polymers-14-03551-v2

polymers
Article
Industry 4.0 In-Line AI Quality Control of Plastic Injection
Molded Parts
Saeid Saeidi Aminabadi 1, * , Paul Tabatabai 2 , Alexander Steiner 2 , Dieter Paul Gruber 2 , Walter Friesenbichler 1, *,
Christoph Habersohn 3 and Gerald Berger-Weber 4
1 Department of Polymer Engineering and Science, Montanuniversitaet Leoben, Otto Gloeckel str. 2,
8700 Leoben, Austria
2 Polymer Competence Center Leoben GmbH, Rosegger str. 12, 8700 Leoben, Austria
3 Institute for Production Engineering and Photonic Technologies, Vienna University of Technology,
Getreidemarkt 9/311, 1060 Vienna, Austria
4 Institute of Polymer Processing and Digital Transformation, Johannes Kepler University Linz,
Altenberger str. 69, 4040 Linz, Austria
* Correspondence: s.saeidi-aminabadi@stud.unileoben.ac.at (S.S.A.);
walter.friesenbichler@unileoben.ac.at (W.F.)
Abstract: Automatic in-line process quality control plays a crucial role to enhance production effi-
ciency in the injection molding industry. Industry 4.0 is leading the productivity and efficiency of
companies to minimize scrap rates and strive for zero-defect production, especially in the injection
molding industry. In this study, a fully automated closed-loop injection molding (IM) setup with a
communication platform via OPC UA was built in compliance with Industry 4.0. The setup included
fully automated inline measurements, in-line data analysis, and an AI control system to set the new
machine parameters via the OPC UA communication protocol. The surface quality of the injection
molded parts was rated using the ResNet-18 convolutional neural network, which was trained on
data gathered by a heuristic approach. Further, eight different machine learning models for predicting
Citation: Aminabadi, S.S.; Tabatabai,
the part quality (weight, surface quality, and dimensional properties) and for predicting sensor data
P.; Steiner, A.; Gruber, D.P.;
were trained using data from a variety of production information sources, including in-mold sensors,
Friesenbichler, W.; Habersohn, C.;
injection molding machine (IMM) sensors, ambient sensors, and inline product quality measure-
Berger-Weber, G. Industry 4.0 In-Line
AI Quality Control of Plastic Injection
ments. These models are the backbone of the AI control system, which is a heuristic model predictive
Molded Parts. Polymers 2022, 14, 3551. control (MPC) method. This method was applied to find new sets of machine parameters during
https://doi.org/10.3390/ production to control the specified part quality feature. The control system and predictive models
polym14173551 were successfully tested for two groups of quality features: Geometry control and surface quality
control. Control parameters were limited to injection speed and holding pressure. Moreover, the
Academic Editor: Luigi Sorrentino
geometry control was repeated with mold temperature as an additional control parameter.
Received: 1 August 2022
Accepted: 22 August 2022 Keywords: injection molding of plastics; closed-loop quality control; in-line quality control; AI
Published: 29 August 2022 quality control; predictive control; deep neural network; deep residual learning; surface quality
Publisher’s Note: MDPI stays neutral prediction; dimensional features prediction; weight prediction
with regard to jurisdictional claims in
published maps and institutional affil-
iations.
1. Introduction
The plastics industry is an increasingly demanded market which is known as the third
largest manufacturing industry worldwide with a prospective market of 750.1 billion U.S.
Copyright: © 2022 by the authors.
dollars in 2028 [1]. Most of the products are combined with plastic parts, in which the
Licensee MDPI, Basel, Switzerland.
injection molding manufacturing process contributes approximately 80% of this production
This article is an open access article
distributed under the terms and
in the modern plastics industry [2]. Thus, efficient production in such a big industry is
conditions of the Creative Commons
inevitable [3].
Attribution (CC BY) license (https://
Digitized production will dramatically change the value chain of most industrial
creativecommons.org/licenses/by/ producers. In Europe, this goal is referred to as Industry 4.0, while in the USA it is focused
4.0/). on by the Industrial Internet Consortium (IIC), in Asia by the Industrial Value-Chain
Polymers 2022, 14, 3551. https://doi.org/10.3390/polym14173551 https://www.mdpi.com/journal/polymers

Polymers 2022, 14, 3551 2 of 24
Initiative (IVI), and in China by the initiative “Made in China 2025” [4]. There is an
increasing demand for inline and real-time control of the quality of injection molded parts.
These requisitions are defining the term ‘injection mold as a cyber-physical system’ in the
literature on injection molding technology. This subject designates the requirements for the
inline quality feature measurements, machine learning for quality feature prediction, and
smart adaptive control of the injection molding process. [5–7].
Plastic injection molding (PIM) is a non-linear and very complex process with very
dependent process parameters characterizing the quality of the produced parts [8]. Material
selection, part geometry, mold design, and process parameters are recognized as the
effective parameters for the final quality of the part.
1.1. Process Parameters

The quality of the plastic molded part depends on the process parameters in addition
to the part design [9]. The subject of which parameters affect the PIM process significantly
is the first question to be answered. The significance of process parameters’ effects depends
on the expecting quality features. Therefore, a common way is to study the significance of
the process parameters by screening before starting a design of experiment (DOE) [10–12].
Some researchers select a wide range of process parameters and sensors initially, then,
exclude the least significant variables to reduce the number of required experiments [13].
Dimensional properties of the part and the affecting parameters have been one of
the most important quality features in PIM [14–17]. Ferreira Alves [18] and Kapadia [19]
provided a comprehensive study on shrinkage and warpage behavior. The history of the
most important parameters affecting the shrinkage of the part (or dimensional properties)
was studied by Annicchiarico and Alcock [20]. They investigated material properties,
processing parameters, and design effects on the shrinkage of the part and concluded the
process parameters as the main resource for affecting the shrinkage. The most important
process parameters were recognized to be holding pressure, mold temperature, and melt
temperature [20]. Recently, many researchers studied shrinkage and warpage of plastic
molded parts [19,21–26]. In addition, Nurul Hidayah Mohamad Huzaim et al. provide a
review on rapid tooling for heat cycle molding [27], while Tim Evens et al. [28] studied the
replication of microneedles, in other words, micro-molding. Both technologies can signif-
icantly alter the specific effects of injection molding process parameters on dimensional
stability. However, in our study, we analyzed a macroscopic part and used conventional
mold temperature control.
Selvaraj et al. [29] extracted a list of defects-causes-preventions from the works of
literature showing the significant variables over each defect. Among the list, sink marks,
weld lines, and flow lines are the relevant defects to this paper. The significant process
variables on these defects were reported by Selvaraj et al. as holding pressure, injection
speed, and mold temperature.
We consider the weight of the part as a pertinent variable to both other quality features
(dimensional properties and surface quality). Therefore, holding pressure time is also
considered as affecting the process parameter on the weight of the part.
1.2. Process Sensors

The mold as the heart of the injection molding process and part quality provides the
most important information about the process. The type and position of the sensors to
monitor the process are important topics to be studied. Researchers use in-mold sensors
universally [13,30,31]. Florian [32], Ageyeva et al. [33], and Zhao et al. [2] provided com-
prehensive reviews of the in-mold sensors for injection molding. The multivariable sensor
(MVS) is a prevalent sensor in injection molding research. We exerted three MVS sensors in
our mold (described in Section 3.3.3) to monitor the behavior of melt during the injection.
Chang et al. [34] propose a real-time online monitoring method to predict the weight
of injection molded products made of recycled materials. We used some features of
Polymers 2022, 14, 3551 3 of 24
the clamping force since tie-bars elongation is measured as the clamping force by our
experiment injection molding machine.
1.3. Application of Artificial Intelligence (AI) in PIM

The application of artificial neural networks (ANN) in the injection molding of plastics
has been undertaken by many researchers since 1995 [35–37]. S. L. Mok et al. reviewed the
studies on the process parameters optimization approaches till 1998 [36].
Where support vector machines have been rarely used [38], types of ANNs have been
employed widely for the prediction of quality features such as warpage [39], weight [40],
shrinkage [41], or cycle time minimization [42].
Ogorodnyk et al. [43] investigated ANN methods and decision trees versus the number
of input features. Gim and Rhee [44] employed MLP neural network to predict the weight
based on the given features extracted from cavity pressure for part weight optimization.
Convolutional neural networks (CNNs) and recurrent neural networks (RNNs) were
employed by Nagorny et al. [31] recently. They applied the CNN on thermographic
images for part dimension (width) prediction. Despite very few amount of samples (only
177 samples for training and 27 samples for testing), the CNN worked with an accuracy of
87 percent. Guo et al. [45] trained ANN with a reinforcement learning (RL) decision system
to optimize the dimensions of an ultrahigh-precision lens product.
We listed the previous modeling and optimization approaches in another study to
be support vector machine (SVM), least-square support vector machine (LSSVM), genetic
neural fuzzy system (GNFS), fuzzy neural network (FNN), back propagation neural net-
work (BPNN), a combination of genetic algorithms and neural network (GA-NN), deep
restricted Boltzmann machine (DRBM), Feedforward neural network (FFNN), MLP, PSO-
based BPNN, FFBPNN, hybrid neural network (HNN), Decision trees, CNN, radial basis
function (RBF), RL, and Forgetting Factor Recursive Least Square Algorithm (FF-RLS) [4].
Given a review, we conclude that multiple AI solutions may provide an acceptable
accuracy of predictions. The application of the AI method type not only depends on the
accuracy of the predictions but also on the number of required samples and the time of
calculations. Generalizability, transferability, and reversibility of the predicting models are
the other important challenges nowadays in this field.
We investigated the applicability of the central composite design, simple regressions,
and AI models for plastic injection molding in simulations during a prior study [4] resulting
that the ANNs were superior in weight and dimensional properties predictions.
1.4. Non-Iterative Optimization Methods

Dang [46] classified the optimization methods into two groups of Meta model-based
(for AI methods) and direct discrete methods (for regression), whereas, Zhao et al. [2]
categorize the optimization methods into iterative and non-iterative methods. Design of
experiment (DOE) as a non-iterative method is used traditionally in research and industry
for finding an optimal injection molding process setting in an efficient non-iterative way.
In injection molding of plastics, it takes 10 to 50 shots to a stable point of production
based on the changes in the machine parameters, and it makes numerous experiments
unaffordable [11]. A DOE method is selected based on the complexity of the problem and
the affordable amount of experimental production.
The Taguchi method has been used widely by researchers [40,41,47–49]. Taguchi’s
designs are highly fractionated which makes them very attractive to experimenters [50].
Some researchers use the Taguchi method initially for screening the process variables
because of a lower number of experiments [51,52].
Provided that Taguchi is useful to compromise on the number of experiments finding
the optimal parameter, the response surface method (RSM) involves more experimentation
to get closer to the global optimum. There are two popular classes of RSM designs: central
composite design (CCD) and Box-Behnken design (BBD). CCD studies five levels for
each experimental variable, which contain two-level factorial or full factorial designs, an
Polymers 2022, 14, 3551 4 of 24
additional two axial (star) points for each variable, and a replicated center point [10,50].
Therefore, we have employed a CCD, type inscribed (CCI), in this paper.
1.5. Iterative Optimization Methods

