Sie sind auf Seite 1von 139

AI models and methods in

automotive manufacturing:
a systematic literature review

Master Thesis
International Industrial Management

Author:
Christoph Müller, BA MSc

Supervisor:
Vitaliy Mezhuyev, PhD, ScD

Kapfenberg, January 2022


Kurzfassung

Das Interesse an digitalen Technologien mit künstlicher Intelligenz (KI) hat in den letzten
Jahren rasant zugenommen. Während sich Anwendungen von KI in der Automobilindustrie
vorwiegend mit autonomen Fahrzeugen befassten, rückte die Optimierung von Fertigungs-
prozessen immer mehr in den Fokus der Forschung und Entwicklung. Diese Arbeit schließt
eine Lücke in der aktuellen Literatur und befasst sich mit KI im Kontext der Automobilin-
dustrie und der Industrie 4.0. Ziel war es, aktuell eingesetzte Modelle und Methoden der KI
und konkrete Praxisanwendungen in der Automobilproduktion zu erheben, sowie Problem-
stellungen und Zukunftsperspektiven aufzuzeigen. Um eine breite Abdeckung des aktuellen
Wissensstandes zum Thema zu gewährleisten, wurde die Arbeit in der Form einer wissen-
schaftlichen Übersichtsarbeit mit systematischem Ansatz zur Literaturrecherche umgesetzt.
Die Literaturrecherche beschränkte sich auf Fachliteratur ab dem Jahr 2015, wobei die Da-
tenbanken IEEE und ScienceDirect als primäre Quellen genutzt wurden. Zusätzliche Litera-
tur wurde über ResearchGate unter Berücksichtigung von qualitativen Quellen abgerufen.
Es wurden insgesamt 359 Arbeiten ermittelt, die in weiterer Folge auf ihre inhaltliche Eig-
nung geprüft wurden. Davon wurden 84 Beiträge für die quantitative Analyse und 79 Bei-
träge für die qualitative Analyse ausgewählt. Die Ergebnisse der Literaturanalyse bestätigten
das steigende Interesse am Thema: 2015 wurden lediglich 3 Arbeiten veröffentlicht, 2021
waren es bereits 33 Arbeiten. Die Beiträge befassten sich mit der Anwendung von KI in den
Bereichen Produktion (39,29 %), Qualitätssicherung (35,71 %), Montage (16,67 %), Supply
Chain (5,95 %) und Business Intelligence (2,38 %). Die Ergebnisse zeigten auch, dass Me-
thoden des maschinellen Lernens in der aktuellen Forschung und bei Anwendungen in der
Automobilproduktion dominieren, wobei neuronale Netze von den mehr als 70 genutzten
KI-Modellen am häufigsten eingesetzt wurden. In 34 Beiträgen (40,40 %) wurde ein Multi-
Methoden-Ansatz genutzt, um das beste KI-Modell zu ermitteln oder Kompromisse einzel-
ner KI-Methoden zu vermeiden. Der erfolgreiche industrielle Einsatz wurde durch viele Pra-
xisbeispiele bestätigt, bei denen eine Genauigkeit von über 90 Prozent für die Qualitätsprü-
fung, die Robotermontage, die Mensch-Roboter-Kollaboration, die Materialbedarfsermitt-
lung oder die KI-gestützte Entscheidungsfindung erreicht werden konnte. Die Probleme sol-
cher Anwendungen sind vor allem auf die Datenverfügbarkeit und -qualität, die Entwicklung
von KI-Modellen, Schwächen in der Simulation und der Systemintegration zurückzuführen.
Des Weiteren führen die Komplexität der Automobilprozesse, die physikalische Indust-
rieumgebung und der dynamische Wandel zu Herausforderungen. Für industrielle Anwen-
dungen wird daher empfohlen, die KI-Methoden weiter zu optimieren, um eine breitere Sys-
temintegration in CPPS zu ermöglichen und das volle Potenzial von Big Data sowie Edge-
und Cloud-Computing auszunutzen. Auf diese Weise kann KI mehr übergeordnete Aufga-
ben übernehmen, um eine KI-gestützte Produktionssteuerung ohne menschliches Eingreifen
zu ermöglichen. Dies führt zu einer wesentlichen Effizienzsteigerung, Kosteneinsparung
und Erfüllung der Kundenanforderungen, um langfristige Wettbewerbsvorteile zu erzielen.

I
Abstract

While AI has experienced an increasing interest in both research and industry during the past
decade, the true potential and industrial applicability of AI for automotive OEMs and sup-
pliers in real-world scenarios have not been clearly recognised. Most applications of AI fo-
cus on the development of autonomous cars, rather than the optimisation of automotive op-
erations and manufacturing processes. This thesis therefore bridged this gap and shed light
on the topic of AI in the context of automotive manufacturing and Industry 4.0. It aimed to
promote understanding and provide up-to-date insights on specific models and methods of
AI, applications that have been achieved with best practices as well as the problems that
were encountered, underpinned with possible future prospects. A systematic literature re-
view approach was adopted to ensure a broad and thorough coverage of current knowledge
and the identification of relevant literature on the topic. The literature search was confined
to papers that were published from 2015 onwards using the databases of IEEE and Sci-
enceDirect as primary sources, with a three-keyword search phrase to narrow down the re-
sults and increase specificity. Additional literature was manually retrieved using Re-
searchGate and referring to qualitative literature sources. A total of 359 papers were thus
identified that were subsequently screened for eligibility, of which 84 papers were selected
for quantitative analysis and 79 papers for qualitative synthesis. The results of the quantita-
tive analysis confirmed that the topic has gained significant interest, with a mere 3 papers
published in 2015 and 33 papers in 2021. The majority of papers dealt with solving problems
in production (39.29 %), quality (35.71 %) and assembly (16.67 %), whereas supply chain
(5.95 %) and business intelligence (2.38 %) were inadequately represented. The results of
the qualitative synthesis revealed that machine learning methods dominate current research
and automotive applications, with neural networks as the most used out of more than 70
identified models. A multi-method approach was described in 34 papers (40.40 %) to deter-
mine the best model or to avoid any trade-offs entailed by the use of single-methods. The
industrial applicability was confirmed by many use cases where an accuracy of more than
90 percent could be achieved for quality inspection, robot assembly, human-robot collabo-
ration, material demand prediction or AI-enabled manufacturing decision making. The prob-
lems of such applications were mainly attributed to data availability and quality, model de-
velopment and gaps in simulation, system integration, the complexity of automotive pro-
cesses, the physical conditions of the system environment and dynamic change. For indus-
trial applications it is recommended to further optimise AI methods and models, enabling a
wider system integration in CPPS by harvesting the potential of big data and both edge and
cloud computing. AI is thus enabled to accomplish more tasks in an effort to achieve AI-
enabled production control, where the system is capable of analysing, predicting and adapt-
ing of processes without human interference. This leads to increased efficiency, reduced
costs and high conformance with customer requirements, in order to achieve long-term com-
petitive edge.

II
Obligatory Signed Declaration

I hereby declare that the present master's thesis was composed by myself and that the work
contained herein is my own. I also confirm that I have only used the specified resources. All
formulations and concepts taken verbatim or in substance from printed or unprinted material
or from the Internet have been cited according to the rules of good scientific practice and
indicated by footnotes or other exact references to the original source. The present thesis has
not been submitted to another university for the award of an academic degree in this form.
This thesis has been submitted in printed and electronic form. I hereby confirm that the con-
tent of the digital version is the same as in the printed version. I understand that the provision
of incorrect information may have legal consequences.

Christoph Müller 19th January 2022, Kalsdorf bei Graz

III
Table of Contents
Kurzfassung .......................................................................................................................... I
Abstract ............................................................................................................................... II
Obligatory Signed Declaration ......................................................................................... III
List of Tables ...................................................................................................................... VI
List of Figures .................................................................................................................. VII
List of Abbreviations ...................................................................................................... VIII
1 Introduction .................................................................................................................. 1
2 Research Scope, Objective and Questions ................................................................. 2
3 Research Methodology ................................................................................................. 3
3.1 Systematic Literature Review ................................................................................. 3
3.1.1 Methodology Outline....................................................................................... 3
3.1.2 Search Strategy and Search Phrase .................................................................. 4
3.1.3 Literature Sources ............................................................................................ 6
3.1.4 Selection Criteria ............................................................................................. 7
4 Background and Theoretical Foundations ................................................................ 8
4.1 From Industry 1.0 to Industry 4.0 ........................................................................... 8
4.1.1 Key Historical Developments and Aspects ..................................................... 8
4.1.2 Politics in the Context of Industry 4.0 and AI ............................................... 18
4.1.3 Definitions of Industry 4.0 ............................................................................ 19
4.1.4 Aspects and Components of Industry 4.0 ...................................................... 22
4.1.5 Concept of Real-Time and Live Enterprises ................................................. 28
4.2 Principles of the Automotive Industry .................................................................. 30
4.2.1 Key Milestones in Automotive History ......................................................... 30
4.2.2 Characteristics of the Automotive Industry ................................................... 32
4.2.3 The Automotive Development Process ......................................................... 35
4.2.4 The Toyota Production System ..................................................................... 39
4.2.5 Automotive Megatrends and Implications .................................................... 46
4.2.6 A System Dynamics Perspective ................................................................... 47
4.3 The Paradigm of AI .............................................................................................. 50
4.3.1 History and Evolution of AI .......................................................................... 50

IV
4.3.2 Definitions of AI............................................................................................ 57
4.3.3 Classifications of AI ...................................................................................... 59
4.3.4 Contemporary Approaches Towards AI ........................................................ 61
4.3.5 General AI Methods and Intelligent Agents .................................................. 64
4.3.6 Machine Learning Methods ........................................................................... 65
4.3.7 Selection and Implementation of AI Models and Methods ........................... 67
4.3.8 AI Use Cases in the Automotive Industry ..................................................... 70
4.3.9 AI in the Context of Automotive Manufacturing .......................................... 71
5 Systematic Literature Review ................................................................................... 74
5.1 Literature Identification and Evaluation ............................................................... 74
5.2 Research Results ................................................................................................... 77
5.2.1 General Results .............................................................................................. 77
5.2.2 Current Models and Methods ........................................................................ 79
5.2.3 Contemporary Applications and Best Practices ............................................ 88
5.2.4 Issues and problems ....................................................................................... 94
5.2.5 Future Trends and Potential Innovations ....................................................... 98
6 Conclusion ................................................................................................................. 103
References........................................................................................................................ VIII

V
List of Tables

Table 1: Research methodology overview ............................................................................ 3


Table 2: SLR search phrase keywords .................................................................................. 5
Table 3: SLR selection criteria .............................................................................................. 7
Table 4: Live Enterprise operating model elements (Kavanaugh and Tarafdar, 2021) ...... 29
Table 5: Toyota Production System core methods (Cf. Schwickert et al., 2011) ............... 40
Table 6: Definitions of AI (Russel and Norvig, 2016) ........................................................ 57
Table 7: 4-stage AI classification (Cf. Hassani et al., 2020) ............................................... 59
Table 8: AI Approach Overview by Stanford University AI Panel (Cf. Stone et al. 2016) 63
Table 9: General AI methods (Cf. Russel and Norvig, 2021) ............................................. 64
Table 10: Machine learning methods (Cf. Russel and Norvig, 2021) ................................. 65
Table 11: Automotive AI use cases (Cf. Kozlowski and Wisniewski, 2021) ..................... 70
Table 12: Levels of AI-enabled manufacturing (Plattform Industrie 4.0, 2019) ................. 73
Table 13: Literature distribution per year ............................................................................ 75
Table 14: Types of literature ............................................................................................... 76
Table 15: AI-enabled processes in automotive manufacturing ........................................... 77
Table 16: Distribution of AI-enabled processes in automotive manufacturing................... 77
Table 17: AI methods in automotive manufacturing ........................................................... 79
Table 18: Distribution of single-/multi-method approach ................................................... 80
Table 19: AI models and algorithms used in automotive manufacturing............................ 81
Table 20: Current models and methods of AI in automotive manufacturing ...................... 84
Table 21: AI use cases per process ...................................................................................... 88
Table 22: Contemporary applications and best practices of AI........................................... 89
Table 23: Key issues and problems of AI............................................................................ 95
Table 24: Issues and problems of AI in automotive manufacturing ................................... 95
Table 25: Key future trends and potential innovations of AI .............................................. 99
Table 26: Future trends and potential innovations of AI ................................................... 100

VI
List of Figures

Figure 1: Industry 3.0 and 4.0 architecture (Frere et al., 2018) ........................................... 15
Figure 2: AI-enabled smart factory architecture (Wan et al., 2021).................................... 17
Figure 3: Schumpeter's Innovation Cycles (Jovanovic et al., 2018) ................................... 17
Figure 4: BCG Industry 4.0 technological aspects (Rüßmann et al., 2019) ........................ 21
Figure 5: Industry 4.0 architecture (Liu et al., 2019) .......................................................... 26
Figure 6: Development from Industry 1.0 to Industry 4.0 (Frere et al., 2018) .................... 28
Figure 7: Historic development of Toyota (Toyota, 2021) ................................................. 31
Figure 8: Global automotive car makers by revenue in $ (Statista, 2021) .......................... 32
Figure 9: Global automotive sales figures (Statista, 2021) ................................................. 33
Figure 10: Automotive OEM positioning strategies (Diez and Rodriguez, 2010) .............. 33
Figure 11: Shift of the automotive supply chain (Graenzle et al., 2016) ............................ 35
Figure 12: Automotive core processes (Brunner et al., 2017) ............................................. 35
Figure 13: Automotive manufacturing process (Toyota, 2021) .......................................... 37
Figure 14: Toyota‘s production process milestones (Toyota, 2021) ................................... 43
Figure 15: Toyota's autonomation process (Toyota, 2021) ................................................. 44
Figure 16: Automotive OEM innovation ranking 2021 (Bratzel, 2021) ............................. 45
Figure 17: Automotive industry trends and implications (Kuhnert et al., 2018)................. 46
Figure 18: Ecosystem cause-loop diagram (Khakifirooz et al., 2018) ................................ 48
Figure 19: Development history of AI (Xue-bo et al., 2018) .............................................. 54
Figure 20: AI history and technological focus (Copeland, 2016) ....................................... 55
Figure 21: Exponential growth of computing (Tekinerdogan, 2017) .................................. 56
Figure 22: 3-stage AI classification (Jylkäs, 2020) ............................................................. 60
Figure 23: High-level flow of supervised learning (adapted from Awad et al., 2015) ....... 66
Figure 24: High-level flow of reinforcement learning (adapted from Awad et al., 2015) .. 66
Figure 25: AI model map for learning, discovering and reasoning (Contreras et al., 2018)69
Figure 26: AI model map for enabling optical networks (Mata et al., 2018) ...................... 69
Figure 27: SLR PRISMA flowchart .................................................................................... 74
Figure 28: Literature distribution after screening phase ...................................................... 75
Figure 29: Literature distribution before screening phase ................................................... 75
Figure 30: Literature distribution per year .......................................................................... 75
Figure 31: Types of literature .............................................................................................. 76
Figure 32: Literature rating distribution per year ................................................................ 76
Figure 33: Distribution of AI-enabled processes in automotive manufacturing ................. 78
Figure 34: Number of annual publications and distribution of the processes ..................... 78
Figure 35: AI methods in automotive manufacturing ......................................................... 79
Figure 36: AI method application per process .................................................................... 80
Figure 37: AI model and algorithm application per process ............................................... 83
Figure 38: Key issues and problems of AI .......................................................................... 94
Figure 39: Key future trends and potential innovations of AI............................................. 98

VII
List of Abbreviations
Abbreviation Description
5G Fifth Generation Mobile Network Technology
AGI Artificial General Intelligence
AGV Automated Guided Vehicle
AI Artificial Intelligence
AI HLEG AI High-Level Expert Group of the EU
ANI Artificial Narrow Intelligence
AR Augmented Reality
ASI Artificial Super Intelligence
AUROC Area Under the Receiver Operating Characteristic Curve metric
CAD Computer Aided Design
CNN Convolutional Neural Network
CPS Cyber-physical System
CPPS Cyber-physical Production System
CPU Central Processing Unit
DL Deep Learning
DSRPAI Dartmouth Summer Research Project on AI
DT Decision Tree
EC Exclusion Criteria
ENIAC Electronic Numerical Integrator and Computer
ERP Enterprise Resource Planning
FPGA Field Programmable Gate Array
GPU Graphic Processing Unit
HPC High Performance Computing
IC Inclusion Criteria
ICT Information and Communications Technology
IEEE Institute of Electrical and Electronics Engineers
IoT Internet of Things
IP Internet Protocol
IPU Intelligence Processing Unit
IRL Inverse Reinforcement Learning
IT Information Technology
KNN K-Nearest Neighbors algorithm
LAN Local Area Network
LBP Local Binary Patterns descriptor
M2M Machine to Machine
MCC Matthew's Correlation Coefficient metric
ML Machine Learning
MLOps Machine Learning Operations
MENACE Machine Educable Noughts And Crosses Engine
MES Manufacturing Execution System
MQTT Message Queuing Telemetry Transport

VIII
NFC Near Field Communication
NLP Natural Language Processing
OEM Original Equipment Manufacturer (car maker)
OPC UA Open Platform Communications – Unified Architecture
PLC Programmable Logic Controller
PoC Proof of Concept
PRISMA Preferred Reporting Items for Systematic Reviews and Meta-Analyses
RFID Radio Frequency Identification
R&D Research and Development
SCADA Supervisory Control and Data Acquisition
SLR Systematic Literature Review
SME Small and Medium-sized Enterprise
SVDD Support Vector Data Description
SVM Support Vector Machine
TCISS Toyota Creative Idea and Suggestion System
TQM Total Quality Management
TPM Total Productive Maintenance
TPS Toyota Production System
WIFI Wireless Fidelity – Standard for wireless data transmission
QC Quality Criteria
QR Code Quick Response Code
VR Virtual Reality
WWW World Wide Web

IX
1 Introduction

The first industrial revolution that commenced in the 18th century ignited the spark that en-
tailed unforeseen technological advancements and global change. The invention of the steam
engine together with electricity set the cornerstone of the industrialisation of many indus-
tries, with the automotive industry positioned among the first movers in adopting new tech-
nologies. Henry Ford first introduced a conveyor-belt assembly line producing his standard-
ised Model T, whereas Toyota introduced its renowned production system that became in-
dustry best practice. Decades later, the development of information and communication tech-
nologies, with sophisticated hardware and software, marked another milestone in industrial
history that could only be surpassed by the invention of the internet. These and many other
technological breakthroughs led to Industry 4.0 that is characterised by the internet of things,
machine to machine communication and ultimately cyber-physical production systems.
Reaping the benefits of today’s digital transformation will play a key role in further enhanc-
ing manufacturing processes and gaining a competitive advantage. The implementation of
AI might be considered the appropriate strategic move to ensure sustainable growth. Even
though the automotive industry has long paved the way for innovative manufacturing tech-
nologies, it is questionable to what extent AI has already been deployed in manufacturing.
This thesis thus aimed to outline the extent of AI adoption in automotive manufacturing in
the context of Industry 4.0 by means of a systematic literature review. Four research ques-
tions were defined to guide the research process, focussing on (1) the use of AI models and
methods, (2) applications and best practices as well as (3) limitations and (4) possible future
prospects. Given that previous research and industrial applications have been largely con-
fined to the development of connected and autonomous vehicles, there is a significant gap
in research with respect to automotive operations and manufacturing. The results of this re-
search are therefore considered to provide valuable insights and gain relevant knowledge
that can guide future research and applications in the field of AI. The limitations of the results
are mainly attributed to the research methodology adopted and the use of a predefined search
phrase, which narrows the literature and does not allow making generalisations from the
quantitative analysis. The qualitative synthesis and interpretation of results are in part subject
to research bias and can thus provide input for deeper analysis and further research.

This thesis is divided into six chapters: Chapter one has introduced the research context and
approach, chapter two highlights the research scope and questions, chapter three describes
the principles of the adopted research methodology, chapter four discusses the background
and theoretical foundations of the topic area focussing on the industrial revolution, the auto-
motive industry and AI, chapter five presents and discusses the research findings. Finally,
chapter six summarises the key aspects, findings and implications of this thesis.

1
2 Research Scope, Objective and Questions

The research scope of this thesis encompasses technologies of AI that are implemented and
used in the automotive industry to ensure manufacturing excellence. In this respect, Industry
4.0 and the corresponding components and approaches are considered state-of-the-art in au-
tomotive manufacturing and are therefore dealt with in more detail in chapter four. The major
objective of this thesis was to outline the contemporary state and extent of AI adoption in
automotive manufacturing, drawing attention to specific AI approaches and use cases, in-
herent limitations and prospects for the future. This objective was converted into four distinct
research questions that guided the research process and the application of the systematic
literature review approach.

The topic of the thesis outlines the methodological approach and the scope in terms of
technology and industry focus:

“AI models and methods in automotive manufacturing: a systematic literature review”

The following research questions focusing AI models and methods in automotive manufac-
turing have been defined:

(1) What are the current models and methods of AI in automotive manufacturing?

(2) What are the contemporary applications and best practices of AI in automotive man-
ufacturing?

(3) What are the issues and problems that limit the potential of AI applications in auto-
motive manufacturing?

(4) What are the future trends and potential innovations of AI in automotive manufac-
turing?

2
3 Research Methodology

3.1 Systematic Literature Review

A systematic literature review (SLR) approach was adopted to answer the predefined re-
search questions and provide valuable insights into the adoption of AI in automotive manu-
facturing and the respective contemporary developments as well as future prospects. This
approach ensured a high degree of focus, integration and transparency. It was based on a
research plan and search strategy that state both exclusion and inclusion criteria to select
potential literature and provide a holistic overview and thorough discussion.1 For this pur-
pose, specific search terms, databases and platforms were defined as part of the methodo-
logical planning process. Only literature from 2015 onwards was considered to deliver up-
to-date information, in order to cater for the fast pace of digital development and transfor-
mation. Table 1 provides an overview of the research approach.

Table 1: Research methodology overview

Research methodology: Systematic literature review


Methodology type: Secondary data research
Primary literature sources: IEEE, ScienceDirect
Search phrase focus: AI, automotive, manufacturing
Selection criteria: Predefined minimum and specific criteria
Research question focus: AI models and methods, applications and best practices, issues
and problems, future trends and potential innovations
Industry focus: Automotive Industry
Functional area focus: Manufacturing and related processes

3.1.1 Methodology Outline

The systematic literature review has increased in importance as a research methodology dur-
ing the late 20th and early 21st century. In effect, it has been used as a key method in medical
research, given the need to derive at appropriate diagnostic and therapeutic methods based
on a significant number of medical papers and studies. While the approach is commonly
known as “systematic review”, it is also synonymously known as “meta-analysis”, whereas
the latter is more focused on statistical analyses of data.2 The SLR methodology was con-
sidered highly appropriate to be used in the context of this thesis and the topic of AI due to
the wide variety of models and methods that are available to solve manufacturing problems.
Higgins and Green (2011) argue that the major difference between a traditional literature
review and a systematic review is the objective and standardised approach. The basis of

1
Cf. Saunders et al. 2009, p. 82f.
2
Cf. Purssell et al. 2020, p. 10f.

3
every SLR therefore are predefined research questions that outline the focus as well as the
expected results of the research project. The questions should be properly formulated and
cover all aspects and goals that are to be achieved. The adoption of the principles of a sys-
tematic literature review provides a distinctive set of characteristics for the respective re-
search project, as cited by Higgins and Green (2011):3

• A clearly stated set of objectives with pre-defined eligibility criteria for studies.

• An explicit, reproducible methodology.

• A systematic search that attempts to identify all studies that would meet the eligi-
bility criteria.

• An assessment of the validity of the findings of the included studies, for example
through the assessment of risk of bias.

• A systematic presentation, and synthesis, of the characteristics and findings of the


included studies.

Purssell and McCrae (2020) summarise the characteristics of a systematic literature review
as follows: “A systematic review is an original report of research evidence with interpreta-
tion and recommendations, amounting to a whole greater than the sum of parts.”4 The
knowledge gained through a systematic literature review can thus highlight significant gaps
in previous research and shed light on areas that are yet to be given more attention in future
research activities.

Given that research literature and studies in the field of AI and the automotive industry have
gained significance during the past few years, the adoption of the SLR approach was highly
suitable to answer the research questions of the present thesis and derive meaningful findings
and recommendations. In addition to the SLR, desk research was conducted to provide back-
ground information and theoretical foundations on the topic to support the SLR results.

3.1.2 Search Strategy and Search Phrase

The SLR was carried out with a predefined search strategy. All the content to be considered
in an SLR should be retrieved from well-established and highly regarded literature sources
that meet the needs of common research standards. The use of credible database sources,
such as IEEE or ScienceDirect, ensures a high degree of literature coverage.

3
Higgins and Green 2011
4
Purssell and McCrae 2020, p. 11

4
Due to the high number of sources and studies that deal with AI in the automotive industry,
in particular focussing on autonomous driving, it was critical to define a clear search phrase
in order to find relevant and qualitative papers. The search phrase that was initially used to
retrieve papers and studies was confined to a combination of three main key phrases: (1) AI,
(2) automotive industry and (3) manufacturing. First search tests with these keywords on
different databases did not deliver the desired output. Since the titles used for any research
literature available might use synonyms, such as production instead of manufacturing, dif-
ferent variants of the primary search phrase were used to gain more search results by using
the search operators “AND” as well as “OR”. The use of three variables in the search phrase
and the focus on both title and abstract reduced the total number of literature to be retrieved,
yet it provided a clear and focused overview on the current research state in the topic area
and potential opportunities for future research at the same time. Furthermore, due to the
limited number of papers, it was possible to screen all that could be identified.

The desk research carried out revealed that machine learning is a core method of AI and
widely used in the automotive industry, it was therefore added as a separate keyword in
addition to AI. In this respect, it is worth noting that an initial search using the search phrase
below omitting the term machine learning resulted in 238 papers through the use of the Sci-
enceDirect database. By adding the term machine learning, additional 39 papers were found.
The papers that were identified did not always refer to all relevant keywords to be found,
such as AI, machine learning and supervised learning. This was also confirmed by the results
of manual searches. However, despite the use of a combined search phrase, there was still
an inherent lack of full literature coverage. A search phrase with more keyword combina-
tions for each category would have led to more results and more reliable generalisations to
be made, yet this approach was out of the scope of the present thesis due to limited resources.

Table 2 shows the keywords that were used with operators OR/AND to conduct the final
literature search of this thesis:

Table 2: SLR search phrase keywords

Keyword 1 Keyword 2 Keyword 3


Artificial intelligence Automotive Manufacturing
Machine learning Production

The following search phrase was used combining the above-mentioned keywords and oper-
ators:
('artificial intelligence' OR 'machine learning') AND automotive AND
('production' OR 'manufacturing')

5
3.1.3 Literature Sources

When adopting the SLR methodology, the use of just a single literature database would limit
the potential of meaningful literature coverage, due to the broad application of AI in any
given industry. Hence, the literature search was carried out by using different databases with
the predefined multi-keyword search phrase. Any database considered for the literature
search must meet the standards of contemporary research practice and include widely ac-
cepted materials. The research databases of IEEE Xplore and ScienceDirect were used as
primary literature sources, given that they are considered highly credible and reliable. Re-
searchGate was used for an additional manual literature search, since it also covers a broad
spectrum of articles, papers and studies. In order to ensure a high degree of literature quality,
any content found through ResearchGate to be used for the SLR was thoroughly evaluated
and the root sources were retrieved.

3.1.3.1 IEEE Xplore


The Institute of Electrical and Electronics Engineers (IEEE) describes itself as “the world’s
largest technical professional organization for the advancement of technology”.5 It hosts the
Xplore Digital Library, which is a highly regarded database that provides access to research
materials and content published by the IEEE. It ranks on top of the most important databases
for computer science related topics. As of 2021, the database contains more than 5 million
records from journals, magazines, ebooks, conference proceedings and courses.6

3.1.3.2 ScienceDirect
ScienceDirect is an online database provided by the Dutch publisher Elsevier that comprises
science-related and peer-reviewed literature in four main sections, including physical sci-
ences and engineering. The database counts more than 18 million materials ranging from
academic journals and e-books.7

3.1.3.3 ResearchGate
ResearchGate is a commercial social network that connects more than 20 million researchers
around the globe. It also provides a database for research papers, studies and articles.8 How-
ever, it is worth noting that it does not reach the high standards and credibility of IEEE
Xplore and ScienceDirect. ResearchGate was therefore used to manually search for relevant
research papers that meet the selection criteria.

