Quality Engineering

ISSN: 0898-2112 (Print) 1532-4222 (Online) Journal homepage: https://www.tandfonline.com/loi/lqen20

Statistical Process Control for Monitoring

Nonlinear Profiles: A Six Sigma Project on Curing
Process

Shing I. Chang , Tzong-Ru Tsai , Dennis K. J. Lin , Shih-Hsiung Chou & Yu-Siang
Lin

To cite this article: Shing I. Chang , Tzong-Ru Tsai , Dennis K. J. Lin , Shih-Hsiung Chou & Yu-
Siang Lin (2012) Statistical Process Control for Monitoring Nonlinear Profiles: A Six Sigma Project
on Curing Process, Quality Engineering, 24:2, 251-263, DOI: 10.1080/08982112.2012.641149

To link to this article: https://doi.org/10.1080/08982112.2012.641149

Published online: 26 Mar 2012.

Submit your article to this journal

Article views: 613

View related articles

Citing articles: 16 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=lqen20
Quality Engineering, 24:251–263, 2012
Copyright # Taylor & Francis Group, LLC
ISSN: 0898-2112 print=1532-4222 online
DOI: 10.1080/08982112.2012.641149

Statistical Process Control for Monitoring

Nonlinear Profiles: A Six Sigma Project
on Curing Process
Shing I. Chang1,
Tzong-Ru Tsai2, ABSTRACT Curing duration and target temperature are the most critical
Dennis K. J. Lin3, process parameters for high-pressure hose products. The air temperature
Shih-Hsiung Chou1, collected in the curing chamber is represented in the form of a profile. A
Yu-Siang Lin4
1
proper statistical process control (SPC) implementation needs to consider
Quality Engineering Laboratory,
both numeric as well as profile quality characteristics. This article describes
Department of Industrial and
a successful Six Sigma project in the context of statistical engineering for
Manufacturing Systems
Engineering, Kansas State integrating SPC, a statistical method, to the existing practice of engineering
University, Manhattan, Kansas process control (EPC) according to science. A case study on a real pro-
2
Department of Statistics, duction curing process is thoroughly investigated. It is shown that the
Tamkang University, Tamsui new findings could potentially result in significant energy savings. The solu-
District, New Taipei City, Taiwan tions provided in this study can be generalized into other curing processes
3
Department of Statistics,
and applications subjected to both EPC and SPC.
Pennsylvania State University,
University Park, Pennsylvania
4
Department of Industrial KEYWORDS autoclave, biplot chart, DMAIC, Hotelling T 2 chart, profile analysis
Management, National Taiwan
University of Science and
Technology, Taipei, Taiwan

