Depar tment of Mechanical Engineering

QUALIT Y ASSURANCE
THROUGH ADVANCED
MANUFACTURING IN
METAL CASTING
FOUNDRIES
Tekin Çağın Uyan

DOCTORAL
THESES
Aalto University publication series
DOCTORAL THESES 206/2022

QUALITY ASSURANCE THROUGH

ADVANCED MANUFACTURING IN
METAL CASTING FOUNDRIES

Tekin Çağın Uyan

A doctoral dissertation completed for the degree of Doctor of

Science (Technology) to be defended, with the permission of the
Aalto University School of Engineering on 27 January 2023 at 12pm.

Aalto University
School of Engineering
Department of Mechanical Engineering
Advanced Manufacturing and Materials
Supervising professor
Prof. Kevin Otto, University of Melbourne, Australia

Preliminary examiners
Prof. Anders E. W. Jarfors, Jönköping University, Sweden
Prof. Alberto Molinari, University of Trento, Italy

Opponent
Prof. Alberto Molinari, University of Trento, Italy

Aalto University publication series

DOCTORAL THESES 206/2022

© 2022 Tekin Çağın Uyan

ISBN 978-952-64-1099-9 (printed)

ISBN 978-952-64-1100-2 (pdf)
ISSN 1799-4934 (printed)
ISSN 1799-4942 (pdf)
http://urn.fi/URN:ISBN:978-952-64-1100-2

Unigrafia Oy
Helsinki 2022 N
SWA ECO
C
DI

LA
NOR

BEL

Finland
Printed matter
Printed matter
1234 5678
4041-0619
 Ab s t ra c t
Aalto University, P.O. Box 11000, FI-00076 Aalto www.aalto.fi

Author
Tekin Çağın Uyan
Name of the doctoral thesis
QUALITY ASSURANCE THROUGH ADVANCED MANUFACTURING IN METAL CASTING
FOUNDRIES
P u b l i s h e r School of Engineering
U n i t Department of Mechanical Engineering
Series Aalto University publication series DOCTORAL THESES 206/2022
F i e l d o f r e s e a r c h Advanced Manufacturing and Materials
M a n u s c r i p t s u b m i t t e d 22 Se p t e mbe r 2022 D a t e o f t h e d e f e n c e 27 January 2023
P e r m i s s i o n f o r p u b l i c d e f e n c e g r a n t e d ( d a t e ) 17 November 2022 L a n g u a g e English
Monograph Article thesis Essay thesis
Abstract
Quality control is used in metal casting foundries to identify special cause defects and root causes
by correlating process input variations with casting defects. A difficulty exists in
associating collected process data with individual cast parts as the parts are processed through the
foundry and then out into the supply chain. Typically, manual identification of the castings with
route-paper based tracing approaches have been used. Such manual-based systems make root
cause analysis of quality defect issues tedious, prone to error, and limited in extent of supply chain
abled to be studied. This thesis aimed to provide improvements in three main problems which
remain as obstacles for the industrial 4.0 transformation of the metal foundry industry which are
respectively: providing understanding of the production means to form digitally readable
permanent cast part markings, of foundry process data collection and traceable association with
these cast parts, and of the necessary statistical quality control methods utilizing the newly
available comp le te p roduction data. In this the sis, a nove l se mi-automate d app roach is de ve lop e d
using additively manufactured 2D matrix code (AM2D) tags to be used as sand mold inserts that
forms a directly cast identification code into the parts. Furthermore, an Industry 4.0 digital foundry
serial production digital tracking and process data collection of the individual cast parts is
demonstrated utilizing the AM2D tags and captured part-by-part data, to enable a foundry wide
root cause analysis. Finally, state of art supervised machine learning (ML) classification models
are demonstrated using the data extracted by the cloud-based system to identify conditions that
predict defectives. The novel developments are demonstrated in a real foundry, manufacturing
automobile wheel via an aluminum low-pressure-die-casting process. The results in this thesis
ind i c at e t h a t A M 2D t ag s e na b l e a u t o mat e d p a rt t ra c ki ng t h ro u g h o u t t h e c a s t i ng p ro c e s s , and h e re
uniquely at the very early operations including mold making, enabling data collected in these early
operations to also be associated with each part processed. It was further confirmed the tag solution
worked within a foundry operation. An operator can uniquely mark, identify, and track the parts
throughout the foundry. This application enables a smart foundry quality management system:
simp le digital tracking ope ration via industrial table ts or mobile p hone s utilizing re adily available
barcode reader applications. Finally, the ML modelling approach presented maps the complex
relationship between adverse process conditions and defective wheel rim castings, which was
helpful in assisting process parameter tuning on new product pre-series production to lower
defectives. Overall, novel digital part marking, industry 4.0 part tracking and process data
collection, and ML based defective classification provides greater capability and understanding
of foundry quality management.

K e y w o r d s Industry 4.0, Smart Foundry, Part Marking, Machine Learning of metal castings
I S B N ( p r i n t e d ) 978-952-64-1099-9 I S B N ( p d f ) 978-952-64-1100-2
I S S N ( p r i n t e d ) 1799-4934 I S S N ( p d f ) 1799-4942
L o c a t i o n o f p u b l i s h e r Helsinki L o c a t i o n o f p r i n t i n g Helsinki Y e a r 2022
Pa ges 83 u r n http://urn.fi /URN:ISBN: 978-952-64-1100-2
 Tiivistelmä
Aalto-yliopisto, PL 11000, 00076 Aalto www.aalto.fi

Tekijä
Tekin Çağın Uyan
Väitöskirjan nimi

J u l k a i s i j a Insinööritieteiden korkeakoulu
Yksikkö
Sarja Aalto University publication series DOCTORAL THESES 206/2022
Tutkimusala
K ä s i k i r j o i t u k s e n p v m 22.09.2022 V ä i t ö s p ä i v ä 27.01.2023
V ä i t t e l y l u v a n m y ö n t ä m i s p ä i v ä 17.11.2022 K i e l i Englanti
Monografia Artikkeliväitöskirja Esseeväitöskirja
Tiivistelmä

A v a i n s a n a t torquent, per conubia, per inceptos, nostra, himenaeos, nulla

I S B N ( p a i n e t t u ) 978-952-64-1099-9 I S B N ( p d f ) 978-952-64-1100-2
I S S N ( p a i n e t t u ) 1799-4934 I S S N ( p d f ) 1799-4942
J u l k a i s u p a i k k a Helsinki P a i n o p a i k k a Helsinki V u o s i 2022
Si v u mä ä rä 83 u r n http://urn.fi /URN:ISBN: 978-952-64-1100-2
Acknowledgements

First, I would like to extend my heartfelt thanks to my supervisor, Professor

Kevin Otto, for granting me the privilege of working with him. His supervision
and teachings have been invaluable to my vision in scientific research and have
contributed significantly to my doctoral thesis work. Special thanks to Professor
Juhani Orkas for guiding me through my studies from the beginning, which es-
tablished the foundation for this doctoral research. I am also deeply grateful to
my colleague, Dr. Kalle Jalava, for serving as a great advisor and for his
knowledge and efforts in supporting my thesis work. Finally, yet importantly, I
would like to express my gratitude to my significant other, Yasemen, for encour-
aging me to pursue my dreams and for being my greatest source of support and
motivation during the highs and lows of this journey.

I am thankful for all the support I have received from colleagues in the Depart-
ment of Mechanical Engineering at Aalto University. I am also grateful for the
support I received from the FIN3D project. I would like to thank Cevher wheels
for providing the datasets and field support for a portion of my research. I am
also grateful to my lecturers for helping me advance my knowledge in my aca-
demic work. This doctoral work was made possible with the support of grants
from the Academy of Finland (Project Number 310252).

Espoo, 28 August 2022

Tekin Çağın Uyan

1
Contents

Acknowledgements ..................................................................................... 1
List of Abbreviations and Symbols............................................................. 5
List of Publications ...................................................................................... 6
Author’s Contribution ................................................................................. 7
1. Introduction ..................................................................................... 9
1.1 Motivation of the research........................................................... 9
1.2 Background and research gap ................................................... 10
1.2.1 Methods for DPM metal castings ............................................... 11
1.2.2 Data collection and Machine learning applications for quality
assessment in Foundries ....................................................................... 15
1.2.3 Challenges in Foundry Quality Control and Root Cause Analysis
17
1.2.4 Related Work .......................................................................... 18
1.3 Research gap and the objectives of the thesis .......................... 18
1.3.1 An alternative DPM for foundry part tracking and data collection
18
1.3.2 Foundry data management and machine learning in LPDC quality
improvement.......................................................................................... 19
1.4 Contribution of the thesis ..........................................................20
2. Methods .......................................................................................... 22
2.1 An alternative DPM- AM2D code inserts ................................. 22
2.1.1 2D Code generation, tag design and manufacturing ............... 23
2.1.2 Sand molding and casting and quality characteristics ........ 26
2.2 Foundry Data Collection and Part Tracking Using AM2D Tags26
2.3 Defective Part Classification and Process Optimization via Supervised
ML 28
2.3.1 Feature selection .................................................................... 28
2.3.2 Foundry data collection ......................................................... 30
2.3.3 Data preprocessing................................................................. 30
2.3.4 Model evaluation .................................................................... 31

3
 2.3.5 XGBoost binary classification implementation .................... 31
3. Results and Discussion.................................................................. 33
3.1 An alternative DPM- AM2D code inserts ................................. 33
3.1.1 Verification Across Multiple Identifiers, Parts and Operations35
3.2 Foundry Data Collection and Part Tracking Using AM2D Tags36
3.3 Defective Part Classification and Process Optimization via Supervised
ML 38
3.3.1 Feature Importance ............................................................... 40
3.3.2 Model trade-off between good and defectives ..................... 42
3.3.3 Model comparison ................................................................. 42
4. Conclusion...................................................................................... 43
4.1 Research question 1 ................................................................... 43
4.2 Research question 2 ................................................................... 44
4.3 Research question 3................................................................... 44
References ................................................................................................. 47

4
List of Abbreviations and Symbols

DPM Direct part marking

IoT Internet of things

OCR Optical charagter recognition

ML Machine learning

RFID Radio frequency identification

AM2D Additively Manufactured Two-Dimensional Matrix Code

5
List of Publications

This doctoral dissertation consists of a summary and of the following publica-

tions which are referred to in the text by their numerals.