5
IEEE 2021
6
Cf. IEEE 2021
7
Cf. ELSEVIER 2021
8
Cf. ResearchGate 2021

6
3.1.4 Selection Criteria
The literature retrieved using the above-mentioned research databases and web services is
scrutinised based on predefined selection criteria. This ensures a high level of focus and
cohesion to properly answer the research questions. The selection criteria are divided into
three parts: (1) the exclusion criteria (EC) are used to identify suitable literature and exclude
non-relevant literature, (2) the inclusion criteria (IC) are used to screen the literature accord-
ing to the focus of the content to be included in the research, (3) the quality criteria (QC) to
provide a measure of paper significance. For the quality criteria, a distinction was made
between highly-rated conference papers and articles as well as other types of literature, in-
cluding chapters. The papers, studies and research materials were documented in a dedicated
master data file using Microsoft Excel. The implemented filter function allows flexible
search and identification of literature according to the selection criteria rating. The selection
process using predetermined criteria ensures that the papers selected for quantitative analysis
as well as qualitative synthesis contain the relevant content breadth and depth to adequately
answer the research questions. The minimum rating threshold for papers to be selected for
qualitative synthesis was determined at a score of 15. The predefined criteria are listed in
table 3.
Table 3: SLR selection criteria

No. Scope Criteria Score


Yes (2)
1 EC Title, abstract or full text contain the defined keywords Partial (1)
No (0)
Yes (2)
2 EC Content presented is relevant to the defined research topic Partial (1)
No (0)
Yes (2)
3 EC Date of publication since 2015 Partial (1)
No (0)
Yes (2)
4 EC Availability of full-text Partial (1)
No (0)
Yes (2)
5 EC Published in English language Partial (1)
No (0)
Yes (2)
6 IC AI models, methods and technologies are described Partial (1)
No (0)
Yes (2)
7 IC Applications of AI, best practices or use cases are described Partial (1)
No (0)
Yes (2)
8 IC Negative results, issues or limitations are discussed Partial (1)
No (0)
Yes (2)
9 IC Future prospects of AI or recommendations are discussed Partial (1)
No (0)
>10 (2)
10 QC Number of citations <9 (1)
None (0)
Conf.
Type of paper paper (2)
11 QC Article (1)
Conference paper (2), article (1), other (0)
Other (0)

7
4 Background and Theoretical Foundations

The topic of this thesis taps into three knowledge areas that have had a supplementing and
reinforcing effect on each other. Given the distinct characteristics and the noticeable devel-
opment of each area, the background and theoretical foundations are divided into three parts
in order to provide a holistic overview and promote understanding with in-depth information
and knowledge: (1) industrial revolution, (2) automotive industry and (3) AI. This chapter
highlights the technological development of advanced mechanics and steam engines in the
18th century, the emergence of early computational devices and robots, up to AI in manufac-
turing and smart factories as part of the fourth industrial revolution. It will highlight the key
milestones and characteristics of the automotive industry and the major approaches of auto-
motive supply chains and car manufacturing. Significant attention will be devoted to the
paradigm and principles of AI. This approach sheds light on the underlying causes and cor-
relations of the technological advancements that led to contemporary AI-enabled automotive
manufacturing and the prerequisites for successful technology adoption. A system dynamics
use case outlines all three theoretical aspects and significant implications.

4.1 From Industry 1.0 to Industry 4.0

4.1.1 Key Historical Developments and Aspects


It took 13.8 billion years for human beings to arise in cosmic history, while the emergence
of the homo sapiens dates back more than 300,000 years.9 The technological advancements
mankind takes for granted nowadays, including automobiles, robots and AI, developed
within a rather short period considering that the advent of the first industrial revolution dates
back to the 18th century. In fact, it took 300 years for traditional craftsmanship and the mere
focus on manual labour to reach the state of automated and AI-enabled smart factories and
autonomous vehicles that can be further optimised by enhanced technologies, such as ma-
chine learning algorithms.

The first industrial revolution of the 18th century was not triggered on purpose. As such, it
was caused by a number of factors that had a reinforcing effect on the development of soci-
ety, economy and technology at the same time. One aspect built on the other and formed a
new way of living and working as well as the way education and science found their way
through the layers of the population. According to Phyllis Deane, an often-cited professor of
Cambridge University in economic history, the major influencing factors to drive the indus-
trial revolution included (1) the optimisation of former production systems by means of sci-
ence and empirical knowledge, (2) the tendency towards production specialisation and mar-
ket development, (3) the increase in urbanisation, (4) the move from small family businesses

9
Cf. Siegel 2019

8
to bigger enterprises, (5) the enhancement of labour skills from raw material production to
final goods and services, (6) the availability and investment of financial resources, (7) the
process of societal adaption and formation of new social classes, such as workers. Given that
the above-mentioned aspects and others occurred simultaneously, they paved the way for the
first industrial revolution to unfold.10 In effect, the first signs of what is called Industry 1.0
today could be found in Great Britain in the 18th century. The reasons why Great Britain
outpaced other countries at that time were manifold and yet subtle. Adding to Deane’s notion
of industrial revolution influencing factors, Robert Allen, professor at Oxford University,
argues that the most critical influencing factor was economic by nature. The starting point
was the expansion of the European market and trade beyond national borders commencing
in the 16th century. In particular, Great Britain dominated the textile industry and was capable
of outpacing rival countries such as Italy. The focus on the European market was soon ex-
panded to go international and establish trade relationships with bigger countries including
the Americas and India. This significantly enhanced Great Britain’s strength and knowledge
base on a global level, as did the acquisition of colonies to tap into foreign resources. The
ongoing industrial leadership led to an increase in Great Britain’s wages and a reduction in
costs for energy, which in turn caused a demand for new technologies and inventions. In
fact, the advancements made in the cotton industry and mechanised production were mainly
owing to the development of highly productive spinning machines that enabled workshops
to increase productivity and output.11

Thomas Newcomen, an English inventor of the first functional steam machine, unintention-
ally set the cornerstone of modern industrialisation and bridged the gap between intense
manual labour work and the need for a high level of productivity at a lower cost base. In
effect, his efforts to invent the steam machine were driven by the demands of mining rather
than industrial production. As coal mines grew in size and depth, traditional water wheels
pushed by animals lacked the power necessary to pump up rain water. Hence, it was New-
comen’s invention that was the solution to the miner’s underlying problem.12 As a result, the
steam machine was further optimised and used for textile production and other industries to
significantly boost production. Another noteworthy inventor of steam machine technology
was James Watt, who applied for a patent in 1769 that would further increase the adoption
of steam machines from production to transportation, including the steam locomotive, which
further enhanced national and international trade and travel.13 The success of the steam ma-
chine and its adoption was confined to Great Britain at first, since the input resources were
less expensive compared with other countries. When it comes to the raw materials needed
for steam engines, such as coal, the process of iron making was significantly enhanced in

10
Cf. Deane 1979, p. 1ff.
11
Cf. Allen 2009
12
Cf. Lira 2013
13
Cf. Allen 2009

9
Great Britain to be able to supply cheap coal and significantly reduce costs associated with
machine operations. As Robert Allen sums it up: “The Industrial Revolution was invented
in Britain in the eighteenth century because that was where it paid to invent it.”14 While the
technology was meant to revolutionise the whole world, it remained exclusive to Great Brit-
ain at the beginning of its dominance, but was soon adopted worldwide as circumstances
changed.15

The first industrial revolution prevailed until the middle of the 19th century and was brought
to a halt due to a lack of inventions. It exemplified that necessity begets ingenuity, taking
into account the effect of Thomas Newcomen’s invention of the first steam machine. Similar
to the first industrial revolution, the second industrial revolution was again triggered by a
number of influencing factors that had a reinforcing effect on each other. As such, Great
Britain was no longer the leader in industrial manufacturing and technologies, given that
countries including Germany, USA, France or Japan were keen to follow suit. As a matter
of fact, the technologies that sparked the second industrial revolution were not invented and
were driven solely in Great Britain. For example, the United States made a huge contribution
to the expansion of rail networks, which further enhanced trade and travel. The transportation
of raw materials, supplies and goods was enabled at lower costs and farther distances at a
reduced time span. The bigger the rail network grew, the bigger the effect on the develop-
ment of technologies and whole industries.16

In effect, operations at farther distances, in turn, required a more efficient and effective way
of communication. Messages that were delivered via postal services were time-consuming
and insecure. The so-called electrical telegraph invented by US-born Samuel Morse and oth-
ers marked a significant cornerstone of long-distance communication. By using a wired con-
nection between two ground stations, individual electrical signals could be transmitted. It
was Samuel Morse who described the Morse Code as a simple means to convert messages
into a code that could be transmitted via telegraphs.17 Nevertheless, the key driver of the
second industrial revolution was not steam-powered transport or telegraph-enabled commu-
nication. The discovery of electricity and its potential to power industrial machines and
equipment pushed production to the next level. Per se, electricity was discovered by Benja-
min Franklin as early as 1752, when the first industrial revolution commenced to gain mo-
mentum and spread from Great Britain to the world. With his kite-experiment, Franklin was
able to pick up an electric charge from a flash of thunderstorm lightning. As a result, it took
eighty years for electricity to be made usable for industrial operations. In 1831 the English
scientist Michael Faraday discovered the electric current that occurred due to magnet

14
Allen 2009
15
Cf. Deane 1979, p. 1ff.
16
Cf. Engelmann 2021
17
Cf. HISTORY 2019

10
movement next to a metal wire. Thus, owing to this experimental approach, the principle of
induction was understood and further translated into what is known as an electric generator.
The concept of the electric generator was further optimised to be used in the industry by
smart and eager scientists and inventors and caused the second industrial revolution to reach
its full potential in the late 19th century. Apart from technological breakthroughs since the
invention of the steam machine, telegraphs or electric generators, the approach towards pro-
duction management and operations was influenced in parallel by great management think-
ers such as Adam Smith. In particular, he was the first proponent of what is known as the
theory of division of labour. His main work on the wealth of nations, published in 1776,
marked the early approach towards management thinking. In particular, he made a signifi-
cant contribution to the optimisation of production efficiency and process optimisation. By
dividing the manufacturing process of a product in different tasks and allocating these dis-
tinct tasks to a number of workers who just repeat them over and over again, the throughput
time could drastically be reduced. Traditional craftsmanship, on the other hand, was focused
on well-trained manual work and the manufacturing of the whole product by a single worker.
Adam Smith, with his notion of the division of labour, was a key driver of the industrial
revolution and contributed management knowledge and optimisation insights. It can also be
seen as the first move towards a holistic view of production systems, making them smart by
design. In effect, in the past, it was solely human capacity and intelligence as well as trial
and error, that enabled the technological and operational advancement of the first and second
industrial revolutions. This is said to change by adopting AI to industrialisation beyond to-
day’s notion of Industry 4.0.18

With both advancements in technologies and production optimisation methods, the second
industrial revolution set off the shift from small family-run workshops to fully industrialised
factories focussing on mass production of products, rather than products tailor-made to cus-
tomer needs. The increasing rail network allowed nationwide and cross-country shipping of
products to reach a greater customer base, as did the wide-spread use of steam-powered
merchant ships. The noticeable contribution was devoted to industrialisation by the founder
of the Ford Motor Company, US-born Henry Ford. After developing his own concept of a
steam machine in 1878 at the age of 15, he revolutionised the manufacturing process of his
Model T motor car by implementing the first moving assembly line in his Michigan plant in
1913.19 Not only did Ford adopt the principles of the division of labour, he also optimised
the mass production of the Model T by implementing a fixed throughput time to boost the
total output of cars. He is famous for having said that his customers can have their Model T
in any colour, as long as it is black, underpinning his one product mass production strategy.20

18
Cf. Martin 2019
19
Cf. Ford 2021
20
Cf. Vlaskovits 2011

11
Ford’s approach was subsequently adopted worldwide and had a significant influence on
many industries way beyond automotive manufacturing.

Only a few years after Ford’s assembly line was introduced, the first world war brought
industrialisation to a sudden halt and, to a great extent, shifted production towards military
goods and equipment. After the golden years in the 1920’s the second world war again dis-
rupted the inherent economic growth and further pushed military needs on top of the indus-
trial and political agenda. The late 1940’s onwards, after signing numerous peace treaties to
ensure political stability, were mainly characterised by the building and rebuilding of nations
and society, economic growth and development. Despite the adverse effects of two global
wars on society as a whole, from a technological point of view it evidently enhanced indus-
trial science, research and development and had a significant effect on the development of
manufacturing technologies and methods. Similar to the expansion of rail transport in the
middle of the 19th century, the invention of the gas turbine and its use for different types of
aircraft was soon translated into the foundation of civil airlines, which exerted a marked
contribution to transport optimisation and the beginning of globalisation from the middle of
the 20th century onwards. Goods could thus be shipped effortlessly worldwide and manufac-
turing sites could be located abroad to harness the cost and location-related benefits of for-
eign countries.21

Apart from the shift from rail transport to aviation, early computational devices used during
and after the second world war reached their technical limits owing to the increasing need
for vacuum tubes, which entailed a lengthy process if errors arose and replacement was
needed. This is exemplified by ENIAC, the Electronic Numerical Integrator and Computer
that was developed on behalf of the US Army in 1942 and publicly released four years later.
Back then, ENIAC was considered the first “Turing-complete” programmable computa-
tional device, given that it could compute trivial numerical problems. It weighed 27 tons and
occupied 167 square meters of space.22 However, despite the promising functionality, the
more components were used in any computational device and system, the more difficult,
complex and costly it became to maintain. Similar to the underlying problems of the first
and second industrial revolution, this problem arguably sparked new ideas and inventions.
For example, the invention of the transistor in 1947 enabled the development of smaller and
more energy-efficient devices.23 Twelve years later, in 1959, the inventors Robert Noyce
and Jack Kilby successfully developed the integrated circuit, commonly known as a micro-
chip, which solved the inherent problem of building large-scale computational devices that
were functional and easy to maintain. The integrated circuit and microchip revolutionised

21
Cf. Davies et al. 1975
22
Cf. Swaine 2021
23
Cf. CHM 2021

12
the whole electronics industry and set the cornerstone for new innovations in the field of
computer technology. Furthermore, Robert Noyce was also co-founder of Intel in 1968, a
key industry player of the 21st century, and pushed the development of Intel’s first micro-
processor in 1971.24 At the same time of the development of integrated circuits and micro-
chips, first attempts were made to automate human work by using a so-called reprogramma-
ble manipulator, most commonly noted as robot. It was the businessman and engineer Joseph
Engelberger who successfully transferred early concepts of robots into an industrial robot
that could be implemented in manufacturing to automate manual processes, increase produc-
tivity and reduce human error.25 The Robot Institute of America, as cited by Stanford Uni-
versity, defined a robot as "a reprogrammable, multifunctional manipulator designed to
move material, parts, tools, or specialized devices through various programmed motions for
the performance of a variety of task." 26

The crucial hardware device that enabled robot movements along multiple axes and indus-
trial applications in the first place, was the programmable logic controller (PLC). The PLC
was invented by Dick Morley and like-minded engineers in 1968 who envisaged the ad-
vantages of a device that could handle input and output signals according to a predefined
programme. Additionally, it should be capable of being used for industrial applications
where high temperatures, vibrations, electrical noise or any external influencing factors have
an adverse effect on operations. Only one year later, the US-based automotive OEM General
Motors invested approximately one million dollars to equip its factories with the innovative
PLC to optimise production. The early adoption of PLC devices in the automotive industry
exemplifies the first-mover philosophy of automotive OEMs and suppliers which has pre-
vailed until today. The PLC has undergone numerous improvements since its first applica-
tions and is considered standard in contemporary manufacturing.27 Considering the above-
mentioned technological innovations, it is arguable that the 1970’s onwards marked the up-
swing of advanced computer technology and industrial automation, hundred years after the
kick-off of the second industrial revolution. In effect, the availability and ongoing develop-
ment of computer hardware had a marked effect on industrialisation. It also formed new
areas of expertise, such as software development and information and communications tech-
nology (ICT). The need for OEMs to further optimise manufacturing processes by harness-
ing the opportunities of automation, called for a means to effectively monitor, control and
handle all processes from business to manufacturing. Paper-based systems had been domi-
nating engineering and management since the Egyptian high culture, yet it was the increasing
complexity that required systems to handle an increasing amount of data and information.
As David Wilson, a famous author on information management, argued that “flexibility and

24
Cf. Intel 2021
25
Cf. Stanford University 1998
26
Stanford University 1998
27
Cf. Dunn 2009

13
responsiveness are paramount, and these abilities of course depend on information – the way
it is collected, processed and acted upon.“28 As a matter of fact, it is the systems put in place
that make the difference in collecting, processing and making use of data and information.

In his personal account, former engineer Dale Fisk of IBM recalls the challenges of early-
stage programming in the 1970’s as follows: “Writing a program began with a paper tablet
of coding forms. Each page of the tablet had about fifty lines on it, and each line on the form
would eventually be converted into a punched card and stowed away in a box with a bunch
of other cards.”29 At the same time, the software company SAP was founded, today the big-
gest player for industrial software solutions to be used in enterprise resource planning (ERP)
and the like. SAP in particular was one of the frontrunners to harness the benefits of compu-
tational devices with its software solution for financial accounting. Given that accounting
departments had to deal with numerical problems, any computer-based system that was able
to support the paper-based process would markedly increase productivity and minimise the
risk of human error. In the beginning, the system fully relied on punched paper cards to be
used. The principle of using programme codes, though digital later, had remained the major
method of SAP’s ERP software for decades. The use and implementation of new technology,
especially computational devices as well as innovative software, experienced a steady
growth across all industries from the 1970’s onwards.30 In effect, companies came quickly
to realise that “technology affects competitive advantage if it has a significant role in deter-
mining cost position or differentiation”.31

Due to the widespread adoption of new technology, further improvements and optimisations
made hardware significantly smaller and more affordable, whereas software became more
powerful and covered more aspects and applications of business and industrial manufactur-
ing. In 1975 the co-founder of Intel, Gordon Moore, coined his famous Moore’s Law, when
he put forward, that with hardware becoming increasingly smaller and less expensive, tran-
sistor components in particular, the computing performance would double every year.32 In
this respect, it was Bill Gates, founder of Microsoft, who made a significant contribution to
the advent of personal computing by releasing his first operating system MS DOS in 1981.
It laid the cornerstone for MS Windows as the most successful and globally used operating
system for personal computers ever since. What followed were numerous client and server
applications with more sophisticated functions, features and even graphic user interfaces.
The availability of hardware and peripherals, from screens up to user input devices, grew

28
Cf. Wilson 2000, p. 15
29
Fisk 2005, p. 2
30
Cf. Lazarevic 2019; Cf. Fisk 2005, p. 2ff.
31
Porter 2004, p. 169
32
Cf. Simonite 2016

14
rapidly from then.33 Further contribution to the adoption of computer technology for indus-
trial applications was the development of communications technology. As the supplier net-
work spread globally and complexity increased to achieve low lead-times and a high level
of operational efficiency, there was an inherent need for IT-driven communication. Ap-
proaches towards transmitting data electronically started in the early 1970s, using the so-
called electronic data interchange (EDI) standard. It allowed businesses to transmit data be-
tween senders and receivers.34 The most critical gamechanger arguably was the implemen-
tation of the global internet and the world wide web (WWW) that experienced its break-
through owing to the contribution of Tim Berners Lee, who is accounted for the invention
of the WWW.35 The internet, as such, entailed many services, standards and functions, which
could be seen as a revolution by itself and was meant to transform many existing industries
and set off many new ones. The development of information and communications technol-
ogy had drastically transformed the way factories operated and further entailed new oppor-
tunities to enhance the use of computational devices and industrial technologies, such as
automated robots.

There is no doubt that all the technological breakthroughs of the Industry 3.0 era, including
both sophisticated hardware and software, had a marked effect on both industrial science
and best practice. The accomplishments are best exemplified and summarised by the so-
called Industry 3.0 automation pyramid, as shown in figure 1, that puts a layer perspective
on industrial operations and automated factories, placing emphasis on the systems and tech-
nologies used. It can also be regarded as the result of global collective wisdom in industrial-
isation. By contrast, contemporary Industry 4.0 adopts a connected network perspective.36

Figure 1: Industry 3.0 and 4.0 architecture (Frere et al., 2018)

33
Cf. Hall 2021
34
Cf. Segev et al. 1997, p. 157ff.
35
Cf. Cerf et al., 1997, p. 14
36
Cf. Frere et al. 2018, p. 8f.

15
The automation pyramid acknowledges a bottom-up approach with a technical focus on data
generation as well as a top-down management perspective focussing on operations planning,
depending on the respective level of management and control or relative to the manufactur-
ing process. The latter contains the machinery and devices on the factory floor that carry out
all manufacturing tasks. The upper level, level one, comprises both sensors and actuators
that are connected to PLCs that process input and output signals. The PLCs enable the ma-
chinery and robots to carry out the predetermined tasks and movements according to input
signals, deliberate control and the respective PLC programme. In effect, the key technolog-
ical advancement of Industry 3.0 can be found in the levels above. Level 2 includes computer
systems for automation and control, so-called supervisory control and data acquisition
(SCADA) systems. These systems enable the monitoring and controlling of the manufactur-
ing process, in order to ensure that it runs as planned and that failures as well as any devi-
ances are detected and taken care of as early as possible. Level 3 includes manufacturing
execution systems (MES) that enable proper planning in terms of machine and worker allo-
cation according to customer orders and contracts. In addition, MES systems allow opera-
tional data acquisition to monitor each individual manufacturing process and machine in
detail, including running time, failures as well as staff working hours. The top level com-
prises enterprise resource planning (ERP) systems, such as SAP ERP, that enable OEMs and
suppliers to fully plan and manage their operations, from customer order to delivery.37

The automation pyramid thus exemplifies the state-of-the-art technologies of Industry 3.0.
In fact, the widespread application of Industry 3.0 together with global R&D initiatives to
further improve operational efficiency and effectiveness using sophisticated internet-enabled
hardware and software, led to the fourth industrial revolution. Industry 4.0 is therefore char-
acterised by the next level of manufacturing technologies including cyber-physical systems,
internet of things, enhanced connectivity and global networks as well as AI. The traditional
notion of the automation pyramid and a level perspective shifted towards a network perspec-
tive with a seamless integration of smart devices, services and products in a smart factory.38.
This new perspective calls for an enlightened approach towards defining a smart factory’s
architecture and system landscape. According to Wan et al. (2020), AI is at the core of future
smart factories and the key enabler for managing complex and interconnected processes in
a dynamic global environment. Figure 2 shows a smart factory architecture that takes ac-
count of new challenges and components of Industry 4.0. It illustrates the respective interre-
lationships with AI being considered as the digital brain for monitoring and controlling the
smart factory through the use of AI models, big data as well as edge and cloud computing.39

37
Zühlke et al. 2011, p. 567ff.
38
Cf. Frere et al. 2018, p. 8ff.
39
Cf. Wan et al. 2020

16
Figure 2: AI-enabled smart factory architecture (Wan et al., 2021)

The highly regarded Austrian management thinker Joseph Schumpeter has made a signifi-
cant contribution to the understanding of the forces and influencing factors that led to Indus-
try 4.0, as cited by researchers Jovanovic et al. (2018) in a recent paper. In particular, Schum-
peter referred to so-called innovation cycles, that all entail significant economic upswing
due to the widespread adoption of innovative technologies that push whole industries and
markets. Figure 3 illustrates the development of the innovation cycles from 1785 until the
fifth wave that commenced from 2020 onwards, with digital networks, software and new
media regarded as the new driving forces.40 It is evident that every cycle since the first in-
dustrial revolution has occurred within a shorter period of time. The wide adoption of AI
might cause the sixth cycle, which could lead to an even shorter time span.

Figure 3: Schumpeter's Innovation Cycles (Jovanovic et al., 2018)

40
Cf. Jovanovic et al. 2018, p. 4

17
4.1.2 Politics in the Context of Industry 4.0 and AI
The wide-spread adoption of new technologies, especially in the digital era, is also driven
by politics and government support in order to foster economic development. In fact, the
term Industry 4.0 has gained increasing significance in Europe since the speech of German
chancellor Angela Merkel at the Bundestag in 2015. Merkel continued to stress that the dig-
ital transformation will be a crucial determinant of industry competitiveness, with autono-
mous production systems, machine to machine communication as well as changing and
adapting supply chains as the critical success factors. Merkel argued, that Industry 4.0 will
not only have an impact on manufacturing, rather it will have an effect on society as a whole.
The rapid adoption of Industry 4.0 technologies and the advancement of manufacturing is
considered the key lever of success and long-term prosperity.41 While a certain degree of
autonomy arguably demands the use of an AI system, this was not explicitly stated in the
early years of Industry. 4.0.

As emerging countries, such as China, have gained edge in both manufacturing efficiency
and quality, it becomes ever more important to contemplate ways to stay ahead, with Indus-
try 4.0 on top of the agenda. In this respect, in a recent report Meyer et al. (2021) argue, that
Asian manufacturers have already envisioned the adoption of Industry 4.0 themselves, with
an estimated growth of manufacturing from 400 billion dollars in 2020 to 600 billion dollars
in 2030, considering the successful implementation of new technologies. Meyer et al. further
claim that “the many forces that are redefining the economics of global manufacturing pre-
sent Southeast Asia with a golden opportunity to move up the value chain in a range of
sectors.”42 This necessitates appropriate action in Europe and well-planned initiatives and
funding if European manufacturers seek to remain competitive. Similar to the German gov-
ernment movement, the Styrian government had followed suit in promoting the shift towards
digital transformation. The former member of the government and economic council, Chris-
tian Buchmann, termed this shift smart production, rather than Industry 4.0. Buchmann pos-
tulated, that the global competition of manufacturers and the increasing need for innovation
pushes the development of connected machines and services along the whole value chain. In
addition, he placed emphasis on the human factor, by stressing that the adoption of smart
production doubtlessly entails new opportunities to enhance industrial education and training
as well as the creation of new jobs. Contrasting Merkel’s mere focus on Industry 4.0 manu-
facturing technologies, Buchmann highlighted the importance of services. The latter argua-
bly are a distinguishing factor for OEMs and suppliers, yet they can also enable small and
medium enterprises (SME) to jump on the bandwagon and offer innovative services related
to smart production.43 The German and Styrian governments exemplify the need for a liaison

41
Cf. CDU 2014
42
Meyer et al. 2021
43
Cf. Land Steiermark 2014

18
between politics and industry in order to foster the development and adoption of Industry
4.0 technologies. While both big industry players and SME companies have been reluctant
to adopt new technologies unless there was an inherent need or proven benefit, the digital
transformation is actively promoted by politics and supported by national and European Un-
ion economic development and funding initiatives. For example, the Austrian government
has been promoting a funding programme called “Production of the Future” since 2011 that
provides annual financial support for industrial research projects dealing with innovative
Industry 4.0 technologies.44 More recently, in 2021 a specific funding programme was in-
troduced to foster the development of AI technologies in the field of sustainable production
to increase production efficiency and minimise resource and energy consumption.45 Consid-
ering the above-mentioned aspects, it is evident, that politics play a more important role for
driving the fourth industrial revolution, compared with the revolutions before.

4.1.3 Definitions of Industry 4.0

Given the sheer complexity of Industry 4.0 and the great variety of technologies and appli-
cations, there is no common ground when it comes to the definition of the term. Recent
research conducted by Culot et al. (2020) shows, that the “lack of agreed-upon definition
poses limitations to theory building and research comparability.”46 While definitions of In-
dustry 4.0 can shed light into the key characteristics of the underlying technological revolu-
tion, it is also evident that any definition tends to be highly focused, specific and to a certain
extent omits critical state-of-the-art technologies. Hence, definitions arguably depend on the
perspective that is adopted and pose the inherent risk of both bias and subjectivity.