INTRODUCTION AND MOTIVATION

Hoerl and Snee (2010) proposed the term statistical engineering defined
as ‘‘the study of how to best use statistical concepts, methods and tools, and
integrate them with information technology and other relevant sciences to
generate improved results’’ (p. 52). In this article we discuss a Six Sigma
project to implement statistical process control (SPC) on high-pressure hose
products produced in a factory, PH Corporation, that is undergoing a lean
Six Sigma transformation. We demonstrate how the proposed SPC solution
is integrated into the current engineering process control operation and its
quality assurance system in the statistical engineering framework. Through
a case study using real-life data, we discuss its potential impact toward
continuous improvement.
Address correspondence to Shing I. High-pressure hose products are made out of alternate layers of rubber
Chang, Quality Engineering
Laboratory, Department of Industrial and metal wires. Toward the end of the production process, various
and Manufacturing Systems high-pressure hose reels are loaded and cured. A curing process typically
Engineering, Kansas State University,
Manhattan, KS 66506-5101, USA. consists of a sealed, heated chamber, called an autoclave or vulcanizer.
E-mail: [email protected] Most autoclaves or vulcanizers are equipped with multiple thermocouples
251
 many pending projects throughout its factory. One of
the process improvement projects that the quality
manager and process engineers of PH Corporation
face is implementation of statistical process control
on the curing process because all high-pressure hose
products go through this critical process before the
final quality check is applied. It is very important to
study the variability of the curing process and then
to maintain its stability. The proposed SPC project
would help ensure that the curing production process
is under both engineering control and statistical con-
trol. When the final quality check identifies potential
quality issues, process and quality engineers can rule
FIGURE 1 Air temperature profile of a typical curing cycle. out the curing process as a cause if a proper SPC
procedure is in place. This study demonstrates how
inside their chambers and=or parts. The housing that statistical thinking and statistical engineering can be
contains an autoclave or vulcanizer is often called the integrated into an existing engineering control
vessel. The most important information gathered from application.
the thermocouples is the air temperature, represented Process engineers often presume that the curing
as a nonlinear profile during the curing cycle. A pro- process must be under control because it is governed
grammable logical controller (PLC) is used to control led by a PLC according to a recipe used for many
temperature based on sensor readings or a fixed time years. This argument comes from engineering pro-
interval according to a curing recipe. Figure 1 shows a cess control (EPC) instead of SPC. The practice of
typical air temperature profile divided into three EPC may keep a malfunctioning process operating
sections—heat-up stage, curing stage, and cool-down for a while before it eventually spirals out of control
stage. It is important for the flat section (the curing when a root cause starts to emerge. An SPC imple-
stage) of this profile to hold for a fixed period of time. mentation combined with an EPC practice can pre-
The thermocouples are calibrated from time to time vent this from happening (see Del Castillo 2002;
so that the measurement error is negligible. The data Montgomery et al. 1994). The cost of scrapping
collection unit of the PLC is capable of collecting an entire load of cured product is prohibitively
multiple temperature readings per second. high. Therefore, this is one of the high-impact
A closer examination of the data set reveals that the projects during the initial stage of Six Sigma
temperature readings do not change very swiftly. The transformation.
data presented in Figure 1 are based on temperature The key question concerning process monitoring
readings taken every few seconds. The variables X1 to is ‘‘Is a curing process in control given the air tem-
X6 are defined in a later section. Due to the confiden- perature profile generated at the end of a curing
tiality agreement with PH Corporation, we altered the cycle?’’ One of our objectives involves the design
numerical values of the temperature and time units. and implementation of a SPC plan on this process
The target air temperature during the curing stage is to make sure that the process is in control in all
500
F and should last for 480 time units. aspects. A process is assumed to exhibit statistical
control when only common cause variability exists
in process parameters of the system. If a curing cycle
PROBLEMS AND CHALLENGES is in control, the quality of products being cured
Under the framework of Six Sigma, quality and would be satisfactory given that the raw material
process improvement is most effective on a project- meets its intended specifications. On the other hand,
by-project basis. One of the main emphases in statisti- if a process is out of control, it is very likely that
cal engineering addresses the need to identify projects products inside the vessel will not be cured properly.
with high impact. PH Corporation is in its initial stage The major benefit of an in-control process is product
of experiencing a Six Sigma transformation. There are consistency that translates into better product quality

S. I. Chang et al. 252

 process. The current practice involves one curing
recipe applied to all high-pressured host products
regardless of their individual specifications. Each type
of hose product requires a minimal period of curing
time at a set temperature for the major functional
requirement; that is, the proper cross-linking of the
rubber material. A hose product without a proper
curing will fail either in the pressure test or in the
field more quickly than expected. In addition, a
gradual cooling procedure is designed to maintain
the cosmetic requirement on the outside of a hose
product; that is, a smooth look without any blis-
tering. It is critical that both aspects are met during
FIGURE 2 A typical air temperature profile during the
cool-down stage. the curing production for customer satisfaction. The
air temperature profiles are recorded to reflect
the production quality from batch to batch in terms
and thus customer satisfaction. Reduced waste is also
of the most critical process parameter—the air
an important benefit.
temperature.
The current process monitoring practice involves
After consultation with the quality manager,
manual visual inspection of a plot similar to Figure 1
process engineers, and information technology (IT)
at the end of a curing process. All temperature read-
specialist, we define critical-to-quality characteristics
ings are recorded in a database, but a hard-copy
(CTQ), often defined in the define–measure–
printout is generated for quality assurance purposes.
analyze–improve–control (DMAIC) process, as
The operator attending the vessel would examine this
follows:
plot and decide whether the process is in control or
not. If the curing is believed to be successful, a ‘‘pass’’
. QC1: the time to reach the target air temperature
would be written on the printout. The current curing
500
F must be swift although the PLC does not
recipe calls for the following sequence: (1) open the
specify a time
heat steam valve rapidly to heat the chamber air tem-
. QC2 and QC3: the air temperature must maintain a
perature to 500
F, (2) maintain this target temperature
target value of 500
F (QC2) for a fixed target
for 480 time units, (3) open a valve to vent off the
duration 480 time units (QC3)
pressure for a fixed period, (4) spray water for a fixed
. QC4: the air temperature must be reduced
period, (5) rest for a fixed period, (6) spray water a
‘‘gradually’’ according to sequences 3 to 7 defined
second time for a fixed period, and then (7) open
earlier
the chamber door to unload the cured products.
The impact on temperature for sequence of actions
After a curing cycle, an air temperature profile similar
from steps 3 to 7 is summarized in Figure 2. Because
to Figure 1 is given. Based on the profile, quality engi-
the timing and duration for each stage is also
neers determine whether the CTQs—that is, QC1 to
controlled by a PLC, operators would glance at the
QC4—are in control. A Phase I data set of a curing
general pattern for quality assurance. The current
process was collected over a one-month period. Most
practice is very subjective and operator dependent.
quality engineers are equipped to implement SPC for
There is a great opportunity to introduce quality
QC1 defined above. However, examining Figure 3
monitoring for continuous quality improvement.
during the curing stage, quality engineers were
puzzled by multiple types of patterns among 153
DEFINE CRITICAL-TO-QUALITY profiles related to QC2 and QC3. It is not clear which
existing control charting methods could be directly
CHARACTERISTICS applied for process monitoring purposes.
In this study, we focus on the air temperature Another challenge is to implement SPC for QC4 in
setting considered to be the most critical to a curing which numerical standards are ill defined. Quality

