Upload
shi-jie-gary-chen
View
213
Download
1
Embed Size (px)
Citation preview
Knowledge-based support for simulation analysis ofmanufacturing cells
Shi-Jie (Gary) Chena, Li-Chieh Chenb, Li Linc,*
aDepartment of Industrial Engineering, National Taipei University of Technology, Taipei, TaiwanbDepartment of Industrial Design, Tatung University, Taipei, Taiwan
cDepartment of Industrial Engineering, State University of New York at Buffalo, Buffalo, NY 14260, USA
Received 28 September 1998; received in revised form 29 January 2000; accepted 11 August 2000
Abstract
Simulation is a widely used approach for assisting design and improvement of manufacturing systems. It is a complex
activity and needs a great deal of human expertise. Since the knowledge of analyzing simulation output for decision making is
not inherently captured in the simulation modeling methodology, a framework that integrates simulation and knowledge-based
decision analysis is needed. In this paper, we develop a knowledge-based system that cooperates with simulation for
improving the performance of manufacturing cells. Using Axiomatic Design as a guideline, a hierarchical knowledge base
structure that corresponds to the decision process is built. Our proposed knowledge-based system consists of a set of facts and
three levels of rules in a hierarchy that is consistent with the manufacturing cell system con®guration. The system
demonstrates the effectiveness of utilizing Axiomatic Design concept when developing a knowledge-based system. The
results of an industrial study show that our method contributes to improving the performance of manufacturing cells.
# 2001 Elsevier Science B.V. All rights reserved.
Keywords: Axiomatic Design; Simulation; Knowledge-based system; Manufacturing cells
1. Introduction
A manufacturing cell is a cluster of machines or
processes in close proximity and dedicated to the
manufacturing of certain identi®ed part families that
share similar manufacturing requirements. To
improve design and performance of manufacturing
cells, simulation has become an effective method for
its versatility in modeling complex and dynamic
operations. Nevertheless, improving the performance
of a manufacturing cell is still a complex activity that
not only is time consuming but also demands much
human expertise in its decision making. In addition,
the skills required to conduct simulation studies cor-
rectly and accurately are not widespread [13]. By
using knowledge-based system techniques, these
skills and knowledge for the simulation analysis pro-
cess can be captured in a computer. This calls for the
need of a framework that integrates simulation and
knowledge-based decision analysis. According to the
simulation outcome, the knowledge-based system will
assist the decision process for the improvement of the
manufacturing cell performance. However, since
human experts typically do not express their knowl-
edge in a well-structured manner during system devel-
opment, knowledge-based systems often suffer from
Computers in Industry 44 (2001) 33±49
* Corresponding author. Tel.: �1-716-645-2357/ext. 2119;
fax: �1-716-645-3302.
E-mail address: [email protected] (L. Lin).
0166-3615/01/$ ± see front matter # 2001 Elsevier Science B.V. All rights reserved.
PII: S 0 1 6 6 - 3 6 1 5 ( 0 0 ) 0 0 0 7 1 - 3
the problems of poor structure, redundancy, and dif®-
culty in maintenance [6,8,16]. To develop such a
decision support system, a well-organized knowledge
base structure that re¯ects how the human experts
solve problems is essential.
To meet this critical need, our research aims at the
following objectives:
1. To develop a knowledge-based system that
cooperates with simulation to support decision
making for manufacturing cell performance im-
provement.
2. To construct a knowledge base structure in
assisting the systematic development of our
proposed knowledge-based decision support sys-
tem.
3. To demonstrate the effectiveness of the knowledge
base for decision support of manufacturing cell
performance improvement.
The research focuses on ¯ow-line type manufacturing
cells where parts travel from upstream to downstream
workstations sequentially in a ®xed route. Every
workstation consists of machines, loaders (i.e. opera-
tors or robots), and a conveyor. The proposed knowl-
edge-based system analyzes outputs from a simulation
model of a manufacturing cell, determines whether the
speci®ed objectives are achieved, and identi®es oppor-
tunities for improvement.
2. Related literature review
2.1. Simulation and knowledge-based systems
An effective approach for improving manufacturing
cell performance is to develop a simulation model for
testing and selecting the con®guration that meets the
desired objectives [2,18]. The primary objective faced
by engineers is to obtain a superior solution by
analyzing manufacturing cell simulation outputs that
include throughput, utilization, time/number in queue,
and time/number in system [9]. Based on this analysis,
engineers would improve the initial system by chan-
ging certain parameters, such as number of machines,
speed of robots or conveyers. This process repeats
until satisfactory results are obtained. However, even
the procedure of analyzing simulation results could
rely on various guidelines and rules, the decision
making still requires signi®cant human expertise
and computer resources. To use simulation ef®ciently
in the decision process, the integration of knowledge-
based systems (also termed as expert systems) with
simulation has been emphasized [4,9,10,13].
O'Keefe developed a taxonomy for combining
simulation models and knowledge-based systems
[10]:
1. Embedded model: The simulation may be em-
bedded within a knowledge-based system, or vice
versa. A knowledge-based system sometimes
needs to run a simulation to obtain results for
the users. On the other hand, a simulation model
may need heuristics for choosing parameters
during the execution.
2. Intelligent-front-end model: A knowledge-based
system functions as an intelligent interface
between the user and a simulation package. It
generates necessary instructions, executes the
simulation, and interprets the results to the user.
3. Parallel model: The simulation and the knowl-
edge-based system are designed, developed, and
implemented as separate software in parallel.
Additional links are built for their communica-
tions.
4. Cooperative model: The simulation and the
knowledge-based system cooperate in performing
the task and sharing the data. The user is able to
access both the simulation and the knowledge-
based system sequentially or concurrently.
In the ®rst three models (embedded, intelligent-front-
end, and parallel models), the user interacts with only
one tool (either simulation model or knowledge-based
system). For instance, Ford and Schroer [4] developed
a system that combines a knowledge-based system
with a commercial simulation language for simulating
an electronics manufacturing plant. Their efforts
focused on providing a natural language interface
so that decision-makers do not have to learn the
simulation language. However, natural language inter-
face is not necessary for the engineers if they could
acquire knowledge and skills in simulation.