Numerous researchers applied iterative optimization methods to find the optimal
point of production variables, on which Zhao et al. [2] gave a review. Few of them tried to
develop an adaptive online process control strategy. The adaptive control methodologies
were conventionally adaptive PID [53] to fuzzy-PID [54,55] and Gray fuzzy PI [56].
Chen and Tung [37] reviewed the history of modeling and closed-loop techniques in
2005. Ogorodnyk [43] provided a general review of the closed-loop control process and
AI application for plastic injection molding (PIM) in 2018. Selvaraj et al. [29] presented a
comprehensive review of AI applications in PIM recently. We listed the in-line iterative
methods for PIM [4] to be Genetic Algorithm (GA), Particle Swarm Optimization (PSO), In-
vasive Weed Optimization (IWO), Fuzzy Logic, Reverse Neural Network, Model Predictive
Control (MPC), and Sequential Approximation Optimization (SAO), and Dynamic Ma-
trix Control (DMC). However, few researchers tried to implement a real-time closed-loop
quality control for plastic injection molded parts [55,57,58].
Among all, there are attempts such as Johnston et al. [13] and Hopmann et al. [59] using
innovative methods, combining iterative methods with AI models. As a recent example,
Tsai et al. [51] developed an adaptive online control system to maintain the consistency of
the weight as the quality feature employing BPNNs for prediction.
In this research, we develop a heuristic model predictive control algorithm (MPC) to
predict a new set of machine parameters for the desired quality feature in a real-time and
in-line way.
1.6. OPC UA and Industry 4.0

The key technologies of Industry 4.0, e.g., the Internet of Things, IPv6 and OPC UA,
cloud services, big data and artificial intelligence, virtual technologies, and intelligent
sensors and actuators are intended to enable manufacturers to remain competitive in the
global economy [4]. In November 2016, the Industry 4.0 Platform published a checklist
for classifying and advertising products as Industry 4.0 “Basic,” “Ready,” or “Full”. To
comply with the “Industry 4.0 communication” criterion, even the lowest category requires
the product to be addressable over the network via TCP/UDP or IP and to integrate at
least the OPC UA information model [60]. EUROMAP 77 describes the interface between
injection molding machines (IMM) and manufacturing execution systems (MES) for data
exchange [61]. While OPC UA provides the technology for the transfer of information,
EUROMAP 83 defines the definition of which information is transferred and in which
form [62].
Therefore, the application of OPC UA platform as an important part of Industry 4.0
was inevitable in our research. We employed three network architectures of the OPC UA
platform for three purposes-communication between IMM and stations, sampling the data
from IMM and peripherals, and manipulating the closed-loop control. The details of the
OPC UA networks are described in Section 2.2.
1.7. Research Objective

Our vision is to improve the performance of industrial manufacturing processes
for plastics products in terms of product quality as well as process flexibility, efficiency,
reliability, and performance through comprehensive process analysis, modeling, control,
and digitization. We consider an intelligent injection molding tool to be an indispensable
building block in this process: By continuously monitoring the process history in the mold
cavity, injection molding machine, and ambient conditions, disturbances can be recognized
in an early stage.
The process is modeled using artificial intelligence (AI) for individual quality factors
during a training phase. The AI knows the relationships between process parameters,
Polymers 2022, 14, 3551 5 of 24
sensor signals, and molded part quality, and uses this knowledge to promptly recognize,
evaluate and react adaptively to disturbances combined with a model predictive control
(MPC) algorithm to minimize scrap and defective injection molded parts.
2. Materials and Methods

2.1. Plastics Injection Molding Process
Injection molding of plastics is a cyclic process for manufacturing plastic components.
There are different types of injection molding processes, however, this process generally
includes plasticization (metering), injection (filling), packing (holding), cooling, and ejecting
(demolding). The polymer melts from solid granules through the combination of heat
conduction from the heated barrel and heat caused by molecular deformation with the
rotation of the screw (shear heating). By closing the mold (clamping), the polymer melt
is forced from the barrel (cylinder) into the mold cavity(s) during the injection stage. The
molten resin travels through the gate(s), and throughout one or more mold cavities where
it will form the desired product(s) shape in the mold. Due to the heat conduction between
the mold walls and the melt, the melt temperature decreases, and solidification progresses.
Afterward, the cooling stage provides additional time for the resin in the cavity to solidify
and become sufficiently rigid for ejection. Finally, the molding machine actuates the
necessary cores, slides, and pins to open the mold and eject the molded part(s) [63,64].
2.2. OPC UA Communication Platform

Each peripheral device or measurement device comprises an OPC UA server to
provide its information and data for the clients. The communication platform is developed
based on three main OPC UA clients.
First, A BECKHOFF® IPC model CP6600 (BECKHOFF Automation GmbH, Buers,
Germany) is programmed for controlling the axes of the cylindrical measurement system
and communications with the other peripheral devices. This IPC is the center of the au-
tomation process and communicates with the DAQ center, measurement devices, injection
molding machine, and robot. An OPC UA server on this IPC provides the required data
(about dimensional measurement status, camera capturing status, and cylindrical moving
system status) for communication with the DAQ center (computer). Simultaneously, an
OPC UA client on this IPC checks the status of the DAQ center programs and manages the
measurements in full compliance with the software on the DAQ center for storing the data.
Additionally, this IPC communicates with the robot through digital I/Os (input/output) to
guarantee the robot’s safe movements. The OPC UA client on this IPC reads out the status
of the injection molding machine, the position of the clamping unit and ejector6pins,
Polymers 2022, 14, x FOR PEER REVIEW of 24 and
the maximum stroke, and manages the permissibility of robot entrance and movements in
the machine and mold area. This platform is illustrated in Figure 1.
OPCUA
Figure1.1.OPC
Figure UAplatform
platformforfor management
management of of
thethe measurements
measurements andand
the the process
process automation.
automation.
Second, the OPC UA client on the DAQ center is responsible for the acquisition of
the injection molding machine data and its peripherals, see Figure 2. The program of this
OPC UA client was provided by WITTMANN BATTENFELD GmbH (Kottingbrunn, Aus-
tria).
OPC UA client was provided by WITTMANN BATTENFELD GmbH (Kottingbrunn, Aus-
Figure 1. OPC UA platform for management of the measurements and the process automation.
tria).
Polymers 2022, 14, 3551 6 of 24
Second, by
OPC UA client was provided theWITTMANN
OPC UA client on the DAQ center is
BATTENFELD responsible
GmbH for the acquisition
(Kottingbrunn, Aus-of the
injection molding machine data and its peripherals, see Figure 2. The program of this OPC
tria).
UA client was provided by WITTMANN BATTENFELD GmbH (Kottingbrunn, Austria).
Figure 2. OPC UA communications platform to collect data from injection molding machine and
peripherals.
Third, the controller synchronizes and applies the outputs of the control process over
theFigure
OPC2.UAOPC platform. The
Figure DAQ
2. OPC
UA communications UAcenter
platform analyses
communications
to collect the
platform sensors
to collect
data from andinjection
datainjection
from measurements
moldingmachine
molding machinedata
and after
peripherals.
and
each cycle of production. The results of the analyses are uploaded onto the ownCloud,
peripherals. Third, the controller synchronizes and applies the outputs of the control process
afterward. The controller
over theOPC
OPC UA client detects
UA platform. The DAQ the center
availability
analysesof thethe dataand
sensors on measurements
the own-
Cloud Third,
and after processing
the controller
data after the
eachdata,
synchronizes theand
cycle of control
applies
production. process
theresults
The takes
outputs place.
of
of the the Theare
control
analyses new set of
process
uploaded ma-the
over
onto
chine parameters (controlling parameters) is written via controller OPC UA client todata
the OPC UA ownCloud,
platform. The afterward.
DAQ center The controller
analyses OPC
the UA
sensors client
anddetects the availability
measurements of
data the
afterthe on
each cycle the ownCloud
of production. and
The resultsafter processing the
of theinanalyses data, the control
are uploadedThe process takes place.
ontodiagram The
the ownCloud, new set of
injection molding machine OPC UA server, case of requirement. is shown
machine parameters (controlling parameters) is written via controller OPC UA client to the
in afterward.
Figure 3. TheTheentire
controller
process
injection
OPC of UA
molding
client detects
measurement,
machine OPC UA datatheanalysis,
server,
availability
in case of data ofprocessing,
the data
requirement.
onandthecontrol
The diagram
own-
is shown
Cloud and after
decision-making takes processing the
place3.within
in Figure data,
one
The entire the control
production
process process takes place. The new set of
cycle. data analysis, data processing, and control
of measurement, ma-
chine parameters (controlling
decision-making parameters) is written
takes place within via controller
one production cycle. OPC UA client to the
injection molding machine OPC UA server, in case of requirement. The diagram is shown
in Figure 3. The entire process of measurement, data analysis, data processing, and control
decision-making takes place within one production cycle.
Figure 3. OPC UA
Figure 3. OPC UA communication communication
platform platform for synchronizing
for synchronizing the control theprocess.
control process.
2.3. Quality Feedback

In our research, in addition to the weight, dimensional properties, and surface quality
are considered to be the crucial quality features for in-line quality control.
Figure 3. OPC UA communication platform for synchronizing the control process.
2.3.1. Weight
A Sartorius scale model Sartorius MSA 2203S-100-DR (Sartorius Lab Instruments
GmbH & Co. KG, Goettingen, Germany) measures the weight of the parts with an accuracy
of 0.001 g (see Figure 4). A software written in LABVIEW (scale software) communicates
with the scale over the RS232 serial protocol. Scale software manages the measurement
process and communications with the robot for placing and picking the part. A fixture is
positioned on the scale to guarantee a repeatable position for the part during the robot
entry. Since the scale is very sensitive to air movements in the measurement room, the scale
software manages the stability of the scale, before and after part placement.
GmbH & Co. KG, Goettingen, Germany) measures the weight of the parts with an accu-
racy of 0.001 g (see Figure 4). A software written in LABVIEW (scale software) communi-
cates with the scale over the RS232 serial protocol. Scale software manages the measure-
ment process and communications with the robot for placing and picking the part. A fix-
ture is positioned on the scale to guarantee a repeatable position for the part during the
Polymers 2022, 14, 3551 7 of 24
robot entry. Since the scale is very sensitive to air movements in the measurement room,
the scale software manages the stability of the scale, before and after part placement.
Figure 4.
4.The
Thescale
scalefor
formeasuring
measuringthe weight,
the thethe
weight, dimensional measurement
dimensional system,
measurement and the
system, andcamera
the camera
for surface quality inspection.
for surface quality inspection.
2.3.2. Dimensional
DimensionalProperties
Properties
AA special
specialthree-dimensional
three-dimensionalcylindrical
cylindrical measurement
measurement system waswas
system builtbuilt
to measure
to measure
the dimensional
dimensionalproperties
propertiesofofthethemolded
molded part
part(illustrated in in
(illustrated Figure 4) with
Figure ±0.005
4) with mm mm
±0.005
precision. Details
precision. Detailsofofthis
thismeasurement
measurement system
system werewerepublished
publishedin our paper
in our [5] recently.
paper [5] recently.
2022, 14, x FOR PEER REVIEW
Due to
Due to in-line
in-lineand
andreal-time
real-timemeasurements
measurements (a (a
complete
complete measurement
measurement should be finished
should be finished
within a production cycle), only the lengths of three linear and three rotary
within a production cycle), only the lengths of three linear and three rotary profiles profiles of the
of the
molded part were measured (see Figure 5). Later, the controller makes an averageeach
molded part were measured (see Figure 5). Later, the controller makes an average of of each
three-line group
three-line groupfor
forsimplicity
simplicityofofthe
thecontrol
control process.
process.
Figure 5. The position of the six scanning lines.