1. Uyan, Tekin; Jalava, K., Orkas, J., & Otto, K. (2021). Sand Casting Imple-
mentation of Two-Dimensional Digital Code Direct-Part-Marking Using Addi-
tively Manufactured Tags. International Journal of Metalcasting, 1-12.
DOI: 10.1007/s40962-021-00680-x

2. Uyan, T., Jalava, K., Arasola, L., & Otto, K. (2022). Foundry Data Collec-
tion and Part Tracking Using Additively Manufactured Digital Code Direct-
Part-Marking Tags. Applied Mechanics and Materials, Journal Vol. 911

3. Uyan, T. Ç., Otto, K., Silva, M. S., Vilaça, P., & Armakan, E. (2022). Indus-
try 4.0 Foundry Data Management and Supervised Machine Learning in Low-
Pressure Die Casting Quality Improvement. International Journal of Metal-
casting, 1-16.
DOI:10.1007/s40962-021-00680-x

6
Author’s Contribution

The author is the corresponding author of the present Thesis and the Publica-
tions 1-3. The corresponding author mentioned as the first author is Tekin Çağın
Uyan.

Publication 1: Sand Casting Implementation of Two-Dimensional Digital

Code Direct-Part-Marking Using Additively Manufactured Tags

In this study together with co-author Kalle Jalava the first author’s contribu-
tion consists of ideation and development of a novel cast part marking experi-
ment and methodology. The first author prepared the setup and performed the
tests. The results were interpreted and commented on together by the co-au-
thors.

Publication 2: Foundry Data Collection and Part Tracking Using Additively

Manufactured Digital Code Direct-Part-Marking Tags

The first author's contribution consists of the design of experiments, sand-

casting Smart Foundry operation setup including automated scan-based track-
ing of cast parts through the foundry and supply chain, and data analysis inter-
pretation. The co-authors took part in discovering the process features, AM2D
tag production, testing, casting, and results interpreted and commented to-
gether by the co-authors.

Publication 3: Industry 4.0 Foundry Data Management and Supervised Ma-

chine Learning in Low-Pressure Die Casting Quality Improvement.

The first author's contribution consists of the arranging necessary protocol to

start an industry-academic collaboration. The first author performed, foundry
production data preprocessing, foundry field visits, data analysis, and machine
learning model implementation. The co-authors took part in metallurgical and
statistical expertise, and field support, and the results were interpreted and
commented on together with the co-authors.

7
1. Introduction

1.1 Motivation of the research

The metal casting foundry industry causes major environmental concerns air
and solid waste pollution, and natural resource depletion becomes major con-
cerns for society. There are serious health risks associated with poor air quality,
especially around large cities caused by point sources such as metal foundries.
On the other hand, primary metals from ores are used across a wide variety of
casting products from the torque arm of a wind turbine to an automobile wheel
rims. The metal resources on the earth are also limited and therefore recycling
of metal products is a must. One bottleneck is that the recycling industry has
difficulty tracing the metal waste and separating it into raw metals stock. Mini-
mally, know what elements are within any recycled part could enable more tai-
lored use of that metal content.
Moreover, metal casting is quite costly to the foundries. Running the heavy in-
dustrial equipment requires a vast amount of energy and manpower. Also, the
quality requirements are increasing by massive scale of production and growing
industrial demand for more specifical metal products. There is a need of defect
reduction and determining and eliminating the root causes of defects in metal
castings to prevent at simplest level; the energy waste and manpower used in
reworking defectives, handling scrap, as well as to meet the high-quality stand-
ards of the industrial society. A robust digital unique identification system and
part data repository is therefore required to provide the relevant information on
the metal cast product. Furthermore, to decrease the pollution composed by the
production failures of the Foundries, a sustainability improvement needs to be
introduced via implementing Industry 4.0 tools such as digital part marking,
data collection, and data analysis through machine learning models that would
minimize production failures. A Smart Foundry concept that collaborates with
IoT sensors, part marking systems, process data is needed to achieve the high
standard sustainability, quality assurance and green industrial goals. Such im-
provements would help to decrease heavy industry’s negative impact on the en-
vironment and foundrymen.
This thesis goes as a demonstrator of smart foundry applications.

9
1.2 Background and research gap

It is quite challenging to trace once the metal is extracted from the ore through
the refineries, foundries, OEM, to the end users, and recycling facilities. One of
the main agreed indicators of such challenge is global supply chain where the
interaction in main processing units of metals is quite complex. Even in con-
trary, internal recycling of failed products’ traceability can also be quite chal-
lenging e.g inside a metal foundry department. Figure 1 represents a metal’s
journey from cradle to grave. One can imagine the potential if it would be pos-
sible to know everything across the earth’s metals lifecycle and to track each
phase of the metals. It could have been used for a more sustainable metal pro-
cessing and tinier carbon footprint. However, there are big gaps in metals pro-
cessing in terms of information loss. What is the bottleneck in data collection?
Where do all the useful data go? How can big data be used to improve the sus-
tainability of metal resources on our planet? How to make efficient recycling or
even preventing metal usage? To find sophisticated answers to these questions,
a digitalized web of metals and big data is needed, and metal foundries are a big
part of this web.

Figure 1. Lifecycle assessment of metals through digital identification from

cradle-to-grave

A brief explanation of a metal foundry is a factory that produces metal castings

which are cast into dedicated shapes (of a wheel or a valve) by melting the raw
materials (ores and turnovers) into a liquid, pouring the metal into a mold (sand
or die), and removing the mold material after the metal has solidified as it cools.
The Smart Foundry concept is being developed in the metal casting industry to
improve this workflow and quality. An example of this would be the implemen-
tation of such technologies as the internet-of-things (IoT) and data collection,
along with process data analytics, to track and improve the quality capability
data and obtain higher quality metal castings [1]-[2]. A further aspect of this is

10
 Introduction

process automation, which enables parts themselves to indicate the need for any
necessary process setups. In order to facilitate these efforts, object identification
and tracking technologies are needed to enable part interaction with production
systems with real time data transfer. To do so, an intelligent object must be
equipped with a unique digital identification, such as barcodes, 2D matrix
codes, RFID labels or other [3]. However, still nowadays many foundries con-
tinue to use manual based route papers for part identification, tracking and data
collection [4]. As a result, statistical quality control (SQC) in foundries is tedious
and statistical process control (SPC) are made mainly on workstation base
which makes it hard to be generalized. Yet, workstation base process data col-
lections is tremendously valuable in enabling SQC and requires high technology
equipment that are capable collecting and monitoring real-time data in harsh
foundry conditions such as robust sensors working in extreme temperatures. On
the other hand, workstation process data are often remained in limited ability
to being utilized in SPC when it is not matched up with actual end products
which are the key quality indicators. When the target is missing the data re-
mains unsupervised. Therefore, the part tracking and process measurements
offer correlation which enables the state of art supervised machine learning and
artificial intelligence.

1.2.1 Methods for DPM metal castings

Foundry process steps such as pattern making, molding, melting, casting, etc.
have extreme environments and heavy industrial conditions that make tracing
individual materials, consumables, and castings difficult. Given this difficulty,
cast part markings are often made using identifiers on the pattern plate or mold
inserts. Figure 2 shows an example of casting with classical alphanumerical la-
beling which has been the marking method of many foundries regardless of pro-
duction type and size. This method creates challenges in almost every stage of
the casting. For example, each casting can have an identification number in-
cluding date and logos, where marking patterns must be adjusted for each of the
castings. As such, most part tracking does not initiate until after casting is made.
As can be seen in Figure 2(right) the failure examples of the classical part mark-
ing method, the part is shown after its sand mold shakeout as Part ID is spread
in a long area and has partly sand-metal penetration defects on ID letters. In
such cases, next to identification challenges this also creates an extra challenge
for the fettling and cleaning process. One other negative side of this approach
is, that when the product is ready for shipping, there are plenty of letters to carry
and check on a foundry product, which causes inefficiencies in foundry-cus-
tomer communications as well.

11
Figure 2, Plastic letters, and numbers for metal identification in sand molding
(left). Traditionally marked part (right).[5]

A method to permanently label parts including serial numbers and other nec-
essary information called a Direct Part Marking (DPM). A direct part marking
is looked for that allows labeling on the early stages of the casting operations
and provide access to digital identification for IoT process control. Identifica-
tion of each cast component enables selection and individual handling of the
castings. The most advanced part marking systems are tending to use RFID tags
[6]-[7] and 2D matrix codes [8] directly on the metals to digitize the ID numbers
into the castings. These methods create benefits over traditional alphanumeric
part marking methods. A specially designed RFID is embedded in the metal
casting during the die casting process (see Figure 3). A laser direct part marking
(see Figure 4) with compact 2D matrix code enable fast marking during the pro-
duction and identifying the parts via DPM readers while enabling digital envi-
ronment for castings. An other important benefit of this is, ID numbers occupies
less space on a cast part compare to traditional alphanumeric numbering. Ad-
ditionally, the technologies used for creating digital DPMs enable the automated
ID dedication, by other means dedicating ID numbers through ERP systems.
This decreases the human error risk and confusions between the departments
of the Foundry.

Figure 3. Positioning and Fixing the RFID Transponder with Insulation in the
Die Casting Mold [7]

12
 Introduction

Figure 4. A data matrix and letters etching via laser on metal.[8]

With laser marked parts, optical character recognition (OCR) technology can
be used to read the alphanumeric characters (see Figure 5). Also, 3D tech-
niques are developed utilizing a stripe projection to allow reading of shape in-
formation rather than OCR [9].

Figure 5. Laser marked sand cores (top); a laser engraved sand mold casted
part (middle) and reading result (down) [9]

On the other hand, the difficulties in adopting direct part markings include
potential for failure of such marking systems within the harsh foundry environ-
ment. For instance, shiny metal surface reflectivity can also cause unacceptable
reading rates for OCR. One other example, laser marking of the castings by car-
rying the big and heavy parts to dedicated laser marking rooms or carrying the
marking device inside the factory may not be suitable for such foundry condi-
tions. Further, laser safety risk can become important in a human-powered
foundry environment [10]. Moreover, embedded RFID technology is still under
development for metal castings. The vulnerability of the RFID technology over
high temperatures, lower signal distance, and distracting metallic environment
create quite much of limitations and challenges for its feasibility applications in
foundries.