In their research, Culot et al. (2020) acknowledge that Industry 4.0 as a concept has its roots
in Germany, dating back to 2011 and the famous Hannover Technology Fair.47 In particular,
according to the German Federal Ministry for Economic Affairs and Energy and its Platform
Industry 4.0, it can be defined as “the intelligent networking of machines and processes for
industry with the help of information and communication technology.”48 The major benefits
are considered to be enhanced flexibility, more efficient processes, customer centricity and
resource efficiency. Thus, much emphasis is placed on a smart approach by connecting man-
ufacturing equipment using ICT to achieve certain business benefits. While the definition of
the German Platform Industry 4.0 does give an idea of the underlying principle of the fourth
industrial revolution, it is not clear what constitutes Industry 4.0 in terms of the technologies
used. In this respect, further contribution comes from the association of German engineers,
which issued an announcement on the topic in 2011, preceding the Hannover Technology

44
Cf. FFG 2021
45
Cf. Mosnik et al. 2021
46
Culot et al. 2020
47
Cf. Culot et al. 2020
48
Plattform Industrie 4.0 2021

19
Fair. In particular, they provided a short definition of Industry 4.0, calling it the era of IoT-
enabled cyber-physical systems. While the internet of things represents the connected eve-
rything and everywhere, cyber-physical systems dig deeper into the technological advance-
ment of the fourth industrial revolution by representing complex connected and smart sys-
tems that are capable of carrying out multiple tasks, in part autonomously without direct
human influence and control. 49

The focus on the internet of things as part of the fourth industrial revolution can be further
translated into its role for the business strategy as a whole. Researchers at Hochschule Wis-
mar and Tallinn University of Technology respectively, as cited by Piccarozzi et al. (2018),
argue that “Industry 4.0 refers to the integration of Internet of Things technologies into in-
dustrial value creation enabling manufacturers to harness entirely digitized, connected,
smart, and decentralized value chains.”50 This definition highlights the important role of
value creation as a key driver of competitive advantage and operational excellence. The im-
plementation of any new technology in manufacturing requires a demonstrated return on
investment in order to ensure commitment from the shop floor to top management.

Opposing the focus on advanced technology implementation and value creation, Telukdarie
et al. (2018) provide an enterprise definition of Industry 4.0, postulating that the seamless
exchange of information, both vertically and horizontally, lies at the core of the fourth in-
dustrial revolution. As a result, the major goal is to establish a “value chain that self-adjusts,
generates and shares information.”51 This definition highlights the importance of data to un-
tap the full potential of Industry 4.0. On the other hand, BCG, a big three global consultancy
firm, adopts a more practical approach and defines Industry 4.0 as the digital transformation
of industry that is driven by nine distinct technologies, including autonomous robots, simu-
lation, horizontal and vertical system integration, the internet of things, cybersecurity, the
cloud, additive manufacturing, augmented reality as well as big data and analytics.52

While it is argued that the quality and speed of information is the critical success factor in
the digital era, there is no doubt that AI should be considered a key lever, and should there-
fore complement the nine core technologies of Industry 4.0 proposed by BCG, which are
illustrated in figure 4.

49
Cf. VDI Nachrichten 2011
50
Piccarozzi et al. 2018
51
Telukdarie et al. 2018
52
Cf. Rüßmann et al. 2019

20
Figure 4: BCG Industry 4.0 technological aspects (Rüßmann et al., 2019)

In this respect, Mhlanga (2021) postulates that “AI in the fourth industrial revolution is be-
ginning to live up to its promises of delivering real value necessitated by the availability of
relevant data, computational ability, and algorithms.”53 As the concept of big data emerges
and grows of importance in business, so does AI as a means to deal with it to improve oper-
ational effectiveness and efficiency.54

An enlightened approach towards defining Industry 4.0 is adopted by author David


Siepmann, who argues that Industry 4.0 is not just a set of innovative technologies that can
be implemented by manufacturers to boost efficiency and effectiveness. Rather, it is up to
top management to not only drive the digital transformation process to successfully outpace
their peers. The company-wide implementation of Industry 4.0 requires amendments to a
company’s vision and culture as well as its business strategy, and it therefore necessitates
the development of new business models and processes. 55 In effect, it is the selection of
appropriate technologies and the well-planned alignment of a company’s strategy and re-
sources which make the difference between success and failure. Companies that jump on the
band wagon of Industry 4.0 and are keen to adapt themselves appropriately, will make the
most out of the digital era to ensure sustainable growth and prosperity from a long-term
perspective. Considering the above-mentioned aspects, there is no doubt that a uniform and
universally applicable definition of Industry 4.0 does not exist. It was shown that the defini-
tion, to a great extent, depends on the respective focus that is applied. It primarily builds on
the subjective perspective in terms of technologies, core values and benefits as well as busi-
ness strategy. In effect, Industry 4.0 as the fourth industrial revelation entails not only ad-
vanced technologies that are implemented to increase operational effectiveness and

53
Mhlanga 2021
54
Cf. Shani 2015
55
Cf. Siepmann 2016, p. 20ff.

21
efficiency, it can also be regarded as a completely new mindset of manufacturing and man-
agement that is required in the light of the digital era, and that is based on real-time data
generation and decision-making as well as continuous improvement based on solid infor-
mation. This arguably is where AI can unleash its potential and make the difference between
average implementations of Industry 4.0 or sustainable operational excellence to achieve
competitive edge.

4.1.4 Aspects and Components of Industry 4.0

The industrial revolutions that preceded the era of the digital transformation were character-
ised by just a few technological breakthroughs, as was outlined in the chapter beginning. By
contrast, Industry 4.0 entails a much greater variety of aspects and components that comple-
ment each other to achieve an unprecedented operational excellence along the whole value
chain. The sheer pace of technological advancement since the invention of computer devices,
sophisticated software and the internet during the third industrial revolution, brought about
unforeseen opportunities in the context of manufacturing. In this respect, it is worth noting
that similar to the contrasting definitions of Industry 4.0, there are different views on what
components constitute Industry 4.0 and the digital transformation of manufacturing respec-
tively.
4.1.4.1 Technological Phases of Industry 4.0
When it comes to the principles of Industry 4.0, there are three distinct technological phases
that are considered a prerequisite in contemporary manufacturing. In particular, according
to Siepmann (2016), these include (1) ubiquitous computing, (2) internet of things and (3)
cyber-physical production systems. It is worth pointing out that the advancements of the
third industrial revolution, such as computing devices, software or ICT in general, are part
of the so-called ubiquitous computing environment. The technological development has led
to ever smaller and higher performing computer systems. Furthermore, nowadays even the
smallest components are enriched using innovative microelectronics, which constitute a shift
towards so-called embedded computing as well as the development of embedded devices.
The latter form the smallest parts of a bigger system. The combination of embedded devices
with advanced communication technology paves the way for the internet of things. This has
been achieved by the upgrade from IPv4 to IPv6, the internet protocol (IP) for assigning an
IP address to any single device that is connected to the internet. The latest protocol allows
up to 340 sextillion IP addresses and makes sure that the internet of things is feasible in
practice, not just in plain theory. In particular, if every single device is IoT-enabled and
connected to the internet, new opportunities emerge that can be harnessed in manufacturing.
It enables the development of smart devices, machines and products, up to whole smart fac-
tories that have the capability to communicate with each other, perform data analyses and
carry out specific tasks and processes autonomously and self-adjusted. All stakeholders of
this new eco system form a fully connected value chain, which is characterised by both hor-
izontal and vertical integration through seamless exchange of data and information. There is

22
no doubt that this approach has a decisive effect on operational effectiveness and efficiency.
In order to achieve this effect a number of technologies need to be aligned appropriately. In
effect, apart from embedded devices and the IPv6 internet protocol, there is a need for high
performing communications technology, such as LAN, WIFI or mobile internet, including
5G. The communication between devices or machines with humans as well monitoring and
controlling requires sophisticated user interfaces and means for identification, such as RFID
or NFC. Through the implementation of sophisticated sensors in any device or machine,
important data and information can be gathered and processed. In addition, the integration
of actuators allows translating digital signals into mechanical movement or physical quanti-
ties.56 Thus, a machine can be enabled to process data generated by sensors, such as temper-
ature sensors or cameras, to perform certain movements and tasks through integrated actua-
tors, such as a gripper, which all forms a part of the production process.

With the appropriate technologies in place, manufacturers can move one step and phase fur-
ther and establish what is called a cyber-physical production system (CPPS). This system is
based on a well-planned combination and integration of individual cyber-physical systems,
which transform a traditional factory into a true smart factory. In fact, traditional factories
have long been focussing on fragmented systems and strict hierarchies, both vertically and
horizontally, which inhibited operational efficiency due to an increase in communication
effort, lengthy processes and lack of integration. Sensors and actuators implemented in CPS
that are capable of gathering and transmitting data, or translating input data into mechanical
actions, can directly influence the production process as well as enhance management deci-
sion making. CPS are usually connected via communication technologies, such as LAN or
WIFI, and are monitored and controlled using advanced user interfaces, up to data glasses
harnessing both virtual reality (VR) and augmented reality (AR). CPS in the context of In-
dustry 4.0 consist of a three-layer architecture: (1) Physical objects including machinery,
smart devices or 3d printers, (2) data storage and (cloud) database systems for documents,
production data or models, (3) application layer, with monitoring, reporting and controlling
features and functions. 57 This layer is where data are processed and used for evaluating and
improving the manufacturing process. In this respect, the implementation of AI, advanced
algorithms and analytics have the potential to increase the quality of information, reduce the
total process time and manpower needed, and deliver unprecedented analytical results and
performance.

56
Cf. Siepmann 2016, p. 47ff.
57
Cf. Siepmann 2016, p. 35ff.

23
4.1.4.2 Key Components of Industry 4.0
From a manufacturing point of view Industry 4.0 is built on three pillars that comprise a
diverse set of technological components: The pillars include (1) data generation and pro-
cessing, (2) machine to machine communication, (3) human machine interaction. In fact, the
traditional notion of the automation pyramid that represents the achievements of the third
industrial revolution is challenged in the light of Industry 4.0. It is argued that each layer of
the automation pyramid represents a distinct technology, process and system, while each
layer connected, yet logically separated from others. While the top layers, such as ERP or
MES, are devoted to planning functions, the bottom layers focus on the manufacturing pro-
cess and the generation of data. In effect, the components of traditional ubiquitous compu-
ting more often than not do not live up to the high standards and the required performance
for generating and processing huge amounts of data (big data), primarily owing to the in-
creasing connection and integration of devices, machines or products as well as players of
the whole value chain. This poses challenges when there are no technologies and processes
in place to effectively handle big data. Industry 4.0 enables to overcome this by implement-
ing cloud services to directly transmit data from sensors of the shop floor and process them
using analytics services. The data can then be transmitted to the control layer and the respec-
tive systems via high-speed internet. Another setback of the traditional automation pyramid
is caused by the shift from centralised to decentralised manufacturing control. In effect, the
implementation of CPPS and autonomous smart factories changes the requirements and in-
terdependencies of the respective layers. For example, it is expected that production planning
is increasingly integrated in ERP systems, whereas some MES functions move down to the
control and process layer respectively. The programmable logic controller of the past can be
substituted by highly sophisticated cloud-based PLC services. All these changes to the tra-
ditional view of the automation pyramid call for an adaption according to the characteristics
of Industry 4.0.58

It is argued that big data is among the aspects that distinguishes Industry 4.0 from the third
industrial revolution. A key component and practical example thereof is data acquisition
driven by advanced communication technologies, including RFID and NFC. The continuous
development of microelectronics, chips and transponders in particular, has led to a signifi-
cant reduction in size, costs and an optimisation of the transmission range. This, in turn,
enables more versatile applications and a wide-spread adoption in manufacturing as a means
for localisation, identification, monitoring, documentation and authentication. As opposed
to barcodes or QR tags, RFID chips are capable of withstanding demanding manufacturing
conditions and are usable without visual contact. In effect, products that are RFID-enabled
and thus made smart, can document crucial process and quality data during the manufactur-
ing process, which are further processed using advanced IT systems. The production team

58
Cf. Siepmann 2016, p. 49ff.

24
can thus determine the location and quantity of the respective product any time. By integrat-
ing stakeholders along the value chain, products can be seamlessly traced across a manufac-
turer’s boundaries. In cases of quality issues, defective products can also easily be tracked
according to the data documentation. In addition, process data can deliver clues on the un-
derlying causes of the problem, such as deviating machine parameters or faulty tools used
during the process. The above-mentioned technologies have also had a decisive impact on
the development of mass customisation and lot size one, a move to further achieve compet-
itive advantage by excelling at meeting customer expectations. Integrated CPPS can thus
autonomously adapt to changing circumstances and needs to manufacture products that are
tailor-made to customer needs and demands.59 In this respect, it is worth noting that signifi-
cant improvements have been made in the field of 3d printing to support this trend. The
combination of advanced 3d modelling software and algorithms, innovative materials and
3d printing technologies, have paved the way for a new wave of customisation opportunities
and the production of products with distinguished shapes and special features, that were not
feasible with traditional production methods.60

When it comes to data processing and information technology, CPPS harness the full poten-
tial of cloud computing, which comprises sophisticated and high-performance applications
and services provided via the internet.61 In fact, the development of high-speed and stable
internet, including mobile 5G technology, has had significant influence on the use of cloud-
based solutions. In the past, the maximum bandwidth was a critical limitation of the perfor-
mance of any internet-based application. When the world wide web (WWW) was introduced
in the 1990’s, the bandwidth available peaked at 56 kilobit per second. Nowadays, the inter-
net bandwidth can reach more than one gigabit per second, which is almost 20 times faster
with markedly lower latency. The current mobile communication standard 5G as of 2021
enables data rates of up to ten gigabit and is therefore considered a key gamechanger.62 In-
dustry players have previously invested in on-premise solutions, which required substantial
resources to be planned, developed and maintained. Nowadays, the benefits of cloud com-
puting can outweigh those of on-premise solutions, especially where companies are not able
to establish state-of-the-art systems themselves. According to Avram, cloud computing en-
ables even small companies to benefit from immediate access to high performance hardware
and software solutions and innovative technologies, given lower entry barriers and signifi-
cant cost advantages. On the other hand, big enterprises are enabled to quickly scale up their
services and enhance their business processes and operations when needed. Nevertheless,
cloud computing poses adverse effects that are to be taken into account, including data se-
curity, dependability and flexibility, reliability of service, process fit and change as well as

59
Cf. Siepmann 2016, p. 51ff.
60
Cf. Rüßmann et al. 2019
61
Cf. Siepmann 2016, p. 54ff.
62
Cf. Murphy 2019

25
the business benefit and value in the long run.63 A recent technological development is mak-
ing AI services and applications available through cloud computing. This enables both big
industry players and SME with limited experiences and resources to be among the frontrun-
ners of adopting and implementing AI models and methods. They can thus benefit from
enhanced operations using machine learning and big data to achieve competitive edge. 64 In
effect, the combination of advanced communications technologies, cloud computing, big
data and analytics allows the establishment of an intelligent operations and manufacturing
network. Consequently, from a manufacturing point of view, the major element of Industry
4.0 is a more efficient and effective convergence of all layers of the traditional automation
pyramid, including machinery on the shop floor level, up to actual products in the field. In
this respect, an essential component and lever to achieve manufacturing excellence in the
light of Industry 4.0 is machine to machine communication (M2M), which allows automated
information exchange and self-adjusting operations. Research and development as well as
science have devoted much effort to establish industry-wide standards that enable reliable
and performant M2M. In the past, the lack of compatibility of systems, missing links and
interfaces as well as harmonised standards inhibited the potential and implementation of
M2M. The development of communication protocols and industry standards, such as OPC
UA or MQTT had a decisive effect on the adoption of M2M in manufacturing.65

Figure 5: Industry 4.0 architecture (Liu et al., 2019)

63
Cf. Avram 2014, p. 531ff.
64
Cf. Amanath et al. 2020
65
Cf. Siepmann 2016, p. 59ff.

26
The implementation of M2M and a sample architecture is depicted in figure 5, as cited by
Liu et al. (2019). In particular, it shows the use of sensors and machinery on the shop floor
level, the connection via WIFI or Ethernet to the on-premise OPC UA server, up to the link
and connection to cloud-based systems. 66 In effect, communication standards and protocols,
such as OPC UA, enable the reliable and real-time connection of industrial devices and ma-
chines and the exchange of data and information, which can then be stored in cloud databases
and processed by analytical and manufacturing control systems, enabling technologies of
big data. The data can then be used to monitor manufacturing and operations performance,
engage in predictive analytics, derive at qualitative data-driven forecasts or develop appro-
priate measures for further development and optimisation.67

A ground-breaking approach towards seamless full-scale integration and M2M is exempli-


fied by the German OEM and premium car maker Mercedes-Benz and its so-called “Mer-
cedes me” service, that is also relies on state-of-the-art cloud computing. The car itself acts
as a smart product equipped with advanced sensors and communication technologies, and is
connected to the Mercedes-Benz cloud system. While on the road, any errors that occur are
transmitted to the OEM cloud system and passed through to manufacturing and R&D for
evaluation in order to avoid any adverse effects. This enables Mercedes-Benz to actively
engage in quality assurance, product development and manufacturing optimisation based on
real-time field data. Also, the customers are enabled to immediately book an appointment at
the local dealership using the in-vehicle touchscreen user interface, in order to have the prob-
lem fixed.68 This approach ensures a high level of customer service, and at the same time,
allows stakeholders to generate data for further processing, analytics and documentation.
When it comes to user interfaces and human machine interaction, there has been noticeable
technological development in contemporary manufacturing compared with passenger cars.
In fact, the workers, machine operators and production managers need to interact with man-
ufacturing equipment and software applications on a daily basis. Any machine or system in
place should allow easy access and control as well as a proper visualisation of necessary data
and information. In recent years, much progress has been made in the development of both
virtual reality (VR) and augmented reality (AR) using advanced devices, such as smart
glasses or tablet computers. In particular, VR can be used to virtually visualise any 3D object
or even a full factory with manufacturing equipment in the original setting, shape and size.
AR, on the other hand, enables the emergence of virtual models and digital information with
the real world.69 The use of both VR and AR creates new opportunities to enhance both
engineering and manufacturing. However, the key to success is the seamless integration of
multiple technologies of Industry 4.0, rather than adopting single technologies.

66
Cf. Liu et al. 2019
67
Cf. Siepmann 2016, p. 56ff.
68
Cf. Mercedes-Benz 2021
69
Cf. Siepmann 2016, p. 63ff.

27
The above-mentioned components of Industry 4.0 represent the high technological advance-
ment since the previous industrial revolutions. Cyber-physical production systems and the
convergence of real with virtual worlds are considered the new norm in the Industry 4.0 era,
as is illustrated in figure 6.70 A concept that has emerged in the light of the digital transfor-
mation is the real-time or live enterprise, which may constitute the next technological ad-
vancement in combination with sophisticated AI.

Figure 6: Development from Industry 1.0 to Industry 4.0 (Frere et al., 2018)

4.1.5 Concept of Real-Time and Live Enterprises


The full and seamless integration of Industry 4.0 technologies including IoT, M2M, CPS
and ultimately AI can push manufacturers to become a real-time or live enterprise that is
managed through information and knowledge gained from real-time data involving all busi-
ness processes. While the term real-time enterprise was first coined by Gartner analysts in
2003 as a future vision of business, it has recently been adapted and promoted by the global
IT firm Infosys Technologies as “live enterprise”. The latter is a concept that evolves around
a company’s “digital brain” that “manages the connective knowledge and ecosystem, learns
from interaction, and drives value exchange in each interaction.” 71 In this respect, AI argu-
ably plays a key role to cope with the sheer volume and complexity of big data that is gen-
erated within the whole ecosystem and enable the concept of live enterprises.

70
Cf. Frere et al. 2018, p. 7
71
Cf. Kavanaugh and Tarafdar 2021, p. 4ff.

28
The marked shift from traditional towards live enterprises has significant implications for
various elements of a company’s way of operating, such as the organisational setup, value
chain, decision making process and IT systems. This notion implies that Industry 4.0, AI and
applications that go beyond what is conceivable today, have the potential to entail substantial
change and lead to an even more dynamic environment. Table 4 shows a comparison of the
major aspects of the live enterprise transition including organization, value chain, decisions
and IT systems with traditional operations, as cited by Kavanaugh and Tarafdar (2021):72

Table 4: Live Enterprise operating model elements (Kavanaugh and Tarafdar, 2021)

Operating
Traditional (from) Live Enterprise model (to)
Model Element

Organization ▪ Command and control ▪ Self-organizing work teams


▪ Functional, line-of-busi- ▪ Cross-functional and platform-mind-
ness-based large teams set-based small teams
▪ Physical workplace ▪ Anytime, anywhere workplace
▪ Waterfall processes ▪ Distributed agile processes
▪ Technology as enabler ▪ Technology as strategic differentiator

Value Chain ▪ Distant from input ▪ Proximity to source


▪ High latency ▪ Zero latency
▪ Offline, periodic analysis ▪ Instant simulation
▪ Periodic surveys ▪ Instant micro-feedback
▪ Complex process ▪ Guided practice

Decisions ▪ Deterministic ▪ Stochastic


▪ Manual ▪ Semi-automated
▪ What-if analysis ▪ Predictive and prescriptive
▪ Distributed organizational ▪ Connected and curated organizational
knowledge knowledge
▪ Insights for review ▪ Insights to actions

IT Systems ▪ Features and functionality ▪ Platform mindset


mindset ▪ Designed for evolvability
▪ Designed for known re- ▪ Unbundled services
quirements ▪ Extreme automation
▪ Monolithic systems ▪ Observable
▪ Moderate automation
▪ Limited telemetry

72
Kavanaugh and Tarafdar 2021, p. 7

29
4.2 Principles of the Automotive Industry

The automotive industry ranks on top of the most innovative, complex and globally dis-
persed industries in today’s business world and is considered an early adopter of Industry
4.0 technologies. More than hundred years of innovation and continuous improvement have
contributed to the production of automobiles that initially were petrol engine-powered horse-
drawn carriages in the late 19th century, which developed to digital and connected cars that
can even respond to spoken language and are capable of driving autonomously, as is exem-
plified by many OEMs and their products today. While cars had traditionally been hand-
crafted in small workshops, they are manufactured as fully individualised and industrialised
mass products today. The history of the automotive industry has experienced significant
milestones and developments until the adoption and implementation of both AI and Industry
4.0. It is arguable that the inherent complexity of the automotive industry has positioned
both OEMs and suppliers as steady frontrunners throughout the recent industrial revolutions
and the digital transformation in particular.

4.2.1 Key Milestones in Automotive History


The invention of the automobile dates back to the late 19th century, with the German inventor
Carl Benz being the first to register a patent for an automobile powered by a petrol engine
in 1886. While Benz’ invention was an actual three-wheeler, the four-wheeler that has pre-
vailed until today was developed at the same time by Gottlieb Daimler. Both inventors sub-
sequently founded their own company, Daimler AG, which still persists with headquarters
in Germany and its well-known brand for passenger vehicles, Mercedes-Benz.73 Back then,
automobiles were primarily handcrafted in workshops with a high amount of manual labour.
As was outlined in previous chapters, the 2nd Industrial Revolution was just in its forming
stage and set to unleash its full potential. Companies had to optimise their production pro-
cesses if they were to expand their market share and lower operating expenses to earn rea-
sonable profits. Any optimisations in manufacturing relied on the ingenuity of individual
people alone. In this respect, Henry Ford set the cornerstone of automotive mass production.
His Model T was perfectly suitable to be shifted from traditional craftsmanship in workshops
to industrialised production, given that it was highly standardised. What contributed to its
success was that Henry Ford focused on division of labour and the use of a conveyor-belt
production line to set his company apart from its competitors and reach a significant mile-
stone in industrial history.74

Since the market introduction of the Model T in 1908 and the mass production that com-
menced from 1913, Henry Ford could thus build approximately 15 million Model T cars
until 1927. Ford’s Model T was the car to be in production for the longest time until it was

73
Cf. Daimler 2011
74
Cf. Rae 2021

30
followed by Volkswagen’s Beetle in 1972. Ford’s manufacturing principles were subse-
quently adopted by car makers around the globe and automotive manufacturing supremacy
shifted to Europe and Japan in particular.75

While European car makers later managed to excel in research and development of automo-
tive technologies, it was the Japanese car maker Toyota that led the way when it came to
improving manufacturing processes and reaching manufacturing excellence, well before the
3rd Industrial Revolution led to the implementation of sophisticated computer systems and
software. Figure 7 illustrates the key milestones of Toyota, including historic aspects of both
the company and major world events from 1935 until 2017. As early as 1950 Toyota estab-
lished its so-called Toyota Production System, which made Toyota more effective and effi-
cient as manufacturing progressed. The Japanese car maker is thus the most prominent ex-
ample of automotive manufacturing excellence of the 20th century. Furthermore, it is worth
noting that Toyota had even excelled at research and development by introducing the first
hybrid car in 1997, the Toyota Prius, and the first series-production fuel cell car in 2014.76

Figure 7: Historic development of Toyota (Toyota, 2021)

75
Cf. Rae 2021
76
Cf. Toyota 2021

31
4.2.2 Characteristics of the Automotive Industry
The global automotive market is primarily dominated by a few OEMs that hold the biggest
share. Among the top five automotive players as of 2021 are Volkswagen, Toyota, Daimler,
Ford and General Motors (see figure 8).77 While European, Japanese and US car makers
have led the global market for decades, it is evident that OEMs from China have caught up
significantly due to advancements in both R&D and manufacturing excellence. They have
gained reputation in developing and producing cars of comparable quality in the low to mid
segment, especially in the category of battery-electric vehicles. Some of these OEMs have
already made inroads into European markets or plan to do so in the future, including BYD,
SAIC, Great Wall or Geely via its acquisition of Volvo. This development is also exempli-
fied by new Chinese OEMs, notably start-ups such as Aiways, Byton or Nio.78 Considering
these market developments and external threats, it is arguable that Industry 4.0 in combina-
tion with AI can be both an opportunity and a critical determinant of competitive advantage
for European automotive OEMs and suppliers from a long-term perspective. In effect, coun-
termeasures are highly necessary, given that the COVID-19 crisis that commenced in 2020
led to a sharp decrease in global vehicle sales from 74.9 million in 2019 to 63.8 million in
2020 and 71.4 million in 2021 respectively (see figure 9).79

Figure 8: Global automotive car makers by revenue in $ (Statista, 2021)

77
Cf. Statista 2021
78
Cf. Ewing 2021; Cf. Carey and Yilei 2021
79
Cf. Statista 2021

32
Figure 9: Global automotive sales figures (Statista, 2021)

Each OEM has positioned itself with a specific customer value focus considering the per-
ceived product benefits or price range, as is indicated in figure 10. While big industry players
such as Volkswagen include brands that address customer groups from mid-price to high-
price, from Seat to Porsche, OEMs such as Lamborghini or Rolls Royce solely offer cars at
high prices on the top end. Toyota positioned itself between the low-price and high-price
segment by adopting a hybrid positioning strategy. This necessitates efficient processes
along the whole vehicle development and production process in order to gain profits and
remain competitive.80

Figure 10: Automotive OEM positioning strategies (Diez and Rodriguez, 2010)

80
Cf. Diez and Rodriguez 2010, p. 2

33
Given that OEMs rely on suppliers for both vehicle development and manufacturing, the
management of complex supply chains is at the core of the automotive value creation process
and is considered a key focus area in the light of Industry 4.0 and AI. Supply chain manage-
ment has thus been dealing with “relations and integrated business processes across the sup-
ply chain that produces products, services and information that add value for the end cus-
tomer”.81 With Industry 4.0, companies should be enabled to better meet the needs of their
customers and excel at adapting to changing circumstances given today’s dynamic global
business environment.