Figure 5. The position of the six scanning lines.
2.3.3. Surface Inspection

For the aspect of surface inspection, extensive expe
several previous research projects, inspection methods f
Polymers 2022, 14, 3551 8 of 24
2.3.3. Surface Inspection

For the aspect of surface inspection, extensive experiences could be drawn upon. In
several previous research projects, inspection methods for complex 3D components were
developed at the Polymer Component Center Leoben GmbH (PCCL) [65–67]. These include
methods that can cope with very difficult reflective or highly reflective surfaces [68–70].
In our research, a monochrome vision camera (AVT Mako U-130B) was installed at
an angle of about 45◦ to the horizon looking to the center of rotations. An LED-bar light
source with a length of 20 cm illuminated the center of the part under the camera (shown
in Figure 4). The camera recorded 250~300 images during both rotations for a complete
surface scan of the part.
The captured pictures from the surface of the part were analyzed and quantified
for three groups of surface defects: flow line thickness, streaks, and sink marks. For
this purpose, an offline model-based supervised learning method was developed using
the produced samples from the preliminary experiments [71]. The levels of the holding
pressure were defined as the numeric classes (labels) and a multi-class classifier based on
ResNet-18 [72] was trained. To transform the pressure-level classification result of a model
PEER REVIEW 9 of 24
into continuous numerical results, the following equation was applied:
QS = (CP − min(CP ))/(max(CP ) − min(CP )), (1)
Table 1. The time duration QS iseach

where for the surface qualitystation
automation value defined 0 and 1, and CP is the class of the
betweentasks.
and parallel
pressure level. It’s defined so that a higher value of QS indicates good quality and a lower
on value means aUnit
Duration surface quality containing large effects of flowActions
Parallel lines, streaks, and sink marks.
measurement 44 2.4. Process Automation

Second Robot handling and weight measurement
a scan 9 A KUKA Second
robot (KUKA CEE Robot
GmbH, handling and weight
Steyregg, Austria) measurement
was programmed to handle
the part between the measurement stations. After taking the part from the open injection
asurement 8 mold, the partSecond Robot handling
was tested initially for a possible rotation during ejection using a fork sensor,
andling 42 then placed Second Measurements
on the scale for weight measurement. and the
Afterward, part production
robot handles the part to
the cylindrical measurement system for dimensional measurements and surface inspection.
of the part 76~95 Second A complete measurement process
The sequence diagram is shown in Figure 6 and the procedure is illustrated in Figure 7.
Figure 6. AutomationFigure
process sequenceprocess
6. Automation diagram.
sequence diagram.
Since the time consistency of the production is important, the part manipulations and
measurements should be accomplished during one cycle of injection molding production.
The table of robot station actions and parallel movements is given in Table 1 (also see
Figure 6).
Polymers 2022, 14, 3551 9 of 24
Figure 6. Automation process sequence diagram.
Figure 7. The procedure of handling and measurement of the injection molded part [5].
Figure 7. The procedure of handling and measurement of the injection molded part [5].
Table 1. The time duration for each automation station and parallel tasks.
2.5. AI Models
Action TheDuration
process was modeled Unit for two groups of outputs.
Parallel Actions
The first group is the quality
Dimensional measurementfeatures including
44 surface Second Robot
quality, weight, and handling and weight
dimensional measurement
properties, which consist of
Camera scan the length of 9 6 profiles of the part. The second
Second group
Robot is the
handling andsensor
weightdata including the dig-
measurement
Weight measurement ital sensor data
8 and analog sensor
Second data. Robot handling
Robot handling Therefore,
42 two groups of models
Second were trained. The
Measurementsfirst
andgroup includes five regres-
part production
Production of the part
sion ANNs consisting
76~95
of four hidden
Second
linear layers, where the first three
A complete measurement process
layers were acti-
vated with a ReLU function. The input layers of ANNs were considered as five machine
parameters based on the preliminary experiment (5 inputs), the time difference from the
2.5. AI Models
start of production on the same day of the experiment to each sample production (one
The process was modeled for two groups of outputs. The first group is the quality
input), and ambient sensors data (19 inputs). The predicted outputs of each neural net-
features including surface quality, weight, and dimensional properties, which consist of the
work
lengthwere:
of 6 profiles of the part. The second group is the sensor data including the digital
1.sensor
Surface quality
data and (single
analog sensoroutput),
data.
Therefore,
2. Weight two groups
(single output),of models were trained. The first group includes five regression
3. Profile lengths on the part linear
ANNs consisting of four hidden layers, where the first three layers were activated
(six outputs),
4. Analog sensor data (163 outputs), ANNs were considered as five machine parame-
with a ReLU function. The input layers of
ters based on the preliminary experiment (5 inputs), the time difference from the start of
5. Digital sensor data (136 outputs).
production on the same day of the experiment to each sample production (one input), and
ambient sensors data (19 inputs). The predicted outputs of each neural network were:
1. Surface quality (single output),
2. Weight (single output),
3. Profile lengths on the part (six outputs),
4. Analog sensor data (163 outputs),
5. Digital sensor data (136 outputs).
The errors of the outputs were calculated by the mean squared error, and the networks
were trained using the Adam optimization algorithm. The second group of models was
developed based on random forest regression models. Random Forest is a classification and
regression algorithm based on the creation of multiple Decision Trees [73]. The inputs for
the second group of the models were analog sensor data (163 inputs), the time difference
from the start of production on the same day of the experiment to each sample (one input),
and the ambient sensors data (19 inputs). Predicted outputs were:
1. Surface quality
2. Weight
(oneinput),
(one input),and andthetheambient
ambientsensorssensorsdata data(19(19inputs).
inputs).Predicted
Predictedoutputsoutputswere: were:
1. Surface quality
2.1. Surface
1. Surfacequality
Weight quality
3.2. Weight
2. Weight
Six lines’ lengths
3. Six
3. Six lines’ lines’lengths
lengths
Polymers 2022, 14, 3551 10 of 24
2.6. Model Predictive Controller (MPC)
2.6.Model
2.6. ModelPredictive
PredictiveController
Controller(MPC) (MPC)
The system in the injection molding process is a combination of an injection molding
The
The system
system inin the
3.
the
machine (IMM), material, and mold. injection
Six lines’
injection molding
lengths
molding process
Theprocess
qualityisofisaathe
combination
combination
production ofofis
anan injectionmolding
injection
evaluated molding
through
machine(IMM),
machine
measurements, (IMM), material,
material,
or predictions andbased
and mold.on
mold. The
The quality
quality
the sensor ofofdata
theproduction
the production
acquired isisevaluated
during evaluated through
through
the production
2.6. Model Predictive Controller (MPC)
measurements,
measurements,
(Figure 8). orpredictions
or predictions
The system
based
based in the
on
on thesensor
the
injection
sensor
molding
data
data acquired
acquired
process
during
during
is a combination
theproduction
ofthe
production
an injection molding
(Figure
(Figure We8). 8).
trained the machine
AI models (IMM), onmaterial,
the quality
and mold.features using
The quality a production
of the preliminary experiment
is evaluated through
data We We
set and trained
trained the AI models
measurements,
the AI modelssystems
the measurement on
or the quality
predictions
on the quality basedfeatures
on the
features
were applied using
sensor a
data
using a to
in parallel preliminary
acquired
preliminary during experiment
the
investigateexperiment production
the coher-
(Figure 8).
data
data
ence set set and
and the
between the measurement
themeasurement
predictions systems
systems
and actual were
were applied in parallel
applied in parallel
measurements. The block to investigate
to investigate thecontrol
the coher-
coher-
We trained the AI models on the quality features usingdiagram
a preliminaryof experiment
the data
encebetween
ence
system between thepredictions
the
is illustrated predictions
in and theand
set Figure and
9. An actual
actual
measurement measurements.
measurements.
algorithm
systemsdecides
were applied The
onThea in
newblock
block diagram
diagram
machine
parallel ofofthe
the
parameters
to investigate control
thecontrol
cal-
coherence
systemisisbased
system
culations illustrated
illustrated
on the inquality
in Figure
Figure
between 9.9.An
the error,Analgorithm
predictions algorithm
the and
difference decides
actualdecides onaathe
on
measurements.
between new
new
The machine
machine
block
predicteddiagram parameters
parameters
andof the control
actual cal-
cal-
system
qual-
is illustrated in Figurethe9. An algorithm decides on athe
culations
culations
ity values,based based
and theoncycle
on the
the quality
quality
counter. error,
error,Thethe difference
difference
block diagram between
between thenew
of this process
machine parameters
predicted
predicted
is shown andin
and actual calculations
actual
Figure qual-
qual-
10.
based on the quality error, the difference between the predicted and actual quality values,
ityvalues,
ity values,and andthe
thecycle
cycle
and the
counter.
counter. Theblock
The
cycle counter.
blockdiagram
The block
diagram of this
thisprocess
diagramofofthis
process
process isisshown
is shown
shown ininFigure
in Figure
Figure
10.
10.
10.
Figure 8. The general block diagram for an injection molding process.

Figure8.8.The
Figure Thegeneral
generalblock
block diagram
Figurediagram
8. forblock
for
The general aninjection
an injection molding
molding
diagram for process.
process.
an injection molding process.
Figure 9. The block diagram of the control system.

Figure9.9.The
Theblock
blockdiagram
diagram The
Figure 9.of block
the diagram
control of the control system.
system.
Figure of the control system.
Figure 10. The controller block diagram in detail.

