Expert Systems with Applications 42 (2015) 2919–2927

Contents lists available at ScienceDirect

Expert Systems with Applications

journal homepage: www.elsevier.com/locate/eswa

A principal component analysis model-based predictive controller

for controlling part warpage in plastic injection molding
Songtao Zhang, Rickey Dubay ⇑, Meaghan Charest
Laboratory of Advanced Controls and Intelligent Sensors, Department of Mechanical Engineering, University of New Brunswick, 15 Dineen Drive, P.O. Box 4400, Fredericton,
New Brunswick E3B 5A3, Canada

a r t i c l e i n f o a b s t r a c t

Article history: Quality control is an important aspect of manufacturing processes. Product quality of injection molded
Available online 23 November 2014 parts is influenced by the injection molding process. In this study statistical tools were used to develop
a model that relates injection molding process variables to part quality. A statistically based model pre-
Keywords: dictive control algorithm was developed for controlling part quality with manipulated variables coolant
On-line quality control flow rate and coolant temperature. This approach replaces the need of off-line quality measurement and
Data mining provides real-time injection modeling quality control.
MPC
Ó 2014 Elsevier Ltd. All rights reserved.
Injection molding
Mathematical modeling

1. Introduction
Injection molding (IM) has been widely used to produce plastic
parts at an economical price. The IM process has proven to be suit-
able for molding thin walled parts with complex geometries where
a degree of precision is required. This is feasible because the IM
process can be adapted to the characteristics of the molding mate-
rial so the desired quality and repeatability can be maintained
(Wang, Zeng, Xiaoxin, & Yuejun, 2013b).
Research concerning IM parameters that effect part quality is an
ongoing field of study. An example of a quality feature is warpage,
the main cause of undesirable warpage in an IM product is the
imbalance of thermal residual stresses. These are caused by a
non-uniform temperature distribution through the thickness of
the moldings resulting from varying cross sections, part geome-
tries, and temperature differences between the mold surfaces,
etc. (Akay, Ozden, & Tansey, 1996). A variety of measures based
on optimization techniques have emerged to improve IM process
parameter settings, injection mold design, and IM machine perfor-
mance with the objective of improving part quality. Optimization
can be achieved in simulation or experimentally using classical
design of experiment tools such as orthogonal experiments and
the Taguchi method (Altan, 2010; Nie, Zhang, & Niu, 2013).
Simulation tools and expert systems such as Moldflow Plastics
Xpert(MPX) (Jauregui-Beckera, Tosellob, van Houtena, & Hansen,
2013; Stanek, Manas, Manas, & Suba, 2011), and neural-network
⇑ Corresponding author.
E-mail addresses: Songtao.Zhang@unb.ca (S. Zhang), dubayr@unb.ca (R. Dubay),
charest.meaghan@unb.ca (M. Charest).
techniques (Fei, Huajie, Lin, Wei, & Maosheng, 2011) outline the
trend to integrate IM process prediction and IM parameter
monitoring into the quality improvement.
Process control is an important aspect of the IM machine
performance. Various modern control techniques have been
applied to the IM process, such as adaptive control, model predic-
tive control (MPC), and other intelligent control algorithms. This
results in more segments of the IM process being controlled and
helps maintain part quality (Wang & Mao, 2012). Examples of
these include the reported study and design of a barrel tempera-
ture controller (Diduch, Dubay, & Li, 2004; Dubay, Diduch, & Li,
2004). The work sought to reduce the melt temperature fluctua-
tions during start-up of the injection molding machine (IMM)
and during continuous part molding. The focus centered on a single
parameter without considering other process parameter interac-
tion, which are important for the part quality. In another investiga-
tion, a temperature-independent adaptive controller was proposed
for injection velocity during the filling phase (Dubay & Lakhram,
2004). They developed a controller which is immune to variations
of processing condition such as melt temperature fluctuations in
order to ensure consistent cavity pressure profiles and hence con-
sistent part quality. Again, the interactions of process parameters
are not being considered.
Despite the efforts of controlling the many parameters that
effect final part quality periodic off-line quality inspection is neces-
sary. Min showed that the quality of IM parts is dependent on both
mold design and processing conditions (Min, 2003). Various factors
including individual IM process settings and variations that relate
the final IM product quality have been investigated in detail and
various control methods have been proposed for the IM process.