253 SPC for Monitoring Nonlinear Profiles

 OVERVIEW OF A SYSTEM
FRAMEWORK FOR DATA PROCESSING
In the framework of statistical engineering, it is
important to integrate statistical methods and tools
into IT practices. Figure 5 shows a proposed system
framework related to data processing before SPC can
be implemented on the curing process. This frame-
work has four components; that is, middleware, data
warehouse, SPC, and decision making. The process
data are first retrieved from various databases corre-
sponding to the vent valve, vulcanizer, and water
FIGURE 3 Patterns of the air temperature profiles during the spray valve. Next, the middleware component pro-
curing stage. vides a data preprocessing function; that is, data
loading, data cleaning, transforming, integrating,
and refreshing on the raw data. The data warehouse
engineers were overwhelmed by multiple segments is a repository that stores, arranges, and organizes
and the large amount of data when they examined information for SPC implementation or any other
Figure 4 containing multiple temperature profile analysis purposes. Furthermore, the SPC component
segments during the cool-down stage. provides a basic process monitoring for decision

FIGURE 4 Patterns of the air temperature profiles during the cool-down stage: (a) the duration of vent open (X3), (b) the duration of
spray 1 (X4), (c) the duration of rest 1 (X5), and (d) the duration of spray 2 (X6).

S. I. Chang et al. 254

FIGURE 5 Diagram of system framework for data processing.

making. Currently, the front-end databases are fed by they are not exactly random variables in the usual
the PLC. Our tasks involve the development of the sense because a PLC is used to make sure that each
middleware, SPC, and decision-making components cycle achieves the targets for QC2 and QC3; that is,
that will be integrated into the existing quality assur- 500
F for 480 time units. When the air temperature
ance database. reading gets close to 500
F, the PLC starts the timer
PH Corporation, like many other manufacturers, is for 480 time units to achieve the required fixed dur-
experiencing a modern-day factory phenomenon; ation. Based on the data recording tag shown in the
that is, a data tsunami. Lots of data are generated by database, the PLC would start the curing timer when
various operations, but their use in decision making a temperature reading first reaches 492
F. Each air
is ignored. The PLC that controls the curing process temperature profile exhibits a different pattern initi-
is capable of generating a large volume of data, most ally as shown in Figure 3. The temperature would
of which are simply stored, archived, and eventually eventually reach the steady state of 500
F. Therefore,
deleted. None or very little of it is transformed into the true quality characteristic for the curing time
useful information or statistics on which insightful requirement is how much time each profile stays at
decision makings can be made. There is really a miss- or above 500
F.
ing link between the statistical thinking=methods that We define X2 as the cumulative time that tem-
suggested collecting these data and engineering perature readings exceeding the target temperature
practice. In this study, we propose to integrate IT (500-d), where d is the standard deviation of the
and the proposed SPC procedure. steady-state temperature readings. Note that d is small
because of the PLC control actions. Based on the
steady-state segment of the temperature profiles
ESTABLISH MEASURE OF
shown in Figure 3, d is estimated to be 0.05
F. This
CRITICAL-TO-QUALITY quality characteristic can be obtained by an algorithm
CHARACTERISTICS written in MATLAB to process any profile during the
curing stage. It captures the key characteristics impor-
This measure step is the second step of the DMAIC
tant for the curing recipe; that is, QC2 and QC3.
process. Based on the CTQs QC1 to QC4, we would
like to establish statistics that can directly measure
the CTQs and be fed into control charts. It is straight- A SOLUTION TO PROFILE ANALYSIS
forward to quantify QC1 as X1, which is the duration
between the start of a curing stage and the start of the
DURING THE COOL-DOWN STAGE
countdown timer of QC3. This statistic is relatively We propose the use of profile monitoring techni-
easy to obtain via an algorithm written in MATLAB ques for tackling the SPC implementation issues for
(Manhattan, KS). QC4. Profile monitoring techniques have gained
It is a challenge to derive statistics based on QC2 much attention since Jin and Shi (1999) introduced
and QC3. A closer look at QC2 and QC3 reveals that their application to the force of a stamping process.