Figure 10. The controller block diagram in detail.
Figure10.
Figure 10.The
Thecontroller
controllerblock
blockdiagram
diagraminindetail.
detail.
The control system was built based on the AI models. To simplify the controlling
The
The
process control
control
for systemwas
system was
the dimensional builtbased
built
control basedon
onthe
problem, theAI
the AImodels.
models.
profile Tosimplify
To
lengths simplify
were thecontrolling
the
averagedcontrolling
into two
processfor
process forthe
thedimensional
dimensionalcontrol
controlproblem,
problem,the
theprofile
profilelengths
lengthswere
wereaveraged
averagedinto
intotwo
two
Polymers 2022, 14, 3551 11 of 24
The control system was built based on the AI models. To simplify the controlling
process for the dimensional control problem, the profile lengths were averaged into two
groups linear line dimension and rotary line dimension, i.e., the bow length. The designed
controller is capable of controlling each of the quality features, “surface quality”, “line
dimension”, and “weight”.
2.6.1. Controller Algorithm

A short summary of some of the computational methods is described in our other
report [71]. The surface quality, weight, and dimensions of the part are denoted by Q1, Q2,
and Q3. as the outputs of the first group of neural network models which were described
in Section 2.5. These networks were used to perform a grid-search sub-routine of the
controller, described in Algorithm 1.
2.6.2. Extending Grid Search

Since model inference was very fast (depending on the discretization step size), a full
enumeration of all possible vectors of the machine parameters was possible. However,
specifying certain combinations of the machine parameters reduces the grid search time
significantly in cases where the search space is very large. Additionally, some machine
parameters (like melt temperature) are known as the slow-changing parameters which
require larger n shots between the controller actions to observe the controlling effect
significantly in the output quality. Therefore, the controlling parameters were limited to
injection speed and holding pressure (and mold temperature in one experiment) in the
experiments of this research.
Algorithm 1: Grid search

INPUT: The current machine parameters M, the current machine run time r, the vector E for
ambient sensor data, allowed machine parameters ranges, and a preference function f which rates
the desirability of a prediction goal and new machine parameters.
OUTPUT: New machine settings M’best
FOR each machine parameters vector M’ so that each element is within the defined range DO
| Compute the predicted surface quality s = Q1 (M’, r, E)
| Compute the predicted weight w = Q2 (M’, r, E)
| Compute the predicted dimensions d = Q3 (M’, r, E)
| IF the current loop iteration is the first or if f(s, w, d, M’, M) > f(sbest , wbest , | dbest ,
M’best , M) THEN
| | Set M’best = M’.
| | Set sbest = s.
| | Set wbest = w.
|_ |_ Set dbest = d.
return M’best
2.6.3. Controller Confidence Scaling

We denote the neural network regressions by S1 and S2 for predictions of analog and
digital sensor data, respectively, with the same input data as the models Q1 , Q2 , and Q3 .
Then, the confidence of the machine parameters prediction could be defined based on the
accuracy of S1 and S2 (analog and digital sensor data) predictions. The result could be
applied as a proportional delta value to scale the controller’s suggested machine parameters.
The MPC controller is described in Algorithm 2.
Polymers 2022, 14, 3551 12 of 24
Algorithm 2: Controller
INPUT: Quality threshold for the quality goal (surface quality, weight, or dimensions), integer n
for the minimum shot distance between the controller actions, and the reference function f as
described in Algorithm 1.
OUTPUT: No output, the algorithm runs perpetually.
WHILE forever DO
| Part Production
| |_ A new part is produced
| Quality Measurement
| | The surface quality smeasured is inspected using CNN.
| | The weight wmeasured is measured using the scale.
| |_ The dimensions dmeasured are measured using the cylindrical dimension
| measurement system.
| Quality Prediction
| | Let M be the vector of current machine parameters, r the running time of
| | the machine, E the vector of current ambient sensor data, and A the vector
| | of current analog sensor data.
| | The surface quality spred is predicted by R1 (M, r, E, A).
| | The weight wpred is predicted by R2 (M, r, E, A).
| |_ The dimensions dpred are predicted by R3 (M, r, E, A).
| Quality Control
| | IF the measured quality value is outside of the given threshold THEN
| | | IF there has been a machine parameters update within the last n shots THEN
| | | |_ Terminate this loop
| | |IF the measured quality deviates strongly from the predicted quality THEN
| | | |_ Warn the user: an external factor might influence the part quality.
| | |_Use Algorithm 1 to calculate the new machine parameters based on M,
| | r, E, and f.
|_ |_ |_Set the new machine parameters.
3. Experimental Setup
3.1. Experiment Devices and Material
Standard white ABS material (acrylonitrile butadiene styrene) Novodur HH-112 from
INEOS Styrolution is selected as a popular material in the injection molding industry.
Masterbatch Renol-black CAV 80036 from Clariant Masterbatches GmbH was added by
DOSIMAX color dosing unit with a ratio of 4% of the metering material into the hopper end.
An EcoPower 110 injection molding machine with a screw size of 35 mm from WITTMANN
BATTENFELD GmbH was used to process the material. The mold temperature was
controlled by a combination of two TEMPRO plus D180 and D90 for fast and wide-ranging
mold temperature control (between 80 and 110 degrees Celsius). A KUKA robot KR 5 arc
was programmed for handling the molded part between the injection molding machine
and stations. The setup is illustrated in Figure 11.
3.2. Case Study

The part is a highly reflective and partially cylindrical-shaped sample with the cavity
dimensions of L = 120.2 mm, outer radius r = 123 mm, curve (bow) length C = 125.36 mm,
and arc angle α = 58.39 ◦ , as illustrated in Figure 12. The post-molded volumetric shrinkage
ratio for ABS (Acrylonitrile Butadiene Styrene) is expected to be 0.4 to 0.7%. The aggregation
of black color, very low surface roughness, and cylindrical shape cause the problem as the
hard example for surface inspection and surface quality control.
3.3. Process Sensors

3.3.1. Digital Sensor Data
The data from the injection molding machine, DRYMAX dryer, and TEMPRO plus
devices were sampled over the OPC UA network platform. The injection molding machine,
DRYMAX dryer, and TEMPRO plus D devices have each an OPC UA server inside which
Polymers 2022, 14, 3551 13 of 24
provides the information of the internal sensor and variables to be read or written. A
program named DataRetriever provided by WITTMANN BATTENFELD reads the data
Polymers 2022, 14, x FOR PEER REVIEW 13 of 24
from the OPC UA servers with a sampling rate of 60 Hz for every cycle of production.
These data are called digital sensor data.
Figure 11. Experimental setup.
3.2. Case Study

The part is a highly reflective and partially cylindrical-shaped sample with the cavity
dimensions of L = 120.2
Figuremm,
Figure 11. outer radius
11. Experimental
Experimental setup.r = 123 mm, curve (bow) length C = 125.36 mm,
setup.
and arc angle α = 58.39 °, as

3.2. Case
illustrated in Figure 12. The post-molded volumetric shrink-
Study code extracted a set of 136 features from the digital sensor data. Each
A python
age ratio for ABS (Acrylonitrile
The part
extracted is aButadiene
feature highly
is a single
Styrene)
reflective
valueand
that
is expected
partially to be 0.4
cylindrical-shaped
represents
to 0.7%.
sample
a characteristic withof
property
The
the ag-
cavity
the process.
gregation of black color,
These very
dimensions low
features =surface
of Lwere,120.2
formm, roughness,
outer radius
instance, rand
maximum, cylindrical
= 123average,
mm, curve shape
(bow)
integral, cause
length the prob-
C = 125.36
difference, time mm,
of action,
and
lem as the hard example arc
length forangle α
surface
of action, = 58.39 °, as
and inspectionillustrated
time difference and in Figure 12. The post-molded
surface quality control.
properties. volumetric shrink-
age ratio for ABS (Acrylonitrile Butadiene Styrene) is expected to be 0.4 to 0.7%. The ag-
gregation of black color, very low surface roughness, and cylindrical shape cause the prob-
lem as the hard example for surface inspection and surface quality control.
(a) (b)
Figure 12. The illustration of the measuring sample, (a) Isometric view of the sample, (b) The highly
(a) surface of the sample is called a piano-black surface. (b)
reflective
Figure 12. The illustration

Figureof12.
the
3.3. Process measuring
The sample,
illustration of
Sensors (a) Isometric
the measuring sample, (a)view of the
Isometric sample,
view (b) The
of the sample, (b) highly
The highly
reflective surface of the sample
reflective is called
surface of thea
3.3.1. Digital Sensor Data piano-black
sample is called surface.
a piano-black surface.
3.3.2.The data Sensor

Analog from the injection
Data molding machine, DRYMAX dryer, and TEMPRO plus
(Machine)
3.3. Process Sensors devices were sampled over the OPC UA network platform. The injection molding ma-
Since a persistent sampling rate of the OPC UA data was limited to 60 Hz, a set of
chine, DRYMAX dryer, and TEMPRO plus D devices have each an OPC UA server inside
3.3.1. Digital Sensorimportant machine signals such as screw pressure (i.e., pressure in the screw antechamber),
Dataprovides
which the information of the internal sensor and variables to be read or written.
The data from A program
the namedmolding
injection DataRetriever providedDRYMAX
machine, by WITTMANN BATTENFELD
dryer, and TEMPRO reads the data
plus
devices were sampled over the OPC UA network platform. The injection molding ma-
chine, DRYMAX dryer, and TEMPRO plus D devices have each an OPC UA server inside
ber), screw position, clamping force, and digital signals of the machine process pha
were measured using an HBM MX 840B 840B (Hottinger Brüel & Kjaer Austria Gmb
Vienna, Austria) for data acquisition and CATMAN software 840B (version 5.2.1, H
Polymerstinger Brüel & Kjaer Austria GmbH, Vienna, Austria) for storing the data with
2022, 14, 3551 14 ofa24sampli
rate of 600 Hz.

screw position, clamping force, and digital signals of the machine process phases were
3.3.3. Analog Sensor Data
measured (Mold)
using an HBM MX 840B 840B (Hottinger Brüel & Kjaer Austria GmbH, Vienna,
Austria) for data acquisition and CATMAN software 840B (version 5.2.1, Hottinger Brüel &
Four sensorsKjaer
areAustria
installed
GmbH,in the moving
Vienna, Austria) forside ofthe
storing the mold
data with ato monitor
sampling the
rate of 600cavity
Hz. pr
sure and polymer temperature during the injection, as shown in Figure 13. Three of
3.3.3. Analog Sensor Data (Mold)
sensors were typed on Fouran MTPS
sensors 408-IR-BTS-XSR
are installed in the moving sidefrom
of theFOS Messtechnik
mold to monitor the cavityGmbH
pressure(Schac
Audorf, Germany) andand thetemperature
polymer fourth sensor is ainjection,
during the Kistleraspressure sensor.
shown in Figure MX840B
13. Three and MX16
of the sensors
were typed on an MTPS 408-IR-BTS-XSR from FOS Messtechnik GmbH (Schacht-Audorf,
DAQ systems (from HBM) were used for sampling the mold sensor data with a sampli
Germany) and the fourth sensor is a Kistler pressure sensor. MX840B and MX1601 DAQ
rate of 600 Hz and CATMAN
systems (from HBM) software was
were used for usedthe
sampling formold
storing the with
sensor data data.
a sampling rate of
600 Hz and CATMAN software was used for storing the data.
(a) (b)
Figure 13. The positions
Figure 13.of
Thethe FOS ofP/T
positions sensors
the FOS are marked
P/T sensors are markedinin red and
red and the the position
position of the Kis
of the Kistler
pressure sensor is marked in blue, (a) marked positions on the part, (b)
pressure sensor is marked in blue, (a) marked positions on the part, (b) marked positions marked positions on the on
moving side of the mold.
moving side of the mold.
3.3.4. Ambient Sensors
Three special ambient sensor packages were built based on the ESP8266 module
3.3.4. Ambient Sensors
(Figure 14). Three sensors (BME280, MLX90614, HYT 271) were installed in the ambient
Three special ambient
sensor package sensor
to ensure packages were
the safety of the built
data and based on
the precision the
of the ESP8266 Three
measurements. module (F
sensor packages were installed in three zones: room, above mold area, and part quality
ure 14). Three sensors (BME280, MLX90614, HYT 271) were installed in the ambient sen
measurement area. These sensors sample the temperature and relative humidity every 30 s
package to ensureand the safety
upload of the data
the measurement data and theminutes
every ten precision
onto anof the measurements.
ownCloud (developed by Th
Information Center of Montanuniversitaet Leoben, Leoben, Austria).
sensor packages were installed in three zones: room, above mold area, and part qual
measurement area. Theseof sensors
3.4. Design Experiment sample
(DOE) the temperature and relative humidity every
A central composite
s and upload the measurement data inscribed
every(CCI)ten[50] type of CCD
minutes onto wasanselected to plan the(developed
ownCloud experi-
ments. The levels of the experimental points are shown in Table 2. The combinations of
Information Center
melt of Montanuniversitaet
temperature, Leoben). and mold temperature, i.e., the set TCU
i.e., the set nozzle temperature,
output temperature, were blocked separately since changes in the melt temperature and
mold temperature take a long time to reach stability. Therefore, 56 experiments were carried
out randomly. To approximate a stable process, a minimum of 40 samples were produced
Polymers 2022, 14, 3551 15 of 24
PEER REVIEW within each experiment. The last 10 shots were used to feed the statistical analysis of the15 of 24
designed DOE, however, all the shots (about 2500 samples) were used for training the AI
models to capture the trend of changes in the quality features.
(a) (b)
Figure 14. (a) Two ambient
Figure 14.sensor packages
(a) Two ambient sensorclose to an
packages industrial
close temperature
to an industrial and
temperature and humidity
humidity sensor sensor
for calibration,
for calibration, (b) The ESP8266(b) The ESP8266
module andmodule andboard.
sensor sensor board.
Table 2. Factor levels in designing the CCI experiments.
3.4. Design of Experiment (DOE)