13
Reconfigurable pin-type tooling, which enables direct part markings in green
sand permanent molding machines, is another part marking technology [11].
Figure 6 depicts the procedure of applying a pin tool to the inside of the mold
to generate a Datamatrix symbol design. Each casting can be assigned a unique
DPM using this manner. The robustness of the matrix patterns created by em-
bossing the sand patterns in this method was also investigated. It has been re-
ported that dots formed as bumps fared well during the casting process (see
Figure 7). Part markings made as dimples, on the other hand, were less struc-
turally sound and caused casting errors. Because of the compression type of
molding equipment is vertical, this technology is only possible for one type of
production method. It has dimensional constraints regarding the molding ma-
chine since bigger parts still require hand molding. However, it can still be
used as a marking core, which may also create an additional step in the casting
process. Nonetheless, it must be considered if the surface embossed(bumps)
sand cores could survive in a foundry environment with the millimeter data
matrix pattern since dimple data matrix forms are problematic with this
method.

Figure 6. Pin tool DPM application for vertical sand molding machines. [11]

Figure 7. Pin tool DPM application for vertical sand molding machines. [11]

Using stencils composed of mineral fiber with clay and latex as a binder is an-
other method of direct part marking. The hot-melt adhesive is used to adhere
this to the sand, and the stencil can be applied to a variety of metal kinds. This

14
 Introduction

offers DPM patterns that are permanent and can be used for castings in sand,
lost foam, investment, and permanent molds. Using an automated laser engrav-
ing device and a specific box, the stencils can be prepared as pre-encoded [12].
The preparation of the mold surface, spraying heat-resisted adhesives on the
stencil, and the contrast painting of the DPM codes after manufacturing are a
few manual subprocesses that are final but crucial. Additionally, to sub pro-
cesses, consumables such as thermal spray prints need to be used to provide
contrast to DPM codes.
Utilizing CNC machined pattern inserts is a different approach. Contrast dif-
ferences for two-dimensional code readability when using bumps or dimples in
CNC machined plastic permanent patterns for sand molding is studied [13]. It
is discovered from the readability tests that dimples consistently had higher
contrast and better readability than bumps, in contrary to the pin type tooling
method discussed previously.
In addition to markings embedded in the part at the onset of production, there
are also post-production marking methods such as laser-engraving technology
[14], pin impact marking [15] or even simple painted stencils. In contrast to the
methods has been discussed so far, such methods are used at the end of the pro-
duction phase and used for part tracking in production steps after casting oper-
ations.

1.2.2 Data collection and Machine learning applications for quality as-
sessment in Foundries

In this first section Low Pressure Die Casting (LPDC) is explained over a specific
product – automobile wheel, to give an insight about the machine learning study
as a part of the thesis research. In the second section, the previous studies on
the topic of ML applications in castings are explored.

Low Pressure Die Casting Process

Generally, two die cavities are arranged above a holding furnace that is heated
electrically and installed in the lower portion of LPDC machines. One cycle of
the casting process entails pressurizing the holding furnace, which contains the
molten aluminum that is pushed to fill the mold cavity, followed by mold cooling
increased by air and water entering through cooling channels and ultimately di-
rectional solidification of the cast [16]. The cooling channel’s locations em-
ployed in this study are shown in Figure 8 as intersection of the LPDC mold.

15
Figure 8. Intersection of a die mold of LPDC machine

Material loading, melting in a crucible, degassing, batch transfer to an LPDC

machine, casting into a die mold, directed cooling, X-ray inspection, heat treat-
ment, and further component cleanup and inspections are all part of the found-
ry's entire production flow. The LPDC system is the foundry process that has the
most impact on component quality, hence this study focuses on it. Figure 2 de-
picts the primary process flow for an LPDC machine.

Figure 9. Low pressure die casting process cycle.

Three primary phases consist in the casting process in Figure 9. Phase 1 of the
process involves applying pressure to the molten metal in the fill tube to force it
up to the runner, and Phase 2 involves pouring the molten metal into the cavity
of the die mold. On the length of these phases, there is a threshold that may be
changed. To prevent any turbulence of the liquid metal that might result in en-
trained air and porosity problems, it is crucial to fill the mold cavity gently.
However, if the metal fills too slowly, temperature variations between the metal
and the mold will alter the liquidity and lead to both cold and hot tearing fail-
ures. Increased feeding pressure is used during Phase 3, and directed solidifica-
tion is integrated using cooling channels inserted into the die mold (Figure 8).
These cooling conduits may have passageways or bubbler-style designs.

Many time series data, including pressure and temperatures in the cooling
channels used for directional solidification, are tracked, and archived as part of
the LPDC process. These measurements are made every second of a cycle time

16
 Introduction

that lasts around five minutes for each wheel-rim. Figure 10 depicts an example
time series data for the molding pressure that spans two cycles that produced
two castings. The operator and the environment can affect the dwell duration
between cycles.

Figure 10. Time series pressure measurement on LPDC machine

1.2.3 Challenges in Foundry Quality Control and Root Cause Analysis

Traditional quality control methods include visual inspection, destructive and

non-destructive testing, also part by part dimensional specification control
combined with manual and digital coordinate measuring tools and machines.
SPC adjustments on process variables are done usually by the operator in work-
station level according to the feedback received from quality control indications
of upper and lower specs studies on parts. The correlation analysis process var-
iables and related cast wheel is often missing. The defectives are, therefore can-
not been properly classified in root causes.

It is difficult to build machine learning models using manufacturing data to

forecast the commencement of faulty components for several reasons. There are
barriers of data silos within departments of a foundry, meaning different stor-
age, datatypes, timestamps and tracking methods applied in departments from
melt-shop to quality control. Additionally, there plenty of available process
measurement points from controllable and uncontrollable process variables,
but a limited set of part inspection points.

The first challenge is the difference in time scales of the data and locations
within the foundry where the data is generated and stored. Material data is at
the batch level (not part by part) and held within the melt shop. The process
data is at a one second time interval and held within each separate LPDC ma-
chine. The inspection data is at the part level and held within the x-ray machine.
These datasets are typically siloed within the departments and rarely shared.

17
The data's intricacy itself presents another difficulty. The manual entry of some
data, such as material qualities, into excel spread sheets. LPDC machines, on
the other hand, process terabytes of time series data per day. The data record
formats are stratified, which makes it difficult to process the data into a coherent

Finally, as quality improvements are applied, the data structure of passed and
fail parts become more imbalanced with fewer failed parts. Machine learning
algorithms have a harder time analyzing such imbalanced and biased datasets.
Further, any data entry errors increase the sensitivity of the machine learning
system to accuracy.

1.2.4 Related Work

Based on the experimental data, the equations are developed for high pressure
die casting (HPDC) [17]. These equations were used to artificially generate a
sufficiently large data for training the Neural Networks by selecting the values
of the input variables randomly. Also, an artificial neural network (ANN) model
is trained using data generated by FEM-based flow simulation software [18].

In recent sophisticated work, HPDC mechanical properties of a die cast tensile

testing machine bar, and the relationship to process features is studied using
large dataset spanning many months of production [19]. The regression models
of Random Forest, Support Vector Machine (SVM), Neural Network, and
XGBoost are compared. The focus in this study was on large datasets spanning
many months of production. Random Forest Regression found to be best per-
forming model for regression fitting the ultimate tensile strength of produced
parts. Also, classifying good parts and defectives for HPDC of tensile testing ma-
chine bars is studied [20], as well as illustrated machine learning application by
a case study using sand cast foundry data [21].

One of the earlier studies on LPDC, computational process simulation modeling

is used, and minimizing various defects in LPDC of aluminum alloy automotive
wheels via simulating heat transfer and fluid flow during die casting is studied
[22]. An experimental L-shape thin-walled casting is studied with a laboratory
LPDC machine to predict and optimize the part quality [23]. This work was early
demonstration of using machine learning in LPDC. The artificial neural net-
works are used for modelling and genetic algorithms for optimization. The
training data was collected from numerical simulation and tested on an experi-
mental part under laboratory conditions.

1.3 Research gap and the objectives of the thesis

1.3.1 An alternative DPM for foundry part tracking and data collection

The foundry industry is using high-technology process equipment for quality

purposes. Generally, the data collection from the equipment and sensors is quite
sophisticated. On the other hand, a limitation arises in using the process data

18
 Introduction

for quality purposes when individual parts cant be associated with the process
data. To implement the digitalized quality control and part compliance certifi-
cation, there is a need for permanent digital part markings. Along with many
challenges for marking a cast part explained in section 1.2.1, there is a need for
an in-cast part marking for the casting industry. Two objectives of the thesis are
creating a novel and in-cast DPM method as an alternative to its peers, and
demonstrating an Industry 4.0 data collection in a foundry with this new
method that covers all in one below:

- part marking starts in the early stage of the production.

- can be applied in multiple casting methods.

- relatively cheap to use, robust for foundry environment, and rapid system that
copes with production speed.

- easily readable DPM e.g via common tools such as mobile phones or indus-
trial tablets,

- the data collection and association between process inputs and individual
parts for improved quality assurance purposes.

1.3.2 Foundry data management and machine learning in LPDC quality

improvement

This part of the research aims to demonstrate the usage of foundry process data
for modern ML algorithms to identify causes of casting defects. The research
makes use real foundry data from LPDC foundry. The objective is to improve
quality assurance with industry 4.0 data management and machine learning
tools by filling the gaps in the related research field. The objective’s scope is ex-
plained below:

- The related studies were conducted for mostly injection molding and HPDC
which is different than LPDC in terms of process parameters, casting physics
and production parts. There is relatively less application of ML in LPDC. Our
objective here is to focus on LPDC initial production to help set up the process.

- Studies showed porosity defects were related to input parameters such as melt
and mold initial temperatures and first and second phase velocities. These input
quantities are insufficient to cover all the effects of real LPDC machines. The
scope of research is covering complex relations among the process variables
with measurement points both from machine and actual sensor values.