http://dx.doi.org/10.1016/j.eswa.2014.11.030
0957-4174/Ó 2014 Elsevier Ltd. All rights reserved.
2920 S. Zhang et al. / Expert Systems with Applications 42 (2015) 2919–2927
Controlling combinations of key process parameters is therefore a
reasonable idea for comprehensive IM quality improvement.
Collecting and analyzing large amounts of data to identify and
understand key information is becoming more prominent in indus-
try (Gandini, Lombardi, & Vaccarino, 2011). An example of this is
the relation of production cost to a variety of injection molding
production factors using artificial neural networks (Wang, Wang,
& Wang, 2013a). In this investigation process parameter data was
collected and analyzed since it express process information and
can be used to identify the quality and control models necessary
for quality based model. Extracting the correct information from
the large amount of process data comprised of unavoidable noise
is the key point for this approach. Principal component analysis
(PCA) was used to construct a quality based regression model. This
model combined with independent component analysis (ICA) is
used to develop a Dynamic Matrix Control Scheme. This makes it
possible to maintain product quality on-line during production.
Recently, implementing powerful statistical tools and data min-
ing techniques has been gaining popularity in industry (Gandini
et al., 2011). When large amounts of data can be collected and ana-
lyzed, it becomes possible to identify and understand trends in the
data. An example of this is the relation of production cost to a vari-
ety of injection molding production factors using artificial neural
networks (Wang et al., 2013a). In another study, a less complex
method was used, a principal component analysis (PCA) based
pattern matching method was applied to process monitoring and
fault detection using independent component analysis (ICA). This
showed the importance of PCA and ICA technique as data mining
tools (Chen et al., 2013).
In this investigation process parameter data was collected and
analyzed in order to identify the quality and control models
necessary for quality based MPC. Extracting the correct informa-
tion from the large amount of process data collected that included
unavoidable noise was a key point for this approach. Principal
component analysis was used to construct a quality based regres-
sion model. This model combined with ICA was used to develop a
MPC scheme. The unique combination of PCA, ICA, and MPC make
it possible to automatically affect part quality on-line while vary-
ing key process parameters. This novel methodology demonstrates
good performance.
2. Background
In order to directly control part quality with on-line process
data (cavity pressure, cavity temperature, etc.), an accurate model
to estimate part quality is built in two steps:
1. Identifying the important components of the IM process using
PCA.
2. Using the principal components obtained as predictors to build
a regression analysis model to evaluate part quality.
The background theory required to obtain the model is
described in this section.
2.1. PCA based quality model
Principal component analysis is a well known statistical method
and has been widely used in data analysis and compression. For
example consider the vector ~ Z which represents the injection
molding process output as
Z ¼ ½z1 ; z2 . . . zn T
~ ð1Þ
where zi with i ¼ 1; 2 . . . n are sensor outputs with noise and its
expectation of ~
Z denoted by
ZÞ ¼ ½Eðz1 Þ; Eðz2 Þ; . . . ; Eðzn ÞT
l~Z ¼ Eð~ ð2Þ
its covariance matrix is
T
C Z ¼ Eðð~
Z  l~Z Þð~
Z  l~Z Þ Þ ð3Þ
The variance of a component of ~ Z indicates the spread of the
possible values of outputs zi around its mean value Eðzi Þ . If two
components zi and zj of the data are un-correlated, their covariance
is zero. The covariance matrix is, by definition, always symmetric.
From a set of sample vectors of ~Z : Z~1 . . . Z~M , the sample expec-
tation and the sample covariance matrix can be calculated as the
estimates of the real expectation and the covariance matrix of ~ Z.
Each sample of ~Z represents the cavity pressure and cavity temper-
ature data from one injection molding cycle. From a symmetric
matrix such as the covariance matrix, a set of orthogonal basis
can be determined by finding the eigenvalues and eigenvectors
of the covariance matrix. The eigenvectors e~i and the correspond-
ing eigenvalues ki are the solutions of the equation
C Z e~i ¼ ki e~i ; i ¼ 1; 2 . . . n ð4Þ
Ordering the eigenvectors in the descending order of eigen-
values (i.e., largest first), an ordered orthogonal basis can be cre-
ated in which the first eigenvector relates to the direction of
largest variance of ~Z. Suppose there exists a data set of which the
sample mean and the covariance matrix have been calculated.
Let A be a matrix consisting of eigenvectors of the covariance
matrix (the eigenvectors are the row vectors of A). The relation
of vector ~
Z and the vector X ~ can be expressed as
~
X ¼ Að~
Z  l~Z Þ ð5Þ
which is a point in the orthogonal coordinate system space defined
by the eigenvector. Components of X~ can be seen as the coordinates
of the orthogonal base.
Using all the eigenvectors of the covariance matrix is not
necessary, the output data can be approximated using some of
the vectors from the orthogonal basis. If the matrix having the first
K (significant) eigenvectors as rows is denoted by AK , a similar
expression can be created as
~
X ¼ AK ð~
Z  l~Z Þ ð6Þ
The vector of the original data has then been projected on the coor-
dinate system having dimension K and the resulting vector can be
expressed by linear combination of the basis vectors to its original
value. This transformation minimizes the mean squared error
(MSE) between the process output data and its approximation.
Thus, by selecting a given number of eigenvectors with the largest
eigenvalues, PCA offers a convenient way to control the trade-off
between loosing information and reducing the size of ~ Z. More
details of PCA can be seen in Jollife (2002), MacGregor and Kourti
(1995) and Burnham and Viveros (1996). Based on the K principle
components obtained, a linear regression method can be applied
to build a prediction model of IM product quality.
2.2. Independent component analysis
Once the PCA has transformed the high dimension IM process
output into an orthogonal coordinated system in which the princi-
ple components of the data are un-correlated, the principle compo-
nents can be viewed as a combination of independent events in the
IM process. As an example, the event can be triggered by the
machine operator who adjusts settings such as the coolant flow-
rate (as the manipulated variables). To identify or separate these
events, ICA is applied on the principal component subspace X.
As the Central Limit Theorem indicates, the distribution of a
sum of independent random variables tends toward a Gaussian
 S. Zhang et al. / Expert Systems with Applications 42 (2015) 2919–2927 2921

