently
need a higher level ability that maps processes from design requirements and constraints to solution
spaces. Thus, an intelligent design retrieving system should have the characteristics detailed in the
following subsections.

7.2.1 Retrieving “Similar” Designs Instead of Identical Designs

Most designers start the conceptual design process by referring to similar designs that have been developed
in the past. Through the process of association to similar designs, designers selectively retrieve reference
designs, defined as existing designs that have similar geometric features and technological attributes.

©2001 CRC Press LLC

They then modify these referenced designs into a desired design. Designers also get inspiration from the
relevant design information.

7.2.2 Determining the Extent of Reference Corresponding to Similarity
Measures

Design tasks comprise a mixture of complicated synthesis and analysis activities that are not easily
modeled in terms of clear mathematical functions. Defining a clear mathematical formula or algorithm
to automate design processes could be impractical. Thus, methods that retrieve “the” design are not
compatible with conceptual design tasks.
Moreover, features of a conceptual design can be scattered throughout many past designs. Normally
designers would start to observe a few very similar designs, then expand the number of references until
the usefulness of design references diminishes. An intelligent design retrieving system should be able to
facilitate the ability to change the number of references during conceptual design processes.

7.2.3 Relating to Manufacturing Processes

An integrated system for CAD/CAPP/CAM includes modules of object indexing, database structure,
design retrieving, graphic component, design formation, analysis and refinement, generation for process
plan, and finally, process codes to be downloaded. Most computer-aided design (CAD) systems are
concentrated on the integration of advanced geometric modeling tools and methods. These CAD systems
are mainly for detailed design rather than conceptual design. Their linking with the next process planning
stage is still difficult. An intelligent design retrieving system should aim toward a natural linking of the
next process planning and manufacturing stages.

7.2.4 Conducting Retrieval Tasks with a Certain Degree of Incomplete Query
Input

Currently, users are required to specify initial design requirements completely and consistently in the

design process utilization CAD systems. During a conceptual design stage, designers usually do not know
all required features. Thus a design retrieving system that relies on complete query input would not be
practical. It is necessary to provide designers with a computer assistant design system that can operate
like human association tasks, using incomplete queries to come up with creative solutions for the
conceptual design tasks.

7.3 Structure of an Intelligent System

There have been some studies to facilitate design associative memory, such as case-based reasoning,
artificial neural networks, and fuzzy set theory. As early as two decades ago, Minsky at MIT proposed
the use of frame notion to associate knowledge, procedural routines, default contents, and structured
clusters of facts. Researchers have indicated that stories and events can be represented in memory by
their underlying thematic structures and then used for understanding new unfamiliar problems.
CASECAD is a design assistance system based on an integration of case-based reasoning (CBR) and
computer-aided design (CAD) techniques (Maher and Balachandran, 1994). A hybrid intelligent design
retrieval and packaging system is proposed utilizing fuzzy associative memory with back-propagation
neural networks and adaptive resonance theory (Bahrami et al.,



1995). Lin and Chang (1996) combine
fuzzy set theory and back-propagation neural networks to deal with uncertainty in progressive die designs.
Many of these presented methods do not integrate associative memory with manufacturing feature-
based methods. Others still use GT-based features as their indexing methods and suffer the drawbacks
inherited from GT systems. These systems try to use a branching idea to fulfill the need for “similarity”
queries. This approach is not flexible enough to meet the need in conceptual design tasks.

©2001 CRC Press LLC

7.3.1 Associative Memory for Intelligent Design Retrieval

According to these recent experiences, the fuzzy ART neural network can be adopted as a design associative
memory in our intelligent system. This associative memory is constructed by feeding all design cases
from a database into fuzzy ART. After the memory has been built up, the query of a conceptual design
is input for searching similar reference designs in an associative way. By adjusting the similarity parameter
of a fuzzy ART, designers can retrieve reference designs with the desired similarity level. Through the
process of computerized design associated memory, designers can selectively retrieve qualified designs
from an immense number of existing designs.

7.3.2 Design Representation and Indexing

Using a DSG or CSG indexing scheme, a raw material with minimum covered dimension conducts
addition or subtraction Boolean operations with necessary form features from the feature library

ψ

.
Based on either indexing scheme, design case

d

k

can be represented into a vector format in terms of
form features from

ψ

. Accordingly, this indexing procedure can be described as
[

π

F

(

k,

1),…,

π

F

(

k,i

),…,

π

F

(

k,M

)] Equation (7.1)

where

π

(

k

,

i

) [0,1] is a membership measurement associated with the appearance frequency of form
feature

i

ψ

in design case

k

and

M

is the total number of form features.
After following the similar indexing procedure, all design cases in vector formats are stored in a design
database

A

:

A

={

d

1

,…,d

k

,…,d

N

} Equation (7.2)
where

N

is the total number of design cases.
The query construction procedure can be represented as
Equation (7.3)
where

π

F

(

c,i

) [0,1] is a membership measurement defined in Equation 7.1 for conceptual design

c

.

7.3.3 Using a Fuzzy ART Neural Network as Design Associative Memory

Introduced as a theory of human cognitive information processing, fuzzy art incorporates computations
from fuzzy set theory into the adaptive resonance theory (ART) based models (Carpenter et al. 1991;
Venugopal and Narendran, 1992). The ART model is a class of unsupervised as well as adaptive neural
networks. In response to both analog and binary input patterns, fuzzy ART incorporates an important
feature of ART models, such as the pattern matching between bottom-up input and top-down learned
prototype vectors. This matching process leads either to a resonant state that focuses attention and triggers
stable prototype learning or to a self-regulating parallel memory search. This makes the performance of
fuzzy ART superior to other clustering methods, especially when industry-size problems are applied
(Bahrami and Dagli, 1993; Burke and Kamal, 1992).
Mathematically, we can view a feature library as a universe of discourse. Let

R

be a binary fuzzy
relation in

ψ × ψ

if

R

={(

x

,

y

),

π

R

(

x

,

y

)|(

x

,

y

)

ψ × ψ

} Equation (7.4)
where

π

R

(

x,y

)



[0,1] is the membership function for the set

R

.
v
d
k
dd
kk
a
v
≡
∈
∈
qq c ci cM
FFF
a ≡……[()()()]
πππ
,,, ,,, ,1
∈
∈
∈