- Studies used computational process simulation as datasets. However, these

input quantities are quite difficult to measure for real LPDC machines. The ob-
jective is to include machine measurement points such as actual pressure, cool-
ing channel’s activation time, silicon content of molten metal etc.

- Studies used data from a casting laboratory and tested on an experimental part
under laboratory conditions. This study aims to analyze an automotive wheel
that is widely used engineering product.

19
- As a reason of high-quality production, the real foundry datasets have imbal-
anced data structure that is skewed with fewer failed parts, this research focuses
on imbalanced datasets.

- The related work made use of bigger datasets which is collected through
months and years. However, in case of a new product is in pre-series production
phase for process parameters adjustment purposes further serial production,
foundries generally can not collect big amount of data to generalize the machine
learning models. Therefore, there is a need for a study with relatively smaller
datasets. This study aims to use a pre series production dataset.

Research questions

• What are limits in DPM size and process conditions for mark a cast
part? What are the experimental steps to improve DPM to cover the
weakness of the current systems and can be efficiently used in the
foundry process?
• What are the process requirements to implement DPM collabora-
tively for data collection and part tracking in serial production to im-
prove understanding of statistical quality control in the foundry
products? What are the necessary frequency and parameters to be
monitored for statistical quality control?
• What would be the ML model implementation to classify defectives
and highlight the important process variables that have the most in-
fluence to discover the complex process variables causing of defective
castings.

A summary of the scope in three main objectives are fundamental of this re-
search structure and article per each objective is aimed.
OB1: Investigating the technologies to propose a pre-cast direct part marking
method for casting products and their digital identification via unique code.
OB2: Building a unique system to demonstrate data collection and part track-
ing inside a Foundry.
OB3: Advanced machine learning algorithms to root cause analysis of the fail-
ure of the metal casting and avoid failures by supervised machine learning based
prediction models.

1.4 Contribution of the thesis

Publication 1.
The size limits and process conditions whereby metal parts can be successfully
marked using additively manufactured 2D matrix code shaped plastic tags are
understood. A novel marking system was demonstrated by making use of these
AM2D tags for directly in the molding process. Marking limitations of several
industrial parts with different dimensions are discovered. Feasibility limits of
different 2D matrix coding standards are examined via readability tests on the

20
 Introduction

produced parts. This work led to understanding of how small a dot can be sand
casted, what format (dimple or bumps) and what process conditions (tempera-
tures, part thickness, etc.) are necessary to create a DPM with AM2D tags. Over-
all, developed method proved to be a successful DPM method and showed sig-
nificant advantages over the other available pre and post DPM methods.

Publication 2.

Several DPM techniques, data management in foundries or root cause analysis

of casting defects are available from the literature, but collaborative research is
missing in terms of how part marking, tracking and data collection can enable
root cause analysis. This work is demonstrating an Industry 4.0 data collection
in a foundry and first time in the literature a required part marking, process
measurements and tracing of data all together utilized collaboratively to execute
root cause analysis of casting defect. A simple method for part marking, data
collection and part tracking through quality control are successfully presented.
This work led to understanding of what variables to collect in what frequency,
what statistics of the data and how to time stamp to enable correlation of the
input and output data features. Improved understanding on causes of defects in
castings with help of the data collection, effect of variation analysis and part
tracking through the foundry. Making castings digitally readable with AM2D
tags technique may help solving the individual digital part tracking accessibility
problem, which remains an obstacle for many small to medium-sized foundry’s
industrial 4.0 transformation.

Publication 3.

Supervised ML for classification of casting porosity defects using real-time

process data of the casting equipment have not previously been studied to un-
derstand how porosity defects are related to the LPDC process variables. Rela-
tively little data on ML application on LPDC. Many a work on the topic uses
computational process simulations as datasets and experimental laboratory
parts and equipment are used in the studies. In this work machine and metal-
related features were utilized to predict porosity defectives of LPDC processing
of aluminum alloy high-performance automotive wheels. This work led to un-
derstanding of what algorithms work and do not work for classifying LDPC de-
fects and what algorithms work to identify prime causes with non-orthogonal
data. The presented machine learning models on production process data of-
fered the ability to improve statistical quality control. The causes of casting de-
fects were identified. This study differs as being representative of an industrial
foundry with a machine learning application on a small dataset to help establish
process settings.

21
2. Methods

The methods developed and employed during the study are discussed in this
chapter. In the section 2.1, the AM2D methodology developed in Article I and
its foundry implementation in Article III is presented in sections 2.2. In section
2.3, methods adopted in Article II for implementing a supervised ML applica-
tion for LPDC quality improvement is explained.

2.1 An alternative DPM- AM2D code inserts

The process of creating an additively manufactured wax and plastic tags for dig-
ital identification metal castings is explained in this section. The main goal was
to develop a rapid part marking system with consumable tags that can complete
as a two-dimensional matrix code for direct application in traditional evapora-
tive-pattern casting. There are several methods available for augmenting a tra-
ditional casting process with a consumable tag. With investment casting using
wax patterns, a consumable tag can be easily added to the pattern and the tra-
ditional lost wax process used. With sand casting, a plastic/wax tag can be added
to the pattern and a sand mold created. Yet, a third option is to apply sand mold-
ing with a pattern to which the tag is inserted but not attached. The tag can re-
main in the sand mold and burn away upon casting. This third option of using

22
 Methods

tags is explored here, using wooden patterns with the tag remaining in the sand
molding.
The main issues addressed during the development of a part marking research
methodology starts with to exploring the necessary steps including:

1. 2D matrix code generation

2. Converting to 3D tag design
3. Additively manufacturing of the tags
4. Traditional sand casting
5. Readability testing of the unique ID numbers
6. Conformation and verification of unique ID codes in several parts

2.1.1 2D Code generation, tag design and manufacturing

The 2D code matrix design includes symbology concerns over size and complex-
ity. The 2D code concerns for cast part markings include factors such as the di-
mension of the DPM frame and the number of small squares within the matrix
needed to digitally represent an ID code. Smaller dimensions are better for
small castings, but larger dimensions may be necessary for readability and
castability of a marking. Further, part identifiers with more digits require more
dots and hence a more detailed 2D code.
Another concern for cast part markings using 2D codes is the coding standard.
There are many coding standards available, both proprietary and public do-
main. In this study, the Data Matrix Code and Dotcode standard is explored for
several reasons. The DataMatrix code algorithm has error correction, which
means it can still be readable via DPM readers if some part of the code is dam-
aged during harsh foundry environment or for other reasons. Also, a dotted pat-
tern is available with these standards. The square to dot conversion provides a
better castability than edged patterns from casting physics point of view. It is
known from the literature that square Datamatrix creates contrast issues which
have a negative effect on readability when etched onto metal surfaces. The
standards do not expand the dimensions linearly with increasing a ID code digit
up to the required ten digits of encoding. Foundry surveys indicated an alpha-
numeric string of ten digits are adequate for identification purposes [4]. Dot
choice is an option in Datamatrix code generation software. Figure 11 shows the
difference between the square version and the dot version of the Datamatrix.

23
 Figure 11. Data matrix code square and dot patters nfor the same encoded
string.

On the other hand, Dotcode standard is inherently made of dot. It has diago-
nal and adjustable dot spacing. This allows for improved sand molded surface
quality given the prevention of possible molding defects from loose sand at-
tached to any. This coding, therefore, expected to be most readable as a casting.
Based on this, these two coding standards for direct part markings on castings
are explored and compared. As a demonstration, several part identifiers en-
coded with 10 decimal digit complexity are considered. For example, a ten dec-
imal digit number might be ‘1234567890’. The difference of the Datamatrix dot
pattern and the Dotcode matrices for the same 10 digit can be learned from Ar-
ticle 1.
The addition of alphanumeric including letters and numbers also changes the
digital coding complexity. More complex strings can be used to include text such
as a URL, but the tag size grows accordingly. One comparison is for ten digits,
the Datamatrix have less dots than Dotcode and it gives an advantage over
Dotcode since less dots to fill and fail during the casting. The Dotcode’s default
rectangular pattern with an aspect ratio of 12 columns by 19 rows while the
Datamatrix has 12 dots per side can be seen in Figure 12. The two code-type
matrices were generated for testing purposes using common matrix code gen-
erator software.

Figure 12. Data matrix (left) Dotcode (right) codings.

24
 Methods

One obvious consideration is the dot size: the dot diameter and the dot spacing.
Smaller and tighter dot geometry results in smaller tags, but also a tag that is
more difficult to consistently fabricate and cast into a resulting part marking.
The sand-casting process and sand used had a grain size of 0.2 mm, and a min-
imum of three sand grains creates a castable shape, and therefore, a minimal
castable thickness was determined 0.6 mm. Therefore, the minimum dot and
bump size were deemed 0.6 mm, and the minimum separation between dots
was deemed 0.6 mm. These dimensions were applied for a minimum tag size,
with larger dots and spacing also studied. Given the tolerance capability of sand
casting at less than a millimeter over the tag dimensions, a dot size and space at
2x larger were chosen. Also, tags at smaller and larger dot diameters of 1.5, 2
mm and 2.5 mm were created. Another geometrical decision is whether to use
dimples out of the tag or hole bumps into the tag. Either can generate visual
outline shadow pattern that an optical code reader could image and decode.
There are also further refined detailed geometric dimensions necessary to a tag.
More detail in dimensions can be found in Publication 1.

For rapid and mass-customized generation of a wax tag that can be added to a
traditional wax pattern, the idea of a rapid additive manufactured tag was ex-
plored. The wax tag was printed using a 3D wax printer with dimensional accu-
racy of ± 0.0254 mm, the ProJet MJP 3600W printer. Figure 3 depicts the bump
and dimple alternatives of the Datamatrix tags designed and printed as wax
tags. The wax material used was a purple-colored wax material.

25
Figure 13. Upper figure- Bump (left) and dimple (right) datamatrix tag designs.
Bottom Figure- Bump dotted (left) and dimple dotted (right) data matrix wax
printed tags.

The resulting tags appeared consistently high in surface quality. In bottom Fig-
ure 13, notice the contrast differences in the image of the tags themselves, as
bumps or dimples. As a check, the printed wax tags for readability were tested
with consistently positive results for both the Datamatrix and Dotcode tags.
Read attempts were successful with either dimple or bump dots.