distribution, that is output data is the combination of the results of
different manipulated variables. Thus, a sum of two independent
random variables usually has a distribution that is closer to
Gaussian than any of the two original random variables
(Hyvrinen & Oja, 2000). Selecting n independent components
s1 ; s2 . . . sn , and n linear combinations of principal components
~ can be written as
x1 ; x2 . . . xn ; X
~ ¼ B~
X s
where X~ ¼ ½x1 ; x2 . . . xn T , ~
s ¼ ½s1 ; s2 . . . sn T and B is the transforming
matrix. By maximizing non-Gaussianity B is estimated and its
inverse is used to obtain independent components
~ ~
s ¼ B1 X
Using this procedure, the results of varying coolant flow rate and
temperature can be extracted from the process output (cavity
pressure and temperature). More details of ICA can be found in
Hyvrinen and Oja (2000).
adjustment parameter /ðtÞ in general is the difference between
the first element of by ðtÞ and the current state of the controlled var-
iable y.
Quality control is therefore achieved by using a DMC scheme
with a dynamic matrix. Based on the open-loop test data (cavity
temperature, cavity pressure, and flatness), the dynamic response
needed to populate the dynamic matrix can be represented by
the independent components identified from the principle compo-
ð7Þ nents by ICA. Meanwhile, the inputs of this test are the changes of
coolant flow rate and temperature. Furthermore, the quality
regression model gives the relation between flatness and the input.
Based on them, auto-regression with exogenous input (ARX) mod-
els can be built, where the manipulated variables are inputs and
ð8Þ the part quality is the modeled output. Hence, the predictive con-
troller (in DMC form) can be designed to track the quality set point
with a quality feedback observer based on process variables.
Details pertaining to ARX modeling and MPC can be found in
Ljung (1999) and Pandelidis and Agrawal (1987).

