Documente Academic
Documente Profesional
Documente Cultură
IMEC2004-FM201-CP
Abstract
Many researchers have tried to interpret control charts to build a good Decision
Support System (DSS) for manufacturing systems. This concept is concerned with
how computer technology can be utilized in the quality control area. Researchers are
trying to build a new philosophy of quality control based on the computer technology
that is available in many systems of manufacturing such as FMS. Many techniques
were suggested to build a good DSS, such as: Neural Networks, Expert System (ES),
Computer Vision, and Feature-Based Method No computer implementations have
been done in the previous works until now. Also the commercial packages did not
give the industries a good solution for the quality control decisions. The main
objective of the present research is to characterize how the system will identify the
causes of the out-of-control situation to take the reasonable corrective action to adjust
the process. The other objective is how to reach a decision about process status i.e.
“in- or out-of-control”. In this paper, a proposed Expert Diagnostic Quality Control
System (EDQCS) for diagnosing problems in flexible manufacturing system will be
presented. This system will help quality engineers to identify the possible causes of
out-of-control situations. This system depends on the interpretation of the control
charts (X-bar and R-charts) using AT&T rules and a pattern recognition technique. A
complete case study is also presented.
Key words
On-line Quality control, Control Charts, Expert Diagnostic System, Variable Control
Charts, Flexible Manufacturing Systems
The proposed Expert Diagnostic Quality Control System (EDQCS) will help the
quality engineer detect bad quality when it happens on time, identify the sources of
error, and take appropriate corrective action. The proposed system allows efficient
process control; it consists of a systematic procedure, and has a friendly user-interface
during analysis.
Evans and Lindsay (1988) had proposed a framework for developing ES for statistical
process control applications. The knowledge can be constructed into three sets,
domain independent which have analysis rules, interpretive rules that analyze the
patterns in the control charts in terms of process changes, and domain-dependent
diagnostic rules which assist in determining assignable causes and corrective actions.
Pham and Ozteme (1992) described an online ES for statistical process control called
XPC. The XPC produces mean and range charts and interprets them automatically
from process capability point of view. The system is based on Leonard’s ES shell.
This ES is designed to employ process data and information regarding assignable
Vinod and Joseph (1992) developed a coupled ES to select a proper control chart,
optimize the design of the control chart and analyze the chart to identify an out-of-
control situation. A coupled ES was developed to automate the various steps used in
building control charts. The authors developed the knowledge base for the selection of
an appropriate control chart and to identify generic reasons for an out-of-control
process.
Jury, et al, (1996) described a statistical process control system that can provide
operators with the probable causes of process problems before they occur. This
system has reduced downtime caused by out-of-control situations. Product waste has
declined because projection of product quality has allowed operators to make
adjustment before non-conforming parts were produced. New information is always
being received from online sensors that measure quality characteristics of the part.
Statistics are computed continuously to predict upcoming process conditions
Chuen and Norma (1992) concentrated on the important aspects of the ES design and
knowledge organization. A new similarity measure, useful in a clustering approach to
the knowledge organization is proposed. ES was constructed with the capability of
automatically identifying the out-of-control situations revealing the source of
variation, and suggesting further actions. This system does not take into account the
AT & T rules. It depends only on the pattern definition by defining unnatural pattern
(signal out-of-control), and then observing irregularities in the process to define the
cause.
Walburn and Powner (1995) presented details of the format of the automatically
gathered production data and how to identify the variables that are derived from the
production data in order to extract knowledge in the form of production rules. The
analysis presented concentrates on the number of parts produced in the production
process and the relationship of the number of parts produced to a number of other
factors derived from the production machine data. Knowledge is extracted from
production machine data by processing the data and inducing rules on the process
data. They concluded that the induced rules provide extracted knowledge, which
facilitates understanding of the production process.
There are many ways to represent the knowledge for ES. James (1991) discussed
different modes to represent the knowledge. These modes are, Object-Attribute-Value
Triplets (OAV), OAV Network, Semantic Networks, Frames, Propositional Logic,
Predicate Logic, and Neural Networks.
EDQCS Assumptions
The proposed system is developed as a decision support system. The analytical tool
part in this system is concerned with the analysis of the control charts for variables
(X-bar and R charts). The analysis of these charts is done in two steps. The first one is
to apply AT&T rules on the control chart. This test is done after the analytical tool
(computer program) computes the characteristics of the charts and plots them.