In the cooperative model, the user could interact
with both simulation and the knowledge-based sys-
tem. Sagi and Chen [11] proposed a framework that
integrates simulation, neural networks, and knowl-
edge-based system tools for manufacturing cell
34 S.-J. Chen et al. / Computers in Industry 44 (2001) 33±49
design. Simulation is used to estimate performance
measures based on input parameters and given cell
con®gurations. Neural networks are applied to predict
the cell design con®guration and the corresponding
complexities of control functions. Training of neural
networks is performed with both forward and back-
ward methods by using the same pair of data sets for
inputs and outputs, such as performance measures,
cell con®gurations, and cell control function ratings.
The rule-based system is employed to store the expert
knowledge regarding the relation between cell control
functions, cost of controls, performance measures and
cell con®guration. Increasingly, the manufacturing
cell designer or engineer possesses some knowledge
of simulation. In cooperative systems, since the
knowledge-based system does not connect with the
simulation directly, engineers will have a better oppor-
tunity to monitor the procedure of decision analysis as
well as to justify the system's results, if necessary.
This gives the cooperative system more ¯exibility for
making decisions.
2.2. Knowledge base structure
Even though combining simulation and knowledge-
based systems is an effective approach to improving
the performance of manufacturing cells, capturing the
domain knowledge still remains a challenge to the
knowledge-based system developers. It has been
emphasized that the most important issue in develop-
ing a knowledge-based system that cooperates with
simulation is capturing of manufacturing experts'
domain knowledge [13]. This is vital to a good under-
standing on the interactions of manufacturing system
components. However, without an explicit and well-
organized knowledge base structure, the knowledge
rules for decision analysis will be dif®cult to de®ne
and develop. Here, the authors would like to empha-
size the importance of a good structure of a knowl-
edge-based system, which is often overlooked in the
design stage.
A knowledge-based system with well-structured
rules is easy to understand, verify, validate, and hence
to maintain [6,8,16]. Higa [6] emphasized that a rule
base is dif®cult to maintain usually due to its complex
rule structure. As the number of rules increases, the
amount of possible interactions among rules increases
rapidly. This would considerably reduce the ef®ciency
of the system if too many unsubstantiated relation-
ships between rules exist. The integrity of the system
even suffers if invalid or inconsistent results cannot be
prevented. In order to remedy the problems and there-
fore improve the performance of the knowledge-based
system, various methods or algorithms have been
proposed to simplify the rule bases. Higa [6] proposed
a rule simpli®cation procedure to eliminate potential
inconsistency in a rule base and thus improve its
maintainability. His procedure contains four condi-
tions to detect and simplify the structures of the rule
bases: (1) no duplicated rules; (2) no rule that sub-
sumes other rules; (3) no overlapping rules; and (4) no
rule that has the same consequent value and adjacent
antecedent value as any other rules. The procedure
reduces the complexity of the knowledge base,
because the structure of rules and the relationships
between them are simpli®ed to manageable units.
Likewise, Vanthienen and Dries [16] developed a
procedure to restructure and simplify rule bases by
transforming rules into an ef®cient rule base using
decision tables. Their procedure reduces redundancy,
con¯ict and incompleteness of rule bases. Other than
the rule simpli®cation procedure, Mehrotra and Wild
[8] developed a multiviewpoint clustering analysis
method to reveal signi®cant structures within the rule
base and partition the rule-based system into a number
of meaningful units. The method is able to enhance the
comprehensibility, maintainability, and reliability of
knowledge-based systems.
In the aforementioned research, the authors
intended to capture both the explicit and implicit
knowledge by clarifying the existing rule sets in the
knowledge base and then simplifying the rule struc-
ture. For instance, a complete knowledge-based sys-
tem is developed in front, then a series of subsequent
actions are followed in restructuring, simplifying, and
optimizing the rule sets of the built system. We argue
that to develop a knowledge-based system more effec-
tively and ef®ciently, a well-organized knowledge
base structure should be established at the very begin-
ning and be followed throughout the system develop-
ment stages. This will signi®cantly reduce time and
cost in modifying the built rule bases. Hence, it will
increase the quality of the knowledge-based system.
Clearly, there is a strong need of a systematic method
that can guide the development of such a structured
knowledge base.
S.-J. Chen et al. / Computers in Industry 44 (2001) 33±49 35
2.3. The Axiomatic Design method
In the real world, engineers tend to tackle a complex
problem by decomposing it into sub-problems and
attempting to maintain independent solutions for these
smaller problems. This calls for an effective method
that provides guidelines for the decomposition of
complex problems and independent mappings
between problems and solutions. Axiomatic Design
(AD) [15] developed by Suh offers such a good
decomposition mechanism. Two axioms, the founda-
tion of the Axiomatic Design, are formally de®ned as
follows:
Axiom 1. The Independence Axiom
Maintain the independence of the Functional
Requirements (FRs). In an acceptable design, the
mapping between Functional Requirements (FRs)
and Design Parameters (DPs) is such that each
requirement can be satis®ed without affecting any
other requirements.
The mapping between FRs and DPs can be
expressed by the following equation
FR1
FR2
..
.
FRn
8>><>>:9>>=>>; �
A11 A12 � � � A1n
A21 A22 � � � A2n
..
. ...
} ...
An1 An2 � � � Ann
2666437775
DP1
DP2
..
.
DPn
8>><>>:9>>=>>;
Where {FRs} is an n� 1 column matrix (or a vector);
{DPs} is an n� 1 column matrix; and [A] is an n� n
matrix with its component, Aij � @FRi=@DPj, indicat-
ing the relation between FRi and DPj. To satisfy the
Independence Axiom, the design matrix [A] must be
either diagonal or triangular so that the relationships
among FRs and DPs can be either uncoupled or
decoupled which are claimed as good or acceptable
design in AD. An uncoupled design matrix is in the
following form. Design solutions can be performed
concurrently or in any order without affecting each
other in an uncoupled design matrix.
Diagonal Design Matrix
�A� �A11 0 � � � 0
0 A22 � � � 0
..
. ...
} ...
0 0 � � � Ann
2666437775 �Uncoupled�
A decoupled design has the following design matrix.
Tasks can be accomplished sequentially in a
Decoupled Design Matrix.
Triangular Design Matrix
�A� �A11 0 � � � 0
A21 A22 � � � 0
..