2.3. Dynamic Matrix Control 3. IM quality control scheme

The Dynamic Matrix Control (DMC) strategy is selected as a In this section the construction of an on-line quality control
control method since it has been used extensively in IM control scheme is described. The target quality feature being controlled
(Clarke, Mohtadi, & Tuffs, 1987). The DMC algorithm consist of a is part warpage. The warpage must fall within an acceptable range
model that is used to make process prediction b y (i.e. part flatness). and must be consistent and repeatable. Problems in this quality
The prediction is used to determine future errors based on a feature are attributed to residual stress and part shape, it is gener-
setpoint ysp . Optimization is used to calculate control actions that ally unavoidable but can be kept within limits. The control meth-
will minimize future errors, the difference between ysp and the odology is outlined with the following three steps:
estimated quality index, e. The cost function for the optimization is
X
N X
nu 1. Formulate a regression IM product quality model that relates
2 2
JðDu; tÞ ¼ ½ysp ðt þ kÞ  b
y ðt þ kjtÞ þ k½Duðt þ k  1Þ ð9Þ the part surface flatness to cavity pressure and temperature.
k¼1 k¼1 2. Use the above mentioned procedure to build the dynamic
matrix from the ARX model.
where by ðt þ kjtÞ is a k-step ahead prediction of the process output at
3. Design a MPC structure that can evaluate part quality and hence
time instant t, ysp ðt þ kÞ is a reference trajectory set point at instant
achieve on-line quality control.
k into the future. The control horizon nu represents the number of
future control moves Duk that are considered in the optimization.
A measurable indicator of warpage is part flatness. Uneven tem-
A move suppression factor k is imposed on Du to ensure a viable
peratures caused by cooling can cause the plastic part shown in
solution to the objective function.
Fig. 1 to warp. The quality index in this study is primarily related
The necessity of k stems from the construction of the model
to a flatness measurement.
used to evaluate b y and is only necessary when using the standard
least squares optimization approach in solving Eq. (9). The model
component of the DMC algorithm is a dynamic matrix constructed 3.1. Experimental setup
with a vector representation of an open loop test.
0 1 A series of experiments were conducted using a 150 tonne
p1 0  0
hydraulic Engel IMM in order to collect data. During the injection
Bp p1  0 C
B 2 C process, the polypropylene melt is injected into a closed mold
P¼B
B .. .. .. .. C
C ð10Þ through a sprue and runner. The part is allowed time to cool and
@ . . . . A
then is ejected. The part being molded in these experiments is
pN pðN1Þ pðNnu Þ shown in Fig. 1. Cavity temperature and pressure are recorded dur-
where pi is the ith element of the open loop test and N is the ing the process with pressure transducers and thermocouple sen-
prediction horizon. The open loop test can be expressed as actual sors. Two pressure sensors and two temperature sensors are
or simulated plant response. This model is then used with mounted on the inner wall surface of the stationary half of the
Eq. (11) to determine Du. In this case k conditions P so that its mold. One pressure is located at the mold gate (Pg) and another
inverse can be taken. pressure is located at the remote end of the part (Pr). The thermo-
1
couples are located beside the pressure sensors and are analo-
Du ¼ ðP T P þ kIÞ P T e ð11Þ gously named Tg and Tr. The mold is cooled with water, a
The prediction is calculated using Du. The expression for solving b
y is dedicated chiller provides flow that is regulated with a control
X
k valve. All data acquisition and actuation signal communication is
b
y ðt þ kjtÞ ¼ b
y ðtÞ þ Pi Duðt þ k  ijtÞ done through a field terminal and DAQ card housed in a PC. A user
i¼knu þ1 interface was constructed using LabWindows CVI software.
X
Nþk1 The quality index y of the molded part flatness is defined as the
þ ðPi  Pik ÞDuðt þ k  ijtÞ þ /ðtÞ ð12Þ standard deviation of the part thickness at 12 points. The measure-
i¼kþ1 ments were made at equidistant points on both sides of the parts
Here, b
y ðt þ kjtÞ is a k-step ahead prediction of the process output y surface. All measurements were taken 24 h after production to
at time instant t, in this case b y ðtÞ is a product quality index. The allow the polymer to set. A Coordinate Measuring Machine CMM
2922 S. Zhang et al. / Expert Systems with Applications 42 (2015) 2919–2927

Fig. 1. IM plastic bar sample.

Flow rate
35
was used to collect the measurement data which was recorded and
associated with the process data. 30

3.2. Component identification

25

First PCA and regression is used to extract the K principal com-

Flow rate (L/min)
20
ponents from measurements on cavity pressure and temperature
to predict the quality index y. In this investigation, cavity pressure
15
and temperature are chosen as the process variables for relating
the IM process and part quality. Cavity pressure was selected
10
because it is a good representation of packing pressure which
has direct influence on part flatness.
5
Coolant flow rate significantly influences warpage of IM parts
and is chosen as a manipulated variable to control part quality
0
(Collin, 2003). Coolant flow rate can be easily varied using a control
valve. Mold temperature is greatly influenced by the coolant tem-
perature, in this case coolant temperature is considered a distur- −5
0 50 100 150 200 250 300 350 400
bance to mold temperature which is modeled and accounted for. Cycle No.
Here the coolant temperature is measured at the coolant line going
Fig. 2. Coolant flow rate in experiment.
to the mold inlet. This is a good representation of the temperature
fluctuation of the coolant fluid since the chiller works with on/off
control. ~ ¼ ½x1 ; x2 ; . . . ; x6 T is the principal component vector variable
where X
Using PCA, six uncorrelated principle components (PCs) were
for each injection cycle. The residue e~ is the influence of unconsid-
extracted from cavity pressure and temperature profiles, which
ered factors and noise expressed by the principal components. The
consist of 600 measurements. The data was collected while the ~i ¼ ½Bi1 Bi2 . . . Bi6 T ; i ¼ 1; 2 . . . 6 are corresponding mixing
vectors B
manipulated variables were changed as shown in Table 1 and in
vectors which is a weight applied to each principal component for
Figs. 2 and 3. Independent component analysis was then applied ~i can be
obtaining the independent component. These weights B
on the six PCs to provide the six statistically Independent Compo- ~
determined by regression when X and si are known.
nents (IC) that are shown in Fig. 4. Two of the six ICs matched the
changes made in the manipulated variables while the other ICs are
3.3. ARX model construction
representative of noise. Noise is unavoidable in practice and can be
fitted to ARX model without much difficulty. It is therefore reason-
To develop a relationship between part quality and the ICs at
able to view the two selected ICs as the dynamic responses to
any kth molding cycle, the relation between ~
X and b
y is expressed
changes in the manipulated variables. The ICs labeled as IC.4 and
as
IC.5 in Fig. 4 are related to coolant temperature and coolant flow
rate respectively. b
y ðkÞ ¼ b0 þ b1 x1 ðkÞ þ b2 x2 ðkÞ þ . . . þ b6 x6 ðkÞ ð14Þ
To develop a process model for DMC, variables s1 and s2 related ~ and is expressed in Eq. (13). Trans-
where x1 ; x2 . . . are elements of X
to IC 4 and 5 are introduced and their respective open loop test
formation Eq. (14) yields the following
results are shown in Figs. 5 and 6. The IM cavity pressure and tem-
perature within the mold cavity for each cycle can be expressed as b
y ðkÞ ¼ b0 þ ðb1 B11 þ    þ b6 B16 Þs1 ðkÞ þ ðb1 B21 þ   
~ ¼ s1 B~1 þ s2 B~2 þ e~
X ð13Þ þ b6 B26 Þs2 ðkÞ þ ey ðkÞ ð15Þ

