An Expert System for Conceiving Company
Wide Quality Control Strategies
Julio Macedo
University of Montreal Business School
and Institut de Stratégies Industrielles,
229 Forest, Pincourt, P.Q., J7V8E8, Canada
Rafael Ruiz Usano
University of Sevilla, School of Industrial
Engineering, R. Mercedes, 41012 Sevilla, Spain
Abstract
Currently there are three methods for conceiving company wide quality control
strategies, the participative approach, the simulation approach and the benchmarking
approach. All of these methods present shortcomings. This paper presents a new
qualitative method based on system dynamics and an expert system that includes this new
method. The expert system is applied to conceive the company wide quality control
strategy of a manufacturing firm.
1. INTRODUCTION
A product of quality is the one that satisfies the physical and managerial characteristics
expected by the customer (Mizuno, 1989). The physical characteristics of the product are
the ones associ; atéd to the structure of the product (precision, strength, dimensions) while
the managerial “characteristics are the ones associated to the management of the system
that makes the product (delivery delay, cost, variety of models). A company-wide quality
control (CWQC) strategy or total quality control strategy "involve the whole company
-every division and every worker at every level- and requires the integration of such
formerly independent functions as raw material purchasing, ‘work procedure analysis,
work procedure management and inspection" (Mizuno, 1989, p.16). Hence, the
conception of a CWQC strategy is a conceptual step before the detailed design and
implementation of the improvements. A CWQC strategy includes only the profiles of the
improvements because its main goal is to guarantee that these improvements will together
produce a product of quality.
What the customer desires exactly is a difficult question and it is not the goal of this
paper to answer it. It is assumed that a product is of quality when it shows "zero"
defects, "zero" delivery delay, "zero" cost and an “infinity” variety of models, These
desired characteristics come from the Just-in-Time systems goals (Gilbert, 1990).
However, the firms that best satisfy the customer’s desires (the industry leaders) have
- 385 -
customer's _ products sold by the
F desires industry leaders
™ —V
H
New -—-» _ [product proble- desired product
f matic behaviors behaviors:
( , mar delay zero delay
p——” x defects zero defects
J x cost zero cost
B x variety max variety
Ga Ca
poor quality product quality product
Problematic system:
Variables that interact in order to produce
the problematic product behaviors in the
firm analyzed.
Figure 1. A company wide quality control strategy: Given a set of problematic product behaviors,
identify what variables of the problematic system modify and how to modify them in order that the
desired product behaviors be reached,
Methods to conceive
company wide quality Principal characteristics and shortcomings
control strategies
Participative approach | A cause effect diagram that describes the problematic system is built by a
(Ozeki, 1990) participative approach. The CWQC strategy is found by identifying the
sources of the problematic product behavior using the cause-effect diagram
and brainstorming sessions. This method has no tools to quantify the effect of
the CWQC strategy on the problematic system.
Simulation approach __| A model of the problematic system is built using system dynamic principles
(Macedo, 1992) and a continuous/discrete simulation language. The CWQC strategy is found
by intensive simulation of the model. Although this method quantifies the
effect of the CWQC strategy (and particularly the counterintuitive effects) on
the problematic system the construction of the simulation model is long.
Benchmarking approach | A set of indicators that mesure the gap betwen the performance of the
(Camp, 1989) problematic system variables and the industry leader system variables are
defined. Next, these indicators are measured and the reasons of the superior
performance of the industry leader investigated. The CWQC strategy consists
of adopting the succesful solutions implemented by the industry leader. The
principal problem with this approach is the measurement of the indicators. In
addition, the use of quantitative indicators can produce the refusal of the
industry leader to cooperate in explaining its superior performance.
Table 1. Principal characteristics and shortcomings of the methods used to conceive company wide
quality control strategies.
- 386 -
total or partially implemented Just-in-time systems (Hayes, Wheelwright and Clark,
1988). 1,
As indicated in figure 1, the conception of a CWQC strategy begins when a product
is of poor quality and consists of two steps. First, analyzing the current product and
manufacturing system in order to identify the set of variables that interact in order to
produce a product of poor quality (problematic system). Second, specifying what
variables to modify and in what form so that the product becomes one of quality.
Currently there exists three methods to conceive CWQC strategies: The participative
approach (Ozeki, 1990), the simulation approach (Macedo, 1992) and the benchmarking
approach (Camp, 1989). As indicated in table 1, all these methods present shortcomings.