255 SPC for Monitoring Nonlinear Profiles

Three general approaches often used to implement each segment for the cool-down steps 3–6, respect-
SPC on profiles as quality characteristics include ively. These statistics are chosen to reflect the general
the use of process parameters, the use of projected trend of the cool-down process. Each statistic ident-
space, and the use of critical spatial information. ifies the proper behavior within a functional segment
Most literature on this topic follows the first category so that it is easy for operators to diagnose problem-
because the process parameters are often indepen- atic profiles. As shown in Figure 2, X3 to X6 are
dent of each other, which is one of the conditions derived from air temperature readings yij where the
required to be used for simultaneous charting of index i is defined according to the cool-down recipe
univariate control charts (see Kang and Albin 2000; and j ¼ 3; 4; 5; 6. Specifically, the index i of yij takes
Kim et al. 2003; Noorossana, Eyvazian, Amiria, and the integer values in [1, 60], [61, 88], [89, 116], and
Mahmoud 2010; Noorossanna, Eyvazian, and [117, 140], respectively depending, on j ¼ 3, 4, 5, or 6.
Vaghefi 2010). The second category is the use of pro- Process engineers do not know the shape of the
jected space in which original profiles are trans- nominal profile of the air temperature during the
formed from one domain into another domain with cool-down cycle. We need to estimate it from a
axes orthogonal to each other. The principal compo- Phase I data set. Similar to the practice reported by
nent analysis (Jones and Rice 1992) is often used for Chang and Yadama (2010), X3 to X6 can be obtained
this purpose. Other work in this category includes from the following function:
Vapnik (1998), Walker and Wright (2002), Woodall
et al. (2004), and Moguerza et al. (2007). Finally, Pnj





i¼1
yij  y
j

the third category is the use of spatial information. Xj ¼ j ¼ 3; 4; 5; 6 ½1
The core idea is to reduce the number of observa- nj
tions in the original profile but still maintain crucial
profile information. Those methods can be found where yij is the ith data point of the jth profile seg-
in Grossmann and Morlet (1984), Mallat (1989), ment, nj is the number of points within segment j,
Strang and Nguyen (1997), Jin and Shi (1999), and yj is the jth segment of the ‘‘average’’ profile
Chicken et al. (2009), and Chang and Yadama (2010). fitted over all profiles in a Phase I data set. Judging
The main activities during the cool-down stage from Figures 2 and 4, linear functions should be suf-
have to do with the proper temperature drops in four ficient to model yj for j ¼ 4, 5 and 6 but perhaps not
segments. Depending on the load—that is, heat mass for j ¼ 3. In this study, we choose to use a B-spline
in the vessel—these temperature profiles exhibit dif- function to fit each profile segment because it is
ferent characteristics but hold the general patterns as capable of fitting either linear or nonlinear
shown in Figure 4. The geometric shapes from vari- profiles. For details of B-spline fitting, please refer
ous cycles deviate from each other greatly in each to Mortenson (2006).
segment. The steepest temperature drops usually
take place during the segments of vent opening ANALYZE CURING PROCESS VIA
and spraying. This pattern will break if the vent does
not open or the water tank runs out of water during a
CONTROL CHARTING
spraying cycle. Instead of using one model for these The third step of the DMAIC process involves the
complicated profiles at this stage, it would be much use of a Phase I study of control charts to understand
easier to break this stage into smaller segments causes of variation and identify potential root causes.
according to the PLC recipe. The current recipe con- One approach for implementing SPC combines X1 to
tains several fixed intervals—venting, water spray X6 into one multivariate control chart. This design
cycle 1, resting, and water spray cycle 2. The last seg- allows the monitoring of the curing process in a
ment (not labeled) in Figure 2 is not important single chart instead of spreading the task into six
because a load may sit in the vessel for a period of individual control charts. However, this approach
time before it is unloaded. The length of this period will delay fault diagnosis when a process is out of
depends on factory schedules. control. It would need the involvement of a quality
We propose statistics X3, X4, X5, and X6 as the engineer to examine the individual data of X1 to X6
average of deviations from the nominal profiles for to identify which variables are responsible for an