Table 1
Open loop test schedule.

Cycle No. 1–75 76–150 151–225 226–300 301–375 376–450

Coolant temperature (°C) 25 25 25 15 15 15
Coolant flow rate (L/min) 5 30 5 5 30 5
 S. Zhang et al. / Expert Systems with Applications 42 (2015) 2919–2927 2923

Coolant temperature Independent component of coolant flow rate

30 2.5

28 2

1.5
26

Independent component S1
1
24
Temperature (C)

0.5
22
0
20
−0.5
18
−1

16 −1.5

14 −2

12 −2.5
0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400
Cycle No. Cycle No.

Fig. 3. Coolant temperature in experiment. Fig. 5. Independent component of coolant flow rate.

A part quality index estimate can then be made where, by 1 is the quality index related to coolant flow rate shown in
Fig. 7 and u1 is the coolant flow rate. Fig. 8 shows the quality index
b
y ðkÞ ¼ b0 þ b1 s1 ðkÞ þ b2 s2 ðkÞ þe 
y ðkÞ ð16Þ response to a unit step change in the coolant flow rate demonstrat-
P6 ing a relatively large time constant which is expected.
where, b0 ¼ b0 ; bi ¼ j¼1 bi Bij . In this case b1 s1 ðkÞ and b2 s2 ðkÞ
Coolant temperature was also found to be an important influ-
express the coolant flow rate and temperature effects on the part
encing factor on the quality index. In this experiment the coolant
quality index y. The noise, ey ðkÞ describes disturbances of the IM
temperature is treated as a disturbance to the process and is mod-
process that cause unexpected deviations in factors that influence
eled with the 4th order ARX model.
the actual part quality y, such as ambient temperature and polymer
purity.
X
4 X
3
Considering the quality response b y 1 as the system output due to b
y 2 ðkÞ ¼ a2i b
y 2 ðk  iÞ þ b2j u2 ðk  jÞ ð18Þ
flow rate value input, a 4th order ARX model can be identified as i¼1 j¼1

X
4 X
3
b
y 1 ðkÞ ¼ a1i b
y 1 ðk  iÞ þ b1j u1 ðk  jÞ ð17Þ where, b
y 2 is the quality index related to coolant temperature and u2
i¼1 j¼1 is the coolant temperature setting. In Fig. 9 the quality index

Independent Components
5
IC.1

−5
0 50 100 150 200 250 300 350 400 450
20
IC.2

−20
0 50 100 150 200 250 300 350 400 450
10
IC.3

−10
0 50 100 150 200 250 300 350 400 450
5
IC.4

−5
0 50 100 150 200 250 300 350 400 450
5
IC.5

−5
0 50 100 150 200 250 300 350 400 450
5
IC. 6

−5
0 50 100 150 200 250 300 350 400 450
Cycle No.

Fig. 4. Independent components.

2924 S. Zhang et al. / Expert Systems with Applications 42 (2015) 2919–2927

Independent component of coolant temperature Step response for coolant flow rate
2.5 x 10
−4

1.5
Independent component S2

1
3

Quality index (mm)

0.5

−0.5
2
−1

−1.5

−2
0 50 100 150 200 250 300 350 400
Cycle No. 1
0 50 100 150 200
Fig. 6. Independent component of coolant temperature. Cycle No.

Fig. 8. Step response for coolant flow rate.