. ...
} ...
An1 An2 � � � Ann
2666437775 �Decoupled�
A coupled design matrix, which has non-zero
elements in both upper and lower triangles of the
design matrix, is not recommended by AD because
much iteration will be involved in the design
process.
Axiom 2. The Information Axiom
Minimize the information content of the design.
Among all proposed solutions that satisfy Indepen-
dence Axiom, the best design has the minimum
information content.
The axiomatic approach to design consists of the
following key concepts [5]:
1. The design world consists of distinct domains,
such as the `̀ consumer,'' `̀ functional,'' `̀ physi-
cal,'' and `̀ process'' domains.
2. The design process involves mapping between the
domains.
3. Each domain is de®ned (or characterized) by a
characteristic vector, which can be decomposed
by zig-zagging between functional domain and
physical domain. The physical solutions (i.e. DPs)
should be found before decomposing the corre-
sponding FRs at the same level in the hierarchy.
That is, the entire FR hierarchy cannot be
constructed without referring to the DP hierarchy
at each corresponding level.
4. The mapping process involves creative concep-
tualization, which must satisfy the design axioms,
i.e. the Independence Axiom (Axiom 1) and the
Information Axiom (Axiom 2).
The ®rst axiom facilitates concurrent design without
interactions. The second axiom is a variation of the old
adage `̀ keep it simple.'' They represent two quality
characteristics of the design [3].
Due to its usefulness of basic principles for ana-
lyzing, comparing, and selecting solutions, AD has
36 S.-J. Chen et al. / Computers in Industry 44 (2001) 33±49
been applied in various design ®elds such as
manufacturing, materials, software, organization
and systems since 1990 [14]. In a software system
design [7], FRs are the outputs of a software and DPs
are the key inputs to the software which can char-
acterize or control the FRs. The process variables
(PVs) in the process domains are in the form of
subroutines, operating systems, and compilers. In
addition, Axiomatic Design provides the decision-
making criteria to evaluate different designs. Even
though, in the context of knowledge-based system
design, FRs and DPs may be domain-dependent,
the Independence Axiom and Information Axiom
are general guidelines that are applicable for different
domains.
Other applications of Axiomatic Design include
manufacturing system process improvement [15],
arti®cial skin design [5], software system design
[7], design of paper handling mechanisms of an
ATM (Automatic Teller Machine) [12], structural
design in civil engineering structures [1], and envir-
onmental problem solving [17]. These studies have
convincingly shown the applicability and bene®ts
of AD in solving industrial problems. In addition,
since AD provides the independent mapping bet-
ween each set of FR and its corresponding DP, it
would help relieve the burden of system development
processes.
2.4. Summary of literature review
The research reviewed has shown that a cooperative
model of combining the simulation and knowledge-
based systems is an ef®cient approach. However, the
developed rules are dif®cult to manage due to the
complexity of manufacturing processes. Without a
well-organized knowledge base structure, the true
root causes (e.g. the bottleneck workstations or
machines) may not be clearly exposed by these
rules. Therefore, we need a tool to guide the devel-
opment of a knowledge base systematically. AD is
a useful tool for assisting the system design and
development, in which the framework, criteria, and
methodologies are well established. Numerous
applications have demonstrated that Axiomatic
Design is applicable to solving various engineering
design problems. Using AD as a guideline, the
domain knowledge of simulation and manufacturing
cell can be well structured in the development of
knowledge-based systems.
3. Methodology
3.1. Description of manufacturing cells
In this research, the manufacturing cells consist of
several workstations where parts travel in a ®xed
sequence. A workstation may be a dial-type index
machining center or a stand-alone machine. These
machining resources are linked by a conveyor of a
certain type as part transport means, such as a hook
conveyor, a belt conveyor, or a pallet conveyor. Opera-
tors are assigned to workstations to serve as a part
loader as well as an inspector. Some workstations may
use a robot to automate part loading and unloading.
Hence, each workstation in our case study will incor-
porate machines, loaders (e.g. operators or robots),
and a conveyor.
3.2. System development procedure
Following O'Keefe's taxonomy [10], a cooperative
model of combining simulation and knowledge-based
decision support system is developed to improve the
performance of manufacturing cells. An iterative sys-
tem development procedure shown in Fig. 1 illustrates
how our proposed knowledge-based system coop-
erates with simulation.
First, the current con®guration of the manufacturing
cell is used as the input data to build a simulation base
model. The simulation run of the base model will
produce the current cell performance, such as resource
utilization, blocking percentage, and throughput. They
will be the input to the knowledge-based decision
process. If the performance target is not achieved, the
knowledge-based system will recommend how to
modify the simulation model by varying the number
of machines, the number of operators, conveyor speed,
etc. The iteration continues until a satisfying cell
con®guration is reached, i.e. the performance target
of the cell is met.
In order to provide decision support for identifying
and improving bottlenecks, the working procedure of
our knowledge-based system contains four steps that
identify the bottlenecks from the workstation level to
S.-J. Chen et al. / Computers in Industry 44 (2001) 33±49 37
the resource level. These steps in the knowledge-based
system are as follows:
Step 1 to collect sets of facts at the initial step
including cell configurations, simulation
output, as well as performance criteria,
Step 2 to identify the bottleneck workstations one
by one,
Step 3 to identify the bottleneck resources (i.e.
machine, operator, robot, or conveyor) within
each identified bottleneck workstation,
Step 4 to provide recommendations for improving
the bottleneck resources.
Once the knowledge-based system has made the
recommendations, the simulation model is adjusted
accordingly and rerun. The simulation component and
knowledge-based system component are cooperating
with each other until the objective (target throughput)
is achieved.
Fig. 1. System development procedure.
38 S.-J. Chen et al. / Computers in Industry 44 (2001) 33±49
3.3. Structure of the knowledge base
To assist the development of the proposed knowl-
edge-based system, we introduce a systematic struc-
ture with a four-level knowledge base. The Axiomatic
Design approach helps the design of this four-level
knowledge base structure. There are several advan-
tages for developing the knowledge base structure
using AD as design guidance.
1. Tasks are decomposed top-down so that they can
be achieved with smaller scale.
