Professional Documents
Culture Documents
Chen e Turng, 2005
Chen e Turng, 2005
A Review of Current
Developments in Process
and Quality Control
for Injection Molding
ZHONGBAO CHEN, LIH-SHENG TURNG
Polymer Engineering Center, Department of Mechanical Engineering,
University of Wisconsin–Madison, Madison, WI 53706
Received: May 21, 2004
Accepted: March 3, 2005
that the quality sensor and the process and quality model are the two most
important areas for further advancement in injection molding control. C 2005
Wiley Periodicals, Inc. Adv Polym Techn 24: 165–182, 2005; Published online in
Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/adv.20046
tabulated in Table I (after Wang et al.2,3 ). These vari- monly referred to as process control, has been widely
ables have a direct impact on the final part quality studied and reported in various publications.
and process economics. The level 3 variables, known as quality variables,
In particular, the variables at level 1 deal with ma- are a collection of various quality criteria. The control
chine operation and normally can be well and in- of level 3 variables is the ultimate focus of the con-
dependently controlled with proper controllers and trol system and of prime importance to the injection
sensors in the feedback loop. These variables are molding operation. Depending on the applications
called machine variables, and the control of these and functional requirements of the part, different
variables is generally referred to as machine con- quality indices may be selected. Nevertheless, all of
trol. Nowadays, it is very common for modern ma- the quality variables are the responses of levels 1 and
chine manufactures to equip the machine with cer- 2 variables. As mentioned previously, the quality dy-
tain machine controllers such as the programmable namics, which govern the relationship between ma-
logic controller (PLC) and proportional, integral, and chine and process variables and quality variables,
derivative (PID) controller to facilitate the machine is not well understood at present. This in part con-
control and operation. tributes to why quality control is lagging far behind
The level 2 variables (process or dependent vari- machine and process control.
ables) depend on not only the process conditions (the The three categories of variables that describe the
level 1 variables) but also the material, machine, and hierarchy and dependency of injection-molding con-
mold configuration used. They reflect the character- trol fit in a multiple-level control system architecture,
istics of the resin being processed and the machine as shown in Fig. 1. Specifically, the feedforward and
dynamics, among other process and design param- feedback loops at the three levels encompass and
eters. The control of these variables, which is com- enable the overall quality control. In the past, Gao
and Yang4 introduced a similar total control system
structure with three feedback loops. However, a sys-
TABLE I tem that thoroughly implements online quality con-
Three-Level Variables in Injection Molding2 trol has not been reported in the literature. Most of
Level 1—Machine variables (independently controllable) the prior and current work deals with the control
Temperature tasks at the first two levels, usually by controlling
Barrel temperature (in several zones) one or several important machine and process vari-
Nozzle temperature ables. For quality control, this is often achieved in-
Coolant temperature directly through online dynamic process control or
Pressure off-line statistic process control (SPC).
Pack/hold pressure There are two major challenges to the implemen-
Back (recovery) pressure
tation of online, real-time quality control, as listed
Maximum injection pressure
Sequence and Motion
below.
Clamp/fill/pack/hold/recovery/eject switchover point
Injection (ram) speed (constant or profiled)
1. The complex quality dynamics in injection molding.
Screw (rotation) speed Insufficient quantitative information is avail-
Shot volume and cushion (via screw displacement) able regarding the relationship among ma-
chine, process, and quality variables. Even for
Level 2—Process (dependent) variables
Melt temperature (in the nozzle, runner,
the process level control in which the process
or mold cavity) dynamics could be described by a set of gov-
Melt pressure (in the nozzle, cavity) erning equations, it still lacks degrees of free-
Melt-front advancement dom to manipulate the state of the polymer
Maximum shear stress melt through space and time.
Rate of heat dissipation and cooling 2. The lack of online quality feedback information.
Level 3—Quality definitions (final response) With existing technologies, direct and online
Part weight and part thickness measurement of part quality within the cycle
Shrinkage and warpage time remains a challenging task for commercial
Sink marks applications. In addition, some of the quality
Appearance and strength at the weld lines criteria of a qualitative nature, such as aes-
Other aesthetic defects: burn marks, gate blushes,
thetic features, are difficult to measure quan-
surface texture, etc.
titatively by regular instrumentation methods.
Moreover, the diversity of quality requirements pressure, coolant temperature, and barrel tempera-
across multiple molding applications is an- ture) depends on the machine, mold, and material.
other major barrier to achieving online quality Typically, these parameters can be set at the ma-
feedback. chine’s operating console. The relationship between
process conditions and the final quality of molded
This paper aims to review and summarize the cur- parts can be expressed as a mapping in the follow-
rent research and developments in injection mold- ing form:
ing process and quality control so that interested
readers will be able to identify proper research areas Q(n, t) = f ( ps (t), n, t) + v(n, t) (1)
for further advancement, pinpoint a feasible path to
real, online quality control, and realize the dream of where Q is the collection of level 3 variables, which
“lights-out” production. In the following sections, include quantitative and qualitative variables such
the machine setup methods will be presented first as part weight, volume shrinkage, burn marks, sink
since they are the starting point of control to deter- marks, etc.; n is the cycle number; t is the time from
mine the setpoint values (SVs) for the process, even the starting point of the cycle; ps (t) is the collection of
though the optimal SVs for level 1 and/or level 2 process conditions; v is the disturbances; and f is the
variables might not be constant in injection molding. mapping function from process conditions to qual-
Then machine control is introduced, with empha- ity without considering the disturbances. Unfortu-
sis on advanced control algorithms that need to be nately, f is typically very complicated or unknown
implemented on microprocessors. Next two sections a priori, making the setup of process conditions a
are devoted to process control and quality control, re- difficult and cumbersome job.