2.1.2 Sand molding and casting and quality characteristics

The next step is to combine the created wax tag with the pattern fabricated to
cast the part. Our pattern is a CNC machined wooden block structure. The wax
tag was attached to the wooden pattern with a small amount of adhesive tape.
With this pattern complete with wax tags, the sand mold can be fabricated. Our
molding sand was 0.2 mm quartz sand with an ester hardened alkaline phenolic
binder, hand molded. When the pattern was removed from the upper mold, the
wax tags remained with the sand mold and gating is arranged accordingly to the
bottom mold. During the pour, the wax simply burned away. Subsequent to the
molding process, the test parts were cast with molten recycled Aluminum at 740
̊C. The casting result was air cooled in the foundry laboratory.
An industrial scanner (Honeywell 1920i) and a mobile phone camera (Sam-
sung S4) with a free barcode reader application were both used as image sensor
for readability tests of the DPMs. An enclosed photo booth was used capture
images according to ISO/IEC DIS 29158 standards [24]. A Pass/Fail read test
was used to assess the readability of DPMs. Two randomized repeat tests were
done per condition, both equally for the industrial reader and the mobile phone.

2.2 Foundry Data Collection and Part Tracking Using AM2D Tags

The process of creating an AM2D identification tags and its implementation

on to different type of castings parts have been explained in the previous section.
Publication 1 developed the process parameters and settings needed to effec-
tively cast direct part markings using sand casting. In this section, the serial
manufacturing of a rotor using AM2D tags in a sand-casting is studied in a
foundry operation.

The casting produced is a small aluminum rotor, as shown as a drawing in

Figure 14 as well as a depiction of tag housing and casting gating system used.
The parts were cast in batches of four.

26
 Methods

Figure 14. Rotor Casting used for study.

The foundry operations used to create this part include pattern making, mold-
ing, casting, and quality inspection. The first step is pattern making, where the
rotor pattern was fabricated using additive manufacturing and combined with
the gating required for four batch casting. Next the 2D codes are read into 3D
solid modelling CAD software and extruded into 3D tag shapes. Then the tags
represented by the 3D solid model are additively manufactured (AM). AM al-
lows for very easy manufacturing of many different and unique tags. The AM
process is also sufficiently precise to enable tags that can be added to casting
patterns. Here, the more environmentally friendly PLA (poly-lactic acid) is used.
PLA has a relatively low melting point, 150–160 °C, thus requiring less energy
to print with the material. In addition, PLA- plastic has been shown to be a safer
alternative to the possibly toxic ABS plastic [25].

Plastic tags were printed to a dimensional accuracy of approximately ±0.1 mm

using an Ultimaker 2+ printer. The AM2D digital identification tags were added
to the pattern, one unique tag for each part. The molding is a two-part mold
where the upper includes the AM2D identification tags inserted on the mold,
and bottom mold includes the rotor surface. A set of four rotor patterns were
used in each mold, with the tag placed at the top of the mold. The riser supports
are included next to each tag to ensure adequate venting and feeding, to allow
the AM tag gases to vent as the tag burns away. After molding, the next step is
the casting process which includes melting and pouring. The melting process
was done using an induction furnace and the raw material stock was secondary
aluminum ingots and foundry returns. A batch of molten aluminum was used
sequentially over multiple molds. After pouring, the metal solidifies. Once solid,
the casting was removed in shake out. It was cleaned and fettled then each part
was cut from the gating. After completing the fabrication, the parts were quality
inspected. This includes AM2D DPM readability through image processing for
dimensional accuracy.
Variations occur throughout the part casting process. Each step includes a
separate set of variables that can impact ultimate casting quality. A process fail-
ure modes and effects analysis (FMEA) were done to identify possible sources
of part quality variability. From the FMEA, 56 input variables were selected for

27
data recording. The important variables are summarized in Figure 15 which pro-
vides a sequential view of the process and factors that influence each step. De-
tails of the affects of process inputs and variations are provided in Publication
2.

Figure 15. Process Variations as Causes and Their Effect on Part Quality

2.3 Defective Part Classification and Process Optimization via

Supervised ML

The objective here is a method to collect LPDC foundry process data and use
machine learning models to correlate the process parameters and tolerances
with the occurrence of critical defects that will invalidate a casting part from
being used. The method includes machine process data collection, part tracking,
and quality inspection of defect data.

2.3.1 Feature selection

Numerous casting defects might appear during the LPDC of aluminum wheel
rims. Gas and shrinkage porosity, shrinkage, cold shut, and foreign materials
are among the most frequent failures [26]. Instead of focusing on all defects,
porosity defects are the subject of this study because isolated defects that statis-
tically fit will correspond better to certain process conditions [27].
Using X-ray scanning, the whole surface of the wheel rims is checked for po-
rosity void defects, which are visible as elongated and circular dark spots. The
X-ray system in use conforms with the specifications outlined in, two non-de-
structive testing standards DIN EN ISO 19232-1 [28] and DIN EN ISO 19232-5
[29]. In Figure 16, an example defective wheel rim is displayed.

28
 Methods

Figure 16. X-ray scan of die cast wheel rim with a porosity defect on a spoke.

As such porosity defects are often more prone to happen at specific periods of
the year, such as summertime with overall higher temperatures and higher hu-
midity levels [30]. Therefore, the data shown in this example are from wheel
rims manufactured six days in a row. The foundry runs for 24 hours with three
shifts. The result was 1077 parts.
Studies on the process parameters have indicated that the amount and size of
the porosity may be reduced by lowering the melting temperature. Additionally,
a shorter holding period and more pressure applied encourage smaller grain
size, which minimizes porosity [31]–[32]–[33]. The metal's temperature and
pressure are measured factors in this investigation.
Elemental silicon is the alloying component of aluminum that is utilized the
most commonly to balance off aluminum's shrinking during solidification [34]–
[35]. The silicon concentration is a monitored variable that is significant in po-
rosity failures.
Another set of defect causes include directional solidification effects. It is im-
portant to ensure that a controlled, nonturbulent fill exists to eliminate air en-
trapment. To do this, a ready supply of molten metal in the direction of solidifi-
cation is needed to reduce the shrinkage. The use of a metal die with integral
cooling passages allows for controlled cooling of the casting. This can further
improve mechanical properties through refinement of the microstructure [36].
In this process there are both water and air cooling channels whose flow rate
and activation times are monitored and -in some cases- can be adjusted accord-
ing to the most recent X-ray results. This operator-to-operator skill at adjust-
ment compensation of the water and air cooling could be a source of defects or
lack of defects. Machine learning can help quantify proper water and air cooling
times.
Finally, molten metal properties can also affect the generation of porosity.
Therefore, the quality of an aluminum sample ought to be evaluated through the
density index (DI) [37]. In this study the DI is included as monitored variables.

29
2.3.2 Foundry data collection

The data discussed here was as collected from a particular LPDC machine and
die mold in the production of automotive wheels castings. The foundry process
is monitored at several locations such as the metal chemical analysis, LPDC ma-
chine, and X-ray.
The LPDC machine is setup and adjusted by the machine operators in accord-
ance to process control rules established over time. There are default settings
but also inherent flexibility to adjust these by the operator according to the rules
and given tolerances. Therefore, the actual values of process parameters in the
LPDC machine are time series monitored and available for statistical machine
learning analysis.
The LPDC machine time series monitors values during serial production in-
clude molding pressure, temperature, time and air and water channel cooling.
These values are continuously monitored at one and two second intervals to as-
sure they remain within the LPDC machine tolerances. The data itself is stored
as time series data with each machine’s identifier, material batch identifier, and
the monitored LPDC machine values. This forms the first database of LPDC
timeseries data.
The other set of variables important to porosity are the Aluminum alloy metal
batch properties. The data is stored with each batch sample’s time, material
batch identifier, and chemical properties. This forms the second database of
Aluminum alloy properties, time stamps, and identifiers.
The last set of measurements are the X-ray inspection results of each casting.
These results are stored as a dataset of the time of inspection, the part identifier,
the X-ray image and the pass/fail result.
These datasets themselves are insufficient for the machine learning study. Ad-
ditionally, part tracking data is needed to identify when the X-rayed part was in
the LPDC machine. To do this, a part tracking system that tracks when each rim
was in the LPDC machine and in the X-Ray is introduced in this study. This
database consisted of time as rows and part number at the LPDC machine and
a part number at the X-Ray machine. With this, LPDC machine process param-
eters and Aluminum metal properties with the X-ray results can be associated.
In summary, the result of these data collections are process time series data,
batch metal material properties, part level tracking location data, and porosity
quality results.

2.3.3 Data preprocessing

The initial dataset consisted of 1077 parts in total of which 62 of were porosity
failures. The process time series data is taken and partitioned according to the
rim tracking data. The result is approximately 300-time increments (seconds)
of processing time data points for each of the 1077 rims.
The preprocessed features that can be used to classify good and defective rims
formed by the time series data for each wheel rim. For example, features might
include the average and variance of a pressure over a phase of the casting cycle.

30
 Methods

Different phases of the pressure cycle are more important than others, for in-
stance phase 3 intensification pressure disturbances can affect porosity defects
more than other phases. A detailed figure about this disturbance is provided in
Publication 3.
The result of the feature extraction is 36 features from the 13 measurement
points. There are more features than the measurement points since a single time
series measurement might be extracted into multiple phases each with means,
standard deviations, and durations.
From this, a dataset for machine learning analysis was prepared. This dataset
consists of a 1077x37 matrix, where each row represents a production rim cast-
ing, and 36 columns of input variable feature data and 1 inspection pass/fail
column. The 36 input features are shown in Article 3, Table 1. Eleven of the ac-
ceptable castings had process settings far out of specification as outliers due to
maintenance and adjustment. These outliers were removed from the data. The
result was a dataset of 1066 parts in total.