2. Root causes of the problems at each level are
narrowed down and clearly identi®ed so that
engineers can focus on solving sub-problems one
by one.
3. Tasks are completed systematically with the guide
of Independence Axiom at each level to ensure
ef®ciency of the solution procedure.
AD decomposes a design problem by the mappings
between Functional Requirements (FRs/problems)
and their corresponding design parameters (DPs/solu-
tions). The Independence Axiom maintains the inde-
pendence of Functional Requirements. That is, in the
idea case of uncoupled design, each DP can satisfy
only one FR without affecting any other FRs in the
mapping processes. It is the essence of AD in which a
complex problem is decomposed into sub-problems
(FRs at each level) and maintaining the independent
solutions (one DP for one corresponding FR) for these
smaller problems. Using this `̀ independence'' philo-
sophy in AD, a four-level knowledge base structure is
built and shown in Fig. 2.
The four levels in the hierarchy are formed in terms
of the mapping between FRs and the corresponding
DPs. We intend to improve the performance of man-
ufacturing cell by the cooperation of simulation and
the knowledge base system. The problem sources
(FRs) are those physical components in the manufac-
turing cell that need to be improved. The feasible
solutions (DPs) are the facts and rules in the knowl-
edge base to make the improvement for the cell
performance. These four levels of facts and rules
are systematically applied in the knowledge-based
decision process.
3.3.1. Base level (facts to declare cell con®guration,
simulation output and performance criteria)
Since the primary objective in our study is to
improve the throughput of a manufacturing cell, a
simulation analysis assisted by the knowledge-based
decision process is carried out. A fact list that supports
Fig. 2. Structure of the knowledge base.
S.-J. Chen et al. / Computers in Industry 44 (2001) 33±49 39
all the required information for decision making
before executing the knowledge-based system needs
to be developed ®rst. Details of the three types of facts
needed for the proposed knowledge-based system are
as follows:
1. The ®rst type of facts represents the manufactur-
ing cell con®guration of the simulation model,
including:
1.1. total number of workstations in the cell;
1.2. number of machines within each workstation;
1.3. number of loaders and their types (e.g. robot
or operator); and
1.4. maximum number of machines and operators
allowed due to the space constraint.
2. The second type of facts correspond to the
simulation outputs, containing:
2.1. machine and loader utilization;
2.2. time between part departure at each of the
workstations;
2.3. blocking percentage of each workstation; and
2.4. throughput of the entire cell.
3. The third type of facts indicates the performance
criteria for evaluating the system, including:
3.1. target throughput;
3.2. target average time between departure; and
3.3. maximum utilization and minimum utiliza-
tion limitations.
At this initial level, FR is a manufacturing cell that
needs to increase its throughput to the required target.
The corresponding DP is the facts that collect the cell
con®guration, simulation output, and target criteria.
The FR and the corresponding DP are as follows:
Base level:
FR to improve manufacturing cell throughput,
DP provide sufficient information regarding
the manufacturing cell configuration, simu-
lation output, and performance criteria in a
fact list.
FR� � � A� � DP� �Element A in the design matrix represents the mapping
between FR and DP. Since this level has only one FR
and one DP, the design matrix [A] has only one
element A and thus is uncoupled. The Independence
Axiom is automatically satis®ed.
The design matrix [A], showing the mapping rela-
tionship between FRs and DPs, is an indicator that
determines whether the current level maintains the
independent solutions and satis®es the Independence
Axiom in AD. An uncoupled design matrix [A] with
only non-zero elements on diagonal represents the
best design matrix in AD in which each FR can be
satis®ed by its corresponding DP independently with-
out affecting or being affected by other FRs or DPs. A
decoupled design matrix [A] with non-zero elements
on upper or lower triangle is an acceptable design
because DPs can be performed sequentially. A
coupled design matrix [A] with non-zero elements
on both upper and lower triangles is a bad design
and not acceptable due to many iterations.
3.3.2. Level 1 (rules to identify the bottleneck
workstation)
In order to improve the manufacturing cell through-
put, the bottleneck workstations should be identi®ed
®rst. Thereby the cell improvement processes (i.e. at
Level 2 and Level 3) can be initiated from the recog-
nized bottleneck workstations. However, improving
one bottleneck workstation will affect both upstream
and downstream workstations in a ¯ow-line cell. If
one identi®ed bottleneck workstation is alleviated, the
improved throughput of this workstation may change
some other workstations into the next bottleneck. This
will require iterations in solving the bottleneck pro-
blems. Therefore, all the true bottleneck workstations
have to be identi®ed from the outset. By doing so, the
improving processes can be carried out effectively.
A manufacturing cell with various workstations
linked by certain types of conveyors is a complex
scenario. Commonly, simulation output consists of the
performance measures for each workstation, such as
utilization and blocking percentage of resources, and
throughput, etc. The information that needs to identify
the bottleneck workstation is not provided directly.
Therefore, it is dif®cult to distinguish the key bottle-
neck on the outset without the help of a systematic
analysis. So, an effective approach to correctly iden-
tify the key bottleneck workstation is needed. Usually,
machine or workstation utilization may be suitable to
indicate the ef®ciency of the cell. However, it is not
appropriate for bottleneck identi®cation. A highly
utilized workstation may not be the bottleneck in
the cell. Instead, it may be merely in¯uenced
40 S.-J. Chen et al. / Computers in Industry 44 (2001) 33±49
(blocked) by its succeeding workstations that are the
true bottleneck workstations. For example, a highly
utilized workstation may contain a large portion of
waiting time during which the part has to stay at the
current workstation and cannot be delivered to the
next workstation due to the bottleneck of its succeed-
ing workstations with limited buffer capacity.
In AD, each Functional Requirement (FR) in the
problem domain is satis®ed by a corresponding
Design Parameter (DP) in the solution domain, inde-
pendently. For example, `̀ Determine whether work-
station Wi is the bottleneck?'' represents an FRi in the
problem domain and `̀ developing a rule/rules to
answer FRi'' serves as a DPi in the solution domain.