In the case of the Japanese car maker Toyota, each car is built using more than 30,000 indi-
vidual parts.82 These are, to a great extent, developed and delivered by more than 200 global
suppliers.83 All the parts have to be specified in terms of their individual requirements and
go through procurement and meet the predetermined quality standards. In effect, the tradi-
tional automotive supply chain is based on a tier-x hierarchy, with the OEM on the top of
the supply chain. Tier 3 suppliers provide raw materials, intermediate goods or components.
Tier 2 suppliers develop and provide modules and components, whereas Tier 1 suppliers
develop and provide whole systems, such as powertrain systems. One significant exception
are considered Tier 0.5 suppliers, which operate as contract manufacturers and built whole
cars, such as OEMs do. The latter type of supplier is exemplified by the Austrian-based
company Magna Steyr, a contract manufacturer that produces cars for OEMs including
BMW, Jaguar, Mercedes-Benz or Toyota at its production site in Graz, Styria.84 Due to the
digital transformation, this traditional notion of the automotive supply chain is set to undergo
a marked change. The global management consulting firm Kearney postulated in an expert
report, that automotive supply chains will increasingly shift towards a network-oriented and
more collaborative and interconnected form of supply chain. This notion coincides with the
proposed network architecture of Industry 4.0. As new players, such as online and IT players
including Google or Apple, make inroads into the automotive market and digital technolo-
gies experience widespread adoption, the traditional view of the automotive supply chain is
deemed outdated.85 Rather, the supply chain will change from a pyramid and tier-x perspec-
tive to a wheel-like perspective including traditional players and new players, including IT
and online firms such as Apple or Google, as is indicated in figure 11.

81
Jespersen et al. 2005, p. 12
82
Cf. Toyota 2021
83
Cf. Carpart 2021
84
Cf. Magna 2021
85
Cf. Gaenzle et al. 2016

34
Figure 11: Shift of the automotive supply chain (Graenzle et al., 2016)

4.2.3 The Automotive Development Process

The automotive development process contains a diverse and complex set of disciplines and
phases that typically stretch over several years. The main process phases include definition,
concept, pre-development, series development and pre series and series production. All
phases comprise specific procedures, such as production process development as part of the
series development phase, which is where the manufacturing principles and technologies are
laid out and put together. Figure 12 provides an overview on the key components and phases
of automotive development up to the start of production.86

Figure 12: Automotive core processes (Brunner et al., 2017)

86
Cf. Brunner et al. 2017, p. 2f.

35
Once a vehicle has passed the vision and idea stage, the requirements and functions are fur-
ther defined and thoroughly evaluated. In order to support the development process, auto-
motive engineers use IT systems, such as 3D CAD software and simulation tools, that mark-
edly increase their quality of work and reduce the total development time. After several loops
and iterations all details are formed into a full vehicle with loads of data, taking account of
technological, economic and ecological aspects as part of the vehicle concept. The latter, in
particular, has become increasingly important during the past few years. OEMs have thus
integrated a life cycle perspective into the vehicle development phases in order to make their
vehicles more environmentally friendly from development, production, actual vehicle use,
up to recycling.87 This new eco perspective arguably demands a high amount of data to be
processed and evaluated, which could be fostered by the use of AI. For example, vehicles
could be designed in such a way, that they have a lowest possible impact on the environment,
choosing the appropriate mix of materials and technologies, whereas the manufacturing pro-
cess is optimised to take account of efficiency and future recycling demands. Different sce-
narios could then be generated and compared using AI-enabled IT systems. A similar ap-
proach was adopted by researchers Sundaravaradan et al. (2011), who adopted a data mining
approach for enhanced life cycle assessments.88

The Styrian automotive supplier Magna Steyr is credited for developing the passenger vehi-
cle Peugeot RCZ using virtual development methods only, without any physical prototypes
before the start of production. The car could thus be developed in less than two years, which
exemplifies the opportunities of sophisticated IT systems and the use of advanced simulation
tools.89 The latter, in particular, rely on AI to deliver suitable results and speed up the devel-
opment process. In their research paper on AI-enabled simulation for lightweight vehicle
development, Kohar et al. (2020) highlight that “while information from these simulations
can be manually extracted, the vast amount of data lends itself to AI techniques that can
extract knowledge faster and provide more useful interpretations that can be convenient for
designers and engineers.”90 While the Peugeot RCZ was discontinued due to questionable
market acceptance and low sales figures91, it was still a key milestone in vehicle development
and production in the light of the digital era. In this respect, it is worth noting, that in order

87
Cf. Brunner et al. 2017, p. 2ff.
88
Cf. Sundaravaradan et al. 2011
89
Cf. Wingett 2011
90
Kohar et al. 2020
91
Cf. Pollard 2016

36
to further optimise and facilitate the vehicle development process and reduce the time to
production, automotive players can implement a so-called simultaneous engineering ap-
proach, where aspects of manufacturing are taken into account in early development stages.
The Styrian-based automotive engineering firm AVL uses simultaneous engineering to
check the designed components for ease of manufacturing, cost efficiency and to evaluate
suitable technologies and suppliers.92

The above-mentioned examples of automotive players showed that they have consistently
implemented state-of-the-art information systems and digital technologies that seamlessly
integrate and optimize the automotive development and production processes. It is evident
that systems that have previously been highly fragmented are now integrated into digital
automotive ecosystems to harness the full potential of Industry 4.0 and big data, with con-
temporary AI technologies being a key enabler. Once a vehicle concept has successfully
passed the development stages and made ready for the start of production (SOP), the auto-
motive production process involves the manufacturing of the vehicle body parts through a
number of individual process steps, including stamping, welding, painting and plastic mold-
ing or painting up to the general vehicle assembly.93 Figure 13 illustrates the various stages
of Toyota’s production process.

Figure 13: Automotive manufacturing process (Toyota, 2021)

92
Cf. AVL 2021
93
Cf. Toyota 2021

37
The first industrialised production process of Henry Ford was based on standardisation of
assembly tasks as well as a well-planned material flow. His vision was to avoid any non-
value adding activities, which is still elusive in today’s Industry 4.0 environment. OEMs can
optimise the material flow by means of horizontal integration. They would thus develop and
manufacture their own component parts either directly or via separate entities. Yet the auto-
motive industry is characterised by a strong dependency of OEMs and their suppliers of raw
materials, components and systems, which deliver their products to the vehicle assembly
location. Each automotive factory consists of a great number of tools and machines for every
production step, such as stamping presses for body parts. The most complex production pro-
cess of automotive manufacturing is the vehicle assembly, where two lines are separated to
assemble the vehicle’s body and chassis.94 According to Porsche, a German OEM for pre-
mium cars, “car manufacturers around the world speak about a ‘marriage’ when they insert
the chassis, transmission and engine in the body on the production line.“95 After the marriage
the final assembly line focuses on the instalment of individual parts and systems, such as
interior components or a vehicle’s exhaust system. After a vehicle completed the whole as-
sembly process, it is tested and inspected. The automotive industry has been an active player
in adopting Industry 3.0 technologies with regard to automation and computer systems, as
well as technologies of Industry 4.0, including M2M, cyber-physical systems or IoT.

In effect, the continuous strive to cut costs, increase efficiency and productivity has contrib-
uted to the leading position of automotive OEMs and suppliers in the light of the digital
transformation. Nowadays, OEMs increasingly rely on a global supplier network to deliver
individual parts and whole pre-assembled systems, such as gearboxes. Other major compo-
nents, such as engines, are often built by OEMs themselves at separate factories. All parts
and systems, whether produced by the OEM or by suppliers, are delivered to the assembly
line by adopting a just-in-sequence sourcing strategy, at the right time, the right quantity and
quality as well as in the correct sequence depending on the customer vehicle orders. Any
problems that occur along the vehicle supply chain up to the final vehicle assembly or
throughout individual production steps can have adverse effects on individual OEMs or the
whole automotive industry. The former case is exemplified by the conflict between
Volkswagen and its supplier Prevent in 2016, which caused a production stop at six

94
Cf. Rae 2021
95
Porsche 2014

38
Volkswagen factories.96 A case with bigger impact is exemplified by the global chip shortage
during the 2021 corona crisis, which is set to cost the global automotive industry approxi-
mately 210 billion dollars in sales revenue.97

4.2.4 The Toyota Production System


The Toyota Production System (TPS) has become the synonym of lean manufacturing and
it is widely accepted as a best practice approach towards achieving manufacturing excellence
in the global automotive industry. It has therefore been adopted by many leading OEMs
worldwide. In effect, the way production processes are designed and optimised have always
made the difference between success and failure, as was exemplified by the approach of
Henry Ford at the beginning of the 20th century. Toyota was keen to refrain from merely
imitating Henry Ford’s approach, which focused on reducing product costs by means of mass
production and standardisation. This manufacturing approach, indeed, proved successful in
times of economic growth. However, especially in times of economic stagnation, which was
the consequence of the oil crisis of the 1970s, it was considered less profitable. The TPS was
developed by Toyota’s former head of production, Taiichi Ohno, who analysed Henry Ford’s
mass production system and reasoned that it is not suitable for high variance vehicle produc-
tion. The automotive production process thus had to shift from reducing product costs to
minimising operating expenses and increasing production efficiency while enabling the pro-
duction of smaller lot sizes.98

Irrespective of what approach is adopted, the result of each automotive production process
should be an automobile that meets customer demands and is produced as efficiently and
effectively as possible in order to make a contribution to a company’s bottom line and
achieve long-term competitive edge. In fact, Toyota had managed to succeed in effective
production process design and optimisation from their early years on. The production system
of Toyota was invented as a company philosophy with the underlying principle of simply
producing "what is needed, when it is needed, and in the amount needed."99 Toyota’s core
manufacturing strategy has been to reduce waste and work-in-process wherever possible. In
addition, they placed much emphasis on balancing the whole supply chain and material flow
with a low level of inventory. This lean manufacturing approach contributed to Toyota’s

96
Cf. Ruddick 2016
97
Cf. Wayland 2021
98
Cf. Monden 2012, p. 24; Cf. Ohno 1988, p. 124ff.
99
Alukal and Manos 2006, p. 164

39
success in automotive manufacturing and has also become widely known as “lean think-
ing”.100 At the core of Toyota’s production excellence are distinct concepts and methods. In
particular, the concepts of KAIZEN, JIDOKA, KANBAN, TPM, TQM and Poka Yoke are
among the most prominent TPS methods that had markedly been influenced by the philoso-
phies of Taiichi Ohno, as cited by Schwickert et al. (2011). A brief description of each
method is provided in table 5:101

Table 5: Toyota Production System core methods (Cf. Schwickert et al., 2011)

KAIZEN JIDOKA
KAIZEN constitutes the central philosophy for The JIDOKA principle is used to reduce the
continuous improvements of the whole produc- amount of potential failures by stopping the pro-
tion process to increase customer satisfaction duction process if any deviation from the stand-
and reduce any waste in all relevant production ard process is detected. Machines should be
areas. The internal customer perspective is enabled to automatically stop without any hu-
added to highlight potential interdependencies man action. This ensures that problems can be
of production steps and effects on product qual- sorted well before cars are produced that are
ity. below quality standards.
Just-in-time & KANBAN Total Quality Management (TQM)
The just-in-time concept ensures that all re- TQM is a quality system built into the corporate
quired parts are available at the assembly line strategy. It is based on a focus on customer re-
at the right time and right amount. It is based on quirements, quality standards as well as em-
customer orders to manufacture only what is ployee involvement in optimisation initiatives.
needed. KANBAN emphases a pull approach The TQM system is based on the assumption
including all production steps, ensuring that sta- that the level of quality is influenced by all criti-
tions only produce what is needed by the follow- cal stakeholders, such as shop floor and as-
ing station in order to avoid unnecessary inven- sembly, supporting departments and suppliers.
tory.
Total Productive Maintenance (TPM) Poka Yoke
The TPM philosophy strives to increase ma- Poka Yoke is a key method to avoid failures dur-
chine efficiency through reducing unplanned ing the production process, by implementing
stops of machines, caused by maintenance or specific measures or means that ensure work-
defects. A predefined and well-planned mainte- ers avoid mistakes in the first place. It includes
nance programme involving all workers is set up colour coding of materials or tools to ensure that
to ensure 100% machine availability. workers always put them where they belong.

100
Cf. Liker 2004, p. 15ff.
101
Cf. Schwickert et al. 2011, p. 27ff.

40
Taking into account the success of Toyota with the implementation of its TPS, it is arguable
that it primarily stemmed from a perfectly aligned set of concepts and methods that comple-
mented each other within a corporate culture that fostered mutual learning and collaboration.
The adoption of just a single method would, without doubt, not deliver the expected results
in any automotive factory. Also, it is evident that a process that delivers proper results today,
might fail under changing circumstances and new external factors. Henry Ford, for example,
was highly successful with his mass production approach at the beginning of the 20th century,
yet Toyota was able optimise what they learned and continuously improve their processes to
put innovation at the core of their corporate strategy. As a matter of fact, the key to manu-
facturing superiority and above-industry product quality is “performance excellence
throughout the total system – including design, production, distribution, service, and the in-
volvement of all categories of employees, customers, and suppliers in the quality initia-
tive.”102

There is no doubt that in today’s global automotive markets, both OEMs and suppliers need
to be capable of adapting to a dynamic and constantly changing external environment if they
are operate successfully and avoid any adverse effects. Peter Drucker, a renowned Austrian
management philosopher, “had continued to stress over the years that organisations that do
not innovate will not survive.”103 In effect, Toyota realised early on that the innovation po-
tential of a company depends on a number of correlating influencing factors. If Taiichi Ohno
just had copied Henry Ford’s manufacturing process without any amendments and improve-
ments, Toyota would not have been that successful. What made Toyota successful was the
adoption of key characteristics of an inherent innovation philosophy, such as a collective
and shared vision, leadership based on inspiring, the will to be genuinely innovative, a per-
fectly aligned organisational structure fostering key individuals and effective teamwork,
with a high degree of employee involvement in a creative climate.104 In order to put this
innovation philosophy into practice, Toyota placed much emphasis on fostering human in-
telligence and collective wisdom. In effect, the underlying assumption has always been that
“employees themselves more often than not, know what needs to be done to improve oper-
ation.”105

102
Natarajan 2000, p. 70
103
Swaim 2010, p. 78
104
Cf. Tidd et al. 2005, p. 469
105
Kanter 1985, p. 68

41
Contrary to the leading management philosophies of the 20th century, Toyota adopted a bot-
tom-up approach that allowed all employees, from the shop floor to top management, to
question the status quo and put forward their own ideas of how to improve the manufacturing
process and increase both quality and efficiency. This was enabled by the so-called Toyota
Creative Idea and Suggestion System (TCISS). The TCISS was implemented as early as
1951 and continuously developed further to ensure employee engagement and involvement.
In 2011, Toyota had gathered 40 million ideas of its employees, many of which were exclu-
sively honored with an annual award.106 In this respect, Peter Senge, proponent of the learn-
ing organisation, argued that “collectively, we can be more insightful, more intelligent than
we can possibly be individually.”107 However, it is worth pointing out that in the light of the
digital era with an ongoing development of AI, the notion of combining the collective wis-
dom of employees within a learning culture to drive innovation might be subject to change.
An individual and well-developed AI system could thus be more intelligent than humans
could be collectively and derive at solutions and innovations humans would not have envis-
aged. Furthermore, it is arguable that a computer system is easier to maintain than thousands
of individuals that should share the same vision and integrate themselves into an OEM’s or
supplier’s corporate culture.

While the Toyota Production System is primarily known for its continuous improvement,
lean thinking and a strong people focus, it is lesser-known that Taiichi Ohno was an early
proponent of machine intelligence and automation. In effect, the general-purpose machines
Toyota used in its early years of vehicle manufacturing required skilled operators to work
properly, which led to high labour costs and the risk of human error. Maintaining a high
degree of product quality and achieving Toyota’s goal of above-standard customer satisfac-
tion thus posed a huge challenge on the company. Since the 1950s Toyota has been enhanc-
ing its production process from a technological point of view, starting with an optimised
version of Henry Ford’s mass production approach to strike a balance between high-volume
and high-variance customer demand. At a rather early stage, Toyota implemented special-
ised and automated production equipment, and began to push the application of robots from
1985 onwards. In the 1990s, they actively engaged in process innovation through the use of
digital technologies, as is indicated in figure 14.

106
Cf. Clifford 2014
107
Senge 2006, p. 221

42
Figure 14: Toyota‘s production process milestones (Toyota, 2021)

In the early years of Toyota’s Production System, Taiichi Ohno postulated that machines
should be capable of operating independently of workers, by stopping production automati-
cally if any abnormality is detected. This was crucial for Toyota, given that the TQM concept
demanded that high quality standards are met constantly, without any interruption during the
production process. Per se, any broken or defective part would increase costs significantly
and diminish customer satisfaction. More than decades ago, Toyota implemented the con-
cept of autonomation, that significantly reduced the risk of failure by using machines “with
a human touch”, as Taiichi Ohno underpinned its basic principle. The concept was first ap-
plied to Toyota’s weaving machines that instantly stopped in case of material defects. If a
failure is detected, workers could act immediately and implement appropriate countermeas-
ures in order to keep the production line up and running. Furthermore, the process would
subsequently be optimised in order to prevent similar failures in the future. 108 This, without
doubt, was another noticeable technological breakthrough, given that weaving machines al-
ready played a key role in the first industrial revolution, as was mentioned in a previous
chapter. As Taiichi Ohno put it: “In a product like the automobile, safety must always be of
primary importance. Therefore, on any machine on any production line in any plant, distinc-
tions between normal and abnormal operations must be clear and countermeasures always
taken to prevent recurrence.”109 The major steps at Toyota are shown in figure 15. Today,
this process could be supported by an AI-enabled system that automatically detects problems
and root-causes based on historic and real-time data, with subsequent adaption of relevant
parameters without human intervention.

108
Cf. Ohno 1988, p. 29ff.
109
Ohno 1988, p. 31

43
Figure 15: Toyota's autonomation process (Toyota, 2021)

In a recent report, Toyota commented on their manufacturing practices postulating that “ma-
chines and robots do not think for themselves or evolve on their own. Rather, they evolve as
we transfer our skills and craftsmanship to them. In other words, craftsmanship is achieved
by learning the basic principles of manufacturing through manual work, then applying them
on the shop floor level to steadily make improvements.“110 Toyota’s view on the role of
human intelligence in manufacturing shows that AI has not yet evolved to such an extent,
that it could make machines learn on their own on the same level as humans would do. For
Toyota, it is still human input that is necessary to develop machines and robots that fulfil
predefined tasks as they were intended. In effect, it is assumed that any optimisation is car-
ried out by humans alone with no input of the machine itself.111 It is worth noting, however,
that Toyota’s notion does not necessarily coincide with contemporary achievements and ap-
proaches towards AI and Industry 4. Nevertheless, there is no doubt that the whole automo-
tive production process has become highly complex and interconnected.

In fact, all automotive core processes have been subject to change in the light of the digital
era. While the focus has previously been on an iterative process, from development, produc-
tion and sales up to after-sales support, the processes have converged and become highly
integrated due to the implementation of digital technologies.112 What set Toyota apart from
its competitors was its sheer focus on efficiency and the establishment of lean processes
throughout the whole supply chain. At the core of its lean manufacturing approach was a
culture that fostered mutual learning and innovation thinking. In effect, the prerequisite was
a well-trained workforce that adheres to the principles of Toyota’s production system. The

110
TOYOTA 2021
111
Cf. TOYOTA 2021
112
Cf. Wedeniwski 2016, p. 21

44
availability of AI-enabled systems that gather big data from the shopfloor level to the top
have caused a shift from a strong people focus towards the digital transformation. With the
advent of Industry 4.0, new technologies have had a marked impact on how all individual
process steps are integrated. AI further adds a key technology to handle the huge amount of
data generated by all systems throughout the value chain. Bill Gates, founder of Microsoft,
argued as cited by O’Reilly, that “the first rule of any technology used in a business is that
automation applied to an efficient operation will magnify the efficiency. The second is that
automation applied to an inefficient operation will magnify the inefficiency.”113 Considering
Gates’ statement, the principles of the Toyota Production System seem still applicable and
a critical determinant of success in the digital era. However, there is evidence that despite
Toyota’s supremacy in lean manufacturing, the company failed to keep the leading position
in the global automotive market. In fact, as of 2008 Toyota was considered the world’s most
innovative car maker.114 According to recent research conducted by the Center of Automo-
tive Management, Toyota ranks on 13th position as of 2021, with Volkswagen, Daimler and
Tesla leading the way, as is shown in figure 16. The study also revealed that Chinese OEMs
have caught up, including SAIC, Geely and BYD. 115

Figure 16: Automotive OEM innovation ranking 2021 (Bratzel, 2021)

113
O’Reilly 2021
114
Cf. Fast Company 2008
115
Cf. Bratzel 2021

45
4.2.5 Automotive Megatrends and Implications

The competitive factors and impacts of the external environment on the automotive industry
have markedly intensified on a global scale. According to analysists of the global consul-
tancy PwC, “the automotive sector faces an unprecedented change with regard to the far-
reaching effects it will have on the industry and its users“, driven by climate change, demo-
graphic change, urbanisation and technological change.116

Figure 17: Automotive industry trends and implications (Kuhnert et al., 2018)

Today’s macro-environment has become highly dynamic and subject to constant change
with respect to political, economic, social, technological, environmental and legal factors.
Key implications are depicted in figure 17. For example, the European Union demands a
significant reduction in emissions for light-duty vehicles, with the goal of achieving zero
emissions by 2050.117 These standards pose a significant challenge for OEMs to develop
alternative powertrains and enabled new players to make inroads into the global automotive
market, with Tesla having gained the position of the most innovative car maker and Chinese
manufacturers gaining ground. At the same time, the trend towards urbanisation and envi-
ronmental awareness affects customers’ needs, resulting in the development of new forms
of mobility, such as car sharing. Advancements in both software and hardware technologies
have led to the development of connected and autonomous cars, which will further disrupt
the whole automotive industry. There are significant implications for both OEMs and sup-
pliers to cope with all challenges and speed of change, that will require more flexibility,
agility and efficiency than before.118 It is questionable whether AI can be a critical determi-
nant of competitive advantage and the ability to cope with dynamic change.

116
Kuhnert et al. 2018, p. 3
117
Cf. European Commission 2021
118
Cf. Kuhnert et al. 2018, p. 6ff.

46
4.2.6 A System Dynamics Perspective

The wide variety of Industry 4.0 technologies as well as the complexity of contemporary
automotive manufacturing call for an enlightened approach towards the proper management
of the whole ecosystem. In this respect, the discipline of system dynamics has proven to be
highly suitable for highlighting the major aspects of Industry 4.0 in the context of automotive
manufacturing and the corresponding interrelationships of all technologies and components
involved. The US-based society of system dynamics defines the term as a “computer-aided
approach for strategy and policy design”, that aims to “to help people make better decisions
when confronted with complex, dynamic systems. The approach provides methods and tools
to model and analyze dynamic systems. Model results can be used to communicate essential
findings to help everyone understand the system’s behavior.”119

The model that is derived of the system dynamics approach can thus aid both OEMs and
suppliers to be capable of steering their business towards the right direction taking account
of both internal and external influencing factors, such as the effects of the digital transfor-
mation and the development of new ecosystems. Khakifirooz et al. (2018) developed a sys-
tem dynamics use case for the automotive industry, that acknowledges the whole value chain
focusing on interrelationship between the customer, the product and the OEM, up to the
vehicle insurance company. It exemplifies the future automotive ecosystem that is based on
the technological aspects of Industry 4.0 and big data. Systems can thus be monitored, con-
trolled and consequently optimised through the use of interconnected IT systems and AI
technologies from the shopfloor up to top-level management. This interconnectedness and
seamless integration of systems leads to unprecedented opportunities for all stakeholders of
the future automotive ecosystem. On the one hand, insurance services could be based on the
actual mileage of a car and the respective driver profile. On the other hand, OEMs could
gather data during development and production phase. As a result, data gathered during the
use phase of the car is transferred back to product and process development as well as man-
ufacturing. The major goal is the optimisation of both product and process quality in order
to achieve a high level of customer satisfaction while ensuring cost efficiency and competi-
tive advantage.120

119
System Dynamics Society 2021
120
Cf. Khakifirooz et al. 2018, p. 4f.

47
The above-mentioned aspects as well as potential interrelationships and dependencies of the
automotive use case are illustrated in figure 18 using a causal-loop diagram that depicts ma-
jor components of the automotive ecosystem and the corresponding interrelationships and
effects.

Figure 18: Ecosystem cause-loop diagram (Khakifirooz et al., 2018)

The components include cloud computing with a central data warehouse, smart processes
for automated data transfer and monitoring, smart products as connected cars, up to the smart
factory of the automotive OEM that is connected to the players of the automotive supply
chain. The system harnesses the potential of IoT and big data to learn, develop and grow
based on the data that is collected in real-time. The causal-loop diagram thus shows how all
components interact and highlights the effect that they exert on each other. Interfaces be-
tween every component can be defined by a direct positive (+) or negative (-) relationship.
For example, internet outages lead to a disruption of online services and a lack of data avail-
ability, which describe a so-called balancing-loop (B) that can lead to a cycle error and can
therefore not increase exponentially. In turn, big data is used as input for enabling smart
processes, which enhance physical workflows that are a critical determinant of cyber-phys-
ical systems. The combination of sensors and smart products constitutes a so-called reinforc-
ing loop (R), as more smart products and sensors lead to more data that is available to be
processed, which describes a positive relationship. While the development of a system

48
dynamics model requires expertise and experience to cope with complex problems, it can be
used to promote the understanding of the system behaviour.121

Per se, the underlying principle of the above-mentioned system dynamics model has signif-
icant implications for the use of AI technologies. Given the proposed high complexity of
future ecosystems and the sheer amount of data that is gathered in real-time throughout the
whole value chain, it is obvious that human intelligence alone lacks the capacity to monitor
and control the whole system as well as generate meaningful information and decisions. The
implementation of appropriate models and methods of AI could thus push the system to the
next level to make use of the data at hand and achieve the state of a real-time or live enter-
prise.

121
Cf. Khakifirooz et al. 2018, p. 5f.

49
4.3 The Paradigm of AI

4.3.1 History and Evolution of AI

The notion of AI had its roots in science fiction of the first half of the 20th century, when
creative authors envisaged machines, such as autonomous robots, that are able to behave like
humans, with the musical Wizard of Oz of 1939 as a prominent example. 122 When the first
computing devices were introduced in the 1940s during the Second World War, they were
capable of performing simple numerical tasks and any AI-related thinking was pure imagi-
nation and more concerned with the far future. Also, computers could only execute prede-
fined tasks, instead of storing and processing data. The costs for running them amounted to
approximately 200,000 dollars a month, which made them too expensive to be widely set up
and used in practice and business settings. It was Alan Turing, in particular, who is credited
to put forward early AI-thinking. He made a significant contribution to the AI evolution by
developing the well-known Turing-Test in 1950, a test that is still used today to determine a
machine’s or computer’s capability of human thinking.123 In Turing’s 1950 paper “Compu-
ting Machinery and Intelligence”, he called this test “imitation game”, which is meant to
engage humans in conversations with machines, and observe how long a machine can con-
tinue communicating without being discerned as such. Even though computing devices were
just in the early development stage, compared with technologies of Industry 4.0, Turing re-
ferred to the need to develop a “table of instructions”, a kind of computer programme, to be
used in so-called learning machines. Furthermore, Turing was an early proponent of chess
games as suitable use cases for developing and applying AI in practice. 124 In effect, the idea
of learning machines Turing referred to in his paper, resemble advanced computers and sys-
tems of today, that run AI-enabled applications and machine learning algorithms.