spectively. The paper concludes with suggested re- In practice, the expression of f has to be simplified
search work to achieve the ultimate online quality to a certain extent in order to establish a reasonably
control. accurate mapping between process conditions and
part quality. For example, by ignoring the system
dynamics, the so-called optimal process conditions
can be obtained by solving a static optimum prob-
Determination of Machine lem as is usually done. The methods of determin-
ing the process conditions can be categorized into
Variables (Process Conditions) two approaches, depending on how Eq. (1) is inter-
preted; namely, the first-principle model approach
The process conditions are the SVs of level 1 vari- and the empirical model approach. The first ap-
ables, as shown in Table I. The determination of pro- proach involves the use of scientific-based computer-
cess conditions (such as shot size, injection speed, aided engineering (CAE) software5,6 or mathemati-
pack/hold time and pressure, cooling time, back cal models7 to assist the process setup. On the other
and Yuen7 proposed an analytical model for injec- other hand, regression analysis yielded a more con-
tion molding, based on which the filling pressure, cise model, usually expressed in linear or quadratic
clamp force, shear stress, shear rate, and tempera- multivariable polynomials, which facilitated the op-
ture at different time instants and locations can be timization task.
calculated and used to determine suitable process Tan and Yuen22 proposed a fuzzy multiple-
conditions. objective approach to set up the process for mini-
mizing injection molding defects. In their study, the
defects were expressed by a scale number through
fuzzy functions. Additional data of quality defect
MACHINE SETUP BASED ON THE
scales and process parameters were obtained
EMPIRICAL MODEL
through DOE. The relationship between the sever-
Several types of empirical models have been ities of the defects and the machine variables was
developed specifically for certain machine, mold, approximated by a set of quadratic polynomials via
and material combinations. The common techniques regression analysis. Accordingly, the optimal process
for developing the empirical models include ANN, conditions could be obtained by minimizing injec-
case-based reasoning (CBA), DOE, expert system, tion molding defects.
fuzzy logic, and numerical regression analysis. As mentioned above, DOE has been used in con-
Again, these empirical models are often combined junction with other methods for generating mean-
with optimization algorithms for determining the ingful experimental data. Sometimes the data them-
optimal process conditions. Hence, the empirical selves can serve as a model, allowing DOE to
models sometimes are constructed in the reverse be applied directly to determine the process con-
form of f in Eq. (1), which take part quality and pro- ditions. For instance, Liao et al.23 started with
cess conditions as input and output, respectively.18 the process conditions suggested by a CAE tool
Amidst the diverse ways of building the compli- and then optimized them with DOE to minimize
cated empirical models, ANN is one of the widely the shrinkage and warpage of a cellular housing
used methods.18–21 In this approach, some indices part.
of part quality, such as weight, thickness, warpage, Mok and Kwong10 developed a CBA system com-
shrinkage, flash, and/or strength, are established as bined with fuzzy logic and neural networks for
the output of neural networks while the input are ei- determining the initial process conditions. In their
ther the machine variables (process conditions), such study, the model of the process was in the form
as injection speed, injection pressure, holding pres- of a case library. The match of the current prob-
sure, holding time, cooling temperature, and bar- lem with the library was solved through fuzzy in-
rel temperature, or the process variables, such as ference, and the case adaptation was implemented
nozzle pressure, cavity pressure, nozzle tempera- in neural networks. Their research presents another
ture, or a combination of them. It is not surpris- type of empirical model of the injection molding
ing that the results from those previous studies sug- process.
gested that networks based on process variables By comparing the empirical-model methods with
could predict the part quality more accurately than those based on first-principle models, it can be noted
those based on machine variables. This is because that first-principle models give more general instruc-
the process variables reflect more closely how the tions in less time. On the other hand, the empirical
polymer is being processed given the limitations models have to be built with time-consuming data
existing in the machine control system. Note that preparation from case to case; but they generally give
the ANN model has to be trained with a set of more accurate results.
well-prepared data capable of describing the pro- The optimal processing conditions set by the
cess sufficiently accurately. Otherwise, the model above-mentioned methods are assumed to exist and
would have little use for process settings. Often, remain constant for a specific combination of ma-
DOE is employed to generate the required train- chine, mold, and material. However, this assump-
ing experimental data over the feasible processing tion may not be true in a real process given various
range. unexpected disturbances. Therefore, it requires some
In Helps et al.,19 a numerical regression analysis methodology to adjust the conditions from cycle to
was performed to compare with the ANN approach. cycle in order to compensate for the disturbances.
The results showed that ANN could attain a more This will be discussed later in the quality control
accurate prediction than regression analysis. On the section.
Gao36 used a fuzzy logic-based model to develop an ing, which is the switchover point from the injection
FLC for injection velocity. Their experimental results phase to the pack/hold phase. Typical signals used
revealed that the fuzzy logic controller worked well for determining the switchover point include ma-
with different molds, materials, barrel temperatures, chine variables such as time, ram position, and/or
and injection velocity profiles, suggesting that the hydraulic pressure, as well as process variables such
fuzzy logic controller has superior performance over as nozzle pressure and cavity pressure. Chang44
the conventional PID controller in response speed, compared the process capability of five switchover
setpoint tracking, noise rejection, and robustness. modes where part quality of weight and dimension
Because of the cyclical nature of injection mold- were concerned. The research concluded that the de-
ing, it is well suited for a learning control strategy; sirable modes for switchover should be based on, in
namely, ILC, which uses successive repetitions of the descending order, cavity pressure, hydraulic pres-
same action to refine the control input. ILC has been sure, stroke, time, and speed. A similar sequence was
employed to control injection velocity by Gao et al.,37 also suggested in other independent works (see, e.g.,
and to control hydraulic pressure/ram position by Sheth et al.45 ). Due to the abrupt changes that occur at
Havlicsek and Alleyne,32 where the traditional open- the moment of switchover, large disturbances could
loop feedforward compensator was combined with be introduced into the system. Zheng and Alleyne46
feedback optimization control to achieve stable con- developed a comprehensive model representing an
vergence and performance improvement akin to injection molding machine with fill-to-pack dynam-
closed-loop control. The control law of ILC can be ics and proposed a bumpless transfer in order to
expressed in the following form: achieve a smooth transition from filling to packing.