It is our purpose to develop such rule(s) that enable us
to identify any bottleneck workstation without the
in¯uence from other workstation(s), which is the
essence of independence in AD. At this level, we
decompose FR from the base level into n FRs (from
FR1 to FRn) because there are n workstations in the
manufacturing cell. Each FRi (i � 1; . . . ; n) represents
a workstation that needs to be examined for bottle-
neck. The corresponding DPi is the rules that detect
the bottleneck workstations. Two related issues on
criteria of identifying bottlenecks are developed and
shown in Fig. 3 (1) examine the ®rst workstation for
bottleneck (rule 1±1); and (2) check all other work-
stations for bottleneck (rule 1±2). The advantage of
doing so is that the problems (bottleneck worksta-
tions) can be focused and solved step by step without
the in¯uence from other workstations.
Since parts need to pass every workstation in the
cell, the target throughput has to be maintained at each
workstation. That is, the average time between two
consecutive parts leaving a workstation should also be
maintained for each workstation. This is de®ned as
time between departure (TBD). In addition, any inef-
®cient workstation, except for the ®rst workstation,
could block their upstream workstations that signi®-
cantly affect the throughput of their upstream work-
stations. This is termed as the blocking percentage (B),
which is the proportion of machine blocking time of
the total simulation time. These two factors, TBD and
B, will be used to identify the bottleneck workstations
in the knowledge rules.
1. Examine the ®rst workstation for bottleneck (rule
1±1): Since the ®rst workstation W1 has no
preceding workstation and no upstream work-
station can be blocked by W1, the rule does not
need to consider the blocking percentage (B)
caused by W1. The bottleneck of W1 is checked by
Fig. 3. Rules to identify the bottleneck workstation.
S.-J. Chen et al. / Computers in Industry 44 (2001) 33±49 41
comparing the target TBD (dividing simulation
time by target throughput) with TBD1. If the target
TBD is less than TBD1 that means the workstation
W1 cannot reach the target throughput, then W1 is
a bottleneck workstation. Here, we assume that
there is an adequate source of work for the ®rst
workstation.
2. Check all other workstations for bottleneck (rule
1±2): The bottleneck workstation not only in-
creases TBD, but also blocks its preceding
workstation. Hence, two conditions have to be
checked in this rule. One is TBD, the other is
blocking percentage B. If TBDi of workstation Wi
is greater than TBDiÿ1 of its preceding work-
station, then Wi would have less throughput than
its preceding workstation. Therefore, Wi may be a
bottleneck workstation. In addition, if the block-
ing percentage, Bi, of workstation Wi is less than
Biÿ1 of its preceding workstation, then Wi could
also be a bottleneck workstation because Wi
causes a high value of Biÿ1. That is, Wi causes
more blocking to its preceding workstation Wiÿ1
but suffers less blocking resulted from the next
workstation Wi�1. Therefore, Wi is less ef®cient
than Wiÿ1 and Wi�1. Consequently, the bottleneck
workstation is identi®ed if both situations are
present.
The number of FRs and DPs at this level depends on
how many workstations the cell has. The decomposed
FRs and the corresponding DPs are as follows:
Level 1:
FR1 to identify bottleneck of W1
FRi to identify bottleneck of Wi
FRn to identify bottleneck of Wn
DP1 use rule 1±1 to identify bottleneck W1
DPi use rule 1±2 to identify bottleneck Wi
DPn use rule 1±2 to identify bottleneck Wn
where, W1 is the ®rst workstation in the cell, Wi the ith
workstation in the cell (i � 1; . . . ; n) and Wn is the last
workstation in the cell.
FR1
..
.
FRi
..
.
FRn
2666664
3777775 �A11 0 0 0 0
0 } 0 0 0
0 0 Aii 0 0
0 0 0 } 0
0 0 0 0 Ann
266664377775
DP1
..
.
DPi
..
.
DPn
2666664
3777775
Elements A11 to Ann in the design matrix represent the
mapping relationships between FRs (from FR1 to FRn)
and DPs (from DP1 to DPn), respectively. The design
matrix [A] is uncoupled so that Independence Axiom
is satis®ed.
3.3.3. Level 2 (rules to identify the bottleneck
resource within the bottleneck workstation)
Now we have identi®ed the bottleneck workstation,
the next step is to examine each resource in the
workstation to pinpoint the root cause of not meeting
the target throughput. This is accomplished by using
the knowledge-based decision support system and the
established rules at Level 2. Within each bottleneck
workstation i at this level, we decompose FRi
(i � 1; . . . ; n) from Level 1 into four FRs (from
FRi1 to FRi4) because there are four types of resources
(machine M, operator O, robot R, and conveyor C) in
each workstation. FRi1 to FRi4 represent each resource
that needs to be examined for bottleneck. The corre-
sponding DPs from DPi1 to DPi4 are the rules that
detect the bottleneck resources.
Four groups of rules (rules 2±1, 2±2, 2±3, and 2±4)
regarding checking machine, conveyer, robot, and
operator, accordingly, are developed and shown in
Fig. 4. At this level, the desired TBD of parts and
utilization of machine, robot and operator are the
gauges to identify the bottleneck resources.
If the processing time of a machine, a robot, or an
operator is greater than the target TBD times the
speci®ed minimal utilization level, it implies that
the part stays at the workstation longer than the target
TBD when the resources are at their minimal utiliza-
tion. Therefore, the bottleneck resource is the
machine, robot, or operator, respectively (i.e. rule
2±1, rule 2±3.2, and rule 2±4.2). The multiplication
of the minimal utilization is to guarantee the resource
is able to deliver the target throughput as required. For
example, if a machine's processing time is equal to the
target TBD, the machine cannot produce the target
throughputs without 100% of utilization. A full utili-
zation of resources is not recommended in industrial
practice due to the extreme demand on maintenance of
the resources, as well as the dif®culty of production
planning and scheduling of manufacturing cells.
The utilization of machine, robot, or operator has to
be checked (i.e. rule 2±2, rule 2±3.1, and rule 2±4.1) if
their processing time is less than or equal to the target
42 S.-J. Chen et al. / Computers in Industry 44 (2001) 33±49
TBD times the speci®ed minimal utilization level. In
rule 2±2, if the utilization of machine and robot (or
operator) is lower than the minimal level, it implies
that they are always waiting for the part. Therefore,
the conveyer is the bottleneck. In rules 2±3.1 and 2±
4.1, on the other hand, if the utilization level of
machine is lower than the minimal level and the loader
(either a robot or an operator) has higher utilization
than the speci®ed maximal level, the loader is too busy
to handle the incoming parts. In this case, the loader is
the bottleneck.