Open loop response for coolant flow rate

−3
Step response for coolant temperature
0.089 x 10
0

0.0885 −0.2

0.088 −0.4
Quality index (mm)

Quality index (mm)

−0.6
0.0875

−0.8
0.087

−1
0.0865

−1.2

0.086
0 100 200 300 400 500
Cycle No.
−1.4
0 50 100 150 200
Cycle No.
Fig. 7. Open loop response for coolant flow rate.
Fig. 9. Step response for coolant temperature.

X
k
response caused by a unit step increase in coolant temperature is b
y ðt þ kjtÞ ¼ b
y ðtÞ þ P ci Du1 ðt þ k  ijtÞ
shown. i¼knu þ1
Based on the IM process ARX model Eqs. (17) and (18) and the X
Nþk1

quality index estimation model in Eq. (14) an on-line quality þ  ðP ci  Pcik ÞDu1 ðt þ k  ijtÞ
control scheme is achieved. Disturbances to the IM process can i¼kþ1

be difficult and sometimes even impossible to control, they should X

k

always be minimized in order to optimize part quality. Replacing þ Pdi Du2 ðt þ k  ijtÞ þ   
i¼knu þ1
on-line quality measurement with the quality estimation Eq.
(14), the dynamic models of Eqs. (17) and (18) make the task of X
Nþk1
þ ðP di  Pdik ÞDu2 ðt þ k  ijtÞ þ /ðtÞ ð19Þ
compensating for quality loss due to a disturbance feasible. i¼kþ1

3.4. DMC implementation
In Fig. 10 a block diagram representation of the closed loop con-
trolled plant is shown. An estimated IM process model is used to
predict the future part quality index caused by changes in coolant
flow rate and the coolant temperature. The prediction equation is
expressed as
where P c and P d or dynamic matrices obtained from Eqs. (17) and
(18).
When the estimated quality index is compared with the desired
quality index ysp , the DMC/MPC optimizer can evaluate the manip-
ulated variable value in order to minimize the deviation of the esti-
mated quality b y over the prediction horizon N. The Gm block in
Fig. 10 represents the model used to predict b y over N using
 S. Zhang et al. / Expert Systems with Applications 42 (2015) 2919–2927 2925

Fig. 10. Quality predictive control details.

dynamic matrices P c and P d corresponding to coolant flow rate and Table 2

temperature. The output of the quality measurement estimator b ym ANOVA (analysis of variance) for the open loop experiments.
is subtracted from by ðtÞ to obtain the adjustment parameter which Source SS df MS F Probability
is added to all values of the prediction. Temperature 0.02039 1 0.02039 20.85 0
Flow rate 0.02184 1 0.02184 22.33 0
4. Simulation and experimental results and discussions Interaction 0.00091 1 0.00091 0.93 0.3356
Error 0.038522 394 0.00098
Total 0.43025 397
Open loop test were performed for all combinations of coolant
flow rates (5 L/min, 30 L/min) and coolant temperatures (15 °C,
25 °C). These manipulated variable test setting combinations are
provided in Table 1. Experimental test and control simulation are shown in Table 2. The zero probability that y corresponds to
results are presented in this section. the coolant flow rate indicates a strong relationship between flow
rate and quality index of the IM product. Similarly, the coolant
temperature was found to be an influential factor. The quality
4.1. Quality measurements
index can therefore be viewed as the linear combination of coolant
flow rate and coolant temperature.
Fig. 11 shows the quality index mean value y for the manipu-
lated variable settings shown in Table 1. To analyze the variance, In simulation and during open loop test an injection cycle time
of 15 s and a sensor sampling interval of 0.1 s was used. Each sen-
ANOVA test were applied on quality index the results of which
sor Pg; Pr; Tg; Tr produces 150 sampled variables for the cycle.
The dimension of the process condition space is hence 600. Fur-
Quality indices for different settings thermore, the data was found to be strongly correlated meaning
0.11 that is not suitable for creating a regression model. Using PCA,
30L/min, 15C the process data volume was significantly reduced and uncorre-
lated allowing the extraction of suitable variables resulting in a
0.105
more reliable and robust regression. The contribution to the vari-
ance that the different PCs make was found and is shown in Table 3.
0.1
Quality index (mm)

0.095
Table 3
Percentage of variance explained by principal components.
0.09 5L/min, 15C 5L/min, 15C
30L/min, 25C Order of principal component Contribution to variance (%)
1 50.87
0.085 2 29.36
3 3.49
4 1.88
0.08
5 1.42
5L/min, 25C 5L/min, 25C 6 1.18
0.075 7 0.80
0 50 100 150 200 250 300 350 400 8 0.54
Cycle No. 9 0.49
10 0.36
Fig. 11. Mean flatness of experiment.
2926 S. Zhang et al. / Expert Systems with Applications 42 (2015) 2919–2927

Removing the redundant data allows for manageable ICA and Manipulated variable
30
regression analysis. The first six PC were selected in this case to
do the ICA regression.