S. I. Chang et al. 256

out-of-control signal. On the other hand, the interest, which are related to the average deviations
approach of using six individual control charts simul- from their nominal nonlinear profiles. Using conven-
taneously would increase the overall type I error of tional SPC procedure, we first apply two Hotelling’s
control charting. Moreover, the correlation informa- T2 control charts to explore the first segment and
tion between quality characteristics would be second segment on all profiles, respectively.
ignored so that certain out-of-control causes would Although the quality engineers claimed that the 153
not be detected under this approach. profiles in Phase I data set are all in control, Hotell-
From examining QC1 to QC4 and various stages of ing’s T2 control charts may show different results.
curing, we propose a solution to separate variables X1 In order to establish the control charts for Phase II
to X6 into two groups. The first group consisting of X1 monitoring, it is necessary to screen out abnormal
and X2 is used for monitoring the heat-up and curing observations. All plotted points should fall within
stages because both pieces of information are main- the control limits. Otherwise, we should seek for
tained by the heating valve operation. These two vari- explanations and justifications for any out-of-control
ables are highly related because control actions on the point to be included. In order to maintain a com-
heat valve dictate them. The second group consisting bined type I error of 0.005, we establish the control
of the rest of the variables is mainly used for the limit for each T2 control chart accordingly. Results
cool-down stage. Two multivariate control charts for of Phase I studies for both T2 control charts are dis-
individual observations are proposed for these two cussed in the following sections.
groups, respectively. In this case, each observation cor-
responds to a curing cycle. In other words, we would STUDY DURING THE HEAT-UP AND
treat each cycle as one sample with one set of X1 to X6
values that are monitored via two control charts.
CURING STAGE
Possible multivariate control charts for individual We first plot all standardized observations on the
observations include Hotelling’s T2 charts (Hotelling Hotelling T2 control chart with individual observa-
1947) and MEWMA (Multivariate Exponentially tions (see Vargas 2003; Williams et al. 2006). As
Weighted Moving Average) charts (see Lowry and shown in Figure 6, there are 11 points outside the
Montgomery 1995; Tracy et al. 1992). A control chart control limit. We further examine the causes that
for individual observations is used because each cur- make these points ‘‘abnormal’’ before any point is
ing cycle only provides one air temperature profile. discarded from the Phase I data set. Figure 7 shows
Note that the underlying process is controlled by a a biplot that plots all points in a two-dimensional
PLC. Within each cycle it is possible that temperature plot between X1, the heat-up duration, and X2, the
observations may exhibit autocorrelation. However, curing duration above the target 500
F. We draw
points on a control chart are plotted from cycle to the control limits of the IX chart (i.e., X-bar chart
cycle. Autocorrelation, if it exists, may diminish. on individual observations) for X1 as well as the
lower specification limit 360 time units for X2 on
A CASE STUDY—THE CURING Figure 7. Note that this lower specification is more
important than the control limits in that rubber pro-
PROCESS REVISITED ducts require a minimal time at a set temperature
In this study, 153 air temperature profiles (or for proper curing. In this case, the target curing time
cycles) of Phase I data are examined to seek oppor- is set at 480 time units so that most curing cycles
tunities for improvement. Quality and process engi- would have at least 360 time units above the cured
neers deemed that they are in control. Each profile temperature of 500
F. The group of out-of-control
is divided into two segments. The first segment aims points including 23, 77, and 126 indicates that these
for the monitoring of heat-up and curing duration, three cycles take a long time to reach the target cur-
and the second segment is for the monitoring of ing temperature but their curing duration above
the cool-down stage. In the first segment, there are 500
F are above the mean curing duration. On the
two quality characteristics of interest. Both have the other hand, the group of out-of-control cycles
same unit; that is, time duration. The second segment including 1, 9, and 98 shows that the time to reach
of a cycle contains four quality characteristics of the target temperature is shorter than that of most

257 SPC for Monitoring Nonlinear Profiles

FIGURE 6 Hotelling’s T2 control chart for X1 and X2 in the heat-up and curing stages.

cycles. Specifically, cycles 1 and 98 are below the control limit of the Hotelling T2 control chart changes
mean curing duration but above the minimal 360 from 11.594 to 11.565 (based on Tracy et al. 1992).
time units. Another cycle 54 also belongs to this
group. Although this point falls within the IX chart STUDY DURING THE
control limits and is above 360 time units, it is far
from the mean point of the ellipsoid, causing it to
COOL-DOWN STAGE
be plotted outside the control limit of the T2 chart. The Hotelling’s T2 control chart for the cool-down
We keep the cycles of these two groups when estab- stage has four variables X3 to X6, which are the aver-
lishing the control limits of the T2 chart. age deviation statistics from the B-spline fitted nom-
The group of out-of-control points that cause con- inal profile. This is displayed in Figure 8. Cycles 41,
cerns consists of cycles 60, 73, 131, and 137. 136, and 141 are above the control limit of 15.764.
Although they reach the target temperature like the Figure 9 shows detailed plots of each segment with
majority of the other cycles, they fail to keep the these three profiles highlighted. In various segments,
minimal of 360 time units of curing time above the shapes of these cycles are very different from the
500
F. This result is forwarded to process engineers rest of the profiles. In addition, we can also use a
for more detailed study. After discarding the cycles matrix plot to show the correlation structure among
of this group from the Phase I data set, the new X3 to X6. As shown in Figure 10, the outliers corre-
spond to cycles 41, 136, and 141 in various places.
After removing these points, the control limit of the
Hotelling’s T2 control chart changes from 15.764 to
15.732.