For each bottleneck workstation Wi, the four FRs
and their corresponding DPs are as follows:
Level 2:
FRi1 to identify bottleneck resource M (ma-
chine) at workstation i
FRi2 to identify bottleneck resource C (con-
veyor) at workstation i
FRi3 to identify bottleneck resource R (robot) at
workstation i
FRi4 to identify bottleneck resource O (operator)
at workstation i
DPi1 use rule 2±1 to identify bottleneck resource
M at workstation i
DPi2 use rule 2±2 to identify bottleneck resource
C at workstation i.
DPi3 use rule 2±3.1 and rule 2±3.2 to identify
bottleneck resource R at workstation i
DPi4 use rule 2±4.1 and rule 2±4.2 to identify
bottleneck resource O at workstation i
FRi1
FRi2
FRi3
FRi4
2666437775�
Ai1ÿi1 0 0 0
0 Ai2ÿi2 0 0
0 0 Ai3ÿi3 0
0 0 0 Ai4ÿi4
2666437775
DPi1
DPi2
DPi3
DPi4
2666437775
Elements Ai1ÿi1 to Ai4ÿi4 in the design matrix repre-
sent the mapping relationships between FRs (from
FRi1 to FRi4) and DPs (from DPi1 to DPi4), respec-
tively. The design matrix [A] is uncoupled so that
Independence Axiom is satis®ed.
3.3.4. Level 3 (rules to improve the throughput)
After the bottleneck resources within the worksta-
tion are identi®ed, the improvement processes can
then be started. Within each bottleneck workstation
i at this level, we decompose each of the FRi1, FRi2,
Fig. 4. Rules to identify the bottleneck resource.
S.-J. Chen et al. / Computers in Industry 44 (2001) 33±49 43
FRi3, and FRi4 from Level 2 into the four correspond-
ing FRs (from FRi11 to FRi44) because each identi®ed
bottleneck resource needs to improve accordingly.
The FRs from FRi11 to FRi44 represent each kind of
resources that needs to be improved if it is the bottle-
neck resource. The corresponding DPs from DPi11 to
DPi44 are the rules that improve the bottleneck
resources.
Four groups of rules are developed as guidelines for
solving the key bottleneck resources and thus improv-
ing the throughput (see Fig. 5). Due to limited space
capacity, the number of machines in the workstation
needs to be constrained by a maximum allowable
number. If a machine is the bottleneck and the space
is still enough for accommodating extra machines, the
rules will suggest adding one machine (rule 3±1.1). On
the other hand, if there is no space for additional
machine, the rules will suggest replacing the current
machine (rule 3±1.2). If the bottleneck is a conveyor or
a robot, it needs to adjust its speed (rules 3±2 and 3±3).
Otherwise, if an operator is the bottleneck, one addi-
tional operator will be suggested (rule 3±4).
Four FRs and four corresponding DPs to improve
the bottleneck resources are as follows:
Level 3:
FRi11 to improve bottleneck resource M (ma-
chine) at workstation i
FRi22 to improve bottleneck resource C (con-
veyor) at workstation i
FRi33 to improve bottleneck resource R (robot) at
workstation i
FRi44 to improve bottleneck resource O (opera-
tor) at workstation i
DPi11 use rule 3±1.1 and rule 3±1.2 to improve
bottleneck resource M at workstation i
DPi22 use rule 3±2 to improve bottleneck resource
C at workstation i
DPi33 use rule 3±3 to improve bottleneck resource
R at workstation i
DPi44 use rule 3±4 to improve bottleneck resource
O at workstation i.
FRi11
FRi22
FRi33
FRi44
2666437775 �
Ai11ÿi11 0 0 0
0 Ai22ÿi22 0 0
0 0 Ai33ÿi33 0
0 0 0 Ai44ÿi44
2666437775
DPi11
DPi22
DPi33
DPi44
2666437775
Elements Ai11ÿi11 to Ai44ÿi44 in the design matrix
represent the mapping relationships between FRs
(from FRi11 to FRi44) and DPs (from DPi11 to
DPi44), respectively. The design matrix [A] is
uncoupled so that Independence Axiom is satis®ed.
4. An illustrative example
To validate the proposed knowledge-based system,
an example that uses a real manufacturing cell in the
Fig. 5. Rules to improve the throughput.
44 S.-J. Chen et al. / Computers in Industry 44 (2001) 33±49
industry with eight workstations is implemented and
tested. The simulation model is built based on actual
data from this manufacturing cell shown in Fig. 6.
Fig. 7 shows the simulation model generated by the
discrete event simulation software system ProModel.
The simulation time is based on one standard shift,
which equals to 7.33 h (8 h with 40 min break time
from labor contract). The performance criteria for this
example are speci®ed in the following:
1. Target throughput � 400 parts/shift;
2. Target TBD � 7:33h/400 parts � 66 s;
3. Minimal utilization � 60%; and
4. Maximal utilization � 80%.
The required con®guration data are shown in Table 1.
In addition, it is important to note that many of the
tasks in Fig. 6 are stochastic in nature, i.e. the machine
processing times and loader cycle times are various in
Fig. 6. The con®guration of manufacturing cell example.
Table 1
Con®guration of the example
Workstation 1 2 3 4 5 6 7 8
No. of machine 1 0 1 2 1 1 2 0
Loader type Operator Robot Operator Operator Operator Operator Operator Operator
No. of operator 1 0 1 1 1 1 1 1
Max no. of machine 1 0 2 3 1 2 3 0
Max no. of operator 2 0 2 2 2 2 2 2
S.-J. Chen et al. / Computers in Industry 44 (2001) 33±49 45
a range. Table 2 de®nes the necessary stochastic
elements in the simulation, such as processing times
of workstations or machines, cycle times of loaders
(operators or robots), and speeds of conveyors. From
the analysis of collected data, the inter-arrival rate of
incoming parts at workstation W1 is best represented
as a triangular distribution with values of minimum,
mean, and maximum equal to 10, 60, and 1200 s,
respectively.
Simulation run of the current cell indicates that the
throughput is 249. The performances are shown in
Table 3.