Since the nonlinear and time-varying characteris-
Feedback
tics are inherent in the injection molding machine,
term
advanced adaptive process control technologies are
uk (t) = G c (s) · e k (t) + wk (t) t ∈ [to (k), to (k) + T]
capable of reaching higher standard performance
Present
interation
Controller
transfer
Present
iteration
Present
iteration than conventional PID if the machine dynamics are
input function input feedforward
term well modeled. However, it seems that industries are
(3) still reluctant to utilize advanced control technolo-
gies in their machine controllers, which are still dom-
where the present iteration feedforward term is inated by PID. The reasons for their choice lie in
learned from shot to shot. Both simulation and ex- the fact that advanced control schemes rely highly
perimental results showed that the tracking error of on underlying machine dynamic models. Both its
injection speed or ram position decreased apprecia- structure and coefficient parameters need to be care-
bly within the first few learning iterations. fully validated before being implemented into the
In addition to the control of machine variables, closed-loop control to ensure reliable performance.
there is a unique sequence control in injection mold- The modeling work has to be carried out by control
engineering experts for different machines, which is pressure or valve opening, is selected as the ma-
very costly and time consuming. On the other hand, nipulated variable, and the disturbances from the
traditional PID possesses the advantages of simple environments or other parameters are compensated
structure and easy tuning, with no special assump- through adjusting the hydraulic pressure or valve
tions on the machine dynamics even though its per- opening.52–54
formance is moderate. As shown in the multilevel system structure in
Therefore, it would be highly beneficial for the Fig. 1, the process variables serve as the connection
original equipment manufacturer (OEM) of machine between the machine setting and the final part qual-
controllers if high performance could be achieved via ity response. However, no generic quantitative mod-
either advanced model-based controls with the aid of els have been established for the connections from
automated procedures for modeling and validation, the machine setting to the process variables and from
or new control strategies with fewer, more general the process variables to the quality response. The re-
and realistic assumptions on the machine dynamics. lationship between process variables and quality in-
dices is modeled as quality dynamics, which should
be thoroughly analyzed when setting up the ref-
erence profiles for the controlled variables; while
the relationship between machine parameters and
Process Control process variables is governed by process dynamics,
which should be carefully studied in designing and
As indicated previously, the process (level 2) vari- developing the process controller. Many efforts have
ables are more closely related to quality (level 3) been made in modeling the process dynamics, rang-
variables than are the machine (level 1) variables.19 ing from physical models,43,53,55 ANN models,56,57
Therefore, it is rational to control the process vari- to hybrid models.58 Based on the resulting process
ables rather than the machine variables in order models, different control algorithms have been de-
to achieve desirable and consistent part quality. As veloped to control process variables. Due to the com-
early as the late 1980s, Agrawal et al. 47 suggested us- plexity of the process dynamics mentioned above,
ing “plastic variables,” which are referred to as pro- conventional time-invariant controllers such as PID
cess variables here, as controlled parameters since are not good enough to achieve a satisfying perfor-
plastic variables are the true indicators of the condi- mance, and advanced process control algorithms,
tion of the plastic inside the mold. The validity of the such as adaptive control, generalized predictive con-
suggestion has been proven in numerous research trol (GPC), and intelligent control, have to be devel-
efforts which show that the polymer melt temper- oped (see below).
ature and pressure have a very strong influence on
the quality of the molded part. The characteristics
of cavity pressure throughout all phases of injection CAVITY PRESSURE CONTROL
molding play an especially dominant role concern-
Due to its widely recognized importance, there
ing the part quality, which can be characterized by
has been much research on cavity pressure control.
part weight, thickness, or other dimensional features
Gao et al.52 analyzed cavity pressure dynamics in the
like shrinkage and warpage.48–50
filling and packing/holding stages and designed a
In addition to cavity pressure, other process vari-
self-tuning regulator (STR) to control the cavity pres-
ables, such as nozzle pressure, melt temperature,
sure so that it follows a preset profile through ma-
melt viscosity, and mold separation, are widely em-
nipulation of the servo-valve opening. In their study,
ployed in injection molding process control. These
the dynamics of cavity pressure was found to be both
process variables are dependent upon the collective
nonlinear and time varying in relation to the servo-
effect of the machine setting, the specific resin and
valve opening. It was modeled as a simple integrat-
mold used, and the nonlinear, distributed, and time-
ing process with a time-varying gain, which was in
varying process dynamics. In order to control the
parallel with a first-order process with a varying pro-
process variables to follow their desired profiles, one
cess gain and time constant and in series with a pure
can manipulate a set of machine parameters to form
delay. This cavity pressure model is shown in Eq. (4)
a multivariable control system taking into account
the complex process dynamics.51 Alternatively, there
k1 (t) k2 (t)
is another option that is much simpler in the sense y(s) = + e −td s u(s) (4)
that only one machine parameter, such as hydraulic s τs + 1
being able to predict the cavity pressure accurately oped an STR based on pole placement, which was
based on generic physical models like those imple- calculated using the process parameters identified
mented in CAE software is greatly desired. How- online to control the nozzle pressure at the pack-
ever, there are always some discrepancies between ing stage. It is noted that a forgetting factor was
the experimental data and the predictions, which re- used in the RLS identification algorithm to overcome
sult not only from the inherent disturbances in the time-varying and nonlinear features. Furthermore,
physical system but from omission or simplification the control performance was improved through em-
of some important material properties, such as the ployment of an antiwindup estimation, an adaptive
pressure-dependent viscosity, transient heat transfer feedforward, and cycle-to-cycle adaptation.