Flowrate (L/min)
25
The results of the analysis show that part quality index is
directly linked to the cavity pressure and temperature within the 20
mold for each injection cycle. The estimated quality index predic-
15
tion based on pressure and temperature data is shown in Fig. 12.
From Fig. 12 it can be seen that the variation of quality cannot 10
be completely explained by current chosen predictors (six PCs), 0 20 40 60 80 100 120
this is reserved for future study. Quality
0.094

Quality index (mm)

0.092
4.2. Controller simulation
0.09
The entire quality control loop has been simulated in LabWin-
dows environment to demonstrate controller performance. A pro- 0.088 setting
cess variable simulation module for cavity pressure and real value
0.086
temperature was developed as part of the control program. The 0 20 40 60 80 100 120
module replaces the IMM in the simulated control loop. This is Cycle No.
done to avoid the costs and setup associated with producing parts.
Fig. 14. Tracing quality setting.
The simulation scheme is shown in Fig. 13.
The first closed loop simulation involves a step increase in ysp
from 0.0875 mm to 0.0925 mm. Here, the coolant temperature is
fixed at a normal condition of 21 °C. The control scheme is able
to adjust the manipulated variable so that the estimated part qual- Manipulated variable
40
Flowrate (L/min)
ity can attain the target quality as shown in Fig. 14. The controller
is capable of reaching the setpoint ±1% within 30 cycles, or 7.5 min. 30
The second closed loop simulation involves an increase in cool-
20
ant temperature from 21 °C to 26 °C. The results of this disturbance
is shown in Fig. 15, a drop in part quality is apparent. The control 10
0 20 40 60 80 100 120
scheme is able to adjust the manipulated variable so that the qual-
30
Temperature (C)

ity can be returned to the target quality setpoint within 35 cycles.

It should be noted that there is a small steady state error in ys . This
25
is due to the limit on the manipulated variable. The tuning param-

20
0 20 40 60 80 100 120
Quality index (mm)

0.088
Estimated quality index
0.35
Estimated flatness
0.086
Moving average setting
0.3
real value
0.084
0 20 40 60 80 100 120
0.25 Cycle No.
Flatness (mm)

0.2 Fig. 15. Compensating disturbance.

0.15

eters for this simulation are the same as those for the first closed
0.1
loop simulation.
A challenging simulation test was conducted to test the limits
0.05
of the controller, the described setpoint changes are not realistic
in industry. The ysp was changed from 0.097 mm to 0.087 mm
0
0 50 100 150 200 250 300 350 400 while the coolant temperature changed from 18 °C to23 °C. In
Cycle No.
Fig. 16, it can be seen that the high quality index setpoint of
Fig. 12. Quality index estimation of the plastic part.
0.097 mm can be achieved with a low coolant temperature of
18 °C and the steady state flow rate of 34 L/min. High coolant
temperature (23 °C) and low flow rate (21 L/min) result in insuf-
ficient product quality. The flow rate initial drops quickly (3 L/
min) in response to the change in the quality index and then
increases slowly. The change in temperature coolant (21 L/min)
is slow due to large time constant. The flow is appropriately opti-
mized by the DMC controller in order to compensate for the dis-
turbance effects. The desired quality index setpoint is attained. It
should be noted that a linear MPC algorithm was used and that
PCA and ICA proved effective in this case under the assumption
Fig. 13. Simulation of quality control. of process linearity.
 S. Zhang et al. / Expert Systems with Applications 42 (2015) 2919–2927 2927

Manipulated variable further developed to release hardware requirements and reduce

40
processing times to allow for faster process applications.
Flow rate (L/min)

30

20 References

10 Akay, M., Ozden, S., & Tansey, T. (1996). Prediction of process-induced warpage in
injection molded thermoplastics. Polymer Engineering and Science, 36(13),
0
0 10 20 30 40 50 60 70 80 1839–1846.
Altan, M. (2010). Reducing shrinkage in injection moldings via the Taguchi anova
Disturbance
and neural network methods. Material and Design, 31, 599–604.
24
Temperature (C)

Burnham, A., & Viveros, R. (1996). Frameworks for latent variable multivariate
22 regression. Journal of Chemometrics, 10, 31–45.
Chen, J., Yua, J., Moria, J., Rashida, M. M., Hub, G., Yub, H., et al. (2013). A non-
20 gaussian pattern matching based dynamic process monitoring approach and its
18 application to cryogenic air separation process. Computers & Chemical
Engineering, 58, 40–53.
16 Clarke, D., Mohtadi, C., & Tuffs, P. S. (1987). Generalized predictive control – Part I.
0 10 20 30 40 50 60 70 80
The basis algorithm. Automatica, 23(2), 137–148.
Quality Collin, M. (2003). The study of the influence of coolant flow rate on injection
Quality index (mm)