Factor Unit −1 −0.42 0 +0.42 +1
◦C
A central compositeMelt inscribed
temperature(CCI) [50] type240 of CCD 246was selected
250 to plan260the exper-
254
iments. The levels of theHolding pressure

experimental pointsbar
are 300
shown358 in Table400 442
2. The 500
combinations of
Time of holding pressure s 14 18 22 26 30
melt temperature, i.e., the set nozzle temperature, ◦C
and mold temperature, i.e., the set TCU
Mold temperature 84 90 94 98 104
output temperature, were blocked separately
Injection speed mm/s
since 15
changes22
in the
27
melt32temperature
39
and
mold temperature take a long time to reach stability. Therefore, 56 experiments were car-
ried out randomly.3.5.
ToExperiments
approximate a stable
for Controlling process, a minimum of 40 samples were pro-
Strategies
duced within each experiment. The last 10 shotsdifferent
The controller was tested with three werecontrolling
used to strategies.
feed theThe statistical analysis
surface quality
was controlled using the injection speed and holding pressure as the controlling parameters.
of the designed DOE, however,
The linear dimension all(average
the shots
of the (about 2500 samples)
linear dimensions) were
was controlled withused for training
two different
the AI models to capture
strategies.the
Thetrend of changes
first strategy included in onlythe quality
injection features.
speed and holding pressure for the
controlling parameters. The second strategy included the mold temperature in addition to
the first strategy’s controlling parameters.
Table 2. Factor levels in designing the CCI experiments.
4. Results and Discussion
Factor Unit
4.1. DOE Results and Factors Correlations−1 −0.42 0 +0.42 +1
Melt temperatureFigure 15 shows the main
°C effect diagrams
240 for weight,
246 surface quality, and
250 Line 4 (as a260
254
sample of the line lengths) with the influence of the five investigated experimental factors.
Holding pressure
The machine parameters,bar 300 melt358
mold temperature, temperature,400 442 and500
holding pressure,
injection
Time of holding pressurespeed, have a significant
s effect on
14 the quality
18 features. 22 26 30
Mold temperature °C 84 90 94 98 104
Injection speed mm/s 15 22 27 32 39
3.5. Experiments for Controlling Strategies

The controller was tested with three different controlling strategies. The surface qual-
ity was controlled using the injection speed and holding pressure as the controlling pa-
were built using the DOE factors and sensors data. The R2 result for Line4 was 98.63% and
for weight was 99.26%. The accuracy of the model is not being better for surface quality
and the R2 remains at about 70%. Models were validated using the K-fold cross-validation
method with K = 10.
The precision of the dimensional measurement system was assessed as ±0.010 µ m in
Polymers 2022, 14, 3551 16 of 24
a previous study [5] and the length of the lines could be predicted with considerable ac-
curacy by regression models.
Figure
Figure15.
15.Main
Maineffect
effectdiagrams
diagrams from the
the analysis
analysisofofthe
theCCI
CCIexperimental
experimental design.
design.
2
Themain
R results of agree
the regression model (usingofonly
The effects well with the state the the
art DOE
whenfactors) for weight
considering were
a dominant
98.08%
impact and for Line 4 were 97.73% including the block effect in the models,
of the pin gates and their sealing: Higher mold temperature, higher melt temper-however, the
R2 value for surface quality response stayed below 70%. A new set of regression models
ature and higher injection speed (via increased shear heating) keep the pin gate longer
were built using the DOE factors and sensors data. The R2 result for Line4 was 98.63% and
open to melt flow, thus the holding pressure can be more effective and overwhelms the
for weight was 99.26%. The accuracy of the model is not being better for surface quality
greater shrinkage potential with lower cooling of the polymer melt, in other words, more
and the R2 remains at about 70%. Models were validated using the K-fold cross-validation
polymer
method is pumped
with K = 10.into the cavity with rising mold temperature, melt temperature, and
injection speed, making
The precision of thethe parts heavier
dimensional and larger. The
measurement minimum
system holding
was assessed as pressure
±0.010 µmtime
ofin14a sprevious
was chosen to guarantee that the gate is sealed in all experiments within
study [5] and the length of the lines could be predicted with considerable the DOE,
thus, as expected,
accuracy holding
by regression pressure time showed no significant influence on dimensions
models.
and weight.
The mainAs predicted in well
effects agree our former
with thesimulation
state of thestudy [4], the
art when process parameters
considering a dominant prove
im-
non-linear influences on the quality outputs. The surface (appearance) quality comprises
pact of the pin gates and their sealing: Higher mold temperature, higher melt temperature
and higher injection speed (via increased shear heating) keep the pin gate longer open to
melt flow, thus the holding pressure can be more effective and overwhelms the greater
shrinkage potential with lower cooling of the polymer melt, in other words, more polymer
is pumped into the cavity with rising mold temperature, melt temperature, and injection
speed, making the parts heavier and larger. The minimum holding pressure time of 14 s
Polymers 2022, 14, 3551 sink marks, weld-lines, and streaks. Sink marks are shrinkage-dominated surface defects, 17 of 24
thus the same dependencies as for dimensions and weight count. Higher contact temper-
ature, higher compression, and lower time for the formation of the frozen layer reduce the
strength
was of weld
chosen lines and
to guarantee flow
that thelines,
gate thus risinginmold
is sealed and melt temperatures,
all experiments and thus,
within the DOE, injec- as
tion speed reduce the visibility of the weld lines at the surface. Allegedly, the
expected, holding pressure time showed no significant influence on dimensions and weight. replication
of (unplanned)
As predicted in micro-scratches
our former simulationin the mirror-finished mold surface
study [4], the process by theprove
parameters ABS polymer
non-linear
grows with high mold and melt temperatures, making the part surface
influences on the quality outputs. The surface (appearance) quality comprises sink rougher, the marks,
sur-
weld-lines, and streaks. Sink marks are shrinkage-dominated surface defects, thus thethat
face gloss lower and the surface (appearance) quality again worse. A phenomenon same
has been reported,
dependencies as forfordimensions
instance in [74,75].
and weight count. Higher contact temperature, higher
compression, and lower time for the formation of the frozen layer reduce the strength of
4.2. In-Line Closed-Loop Control Results
weld lines and flow lines, thus rising mold and melt temperatures, and injection speed re-
duce During each controlling
the visibility of the weldstrategy, the surface.
lines at the machineAllegedly,
parametersthe were changedof
replication manually to
(unplanned)
make disturbances for the controller and investigate the behavior of
micro-scratches in the mirror-finished mold surface by the ABS polymer grows with high the controller. These
changes
mold andwere
meltapplied to the injection
temperatures, making speed
the partand/or
surfaceholding pressure
rougher, at some
the surface points,
gloss lower andand
to the
the mold(appearance)
surface temperature and meltagain
quality temperature
worse. A at phenomenon
some other points.that has been reported, for
instance in [74,75].
4.2.1. Surface Quality Control
4.2. In-Line Closed-Loop
The result Controlquality
of the surface Results control is shown in Figure 16. The controller was
limited to a distance
During of three shots
each controlling between
strategy, the controlling
the machine parametersactions.
wereAdditionally, it takes to
changed manually
between
make one to twofor
disturbances cycles till the controller
the controller can observe
and investigate thethe effect of
behavior ofchanges on the ma-
the controller. These
chine parameters
changes (depending
were applied on the time
to the injection of changes).
speed and/or holding pressure at some points, and
to theThe
mold experiment
temperaturefor this
and control strategy was
melt temperature at carried out at
some other three different melt and
points.
mold temperatures, respectively 250, 245, and 255 °C melt temperatures and 94, 90, and
4.2.1.
98 °CSurface Quality Control
mold temperatures (shown in blue, red, and green in Figure 16). The goal of the
controller was to achieve
The result of the surface a surface quality
quality value
control of 0.6 and
is shown higher.16.
in Figure The injection
The speed
controller was
and/or holding pressure were changed respectively to 15 mm/s and 300 bar
limited to a distance of three shots between the controlling actions. Additionally, it takes at the orange
triangle-marked
between shots.
one to two cyclesThe
tilldirection of thecan
the controller changes
observe was
thetoeffect
decrease the surface
of changes on thequality
machine
value.
parameters (depending on the time of changes).
Figure 16. Experiment results for surface quality control. Orange triangles are cycles with manual
Figure 16. Experiment results for surface quality control. Orange triangles are cycles with manual
interference in the process.
interference in the process.
The result shows that the controller was able to detect the interference after a maxi-
The experiment for this control strategy was carried out at three different melt and
mum of two cycles. The controller calculated and applied the new set of machine param-
mold temperatures, respectively 250, 245, and 255 ◦ C melt temperatures and 94, 90, and
eters
◦ (injection speed and holding pressure in this control strategy) as soon as it observed
98 C mold temperatures (shown in blue, red, and green in Figure 16). The goal of
the quality reduction. It took a maximum of three cycles for the controller to observe the
the controller was to achieve a surface quality value of 0.6 and higher. The injection
speed and/or holding pressure were changed respectively to 15 mm/s and 300 bar at the
orange triangle-marked shots. The direction of the changes was to decrease the surface
quality value.
Polymers 2022, 14, 3551 18 of 24