coefficient at the plastic–mold interface, viscoelastic- Since the purpose of nozzle pressure control is to
ity, and other variable material properties.63 There- stabilize the process and produce high-quality parts,
fore, improving the accuracy of material models will the reference signal for the nozzle pressure controller
help to predict process variables more accurately, is usually determined by the quality requirement. In
thereby improving the dynamic control of these pro- addition, nozzle pressure combined with the melt
cess variables. temperature in the nozzle can also be interpreted
as an indication that reflects process variations. In
Orzechowski et al.,65 it is reported that the nozzle
NOZZLE PRESSURE CONTROL
pressure peak value and its integration verses time
When installing the cavity pressure transducers can be utilized to monitor the process and detect such
inside the mold is not feasible, as encountered in abnormal situations as throat starvation, cold slug in
many production molds, nozzle pressure is usually the nozzle, and a variation of material properties.
taken as an alternative because it provides the melt
information close to the mold cavity. Moreover, the
MELT TEMPERATURE CONTROL
pressure transducer is installed in the injection noz-
zle and therefore can be used for different molds Melt temperature is another fundamental process
without additional tooling. As with controlling cav- variable whose dynamic models and control meth-
ity pressure, a process dynamic model correlating ods had not been discussed in detail until the late
nozzle pressure with machine parameters is essen- 1980s. As reported in Agrawal et al.,47 intelligent
tial to the design and implementation of an online computation-based methods were used to model the
control system. Due to the lack of a generic phys- temperature process dynamics, and advanced con-
ical model and the wide variation of machine and trol strategies based on a relatively large amount of
process dynamics, different types of empirical mod- online calculation were implemented to achieve per-
els have been developed for various specific appli- formance superior to conventional PID controllers.
cations. Petrova and Kazmer58 developed a hybrid Due to its capability to mimic the behaviors of
neural network that combined the training of neu- complex systems, ANN has been employed in mod-
ral networks with the analytical knowledge of the eling the melt temperature. Zhao and Gao66 com-
molding process to predict the injection pressure at bined ANN with a physical model based simulation
the nozzle as a function of shot size, injection veloc- to predict the melt temperature profile in the noz-
ity, and melt temperature. The prediction capability zle chamber. The melt development process was di-
of the hybrid network was compared with conven- vided into plastication, dwell, and injection stages. In
tional networks trained using the experimental data the plastication stage, an ANN was employed to pre-
and simulated data, respectively. The results showed dict the effect of operating conditions, including noz-
that hybrid networks outperformed the others, espe- zle heater, barrel heaters, screw rotation speed, back
cially when the number of training data points was pressure, and charge stroke, on the melt temperature
limited. This research provides a prospective method profile. Subsequently, a physical model based simu-
to start the process control quickly from the process lation was applied to generate the temperature de-
physical model and then refine the model gradually velopment during the dwelling and injection stages.
when enough online process data become available. The prediction of the proposed method matched the
Yang and Gao64 applied an ARX model to describe experimental measurements very well. The work of
the nonlinear and time-varying dynamics between modeling could be further used in the implemen-
the servo-valve opening and melt pressure in the tation of optimization and advanced control. ANN
nozzle. Following an idea similar to that of cavity was also used by Richard et al.67 to predict the cavity
pressure control methods in Gao et al.,50 they devel- melt temperature based on the data obtained from
the measurement of a thermocouple embedded in- separation profile by manipulating the hydraulic
side an ejector pin. The ANN model was found to pressure in the postfilling stage. Such control over-
predict the real melt temperature history very well. rode the initial settings of switchover point and
As to the method for melt temperature con- packing pressure. Therefore, both cycle-to-cycle and
trol, several researchers have employed the GPC within-cycle adaptive control could be used to com-
algorithm,68–70 which originated from continuous pensate for both long-term and short-term distur-
industrial process control. The melt temperature in bances. The reported experimental results showed
the nozzle is one of the few variables in the cyclic in- that the standard deviation of part weight in 200
jection molding process that can be monitored and consecutive shots could be improved by around one
controlled continuously, while the cyclic movement order of magnitude.
of the ram and screw could be treated as periodic There is no doubt that properly selected process
disturbances. In Dubay et al.,68,70 the heater power of variable control is very helpful in improving the pro-
different heating bands was used to control the melt cess capability, increasing the part quality, and re-
temperature in the corresponding zone for insulated ducing the production cost. There are two conditions
and uninsulated machines through GPC during the that have to be satisfied for the process variables to be
machine start-up period. The performance of decou- qualified for online control. First, the signal should
pling among several zones, temperature overshot, be measurable online via commercially available
and dynamic transient time were the main concerns transducers. Second, the measured variable should
in the study. Furthermore, the online adaptive tun- be closely related to the process and the properties
ing of GPC controller parameters was investigated of molded parts. It is based on these criteria that the
based on the performance.69 cavity pressure and mold separation are preferred
over other variables for process control. However, all
variables have their own strengths and weaknesses.
MOLD SEPARATION CONTROL
For example, the directly flash-mounted transduc-
Mold separation is not a process variable directly ers inside the cavity may not be allowed where the
associated with the polymer melt; rather, it depends surface finish is critical. In this case, there are several
on the integration of melt pressure versus the project methods to overcome the difficulties, namely, using
area in the sprue, runner, and cavity. Moreover, the upstream information in the nozzle or a downstream
mold separation can be measured outside of the cav- signal such as mold separation outside of the cavity.
ity, which is more favorable than installing pressure It is noted that there is extensive research about
transducers inside of the mold. It was found in the applying innovative sensors to measure the process
early research that mold separation, which is on the in a nondestructive way through the development
order of microns at the moment when the mold cav- of associated hardware and software. In addition to
ity is volumetrically filled, was a reliable indictor of mold separation detected by LVDT, the tie bar exten-
part weight or thickness.47,71–73 sion (an alternative measurement for mold separa-
Buja71 and Wenskus et al.72,73 studied the applica- tion) can be measured via a strain gauge.74 A study
tion of mold separation control to improve the con- has shown that the deformation of a tie bar measured
sistence of part weight. In their work, a mold sep- by a pretensioned piezo element was highly corre-
aration signal was directly used to determine the lated with the cavity pressure.75 Compared to mold
switchover point from filling to packing, replacing separation, these measurements contain more use-
the traditional switchover criterion based on ram po- ful information in the sense that they are detectable
sition or hydraulic pressure. In addition, the thresh- during the whole injection cycle. In contrast, mold
old value for switchover was adaptively adjusted separation may not be detectable except at the mo-
from cycle to cycle. With mold separation monitor- ment when the mold is filled, and it requires certain
ing and control, consistency of part weight was im- cavity geometry (with a large project area) to pro-
proved. duce measurable separation with proper resolution.