Fig. 7. The simulation model of manufacturing cell example.
Table 2
Stochastic elements in the simulation
Workstation no./loader no. Machine processing time (s) Loader cycle time (s) (min., mean, max.)
(a) The eight workstations:
W1/operator 1 43.2 � 0.1 (28.0, 33.5, 42.0)
W2/robot ± (8.6, 11.4, 16.8)
W3/operator 2 1.0 � 0.2 (19.0, 24.9, 33.0)
W4/operator 3 30.0 � 0.2 (38.0, 46.7, 60.0)
W5/operator 4 44.5 � 0.3 (35.0, 43.6, 56.0)
W6/operator 5 43.4 � 0.6 (18.0, 22.9, 30.0)
W7/operator 6 17.4 � 0.2 (19.0, 23.9, 30.0)
W8/operator 7 2.7 � 0.1 (15.0, 20.3, 30.0)
Conveyor no. Type Linked workstations Distance (ft) Capacity (pcs) Speed (ft/min)
(b) The seven conveyors:
Conveyor 1 Pallet belt W1 & W2 15 9 32.0
Conveyor 2 Furnace W2 & W3 26.5 72 7.3
Conveyor 3 Circle belt W3 & W4 128 24 32.0
Conveyor 4 Hook W4 & W5 90.7 34 16.8
Conveyor 5 Belt W5 & W6 16 24 32.0
Conveyor 6 Belt W6 & W7 13.7 32 27.5
Conveyor 7 Belt W7 & W8 7 20 23.8
46 S.-J. Chen et al. / Computers in Industry 44 (2001) 33±49
Since the throughput is less than the target, analysis
of the simulation output is necessary for possible
improvement. However, it is dif®cult to know how
to improve the current manufacturing cell by simply
inspecting the output data. For example, from Table 3,
the resource utilization of workstations W1 (95% for
machine) and W8 (100% for loader) are very high so
that W1 and W8 could be the bottlenecks. Additionally,
TBDs of all workstations are greater than the target
TBD (66 s), thus, all workstations in the cell may be
the bottleneck. If only examining the face values of
these outcomes, the true bottleneck workstations can-
not be identi®ed. With the help of our knowledge-
based system, the problems are easily revealed and the
recommendations are provided.
First, the simulation output data shown in Table 3
are formulated as the facts for the knowledge-based
system. With these facts, the rules at three levels are
®red one after another as shown in the following
demonstration.
4.1. From Level 1 (rules to identify the bottleneck
workstation)
For the ®rst workstation W1, since its average TBD1
of parts is longer than the target (TBDtarget � 66
< TBD1 � 83:75), Rule 1±1 is ®red and W1 is iden-
ti®ed as a bottleneck workstation.
Rule 1±2 is used to identify whether the subsequent
workstations are the bottleneck or not. For instance,
the average TBD of workstation W5 is longer than that
of its upstream workstation W4 (TBD5 � 90:59 >TBD4 � 88:47). Also, the average blocking percen-
tage is less than that of W4 (B5 � 50:94 < B4 �66:26). This means workstation W5 not only takes
longer to process a part but also contributes to the
blocking percentage of W4. Therefore, it is identi®ed
as a bottleneck workstation by Rule 1±2. Similarly,
workstations W6 and W8 are identi®ed as bottleneck
workstations by Rule 1±2.
4.2. From Level 2 (rules to identify the bottleneck
resource within the bottleneck workstation)
For bottleneck workstations W1, W5, W6, and W8,
the bottleneck resources are identi®ed as follows:
At workstations W1 and W6, the machine processing
times (MPT1 � 40, MPT2 � 68) are longer than the
target time between departure times the minimal
utilization (TBDtarget � 60% � 39:6). Rule 2±1 is
®red, and the bottleneck is identi®ed at the machine.
At work station W5, although the machine proces-
sing time (MPT5 � 22:5) is less than the target time
between departure times the minimal utilization, the
utilization of machine and operator is less than 60%.
Therefore, Rule 2±2 is ®red, and the conveyer is a
bottleneck resource.
Although there is no machine at workstation W8, the
operator processing time (OPT8 � 40:6) is longer
than the target time between departure times the
minimal utilization. Rule 2±4.2 is ®red, and the
operator is a bottleneck.
4.3. From Level 3 (rules to improve the throughput)
For bottleneck resources found, the recommenda-
tions for throughput improvement are generated as
follows:
For the bottleneck machine at workstation W1, since
the number of machine is already equal to the maximal
value. The only way to reduce processing time is to
replace it with another machine that has shorter
processing time (Rule 3±1.2). For the bottleneck
machine at workstation W6, since the number of
machine is still less than the maximal value. The
processing time could be reduced by adding one
Table 3
Simulation output of the example Ð the current cell
Workstation 1 2 3 4 5 6 7 8
Machine utilization 0.95 N/A 0.20 0.34 0.49 0.74 0.30 N/A
Loader utilization 0.50 0.29 0.29 0.61 0.57 0.31 0.43 1.00
TBD (s) 83.75 83.72 85.25 88.47 90.59 98.09 104.85 105.86
Blocking (%) 4.75 0.00 6.56 66.26 50.94 20.39 53.89 0.00
S.-J. Chen et al. / Computers in Industry 44 (2001) 33±49 47
machine (Rule 3±1.1). For the bottleneck conveyer at
workstation W5, a faster speed is needed to increase
the utilization of the machine and the operator (Rule
3±2). For the bottleneck operator at workstation W8,
since the number of operators is still less than the
maximal value, the processing time could be reduced
by adding one operator (Rule 3±4).
All the recommendations from the proposed knowl-
edge-based system are listed in Fig. 8.
The con®gurations of the initial simulation model
are then modi®ed according to the recommendations.
After the next simulation run, the throughput equals
to 416 for the modi®ed cell model. The target is now
achieved. The throughput has improved by 67% from
249 to 416 parts per shift. The engineers can use the
new set of con®gurations to improve the real manu-
facturing cell.
5. Conclusions
A knowledge-based system that cooperates with
simulation has been developed. They compensate
each other in assisting the decision making for man-
ufacturing cell improvement. With our knowledge
base decision support, the key bottleneck workstations
as well as bottleneck resources are clearly identi®ed.