SUMMARY OF THE SPC TOOLS

USED IN THIS STUDY
The paradigm of statistical engineering links stat-
istical methods and tools at the operational level to
statistical thinking at the strategic level. In this study,
various statistical methods integrated with IT tools
have been adopted to fulfill this goal. The main idea
of statistical thinking in this case is to integrate the
statistical process control of the curing process to
FIGURE 7 Biplot and IX control limits for heat-up time and the current practice, which is operated under the
curing time. engineering control. The implementation of SPC will

S. I. Chang et al. 258

FIGURE 8 Hotelling’s T2 control chart for cool-down variables X3 to X6.

prevent major process catastrophes from taking We now summarize how all of the statistical
place and provide useful information and statistics methods and tools are used in this study. The main
for informative decision making. process monitoring tools proposed are a pair of

FIGURE 9 Cool-down process profiles by (a) segment X3, (b) segment X4, (c) segment X5, and (d) segment X6 (out-of-control profiles 41,
136, 141).

259 SPC for Monitoring Nonlinear Profiles

FIGURE 10 Matrix plot of cool-down variables X3 to X6.

Hotelling T2 control charts. The studies described operators would use the control charts to judge
above show how the control limits for both charts whether the process is in control or not. When a
are established based on the Phase I data set. During point plots outside the control limits of either one
the process monitoring phase, an algorithm written of the T2 control charts, process engineers are then
in MATLAB is used to obtain the quality characteris- notified to diagnose the process.
tics X1 to X6 from a sample temperature profile. Then
variables X1 and X2 form an input vector to be used
in the first T2 control chart as shown in Figure 6 and
IMPROVE THE CURING PROCESS
the rest of the variables are put into the second T2 During the fourth step of the DMAIC process, we
control chart as shown in Figure 8. If the points identify and generate potential solutions for process
plotted on both T2 control charts fall below their improvement. Currently air temperature profiles are
respective control limits, the process is deemed in not used for any process-related analysis or diag-
control. Otherwise, a biplot shown in Figure 7 is nosis. The PH Corporation will benefit from decision
used on variables X1 and X2 for fault diagnoses when making based on objective data analysis as demon-
the first T2 control chart triggered the out-of-control strated in this case study. Furthermore, during this
signal. Furthermore, a matrix plot of cool-down vari- Phase I study, we have identified a process adjust-
ables X3 to X6 should be used when the second T2 ment opportunity that may potentially save produc-
control chart triggered the out-of-control signal. tion cost. The distribution of the curing duration
Operators are responsible for monitoring both T2 above the target temperature reveals that it is poss-
control charts at the end of each curing cycle, and ible to trim the target duration to a smaller value as
quality engineers should use either a biplot or a shown in Figure 7. The curing recipe specifies 480
matrix plot for fault diagnoses when an out-of- time units of curing to accommodate all types of pro-
control signal is triggered by either or both of the ducts so that the actual curing time above 500
F may
T2 control charts. During a routine curing operation, be at the minimum of 360 time units. Most X2 values

S. I. Chang et al. 260

in Figure 7 are actually above 380 time units. A gap the ongoing work flow as shown in Figure 11. A
of 20 time units between 360 and 380 can be simplified process flow shows that the semifinished
trimmed from the current setting of 480 time units, hose products are loaded into a vulcanizer after all
and most actual curing times would still meet the braiding processes are finished. Two major subcom-
minimal requirement. As energy costs rise, this sav- ponents in the curing process consist of the pro-
ings opportunity may dramatically improve the cost posed division of the heat-up=curing process and
of production while satisfying functional require- the cool-down process. Cured products would then
ments. As a numerical example, for every 200 go through quality assurance via acceptance sam-
batches of hose products cured, reducing the curing pling before they are shipped to customers.
time by 20 time units will yield an additional 20  The shaded box in the middle of Figure 11
200=(480  20) ¼ 8.7 batches without using addi- summarizes the proposed SPC implementation. In
tional energy. order to gain long-term success of the proposed
Historically, only one curing recipe is used regard- system, the proposed SPC solutions should be inte-
less of the type of hose product loaded. It is under- grated with the existing quality assurance system.
standable that the current curing time has to be set This is another example of integrating statistical
at the maximum level to satisfy the worst-case sce- tools into IT to maintain the gains achieved in
narios. However, with the X2 data analyzed we have the statistical engineering paradigm. As shown in
demonstrated that this one-size-fits-all practice may the dashed line in Figure 11, we recommend the
be revised to satisfy functional requirements and inclusion of the proposed outputs to be stored in
reduce energy consumption. the existing quality assurance database (QA DB).
The solid line connecting to the QA DB repre-
sents the existing IT system used for quality assur-
CONTROL THE CURING PROCESS BY ance while the dash line is the proposed one.
EMBEDDING THE PROPOSED SPC Depending on the results of on-line monitoring—
SOLUTION INTO THE CURRENT that is, Hotelling T2 charts or biplot—quality engi-
neers may be able to use this information to
QUALITY ASSURANCE PRACTICE decide proper sample sizes for destructive tests
The final step of the DMAIC process involves the or to pay more attention to batches with large var-
integration of the proposed SPC implementation into iations, for example.