Recently, Wang et al.2 developed an integrated Ultrasonic sensing is another widely studied tech-
adaptive injection molding control system, in which nology for measuring variables in the injection mold-
mold separation was measured online by a linear ing process.76–80 It was reported that ultrasonic sen-
variable differential transformer (LVDT) mounted sors had been employed to investigate a wide range
along the parting line outside the mold. The of process characteristics such as the melting behav-
system automatically determined the fill-to-pack ior of the resin in the injection barrel, temperature
switchover point and continued to control the mold and pressure variation in the nozzle, melt-front flow
inside the cavity, cavity pressure, solidification be- models, which act as the observers in the control
havior in the cavity, polymer orientation, and some system, are usually implemented offline using such
quality responses such as part dimension. methods as neural network training or data regres-
It is anticipated that the development of the sion analysis. How well the system can regulate and
new sensing technology will benefit the online pro- achieve the predefined quality index will largely de-
cess control by providing more reliable information pend on the accuracy and the capability of adap-
about the process, thus providing opportunities for tation to time-varying characteristics. The perfor-
advanced control to be applied in injection molding mance of the quality models further relies on the
process control. modeling methods and the measurable variables.
Numerous researchers have investigated possi-
ble methods to construct a generic, reliable, accu-
rate, and economical quality observer for process
Quality Control and quality control. Similar to the approaches used
to determine the process settings, the quality ob-
Quality control is the final goal of machine control servers range from different kinds of empirical mod-
and process control. It takes the quality requirement els to first-principle-based models, or a combination
from the design or process engineers and connects of them. As to the quality index, part weight and
the previous two levels of controllers by feeding the part dimensions are the two most widely selected
output to them as reference inputs. Compared to variables to represent the quality because they can
the process setup stage, where the process condi- be quantitatively measured with high resolution.81
tions are sought through static optimization as dis- Woll and Cooper61 compared different empiri-
cussed in section Machine Control, quality control cal models—namely, the linear regression model,
also requires a certain form of quality model that SPC model, and BPN model—to predict part weight
relates the desired quality response to process vari- and length. They concluded that the neural network
ables and/or machine variables. Here, the process model trained using the cavity pressure profile gave
conditions or process variables have to be changed the best prediction. ANN was also employed by
from cycle to cycle to compensate for the possible Rewal and Toncich82 to predict the part weight from
disturbances coming from the material, machine dy- the melt pressure and temperature in the nozzle.
namics, and/or process dynamics. In general, the The calculated part weight fits the experimental data
goal of quality control is to achieve the desired part very well.
quality by adjusting the lower level parameters. Besides the quality model or quality observer, it
As pointed out before, online quality response is noted that the quality controller is a key element
measurement in injection molding is difficult to im- in the quality control loop as well. The previously
plement using commercially available sensors at a built models provide a direct way to develop the
reasonable cost, which precludes direct online feed- controller, which ideally is the inverse of the quality
back for quality control, at least on the production model. However, there are many other methods to
floor. Thus, strictly speaking, there is no so-called handle the design of the quality controller in practice.
closed-loop quality control for injection molding, Ricketson and Wang83 constructed an empiri-
and all of the implementation strategies dealing with cal model of part thickness as a function of pack-
quality control fall into an observer-based control ing/holding pressure, mold temperature, and barrel
system structure, as shown in Fig. 6. The quality temperature via curve fitting experimental data from
a factorial design of experiments using the following
form:
n
Y = a0 + a i xi (5)
i=1
manually measured thickness was outside of the tol- In this implementation, the controller acts as an au-
erance band, the data were used to update the model tomatic troubleshooting system to reduce the chance
and the whole procedure was repeated. This is a typ- of producing parts with severe defects.
ical implementation of quality control based on the Even though the AI-based quality control can han-
empirical regression model. If nonlinear regression dle specific problems very well, it is well known
models are used, an optimization program is needed that the model development and validation need
to search for the optimal solution. a lot of experimental data, which is time consum-
Nevertheless, when the quality observer is ex- ing and costly. Furthermore, the developed model is
pressed in more complicated forms, such as ANN, susceptible to external disturbances even though it
the direct inverse is not easy to obtain; therefore, the can capture highly nonlinear and complex dynam-
controller should be designed with the help of con- ics. Therefore, it is desirable to find a more generic
trol theory. Liang and Wang84,85 proposed an inverse and science-based description of the quality dynam-
neural optimal control system (INOCS) consisting of ics, from which the rigorous multiple objective op-
a closed-loop controller, a quality predictor based on timization procedure can be developed for online
neural networks, and an optimization calculator of quality control.87
a performance index with a multiloss function. It is Kamal et al.88 converted a part weight control to
noted that the controller was also implemented in the process variable control through the simple relation-
form of a neural network whose structure is shown ship between part weight, pressure, and tempera-
in Fig. 7. In Fig. 7, L(t + 1) is the expected loss in cycle ture, as shown below
t + 1; L(t) is the lost value of the current shot; and Ft,
Pt, and Pp are the filling time, packing/holding time, Vc
W= (6)
and pressure, respectively. It is reasonable to expect v(Pp , Tb∗ )
that the current control action will generate the lost
value of L(t + 1) for the next cycle. If the quality pre- where W is the part weight, Vc is the cavity volume,
dictor and inverse model are accurate enough, the v(Pp , Tb∗ ) is the average specific volume evaluated
calculated loss should decrease exponentially, ensur- at the time when the gate freezes off, and Pp and Tb∗
ing that the optimal part quality is achieved. are the cavity pressure and bulk temperature, respec-
In addition to ANN, other forms of artificial in- tively. What remains is how to control or compensate
telligence (AI) technology were also investigated in for the deviation of cavity pressure and temperature
pursuing quality control of injection-molded parts. away from the reference values.