Hence, the improvement processes can be carried out
precisely. In solving this decision process, a hierarch-
ical structure of knowledge-based system is con-
structed. The Independence Axiom in Axiomatic
Design has been followed during the establishment
of the knowledge base structure. Unlike existing
research that only attempt to simplify the structure
of a knowledge base after it is built, our work empha-
sizes the development of a good knowledge base
structure even before it is built. Such a sound structure
will help build the knowledge base systematically
with good solution ef®ciency and consistency. To
demonstrate the effectiveness of our proposed knowl-
edge-based system, a real industry case is used. The
simulation results show that the suggestions provided
contribute to increasing the throughput. In conclusion,
the engineer can improve the manufacturing cell with
the help of the knowledge-based system and the
simulation. It reduces the burden of engineers by
revealing the problem sources and providing recom-
mendations for solving the problems. Moreover, the
study demonstrates the applicability and usefulness of
AD in the design of knowledge-based decision support
system.
References
[1] L.D. Albano, N.P. Suh, Axiomatic approach to structural
design, Research in Engineering Design 4 (3) (1992) 171±
183.
[2] C.J. Antonelli, R.A. Volz, T.N. Mudge, Hierarchical decom-
position and simulation of manufacturing cells using Ada,
Simulation 46 (4) (1986) 141±152.
[3] B. Bras, F. Mistree, A compromise decision support problem
for axiomatic and robost design, ASME Advances in Design
Automation DE 65±1 (1993) 359±369.
[4] D.R. Ford, B.J. Schroer, An expert manufacturing simulation
system, Simulation 48 (5) (1987) 193±200.
[5] D.A. Gebala, N.P. Suh, An application of axiomatic design,
Research in Engineering Design 3 (3) (1992) 149±162.
[6] K. Higa, An approach to improving the maintainability of
existing rule bases, Decision Support Systems 18 (1) (1996)
23±31.
[7] S.J. Kim, N.P. Suh, S.G. Kim, Design of software systems
based on axiomatic design, Robotics & Computer-Integrated
Manufacturing 8 (4) (1991) 243±255.
[8] M. Mehrotra, C. Wild, Analyzing knowledge-based systems
with multiviewpoint clustering analysis, Journal of Systems
and Software 29 (3) (1995) 235±249.
[9] J.M. Mellichamp, O. Kwon, A.F.A. Wahab, FMS designer: an
expert system for ¯exible manufacturing system design,
International Journal of Production Research 28 (11) (1990)
2013±2024.
[10] R. O'Keefe, Simulation and expert systems Ð a taxonomy
and some examples, Simulation 46 (1) (1986) 10±16.
[11] S.R. Sagi, E.F. Chen, A framework for intelligent design of
manufacturing cells, Journal of Intelligent Manufacturing 6
(3) (1995) 175±190.
Fig. 8. Output recommendations from the knowledge-based
system.
48 S.-J. Chen et al. / Computers in Industry 44 (2001) 33±49
[12] S. Sekimoto, M. Ukai, A study of creative design based on the
axiomatic design theory, in: Proceedings of the 6th Interna-
tional Conference on Design Theory and Methodology,
DE,Vol. 68, 1994, pp. 71±77.
[13] R.E. Shannon, Knowledge-based simulation techniques for
manufacturing, knowledge-based systems in manufacturing,
in: A. Kusiak (Ed.), Taylor & Francis, Philadelphia, 1989,
pp. 305±329.
[14] N.P. Suh, Design and operation of large systems, Journal of
Manufacturing Systems 14 (3) (1995) 203±213.
[15] N.P. Suh, The Principles of Design, Oxford Press, New York,
1990.
[16] J. Vanthienen, E. Dries, Restructuring and simplifying rule
bases, in: Proceedings of the International Conference on
Tools with Arti®cial Intelligence, 1995, pp. 484±485.
[17] D.R. Wallace, N.P. Suh, Information-based design for
environmental problem solving, Annals of the CIRP 42 (1)
(1993) 175±180.
[18] E.F. Watson, R.P. Sadowski, Developing and analyzing
¯exible cell systems using simulation, in: Proceedings of
the 1994 Winter Simulation Conference, 1994, pp. 978±
985.
Shi-Jie (Gary) Chen is an Assistant
Professor of Industrial Engineering at
the National Taipei University of Tech-
nology, Taipei, Taiwan. He received his
BS degree in Automatic Control Engi-
neering (1989) from Feng-Chia Univer-
sity, Taiwan, and completed his MS
degree in Mechanical Engineering
(1995) from State University of New
York at Buffalo. He completed his Ph.D.
in Industrial Engineering (1999) from
State University of New York at Buffalo. His research interests
include concurrent engineering, project management, simulation,
knowledge-based systems and CAD/CAM/CIM.
Li-Chieh Chen is an Assistant Professor
in the Department of Industrial Design at
Tatung University, Taipei, Taiwan. He
received his BS degree in Mechanical
Engineering (1990) from Feng-Chia Uni-
versity and completed his MS degree in
Mechanical Engineering (1992) from
Tatung Institute of Technology, Taiwan.
He completed his Ph.D. in Industrial Engi-
neering (1999) from the State University
of New York at Buffalo. His research
interests include product design methodologies, concurrent engi-
neering, computer-integrated manufacturing and arti®cial intelli-
gence applications in design.
Li Lin is an Associate Professor of Indus-
trial Engineering at the State University
of New York, Buffalo. He received a BS
in Mechanical Engineering (1982), an
MTech (1984) in Graphic Communica-
tions and an MSE in Industrial Engineer-
ing (1986). After his Ph.D. in Industrial
and Management Systems Engineering
from Arizona State University in 1989,
Dr. Lin has been a faculty member at
SUNYat Buffalo. His research interests are in modeling and control
of ¯exible manufacturing systems, product life-cycle design in
concurrent engineering and computer simulation. Dr. Lin's research
has been funded by the National Science Foundation and a number
of industrial companies, including American Axle and Manufactur-
ing, Delphi Harrison Thermal Systems, DuPont, Leica and Praxair.
He is a senior member of the Institute of Industrial Engineers (IIE).
S.-J. Chen et al. / Computers in Industry 44 (2001) 33±49 49