This paper presents a new method for conceiving CWQC strategies that does not show
the shortcomings of the current methods. In addition it presents an expert system whose
knowledge base is nourished by the CWQC strategies obtained using the new method.
In the third part of the paper the use of the new method for conceiving CWQC strategies
and the use of the expert system are illustrated in a manufacturing case.
2. THE EXPERT SYSTEM
The suggested expert system is presented in figure 2 and shows the three classical
components of an expert system: Knowledge base, working memory and inference engine
(@adiru, 1992). The knowledge base stores the reference structures, i.e. the structures
of the problematic systems whose CWQC strategy is known. The working memory stores
the structure of the problematic system under analysis. Finally, the inference engine is
the set of procedures that verify if the structure in the working memory matches one of
the reference structures in the knowledge base. If yes, the CWQC strategy becomes
known from the knowledge base. When the matching is not possible, the inference engine
calls an external program in order that the user finds the CWQC strategy by simulating
the system dynamics model of the problematic system. Once the CWQC strategy is
obtained it is added to the knowledge base as well as the structure of the corresponding
problematic system.
The structure of a system dynamics model is ordinarily represented by a level-rate
model (Richardson, 1981). But, this latter generates a set of nonlinear integral equations
complex enough to be stored in the knowledge base of an expert system. In order to
eliminate this problem, the level-rate model of the problematic system (first box in figure
2) must be built using integral equations that have the following structure:
X, = Xj- dt. RTy . [X; - log(X)] 0)
Xp known, 0<X)S1 ; Q)
RTy= A-X, + B.RT, +C.P, G)
k=kl=t; j=jk=t-1; O<t<T
where:
A: Impact of a level variable on X
X, : Level variable at time k B : Impact of a rate variable on X
RT,;: Rate variable at time Ki C : Impact of a control variable on X
P, : Control variable at time k dt : Interval of time
No
The analysis group tentatively represents the
system that generates the problematic
behaviors as a level-rate model. The cause-
effect relationships of the model and the
impacts betwen its level and rate variables
are specified and stored in the working
Memory
Tool: Interactive DYNAMO interfaced with
the working memory of the expert system.
t
Simulate the level rate-model.
Does the suggestion group agree on that
the simulation outputs represent the
problematic behaviors?
Tool: Interactive DYNAMO
| Yes
Do the cause-effect relationships and the
impacts stored in the working memory
match the ones included in the knowledge
base?
Tool: Inference engine of the expert system
| No
The conception group tentatively identifies
the variables that must be modified and
how to modify them in order that the
problematic behaviors reach the desired
behaviors.
Tool; Checklist (Macedo, Ruiz Usano, 1991).
'
Simulate the level-rate model with the
modifications included. Does the
suggestion group agree that the simulation
outputs represent the desired behaviors?
Tool: Interactive DYNAMO-
y Yes
Knowledge base: Includes the cause-
effect relationships of the known level-
rate models and the impacts between
its level and rate variables
Tool: Knowledge base of the expert
system
The strategy to be used for correcting the
problematic behaviors is the one found
aT |
The strategy to be used is |
the conclusion clause
identified by the inference
engine in the knowledge
base of the expert system
Include the cause-effect
relationships of the level
rate model and the impacts
between its level and rate
variables and its found
strategy on the knowledge
base.
Tool: Rule adjuster of the
expert system
Figure 2. An expert system for conceiving company wide quality control strategies. Note that
the expert system uses a participative approach with two groups: the analysis group and the
conception group.
- 388 -
As noted any level variable has an initial value between 0 and 1 so that the modulating
factor [X;*log(X;)] will produce values for the level variables that are always between 0
and 1, In addition, only two data must be specified in order to build any equation of the
level-rate model: if there is a cause-effect relationship between two variables and the
value of the impact coefficient A, B or C!. These two data for all linked variables define
completely the structure of the level-rate model. These two data can be easily stored in
the knowledge base of an expert system.
The searched CWQC strategy is constituted by the new values of the control variables
P and the coefficients A, B and C that modify the current behaviors of the level variables
so that they reach the desired behaviors. Note that the new values of A, B, C and P can
strengthen or weaken the current links between the variables or create new links.
The suggested expert system was implemented using VP-EXPERT shell (Badiru, 1992)
that uses production rules for representing knowledge and backward and forward chaining
as search procedures.
3. A CASE STUDY
The use of the expert system (figure 2) will be illustrated by applying it to conceive
the CWQC strategy of a sector of a manufacturing firm. Simultaneously additional
characteristics of the expert system will be presented.