Vagelatos et al.86 integrated the quality observer and Equation (6) gives a generic relationship between
controller into a fuzzy-logic-based controller, which part weight and the pvT properties of materials.
contained rules obtained from machine operator ex- There are several quality control studies based on
perts in the following form: material properties combined with first-principle
models to control part weight or dimensions.89–91
IF defect(i) is A(i) THEN the increase (decrease) Wang and Wang89 proposed a thickness control
scheme that employed an online, one-dimensional
of parameter( j) must be B( j)
simulation program, which corresponded to the
quality observer in Fig. 6. The output of the observer
(i.e., the calculated part thickness) was compared to
the manually measured value, and a linear regres-
sion was employed to compensate for the model
mismatch. In another study, Wang and Lin92 avoided
online simulation. The quality information provided
by the CAE program was extracted through DOE,
and an ANN was trained based on the simulation
results. Then fuzzy logic rules similar to those used
in Vagelatos et al.86 were used to search for the pro-
cess parameters from shot to shot in order to achieve
the desired part quality.
Sheth and Nunn91,93 also employed the material’s
pvT properties to develop an adaptive online weight
FIGURE 7. Inverse neural optimal controller.84 control system that detected the melt temperature
variation through the change of filling pressure and by system disturbances and measurement noises.
packing pressure. Then the required change in the Moreover, the model is applicable only to the specific
holding time and holding pressure level to compen- case investigated. As to online simulation, it faces
sate for the change in melt temperature was calcu- the challenges of reducing the computation time to
lated based on mathematical equations. The exper- be comparable to the cycle time while maintaining
imental results showed that part weight variation prediction accuracy given the necessary modeling
caused by the change in melt temperature could be simplifications, especially for large and complicated
reduced. parts. Recently, artificial intelligence represented by
In the aforementioned quality control methods, ANN has been increasingly investigated for creat-
the quality observers are usually predetermined ing quality models. This is because ANNs are more
through off-line work such as neural network train- flexible, have higher fitting precision of available
ing or data regression analysis. They could be fur- data, and are also convenient for generalizing the
ther improved when measured quality data become modeling process. But this method has weaknesses
available in the future. Different from previous mod- similar to the other empirical models mentioned
eling approaches, Ivester et al. developed a virtual above.
search method (VSM), which interleaved process It seems that the difficulty of developing an on-
tuning and model development.94 The quality ob- line quality sensor poses an obstacle to fully auto-
server was expressed as an input–output (I-O) model mated quality control for the injection molding pro-
that mapped the process parameters to quality char- cess. Fortunately, there are quite a few studies that
acteristics. The structure of VSM is shown in Fig. 8. reported encouraging progress along this line. For
In their proposed method, the I-O model was first example, by using ultrasonic transducers, both the
used as a virtual process for parameter tuning. Only melt-front velocity in the filling stage and the gap
when the search for a feasible operation region had between the part and mold wall surface during cool-
been exhausted based on the current I-O model, or ing can be detected.78 It is expected that the inno-
the feasibility of the inputs obtained from the current vative quality sensor will eventually enable qual-
I-O model needed testing, was the input suggested ity control. Otherwise, one has to refine the generic,
by the search algorithm applied to the real process. first-principle model and construct the “virtual sen-
Subsequently, the measurements of part quality at- sor” for quality response, which is not trivial work
tributes obtained from the process could be used to either.95
update the I-O model.
Generally speaking, the development of a quality
observer or model plays a key role in online quality
control due to the lack of online quality measure- Conclusions and Suggested
ments. It should be kept in mind that all models
discussed above have their limitations. An empiri-
Research Topics
cal model needs considerable experimental effort to
build. The model-fitness error highly depends on the Injection molding control has gone through a pe-
experimental results, which are inevitably affected riod of fast development in the past decade, accom-
panied by several other key technologies such as
novel sensors and information technology (IT). It is
useful to review the up-to-date advancements in this
area in order to identify the right path toward the
final, fully automated production with real time, on-
line quality control. Based on the topics discussed
in this paper, the following research issues are sug-
gested for future studies:
involved in injection molding as a single sys- system enters the steady state but still faces
tem, akin to other large scale, continuous disturbances.
industrial processes (e.g., petroleum oil refin-
ery). The integrated control system does not
necessarily reside in a single, “main” computer;
it can be distributed among several computa- Acknowledgments
tion units linked with a high-speed connection.
Moreover, the integrated control system can be
This research is partially supported by the UW-
further seamlessly integrated with other high-
Madison Graduate School I&EDR Award. We would
level production management tools.
also like to thank Professor K. K. Wang and Dr. C.
2. The implementation of online quality control A. Hieber at Cornell University for their input and
largely depends on the availability of qual- support during the early stage of this work.
ity sensors. Although it can be a difficult
task, developing suitable sensors for measur-
ing molded part quality online is still feasi-
ble. Nondestructive transducers for measuring
polymer behavior inside the cavity are pre- References
ferred. In the meantime, it is also desirable
to improve the accuracy of CAE software to 1. The Society of the Plastics Industry Website, “About the In-
such a degree that the discrepancies between dustry,” http://www.plasticsindustry.org/industry/index.
experiments and simulation are minimized htm.
and/or can be accommodated in a systematic 2. Wang, K. K.; Zhou, J.; Sakurai, Y. In ANTEC’99 1999; pp. 611–
way. 615.