FIGURE 11 Diagram of current process flow and the proposed embedded SPC flow. (Color figure available online.)

261 SPC for Monitoring Nonlinear Profiles

 CONCLUDING REMARKS of them into the process monitoring framework.
Future Six Sigma projects should be considered.
Following the DAMIC process commonly used in
a Six Sigma project and the paradigm of statistical
engineering, an SPC implementation solution is ACKNOWLEDGMENTS
introduced here for a production curing process We are grateful to the guest editors and two refer-
where there is a mixture of quality characteristics ees whose constructive comments were extremely
presented in the form of a temperature profile. Using helpful.
both computer algorithms (nonstatistical techniques)
and profile monitoring techniques (statistical techni-
ques), we propose a solution to combine critical
ABOUT THE AUTHORS
quality characteristics and profile monitoring into Dr. Shing I. Chang is an associate professor in the
one SPC framework. A Phase I data set is used to Department of Industrial and Manufacturing Systems
demonstrate how the proposed framework involving Engineering at Kansas State University. His main
the use of two Hotelling T2 charts is implemented. research interests include multivariate statistical pro-
This case study on curing production process is cess control for manufacturing and health care, non-
shown to be successful. The proposed SPC frame- linear profile monitoring, neural networks and fuzzy
work is recommended to be integrated into the exist- set applications in quality engineering, and multivari-
ing quality assurance process to ensure long-term ate experimental designs. He is a senior member of
success. Although the results of this study are based both Institute of Industrial Engineers and American
on the vulcanizer of the PH Corporation, it is straight- Associate of Quality. He served as department editor
forward to generalize them into other curing pro- of IIE Transactions in 2003 to 2009. He was a NASA
cesses. The mechanism and control used in various summer fellow in 2004 and a recipient of SME young
autoclaves are very similar. In addition, the proposed manufacturing engineer award in 1997.
SPC procedure can be more broadly applied to other Dr. Tzong-Ru Tsai is a professor in the Department
EPC=SPC applications. The key of a successful inte- of Statistics at Tamkang University. His main research
gration of EPC and SPC relies on the transformation interests include quality control and reliability analy-
from EPC variables into proper quality characteristics sis. He has served as an executive editor for the Inter-
for SPC. In this case study, the EPC variables are the national Journal of Intelligent Technologies and
curing recipe—500
F for 480 time units. We were Applied Statistics since 2008 and served in the edi-
able to convert it into variable X2. torial board of the Journal of the Chinese Institute
Based on the Phase I study, one important issue of Industrial Engineers from 2007 to 2011. Dr. Tsai
that we explored was the cost aspect. The opport- also served as a senior consultant for the Electronics
unity for improvement lies in the curing recipe. and Optoelectronics Research Lab. of the Industrial
The distribution of the curing duration above the tar- Technology Research Institute in 2010 to 2011.
get temperature reveals that it is possible to trim the Dr. Dennis K. J. Lin is a university distinguished
target duration to a smaller value (the current setting professor of supply chain and statistics at Penn State
is at 480 time units). As energy costs rise, it should be University. His research interests are quality assur-
noted and emphasized that there is a great opport- ance, industrial statistics, data mining, and response
unity for cost saving while satisfying the functional surface. He has published over 150 papers in a wide
requirements. variety of journals. He currently serves or has served
Although this study only focuses on SPC as associate editor for near 10 journals and was
implementation strategy, further investigations that coeditor for Applied Stochastic Models for Business
take the full advantage of the entire set of collected and Industry. Dr. Lin is an elected fellow of ASA
data may be considered. For example, there are and ASQ, an elected member of ISI, a lifetime mem-
other profiles—such as pressure and condensation ber of ICSA, and a fellow of RSS. He is an honorary
water temperature—generated from the same pro- chair professor for various universities, including a
cess that can potentially improve the stability of the Chang-Jiang Scholar of China at Renmin University,
curing process. It is a challenge to incorporate all National Chengchi University (Taiwan), Fudan