AT&T rules are generally used to increase the sensitivity of the control charts to small
process shifts. For example, if a point exceeded a control limit, we would immediately
begin to search for the assignable causes, but if two out of three consecutive points
exceeded only the 2-sigma warning limits, we might increase the sample frequency, if
the process was really out of control it would give us a high probability of detecting
this situation more quickly. It is also increases the overall false alarm rate
Montgomery (1992). Each of these rules was tested and valid based on the normality
assumption of the control charts. The following AT&T rules are applied.
1. One or more points outside the control limits.
2. Two out of three consecutive points plot beyond the two-sigma warning
limits.
3. Four out of five consecutive points plot at a distance of one-sigma or beyond
the centerline.
4. A run of eight consecutive points on one side of the centerline
If any of these states are identified, we conclude that the process is out-of-control. To
identify the causes of this situation, a second test is applied. This test is concerned
with the interpretation of the control chart pattern. It is done even if none of the
AT&T rules is identified. An out-of-control situation could also result from an
instability pattern even if all points of the chart lie within the control limits. This
The second part in this decision support system is the ES part. The production rule
method is used to build the rules of the ES. The inference engine uses the forward-
chaining algorithm. There are two different philosophies of the forward-chaining. In
one, all known data is delivered to the ES at the very onset of the consultation session.
In the second approach, only the specific data element that is requested during
consultation is delivered to the ES. The second approach is used in this work. All the
knowledge is collected from statistical handbooks and textbooks that are related to the
quality control area. Each pattern has a group of assignable causes. Any one of these
assignable causes may cause the pattern. When these causes are represented in the ES
rules, equal chance probability is assumed for all causes for each pattern. This
probability can be changed when the decision support system is applied on the
flexible manufacturing system (FMS).
EXSys Professional shell is used to build the ES part. Then, it is linked with the
analytical part. All tests are done on the R chart first to detect the process variability,
and then on the X-bar chart to detect the process mean.
System Framework
The proposed system consists of two major parts. The first part analyzes the quality
control chart. Users enter the data of the control chart (sample size and sample
number and observations) or the data can be entered to this part using a data file.
After that, the program computes the quality control chart characteristics and plots
these charts (X-bar and R chart). A computer program applies AT&T rules to check
whether the process is in or out of control. The next step is a check on the instability
of the control charts. The results of this analysis are transferred to the second major
part in this work (Knowledgebase). The ES will ask the user some questions about the
situation of the process. The process status and the causes that affected it will be fired.
Figure 1 shows the whole framework of the Expert Quality Control System.
In this work, visual basic was used to build the computer program (analytical part)
and EXSys shell 4.1 under windows was used to build the knowledge base.
The knowledge base for the proposed system is represented based on the production
rules method. Also, forward-chaining was used for the inference engine. EXSys shell
is the ES shell used to report the knowledge base for this system. The knowledge base
receives the data from the computer program (out-of-control situation and pattern type
if there is a pattern). After that the consultation session is started to fire the final
conclusion about the process.
The interpretation of the control charts is the vital part in this work (analytical part).
The procedure is done in two stages. At the first stage, AT&T rules are applied to
check whether the process is in or out of control. The second stage is pattern
recognition to check for process instability. Based on this procedure the process status
will be defined and the possible (assignable) causes will be fired from the knowledge
base (ES part).
In this work OAV triplet are used to build a knowledge base. This mode is suitable for
production rules. The following procedure is applied to build the knowledge base:
1. Collect data. All assignable causes related to instability patterns are collected.
2. Construct an attribute value table. The collected data is developed and listed in an
attribute value table.
3. Develop a rule, in this step a rule base is built using production rule method based
on the attribute value table. Production rule is constructed as follows:
IF Premise (Attribute = specific value)
Then conclusion (Attribute = specific value)
4. Computer implementation. An expert system shell EXYS Professional (EP) was
used to write the rules. It is an easy to apply rule-based ES and an excellent tool to
develop expert systems. It features backward and forward chaining, black
boarding, fuzzy logic, and frames. Also, it links with database, and spreadsheet
programs, it is a Windows platform application. Further more, MS-DOSTM,
Macintosh TM, UNIX TM, and VAXTM platforms are available. It can link with many
Under MS-Windows TM applications.
5. Verification. At the end, validation has to be done to ensure that the ES has
reached a correct decision.