Most of the time the product’s technical characteristics are well known but the product
is of poor quality because the system that makes it is badly organized. Hence, the
. conception of a CWQC strategy consists finally in specifying the profiles of the changes
that must be introduced in the organization of the manufacturing system. These changes
belong to six sectors: product structure, manufacturing flow, set up operations, human
factor, production planning and automation (Macedo and Ruiz Usano, 1991). In order
to shorten this paper the CWQC strategy must be limited to the set up operations in a
small factory producing paper napkins”.
The napkins are made automatically in a machine that stamps, folds and puts in boxes
the pieces of paper that it receives, The production director wants to reduce as
much as possible the current production delay and the percentage of defective napkins
produced by the napkins machine.
' Note that the impact coefficients follow the ordinary rules of system dynamics
methodology. Hence, when the impact coefficient is positive, the growth (decline) of the cause
variable produces the growth (decline) of the effect variable. On the other hand, when the impact
coefficient is negative, the growth (decline) of the cause variable produces the decline (growth)
of the effect variable.
? In practice it is suggested to include the six sectors above in the level-rate model because
they are interrelated. This can be done without problems because the form of the equations (1),
(2), (3) do not limit the size of the level-rate model. In addition, the tools for implementing the
expert systems can accept hundreds of production rules.
- 389 -
c1=03%,
~.
7
7
&
(03 =03
BROKE
Broken tools
Symbols used:
4 TTOOLI
Tools far from the
working place
- FAR
Tools not found _//
Lost.
}
yf C2=03
TTOOL’
Time to obtain tools ~~
The equipment
design Installation operations
facilitates its installation well mastered
‘Time to reach
operating conditions
- ,
A6=08-7
/7) MASTE
7 C5205
»
TSTOP
‘Time that the
principal machine is
stopped
Set up time of the
principal machine
SET UP
‘Stock of products
in process
a
Ci: Impacts
/
‘AT = 0.05
/
rf
Z
DEL
7 Production
7 delay
of the control variables
Impacts of the endogeneous variables
Figure 3. Level-rate model for the set-up operations of the paper napkins machine. Note that the
problematic variables are in dark lines while the strategy is constituted by the simultaneous
increase /\ and decrease
of the indicated control variables.
As the first box of figure 2 indicates, a level-rate model is constructed. This one is
shown in figure 3 and the output of its simulation is in figure 4. Its construction began
by writing on a blackboard the problematic variables DEL, DEFEC and COST. Next,
the workers and managers related to these variables (analysis group) suggested the causes
of their high patterns and the associated impact coefficients. This information was added
to the model and the model was simulated. Following the resultant outputs, some
variables were eliminated and new ones added until the high patterns of DEL, DEFEC
and COST were generated. These high patterns can be explained by looking at figure 4
and following the paths of figure 3: LOST, FAR, BROKE are high enough, DESIG,
MASTE low enough and CONDIT high enough that SETUP is high so that STOCK
grows producing the growth of DEFEC and DEL.
It is important to note that the abstraction level used for representing the set-up
operations in figure 3 is higher than the one used in detailed discrete simulation and in
traditional system dynamics. The level-rate model is a meta-model that does not represent
the time-space details of the current organization but that captures the resultant behavior
patterns using qualitative signs (values and signs of the impact coefficients that emerge
from the past experiences of the participants).
The CWQC consists of decreasing LOST, FAR, BROKE, increasing DESIG, MASTE
and decreasing CONDIT (figure 3). This means reducing the distance that separates the
tools from the working place, reducing the tools unavailable or broken, redesigning the
current equipment so that it becomes easy to install, training the operators so that they
master the installation operations and decreasing the time that the principal machine needs
to reach acceptable operating conditions. As indicated in figure 5, this strategy once
introduced in the level-rate model produces the desired behaviors.
The CWQC strategy was found by a conception group constituted by managers and
workers of the quality leader factory in the paper napkins sector. As indicated in figure
2, this group used simulations of the level-rate model to confirm’ the effects of its
suggestions (this step is necessary because the humans misperceive the effects of the
feedback loops). The high level managers of the leader factory were not opposed to the
participation of their workers in the improvement conception group of a competitor.
These managers considered that the kind of information required from their workers
(values and signs of the impact coefficients) do not represent confidential information.
This event demonstrates that the qualitative level-rate models by avoiding the quantitative
measurements allow to implement the benchmarking foundations (table 1).