S. I. Chang et al. 262

University, and XiAn Statistical Institute (China). He Jin, J., Shi, J. (1999). Feature-preserving data compression of
stamping tonnage information using wavelets. Technometrics,
is also the recipient of the 2004 Faculty Scholar Medal 41(4):327–339.
Award at Penn State University and the Youden Jones, M. C., Rice, J. A. (1992). Displaying the important features of large
collections of similar curves. The American Statistician, 46(2):
Address speaker in 2010.
140–145.
Shih-Hsiung Chou is a Ph.D. candidate and a sys- Kang, L., Albin, S. L. (2000). On-line monitoring when the process
tem administrator in the Department of Industrial yields a linear profile. Journal of Quality Technology, 32(4):
418–426.
and Manufacturing Systems Engineering at Kansas Kim, K., Mahmoud, M. A., Woodall, W. H. (2003). On the monitoring of
State University. His main research interests include linear profiles. Journal of Quality Technology, 35(3):317–328.
Lowry, C. A., Montgomery, D. C. (1995). A review of multivariate control
multivariate statistical process control, high-
charts. IIE Transactions, 27(6):800–810.
dimensional data visualization, data mining, machine Mallat, S. (1989). A theory for multiresolution signal decomposition: The
learning, and artificial intelligence. wavelet representation. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 11(7):674–693.
Yu-Siang Lin is a Ph.D. candidate in the Depart- Moguerza, J. M., Munoz, A., Psarakis, S. (2007). Monitoring nonlinear
ment of Industrial Management of National Taiwan profiles using support vector machine. Lecture Notes in Computer
Science, 4756:574–583.
University of Science and Technology in Taiwan.
Montgomery, D. C., Keats, J. B., Runger, G. C., Messina, W. S. (1994).
His advisor is Dr. Kung-Jeng Wang. He has a double Integrating statistical process control and engineering process control.
major in industrial management and computer Journal of Quality Technology, 26(2):79–87.
Mortenson, M. E. (2006). Geometric Modeling. New York: Industrial
science. His current research interests are in the areas Press.
of supply chain management, warehouse configur- Noorossana, R., Eyvazian, M., Amiri, A., Mahmoud, M. A. (2010). Statisti-
cal monitoring of multivariate multiple linear regression profiles in
ation, production planning, inventory systems, and
phase I with calibration application. Quality and Reliability Engineering
gene algorithms. International, 26(3):291–303.
Noorossana, R., Eyvazian, M., Vaghefi, A. (2010). Phase II monitoring
of multivariate simple linear profiles. Computers and Industrial
REFERENCES Engineering, 58(4):563–570.
Strang, G., Nguyen, T. (1997). Wavelets and Filter Banks, Rev. ed.
Chang, S. I., Yadama, S. (2010). Statistical process control for monitoring Wellesley, MA: Wellesley-Cambridge Press.
non-linear profiles using wavelet filtering and B-spline approximation. Tracy, N. D., Young, J. C., Mason, R. L. (1992). Multivariate control charts
International Journal of Production Research, 48(4):1049–1068. for individual observations. Journal of Quality Technology,
Chicken, E., Pignatiello, J. J., Jr., Simpson, J. (2009). Statistical process 24(2):88–95.
monitoring of nonlinear profiles using wavelets. Journal of Quality Vapnik, V. N. (1998). Statistical Learning Theory. New York: Wiley.
Technology, 41(2):198–211. Vargas, N. J. A. (2003). Robust estimation in multivariate control charts
Del Castillo, E. (2002). Statistical Process Adjustment for Quality Control. for individual observations. Journal of Quality Technology, 35(4):
New York: Wiley. 367–376.
Grossmann, A., Morlet, J. (1984). Decompostion of Hardy functions into Walker, E., Wright, S. P. (2002). Comparing curves using additive models.
square integrable wavelets of constant shape. SIAM Journal on Journal of Quality Technology, 34(1):118–129.
Mathematical Analysis, 15(4):723–736. Williams, J. D., Woodall, W. H., Birch, J. B., Sullivan, J. H. (2006). Distri-
Hoerl, R. W., Snee, R. D. (2010). Closing the gap—Statistical engineering bution of Hotelling’s T2 statistic based on the successive difference
links statistical thinking, methods, tools. Quality Progress, May:52–53. estimator. Journal of Quality Technology, 38(3):217–229.
Hotelling, H. (1947). Multivariate quality control. In: Eisenhart, C., Hastay, Woodall, W. H., Spitzner, D. J., Montgomery, D. C., Gupta, S. (2004).
M. W., Wallis, W. A., Eds., pp. 111–184, Techniques of Statistical Using control charts to monitor process and product quality profiles.
Analysis. New York: McGraw-Hill. Journal of Quality Technology, 36(3):309–320.

263 SPC for Monitoring Nonlinear Profiles