A proof-of-concept (PoC) of Turing’s ideas of AI was developed only a few years after his
paper was published, which became known as the Logic Theorist and was considered one of
the first AI applications for human-like problem-solving. The Logic Theorist was publicly
presented at the Dartmouth Summer Research Project on AI (DSRPAI) conference in 1956,
hosted by the then founding-fathers of AI, the researchers John McCarty and Marvin Min-
sky. The term AI, was officially coined at this conference and it marked the cornerstone of
global AI research and development. In effect, nine years after Turing’s often-cited paper
was published and only three years after the DSRPAI conference on AI, Marvin Minsky co-
founded the first AI laboratory at the Massachussets Institute of Technology (MIT), which
set off to actively contribute to the field of AI and become a frontrunner in research for the
decades to come.125

122
Cf. Rockwell 2017
123
Cf. Shani 2015
124
Cf. Turing 1950, p. 1ff.
125
Cf. Rockwell 2017

50
In the 1960s, progress had been made to develop further AI applications and demonstrate
the achievable level of capabilities of human-like machines, with practical examples includ-
ing MENACE, a machine able to play Tic-Tac-Toe; ELIZA, a language processing machine
able to imitate a psychotherapist and Shakey, a mobile robot that could combine reasoning
with physical actions. While significant financial investments were made by governments to
foster the development of AI, researchers increasingly faced technical problems and limita-
tions as to what they are capable of achieving and delivering in practice. In effect, the theo-
retical concepts of researchers could not live up to their high promises owing to a lack of
computing power and data storage, which had adverse effects on the development and hin-
dered the progress. Nevertheless, in 1970, Marvin Minsky is quoted to have said that “in
three to eight years we will have a machine with the general intelligence of an average human
being.”126 It came with no surprise that the initial optimism to build a working human-like
machine gradually diminished and, at the same time, reduced the continuing efforts that were
made. Despite the inherent technological barriers and limitations, the 25 years of AI research
and development after the initiating DSRPAI conference marked the foundation phase that
was necessary to spark further ideas and innovations.127 At the same time, the third industrial
revolution was just before the tipping point of the wide-spread adoption of sophisticated ICT
in business and manufacturing. This, without doubt, had a significant effect on the future
development of AI, given that the optimisation of computer technologies, microelectronics
and software, that intensified from the 1970s onwards, also paved the way for AI applica-
tions to become more capable and versatile. While the first applications of AI were confined
to predefined programmes that ran on machines to deliver certain results, such as playing
Tic-Tac-Toe, more modern approaches increasingly built on the principle of machine learn-
ing. Machines would rather be fed with principles to apply and examples to learn from, lead-
ing to distinct results that were not predetermined, programmed and conceivable.128

The phase between the 1970s and 1990s was considered to be symbolic and comprised only
a few noticeable breakthroughs in the field of AI. Activities were, to a great extent, confined
to systems that were meant to deliver knowledge as part of a computer programme that could
be run on personal computers. According to Delipetrev et al. (2020), such a system consisted
of a knowledge base including all data and information, as well as a so-called inference
engine, that was capable of data processing according to predefined principles. The systems
were based on symbolic programming languages, such as Lisp or Prolog. Noteworthy pro-
jects during the symbolic phase were “Hopfield net” of 1982, an early form of a neural net-
work to learn and process information as well as “ID3”, a decision tree algorithm, which is
considered a predecessor of the C4.5 machine learning algorithm. In 1997, almost forty years

126
Cf. Rockwell 2017
127
Cf. Delipetrev et al. 2020, p. 7f.
128
Cf. Wadhwa 2016

51
after his first renowned quote on the prospects of AI, Marvin Minsky is cited to have said
that “I'm still a realist: If we work really hard - and smart - we can have something like a
HAL [a fictional spaceship computer of a science fiction series] in between four and four
hundred years. I suppose if we're lucky, then, we can make it by 2001!”129

Similar to early AI optimism during the foundation phase, the knowledge systems could not
live up to their promises either, due to their inherent lack of performance and usefulness in
practice. Also, activities in the field of AI were still primarily confined to academia and
research, rather than industry until the 1990s. This arguably describes the learnings and ex-
periences of the decades of AI research and development that commenced after Alan Turing
developed his famous Turing-Test and set off the advent of AI thinking. The so-called AI
winters from the 1950s until the 1990s represented the setbacks of early AI development and
initiatives and significantly diminished the hopes on what R&D can achieve.130 As Or Shani,
CEO of the tech firm Algorithms, put it: “There will be growing pains as AI technology
evolves, but the positive effect it will have on society in terms of efficiency is immeasura-
ble.“131

During the 1990s the development of AI had gained new momentum, yet researchers re-
frained from high optimism and using the AI acronym for their respective research projects.
The pitfalls and fallbacks after high promises since the 1950s and the two AI winters were
still prevalent and edged into researchers’ minds. Therefore, projects and initiatives were
rather named knowledge-based systems, optimisation algorithms or computational intelli-
gence. Almost 50 years after Alan Turing suggested chess games as the proper use cases for
AI, IBM was able to develop a fully functional chess computer named Deep Blue, that was
capable of beating the then world-leading chess player Garri Kasparow in 1996. This re-
markable breakthrough exemplified the opportunities of AI-enabled computers and their
proven superiority over human beings depending on the respective use case. Deep Blue had
just learned all games and moves of chess, without sophisticated AI methods such as deep
learning and still succeeded. At the same time, many other AI applications were showcased
by researchers and developers, including handwriting recognition, image classification, up
to semi-autonomous driving. In parallel to the development of AI applications, major con-
tribution on the hardware and software side came from tech companies such as IBM, Mi-
crosoft or Apple, that pushed the development of sophisticated and high performant com-
puters and applications. In addition to increased performance, computers were drastically
reduced in size and were available at much lower costs. Hence, the adoption of ICT started
spreading across the globe and through various industries. However, apart from the ongoing

129
Delipetrev et al. 2020, p. 9
130
Cf. Delipetrev et al. 2020, p. 9f.
131
Shani 2015

52
development of ever more powerful hardware and software, it is arguable that the internet
has made a significant contribution to AI development and adoption, as it did for Industry
4.0, IoT and the concept of decentralised cyber-physical production systems. In particular,
in the context of AI the internet made information and knowledge publicly available and
easily accessible. Furthermore, it enabled the sharing of knowledge and exchange of infor-
mation across boundaries in an instant. Informal organisations, working groups and online
communities have been formed to collaborate on the development of AI applications and
technologies.132

As Delipetrev et al. (2020) point out, “the AI evolution is accelerated by the openness of
code, frameworks, datasets, scientific publications, and overall knowledge sharing”133, all
primarily facilitated by the emergence and wide-spread adoption of the internet and the
world wide web. Part of the software and applications that are available over the internet are
so-called open-source software. Which means that there is a dedicated online community
with software engineers and developers who contribute to the ongoing development, opti-
misation and continuous updates. This joint effort approach on a global scale has led to better
results in a shorter period of time. Also, it is worth noting that open-source software or ap-
plications are usually available at no cost, which reduces the barrier of entry and allows even
small players to jump on the bandwagon. As a result, the AI development facilitated by the
internet and open-source working groups in particular, experienced an increasing interest
and made major steps forward since the 2000s.134

According to Xue-bo et al. (2018), the time period since the mid-1990s is also referred to as
the learning period with a significant increase in big data adoption and a strong focus on the
development of machine learning models. During this period, many technological break-
throughs could be achieved, including deep learning, neural networks and visual identifica-
tion.135 Figure 19 illustrates the major development stages of AI since the early years in the
1950s until the learning period of the 2000s.

132
Cf. Delipetrev et al. 2020, p. 11ff.
133
Delipetrev et al. 2020, p. 12
134
Cf. Delipetrev et al. 2020, p. 11ff.
135
Cf. Xue-bo et al. 2018

53
Figure 19: Development history of AI (Xue-bo et al., 2018)

Michael Copeland of Nvidia, a well-established global company in the chip sector, catego-
rises the history of AI into three distinct stages according to the respective technological
focus and advancement as is shown in figure 20. In particular, the period from the 1950s
until the 1980s is marked by the increasing interest in the topic of AI in general, with first
attempts in developing practical applications as mentioned before. Furthermore, the period
between the 1980s and the early 2000s pushed the development and applications of machine
learning and brought about real success stories, contrasting the pitfalls of the two AI winters
decades before. In fact, similar to the underlying causes of the industrial revolutions since
the 18th century, there are a number of reinforcing factors that occurred in parallel and ena-
bled AI to became a true gamechanger. The 2010s onwards caused a global AI boom due to
significant breakthroughs in deep learning, largely facilitated by the internet, cloud compu-
ting and the availability of high performance hardware and software.136 Also, nowadays
many AI algorithms and deep learning frameworks are easily accessible and available as
open-source software and applications, such as TensorFlow, Theano or PyTorch, to name a
few. Education in the field of AI is also facilitated by the internet, for example through online
courses, webinars or videos on platforms such as YouTube.137

136
Cf. Copeland 2016
137
Cf. Delipetrev et al. 2020, p. 15

54
Figure 20: AI history and technological focus (Copeland, 2016)

Throughout the decades of AI development, its successful application and potential has pri-
marily depended on the computing performance and memory available. While the roots of
some algorithms used today date back to the early days of AI development between the
1950s and 1980s, computing power has only now reached a level to live up to the AI prom-
ises of the past. In fact, AI is primarily based on math problems that run through a trial-and-
error phase, also known as brute force, to foster learning from mistakes. A computer system
thus needs to be powerful enough to deal with millions of such math problems per second.
The faster it is, the faster and more accurate the results are delivered. Even today’s
smartphones provide the performance of previous supercomputers at a fraction of the price
and size. Computer chips developed by Intel in 2018, for example, were capable of compu-
ting more than ten trillion math problems per second. One way to improve the computing
performance was to tailor chip technology to the demands of AI. A key technology to deliver
this flexibility is the so-called Field Programmable Gate Array (FPGA), which enables the
reprogramming of chips as needed.138 In 2020, the British start-up Graphcore unveiled the
then world’s most powerful and complex processor called Colossus MK2, which comprised
a total of 59.4 billion transistors, more than Nvidia’s A100 chip with a total of 54 billion
transistors. The Colossus MK2 processor was specifically designed for AI applications and
is therefore termed by Graphcore as Intelligence Processing Unit (IPU).139 This distinguishes
IPUs from the previously known processor acronyms Central Processing Units (CPU) or
Graphic Processing Units (GPU). According to a recent study of McKinsey, a big three
global consultancy, the latter have predominantly been used for system training methods,
whereas CPUs have been used for inference. By 2025 the count of so-called application-
specific integrated circuits, hardware tailored to AI needs and demands, will increase signif-
icantly for both datacenter and edge architecture.140 The substantial breakthrough of Graph-
core’s IPU becomes evident by considering the number of transistors of Intel’s first

138
Cf. INTEL AI 2018
139
Cf. Alsop 2020; Cf. Graphcore 2021
140
Cf. Batra et al. 2018, p. 10

55
processors in the 1970s, which comprised merely between 1,000 and 5,000 transistors. De-
spite the great efforts made to significantly improve computing power, as of 2016, it was
argued that the pace of hardware development in the chip sector had slowed down, come to
stagnation and that Moore’s law became less applicable. The rate of performance optimisa-
tion of computers debatably cannot keep up with the rapid development during the past few
decades.141 In a recent report, researchers Chen et al. (2020) coincide with the notion of
limitations to Moore’s law and postulate that collective and ongoing research is needed to
further improve the development of high-performing computer chips, considering the avail-
ability of new materials and structures as well as atomic level deposition.142 By contrast,
research conducted by OpenAI in 2019, a San Francisco based AI research lab, cited by Hao
(2019), revealed that Moore’s Law held true during the period of 1959 until 2012. Since
then, there is evidence that computing power had doubled every 3.4 months, rather than
every two years.143 Figure 21 shows the exponential growth of computing relative to the
brain capacity of animals and humans, which has not been reached by AI.

Figure 21: Exponential growth of computing (Tekinerdogan, 2017)

Fifty years after Marvin Minsky’s famous quote on the development of AI, Elon Musk, fa-
mous tech entrepreneur and CEO of car maker Tesla, argued in 2020 that AI will overtake
humans by 2025.144 It is questionable whether this can be achieved and proven in practice.

141
Cf. Simonite 2016
142
Cf. Chen et al. 2020
143
Cf. Hao 2019
144
Cf. Cuthbertson 2020

56
4.3.2 Definitions of AI

Similar to the definitions of Industry 4.0, the definitions of AI vary based on the respective
focus and have changed over the years. In his 1950 paper “Computing Machinery and Intel-
ligence”, Alan Turing envisaged AI systems as so-called digital computers, that “may be
explained by saying that these machines are intended to carry out any operations which could
be done by a human computer.” These digital computers would consist of three parts: infor-
mation storage, (2) executive unit and (3) control, resembling the AI systems that are in use
today.145 According to Russel and Norvig (2016), AI definitions evolved around the notions
of thinking and acting, with a distinction to be made between human behaviour and rational
behaviour respectively. The latter presupposes a certain degree and level of intelligence to
weigh different options against each other and derive at an appropriate result and solution.146
In this respect, according to Oxford Learner’s Dictionary, the adjective “rational” is defined
by the ability to “to think clearly and make decisions based on reason rather than emo-
tions.”147 It arguably is the absence of emotions, which contributes to rational decision mak-
ing and is considered a prerequisite for the unaffected application of intelligence. Hence,
there is no doubt that machines would rather excel at this ability, compared with human
beings who tend to be influenced by emotions to a great extent. Russel and Norvig (2016)
provide a comparison of definitions of the period between the 1970s and the 2000s, covering
the knowledge and learning stages, as is shown in table 6:148

Table 6: Definitions of AI (Russel and Norvig, 2016)

Thinking Humanly Thinking Rationally

“The exciting new effort to make computers “The study of mental faculties through the use
think … machines with minds, in the full and lit- of computational models”
eral sense” (Haugeland, 1985) (Chamiak and McDermott, 1985)

“[The automation of] activities that we associ- “The study of the computations that make it
ate with human thinking, activities such as de- possible to perceive, reason, and act.”
cision-making, problem-solving, learning …” (Winston, 1992)
(Bellman, 1978)
Acting Humanly Acting Rationally

“The art of creating machines that perform “Computational Intelligence is the study of the
functions that require intelligence when per- design of intelligent agents.”
formed by people.” (Kurzweil, 1990) (Poole et al., 1998)

“The study of how to make computers do “AI … is concerned with intelligent behaviour in
things at which, at the moment, people are bet- artefacts.” (Nilsson, 1998)
ter.” (Ritch and Knight, 1991)

145
Turing 1950, p. 4
146
Cf. AI HLEG 2019, p. 31
147
Cf. Russel and Norvig 2016, p.2; Oxford Learner’s Dictionaries 2021
148
Russel and Norvig 2016, p. 2

57
While the definitions mentioned above place much emphasis on the difference between an
acting and thinking view on AI, they neglect a holistic perspective on the key prerequisites,
causes and effects of AI on a larger scale. A more enhanced and modern approach towards
defining AI was adopted by the OECD, a global policy forum concentrating on economic
and social development, who define AI as “a machine-based system that can, for a given set
of human-defined objectives, make predictions, recommendations, or decisions influencing
real or virtual environments. AI systems are designed to operate with varying levels of au-
tonomy.”149 Per se, the OECD states that there are certain input variables predefined by hu-
mans to have an effect on any given environment. Furthermore, they point out that there is
a clear goal for developing and implementing AI, by establishing a system capable of oper-
ating autonomously as efficient and effective without human influence.

A more detailed and contemporary definition is provided by the AI High-Level Expert Group
(AI HLEG), which was installed by the European Commission to actively engage in discus-
sion, consultation and initiatives around the topic of AI. In a joint effort the AI HLEG pro-
vided a holistic definition, by postulating that AI “refers to systems that display intelligent
behaviour by analysing their environment and taking actions – with some degree of auton-
omy – to achieve specific goals. AI-based systems can be purely software-based, acting in
the virtual world (e.g. voice assistants, image analysis software, search engines, speech and
face recognition systems) or AI can be embedded in hardware devices (e.g. advanced robots,
autonomous cars, drones or Internet of Things applications).”150

The AI HLEG further elaborates on its definition and provides further insights into the ra-
tional functionality of AI systems. In particular, the systems achieve rationality by percep-
tion of their environment through data, which could be data generated by sensors, and pro-
cess these data to derive at meaningful information. The information is then used to engage
in specific actions that have a direct influence on the respective environment of the system.
If the system is just active or proactive depends on the type of system. Hence, a distinction
needs to be made between rational AI systems and learning rational AI systems. The former
are considered as straightforward AI systems that merely influence their environment, yet
are not capable of adapting to changing circumstances in order to continuously improve their
actions, whereas the latter have a learning capability which allows them to learn through
experience and achieve better results. An AI system can thus be defined as a system capable
of perceiving, reasoning, taking decisions, performing actions, and learning. In an Industry
4.0 context, an AI system could be a CPS in the form of a robot as part of a bigger CPPS.
Such robots would be defined as “AI in action in the physical world” or “embodied AI”.151

149
OECD 2019, p. 7
150
AI HLEG 2019, p. 1
151
Cf. AI HLEG 2019, p. 1ff.

58
4.3.3 Classifications of AI

AI can be classified by determining the level of thinking and acting rationally and humanly.
The classification thus gives an idea of the capabilities and advancement of the respective
AI application and technology. While the early approaches of AI applications from the 1950s
to the 2000s were primarily based on predefined programmes that led to specific actions and
results, today’s AI applications are more sophisticated and complex. The AI HLEG made a
mere distinction between rational and learning rational AI systems in their definition of AI.
In this respect, according to researchers Hassani et al. (2020), the level of AI can also be
distinguished by taking account of four different AI stages: (1) Reactive Machines, (2) Lim-
ited Memory Machines, (3) Theory of Mind, (4) Self-aware. Reactive machines resemble AI
applications that are capable of responding to stimuli based on their predefined programme
without any learning functions, as is exemplified by IBM’s Deep Blue chess computer. Lim-
ited-memory machines, on the other hand, have the ability to learn by taking account of
collected and stored data in order to derive at more informed decisions and actions, which is
exemplified by today’s autonomous cars. Theory of Mind describes the next level of AI,
which emphasises more human-like features, such as individually adapting to a given envi-
ronment. An AI system classified as “Theory of Mind” would thus be able to individually
adapt to changing circumstances or even emotions. This classification is still in its theoretical
phase and requires extensive research. The most sophisticated AI stage is the achievement
of AI self-awareness. If an AI system is fully self-aware, it can think and act rationally, just
like a human being. In effect, it is the phase where the AI system can contribute to its own
optimisation and development. Achieving AI autonomy in an automotive CPPS would ar-
guably require at least the level of “Theory of Mind”. An overview of the four stages and
the respective characteristics is presented in table 7:152

Table 7: 4-stage AI classification (Cf. Hassani et al., 2020)

1. Reactive Machines 2. Limited Memory Machines


Characteristics: Characteristics:
• Automatic response to input • Data collection and storage
• Predefined programme • Learning through experience
• No learning capability • Optimised results and actions
• No data collection and storage • State-of-the-art technology
3. Theory of Mind 4. Self-aware
Characteristics: Characteristics:
• Automation and learning capability • Proactive thinking and reasoning
• Active thinking and reasoning • Fully self-aware of own actions
• Response to environment effects • Human-like behaviour
• Near human-like behaviour • Self-optimisation capability

152
Cf. Hassani et al. 2020, p. 3f.

59
Another way of classifying AI distinguishes between three different stages: (1) Artificial
Narrow Intelligence (ANI), Artificial General Intelligence (AGI) and Artificial Super Intel-
ligence (ASI). Artificial Narrow Intelligence forms the basic level of AI capability, which
has commenced in the 1950’s and was driven by early AI proponents and researchers in-
cluding Alan Turing or Marvin Minsky. ANI systems resemble the state-of-the-art and refer
to the notion of reactive and limited memory machines, which follow a clear predefined
programme and do what they are intended to do. In this respect, given the lack of Theory of
Mind and self-awareness, ANI is also referred to as “weak” AI. The further development of
AI should be facilitated by innovative computer technologies, faster processors and memory,
as well as new methods and approaches. The ongoing enhancement of AI systems is thus
expected to lead to the stage of AGI. This level would mark the paramount shift from “weak”
AI to so-called “strong” AI, which is capable of adopting more human-like behaviour. An
AGI system, in this sense, adopts the Theory of Mind by learning, perceiving and under-
standing what is happening in its environment. This system can therefore reach a certain
degree of human intelligence. The highest achievable level of AI capability is considered to
be ASI, which surpasses human intelligence exponentially and is fully self-aware.153 ASI
arguably poses ethical questions and a threat to mankind, for famous theoretical physicist
Stephen Hawking is attributed to have said: “I fear that AI may replace humans altogether.
If people design computer viruses, someone will design AI that replicates itself. This will be
a new form of life that will outperform humans.“154 While the early developments of ANI
took decades and experienced a rather linear growth, the developments of today with
attempts to achieve AGI level performance are said to spark an exponential growth,
as depicted in figure 22.155

Figure 22: 3-stage AI classification (Jylkäs, 2020)

153
Cf. Miailhe and Hodes 2017, p. 4f.
154
Medeiros 2017
155
Cf. Jylkäs 2020

60
4.3.4 Contemporary Approaches Towards AI

AI initially was a distinctive research area separated from others in the middle of the 20th
century, which is exemplified by the foundation the AI laboratory at the MIT by researcher
Marvin Minsky and others. Compared with the past, nowadays AI is an integral part of com-
puter science, given the inherent relationship between computer technologies, both hardware
and software, and AI technologies. The MIT, as of 2021, combined both areas into its so-
called Computer Science and AI Laboratory (CSAIL). The major goal is to cover all relevant
aspects of AI development in a great variety of applications, from reasoning, perception, up
to behavioural approaches. The aspects that are dealt with include computer graphics and
vision, machine learning and robotics. Research combines both theory and practice, consid-
ering the development of AI systems in terms of computer technologies and programming
languages.156 In effect, AI has become an essential core element of computer science, not
only in research, but also in education. Many of the highly-regarded universities have estab-
lished their own AI facilities and departments to support education in computer science and
AI, such as the so-called AI centers at ETH Zürich in Switzerland or RWTH Aachen in
Germany.157 This further drives the ongoing development and initiatives to discover new AI-
related opportunities and further enhance the technologies that are known and used today.

While AI is attributed to the field of computer science, its approaches and techniques have
markedly been influenced by a number of fields and science branches outside of computer
science, including philosophy, logic and mathematics, psychology and cognitive science as
well as biology and neuroscience or evolution. For example, the principles of neuroscience
are taken into account to develop so-called neural networks that enable AI systems to adopt
human capabilities. Cognitive science, on the other hand, provides valuable insights for AI
developers on the functionality and information processing power of human brains. By con-
trast, the fields of logic and mathematics are at the core of AI development to implement
algorithms and models that enable AI systems to achieve operational efficiency and effec-
tiveness, just like humans. The sophisticated AI systems that are in use today are therefore
the product of a joint effort and collaboration of multiple disciplines in research and science.
Considering the above-mentioned aspects, it is evident that the field of computer science has
adopted a broad focus across various knowledge areas to drive the development of sophisti-
cated software, with a dedicated focus on AI systems and applications. This could lead to an
ever-increasing enhancement of AI approaches and techniques to finally derive at self-aware
AI systems that are capable of outpacing the cognitive abilities of humans.158

156
Cf. MIT 2021
157
Cf. ETHzürich 2021; Cf. RWTH Aachen 2021
158
Cf. Bullinaria 2005, p. 3ff.

61
As a result, over the decades of AI development, a great variety of approaches and techniques
have emerged that also found their way into various industries and industrial use cases re-
spectively. Similar to the definitions of AI, there is a proven lack of consensus regarding the
approaches towards AI and how to categorise them appropriately. In effect, AI experts, sci-
entists and researchers have just come to their own categorisation and conclusion of what
specific techniques constitute AI. For example, Van Duin and Bakhshi (2017) of Deloitte, a
global consultancy, refer to the major AI techniques as means to handle big data, learning
based on specific patterns, reasoning to draw conclusions as well as problem solving. Thus,
the approaches to achieve these AI challenges notably are machine learning and cognitive
analytics. While machine learning should add new knowledge to historic data using distinct
algorithms, cognitive analytics creates meaningful information out of unstructured data.159
It is arguable that this perspective oversimplifies the sheer complexity of AI and the tech-
niques that are available and in use today. Similarly, in their recent paper the researchers Hu
et al. (2019) refer to machine learning and deep learning as the major AI branches.160 On the
contrary, the experts of the AI HLEG refer to AI techniques as either based on learning or
reasoning. The former comprises machine learning with its sub-branches and areas of
knowledge including deep learning, neural networks and decision trees as noticeable exam-
ples. Learning specifically deals with perception problems and tasks, including the percep-
tion of language, vision and behaviour. Furthermore, the AI HLEG acknowledges three dis-
tinct approaches of machine learning: (1) supervised learning, (2) unsupervised learning and
(3) reinforcement learning. As opposed to learning, reasoning primarily deals with decision
making through planning, searching and optimising. It is achieved through gaining
knowledge out of a huge amount of data and deriving at appropriate solutions and actions.

The Platform Industry 4.0 of the German Federal Ministry for Economic Affairs and Energy
argues that today’s AI techniques are mainly enabled by rational or behaviour-oriented mod-
els. They refer to behavioural techniques as the appropriate approach for language pro-
cessing as well as industrial applications including major business tasks such as increasing
cost-efficiency or process optimisation. It also includes perception of vision and behaviour,
which is highly relevant in the field of manufacturing and robotics. Machine learning is said
to be the most important approach to achieve the above-mentioned AI challenges using ra-
tional or behaviour-oriented models. The Industry 4.0 platform panel also acknowledges the
distinction of machine learning approaches between supervised, unsupervised and reinforced
learning. It further highlights the “more specific (partial) applications, [which] are regression
and classification, structure recognition and structure prediction, data generation (sampling),
and autonomous action.”161

159
Cf. Van Duin and Bakhshi 2017
160
Cf. Hu et al. 2019
161
Plattform 4.0, 2019, p. 7f.

62
There is no doubt that the AI approaches presented often refer to machine learning and its
distinct methods that have made significant leaps forward during the past decade. On the
contrary, researcher John Bullinaria of the School of Computer Science at Birmingham Uni-
versity summarises the common techniques based on their underlying principle, including
(1) representation of knowledge, (2) learning to build knowledge, (3) rules to guide expert
systems and (4) search to derive at problem solutions.162 Another and more holistic overview
of AI approaches is provided by the AI panel of Stanford University, a collective of highly
regarded researchers and scientists, as is summarised in table 8.163 The use of specific ap-
proaches largely depends on the underlying problem that is to be solved by the AI system,
such as quality inspection or process optimisation in automotive manufacturing.

Table 8: AI Approach Overview by Stanford University AI Panel (Cf. Stone et al. 2016)

Stanford University AI Panel – Approach Overview

Machine Learning Deep Learning


Algorithms are implemented to deal with huge A sub-class of machine learning that taps into
amounts of data and foster machine learning advanced perception.
to gain new knowledge.

Reinforcement Learning Robotics


The advancement from mere pattern recogni- The shift from predefined PLC programmes to
tion to so-called “experience-driven sequential integrated AI-enabled computer vision and
decision-making”. machine perception.

Computer Vision Natural Language Processing


The capability of perceiving the environment From automatic speech recognition up to au-
and objects just like humans, mainly driven by tomated language translation. Alan Turing’s
deep learning. imitation game is challenged.

Collaborative Systems Crowd Sourcing and Human Computation


Enabling mutual collaboration of autonomous Tap into human knowledge to augment com-
systems that interact intuitively and goal-ori- puter systems. An approach to foster problem-
ented, as efficient and effective as possible. solving skills.

Algorithmic Game Theory and Computational Internet of Things


Choice Connecting every single object and “thing” to
Adding social and economic aspects to AI to collect huge data sets and process them to de-
deal with ethical challenges. rive and information.

Neuromorphic computing
Implement principles of biological neural networks to increase the performance of hardware
and systems.

162
Cf. Bullinaria 2005, p. 14
163
Cf. Stone et al. 2016, p. 9

63
4.3.5 General AI Methods and Intelligent Agents
The AI thinkers Stuart Russell and Peter Norvig (2021) postulate that so-called intelligent
agents are the key component and driver of AI technologies. They refer to such agents as
anything that is capable of perceiving and acting within certain boundaries, whereas the way
agents act is dependent on a specific agent function. Each agent is built upon a predetermined
architecture and program that have an effect on its behaviour. When it comes to automotive
manufacturing, the architecture would enable an autonomous robot to gather perceptions
based on the data that is gathered by a number of sensors, which in turn is computed by the
program to derive at proper actions, such as movements made through the robot’s actuators.
This principle drives the development if AI technologies and is considered as the underlying
principle of AI methods that are in use today.164 In this respect, Russell and Norvig (2021)
distinguish between different types of AI methods based on the capability and purpose of
the intelligent agent, which are further described and exemplified in table 9:165

Table 9: General AI methods (Cf. Russel and Norvig, 2021)

General AI methods

Problem-solving by searching Knowledge, reasoning and planning


Search agents are developed based on a prob- Some approaches of AI rely on knowledge-
lem statement with an initial state, set of actions based agents, which perform reasoning and de-
and a transition model. A search tree is built to cision-making based on facts. The knowledge
find proper solutions considering different base stores sentences about the agent’s envi-
search approaches. The method is applicable ronment, which are used by means of an infer-
for automotive assembly sequencing to limit the ence mechanism to take proper actions.
amount of manual labour or in-car navigation.