3. Wang, K. K.; Hieber, C. A.; Wang, N.; Zhou, J.; Lee, M. C.;
3. Once CAE software is accurate enough, it
Sakurai, Y. In Cornell Injection Molding Program Progress
can be incorporated into the closed-loop, on- Report no. 21, 2000; Ch. XI.
line process, and quality control, where the 4. Gao, F.; Yang, Y. In ANTEC’97, 1997; pp. 480–486.
simulation can communicate with the ma- 5. C-Mold, AC-Technology Co., 2000.
chine and process controllers, accepting initial 6. Moldflow, Moldflow Co., 2003.
and boundary conditions for simulation, even 7. Tan, K. H.; Yuen, M. M. F. In Proceedings of International Con-
adaptively adjusting parameters in the mate- ference on Manufacturing Automation, Hong Kong, 1997;
rial models. It is therefore beneficial to combine pp. 832–837.
the simulation with quality control and fuse the 8. Richard, C.; Helps, G.; Strong, A. B. In Proceedings of IEEE-
information from the process and simulation Electrical Insulation Conference and Electrical Manufactur-
ing and Coil Winding Conference, 1999; pp. 259–265.
program.
9. Johnson, B. P. In ANTEC’02, 2002; no. 190.
4. Control-oriented process and quality models
10. Mok, S. L.; Kwong, C. K. J Intell Manuf 2002, 13(3), 165–176.
are needed. They should be built and vali-
11. Kameoka, S.; Haramoto, N.; Sakai, T. Adv Polym Technol
dated online based on either accurate CAE 1993, 12(4), 403–418.
predictions or experimental measurements. 12. Turng, L. S.; Peić, M.; Bradley, D. K. J Injection Molding
Based on these models, the advanced con- Technol 2002, 6(2), 143–155.
trol algorithms successfully applied in other 13. Moldflow Plastics XpertTM (MPX), Moldflow Co., 2001.
processes can be utilized for injection mold- 14. Dr. C-Mold, AC-Technology Co., 1999.
ing. By doing so, control of injection molding 15. Nirkhe, C. P.; Barry, C. M. F. In ANTEC’03, 2003; pp. 3534–
will be totally based on generic and scientific 3538.
methodologies. 16. Speight, R. G.; Reisinger, L.; Lee, C.; Spence, M. In ANTEC’01,
2001; no. 553.
5. Multiple-objective optimization is necessary
17. Bakharev, A. S.; Speight, R. G.; Thomas, A. R. In ANTEC’02,
for both process setting and control because
2002; no. 567.
of the typical multiple-part quality indices.
18. He, W.; Zhang, Y. F.; Lee, K. S.; Fuh, J. Y. H.; Nee, A. Y. C. J
In fact, they can be unified in the sense of Intell Manuf 1998, 9(1), 17–27.
process optimization; namely, process setting 19. Helps, C. R.; Strong, A. B.; Al-Zubi, R.; Lunt, B. M.; Kohkonen,
is in the start-up period to achieve optimal K. E. J Injection Molding Technol 1993, 3(4), 212–225.
process conditions for high quality, and con- 20. Kim, J. G.; Shin, I. H.; Han, W. H.; Lee, J. W. J Injection Molding
trol is to keep the optimal quality after the Technol 2000, 4(4), 201–210.
21. Rewal, N.; Toncich, D.; Friedl, C. J Injection Molding Technol 52. Gao, F.; Patterson, W. I.; Kamal, M. R. Adv Polym Technol
1998, 2(3), 109–119. 1994, 13(2), 111–120.
22. Tan, K. H.; Yuen, M. M. F. Polym Eng Sci 2000, 40(4), 956–971. 53. Rafizadeh, M.; Kamal, M. R.; Patterson, W. I. In ANTEC’97,
23. Liao, S. J.; Chang, D. Y.; Chen, H. J.; Tsou, L. S.; Ho, J. R.; Yan, 1997; pp. 582–587.
H. T.; Hsieh, W. H.; Wang, J. T.; Su, Y. C. In ANTEC’02, 2002; 54. Gao, F.; Patterson, W. I.; Kamal, M. R. Polym Eng Sci 1996,
no. 872. 36(19), 2467–2476.
24. Pro Set 200/700, Allen Bradley/Rockwell Automation, 1997. 55. Havlicsek, H.; Alleyne, A. In Proceedings of the American
25. MACO R
, Eurotherm/Barber-Colman, 2003. Control Conference, 1999; Vol. 1, pp. 171–175.
26. Xtreem XP, Milacron Inc., 2003. 56. Wang, B.; Gao, F.; Yue, P. Chin J Chem Eng 2000, 8(4), 326–331.
27. Ziemba, R. In IEEE Conference of Electrical Engineering Prob- 57. Mapleston, P. Mod Plast 1999, 29(8), 29.
lems in the Rubber and Plastics Industries, 1989; pp. 12– 58. Petrova, T.; Kazmer, D. Adv Polym Technol 1999, 18(1), 19–31.
19. 59. Dubay, R. J Injection Molding Technol 2001, 5(2), 72–79.
28. Bulgrin, T. C.; Richards, T. H. IEEE Trans Ind Appl 1995, 31(6), 60. Woll, S. L. B.; Cooper, D. J. Polym Eng Sci 1997, 37(5), 801–812.
1350–1357.
61. Woll, S. L. B.; Cooper, D. J. Polym Eng Sci 1996, 36(11), 1477–
29. Tsai, C.-C.; Liu, C.-H. IEEE Trans Ind Appl 1998, 34(2), 310– 1488.
318.
62. Kazmer, D.; Barkan, P. Polym Eng Sci 1997, 37(11), 1880–1895.
30. Liu, C.-H.; Tsai, C.-C. IEEE Trans Ind Electron 2001, 48(5),
63. Xu, G.; Koelling, K. W. In ANTEC’03, 2003; pp. 566–570.
968–975.
64. Yang, Y.; Gao, F. Polym Eng Sci 1999, 39(10), 2042–2063.
31. Shy, K.-Y.; Wang, Y.-T. In Proceedings of the World Congress
65. Orzechowski, S.; Paris, A.; Dobbin, C. J. B. J Injection Molding
on Intelligent Control and Automation (WCICA), 2002; Vol. 4,
Technol 1998, 2(3), 141–148.
pp. 2557–2560.