2.3.4 Model evaluation

Since the machine learning hypothesis whether a wheel can be calssified as

pass or fail, this is approached as a binary classification problem. The dataset is
split into training and testing subsets. Typically, the split uses 70 -80% of the
data for training and the remaining for model validation testing. However, in
this case leaving 20% for validation leaves too few defectives. Therefore a 50%
split was used for a reasonable number of defectives in the validation dataset
which is not included in the model training and hyperparameter optimization.
The dataset was split in a random and stratified fashion to ensure an even allo-
cation of the defectives into the training and testing sets.
For fitting models of defectives, multiple evaluation metrics could be used.
These include accuracy, precision, recall and the f1. While it would seem, accu-
racy is a good measure, for unbalanced datasets with a high number of good
parts over defective parts, total accuracy can be misleading by simply being ac-
curate on the good parts alone and misclassifying many of the defective parts.
In this study, all these scores are utilized to ensure a quality model.

2.3.5 XGBoost binary classification implementation

In this section the set up and implementation of the defective classification

using XGBoost is presented. First, the repartitioning and resampling of the
training data is described to promote a higher concentration of the minority
class defective set. Then, the objective function for finding the best fit using hy-
peropt is described.
A resampling is made with over-sampling to increase and balance the minority
defective class in the training data, and with under-sampling to obtain a cleaner

31
space. This is done with Syntetic Over Sampling (SMOTE) [38] and the Edited
Nearest Neighbor Rule (ENN) under sampling [39]. The original data had a 16.1
to 1 ratio of good parts to defective parts, in minimizing the training error the
oversampled set resulted in a ratio of 0.8 to 1 ratio.
The hyperparameter tuning is a crucial task for optimizing performance of the
XGBoost algorithm used here. Here an informed automated hyperparameter
tuning approach is used, the Bayesian Optimization algorithm from the Hyper-
opt library [40] for model selection and hyperparameter optimization. The ob-
jective function searches over the selected hyperparameters to minimize the val-
idation error of training set. With informed Bayesian optimization, the next val-
ues selected are based on the past evaluation results. The evaluation metric cho-
sen is the area under the curve (AUC) which represents the ability to differenti-
ate between the positive and negative classes (good and defective parts). The
fitting algorithm XGBoost’s built in cross validation is utilized for creating vali-
dation sets to tune the model on the training data. For this study a 10-fold cross
validation was used in a stratified fashion, meaning the hyperparameter set is
validated and trained ten times and the mean of these results is the objective
function. Along with other built-in parameters for Xgboost classification algo-
rithm, the hyperparameter space for cross validation included logistic and hinge
loss functions, and regularisation terms withing a certain range to avoid over-
fitting. The testing data remains reserved for validating the machine learning
prediction model.

32
 Results and Discussion

3. Results and Discussion

3.1 An alternative DPM- AM2D code inserts

The objective of this section of the research was to determine the decodable
symbol that could be sand molded and cast in aluminum, as well as to estimate
the readability relation by size and contrast modules (dimple and bump). It was
clearly observed that bump DPMs have higher surface roughness and are less
readable than dimples (see Figure 17). At the smallest dot size of 1.5 mm, the
bumps were not readable, while the dimple version was completely readable.

Figure 17. The smallest size, 1.5 mm diameter bump (left) and dimple (right)
data matrix DPM.

Figure 18 shows a view of the Datamatrix and Dotcode DPMs after the casting,
mold breaking and cleaning process. These parts were studied at each of these
steps for casting quality considerations, including the separation of dots, miss-
ing dots, adjacent dot wall thickness and surface quality. These can all impact
readability.

33
Figure 18. Mold breaking and after cleaning.

Part quality is considered in terms of tag readability for purposes. Table 1 shows
the Pass/Fail results over all reading conditions, both equally for the industrial
reader and the mobile phone. There was no difference in the readability test re-
sults between the industrial reader and the mobile phone application.

Table 1. DPMs Pass/Fail test results with set up for two different lighting angles (30 ̊, 45 ̊)

Dot diameter 30̊ 45̊

(mm)
Datamatrix Dotcode Datamatrix Dotcode
Dimple Bump Dimple Bump Dimple Bump Dimple Bump
1.5 + - + - + - + -
2.0 + + + - + + + -
2.5 + + + - + - + +

There are 65 dots on the Dotcode and 83 dots on the Datamatrix code for the
same ten digits of numbers encoded as a part identification number
(1234567890). This represents how many dots are needed to encode the part
identification for representation in the production process. The Datamatrix
code is inherently more compact, since it makes use of adjacent dots, whereas
the Dotcode always has diagonal dot spacing, as can be seen from the Figure 19.
In the molding process, wider dot space results in improved DPM form as a re-
sult of better sand filling. Less space between dots also creates higher suscepti-
bility of the tag pattern in the sand of the mold due to possible impact damage
from the working environment. The Dotcode is more robust to the sand molding
process. On the other hand, the Datamatrix code occupies less space on the
metal part and has less 3D printed material consumption accordingly. As the

34
 Results and Discussion

number digits represented increases, the Datamatrix code begins to offer sub-
stantial size advantage.

Figure 19. The largest 2.5 mm dot sized data matrix (left) and Dotcode (right)
of dimple dpms.

Finally, the Datamatrix code has better error correction capability, where the
code remains readable with 17% error in the dots. In our experiment, when 14
of the dots were missing, the code was still readable. No such error correction is
provided with the DotCode standard. It was observed that 2.5 mm casting had
three dots missing and it remained readable, a 5% error rate.

3.1.1 Verification Across Multiple Identifiers, Parts and Operations

The previous experiments developed the process parameters and settings

needed to effectively cast direct part markings using sand casting, in terms of
dot code size, shape and location. To demonstrate this, direct part marking is
applied to a variety of parts.
On the same part, various unique identification codes were attempted. The
readability was tested at several stages of the casting process, from molding to
final part. A collection of variously shaped parts that were labeled in size and
shape based on the findings. Lastly, the readability of codes on various parts was
assessed.
Figure 20 depicts the experimental summary for the various parts and identi-
fications encoded. Even at shakeout, all of the cast parts were readable using a
common mobile phone barcode reading app. A variety of codes could be marked
on a variety of different parts to demonstrate the use of the tags in foundry op-
erations. It was discovered that all parts can be successfully marked with this
method with proper gating
Figure 20a depicts various unique identification codes cast on a flat surface
for multiple identification on the same part. Figure 20b shows a considerably
small engineering part that was successfully marked. Some parts can be more
difficult to mark due to their shape. This included the valve body shown in Fig-
ure 20c. It was found that tag burning can be deficient because of inadequate
feeding and venting from poor initial gating which causes non-readability. This
is dependent on the part shape and gating design to ensure the metal fills the

35
shape of the DPM and gases vent. An example casting of such failure is showed
in Publication 1 Figure 13.
Figure 20d shows the results after sand-mold shakeout on the metal casting.
These tags and tag imprints were all readable even at this early stage in produc-
tion before cleaning and finishing. Cleaning and finishing were done using
metal brushes, cutting the gatings from each part and polishing.

(a) 20 x 30 cm (b) 6cm Rotor (c) 30cm Valve (d) 15cm Pump
Flat Plate Body Housing

Figure 20. A variety of castings with a variety of tags. (a) 20 x 30 cm flat plate,
(b) 6 cm rotor, (c) 30 cm valve body, (d) 15 cm pump housing.

Overall, AM2D tags can form DPMs with sand casting considering the gating
system is adequate for the part. Care is needed to ensure there is sufficient metal
volume to burn the tag (proper gate height) and enough gas escape path (ade-
quate riser location). Also, the well-known conventional mold quality properties
such as sufficient sand binder content is important for the tag to remain in the
mold.

3.2 Foundry Data Collection and Part Tracking Using AM2D Tags

A simple procedure to use a mobile phone app to identify the part for rapid
and easy process data recording is introduced. As shown in Figure 21 , at each
operation the parts scanned into a database and entered process data for the
each part.

36
 Results and Discussion

Figure 21. Data Tracking using 2D matrix codes throughout the foundry.

The 2D datamatrix code was readable throughout the process either in early
stages (see Figure 22) as an AM2D tag and in the later stages as a DPM on the
cast rotor parts. A video snapshot of the cast rotor part tracking with unique
IDs is shown in Figure 23 along with the video link. All tags and DPMs on the
castings were readable using common mobile phone barcode reading applica-
tions.

Figure 22. Sand mold configuration of the parts with different unique datama-
trix id codes (left) and after shakeout (right).

37
Figure 23. Video of DPM captures available by scanning the Datamatrix code.
DPM capture using mobile phone app for part tracking (right).

When the data is collected and associated with the individual parts, several
uses become possible including monitoring and statistical analysis such as pro-
cess control and correlation. The most direct use of the collected data is to create
statistical process control (SPC). The SPC help indicate if the process is in steady
state and not subject to special cause changes.
Since the objective of this section is building a unique system to demonstrate
data collection and part tracking inside a Foundry. The details of the SPC charts
of the individual values and quality defect causal correlation are out of the scope
of this section but can be found in Publication 2.

3.3 Defective Part Classification and Process Optimization via

Supervised ML

The summary results of the model performance are shown using the confusion
matrix in Figure 24. 440 out of 505 good parts were correctly labeled as good,
whereas 65 out of 505 good parts were mislabeled as defective, and 87% recall
rate. Similarly, for the defectives; 23 out of 31 defective parts were correctly la-
beled as defective, and 8 out of 31 defective parts were mislabeled as good, a
74% recall rate.

38
 Results and Discussion

Figure 25. Confusion matrix of the XGBoost model for part quality.

The statistical summary report is shown in Table 2. Note the test data was
biased with many good parts (505 of 536, 94%) and only a small fraction of de-
fective parts (31 of 536, 6%) making classification difficult. The results show that
the classifier was able to correctly predict a good and defective part using the
process and material input data alone.
In the beginning of the model optimization and hyperparameter tuning, ob-
jective function was set to maximize the mean AUC score from the cross vali-
dated training set. There are other metrics tested such as area under the preci-
sion and recall curve and mean average precision to find out the best performing
model with given hyperparameters. Since the evaluation metric scores differs in
numeric base, therefore not comparable. In the end the metrics trained and
tested with validation set and these are not yielded better results that the the
metric remained as AUC.