Uncertain knowledge and reasoning Machine learning


The probability theory has proven to be suitable Contemporary AI is often based on agents that
for reasoning in uncertain conditions, where an learn from experience. Their performance thus
agent encounters a complex world that is diffi- depends on prior knowledge and the data that
cult to ascertain. The agent’s decisions are not is available to the agent. Through data the
based on definite reasoning and real truth, ra- agent builds a model to derive at a certain as-
ther it is the agent’s “beliefs and desires” that sumption and prediction about the environment.
lead to an action considered most suitable. The method is commonly used for image recog-
nition.

Communicating, perceiving and acting


The more practical AI methods that are used in an industrial context as a means for enabling
communication, perception and action include natural language processing, autonomous robots
and the field of advanced computer vision.

164
Cf. Russell and Norvig 2021, p. 54ff.
165
Cf. Russell and Norvig 2021, p. 81ff.

64
4.3.6 Machine Learning Methods
Machine learning, first developed in 1959 by US-born scientist Arthur Samuel, is considered
a key component of contemporary AI and is widely used for industrial applications as was
mentioned above, such as automotive manufacturing. The aim of machine learning is to en-
able an AI system doing more than it was programmed to do, hence to learn from experience
and to gain new knowledge. It is based on an algorithm that is fed with and processes huge
amounts of data. The distinguishing power of machine learning algorithms is that they are
capable of making generalisations based on training data.166 In particular, according to Awad
et al. (2015), “the generalization attribute of ML allows the system to perform well on unseen
data instances by accurately predicting the future data.”167 Table 10 provides an overview of
common machine learning methods as cited by Russell and Norvig (2021):168

Table 10: Machine learning methods (Cf. Russel and Norvig, 2021)

Machine learning methods

Supervised learning Unsupervised learning


Agents in supervised learning are fed with data As opposed to supervised learning, the unsu-
input that is to be matched with a distinct output, pervised approach is based on learning without
such as different sets of images and properly direct learning in terms of true and false feed-
allocated image labels. The pairs formed as part back. The simplest form constitutes the agent’s
of the learning function are used to predict new clustering of images based on perceived simi-
input-output pairs. larities.

Deep learning Reinforcement learning


This approach overcomes the limitations of lin- The agent is enabled by a feedback and reward
ear input-output approaches. It enables the loop, which lets the agent know if actions were
agent to deal with multi-layer dimensions and successful and desirable or led to failure. It thus
high numbers of input variables to compute can draw conclusions on what actions contrib-
complex learning problems. Neural networks uted to the success in order to gain more re-
are at the core of deep learning. wards.

Knowledge-based learning Statistical learning


The combination of prior knowledge with new The use of probability theories enables agents
experiences should allow agents to improve to conduct calculations of averages or more
their learning capabilities. Knowledge-based in- complex Bayesian networks. The latter are
ductive learning is an example of a learning ap- based on probabilistic inference, where obser-
proach that seeks to explain observations with vations made by the agent have an effect on
prior knowledge much faster than others. both prior distributions and the corresponding
hypotheses.

166
Cf. Awad et al. 2015
167
Awad et al. 2015
168
Cf. Russell and Norvig 2021, p. 669ff.

65
Figures 23 and 24 exemplify the process flow of supervised learning and reinforcement
learning respectively. The former is based on raw data with feature vectors that represent
explanatory variables as well as data labels. The algorithm is fed with a training dataset and
a validation dataset for input-output testing and optimisation, with the goal to properly pre-
dict expected labels. The reinforcement learning flow exemplifies the feedback and reward
loop to let the agent know if actions were correct or not as a means to foster learning.169

Figure 23: High-level flow of supervised learning (adapted from Awad et al., 2015)

Figure 24: High-level flow of reinforcement learning (adapted from Awad et al., 2015)

169
Cf. Awad et al. 2015

66
4.3.7 Selection and Implementation of AI Models and Methods
The implementation of machine learning or any other AI method depends on a clear problem
statement that guides the model selection and development of AI systems. Moreover, it is
essential to evaluate further boundary conditions, including requirements in terms of accu-
racy, performance or interpretability.170 This holds true for any AI approach that is consid-
ered for industrial applications. In automotive terms, the problem statement could include
the need to automatically and accurately detect faulty parts or those that need repair in a
production line. In effect, a number of AI methods and models have been developed that
serve a distinct purpose and can be applied to best solve an OEM’s or supplier’s manufac-
turing-related problem at hand. While the AI method is the underlying technique based on
the major AI purpose, such as learning, reasoning or discovering, as described in a previous
chapter, an AI model digs deeper and constitutes the “software program [or algorithm] that
has been trained on a set of data to perform specific tasks like recognizing certain pat-
terns.”171

The above-mentioned automotive manufacturing problem and the need to detect component
defects can be solved through learning methods, since the system should learn the difference
between good and defect parts and apply this knowledge accordingly. This would require a
model or algorithm, that best suits this distinctive AI task and enables the system to deliver
reliable results that can be used in an industrial context. According to Ed Biddle (2021), an
IBM Watson Data Scientist, “modelling is an iterative process where a number of experi-
ments are conducted to find the optimal model, parameter, and feature selection.”172 The
modelling step can thus be seen as a trial-and-error process, rather than a one fits all ap-
proach. Russell and Norvig (2021) highlight the need for optimisation by postulating the
necessary steps of AI system development, including “choosing a model class, training your
model with the training data, tuning any hyperparameters of the class with the validation
data, debugging the process, and final evaluating the model on the test data.” 173 Depending
on the complexity and core requirements of AI tasks and problems, it might be necessary to
use different AI models and adopt the so-called “ensemble modelling“ approach in order to
benefit from advantages and outweigh limitations of a single model approach. 174

It is worth noting that apart from selecting the appropriate AI models to solve the predefined
problem, the successful implementation of any AI model largely depends on the datasets that
are provided as well as a thorough understanding of the requirements. For the above-men-
tioned example in automotive manufacturing, the data that is required for the machine

170
Cf. Taulli 2021
171
Socha 2021
172
Biddle 2021
173
Cf. Russell and Norvig 2021, p. 669ff.
174
Cf. Taulli 2021

67
learning agent to work, could be gathered by optical sensors fitted to a robot, stored in a data
warehouse and evaluated for its applicability to be used for developing the AI model. Given
that data are the lifeblood of any AI technology, it is thus of the utmost importance to ensure
a high level of data quality and availability considering the respective requirements. Accord-
ing to Valeria Sadovykh of PwC Labs, as cited by Taulli (2021), “on average, the data prep-
aration process takes 2X or in some cases 3X longer than just the design of the machine
learning algorithm.” It is suggested by Dan Simion of Capgemini, as cited by Taulli (2021),
to pre-evaluate the data at hand and conduct exploratory analyses to gain a better under-
standing, identify potential data deficiencies and optimise the data as needed.175

Russell and Norvig (2021) postulate that any machine learning system requires proper
maintenance and monitoring to deliver accurate results on a continuous basis. This is specif-
ically important for automotive manufacturing, given that there is a high demand for high
product and process quality with zero defects and failures. If a system is meant to replace
human workers, as Toyota pushed it with its autonomation approach and machines with a
human touch, it needs to be highly reliable and capable of improving based on any errors
that might occur, in order to avoid any adverse effects. Hence, in terms of interpretability of
any given data, it is questionable whether the machine learning system correctly matches
input and output. Furthermore, when it comes to explainability, it is crucial to understand
why a certain output was generated according to the respective input that was provided by
the data. A learning algorithm should be developed and monitored to such an extent, that
these factors are taken into account to ensure stability and accuracy.176

The global consultancy PwC stressed in a recent report that AI development should be seen
in the context of an iterative lifecycle that is leveraged through feedback loops and a process
called “value stewardship” that focusses on ongoing monitoring, evaluation and optimisa-
tion. For AI to work, solve the manufacturing problem at hand and achieve its predetermined
business goals, it thus “requires oversight mechanisms to monitor how it performs over time
and maintenance procedures to update the models to help adapt to unforeseen changes in its
environment.”177 Hence, similar to the Toyota Production System, the application of AI de-
mands an organisational system and process that caters for continuous improvement, encom-
passing all relevant business processes and digital technologies.

175
Taulli 2021
176
Cf. Russell and Norvig 2021, p. 669ff.
177
PwC 2020

68
Figures 25 and 27 illustrate various AI models, (1) categorised according to the purpose of
learning, reasoning and discovering178 and (2) categorised by type of AI method exemplified
by the use in optical networks. 179

Figure 25: AI model map for learning, discovering and reasoning (Contreras et al., 2018)

Figure 26: AI model map for enabling optical networks (Mata et al., 2018)

178
Contreras and Vehi 2018, p. 8
179
Mata et al. 2018, p. 45

69
4.3.8 AI Use Cases in the Automotive Industry
Automotive players have been among early adopters when it came to new ways of operating
to improve both efficiency and effectiveness and enhance the products they make. In recent
years, AI has increasingly been implemented for a number of use cases, from autonomous
driving to supply chain management and vehicle manufacturing. An overview of how AI
can be used in practice with automotive industry examples is shown in table 11:180

Table 11: Automotive AI use cases (Cf. Kozlowski and Wisniewski, 2021)

Automotive OEM AI Use Case

An AI-enabled supply chain monitoring system is used to cope


with high complexity and enable real-time information. In addi-
Audi tion, Audi facilitates quality control by enhanced computer vi-
sion to identify defects in parts to ensure conformance to cus-
tomer requirements.

The latest BMW 3 series models are equipped with a digital


personal assistant system that enables automatic adaption of
BMW vehicle functions. For example, if a driver is feeling tired, the
system can brighten the interior lighting, adapt temperature
control or selection relaxing music.

The latest models of Ford are equipped with connected vehicle


technology that allows gathering real-time data of vehicles in
Ford the use phase, including driving profiles. This data is shared
with insurers and can also be used as valuable feedback for
engineering and manufacturing.

Robots are used to assist workers during the manufacturing


process to provide support and avoid both failures and injuries.
Hyundai
The robots are capable of sensing human work tasks and ad-
just any movements accordingly.

The acoustic testing system “Sounce” is set up with deep


Porsche learning methods for detecting any quality deviances, as used
in endurance tests to provide support for engineers.

The Tesla Autopilot system provides automatic lane keeping


and object detection by using various sensors and advanced
Tesla
AI technologies. In addition, computer vision is used to detect
a driver’s condition to avoid any accidents caused by fatigue.

180
Cf. Kozlowski and Wisniewski 2021

70
4.3.9 AI in the Context of Automotive Manufacturing
The major goals of manufacturing have always been increasing operational effectiveness
and efficiency, ensuring conformance to customer requirements and continuous improve-
ment, while ensuring organisational success and competitive edge from a long-term perspec-
tive. This was exemplified by the development of the automotive industry since Henry Ford
introduced his mass production system and the approach of the Japanese car maker Toyota
by adopting the Toyota Production System. Industry 4.0, as was outlined in a previous chap-
ter, is regarded as the vital revolutionary step for automotive manufacturers to achieve these
goals better than before. This holds true for European companies that face fierce competition
from competitors in emerging markets, such as Asia. In most of the early Industry 4.0 defi-
nitions and descriptions of the respective components and technologies that are presented in
this thesis, AI was not explicitly stated as the driving force to achieve true Industry 4.0 man-
ufacturing and operational excellence. This view has significantly changed in the past few
years as AI development and applications have gained substantial momentum and experi-
enced widespread adoption on a global scale. Per se, in Josef Schumpeter’s terms, the wide-
spread adoption of sophisticated AI systems would arguably mark another innovation cycle
and the sixth wave of innovations, as AI is increasingly being implemented by both OEM
and suppliers and set to reach the next development level towards AGI. The German Industry
4.0 platform stresses that “from an industrial point of view, AI technologies are to be under-
stood as methods and procedures that enable technical systems to perceive their environ-
ments, process what they have perceived, solve problems independently, find new kinds of
solutions, make decisions, and especially to learn from experience in order to be better able
to solve and handle tasks.“181 A major benefit of AI in automotive manufacturing is thus
seen in harvesting real-time data through the deployment of so-called machine learning op-
erations (MLOps), in order to deal with the complexities of today’s dynamic environment
and continuously adapt to change to foster organisational learning and development from a
long-term perspective.182

It is worth pointing out that the theoretical aspects and opportunities of AI in an industrial
context are in parallel confronted with practical challenges. The Industry 4.0 platform panel
thus points out that approximately one third of value adding activities will be supported and
enhanced by AI and affect almost all functional departments. However, as of today, the im-
plementation of AI in manufacturing is still seen in its early trial-and-error stages, primarily
owing to high costs, inherent complexity and the need for organisational change and process
amendment. As a consequence, big industry players would rather focus on AI-enabled ro-
botics and resource management, while SMEs are more prone to use AI to enhance the areas
of knowledge and quality management as well as supply chain operations. Above all, the

181
Plattform Industrie 4.0 2019, p. 6
182
Cf. Gupta et al. 2021, p. 2

71
greatest benefit of AI in manufacturing is considered to be the achievement of a level of
operational excellence that could not be reached by collective human intelligence alone.
While it is expected that AI will enable automotive manufacturers to increasingly optimise
task-intensive and repetitive processes, which would reduce much manual work, AI is also
regarded as a key driver to create business opportunities at the same time.183 In effect, the
opportunities for AI in automotive manufacturing are manifold. Demlehner et al. (2021)
conducted a Delphi study and involved industry experts who identified the following use
cases:184

• AGV routing
• Autonomous optimization of robot trajectories
• Bin picking
• Cut-out waste reduction
• Factory energy monitoring / management
• Human-robot integration / collaborative robotics
• Maintenance knowledge management system
• Matrix production / cellular manufacturing
• Non-visual quality control
• Object labelling / tracking
• Paint bath com- position control
• Predictive maintenance
• Production forecasting / planning
• Reduction of robot energy consumption
• Sequencing optimization
• Staff assignment
• Visual quality control

Gupta et al. (2021) summarise the challenges of the automotive industry in the context of AI
as follows:
“The automotive industry is transitioning from a vehicle-centric view to a state of
deep, sustained customer-centric through Digital Transformation & Modernization –
which requires them to establish a new vision, new culture, and unique expertise.
Auto manufacturers are rethinking their organization structure, products, and pro-
cesses. It is no longer simply automation but the inclusion of AI in all the processes
to continually innovate, increase productivity, reduce costs, and optimize the supply
chain.”185

183
Cf. Plattform Industrie 4.0 2019, p. 6
184
Demlehner et al. 2021, p. 6
185
Gupta et al. 2021, p. 5

72
4.3.9.1 Five Levels of AI-enabled Manufacturing
Similar to the SAE levels of autonomous driving, the application of AI in the context of
Industry 4.0 is seen as an incremental implementation and enhancement of different types of
AI technologies at different levels of autonomy. The definition of such levels allows a clear
distinction between system boundaries and the degree of AI readiness in manufacturing. Per
se, the ultimate goal of implementing AI is to reach a high level of manufacturing autonomy
where production processes are fully monitored, controlled, managed and optimised by AI
without any human influence and interference. The majority of today’s AI use cases in au-
tomotive manufacturing have reached level two and three respectively, as is exemplified by
the use cases presented before. A full AI system integration beyond level three has thus not
been achieved yet. Reaching level 5 would mean the achievement of at least “Theory of
Mind” AI capability, with both autonomy and learning as well as active thinking and re-
sponding to the environment. Table 12 shows the five levels of autonomy as cited by the
platform Industrie 4.0:186

Table 12: Levels of AI-enabled manufacturing (Plattform Industrie 4.0, 2019)

Five Levels of AI-enabled manufacturing


No autonomy
Level 0 All tasks and operations are completed, monitored and
controlled by humans.

Assistance with respect to select functions


Overall responsibility belongs to humans, with minor
Level 1
AI support for decision making.

Partial autonomy
AI is implemented for smaller and specific manufactur-
Level 2
ing areas and processes, while humans remain fully
responsible.
Delimited autonomy
AI is implemented for bigger areas and processes and
Level 3 provides warning systems and solutions to problems
that are evaluated by humans.

System functions autonomously and adaptively


AI takes over whole parts of manufacturing and pro-
Level 4 cesses that are under close human supervision.

Autonomous operations in all areas


Fully AI-enabled manufacturing operations within pre-
Level 5 defined boundaries, without any human interference
needed.

186
Plattform Industrie 4.0, 2019, p. 12ff.

73
5 Systematic Literature Review

5.1 Literature Identification and Evaluation

The process of literature identification and evaluation followed the proposed methodology
and is depicted in figure 27. Relevant literature was identified using the databases of IEEE
and ScienceDirect as primary sources, with ResearchGate used for manual identification of
further relevant literature. The identified literature was subsequently screened for relevance
and eligibility by applying the predefined criteria to derive at a literature rating. Finally, 84
papers were included for the quantitative analysis and 79 papers for the qualitative synthesis.

Figure 27: SLR PRISMA flowchart

Literature identified
Identification

through database search Additional literature


(n=344) identified through
IEEE (n=277) other source
ScienceDirect (n=67) ResearchGate (n=15)

Literature identified for screening


(n=344)
Screening

Literature after duplicate Literature not relevant


removal and screening to topic excluded
(n=91) (n=253)

Total number of full-text


Eligibility

Full-text literature after literature excluded


assessment for eligibility (n=7)
(n=84) Full-text not available (n=3)
Different focus (n=4)

Literature included in
quantitative analysis
(n=84)
Included

Literature included in Literature excluded due


qualitative synthesis to limited content
(n=79) (n=5)

74
Figure 29 shows the literature distribution of the three literature sources before the screening
phase, whereas figure 28 shows the results after the assessment for eligibility (84 papers).

Figure 29: Literature distribution before screening phase Figure 28: Literature distribution after screening phase

n=25 IEEE n=25 IEEE


ResearchGate ResearchGate
n=48 n=48
n=11 ScienceDirect n=11 ScienceDirect

For the literature identification, papers published before 2015 were excluded to ensure that
outdated information and methods are not considered for analysis and synthesis. Figure 30
and table 13 show the number of literature publications per year since 2015. It can be clearly
seen that the research interest in AI and automotive manufacturing has markedly increased,
with 3 publications published in 2015 (3.57 %) and 33 publications in 2021 (39.29 %).

Figure 30: Literature distribution per year

35 33

30

25

20
15 14
15
9
10
6
3 4
5

0
2015 2016 2017 2018 2019 2020 2021

Table 13: Literature distribution per year

Year of publication Number of papers Percentage


2021 33 39.29 %
2020 14 16.67 %
2019 15 17.86 %
2018 9 10.71 %
2017 4 4.76 %
2016 6 7.14 %
2015 3 3.57 %
Total 84 100.00 %

75
Figure 31 and table 14 show that the majority of included 84 papers for the quantitative
analysis are 44 conference papers (52.38 %) and 37 journal articles (44.05 %). Additionally,
2 chapters and 1 technical paper were selected. A total of 5 papers were excluded for the
qualitative synthesis due to lack of suitable content, deriving at 79 papers.
Figure 31: Types of literature

n=2 n=1

Conference paper
Article
n=37
n=44 Chapter
Technical Paper

Table 14: Types of literature

Type of literature Number of papers Percentage


Conference paper 44 52.38 %
Article 37 44.05 %
Chapter 2 2.38 %
Technical Paper 1 1.19 %
Total 84 100.00 %

All papers were assessed according to the predefined criteria with a maximum rating of 22
points to be achieved in order to guide the paper selection process. Most papers were con-
sidered adequate for quantitative analysis and qualitative synthesis, only 12 papers were ex-
cluded after the screening phase due to limited access to the full-text paper or lack of focus
and suitable content. The rating threshold to be achieved for qualitative synthesis was deter-
mined at a score of 15 points. Figure 32 shows the literature rating distribution per year.

Figure 32: Literature rating distribution per year

100% 20
1 3
2 4 2
2 4 19
80%
2
2 4 18
60% 3
1 14
5 1 17
40% 1
1 3 4 16
3
20% 15
1 2 1 3 2 7
1
1 14
0% 1 1 2
2015 2016 2017 2018 2019 2020 2021 13

76
5.2 Research Results

5.2.1 General Results

The papers were categorised according to their relation to a specific manufacturing process
in order to provide an overview of the focus of AI deployment and distribution. Table 15
lists the AI-enabled processes in automotive manufacturing that were identified:

Table 15: AI-enabled processes in automotive manufacturing

Process Description
Assembly of individual automotive parts, components and systems
Assembly
in a dedicated production line to produce the final vehicle.
Methods used for monitoring, controlling and optimising operations
Business Intelligence
and manufacturing processes on the managerial level.
Machines and robots that properly execute production tasks accord-
Production
ing to production planning and customer requirements.
Planning, analysis, evaluation and improvement of the quality of au-
Quality
tomotive parts, components and vehicles.
Provision of required automotive parts, components and systems ac-
Supply Chain
cording to manufacturing requirements.

Table 16 and figure 33 provide an overview of the processes and the respective number of
papers. The quantitative analysis revealed that the majority of papers evaluated focussed on
the application of AI for production and quality related issues, representing 33 papers
(39.29 %) and 30 papers (35.71 %) respectively. Assembly ranks on third position, whereas
supply chain and business intelligence form the minority.

Table 16: Distribution of AI-enabled processes in automotive manufacturing

Process Number of papers Percentage


Production 33 39.29 %
Quality 30 35.71 %
Assembly 14 16.67 %
Supply Chain 5 5.95 %
Business Intelligence 2 2.38 %
Total 84 100.00 %

77
Figure 33: Distribution of AI-enabled processes in automotive manufacturing

n=14
n=5

n=2
Assembly
Business Intelligence
n=30 Production
Quality
Supply Chain
n=33

Figure 34 illustrates the number of publications from 2015 to 2021 and the distribution of
the identified manufacturing-related processes. It can be seen that both production and qual-
ity have been dealt with since 2015, yet the focus on assembly commenced and increased
from 2018 onwards. As was exemplified by the production process of Toyota, the assembly
of a full vehicle is highly complex with a thousand of individual parts that have to be assem-
bled properly. While an AI system has the potential to support the model development and
process of a full vehicle assembly, it requires substantial experience and knowledge to be
put in practice. The same holds true for issues related to supply chain and business intelli-
gence.

Figure 34: Number of annual publications and distribution of the processes

35

3
30

25 10

Supply Chain
20
Quality
Production
15
2 1 12 Business Intelligence
Assembly
10
1 9 8
3
5
4 3 8
1 2 4
1 4
2 2 1 2 1
0
2015 2016 2017 2018 2019 2020 2021

78
5.2.2 Current Models and Methods
This chapter gives an answer to research question 1:
What are the current models and methods of AI in automotive manufacturing?

The quantitative analysis revealed that machine learning methods rank on top of the most
used AI methods for the presented automotive use cases, as is shown in figure 35 and table
17. The top three methods included supervised learning (46.74 %), deep learning (22.83 %)
and unsupervised learning (11.96 %). Both federated learning and genetic programming
were considered more recent and advanced methods, as is deep learning. Other AI methods
and techniques, such as rule-based, case-based or knowledge-based reasoning, were used to
solve specific AI problems and formed the minority.

Figure 35: AI methods in automotive manufacturing

Supervised Learning
Deep Learning
Unsupervised Learning
Reinforcement Learning
Rule-based
Deep Reinforcement Learning
Knowledge-based reasoning
Search-based
Federated Learning
Case-based reasoning
Semi-supervised Learning
Genetic Programming

0 5 10 15 20 25 30 35 40 45 50

Table 17: AI methods in automotive manufacturing

AI Method Number of applications Percentage


Supervised Learning 43 46.74 %
Deep Learning 21 22.83 %
Unsupervised Learning 11 11.96 %
Reinforcement Learning 4 4.35 %
Rule-based 3 3.26 %
Deep Reinforcement Learning 3 3.26 %
Knowledge-based reasoning 2 2.17 %
Case-based reasoning 1 1.09 %
Genetic Programming 1 1.09 %
Federated Learning 1 1.09 %
Search-based 1 1.09 %
Semi-supervised Learning 1 1.09 %
Total 92 100.00 %

79
Figure 36 shows the applied AI method in relation to the respective manufacturing process.
This approach revealed, that deep learning is primarily implemented for solving quality re-
lated manufacturing problems. Case-based, rule-based, knowledge-based or search-based
methods were confined to production. Reinforcement learning was chosen as the proper
method to solve complex problems in assembly, production and supply chain. Federated
learning and genetic programming were considered as revolutionary methods for enhancing
the vehicle assembly process. Semi-supervised learning was used in production to combine
the benefits of both machine learning techniques.

Figure 36: AI method application per process

100% 1 3
90% 1
80% 4
70% 17
13 2 1
60%
50% 1 1 1 2 3 1 1
4 Supply Chain
40%
30% 17
4 2 Quality
20% 1
10% 2 3 Production
3 4
0%
Business Intelligence
Assembly

A total of 50 papers (59.52 %) presented a single-method approach with selection of an


adequate model and technique to solve the predefined problem based on the respective model
characteristics. A multi-method approach or “ensemble modelling” was presented in 34 pa-
pers (40.48 %), either by comparing the performance of different models for proper model
selection and subsequent system optimisation, or to combine multiple models to overcome
any trade-offs with the use of a single-method approach. For quality problems the multi-
method approach was most applied (17 papers). Table 18 shows the distribution of single
and multi-method AI.

Table 18: Distribution of single-/multi-method approach

Process Multi-method Single-method Total Total percentage


Assembly 3 11 14 16.67 %
Business Intelligence 1 1 2 2.38 %
Production 11 22 33 39.29 %
Quality 17 13 30 35.71 %
Supply Chain 2 3 5 5.95 %
Total 34 (40.48 %) 50 (59.52 %) 84 100.00 %

80
A total of 79 AI models and algorithms with 168 applications were identified in the literature.
The most applied models comprise neural networks (37), followed by support vector ma-
chine (14), random forest (9), decision tree (8) and multi-layer perceptron (6). Table 19 lists
all identified models and algorithms. Figure 37 shows the process distribution. The above-
mentioned models were used for different use cases. Both convolutional neural networks
and support vector machines were predominantly used for quality issues. Other techniques
that were specifically used for quality issues include but are not confined to autoencoder,
Bayesian network, principal component analysis, YOLO or XGBOOST-Arima.