The result shows that the controller was able to detect the interference after a maximum
of two cycles. The controller calculated and applied the new set of machine parameters
(injection speed and holding pressure in this control strategy) as soon as it observed the
results
qualityofreduction.
the controlling
It took actions (because
a maximum onecycles
of three part for
is under production,
the controller one part
to observe is on the
the results
scale for weight measurement and one part is under the dimensional measurement).
of the controlling actions (because one part is under production, one part is on the scale for
Themeasurement
weight controller optimized the surface
and one part is underquality for melt temperatures
the dimensional measurement). of 250 and 255 °C
(alongThewithcontroller optimized the surface quality for melt temperatures of 250 andHowever,
mold temperatures of 96 and 98 °C) continuously without failure. 255 ◦ C
the controller
(along underestimated
with mold temperaturesthe of necessary ◦
96 and 98 controlling
C) continuouslyactionwithout
at the melt temperature
failure. However, of
245
the°C (along with
controller the mold temperature
underestimated the necessaryofcontrolling
90 °C). This could
action be melt
at the addressed by correla-
temperature of
245 ◦between
tions C (along the
withsurface
the mold temperature
quality, ◦ C). This could be addressed by correlations
of 90mold
melt, and temperatures.
between the surface quality, melt, and mold temperatures.
4.2.2. Linear Dimension Control, Strategy 1
The result for linear dimension control is given in Figure 17. The controlling goal was
The result for linear dimension control is given in Figure 17. The controlling goal was
a linear dimension of 120 ± 0.025 mm. The tolerance is shown with red dashed lines. The
a linear dimension of 120 ± 0.025 mm. The tolerance is shown with red dashed lines. The
controller
controller was limitedtotoa adistance
was limited distance of three
of three cycles
cycles between
between the controlling
the controlling action. action.
The meltThe
melt temperature and mold temperature were set◦ to 94 °C and
◦ 250
temperature and mold temperature were set to 94 C and 250 C respectively. °C respectively.
Figure
Figure17.
17.Experiment
Experiment results forlinear
results for lineardimension
dimension control
control with
with strategy
strategy 1 for1controlling
for controlling parame-
parameters.
ters.
Orange triangles are cycles with manual interference in the process. Arrows show the directionsdirec-
Orange triangles are cycles with manual interference in the process. Arrows show the of
tions of the change effect.
the change effect.
The
Themanual
manualinterference
interference at points
points 1,1,4,4,and
and55was
wasa adecrement
decrement ofof
thethe injection
injection speed
speed
and
andholding
holdingpressure
pressure respectively
respectively toto 15
15mm/s
mm/s andand300300bar,
bar,and
andwaswasanan increment
increment of the
of the
sameparameters
same parameterstoto3030mm/s
mm/s and
and 450450
barbar
at at point
point 3. 3.
TheThe melt
melt temperature
temperature andand mold
mold tem-
temperature were increased respectively to 98 ◦ C and 255 ◦ C at point 2. The dimension
perature were increased respectively to 98 °C and 255 °C at point 2. The dimension of the
of the molded
molded part waspart was increasing
increasing cycle bycycle bywith
cycle cycleconstant
with constant
machinemachine parameters.
parameters. The
The process
processwas
control control was successful
successful for dimension
for linear linear dimension control.
control. However,
However, it seemed
it seemed thatthat more
more than
than three cycles’ distance between the controlling actions
three cycles’ distance between the controlling actions is required. is required.

The controller was tested with the second strategy and the results are shown in
The18.
Figure controller
Since thewas tested
mold with theissecond
temperature strategy in
a slow variable and
thethe resultsmolding
injection are shown in Fig-
process,
ure
the18. Since the
distance moldthe
between temperature
controllingisactions
a slowwasvariable in the
changed injection
to ten cycles molding
instead ofprocess,
three. the
distance between the controlling actions was changed to ten cycles instead of three.
The manual interference parameters were respectively incensement of injection
speed and holding pressure to 15 mm/s and 300 bar at points 1, 4, 5, and additionally
decrement of the mold temperature to 90 °C at point 4. The injection speed was decreased
to 15 mm/s individually at point 3, and injection speed and holding pressure were respec-
tively increased to 30 mm/s and 450 bar at point 2. The controller was disabled for 9 cycles
after point 4 to reset the stable conditions of the controller to initial conditions, and espe-
cially observe the first new set point for mold temperature after enabling the controller.
remained unchanged. The controller could successfully predict the appropriate machine
Polymers 2022, 14, 3551 19 of 24
parameters (injection speed, holding pressure, and mold temperature) to set the linear
dimension within the set tolerance.
Figure 18. Experiment

Figure 18. Experimentresults
resultsfor
forlinear
linear dimension
dimension control
control with
with strategy
strategy 2 for2controlling
for controlling parame-
parameters.
ters. Orange triangles are cycles with manual interference in the process. Arrows show
Orange triangles are cycles with manual interference in the process. Arrows show the directions of
the direc-
tions of the change effect.
the change effect.
4.2.4.The
Weight
manualControl
interference parameters were respectively incensement of injection speed
and holding
Despitepressure
that the to 15 mm/swas
controller and not
300 bar at points
tested 1, 4, 5,control,
for weight and additionally
the results decrement
of correla-
of the between
tions mold temperature
the weight 90 ◦ Cother
to and at point 4. The injection
controlling speed
strategies arewas decreased
provided hereto to
15 illustrate
mm/s
individually
that weight isata point
more 3, and and
stable injection speed and
controllable holdinginpressure
parameter were molding
the injection respectively in-
process.
creased to 30 mm/s and 450 bar at point 2. The controller was disabled
The surface quality and weight are illustrated in Figure 19. As shown, the weight for 9 cycles after
point 4 toinreset
(colored red)the stablethe
follows conditions
general of the controller
trend to initial
of the changes, conditions,
since the changingand especially
parameters
observe the first new set point for mold
are injection speed, holding pressure, melt temperature, temperature after
and enabling
mold the controller.
temperature which Ataf-
point 6, the melt temperature was increased to 255 ◦ C but the other machine parameters
fect the surface quality and weight in a similar way, see Figure 15. The fluctuations of the
remained unchanged. The controller could successfully predict the appropriate machine
weight over the shots are less than for the surface quality.
parameters (injection speed, holding pressure, and mold temperature) to set the linear
To compare the fluctuations, we define signal-to-noise ratio (SNR) as shown in Equa-
dimension within the set tolerance.
tion (2). Signal-to-noise ratio for the weight in Figure 18 is between 27 to 51 and for the
surface quality
4.2.4. Weight is between 1.3 to 1.75.
Control
The correlations between was
Despite that the controller linearnotdimension and weight
tested for weight arethe
control, shown in Figure
results 20 for the
of correlations
same experiment of the surface quality control. The SNR value for
between the weight and other controlling strategies are provided here to illustrate that the linear dimensions
is between
weight 4 to 9.stable
is a more The same correlation is
and controllable given forinlinear
parameter dimension
the injection control
molding (second strat-
process.
egy) The
in Figure
surface quality and weight are illustrated in Figure 19. As shown, the weightbe-
21. The trend correlations match for the weight and linear dimension
cause of similar main factor
(colored in red) follows effectstrend
the general (as shown in Figuresince
of the changes, 20). the changing parameters are
The speed,
injection SNR value
holding of pressure,
the weight meltshows that weight
temperature, has fewer
and mold fluctuations
temperature which andaffecthigher
the
resolution rather
surface quality andthan surface
weight in a quality and dimensions
similar way, see Figure 15.against the same variations
The fluctuations of the weight of the
machine parameters.
over the shots are less Low fluctuations
than for the surface and high resolutions are two important variables
quality.
To compare the fluctuations,
for the controllability of a quality feature. we define signal-to-noise ratio (SNR) as shown in
Equation (2). Signal-to-noise ratio for the weight in Figure 18 is between 27 to 51 and
𝐴𝑣𝑒𝑟𝑎𝑔𝑒𝑑⁡𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑙𝑒𝑑⁡𝑞𝑢𝑎𝑙𝑖𝑡𝑦⁡𝑣𝑎𝑙𝑢𝑒−𝐴𝑣𝑒𝑟𝑎𝑔𝑒𝑑⁡𝑙𝑜𝑤⁡𝑞𝑢𝑎𝑙𝑖𝑡𝑦⁡𝑣𝑎𝑙𝑢𝑒
for the surface quality𝑆𝑁𝑅 is = between 1.3 to 1.75. , (2)
𝑀𝑎𝑥⁡𝑜𝑓⁡𝑓𝑙𝑢𝑐𝑡𝑢𝑎𝑡𝑖𝑜𝑛𝑠−𝑀𝑖𝑛⁡𝑜𝑓⁡𝑓𝑙𝑢𝑐𝑡𝑢𝑎𝑡𝑖𝑜𝑛𝑠
The correlations between linear dimension and weight are shown in Figure 20 for the
same experiment of the surface quality control. The SNR value for the linear dimensions is
between 4 to 9. The same correlation is given for linear dimension control (second strategy)
in Figure 21. The trend correlations match for the weight and linear dimension because of
similar main factor effects (as shown in Figure 20).
Polymers
Polymers 2022,
2022, 14,
14, xx FOR
FOR PEER
PEER REVIEW
REVIEW 20
20 of
of 24
24
Polymers2022,
Polymers 14,x 3551
2022,14, FOR PEER REVIEW 20
20ofof2424
Figure
Figure 19.
19. Comparing
Comparing the
the correlations
correlations between
between surface
surface quality
quality and
and weight
weight for
for surface
surface quality
quality con-
con-
trol.
trol. 19. Comparing the correlations between surface quality and weight for surface quality control.
Figure
Figure 19. Comparing the correlations between surface quality and weight for surface quality con-
trol.
Figure
Figure 20.
20. Comparing
20. Comparing the
the correlations
between linear
between dimension
linear dimension
linear and
dimension and weight
and weight for
weight for surface
forsurface quality
surfacequality
quality
control.
control.
control.
Figure 20. Comparing the correlations between linear dimension and weight for surface quality
control.
Figure
Figure 21.
21. Comparing
21. Comparing the
Comparing the correlations
the between
correlations linear
between linear dimension
lineardimension and
dimensionand weight
andweight for
weightfor linear
forlinear dimension
lineardimension
dimension
control (with
control (with
control strategy
(with strategy 2 for
strategy 22 for controlling
for controlling parameters).
controllingparameters).
parameters).
Figure 21. Comparing the correlations between linear dimension and weight for linear dimension
control (with strategy 2 for controlling parameters).
Polymers 2022, 14, 3551 21 of 24
The SNR value of the weight shows that weight has fewer fluctuations and higher
resolution rather than surface quality and dimensions against the same variations of the
machine parameters. Low fluctuations and high resolutions are two important variables
for the controllability of a quality feature.
Averaged controlled quality value − Averaged low quality value