Table 2. Statistical Summary Report

Precision Recall F1 Score Support

Good Parts 0.98 0.87 0.92 595
Defective Parts 0.26 0.74 0.39 31

The predictor equation consisted of a set of 77 separate binary decision tree

estimators with maximum depth of four layers. Each tree contributes an incre-
ment to the random variable of a logistic distribution function. One such deci-
sion tree is shown in Figure 26. Interpreting the tree, if the normalized standard
deviation value of Air Channel 2 is less than -0.502 the probability of failure is
around 0.37 which is the probability value of the loaded leaf value -0.5434. Sim-
ilarly, if the normalized standard deviation value of Air Channel 2 is more than
-0.502 and the normalized value of the Density Index smaller than 1.106 and

39
the Mean Value of Phase 3 intensification pressure is larger than -1.284 and the
Silicon Content of the metal is smaller than 1.021, then the probability of failure
increases to larger than 0.64, which is the probability of the loaded leaf value
0.561. In general, large leaf values in any tree indicate combinations of features
give rise to defects.

Figure 26. Example binary decision tree estimator.

3.3.1 Feature Importance

36 input features used in the classifier. the Shapley index of each feature is
computed to determine the relative importance of the features. To control the
feature importance scoring, a random noise variable was added and the analysis
with this added variable was rerun. Such a random noise variable will not con-
tribute to prediction accuracy. This resulted in a Shapley index of 0.3 indicating
all features below this value are also insignificant. The results are shown in Fig-
ure 8, with the features ordered according to contribution as shown by the Shap-
ley index. The result show that 15 of 36 features contribute, and the remaining
21 features combined contribute less than a half of the most contributing varia-
ble. Note these contributions do not sum to one since each Shapley index in-
cludes interaction effects.

40
 Results and Discussion

1,60

1,40

1,20

Shapley Index
1,00

0,80

0,60

0,40

0,20

0,00

Figure 27. Feature importance order.

Looking at the results shown in Figure 27, three out of the top four contribu-
tors are related with air cooling channel flow rates. The first contributing feature
is the standard deviation of the air flow in the channel closest to the center hub
on the bottom mold. The air channel standard deviation varies due to equip-
ment process control during operation of the machine; this analysis shows it
indeed has an impact on porosity defects. The second contributor was the Melt
Silicon Content of the molten metal. This corresponds with known casting phys-
ics, where the silicon content of the aluminum effects the fluidity which in turn
effects porosity formations [34]-[35]. The third highest contributor is again an
air channel flow standard deviation but of the top mold center hub. The Shapley
indices indicate that the upper and lower air channels at the center have highest
influence on porosity, this corresponds with understood physics since that is the
thickest area of the casting as shown in Figure 8.
The fourth contributor is the air channel flow of the top spoke near the center
hub. These air channels are located next to the center and are also at a relatively
thick location. The fifth highest contributor is the Density Index, which refers to
the oxide level in melt quality which is a direct effect on the porosity defects [41].
As a result of the feature importance scoring, the casting defects appears with
the combination of improper air channel operations and variations in material
properties. This work quantifies the combinations of the features needed to re-
sult in defective wheel rims. That is, a quantified prediction model has been cre-
ated to predict when porosity will likely to arise on this machine for this wheel.

41
3.3.2 Model trade-off between good and defectives

The XGboost model also includes hyperparameters to tune the trade-off be-
tween the rate of false positives and false negatives. It is possible to scale the
gradient of the minority class of false negatives. A tradeoff curve by varying the
model to generate a ratio of true positives to true negatives between 0 and 1 is
explored. As can be seen in Figure 28, predication of defectives can get above
90% accuracy; however, it would also suffer about 50% accuracy on the good
parts. If one were equally concerned on pass and fail accuracy, a reasonable
tradeoff value for this dataset would be 80% accuracy on both. For this study,
fail accuracy is of more concern.

Figure 28. Model tradeoff between pass and fail accuracy.

3.3.3 Model comparison

Other methods were explored including Support Vector Machine (SVM) and
logistic regression with performance results shown in Table 3. The logistic re-
gression provided poor accuracy on predicting good and defective parts. The
SVM provided sufficient accuracy on defective parts but very poor accuracy on
predicting good parts. Overall XGBoost did much better.

Table 3. Comparison Classification Algorithms.

XGBoost SVM Logistic Regression

F1 Score 0,39 0,13 0,2

Precision 0,26 0,07 0,12

Recall 0,74 0,94 0,55

42
 Conclusion

4. Conclusion

The present thesis provides tools for quality assurance through advanced man-
ufacturing techniques in foundries within three main aspects: proposing a
method to create permanent digital part markings for metal castings, demon-
strating foundry data collection and part tracking using this method, and data
management and supervised machine learning in foundry statistical quality
control.

4.1 Research question 1

What are limits in DPM size and process conditions for mark a cast part? What
are the experimental steps to improve DPM to cover the weakness of the cur-
rent systems and can be efficiently used in the foundry process?

A rapid marking system was demonstrated by making use of 3D printed tags

for use directly in the molding process. Overall, there are number of important
characteristics to consider in matrix code DPMs. The choice of coding standard
such as Datamatrix and Dotcode has an impact on the tag size and castability.
The results also indicate that dimple versions are more readable. Finally, high
contrast is needed for a readable tag, and dot sizes larger than the inherent
molding tolerance are necessary. In summary, 3D printing of matrix code tags
for sand casting provides a feasible means to provide permanent digital part
markings. This application enables a simple tracking operation via industrial
tablets or mobile phones utilizing free barcode reader application rather than
expensive dedicated industrial DPM readers. Such an application would assist
in resolving an important individual digital part tracking problem that remains
a block to foundry's industrial 4.0 transformation, as well as satisfy customers
by enabling digital code scanning of purchased parts.

Further research recommendations

Additional quality concerns include internal quality such as porosity that is

recommended as future work. It could be the case that burn out of paraffin wax
or PLA plastic tag material could affect the porosity, particularly for aluminum
castings. However, the combustion products would only affect the metal quality

43
if the tag size is not small compared to the casting. Further, proper gating and
risers can mitigate this.
It has also remained unclear how thin a part can be and yet be marked. There
must be sufficient molten metal to burn away the tag, and sufficient feeding
pressure to form the 2D matrix code dot pattern. As future research the forma-
bility limit on minimal part thickness for any casting can be formulated to pro-
vide a means to compute DPM tag formability limits before casting.

4.2 Research question 2

What are the process requirements to implement DPM collaboratively for data
collection and part tracking in serial production to improve understanding of
statistical quality control in the foundry products? What are the necessary fre-
quency and parameters to be monitored for statistical quality control?

The tag solution worked within a sand molding foundry operation. An operator
can uniquely mark, identify, and track the parts throughout the foundry opera-
tion. The AM2D tags are digitally readable and enable tracking in the beginning
of foundry operations including the mold making process showed in Figure 21.
This is a benefit over other postproduction DPM methods such as laser etching.
Associating the process data with individual cast part was made by scanning 2D
matrix codes and DPMs via mobile phone app which is considerably simple and
cheaper option against known applications.

Further research recommendations

The AM2D tags for digital part marking can easily be extended to investment
casting because of the huge similarities with sand casting. Feasibility study
within an investment casting foundry can be performed. The future work now
extends to considering how the embedded codes enable traceability in invest-
ment casting foundries with a cloud-based storage of material, wax patterns,
dewaxing machine, operators etc. This should consider using DPM codes for
workstation setups and consideration over how many setups to employ versus
often a scan is needed in a process flow. Also, future work ought to consider
industry standard coding schemes to allow for company agnostic coding per-
haps also propose a cheaper, faster, and more feasible manufacturing method
of the tags that can cope with for serial production speed in foundries.

4.3 Research question 3

What would be the ML model implementation to classify defectives and high-

light the important process variables that have the most in-fluence to discover
the complex process variables causing of defective castings?

44
 Conclusion

The causes of defective castings were identified. The Extreme Boosted Deci-
sion Tree (XGBoost) model did well in predicting good parts from defective
parts. Keeping the data fixed, it is possible to increase the defective part predic-
tion accuracy but at the expense of reduced good part prediction accuracy. This
enables the foundry managers to decide on which classifier has the higher ex-
pected benefit.
Beyond the cost of quality, identifying a defective part early reduces the carbon
footprint of the foundry by eliminating the unnecessary machining, painting,
and recycling. One of the main carbon footprint impacts is the high carbon emis-
sion of paint removal when recycling rims. Even with false positives and false
negatives, this model approach could possibly be used for predicting porosity
failures in the early stages by taking the model predicted and possibly defective
parts into quarantine for a deeper quality inspection. Dispositioning into such
quarantine parts would not automatically reject the false negatives but rather
quarantine them for deeper defect analysis.
This work did not make use of big data. Rather with a small dataset of a thou-
sand units the quality control of new production lines can be formed. While
large datasets can provide better and more robust models, practically a LPDC
foundry operation needs to make use of early smaller datasets. Foundry opera-
tions can use this machine learning methods to predict failures in short time
periods. For example, the process optimization can be started immediately to
decrease the porosity defect rims for an existing or a new wheel rim production
start. The model can be expanded to a more general and robust as more data is
gathered. Analysis of small datasets offers a means for foundries to begin to
make use of machine learning in their production.
Overall, Industry 4.0 data collection and machine learning worked well to
identify causes of casting defects within process data. This required a sophisti-
cated part tracking and data collection in the foundry as well as application of
state-of-the-art machine learning algorithm. While it is not sufficiently accurate
for dispositioning acceptable versus defective parts in production but is useful
for assisting in identifying root causes.

Further research recommendations

One critical decision point is how to capture features of the time series data.
In this study, averages and standard deviations are considered for each phase in
a cycle. Taking averages of time series data can mask patterns in the data. There
are more sophisticated methods to consider the full time series data. However,
in this study, the statistics are used from the process to compare the average
with machine set points for the operation and compare the standard deviations
with given machine and process tolerances. Future work would include compar-
ing this approach with use of the full time-series data and understanding what
patterns give rise to defectives.
This work studied porosity defects at all locations across the spoke of the rim
to determine a defective. The accuracy could be improved with more refined de-
tail of defects. The porosity failures could be separated by zone of a wheel to

45
enable linking the cooling channels by their locations. This could provide a bet-
ter understanding of the effects of the mean and standard deviation of the flow
rate of the channels as well as the operation times. However, it also requires
substantially more data, and therefore would be more suitable for analysis dur-
ing full production. To enable such high volumes of process and quality inspec-
tion data to be analyzed, automated data collection and tracking is needed. In
this study a semi-automated approach associating the X-Ray and LPDC ma-
chine was used. In future work, a totally automated parts tracking method
should be implemented starting from the earliest stages of production. The cur-
rent laser QR code marking system starts the marking after X-Ray in the middle
of the production and is inadequate. Further, the visual inspection system cur-
rently used at the X-ray could be automated to isolate the defects to types. More
data gathered on failed rims and the coordinates of the failures could be classi-
fied and a multiclass classification algorithm applied.