Table 19: AI models and algorithms used in automotive manufacturing

Model / Algorithm Number of applications Percentage


Neural Network 17 10.12 %
Convolutional Neural Network 17 10.12 %
Support Vector Machine 14 8.33 %
Random Forest 9 5.36 %
Decision Tree 8 4.76 %
Multi-layer Perceptron 6 3.57 %
k-Nearest-Neighbor 5 2.98 %
Predictive Modeling 4 2.38 %
Bayesian Network 3 1.79 %
Autoencoder 3 1.79 %
Principal Component Analysis 2 1.19 %
Gaussian Mixture Modeling 2 1.19 %
Proximal Policy Optimization algorithm 2 1.19 %
Levenberg-Marquardt algorithm 2 1.19 %
k-Means 2 1.19 %
Markov Decision Process 2 1.19 %
Probabilistic Neural Network 2 1.19 %
Bagging 2 1.19 %
Q-learning 2 1.19 %
Expectation Maximization algorithm 2 1.19 %
Data Mining 2 1.19 %
Naive Bayes Classifier 2 1.19 %
YOLO 2 1.19 %
Deep Multi-Layer Perceptron 1 0.60 %
Structural Similarity Index algorithm 1 0.60 %
Resilient Back Propagation 1 0.60 %
Fully Connected Network 1 0.60 %
PoseNet 1 0.60 %
ARIMA 1 0.60 %
Recursive Convolutional Neural Network 1 0.60 %
Gaussian Naive Bayes 1 0.60 %
Search algorithm 1 0.60 %
Gaussian Process Regression 1 0.60 %
Vector Space Model 1 0.60 %

81
Generative Adversarial Network 1 0.60 %
Adaptive Boosting 1 0.60 %
Genetic algorithm 1 0.60 %
Detectron2 1 0.60 %
Gradient Boosted Tree 1 0.60 %
Regression algorithm 1 0.60 %
Iteration algorithm 1 0.60 %
Rule-based algorithm 1 0.60 %
Balanced Random Survival Forest 1 0.60 %
Single Shot Detection 1 0.60 %
ARIMAX 1 0.60 %
Two-step feature learning model 1 0.60 %
Bootstrap Forest 1 0.60 %
XGBoost 1 0.60 %
Linear Discriminant 1 0.60 %
Association Rules 1 0.60 %
Linear Regression 1 0.60 %
Deep Deterministic Policy Gradient 1 0.60 %
Border Tracing algorithm 1 0.60 %
Deep Neural Network 1 0.60 %
XGBoost-ARIMA 1 0.60 %
Recursive algorithm 1 0.60 %
Adaboost 1 0.60 %
Regression algorithm 1 0.60 %
Multi-agent system 1 0.60 %
Regression Neural Network 1 0.60 %
Multi-criteria decision-making 1 0.60 %
rPart 1 0.60 %
Causal model 1 0.60 %
Scaled Conjugate method 1 0.60 %
Multinominal Naive Bayes 1 0.60 %
Simple Logistic Regression 1 0.60 %
Classification algorithm 1 0.60 %
State–Action–Reward–State–Action 1 0.60 %
Association algorithm 1 0.60 %
Back Propagation Neural Network 1 0.60 %
N-step SARSA 1 0.60 %
Variational Autoencoder 1 0.60 %
OSER 1 0.60 %
VGG16 framework 1 0.60 %
Perception algorithm 1 0.60 %
Feature Pyramid Network 1 0.60 %
Petri nets 1 0.60 %
MaxEnt-IRL 1 0.60 %
MobileNet 1 0.60 %

82
Figure 37: AI model and algorithm application per process

Model / Algorithm Count


(1) Neural Network 17
(2) Convolutional Neural Network 17
(3) Support Vector Machine 14
(4) Random Forest 9
(5) Decision Tree 8
(6) Multi-layer Perceptron 6
(7) k-Nearest-Neighbor 5
(8) Predictive Modeling 4
(9) Bayesian Network 3
(10) Autoencoder 3
(11) Gaussian Mixture Modeling 2
(12) Naive Bayes Classifier 2
(13) Expectation Maximization algorithm 2
(14) k-Means 2
(15) Bagging (Bootstrap aggregating) 2
(16) Principal Component Analysis 2
(17) Proximal Policy Optimization algorithm 2
(18) Probabilistic Neural Network 2
(19) YOLO 2
(20) Q-learning 2

83
The findings of current models and methods of AI in automotive manufacturing together
with the corresponding AI purpose are shown in table 20. It can be seen what approaches
and techniques were applied to solve a specific problem in automotive manufacturing.

Table 20: Current models and methods of AI in automotive manufacturing

AI Purpose AI Methods & Models Reference

Quick and reliable detection Machine Learning with combined Principal Component Analysis
(Ciampaglia et al., 2021)
of defects (PCA) and One Class Support Vector Machine (OC-SVM)
Development of vehicle as- Federated learning with Support Vector Machine and smart
(Aronozhi et al., 2021)
sembly model contract integration
Machine learning model comparison including Multi-layer Per-
Pre-calculation of risks for
ceptron neural network, gradient boosted tree and Decision (Evangeliou et al., 2021)
reliable inbound logistics
Tree
Trajectory optimisation for Deep reinforcement learning with Markov decision process and
(Klarmann et al., 2021)
industrial robots proximal policy optimisation algorithm
Monitoring of cutting tool Machine learning with supervised approach and naive bayes al- (Carvalho & Bittencourt,
life span gorithm 2021)
Detection of errors during Multi-agent comparison using machine learning with Random
(Schuh et al., 2021)
assembly Forests, Bagging, rPart, Naive Bayes classifier
Selection of design param-
Supervised learning approach with Decision Tree and associa-
eters for automotive weld- (Gilabert & Arnaiz, 2021)
tion rule techniques
ing
Machine learning with linear regression, kNN, Random Forest,
Quality monitoring of bend-
Multi-Layer Perceptron and deep learning using Convolutional (Mayr et al., 2021)
ing processes
neural network
Adaption of production sys- Reinforcement learning approach with neural networks and use
(Overbeck et al., 2021)
tems of proximal policy optimization algorithm
AGV material handling and Reinforcement learning applied with Q-learning algorithm for
(Jeong et al., 2021)
route planning AGV navigation policy development
Sound detection of electri- Machine learning and multi-layer perceptron compared with
(Espinosa et al., 2021)
cal harness assembly deep learning and convolutional neural network
Predictive quality and main- Machine learning with data drift detection algorithm based on
(Zeiser et al., 2021)
tenance principal component analysis
Control for discrete event Reinforcement learning with state-action-reward-state-action al-
(Zielinski et al., 2021)
system in assembly gorithm
Prediction of machine Machine learning with adaptive ARIMA model for breakdown
(Mohan et al., 2021)
maintentance demand prediction
In-line evaluation of me-
Machine learning with analytical and numerical algorithms (Magro et al., 2021)
chanical properties
Machine learning approach comparing different methods/mod-
Prediction of material dela-
els including XGBoost-ARIMA, neural network, Support Vector (Cui et al., 2021)
mination
Machine
Multi-agent approach with statistical and machine learning
Forecasting of manufac- methods, models including ARIMAX (for launch stage), multi-
(Goncalves et al., 2021)
turing demand layer perception, support vector regression and Random For-
ests
Evaluation of welding pro- Deep learning with 2D and 3D convolutional neural network
(Zhou et al., 2021)
cess quality adapted for temporal and spatial IR intensity changes
Deep learning with Expectation Maximization Algorithm and
Quality inspection of com-
YOLOv3 for object detection through regression and Darknet-53 (Rahimi et al., 2021)
ponents
CNN for class probabilities
Detection of surface de- Machine learning with structural similarity index algorithm and
(Abagiu et al., 2021)
fects neural network
Process control in full-vehi- Multi-agent machine learning with algorithms for classification
(Queiroz et al., 2021)
cle assembly and prediction tasks
Optimisation of human-ro-
Deep learning with convolutional neural network (Chen et al., 2021)
bot collaboration

84
Deep learning with Bayesian network and Object Shape Error
Quality control of free form
Response model, multi-step training with closed-loop, transfer (Sinha et al., 2021)
surfaces
learning, continual learning approach
Inspection of structural ad- Deep learning with generative adversarial networks and
(Peres et al., 2021)
hesives YOLOv4 models
Correction of assembly ob- Deep reinforcement learning with deterministic policy gradient
(Sinha et al., 2021)
ject shape errors algorithm and neural network
Determination of appropri- Machine learning with gaussian process regression and neural
(Qian et al., 2021)
ate assembly parts network
Predictive analytics and Machine learning with classification, regression and association
(Kumari et al., 2021)
maintenance algorithms
Detection of presence/ab-
Deep learning with ResNet-50 convolutional neural network (El Hachem et al., 2021)
sence of objects
Machine learning with use of Detectron2 library for detection
Detection of non-conform-
and segmentation algorithms and feature pyramid network as (Rio-Torto et al., 2021)
ance parts in assembly
feature extractor for accuracy and speed
Root-cause problem solv-
Machine learning with bootstrap forest model with class-probe
ing in electronics wafer pro- (Berges et al., 2021)
and unit probe, data storage in Hadoop as open framework
duction
Single and multi-agent with reinforcement learning and Q-learn-
Modular production control
ing to handle complexity combined with multi-agent system ap- (Gankin et al., 2021)
with AGVs
proach for high robustness
Machine learning with Multi-layer Perceptron algorithm to be
Performance evaluation in
used to analyse data to reveal relations and influences of di- (Dobrota et al., 2020)
supply chains
verse variables
Multi-agent comparison using machine learning with neural net-
Detection of functionality
works, Support Vector Machine, k-Nearest-Neighbour, Decision (Wagner et al., 2020)
deviations
Tree
Prediction of energy effi-
Deep learning with deep Multi-layer Perception algorithm (Serin et al., 2020)
ciency and surface quality
Improvement of quality, reli-
Machine learning with prediction algorithms (Dacal-Nieto et al., 2020)
ability and maintenance
Detection of wafer sawing Multi-agent machine learning with XGBoost and Random For-
(Peres et al., 2020)
defects ests models
Machine learning with Bayesian network parameter learning ap-
Analysis of manufacturing
proach and maximum expectation algorithm combining expert (Zhang et al., 2020)
anomalies
knowledge and new input samples
Quality inspection of com- Deep learning with Convolutional neural network, LeNet-5 archi-
(Muresan et al., 2020)
ponents tecture and Border Tracing Algorithm
Automated inspection of Deep learning with Convolutional neural network and VGG16
(Nguyen et al., 2020)
crimp connections framework
Real-time error detection Deep learning with Convolutional neural network for error detec-
(Vater et al., 2020)
and correction tion and Decision Tree for automated rework
Machine learning with Autoencoder/variational Autoencoder for
Inspection of crimping pro-
anomaly detection, deep learning with Convolutional neural net- (Meiner et al., 2020)
cess
work for process diagnosis
Machine learning with Decision Tree and simple logistic regres-
Prediction of automotive (Salcedo-Hernandez et
sion approach, Area Under ROC Curve used as performance
paint film quality al., 2020)
metric
Part recognition and pose Deep learning with application of Convolutional neural network,
(Li et al., 2020)
estimation PoseNet and use of OpenSceneGraph to generate datasets
Machine learning approach with neural networks and Multi-layer
Optimisation of manufactur- (Patalas-Maliszewska et
Perceptron used for classification, Levenberg-marquardt and
ing decision-making al., 2020)
genetic algorithms used for training process
Prediction of material requi-
Data mining with vector space modelling (Widmer et al., 2018)
rements
Regression algorithms for training and creation of production
Planning of production sys-
configurations, neural networks to link training data with assem- (Hagemann et al., 2019)
tems
bly processes and stations to replicate knowledge of planners
Scheduling and motion Multi-criteria decision-making model used for balancing of work,
planning of robot assembly implementation of digital twin database for simulation/real world (Kousi et al., 2019)
lines data

85
Machine learning with balanced random survival forest model to (Bukkapatnam et al.,
Long-term fault prediction
deal with complex dynamic dependencies 2019)
Consistency of product Supervised machine learning algorithms used to identify pro-
(Park et al., 2019)
quality cess deviations and suggest corrections
Determination of machine
Unsupervised machine learning approach adopting K-means
health status and mainte- (Wescoat et al., 2019)
and Gaussian Mixture Modelling
nance demand
Selection of assembly
Machine learning with neural network (Hagemann et al., 2019)
equipment
Improvement of production
performance through plant Machine learning with recursive and iteration algorithm (Qian et al., 2019)
benchmarking
Prescriptive analytics in
Machine learning with predictive algorithms and models (Vater et al., 2019)
production
Detection of robot gripping
Machine learning with perception algorithm (Poss et al., 2019)
points
Detection of defects in Three-stage deep learning approach with neural networks and
(Ren et al., 2019)
components Autoencoder
Deep learning with two-step approach including (1) Convolu-
Part detection during as- tional neural network for key point prediction and (2) recursive
(Zhang et al., 2019)
sembly process Convolutional neural network for optimisation first prediction re-
sults
Unsupervised learning using Gaussian Mixture Model to cater
Detection of robot faults (Cheng et al., 2019)
for lack of labelled OK/NOK training data
Detection and classification Deep learning with single shot detection and MobileNet as Con-
(Mazzetto et al., 2019)
of assembly objects volutional neural network, implemented with TensorFlow API
Comparison of AI techniques including rule-based and search
Automation of body-in- (Hagemann & Stark,
algorithms, data mining, knowledge discovery and machine
white production design 2018)
learning
Optimsation of production Multi-agent approach with genetic programming and petri nets,
(Denno et al., 2018)
scheduling probabilistic neural nets and causal models
Replicating human sensory Supervised machine learning using deterministic Markov deci-
(Ruiz et al., 2018)
perception sion process for representing expert knowledge
Diagnostic and prediction of Machine learning using neural networks approach with Bagging,
(Luo et al., 2018)
machine health status Adaboost and Random Forests models
Image recognition for man- Deep learning with Convolutional neural network and specific
(Luckow et al., 2018)
ufacturing logstics architectures including Q-Gate, TensorFlow and YOLO
Prediction of screw-fas- Multiple AI models used to train classifiers, including SVM, NN,
(Matzka, 2018)
tening process quality DT, kNN
Teaching-learning-collaboration model for robot learning
Enabling advanced human- through human demonstration with application of MaxEnt-IRL
(Wang et al., 2018)
robot collaboration algorithm, development of collaboration algorithm based on op-
timal assembly policy
Detection and identification Machine learning with Support Vector Machine and neural net-
(Carino et al., 2018)
of faults work
Development of consistent
Data mining algorithms (Hofmann et al., 2017)
and reliable KPI forecasts
Multi-agent deep learning approach with two-step feature learn-
Prediction of component
ing model, back propagation neural network, Autoencoder, re- (Bai et al., 2017)
quality
gression neural network
Automation of robotic as- Machine learning with supervised learning approach and neural (Ortega-Aranda et al.,
sembly operations network 2017)
Detection of component de- Multiple machine learning and computer vision methods used,
(Mery et al., 2017)
fects in x-ray images best results with SVM and hand-engineered LBP descriptor
Deep learning with Convolutional neural network, Ha-
Quality inspection of com-
doop/Spark and GPU-based deep learning frameworks on- (Luckow et al., 2016)
ponents
premise and cloud
Multiple AI models with supervised/unsupervised learning in-
Prediction of failures in pro- cluding Principal Component Analysis, Support Vector Machine
(Zhang et al., 2016)
duction lines and Random Forests, two-stage approach with (1) clustering
and (2) supervised learning

86
Machine learning with neural network approach and use of Le-
Diagnosis of bearing faults (Patil et al., 2016)
venberg-Marquardt algorithm
Machine learning with one-class SVM with fault-free training
Detection of end-of-line
data, hyperparameter tuning of SVM regularization and kernel
faults in combustion en- (Leitner et al., 2016)
parameters, random artificial outliers generated for compensat-
gines
ing low availability of outliers
Recognition of human be- Deep learning with Convolutional neural network with 3D convo-
haviour in industrial workfl- lution raw data or visual content adapted with temporal infor- (Makantasis et al., 2016)
wos mation
Machine learning with comparison of Support Vector Machine,
Optimisation of robot pick-
Multinominal Naive Bayes, Gaussian Naive Bayes, Decision (Semeniuta et al., 2016)
ing and inspection
Tree, Adaptive Boosting, Random Forest, K-Nearest-Neighbors
Tool condition monitoring Machine learning probabilistic neural network implement for
(Huang et al., 2015)
and breakage detection real-time breakage detection
Knowledge and rule-based system with expert input used to de-
Predictability of formability
termine material formability, combination of VB.NET, AU- (Bhatt et al., 2015)
for sheet metal components
TOLISP and AUTOCAD applied
Multiple machine learning approach with Bayesian Network,
Categorisation of surface (Pastor-Lopez et al.,
Support Vector Machine, K-nearest-Neighbour, Random Forest
defects 2015)
and Best Crossing Line Profile features

87
5.2.3 Contemporary Applications and Best Practices
This chapter gives an answer to research question 2:
What are the contemporary applications and best practices of AI in automotive man-
ufacturing?

The results of the qualitative synthesis showed, that AI has already been successfully imple-
mented in automotive manufacturing for a variety of use cases in assembly, business intelli-
gence, production, quality and supply chain. Table 21 shows the applications and best prac-
tices according to the respective process category. While a great variety of AI models was
used to solve the specific AI problems, many papers presented highly positive results of AI
system deployment, with accuracy rates exceeding 90 percent. These results are a substantial
achievement and best practice considering the complex and dynamic characteristics of the
automotive industry. The papers thus confirmed the need for proper model selection, devel-
opment and optimisation according to the distinct requirements of each use case.

Table 21: AI use cases per process

AI applications in vehicle assembly


Control for discrete event system in assembly Enabling advanced human-robot collaboration
Correction of assembly object shape errors Image recognition during assembly
Detection and classification of assembly objects Optimsation of production scheduling
Detection of errors during assembly Part detection during assembly process
Detection of non-conformance parts in assembly Planning of intelligent automation systems for assembly
Determination of appropriate assembly parts Process control in full-vehicle assembly
Development of vehicle assembly model Selection of assembly equipment
AI applications in business intelligence
Development of consistent and reliable KI forecasts Optimisation of manufacturing decision-making
AI applications in production
Adaption of production systems Optimisation of robot picking and inspection
Automation of body-in-white production design Part recognition and pose estimation
Automation of robotic assembly operations Planning of production systems
Cooperative robots in metalworking Predictability of formability for sheet metal components
Detection of functionality deviations Prediction of machine maintentance demand
Detection of robot faults Prediction of material requirements
Detection of robot gripping points Predictive analytics and maintenance
Determination of machine health status and maintenance
Predictive quality and maintenance
demand
Diagnostic and prediction of machine health status Prescriptive analytics in production
Identification of flexibilities for lean manufacturing Quality monitoring of bending processes
Improvement of production performance through plant
Recognition of human behaviour in industrial workflwos
benchmarking
Improvement of quality, reliability and maintenance Replicating human sensory perception
Long-term fault prediction Root-cause problem solving in electronics wafer production
Modular production control with AGVs Scheduling and motion planning of robot assembly lines
Monitoring of cutting tool life span Selection of design parameters for automotive welding

88
Optimisation of human-robot collaboration
AI applications in quality
Analysis of manufacturing anomalies Inspection of crimping process
Automated inspection of crimp connections Inspection of structural adhesives
Categorisation of surface defects Prediction of automotive paint film quality
Consistency of product quality Prediction of component quality
Detection of defects in components Prediction of energy efficiency and surface quality
Detection of end-of-line faults in combustion engines Prediction of failures in production lines
Detection of presence/absence of objects Prediction of material delamination
Detection of surface defects Prediction of screw-fastening process quality
Detection of wafer sawing defects Quality control of free form surfaces
Diagnosis of bearing faults Quality inspection of components
Ensure flexibility and quality in production Real-time error detection and correction
Evaluation of welding process quality Sound detection of electrical harness assembly
In-line evaluation of mechanical properties
AI applications in supply chain
AGV material handling and route planning Performance evaluation in supply chains
Forecasting of manufacturing demand Pre-calculation of risks for reliable inbound logistics
Image recognition for manufacturing logstics

An overview and more detailed description of all applications in automotive manufacturing


that were identified, including best practices and key results, is provided in table 22. The
examples indicate that AI has the potential to replace manual tasks and accomplish more
complex and data-intense tasks humans could not easily cope with, if AI models are devel-
oped and implemented properly with the availability of adequate datasets. The high robust-
ness, reliability and accuracy that was achieved in many use cases makes AI suitable for
industrial deployment.

Table 22: Contemporary applications and best practices of AI

AI application & best practice Reference

Detection of defects and damages in automotive parts, exemplified by carbon fiber reinforced
(Ciampaglia et al., 2021)
suspension arms

AI-enabled Automotive Assembly Model as decision support system for automotive manufac-
turing, data input from energy usage, defects or costs, Trust Threshold Limit introduced as
(Aronozhi et al., 2021)
nominal range and maximum limit of processes to reduce losses, smart contract integration
for manufacturing policies
Material Planning Decision Support System for pre-calculation of external risks and influenc-
ing factors (weather, traffic, disasters) on inbound logistics reliability, input from Audi's internal (Evangeliou et al., 2021)
data and external public data, Multi-layer Perceptron neural network with highest accuracy
Data-based control of industrial robots at an Audi plant to automatically generate robot trajec-
tories overcoming manual efforts and lack of flexibility for dynamic robot adaption, AI system (Klarmann et al., 2021)
creates brown fields - engineers finalise the solution,
Monitoring of automotive CNC cutting tool life span to reduce process costs and increase
(Carvalho & Bittencourt,
productivity, tool and machine data from global tool setting variset (GTSv) for root-cause anal-
2021)
ysis and improvement, 10% life span increase achieved
Machine learning used for early detection of automotive assembly errors to reduce repair
costs, data integrated from assembly times of conveyor belt, 90+% accuracy with Random (Schuh et al., 2021)
Forest model

89
Extraction of information from process and quality data to evaluate parameter influence on
product performance including yield strength, tensile strength, hardness, gap and thickness, (Gilabert & Arnaiz, 2021)
application in automotive welding scenario with quality tests
Quality monitoring approach for hairpin bending for automotive traction motors with feature-
(Mayr et al., 2021)
based models and CNN, use of process data including torque curves, best results with CNN
Use of policy-based algorithm to establish a control logic and ideal work allocation for the pro-
duction system to increase productivity based on the chaku-chaku-principle, data from worker
(Overbeck et al., 2021)
movement in production line, validation through digital twin with Bosch Automotive, stable and
efficient process achieved
AI-enabled flexible and real-time routing of AGVs and forklifts for material handling within a
dynamic automotive manufacturing environment, data input of layout cell state and AGV initial
(Jeong et al., 2021)
position, optimised make span, distance, responsiveness and decreased energy consumption
to be achieved
Quality inspection by means of click-event sound for automotive electrical harness assembly
for engines to detect non-conformance connection between components, data input with (Espinosa et al., 2021)
25,000 click sounds of 25ms length at 22khz, MLP 97.66% and CNN 94.55% accuracy
Production management enhanced through AI-enabled data analytics to enhance decision
making for predictive quality and maintenance, process data enriched with meta info and pro- (Zeiser et al., 2021)
cess expertise,
AI-enabled automatic control for discrete event system in automotive assembly environment,
RL reward systems considers influence of practical tasks on cost, profit and performance to
(Zielinski et al., 2021)
derive at proper decisions, 9x improvement for profit reward compared with sole application of
supervisory control theory
Prediction of foundry machine breakdowns and maintenance demand with focus on zero
downtime, main parameter oil contamination level in machine setup, MTBF increase by 880%, (Mohan et al., 2021)
84% downtime reduction
Real-time monitoring of automotive steel rolling process to ensure uniformity of material prop-
erties, data input incl. supplier/material and sheet geometry, multiple-sensor data with roller
(Magro et al., 2021)
speed, force on constraints, roller position, brindles loads or elongation, model derives at opti-
mum process parameters for quality conformance and possible adaption during process
AI-enabled prediction of delamination in carbon fiber reinforced plastic automotive compo-
nents with multi-sensor data of force, torque, temperature and vibration and with hole exit im-
(Cui et al., 2021)
ages, forces values with highest accuracy, XGBoost-Arima model with 39% lower mean
square error compared with SVM/NN, industrial application confirmed
Multi-agent demand forecasting for assembly operations implemented with Bosch Automotive
(Goncalves et al., 2021)
Electronics, achieved more accurate results across all life-cycle stages
Evaluation of automotive spot-welding process with prediction of nugget shape and size for
quality inspection, data input with IR images recorded during process, model transfer for other (Zhou et al., 2021)
welding parameters and processes, 85-96% accuracy achieved
Quality monitoring and object detection through computer vision to complement manual part
(Rahimi et al., 2021)
inspection process, accuracy and speed achieved
Development of automated defect detection system for automotive manufacturing to for de-
tection of surface defects using engine block x-ray images, data fed from reference and test (Abagiu et al., 2021)
images
Process control system in VW T-Roc assembly line to detect process deviations and defects
enabling dynamic parameter adaption, effectiveness and efficiency increase confirmed, indus- (Queiroz et al., 2021)
trial-level robustness and scalability achieved
Optimisation of human-robot collaboration through AI-enabled robot assistance model for
(Chen et al., 2021)
real-time collaboration in assembly, robot assisted assembly with 98% accuracy
Estimation of dimensional/geometric defects in multi-station assembly systems in automotive
manufacturing caused by robot or end effector issues, 56% less training samples through (Sinha et al., 2021)
multi-step training, avoidance of process variations with integrated root-cause analysis
Inspection of structural adhesives in automotive manufacturing, e.g. hem flange bonding, to
avoid defects including adhesive discontinuities and blobs, synthetic data used to augment (Peres et al., 2021)
small real dataset for generalisation, method confirmed and verified
AI-enabled object shape error correction in multi-station assembly systems to reduce dimen-
sional/geometric shape errors, timely adaption of relevant process parameters, approach con- (Sinha et al., 2021)
firmed with benchmarks, 40% effectiveness increase achieved
Hybrid AI model to improve the automotive steering shaft assembly process by automatically
determining the ideal ball slider size to fit both tube and shaft, 91.32% accuracy and manual (Qian et al., 2021)
system improvement by 40% achieved
Predictive analytics and maintenance through multiple data from temperature, ultrasonic or
(Kumari et al., 2021)
vibration sensors to identify machine patterns and health status

90
AI-enabled quality control and visual inspection system through neural network to detect pres-
ence/absence of objects at Faurecia automotive plant, including joinings, welding, spatters, (El Hachem et al., 2021)
holes or parts, 99% accuracy achieved with ResNet-50
Development of support tool for workers in assembly lines for quality control at final inspection
stage, approach needs no retraining for new model versions, autonomous generation of da-
(Rio-Torto et al., 2021)
taset from annotated 3D models of automotive components, non-conformities to be shown on
mobile device, 65.7% accuracy achieved
AI-enabled root-cause problem solving for electronic wafers by linking observed high variabil-
ity with parameters, e.g. location of weaknesses and potential causes in wafer-manufacturing
(Berges et al., 2021)
process, different data sources used with correlations to constrain root causes, system results
coincided with expert opinion on root causes
Implementation of AI-enabled modular production control utilising AGVs between modular sta-
tions, system used to handle AGV scheduling and routing to increase efficiency and reduce (Gankin et al., 2021)
production costs, industrial scalability confirmed
Using the AI model and data mining approach to link performance KPIs with specific problems
(Dobrota et al., 2020)
in order to derive at a trend line that enables the identification of risks and relevant changes
Function verification of common-rail injectors during production process to minimise defect
parts, reduce costs and cope with complexity, features determined by experts and integrated
(Wagner et al., 2020)
with data from quality control database, conformity evaluation before final assembly using in-
line production data, kNN with good result saving 215k€/p.a.
Prediction of surface conditions and machine energy consumption for slot milling production
(Serin et al., 2020)
process, algorithm performance at 91% accuracy
AI-enabled data analytics system for production performance optimisation focussing on qual-
ity and maintenance processes and improvement of reliability, use case in automotive paint
(Dacal-Nieto et al., 2020)
shop with substantial time savings for analyses and preventive alerts, optimised energy man-
agement
Multi-stage quality control to identify OK/NOK cars in a Volkswagen assembly line with two
(Peres et al., 2020)
automated measurement stages
Analysis of quality deviances in automotive sunroof manufacturing to improve existing quality
control chart and analysis approach, data input with previous anomalies and faults with expert (Zhang et al., 2020)
knowledge, 85.7% accuracy achieved
correct/incorrect placement and missing parts, proposed method proved accurate with weak
(Muresan et al., 2020)
performance of SVM, DT and IPT
Final inspection of automotive electronics crimp connectors to detect different defect classes
automating the quality assessment process, dataset built from laboratory input, 99.87% accu- (Nguyen et al., 2020)
racy achieved
Real-time error detection in automotive manufacturing for hairpin welding of electric motor sta-
tor, combination of edge and cloud architecture, bird eye view of hairpins as input data, ena-
(Vater et al., 2020)
bling automated correction and rework with prescriptive analytics, 99.2% accuracy achieved
with CNN
Quality monitoring of crimping process in automotive wire harness assembly considering
(Meiner et al., 2020)
crimp height and pull-out-force as data measures
Development of quality assurance system to detect enamel layer defects as predictor of po- (Salcedo-Hernandez et
tential automotive paint film non-conformance al., 2020)
AI-enabled part type recognition and pose estimation in production system with automated
(Li et al., 2020)
sorting of parts by robots, 97+% accuracy achieved
AI-enabled decision making in manufacturing maintenance, assessing the degree of automa-
(Patalas-Maliszewska et
tion on strategic, tactical and operational level, input data based on empirical studies, 95.8%
al., 2020)
accuracy achieved
Prediction manufacturing material demand with long lead time by means of data mining using
(Widmer et al., 2018)
historical customer order data, 98% accuracy achieved
Multi-agent approach used to plan assembly processes, e.g. form new assembly stations by
(Hagemann et al., 2019)
choosing suitable components/manufacturing resources
AI-enabled balancing of work for humans and robots, aiming at optimised plans using real-
(Kousi et al., 2019)
time data from the shopfloor
Prediction of potential machine breakdowns through data of manufacturing system compo-
(Bukkapatnam et al.,
nents and plant floor automation systems, capability to predict breakdowns nearly one shift
2019)
ahead, optimisation by 90%
Implementation of algorithm to generate new process parameters for automotive injection
molding to limit process deficiencies, data used from temperature and pressure sensors, ex- (Park et al., 2019)
pert knowledge for optimum values and parameters
Condition-based monitoring and variance analysis for an automotive paint shop machine of
(Wescoat et al., 2019)
BMW to provide machine health status information