0.1 molding products warpage. Senior Year Project Report, Department of

0.095
0.09
0.085
0 10 20 30 40 50 60 70
Cycle No.
Fig. 16. Negative quality step and positive disturbance step.
5. Conclusion
An innovative methodology for effectively controlling part qual-
ity associated with the cooling cycle of the IM process on-line was
developed. The proposed scheme is independent of process vari-
able type and can therefor be easily extend to include other process
variables such as polymer condition or process parameter profiles.
Furthermore, the scheme can be readily adapted for other model
based predictive control algorithms. Overall this methodology
answers the need for adaptable on-line quality control.
Recently, implementing powerful statistical tools and data min-
ing techniques has been gaining popularity in industry. Mean-
while, integration of data processing and advanced control
theory has not reached its frontier. The research presented here
is an example of how to combine statical tools and MPC. The algo-
rithm was designed for a linear control problem and uses basic
data mining methodologies, these points present limitations for
the algorithm in its current form.
A natural extension to this research is the practical application
of the technology to industrial scale quality control problems. To
this end, improvements to the methodology to account for process
nonlinearities should be made. This includes the integration of
advanced PCA and ICA based tools as well as other data mining
techniques. Also, optimization of on-line computation should be
Mechanical Engineering, University of New Brunswick.
Diduch, C., Dubay, R., & Li, W. (2004). Temperature control of injection molding. Part
I: Modeling and identification. Polymer Engineering and Science, 44(12),
2308–2317.
Dubay, R., Diduch, C., & Li, W. (2004). Temperature control of injection molding. Part
II: Controller design, simulation, and implementation. Polymer Engineering and
80 Science, 44(12), 2318–2326.
Dubay, R., & Lakhram, R. (2004). A temperature-dependent adaptive controller. Part
I: Controller methodology and open-loop testing. Polymer Engineering and
Science, 44(10), 1925–1933.
Fei, Y., Huajie, M., Lin, H., Wei, G., & Maosheng, S. (2011). Back propagation neural
network modeling for warpage prediction and optimization of plastic products
during injection molding. Materials and Design, 32, 1844–1850.
Gandini, M., Lombardi, F., & Vaccarino, F. (2011). Blind separation of manufacturing
variability with independent component analysis: A convolutive approach.
Expert Systems with Applications, 38.
Hyvrinen, A., & Oja, E. (2000). Independent component analysis: Algorithms and
applications. Neural Networks, 13, 411–430.
Jauregui-Beckera, J. M., Tosellob, G., van Houtena, F. J. A. M., & Hansen, H. N. (2013).
Performance evaluation of a software engineering tool for automated design of
cooling systems in injection moulding. In Procedia CIRP (Vol. 7, pp. 270–275).
Jollife T. (2002). Principal component analysis: Springer series of statistics (2nd ed.).
Ljung L. (1999). System identification, theory for the user: 1999 PTR prentice hall
information and system sciences series (2nd ed.).
MacGregor, J., & Kourti, T. S. (1995). Process control of multivariate processes.
Control Engineering Practice, 3(3), 403–414.
Min, B. H. (2003). A study on quality monitoring of injection-molded parts. Journal
of Materials Processing Technology, 136, 1–6.
Nie, Y., Zhang, H. M., & Niu, J. T. (2013). Optimization of the injection molding
process parameters based on mold flow and orthogonal experiment. Key
Engineering Materials, 561, 539–543.
Pandelidis, I. O., & Agrawal, A. (1987). Self-tuning control of ram velocity in
injection molding. SPE Annual Technical Conference, 33, 235.
Stanek, M., Manas, D., Manas, M., & Suba, O. (2011). Optimization of injection
molding process. International Journal of Mathematics and Computers in
Simulation, 5, 413–421.
Wang, J., & Mao, Q. (2012). A novel process control methodology based on the pvt
behavior of polymer for injection molding. Advances in Polymer Technology, 32,
E474–E485.
Wang, H., Wang, Y., & Wang, Y. (2013a). Cost estimation of plastic injection molding
parts through integration of pso and bp neural network. Expert Systems with
Applications, 40(2), 418–428.
Wang, R., Zeng, J., Xiaoxin, F., & Yuejun, X. (2013b). Evaluation of effect of plastic
injection molding process parameters on shrinkage based on neural network
simulation. Journal of Macromolecular Science, Part B: Physics, 52, 206–221.