Status
Pattern
QCS
Inference
Characteristics
Engine Computations
Input Data
User
Interface
The computer implementation of the EDQCS was implemented. Three case studies
were tested to verify and validate the EDQS by Youssef (2002). In the present paper,
one case study will be demonstrated.
This case study was taken from the Western Electric Company (1985). This case
consists of 100 measures in 20 samples, each sample has 5 measures. Table 1 shows
the historical data, whereas Table 2 shows X-bar and R Charts results and Figure 2
shows these results.
In this section a demonstration of the features of the program will be presented based
on the case study Figure 3 shows the first menu that appears at the start of the
program. Figures 4 and 5 show the control charts and the characteristic data
corresponding to them.
Figure 6 demonstrates the analytical part of the program. The results of the analysis
show that the process is out of control due to AT&T rules one and two as shown by
the resulting X-bar chart. However the R and X-bar charts are natural patterns. On the
other hand, the variability chart (R Chart) was in control.
Clicking on the cause and corrective action button, the Knowledge base will be
activated and a consultation session will be started as shown in Figure 7. At the end of
this consultation the possible causes of the instability pattern will be fired if the
instability pattern is present in the control charts. The inferred fact is stated as
follows:
• AT&T rules were present
• No Pattern present for both controls charts (X-bar and R).
The results of the consultation are:
• Process is out-of-control with certainty factor 100%.
• Causes can not be defined
The results indicate that rule 34 was fired. Figure 8 shows the result of consultation
session (last menu of the program). In this case, the decisions mean that the quality
control engineer has to refer to the out-of-control periods and investigate what happen
at these points. An explanation feature is added to the program. Figure 9 shows an
example of the reason that rule 34 was fired to obtain these results.
On the other hand, in order to verify whether the decision about the pattern was
correct or not, a normality test is applied graphically. The results of the normality test
indicate that, the control charts (X-bar and R-charts) are normal. Figures 10 and 11
show the graphical results of the normality test on both X-bar and R charts.
2.5
2
1.5
1
0.5
Z(i)
0
-0.5 0 0.5 1 1.5 2 2.5 3
-1
-1.5
-2
-2.5
R(i)
2.5
2
1.5
1
0.5
Z(i)
0
-0.5 0 5 10 15
-1
-1.5
-2
-2.5
X(i)
In this section, the most significant results will be discussed. Namely, we are
comparing the control chart characteristics done using the proposed system with those
that are obtained using the Traditional methods. Table 3 represents the summarizing
results of the demonstrated case study.
Table 3 Results Comparison of Results by Proposed and Traditional Methods
Item R Chart X-bar Chart
Center line No Difference No Difference
Upper Control Limits No Difference No Difference
Lower Control Limits No Difference No Difference
From the previous table, there are no significant difference between the results of the
traditional method and the output of the proposed system. Moreover, the proposed
system interprets all instability patterns and gives satisfactory solution. The main
contribution of this study is that it gives a comprehensive analysis of the quality
control issues and can be summarized as follow:
• A new framework of expert system diagnosis suggested and built to take into
account the AT&T rules plus instability tests to get the accurate process control
decisions and define the possible causes of the out-of-control situation.
• Fourteen patterns are classified into nine classes, and the proposed system can
be recognized.
• The causes of the instability patterns were collected and represented in a
knowledge base based on the production rule.
• Until now no commercial software includes the identification of control chart
patterns. Wise (2000) state that, the last version of the InfinityQS management
depends on normality test and process deviation in its decision. It does not give
clarification or reasons about its decision, does not give a complete
interpretation of the control chart patterns and does not have a supported
knowledge base as per the proposed system
The results of the program matched those of the handbook with no significant
differences. In addition, the program interprets the pattern of the control charts and
the results are sent to the expert system to diagnose the process. The inferred facts of
this case were the process is out of control due to AT&T rules, and no patterns in both
of the control charts. Normality test was used to verify the decisions. The test
indicated that the control charts were normal. That means the pattern decision was
correct. Also, after investigation of the control charts, the AT&T rules were applied
correctly. Further more, these decisions were matched with the decisions that were
documented in the handbook, and no difference was noticed.
In conclusion, it can be seen that, the proposed system makes an on-line quality
control system and the variable sampling plan more effective in the automated manufacturing
systems. An integrated statistical quality control package with expert system was
produced. This hybrid DSS was applied and verified. Also this package can be
When the proposed system is applied, the chance of the human error will decrease
and quality improvement will increase. Also, the cost of production will decrease by
reducing the percentage of nonconforming units.
References