46
 Conclusion

References

1. Jung RP. The Digital Transformation - Chances and Challenges for the Foundry
Industry.23.09.2016; Dresden. The European Foundry Association.
2. Reuter, M. (2016). Digitalizing the Circular Economy. Metallurgical and Materials
Transactions B, 47(6), pp. 3194-3220.
3. Främling, K. (2013). Sustainable PLM through Intelligent Products. Engineering
Applications of Artificial Intelligence, 26(2), pp. 789-799. (2013).
4. T. Uyan, K. Jalava, J. Orkas, K. Otto, Cast part marking with 2D matrix codes us-
ing pre and post processing methods, in 2nd National Foundry Congress by
Tüdöksad Academy (2019), pp. 379–383
5. T. Uyan, (2018). Object identification in smart foundries https://aalto-
doc.aalto.fi/handle/123456789/33748
6. Remote Identification of Metal Castings. J.C. Altimus, V.D. Johnson, M.A. Luegge,
S.N. Ramrattan. 98-44, Kalamazoo, Michigan. : American Foundrymen's So-
ciety, Inc., 1998, May 10-13, Vol. 106
7. Autonomous Control of Intelligent Products in Beginning of Life Processes. Karl A.
Hribernik, Christoph Pille. s.l. : International Conference on Product Lifecycle
Management.
8. Laserax. Industrial Traceability Tracking Codes and How They Work. s.l. : La-
serax.
9. T. Winkel, D. Stein, Casting Inspection and Traceability for Total Quality Assur-
ance (Foundry and Management Technology, 2016)
10. J. Parkas, Improving casting Traceability in Iron Foundry by Laser Engraving of
Moulds, PhD thesis, Lappeenranta University of Technology, 2017
11. Casting Traceability with direct part marking using reconfigurable pin-type
toolin based on paraffin-graffin actuators. Nikolaj Kjelgaard Vedel-Smith*,
Torben Aker Lenau. Denmark : Journal of Manufacturing Systems, 2012.
12. Huskonen, W. “Need to Know: Cast-in-Place ID Marking System,” Foundry Man-
agement and Technology. 2006.
13. Saveraid, G. Cast data matrix symbols performance characterization, PhD thesis,
Iowa State University, 2010.
14. Li, X., He, W., Lei, L., Wang, J., Guo, G., Zhang, T. and T. Yue. “Laser direct
marking applied to rasterizing miniature DataMatrix Code,” Optics and Laser
Technology, Elsevier., 2015
15. Denkena, Berend, Thilo Grove, and Alexander Seibel. "Direct part marking by vi-
bration assisted face milling." Procedia Technology 26 26 (2016): 185-191.
16. B. Zhang, S.L. Cockcroft, D.M. Maijer, J.D. Zhu, A.B. Phillion, Casting defects in
low-pressure die-cast aluminum alloy wheels. JOM 57(11), 36–43 (2005).
https://doi.org/10.1007/s11837-005-0025-1
17. J.K. Kittur, G.C. ManjunathPatel, M.B. Parappagoudar, Modeling of pressure die
casting process: an artificial intelligence approach. Int. J. Metalcast. 10(1),
70–87 (2016). https://doi.org/10.1007/s40962-015-0001-7
18. J.K. Rai, A.M. Lajimi, P. Xirouchakis, An intelligent system for predicting HPDC
process variables in interactive environment. J. Mater. Process. Technol.

47
 203(1–3), 72–79 (2008). https://doi.org/10.1016/J.JMATPRO-
TEC.2007.10.011
19. A. Kopper, R. Karkare, R.C. Paffenroth, D. Apelian, Model selection and evalua-
tion for machine learning: deep learning in materials processing. Integr. Ma-
ter. Manuf. Innov. 9(3), 287–300 (2020)
20. A.E. Kopper, D. Apelian, Predicting quality of castings via supervised learning
method. Int. J. Metalcast. (2021). https://doi.org/10.1007/s40962-021-00606-
7
21. N. Sun, A. Kopper, R. Karkare, R.C. Paffenroth, D. Apelian, Machine learning
path way for harnessing knowledge and data in material processing. Int. J.
Metalcast. 15(2), 398–410
22. C. Reilly, J. Duan, L. Yao, D.M. Maijer, S.L. Cockcroft, Process modeling of low-
pressure die casting of aluminum alloy automotive wheels. JOM 65(9), 1111–
1121 (2013)
23. L. Zhang, R. Wang, An intelligent system for lowpressure die-cast process pa-
rameters optimization. Int. J. Adv. Manuf. Technol. 65(1–4), 517–524 (2013)
24. ISO/IEC. DIS 29158, Information technology—automatic identification and data
capture techniques—Direct Part Mark (DPM) quality guideline (2020)
25. Groenendyk, M., and R. Gallant. "3D printing and scanning at the Dalhousie Uni-
versity Libraries: a pilot project." Library Hi Tech (2013).
26. ASTM E155-15, Standard Reference Radiographs for Inspection of Aluminum and
Magnesium Castings, ASTM International, West Conshohocken, PA, 2015,
www.astm.org
27. Blondheim Jr, D., and S. Bhowmik. "Time-Series Analysis and Anomaly Detection
of High-Pressure Die Casting Shot Profiles." NADCA Die Casting Engineer
(2019): 14-18.
28. ISO 19232-1:2013, Non-destructive testing — Image quality of radiographs —
Part 1: Determination of the image quality value using wire-type image qual-
ity indicators
29. ISO 19232-5:2018, Non-destructive testing — Image quality of radiographs —
Part 5: Determination of the image unsharpness and basic spatial resolution
value using duplex wire-type image quality indicators.
30. B. Zhang, S. L. Cockcroft, D. M. Maijer, J. D. Zhu, and A. B. Phillion, “Casting de-
fects in low-pressure die-cast aluminum alloy wheels,” JOM, vol. 57, no. 11.
Minerals, Metals and Materials Society, pp. 36–43, 2005, doi: 10.1007/s11837-
005-0025-1.
31. Uludağ, M., Çetin, R., Gemi, L. et al. Change in Porosity of A356 by Holding Time
and Its Effect on Mechanical Properties. J. of Materi Eng and Perform 27,
5141–5151 (2018). https://doi.org/10.1007/s11665-018-3534-0
32. S. G. Lee, A. M. Gokhale, G. R. Patel, and M. Evans, “Effect of process parameters
on porosity distributions in high-pressure die-cast AM50 Mg-alloy,” Mater.
Sci. Eng. A, vol. 427, no. 1–2, pp. 99–111, Jul. 2006, doi:
10.1016/j.msea.2006.04.082
33. K. N. Obiekea, S. Y. Aku, and D. S. Yawas, “Effects of Pressure on the Mechanical
Properties and Microstructure of Die Cast Aluminum A380 Alloy,” J. Miner.
Mater. Charact. Eng., vol. 02, no. 03, pp. 248–258, Apr. 2014, doi:
10.4236/jmmce.2014.23029.
34. Di Sabatino, Marisa, and Lars Arnberg. "Castability of aluminium alloys."
Transactions of The Indian Institute of Metals 62, no. 4 (2009): 321-325.

48
 Conclusion

35. Dybowski, B., Poloczek, Ł. and Kiełbus, A., 2016. The porosity description in hy-
poeutectic Al-Si alloys. In Key Engineering Materials (Vol. 682, pp. 83-90).
Trans Tech Publications Ltd.
36. Kridli, G. T., P. A. Friedman, and J. M. Boileau. "Manufacturing processes for
light alloys." In Materials, design and manufacturing for lightweight vehicles,
pp. 267-320. Woodhead Publishing, 2021
37. H. S. Jang, H. J. Kang, J. Y. Park, Y. S. Choi, and S. Shin, “Effects of Casting Con-
ditions for Reduced Pressure Test on Melt Quality of Al-Si Alloy,” Met. 2020,
Vol. 10, Page 1422, vol. 10, no. 11, p. 1422, Oct. 2020, doi:
10.3390/MET10111422.
38. Lemaître, Guillaume, Fernando Nogueira, and Christos K. Aridas. "Imbalanced-
learn: A python toolbox to tackle the curse of imbalanced datasets in machine
learning." The Journal of Machine Learning Research 18, no. 1 (2017): 559-
563.
39. Wilson, D. L. Asymptotic Properties of Nearest Neighbor Rules Using Edited
Data. IEEE Transactions on Systems, Man, and Communications 2, 3 (1972),
408–421
40. Bergstra, J., Yamins, D., Cox, D. D. (2013) Making a Science of Model Search:
Hyperparameter Optimization in Hundreds of Dimensions for Vision Architec-
tures. To appear in Proc. of the 30th International Conference on Machine
Learning (ICML 2013).
41. D. Dispinar, J. Campbell, Porosity, hydrogen and bifilm content in Al alloy cast-
ings, Materials Science and Engineering: A, Volume 528, Issues 10–11, 2011,
Pages 3860-3865, ISSN 0921-5093, https://doi.org/10.1016/j.msea.2011.01.0

49
Mark and track individual casting, colle ct and associate the p roce ss data

Aalto-DT 206/2022
with the metal and let the machine learn the rest of the story.

BUSINESS +
ECONOMY
ISBN 978-952-64-1099-9 (pr int ed)
9 HSTFMG*ebajjj+

ART +
ISBN 978-952-64-1100-2 (pdf)
DESIGN +
ISSN 1799-4934 (pr int ed) ARCHITECTURE
ISSN 1799-4942 (pdf)
SCIENCE +
T E C H N O L O GY
Aalto University
C R O S S OV E R
School of Engineering
Department of Mechanical Engineering DOCTORAL
www.aalto.fi THESES