66. Zhao, C.; Gao, F. Polym Eng Sci 1999, 39(9), 1787–1801.
32. Havlicsek, H.; Alleyne, A. IEEE/ASME Trans Mechatron
1999, 4(3), 312–323. 67. Richard, C.; Helps, G.; Griffen, B. T. Plast Eng 1994, 50(10),
25–27.
33. Tan, K. K.; Tang, J. C. Eng Appl Artificial Intell 2002, 15(1),
6572. 68. Dubay, R.; Bell, A. C.; Gupta, Y. P. Polym Eng Sci 1997, 37(9),
1550–1563.
34. Yang, Y.; Gao, F. Control Eng Practice 2000, 8(11), 1285–1296.
69. Pramujati, B.; Dubay, R. J Injection Molding Technol 2002,
35. Tan, K. K.; Huang, S. N.; Jiang, X. IEEE Trans Control Syst
6(4), 247–258.
Technol 2001, 9(4), 663–671.
70. Dubay, R.; Bell, A. C.; Gupta, Y. P. J Injection Molding Technol
36. Tsoi, H. P.; Gao, F. Polym Eng Sci 1999, 39(1), 3–17.
1998, 2(1), 37–45.
37. Gao, F.; Yang, Y.; Shao, C. Chem Eng Sci 2001, 56(24), 7025–
71. Buja, F. J. US Patent 4904172, 1990.
7034.
72. Wenskus, J. In ANTEC’88, 1988; pp. 305–309.
38. Cheng, J.-W.; Lin, Y.-W.; Liao, F.-S. In ANTEC’03, 2003;
pp. 556–560. 73. Wenskus, J.; Miller, A. B. US Patent 5063008, 1991.
39. Lewis, F. L.; Syrmos, V. L. Optimal Control; Wiley: New York, 74. Müller, N.; Schott, N. R. J Injection Molding Technol 2000,
1995. 4(3), 120–125.
40. Clarke, D. Advances in Model-Based Predictive Control; 75. Mortizer, E.; Neumann, M. In ANTEC’99, 1999; no. 81.
Oxford University Press: New York, 1994. 76. Edwards, R.; Thomas, C. L. Polym Eng Sci 2001, 41(9), 1644–
41. Wang, K. K.; Shen, S. F.; Cohen, C.; Hieber, C. A.; Isayev, A. I. 1653.
Cornell Injection Molding Program Progress Report no. 10, 77. Baek, S. Y.; Edwards, R.; Thomas, C. L. In ANTEC’00, 2000;
Cornell University, 1984. no. 1038.
42. Huang, S.; Tan, K. K.; Lee, T. H. Polym-Plast Technol Eng 1999, 78. Nishiwaki, N.; Hori, S.; Yoshida, K.; Mineo, K. In IEEE Ultra-
38(2), 285–303. sonics Symposium, 1998; pp. 755–758.
43. Rafizadeh, M.; Patterson, W. I.; Kamal, M. R. Int Polym Pro- 79. Wang, H.; Cao, B.; Jen, C. K.; Nguyen, K. T.; Viens, M. Polym
cess 1996, 11(4), 352–361. Eng Sci 1997, 37(2), 363–376.
44. Chang, T. C. J Injection Molding Technol 2002, 6(4), 239–246. 80. Brown, E. C.; Dawson, A. J.; Coates, P. D. In ANTEC’00, 2000;
45. Sheth, B.; Barry, C. M. F.; Schott, N. R.; Higdon, R. D.; Davison, no. 592.
B. In ANTEC’01 2001; no. 306. 81. Wenskus, J. J. J Injection Molding Technol 1997, 1(3), 151–157.
46. Zheng, D.; Alleyne, A. Trans ASME. J Manuf Sci Eng 2003, 82. Rewal, N.; Toncich, D. J Injection Molding Technol 1998, 2(3),
125(1), 154–163. 109–119.
47. Agrawal, A. R.; Pandelidis, I. O.; Pecht, M. Polym Eng Sci 83. Ricketson, R. C.; Wang, K. K. In ANTEC’87, 1987; pp. 231–
1987, 27(18), 1345–1357. 234.
48. Dininger, J. In IEEE Industry Applications Society Annual 84. Liang, J. M.; Wang, P. J. J Injection Molding Technol 2002, 6(1),
Meeting 1994; Vol. 3, 2159–2164. 58–71.
49. Pierick, D.; Noller, R. In ANTEC’91 1991; pp. 252–258. 85. Liang, J. M.; Wang, P. J. J Injection Molding Technol 2002, 6(4),
50. Gao, F.; Patterson, W. I.; Kamal, M. R. Polym Eng Sci 1996, 331–342.
36(9), 1272–1285. 86. Vagelatos, G. A.; Rigatos, G. G.; Tzafestas, S. G. Expert Sys-
51. Collins, C. Assembly Automation 1999, 19(3), 197–202. tems Appl 2001, 20(2), 207–216.
87. Seaman, C. M.; Desrochers, A. A.; List, G. F. IEEE Trans 92. Wang, P. J.; Lin, J. Y. In CAE and Intelligent Process-
Systems Man Cybernetics 1993, 23(2), 414–426. ing of Polymeric Materials, 1997; MD-Vol. 79, pp. 265–
88. Kamal, M. R.; Varela, A. E.; Patterson, W. I. Polym Eng Sci 279.
1999, 39(5), 940–952. 93. Sheth, H. R.; Nunn, R. E. J Injection Molding Technol 2001,
89. Wang, P. J.; Wang, K. K. In ANTEC’91, 1991; pp. 2336–2340. 5(3), 141–151.
90. Michaeli, W.; Lauterbach, M. Adv Polym Technol 1989, 9(4), 94. Ivester, R.; Danai, K.; Kazmer, D. J Injection Molding Technol
337–343. 1998, 2(3), 103–108.
91. Sheth, H. R.; Nunn, R. E. J Injection Molding Technol 1998, 95. Cheng, J.-W.; Chao, T.-C.; Chang, L.-H. In ANTEC’03, 2003;
2(2), 86–94. pp. 392–396.