SNR = , (2)
Max o f f luctuations − Min o f f luctuations
5. Conclusions
In this paper, a novel in-line and fully automated closed-loop process control was
presented for the injection molding process. The automation was implemented over the
OPC UA network platform in compliance with Industry 4.0 framework for the injection
molding process. A novel in-line dimensional measurement system was used to measure
the as-molded dimensions of the parts. Therefore, three quality features of the molded
parts (weight, dimensional properties, and surface quality) were measured in an in-line and
as-molded manner. Initially, a set of experiments through a CCI type DOE were planned to
train eight types of AI models. Afterward, the models were applied along with a predictive
controller to control the injection molding process for a quality feature. The controller was
tested with three different control strategies for the control goal. The surface quality was
controlled to achieve a surface quality value higher than 0.6 (where 1 is the best surface
quality). The linear dimension (the average of three measured linear dimension lines) was
controlled for the goal of 120 mm with a tolerance of ±0.025 mm. The quality features were
predicted through the AI models and applied to make the control decision in comparison
with the in-line measurements. The controller could successfully control the process for all
the strategies. Since injection molding process is a discrete and slow process, the distance
between the controlling actions appears to be important in preventing fluctuations.
It would be possible to exclude the in-line measurements after training the AI models,
in future research. The process is influenced by many parameters including the environ-
mental variables. Monitoring the trend of the changes could be combined with the control
strategy to make decisions for the suitable cycle of control action. Multi-objective process
control is the next stage of this research to investigate the possibility of quality control in
multiple aspects.
Author Contributions: Conceptualization, S.S.A., G.B.-W., C.H., P.T., D.P.G. and W.F.; Data cu-
ration, S.S.A. and G.B.-W.; Funding acquisition, G.B.-W., D.P.G. and W.F.; Investigation, S.S.A.,
G.B.-W. and P.T.; Methodology, S.S.A., D.P.G., G.B.-W., P.T. and A.S.; Project administration, W.F. and
D.P.G.; Resources, G.B.-W., D.P.G. and W.F.; Software, S.S.A., P.T. and A.S.; Supervision, G.B.-W. and
D.P.G.; Validation, S.S.A. and G.B.-W.; Visualization, S.S.A. and P.T.; Writing—original draft, S.S.A.;
Writing—review & editing, S.S.A., P.T., A.S., C.H., D.P.G., G.B.-W. and W.F. All authors have read and
agreed to the published version of the manuscript.
Funding: This research was funded by FFG research promotion agency in Austria as a part of the
project INQCIM.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: The data presented in this study are available upon request from the
corresponding authors.
Acknowledgments: This research has been performed at Montanuniversitaet Leoben, department
of Polymer Engineering and Science, chair of injection molding of polymers, and in partnership
with Polymer Competence Center Leoben (PCCL) as a part of the project INQCIM which has
been supported by the companies WITTMAN BATTENFELD GmbH, MAHLE Filtersysteme Aus-
tria GmbH, EKB Elektro- u. Kunststofftechnik GmbH, Miraplast Kunststoffverarbeitungs GmbH,
Julius Blum GmbH, and the Institute of Production Engineering and Photonic Technologies at TU
WIEN university.
Polymers 2022, 14, 3551 22 of 24
Conflicts of Interest: The authors declare no conflict of interest.
References
1. Statistica. Plastics Industry in Europe; Statistica: London, UK, 2021.
2. Zhao, P.; Zhang, J.; Dong, Z.; Huang, J.; Zhou, H.; Fu, J.; Turng, L.-S. Intelligent Injection Molding on Sensing, Optimization, and
Control. Adv. Polym. Technol. 2020, 2020, 7023616. [CrossRef]
3. Schmidberger, E. Auf Dem Weg Zur Null-Fehler-Produktion. Available online: https://industrieanzeiger.industrie.de/allgemein/
auf-dem-weg-zur-null-fehler-produktion/ (accessed on 10 June 2022).
4. Berger-Weber, G.; Saeidi Aminabadi, S. The Injection Mold as a Cyber Physical System. Using Simulation to Train Its Artificial
Intelligence. In Proceedings of the 29th Leoben Kunststoffkolloquium, Leoben, Austria, 16 September 2021; Volume 29, p. 10.
5. Saeidi Aminabadi, S.; Jafari-Tabrizi, A.; Gruber, D.P.; Berger-Weber, G.; Friesenbichler, W. An Automatic, Contactless, High-
Precision, High-Speed Measurement System to Provide In-Line, As-Molded Three-Dimensional Measurements of a Curved-Shape
Injection-Molded Part. Technologies 2022, 10, 95. [CrossRef]
6. Haman, S. Prozessnahes Qualitätsmanagement Beim Spritzgießen. Ph.D. Thesis, Technischen Universität Chemnitz, Chemnitz,
Germany, 2003.
7. Heinzler, F.A. Modellgestützte Qualitätsregelung Durch Eine Adaptive, Druckgeregelte Prozessführung Beim Spritzgießen. Ph.D.
Thesis, Universität Duisburg-Essen, Duisburg, Germany, 2014.
8. Kenig, S.; Ben-David, A.; Omer, M.; Sadeh, A. Control of Properties in Injection Molding by Neural Networks. Eng. Appl. Artif.
Intell. 2001, 14, 819–823. [CrossRef]
9. Kazmer, D.O. Injection Mold Design Engineering. In Injection Mold Design Engineering, 2nd ed.; Kazmer, D.O., Ed.; Hanser
Publishers: Munich, Germany, 2016; pp. I–XXIV. ISBN 978-1-56990-570-8.
10. Kleppmann, W. Versuchsplanung Produkte und Prozesse Optimierem, 8th ed.; Carl Hanser Verlag GmbH: Munich, Germany, 2013.
11. Kulkarni, S. Design of Experiments for Injection Molding. In Robust Process Development and Scientific Molding; Carl Hanser Verlag
GmbH & Co. KG: Munich, Germany, 2010; Volume 8, pp. 145–167. ISBN 978-3-446-42275-9.
12. Andrisano, A.O.; Gherardini, F.; Leali, F.; Pellicciari, M.; Vergnano, A. Design of Simulation Experiments Method for Injection
Molding Process Optimization. Libr. Cortina 2011, 11, 476–486.
13. Johnston, S.; McCready, C.; Hazen, D.; VanDerwalker, D.; Kazmer, D. On-Line Multivariate Optimization of Injection Molding.
Polym. Eng. Sci. 2015, 55, 2743–2750. [CrossRef]
14. Régnier, G.; Trotignon, J.P. Local Orthotropic Shrinkage Determination in Injected Moulded Polymer Plates. Polym. Test. 1993,
12, 383–392. [CrossRef]
15. Jansen, K.M.B.; Van Dijk, D.J.; Husselman, M.H. Effect of Processing Conditions on Shrinkage in Injection Molding. Polym. Eng.
Sci. 1998, 38, 838–846. [CrossRef]
16. Liao, S.J.; Chang, D.Y.; Chen, H.J.; Tsou, L.S.; Ho, J.R.; Yau, H.T.; Hsieh, W.H.; Wang, J.T.; Su, Y.C. Optimal Process Conditions of
Shrinkage and Warpage of Thin-Wall Parts. Polym. Eng. Sci. 2004, 44, 917–928. [CrossRef]
17. Pomerleau, J.; Sanschagrin, B. Injection Molding Shrinkage of PP: Experimental Progress. Polym. Eng. Sci. 2006, 46, 1275–1283.
[CrossRef]
18. Sónia Ferraz Ferreira Alves, P. Shrinkage and Warpage Behaviour on Injection Moulding Parts; Universidade do Minho: Minho,
Portugal, 2008.
19. Kapadia, D. Shrinkage and Warpage in Polymer Injection Moulding: Models and Simulation; Johannes Kepler University Linz: Linz,
Austria, 2019.
20. Annicchiarico, D.; Alcock, J.R. Review of Factors That Affect Shrinkage of Molded Part in Injection Molding. Mater. Manuf.
Process. 2014, 29, 662–682. [CrossRef]
21. Ahmed, T.; Sharma, P.; Karmaker, C.L.; Nasir, S. Warpage Prediction of Injection-Molded PVC Part Using Ensemble Machine
Learning Algorithm. Mater. Today Proc. 2020, 50, 565–569. [CrossRef]
22. Masato, D.; Rathore, J.; Sorgato, M.; Carmignato, S.; Lucchetta, G. Analysis of the Shrinkage of Injection-Molded Fiber-Reinforced
Thin-Wall Parts. Mater. Des. 2017, 132, 496–504. [CrossRef]
23. Singh, G.; Pradhan, M.K.; Verma, A. Multi Response Optimization of Injection Moulding Process Parameters to Reduce Cycle
Time and Warpage. Mater. Today Proc. 2018, 5, 8398–8405. [CrossRef]
24. Azad, R.; Shahrajabian, H. Experimental Study of Warpage and Shrinkage in Injection Molding of HDPE/RPET/Wood Compos-
ites with Multiobjective Optimization. Mater. Manuf. Process. 2019, 34, 274–282. [CrossRef]
25. Barghash, M.A.; Alkaabneh, F.A. Shrinkage and Warpage Detailed Analysis and Optimization for the Injection Molding Process
Using Multistage Experimental Design. Qual. Eng. 2014, 26, 319–334. [CrossRef]
26. Oliaei, E.; Heidari, B.S.; Davachi, S.M.; Bahrami, M.; Davoodi, S.; Hejazi, I.; Seyfi, J. Warpage and Shrinkage Optimization
of Injection-Molded Plastic Spoon Parts for Biodegradable Polymers Using Taguchi, ANOVA and Artificial Neural Network
Methods. J. Mater. Sci. Technol. 2016, 32, 710–720. [CrossRef]
27. Huzaim, N.H.M.; Rahim, S.Z.A.; Musa, L.; Abdellah, A.E.; Abdullah, M.M.A.B.; Rennie, A.; Rahman, R.; Garus, S.; Błoch, K.;
Sandu, A.V.; et al. Potential of Rapid Tooling in Rapid Heat Cycle Molding: A Review. Materials 2022, 15, 3725. [CrossRef]
28. Evens, T.; Castagne, S.; Seveno, D.; Van Bael, A. Predicting the Replication Fidelity of Injection Molded Solid Polymer Microneedles.
Int. Polym. Process. 2022, 37, 237–254. [CrossRef]
Polymers 2022, 14, 3551 23 of 24
29. Selvaraj, S.K.; Raj, A.; Rishikesh Mahadevan, R.; Chadha, U.; Paramasivam, V. A Review on Machine Learning Models in Injection
Molding Machines. Adv. Mater. Sci. Eng. 2022, 2022, 1949061. [CrossRef]
30. Karbasi, H.; Eng, P.; Reiser, H. Smart Mold: Real-Time In-Cavity Data Acquisition. Ph.D. Thesis, School of Engineering, Conestoga
College ITAL Kitchener, Kitchener, ON, Canada, 2006.
31. Nagorny, P.; Pillet, M.; Pairel, E.; Le Goff, R.; Loureaux, J.; Wali, M.; Kiener, P. Quality Prediction in Injection Molding. In
Proceedings of the 2017 IEEE International Conference on Computational Intelligence and Virtual Environments for Measurement
Systems and Applications (CIVEMSA), Annecy, France, 26–28 June 2017; pp. 141–146.
32. Mueller, F. Analysis, Implementation and Investigation of a Wireless In-Mold Sensor for Injection Molding. Ph.D. Thesis,
Montanuniversitaet Leoben, Leoben, Austria, 2014.
33. Ageyeva, T.; Horváth, S.; Kovács, J.G. In-Mold Sensors for Injection Molding: On the Way to Industry 4.0. Sensors 2019, 19, 3551.
[CrossRef]
34. Chang, H.; Su, Z.; Lu, S.; Zhang, G. Intelligent Predicting of Product Quality of Injection Molding Recycled Materials Based on
Tie-Bar Elongation. Polymers 2022, 14, 679. [CrossRef]
35. Petrova, T.; Kazmer, D. Hybrid Neural Models for Pressure Control in Injection Molding. Adv. Polym. Technol. 1999, 18, 19–31.
[CrossRef]
36. Mok, S.L.; Kwong, C.K.; Lau, W.S. Review of Research in the Determination of Process Parameters for Plastic Injection Molding.
Adv. Polym. Technol. 1999, 18, 225–236. [CrossRef]
37. Chen, Z.; Turng, L.-S. A Review of Current Developments in Process and Quality Control for Injection Molding. Adv. Polym.
Technol. 2005, 24, 165–182. [CrossRef]
38. Ribeiro, B. Support Vector Machines for Quality Monitoring in a Plastic Injection Molding Process. IEEE Trans. Syst. Man Cybern.
Part C Appl. Rev. 2005, 35, 401–410. [CrossRef]
39. Yen, C.; Lin, J.C.; Li, W.; Huang, M.F. An Abductive Neural Network Approach to the Design of Runner Dimensions for the
Minimization of Warpage in Injection Mouldings. J. Mater. Process. Technol. 2006, 174, 22–28. [CrossRef]
40. Chen, W.-C.; Tai, P.-H.; Wang, M.-W.; Deng, W.-J.; Chen, C.-T. A Neural Network-Based Approach for Dynamic Quality Prediction
in a Plastic Injection Molding Process. Expert Syst. Appl. 2008, 35, 843–849. [CrossRef]
41. Altan, M. Reducing Shrinkage in Injection Moldings via the Taguchi, ANOVA and Neural Network Methods. Mater. Des. 2010,
31, 599–604. [CrossRef]
42. Özek, C.; Çelık, Y.H. Calculating Molding Parameters in Plastic Injection Molds with ANN and Developing Software. Mater.
Manuf. Process. 2012, 27, 160–168. [CrossRef]
43. Ogorodnyk, O.; Martinsen, K. Monitoring and Control for Thermoplastics Injection Molding A Review. Procedia CIRP 2018,
67, 380–385. [CrossRef]
44. Gim, J.; Rhee, B. Novel Analysis Methodology of Cavity Pressure Profiles in Injection-Molding Processes Using Interpretation of
Machine Learning Model. Polymers 2021, 13, 3297. [CrossRef] [PubMed]
45. Guo, F.; Zhou, X.; Liu, J.; Zhang, Y.; Li, D.; Zhou, H. A Reinforcement Learning Decision Model for Online Process Parameters
Optimization from Offline Data in Injection Molding. Appl. Soft Comput. 2019, 85, 105828. [CrossRef]
46. Dang, X.-P. General Frameworks for Optimization of Plastic Injection Molding Process Parameters. Simul. Model. Pract. Theory
2014, 41, 15–27. [CrossRef]
47. Chen, W.C.; Lin, Y.F.; Liou, P.H. A Multi-Objective Quality Optimization System for Injection Molding Process Parameters. Adv.
Mater. Res. 2014, 945–949, 478–483. [CrossRef]
48. Lu, N.; Gong, G.; Yang, Y.; Lu, J. Multi-Objective Process Parameter Optimization for Energy Saving in Injection Molding Process.
J. Zhejiang Univ. Sci. A 2012, 13, 382–394. [CrossRef]
49. Mehat, N.M.; Kamaruddin, S. Quality Control and Design Optimisation of Plastic Product Using Taguchi Method: A Comprehen-
sive Review. Int. J. Plast. Technol. 2012, 16, 194–209. [CrossRef]
50. Montgomery, D.C. Design and Analysis of Experiments, 9th ed.; John Wiley & Sons, Inc.: Hoboken, NJ, USA, 2017.
51. Tsai, K.-M.; Luo, H.-J. An Inverse Model for Injection Molding of Optical Lens Using Artificial Neural Network Coupled with
Genetic Algorithm. J. Intell. Manuf. 2014, 28, 473–487. [CrossRef]
52. Chen, W.-C.; Nguyen, M.-H.; Chiu, W.-H.; Chen, T.-N.; Tai, P.-H. Optimization of the Plastic Injection Molding Process Using the
Taguchi Method, RSM, and Hybrid GA-PSO. Int. J. Adv. Manuf. Technol. 2016, 83, 1873–1886. [CrossRef]
53. Zhu, J.; Chen, J.C. Fuzzy Neural Network-Based in-Process Mixed Material-Caused Flash Prediction (FNN-IPMFP) in Injection
Molding Operations. Int. J. Adv. Manuf. Technol. 2006, 29, 308–316. [CrossRef]
54. Vagelatos, G.A.; Rigatos, G.G.; Tzafestas, S.G. Incremental Fuzzy Supervisory Controller Design for Optimizing the Injection
Molding Process. Expert Syst. Appl. 2001, 20, 207–216. [CrossRef]
55. Li, D.; Zhou, H.; Zhao, P.; Li, Y. A Real-Time Process Optimization System for Injection Molding. Polym. Eng. Sci. 2009,
49, 2031–2040. [CrossRef]
56. Wang, S.; Ying, J.; Chen, Z.; Cai, K. Grey Fuzzy PI Control for Packing Pressure during Injection Molding Process. J. Mech. Sci.
Technol. 2011, 25, 1061–1068. [CrossRef]
57. Bangert, P.; Cajaraville, P.; Dormann, B.; Kohler, M.; Imgrund, P.; Haack, J.; Volkert, J.; Lopez, O.; Rodriguez, P.; Martinez, L.; et al.
Real-Time Quality Control of Injection Molding. In Proceedings of the XIX Conference on Machine Tools and Manufacturing
Technologies, San Sebastian, Spain, 10 June 2013; p. 7.
Polymers 2022, 14, 3551 24 of 24
58. Moreira, E.E.; Alves, F.S.; Martins, M.; Ribeiro, G.; Pina, A.; Aguiam, D.E.; Sotgiu, E.; Fernandes, E.P.; Gaspar, J. Industry 4.0:
Real-Time Monitoring of an Injection Molding Tool for Smart Predictive Maintenance. In Proceedings of the 2020 25th IEEE
International Conference on Emerging Technologies and Factory Automation (ETFA), Vienna, Austria, 8–11 September 2020;
pp. 1209–1212.
59. Hopmann, C.; Ressmann, A.; Reiter, M.; Stemmler, S.; Abel, D. A Self-Optimising Injection Moulding Process with Model-Based
Control System Parameterisation. Int. J. Comput. Integr. Manuf. 2016, 29, 1190–1199. [CrossRef]
60. Hoppe, S. There Is No Industrie 4.0 without OPC UA; Opcfoundation: Scottsdale, AZ, USA, 2017.
61. Müllers, B. Euromap 77: OPC UA Interfaces for Plastics and Rubber Machinery—Data Exchange between Injection Moulding
Machines and MES. 2020. Available online: https://www.euromap.org/euromap77 (accessed on 15 March 2022).
62. Müllers, B. Euromap83: OPC UA for Plastics and Rubber Machinery—General Type Definitions. 2021. Available online:
https://www.euromap.org/euromap83 (accessed on 15 March 2022).
63. Kazmer, D.O. Plastics Manufacturing Systems Engineering; Carl Hanser Verlag GmbH & Co., KG: Munich, Germany, 2009;
ISBN 978-3-446-42014-4.
64. Brinkmann, O.B.; Schmachtenberg, O. International Plastics Handbook. In Hanser; Carl Hanser Verlag GmbH & Co., KG: Munich,
Germany, 2006; ISBN 978-3-446-22905-1.
65. Gruber, D.P. Method for Automatically Detecting a Defect on a Surface of a Molded Part. WO2010102319, 16 September 2010.
66. Gruber, D.P. Method and Device for the Optical Analysis of the Surface of an Object. EP14186013, 1 April 2015.
67. Haselmann, M.; Gruber, D.P. Pixel-Wise Defect Detection by CNNs without Manually Labeled Training Data. Appl. Artif. Intell.
2019, 33, 548–566. [CrossRef]
68. Gruber, D.P.; Haselmann, M. Inspection of Transparent Objects with Varying Light Scattering Using a Frangi Filter. J. Imaging
2021, 7, 27. [CrossRef]
69. Macher, J.; Gruber, D.P.; Altenbuchner, T.; Pacher, G.A.; Berger, G.R.; Friesenbichler, W. A Novel Sink Mark Model for High Gloss
Injection Molded Parts—Correlation of Deflectometric and Topographic Measurements. Polym. Test. 2014, 39, 12–19. [CrossRef]
70. Gruber, D.P.; Buder-Stroisznigg, M.; Wallner, G.; Jandel, B.S.L.; Lang, R.W. Characterization of Gloss Properties of Differently
Treated Polymer Coating Surfaces by Surface Clarity Measurement Methodology. Appl. Opt. 2012, 51, 4833–4840. [CrossRef]
71. Gruber, D.P.; Steiner, A.; Saeidi Aminabadi, S.; Friesenbichler, W. A New Quality Predictor and Controller by Means of Artificial
Intelligence for an Injection Molding Process. In Proceedings of the 30th Leoben-Conference on Polymer Engineering and Science,
Leoben, Austria, 14 June 2022.
72. He, K.; Zhang, X.; Ren, S.; Sun, J. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, Las Vegas, NV, USA, 27–30 June 2016; pp. 770–778.
73. Breiman, L. Random Forests. Mach. Learn. 2001, 45, 5–32. [CrossRef]
74. Berger, G.R.; Gruber, D.P.; Friesenbichler, W.; Teichert, C.; Burgsteiner, M. Replication of Stochastic and Geometric Micro
Structures—Aspects of Visual Appearance. Int. Polym. Process. 2011, 26, 313–322. [CrossRef]
75. Liparoti, S.; Speranza, V.; Pantani, R.; Titomanlio, G. Multi-Scale Simulation of Injection Molding Process with Micro–Features
Replication: Relevance of Rheological Behaviour and Crystallization. Polymers 2021, 13, 3236. [CrossRef]