Once the CWQC strategy is known, it is introduced in the knowledge base of the
expert system as well as the structure of the associated level-rate model (figure 3). This
is done using production rules as in figure 6. The last rule of the knowledge base, the
one that includes as a conclusion clause the known CWQC strategy, can have the
following form (see figure 3):
- 391 -
Adimentional
ot
scale ie >
0 10 20 30 40 Time
Figure 4. Pattern behaviors obtained by simulating the level-rate model with the current vafues of the
control variables: LOST = 1; FAR = 1; BROKE = 1; DESIG = 0.1; MASTE = 0.1; CONDIT = 0.3. The
symbol used are: SIZE (size of the production lots), SET UP (set up time of the principal machine),
DEFEC (defective products), STOCK (stock of products in process), DEL (production delay), RUNS
(number of production runs). , §
0.75
0.50
0.25
Adimentional,
scale
0 10 20 30 40 Time
Figure 5. Pattern behaviors obtained by simulating the level-rate model with the CWQC strategy
included: LOST = 0.01; FAR = 0.01; BROKE = 0.01; DESIG = 0.3; MASTE = 0.3; CONDIT = 0.01.
The symbol used are: SIZE (size of the production lots), SET UP (set up time of the principal
machine), DEFEC (defective products), STOCK (stock of products in process), DEL (production
- 392 -
LOST
influences
TTOOL 1
C1403 Cc1=03
FAR
influences
TTOOL 1
No Yes
Impact
c2
C2 #03 C2=0.3
BROKE
influences
TTOOL 1
f No Yes
Impact
C3
RULE1
if LOST - TTOOL 1= Yes C3=0.3
and C1 = 0,3
and FAR - TTOOL | = Yes
and C2 =0.3 LOST, FAR
and BROKE - TTOOLI = Yes and BROKE
and C3 = 0.3 influence
then CAUSE-EFFECT = LOST. FAR . BROKE - TTOOL1 TTOOL 1
Figure 6. Rule 1 of the knowledge base that represents the structure of the level-rate model. The
variables used are defined in figure 3.
IF CAUSE-EFFECT=RUNS_STOCK_DEFEC_DEL-COST
AND RUNS_COSTI=YES
AND A12=0.10
AND STOCK_COST1=YES
AND A11=0.10
AND DEFEC_COST1=YES
AND A10=0.20
AND DEL_COST1=YES
AND A13=0.20
THEN STRATEGY =DECREASE_LOST FAR_BROKE_INCREASE_DESIG_
MASTE_DECREASE_CONDIT
Hence, when another paper napkin factory wants to reduce the production delay and
the percentage of defective products of its paper napkin machine it can profit of the
knowledge base just created. It must introduce the structure of its particular level-rate
model (cause-effect relationships and values of the impact coefficients) in the working
memory of the expert system. If this structure corresponds to the one analyzed in this
paper, the expert system will answer with the CWQC strategy to be used. With the
progressive use of the expert system, the knowledge base will be enriched so that the use
of simulation exercises for conceiving the CWQC strategies will become unnecessary.
REFERENCES
Badiru A. (1992), Expert systems: Applications in engineering and manufacturing,
Prentice Hall, Englewood Cliffs, N. J.
Camp R. C. (1989), Benchmarking: The search for industry best practices that lead to
superior performance, Quality Press, Milwaukee, WI.
Gilbert J.P. (1990), The state of JIT implementation and development in the USA,
International Journal of Production Research, 28(6), p. 1099-1109.
Hayes R. H., S. C. Wheelwright, K. B. Clark (1988), Dynamics manufacturing:
Creating the learning organization, The Free Press, New York, NY.
Mizuno Sh. (1989), Company wide total quality control, Asian Productivity
Organization, Tokyo. “
Macedo J. and R. Ruiz Usano (1991), Designing prototypes of JIT systems using
system dynamics principles. In K. Saeed, System Dynamics 91, p. 319-328. Asian
Institute of Technology, Bangkok, Thailand.
Macedo J. (1992), A tool for computer aided design of company wide quality
control (forthcoming). In F. Vernadat, Proceedings of the 8th International
Conference on CAD/CAM, Robotics and factories of the future, Metz, France.
Richardson G. P., A. Pugh (1981), Introduction to system dynamics modeling using
DYNAMO, MIT Press, Cambridge, MA.
Ozeki K., T. Asaka (1990), Handbook of quality tools: The japanese approach,
Productivity Press, Cambridge, MA.
= 394 -
—
CC BY-NC-SA 4.0