polymers-14-03551-v2

Uploaded by

Copyright:

Available Formats

polymers-14-03551-v2

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

polymers-14-03551-v2

Uploaded by

Copyright:

Available Formats

polymers

Polymers 2022, 14, 3551. https://doi.org/10.3390/polym14173551 https://www.mdpi.com/journal/polymers

1.1. Process Parameters

1.2. Process Sensors

1.3. Application of Artificial Intelligence (AI) in PIM

1.4. Non-Iterative Optimization Methods

1.5. Iterative Optimization Methods

1.6. OPC UA and Industry 4.0

1.7. Research Objective

2. Materials and Methods

2.2. OPC UA Communication Platform

2.3. Quality Feedback

Figure 5. The position of the six scanning lines.

2.3.3. Surface Inspection

2.3.3. Surface Inspection

QS = (CP − min(CP ))/(max(CP ) − min(CP )), (1)

Table 1. The time duration QS iseach

measurement 44 2.4. Process Automation

Figure 6. Automation process sequence diagram.

Figure 8. The general block diagram for an injection molding process.

Figure 9. The block diagram of the control system.

Figure 10. The controller block diagram in detail.

2.6.1. Controller Algorithm

2.6.2. Extending Grid Search

Algorithm 1: Grid search

2.6.3. Controller Confidence Scaling

3.2. Case Study

3.3. Process Sensors

Figure 11. Experimental setup.

3.2. Case Study

and arc angle α = 58.39 °, as

Figure 12. The illustration

3.3.2.The data Sensor

rate of 600 Hz.

Table 2. Factor levels in designing the CCI experiments.

3.4. Design of Experiment (DOE)

iments. The levels of theHolding pressure

3.5. Experiments for Controlling Strategies

Polymers 2022, 14, x FOR PEER REVIEW 18 of 24

4.2.3. Linear Dimension Control, Strategy 2

Figure 18. Experiment

Averaged controlled quality value − Averaged low quality value

Conflicts of Interest: The authors declare no conflict of interest.

You might also like