91
AI system to support selection of assembly equipment in automotive body-in-white production
(Hagemann et al., 2019)
through hyperparameter optimisation, 96.45% accuracy achieved
Development of benchmark framework to compare KPIs of multiple plants and production
lines, identification of differences and success factors allowing critical machines to learn from
best practice with optimisation, root-cause analysis and cost-benefit evaluation for decision- (Qian et al., 2019)
making, feature function learning input through fishbone diagram analysis, 2-8% throughput
increase achieved
Prescriptive AI-enabled analytics based on sensor data gathered during the production pro-
(Vater et al., 2019)
cess
AI-enabled material handling robots with use of object detection, selection and localisation to
enable errorfree applications in automotive manufacturing avoiding process interruptions, de- (Poss et al., 2019)
fects and damages, error rate reduced by 10%

Defect detection for engine X-ray images, e.g. air bubbles, accuracy above 90% (Ren et al., 2019)

Visual detection of automotive parts during assembly process, industrial use confirmed (Zhang et al., 2019)

Detection of faulty robots through information from current between motor drive and motor
with frequency and amplitude modulation, gear-wear faults detected with 96+% accuracy con- (Cheng et al., 2019)
firmed by experiments
Support of identification and selection process of multiple objects in automotive assembly line
to avoid human failure, tested with break disks/callipers and engine sensors, low implementa- (Mazzetto et al., 2019)
tion cost with accuracy from 90-100% achieved
Feasibility analysis of process automation of automotive body-in-white production system de-
(Hagemann & Stark,
sign with data-based generation of configurations, multi-method approach with case-based
2018)
reasoning, data mining, knowledge discovery and machine learning to cope with complexity
AI models used to compute log content for production control tasks, e.g. line balancing and
(Denno et al., 2018)
job sequencing
Colour-adjustment in automotive paint manufacturing with replication of human sensory per-
(Ruiz et al., 2018)
ception to reduce non-conformance costs
Implementation of self-diagnostic capability for machines to determine the health status, pre-
(Luo et al., 2018)
dicting faults and warnings, Random Forests model superior
AI-enabled image recognition for manufacturing logistics including visual inspection of ship-
ping box labels and truck trailer identifiers, critical enabler for Just-in-sequence vehicle as- (Luckow et al., 2018)
sembly
Increase of screw-fastening resource efficiency by predicting process quality in final assem-
(Matzka, 2018)
bly, process repetition rate reduced from 2.8% to 0.36%
Implementation of AI-enabled human-robot collaboration for automotive assembly where ro-
bot sensing and controls are limited to accomplish tasks, high accuracy and robustness (Wang et al., 2018)
achieved
Implementation of machine condition monitoring with continuous learning for end-of-line test-
ing in automotive manufacturing, use of torque signals as data input, SVM delivered best per- (Carino et al., 2018)
formance

Data mining used for optimising analytics in production to achieve CFRP36, reduce energy
consumption, increase throughput, forming technology (conventional as well as for new mate- (Hofmann et al., 2017)
rials), car body manufacture, corrosion protection, painting, drive trains, and final assembly

Prediction of component quality considering multi-stage, multi-parameter and time-varying


(Bai et al., 2017)
conditions that impact on quality
AI-enabled recognition of robot contact states to optimise autonomous assembly processes of
(Ortega-Aranda et al.,
dual-arm robots, F/T signal from two sensors as data basis to feed neural network, learning
2017)
achieved solely through contact force information

Detection of defects in x-ray images of automotive components, 95.2% accuracy achieved (Mery et al., 2017)

Visual inspection of produced vehicles during production process, iPad application developed
(Luckow et al., 2016)
for image data generation, average 81% accuracy achieved
Bosch production line implementation of PCA for machine fault diagnosis, Random Forest for
gearbox fault diagnosis, SVM for tool breakage detection; MCC and AUROC used as perfor- (Zhang et al., 2016)
mance metrics, increase with training data size observed
Diagnosis and classification of bearing defects in rotating machines to reduce failures, im-
prove safety and avoid downtime, vibration signals as data basis passed through wavelet (Patil et al., 2016)
transform, accuracy of 98.7% achieved with LM algorithm
End-of-line testing for automotive combustion engines with the capability of identifying defec-
(Leitner et al., 2016)
tive/nondefective engines

92
AI-enabled recognition of human activity and behaviour based on human-machine interaction
of video capture for safety purposes, cost reduction or scheduling issues, sample data taken (Makantasis et al., 2016)
from Nissan factory, higher performance achieved than state-of-the-art
Implementation of functional prototype for star washer inspection in automotive manufacturing
with identification, classification of orientation and segmentation of washer teeth, most accu- (Semeniuta et al., 2016)
rate results with adaptive boosting and Random Forest approach
Detection of the tool condition of CNC machines with self-learning capabilities to ensure sys-
tem adjustment, multiple data input used incl. cutting force, feed rate, spindle speed and cut (Huang et al., 2015)
depth, 100% total accuracy achieved
Prediction of sheet metal formability for further fault free shape processing considering draw-
ing ratio and die corner radius parameters, method confirmed by experiments, lead time re- (Bhatt et al., 2015)
duction achievable
Detection and categorisation of surface defects in iron casting production considering inclu-
(Pastor-Lopez et al.,
sions, cold laps and misruns, categorised x-ray images with OK/NOK parts and fault types,
2015)
Random Forest highest accuracy with 96.25%, KNN worst with 73%

93
5.2.4 Issues and problems
This chapter gives an answer to research question 3:
What are the issues and problems that limit the potential of AI applications in automo-
tive manufacturing?

The quantitative analysis of the papers clearly showed that the interest in AI in the context
of automotive manufacturing has increased in significance during the last few years. While
many AI systems were successfully implemented to enhance manufacturing operations and
processes, there are still potential issues and problems that limit the application of AI sys-
tems, as shown in figure 38. In particular, the quality and availability of data as well as the
proper selection and training of AI models are regarded as critical limiting factors. Also,
nowadays the level of AI maturity is confined to specific use cases and processes, whereas
a wider system integration is still seen as a key issue that demands substantial efforts in
research and development. Even if the simulation can foster AI development, there are gaps
to be considered when transferring AI to real-world applications. Despite the availability of
high computing power, it can still limit AI performance if hardware was not properly speci-
fied to meet the requirements of real-time data generation and processing. Furthermore, the
high process complexity of automotive manufacturing, together with both physical environ-
ment conditions and dynamic change, can have adverse effects on the industrial applicability
of AI systems. A brief description of each aspect is provided in table 23 and a more detailed
description of all future trends and potential innovations that were identified are provided in
table 24.

Figure 38: Key issues and problems of AI

AI maturity Computing power Data quality Data availability

Environment
Industry applicability Model selection Model training
conditions

Process complexity System integration Simulation gaps

94
Table 23: Key issues and problems of AI

Issue / Problem Description

The performance and reliability of AI systems is directly linked to the quality of data that is
Data quality provided, with necessary expert knowledge for selection, preparation, structuring and valida-
tion.

For applications in automotive manufacturing the appropriate data needs to be made perma-
Data availability nently available, ideally combined from different sources for data fusion, to be used for real-
world scenarios.

The selection of adequate models that are capable of solving the manufacturing problem is
Model selection
essential, multi-method comparisons proved useful to test and evaluate different models.

AI system performance is highly dependent on proper model training to be applied in automo-


Model training
tive manufacturing in order to ensure robustness and accuracy.

Current methods and models of AI can solve complex manufacturing problems, yet there is
AI maturity still the necessity for human operators to interfere, interpret and act upon the information
gained from AI systems.

The true potential of AI systems for effective production monitoring and controlling as well as
System integration
decision-making support only becomes feasible with a wider system integration.

Model simulation has the potential to optimise the robustness of AI systems before deploy-
Simulation gaps
ment, yet poses the risk of gaps with regard to real-world environments.

While computing power is not considered a critical issue anymore, compared with early AI
Computing power developments. For specific and complex use cases proper hardware planning is key to ensure
high performance.

The complexity of automotive manufacturing processes, such as assembly, makes AI system


Process complexity
development demanding and requires substantial expert knowledge and experience.

Physical conditions of the environment have a marked impact on performance, such as lighting
Environment conditions
in vision. Dynamic environments with a unforeseen changes are difficult to cope with.

AI systems that prove viable for confined use cases in research or simulation settings are to
Industry applicability
be scrutinized and evaluated for their applicability in real-world scenarios.

Table 24: Issues and problems of AI in automotive manufacturing

AI Issues and problems Reference

The initial sensor setup with 250 sensors did not deliver significantly better results, a configu-
(Ciampaglia et al., 2021)
ration with 10 sensors proved still viable

Limited data reliability due to reliance on third party software suppliers, expert knowledge re-
(Aronozhi et al., 2021)
quired for simulation scheme

AI system does not create the final solution, expert knowledge is still required (Klarmann et al., 2021)

Human interference is still necessary to adapt processes, further improvement needed for AI (Carvalho & Bittencourt,
system support 2021)
Limited qualitative data available for specific error detection, AI system is not ready for indus-
trial use, overall accuracy too low for single system use, hybrid use recommended, errors (Schuh et al., 2021)
manually reported by workers during assembly
Real-time worker control in practice is considered as elusive, evaluation of real-world applica-
(Overbeck et al., 2021)
tion necessary

Lack of systematic integration of inter-cycle data in machine learning implementation (Zeiser et al., 2021)

Use of uniform probability distribution (Zielinski et al., 2021)

Temperature values showed limited use for accuracy (Cui et al., 2021)

95
Short timeframe for data collection, processing and non-destructive evaluation (Zhou et al., 2021)

Neural network performance limited without extensive training data (Abagiu et al., 2021)

Dynamic nature of assembly lines limits performance, optimisation of model needed (Sinha et al., 2021)

Low AI system performance expected for cases that differ from training dataset (Peres et al., 2021)

Dataset selection depends on AI problem definition, high impact of data quality on perfor-
(Qian et al., 2021)
mance

Gap between simulated and real data due to lighting and environment conditions (Rio-Torto et al., 2021)

Limitation in datasets lies in value rather than volume (Berges et al., 2021)

The Multi-layer Perceptron algorithm used was confined to analysing one single KPI, which
(Dobrota et al., 2020)
limits the opportunity to analyse all relevant KPI

Intense computing power necessary for large datasets (Wagner et al., 2020)

Performance of deep learning method highly dependent on input/output parameters and data
(Serin et al., 2020)
amount
Intense process knowledge needed for data preparation, proper data structuring, filtering, vali-
dation, etc. needed to achieve suitable results, IT system interoperability and culture fit for (Dacal-Nieto et al., 2020)
wider implementation necessary
Concept feasibility is affected by changes to vehicle characteristics, e.g. new parts used in
(Peres et al., 2020)
previous stations, solution through monitoring and model training
Limitation in results due to deviance features determined through expert knowledge, discrete
(Zhang et al., 2020)
data and inconsistency with diagnostics criteria

Extensive classification model training needed to reach high accuracy (Nguyen et al., 2020)

Development of automated cloud system for CNN training and evaluation to add second con-
(Vater et al., 2020)
trol loop to deal with dynamic change in parameters

Limited detection performance of uncritical conditions, yet critical condition detection sufficient (Meiner et al., 2020)

Availability of factory data and speed of data collection, identification of exact location of varia- (Salcedo-Hernandez et
ble occurrence with proper timestamp option is critical al., 2020)
Estimation error in PoseNet Model, to overcome through training dataset and parameter opti-
(Li et al., 2020)
misation
AI module creates the configurations in a machine-readable form, module needed to write
(Hagemann et al., 2019)
configuration as AML file
Theoretical knowledge on AI parameters only leads to insufficient neural network configura-
tion, configuration tool developed for parameter evaluation, neural network fit only for existing (Hagemann et al., 2019)
assembly system, limitation due to random training and test data separation
Limitations encountered with vibration signals as fault information due to sensor positioning
(Cheng et al., 2019)
and environment effects
Limitations usually due to environment conditions, e.g. lighting and reflexes, yet deep learning
(Mazzetto et al., 2019)
to produce acceptable results for industrial settings
Single-method approach not appropriate, complex data basis lends itself to inconsistency and (Hagemann & Stark,
errors 2018)

Limitation in determining the buffer size with petri net and causal model methodology (Denno et al., 2018)

Expert opinion/expert colour matchers are preferable over Euclidean distance, 80% vs. 90%
(Ruiz et al., 2018)
accuracy
Data and model management/deployment of AI systems is considered a critical limitation due
to a lack of useful tools that facilitate the AI implementation process; edge deployments entail (Luckow et al., 2018)
trade-offs in terms of complexity, data security, performance, flexibility and manageability

Limitation when offline training of neural network for online application without historical data (Carino et al., 2018)

Data mining demands provision of integrated and automated process and quality data to be
viable and deliver suitable results; for a holistic approach data of subprocesses is needed;
(Hofmann et al., 2017)
Lack of implementation of ML in virtual development due to missing historical and actual
knowledge, heterogeneity and volume of data makes data mining complex

(Ortega-Aranda et al.,
Speed improvement of insertion times recommended
2017)

96
CNN approach led to lower accuracy rate (85%) despite optimisation, potentially due to over-
(Mery et al., 2017)
fitting

GPU performance good for scaling neural networks, yet weak for higher image data volumes (Luckow et al., 2016)

Limitations of algorithm performance due to dataset that is highly imbalanced and sparse, lack
(Zhang et al., 2016)
of domain knowledge requires algorithm to discover features
Standard feature-selection methods require particular attention, since the feature/fault relation
(Leitner et al., 2016)
is unclear, hence full feature to be used
Manual creation of features difficult and dependent on AI problem, automatic creation of high-
(Makantasis et al., 2016)
level features with DL architecture, for temporal information use of Motion History Image
Data quality for training sets impacts on algorithm structure and the resulting accuracy, sev-
(Huang et al., 2015)
eral cycles needed for achieving proper performance

97
5.2.5 Future Trends and Potential Innovations
This chapter gives an answer to research question 4:
What are the future trends and potential innovations of AI in automotive manufactur-
ing?

The qualitative synthesis revealed future trends and potential innovations that could push the
performance and applicability of future AI systems for automotive manufacturing, as shown
in figure 39. Many aspects are meant to address the critical limitations of contemporary AI.
While huge leaps forward have been achieved since AI was coined in the 1950s, there is still
a long way to go to reach the AI autonomy levels 4 and 5. Much of the development needs
to focus on both model development and simulation to make them more capable for complex
tasks and processes. While data availability and quality were regarded as a critical limitation,
much emphasis is expected to be put on data improvements for more accuracy and reliability
to meet all automotive manufacturing requirements. In this respect, data fusion is seen as a
crucial component of future AI systems, in order to reap the benefits of multiple data sources
that are provided in fully integrated CPPS. Through the increasing deployment of multi-
agent systems, AI should thus be enabled to engage in production control by autonomously
adapting parameters based on data analytics and knowledge generation and learning. Both
edge and cloud computing are expected to become important drivers of AI performance,
security and efficiency. The latter will be a critical competitive factor when it comes to AI
development methods to speed up the implementation of AI in the automotive industry.
Above all, a corporate culture that fosters employee engagement and AI acceptance will play
a key role in the future of AI, as do proper training and education. A brief description of each
aspect is provided in table 25 and a more detailed description of all future trends and poten-
tial innovations that were identified are provided in table 26.

Figure 39: Key future trends and potential innovations of AI

CPPS system
Cloud computing Data fusion Data quality
integration

Development Employee
Data volume Edge computing
efficiency engagement

Multi-methods and
Model development Model simulation Production control
agents

98
Table 25: Key future trends and potential innovations of AI

Trend / Innovation Description

The ongoing development of AI methods and models should make them more flexible,
Model development versatile, transferable and capable to be applied for various automotive manufacturing
problems and use cases.

While simulation has already been successfully implemented for vehicle development, it
Model simulation is expected to be increasingly used for AI to increase the robustness and applicability for
real-world scenarios.

A prerequisite of model performance and accuracy in real-world scenarios are data quality
Data quality and volume and volume to derive at datasets for AI training and development. New methods for data
optimisation are needed.

The integration of data from multiple sources becomes a key enabler for enhancing AI
Data fusion systems in automotive manufacturing, where machines, robots and IT systems interact
across all processes.

Current applications of AI are confined to specific use cases and should be extended
CPPS system integration through a wider integration adopting a CPPS perspective for achieving the AI automation
levels 4 and 5.

Implementing AI systems with multiple methods and agents is expected to further increase
Multi-methods and agents to meet automotive requirements and ensure high model robustness, accuracy, efficiency
and performance.

Further development of edge computing puts AI systems to the shop floor directly where
Edge computing data is generated and processed, enabling an increase in efficiency, performance and
security.

The use of cloud-based AI is expected to drive the development live enterprises, with cloud
Cloud computing processing and storage of big data, generated throughout the whole enterprise for deeper
managerial insights.

The shift from AI analysis and prediction is expected to shift towards AI production control,
Production control where the implemented systems are capable of adjusting production parameters based
on the knowledge gained.

Techniques to increase the efficiency of AI systems and optimise the speed while main-
Development efficiency taining accuracy requirements for industrial applications become highly relevant, including
pruning and quantisation.

Raising the awareness of employees to increase their engagement based on trust and
Employee engagement
commitment is expected to be a critical determinant of successful AI implementation.

99
Table 26: Future trends and potential innovations of AI

Future trends and potential innovations of AI Reference

Implementation of an optimisation algorithm to derive at an appropriate sensor count for accu-


(Ciampaglia et al., 2021)
racy, conduct an experimental test on an AI-analysed part

Integration of cloud-based data storage (Aronozhi et al., 2021)

Development of end-to-end learning platform with potential for wider application across manu-
(Klarmann et al., 2021)
facturing process
Wider integration of CPS, sensors and data network, improvement of classification algorithm
(Carvalho & Bittencourt,
by means of a cloud-based neural network enabling online diagnosis, model optimisation to
2021)
enable automated process control and parameter adaption

Use of feedback data to further improve the models, optimisation of datasets and data amount (Schuh et al., 2021)

Optimisation of model performance to increase reliability, implementation of control loop to


(Mayr et al., 2021)
automatically adjust process parameters

Optimisation of model to cater for changes in production layouts (Overbeck et al., 2021)

Implementation in real-world scenario beyond simulation and development of demonstrator,


optimisation performance through multi-agent approach and application of real-time location (Jeong et al., 2021)
systems for dynamic obstacle handling
Development of 2D CNN with spectrogram and Mel-spectrogram sound images, comparison
of manual and automated feature extraction and transfer of learning techniques, handling of (Espinosa et al., 2021)
noisy sound environment through occlusion and reduction methods
Development of augmented, multi-modal machine learning and information fusion techniques,
building trust and acceptance for machine learning systems alongside technology develop- (Zeiser et al., 2021)
ment
Application of deep learning time series forecasting with multiple data sources and parame-
(Mohan et al., 2021)
ters that affect breakdowns, e.g. oil pressure, vibration

Full automation of monitoring system (Magro et al., 2021)

Consideration of processing parameters for optimised defect prediction, adaption of process


(Cui et al., 2021)
parameters during production
Increase of data quality through expert input, consideration of transfer learning models for de-
(Goncalves et al., 2021)
mand forecasting, integration of more components for a wider perspective

Focus on edge computing for non-destructive evaluation (Zhou et al., 2021)

Analysis of environment impact, such as noise or lack of data quality and volume or sensors (Rahimi et al., 2021)

Provision of higher data volume for neural network training (Abagiu et al., 2021)

Wider integration of AI system across assembly stations, extension of correlation patterns for
(Queiroz et al., 2021)
more assembly stages
Future potential for quantitative modelling of invariant feature considering various multi-sta-
(Sinha et al., 2021)
tion assemblies
Simulation could be improved to include more non-conformance cases, generation of annota-
(Peres et al., 2021)
tions needs improvement to increase accuracy with localisation of bounding boxes
Conversion of determined corrective actions into adjustable process parameters to ensure
(Sinha et al., 2021)
machine availability

Development of correction system for additional trials instead of increase in data volume (Qian et al., 2021)

Extension of datasets for more accuracy and performance (Kumari et al., 2021)

System extension to detect object movement and ensure flexible camera angle adjustment (El Hachem et al., 2021)

Use of more realistic 3D models with high quality textures (Rio-Torto et al., 2021)

Data engineering approaches to increase data value and relate potential features to problem (Berges et al., 2021)

Optimisation through increase in multi-agent system training time in real-world scenarios,


(Gankin et al., 2021)
evaluation and comparison of other deep reinforcement learning approaches

100
A multi-company platform could be established to provide the AI system access to big data,
which should provide more reliable data and conclusions on problems, risks and the respec- (Dobrota et al., 2020)
tive KPIs
Development of AI model application guide necessary, optimisation of accuracy by using ad-
ditional parameters for model training, focus on continuous learning approach to cater for time (Wagner et al., 2020)
effects and changes in functional product evaluation
Application of deep learning prediction models for tool war and chatter to increase machine
(Serin et al., 2020)
quality, optimisation of input parameters through estimation methods
Wider integration of AI system with car instance feature vector across various manufacturing
(Dacal-Nieto et al., 2020)
processes
Implementation of a multi-agent system within the CPPS to constantly update the model
(Peres et al., 2020)
based on new data, adaption to changes are accounted for without model amendments

Improvement of Bayesian network structure learning and focus on data fitting (Zhang et al., 2020)

Improvement of processing performance using GPU and detection through multi-agent ap-
(Muresan et al., 2020)
proach combining learning methods
Increase in AI efficiency through pruning and trained quantisation, use of Grad-Cam++ to re-
veal features and understand deep learning model principles, implementation of alternative
(Nguyen et al., 2020)
CNN models to reduce high training demand, implementation of AI system for fully-automated
machines
Focus on transferability of models across machines considering differing specifications for er-
ror conditions and more quality features, use case for automated optical inspection with deep (Meiner et al., 2020)
learning techniques to deal with higher data complexity
(Salcedo-Hernandez et
Plant digitalisation and system integration to cater for high data and variable requirements
al., 2020)
(Patalas-Maliszewska et
Use larger dataset for training, optimisation of neural network model
al., 2020)
Extension of vector space model approach with K-means clustering algorithm, integration of
economic aspects including price or economies of scale using weighting values for different (Widmer et al., 2018)
part volumes
Impact analysis for data quality recommended to increase amount of training data, future po-
(Hagemann et al., 2019)
tential to integrate data of digital twin
Optimisation of AI systems to be applicable for bigger applications, future potential for imple-
(Kousi et al., 2019)
menting within the IT system landscape
Implementation of learning algorithms for a broader AI-enabled production control system,
(Park et al., 2019)
validation through empirical experiments
Optimisation modelling and testing by using defect components, increase of feature amount to
(Wescoat et al., 2019)
increase performance and accuracy, integration of wider production data
Development of multiple neural networks to use for each resource type, improvement of da-
(Hagemann et al., 2019)
tasets in volume/quality and more attributes for better resource distinction
Research gap in AI-enabled monitoring, controlling and parameter adaption of production pro-
(Vater et al., 2019)
cess, development of framework for prescriptive automation necessary

Application of optimised deep neural network architectures considering small sample sizes (Ren et al., 2019)

Integration of learning and detection techniques to better cope with changes in assembly line (Mazzetto et al., 2019)

(Hagemann & Stark,


Implementation of multi-method approach for industrial use case
2018)
Application of methodology in real-time settings for decision support, evaluation of wider inte-
(Denno et al., 2018)
gration in manufacturing systems
Development of a decision-making application which can propose vectors in pigmentary
(Ruiz et al., 2018)
space to set the correct target colour
Optimisation of self-diagnostic production system for more automotive components, harness-
(Luo et al., 2018)
ing potential of IoT to increase machine robustness, evaluation of machine data correlations
Optimisation of deep learning capability for high volume data and images, improvement of
model management to reduce trade-offs including accuracy and performance in edge and (Luckow et al., 2018)
cloud environments
Optimisation of algorithms with bigger datasets and application of singular spectrum analysis
(Matzka, 2018)
on time series data
Potential for optimisation through dynamic movement primitives model and enhancing multi-
(Wang et al., 2018)
modal robot data including speed, trajectory, object orientation, proximity

101
Connected car to communicated with dealerships and repair shops in case of defects, which
are passed to production and development for analysis and optimisation. Shift of systems
(Hofmann et al., 2017)
from decision-support to decision making, quantum computing will enable more powerful AI
systems,
Potential to enable robot to react to new experiences instead of taught contact states by de-
(Ortega-Aranda et al.,
veloping the primitive knowledge base, knowledge generation could be automated to build
2017)
more capable assembly robots
Optimisation of deep learning network with bigger datasets, use of sliding-windows strategy
(Mery et al., 2017)
for whole x-ray images
Evaluation of distributed deep learning systems for bigger datasets and more complex net-
(Luckow et al., 2016)
works, application of deep learning models including sequence-to-sequence learning
Increase in big data enable the implementation of predictive models for quality assurance, op-
timised classifier parameters can improve performance, extension of the numerical dataset by (Zhang et al., 2016)
categorical and timestamp data recommended

Improvement of support vector data description specificity (Leitner et al., 2016)

Optimised system integration with consideration of robot and vision, including multi-pose im-
aging, adaption of lighting modes, enhancement of image measurement during inspection (Semeniuta et al., 2016)
process
Increase of system robustness by using K-fold cross validation for classifying datasets, appli-
cation of experimental design to identify the required number of trainings datasets for ideal re- (Huang et al., 2015)
sults
Use of Best Crossing Line Profile features in unsupervised learning for distinguishing between (Pastor-Lopez et al.,
regular and defect structures 2015)

102
6 Conclusion

The major objective of this thesis was to outline the contemporary state and extent of AI
adoption in automotive manufacturing by identifying AI models and methods, applications
and best practices as well as limitations and future prospects. Given that there was a lack of
comprehensive and up-to-date meta-studies with regard to AI in automotive manufacturing,
the results of this thesis provide valuable insights for both research and industry. The theo-
retical foundations that were presented on manufacturing principles and the shift towards an
Industry 4.0 network architecture showed, that AI can be regarded as a critical success factor
for reaping the benefits of Industry 4.0. Throughout the past decades, the optimisation of
automotive manufacturing processes has primarily been based on implementing information
and communications technologies, automation as well as lean manufacturing methods de-
veloped by the Japanese car maker Toyota. European automotive manufacturers have there-
fore aimed to improve product quality, reduce costs and ensure customer satisfaction by
adopting continuous improvement, TQM, just-in-sequence or advanced robotics enabled by
sophisticated IT hardware and software. Nowadays, the high operating expenses in Europe
pose a marked risk for manufacturers and urge them to embrace the paradigm of digital
transformation. Not only have manufacturers from Asia gained an increasing competitive
edge, it is also evident that they have devoted much effort to generating know-how and ex-
periences in manufacturing as well as research and development. Hence, the elements of the
fourth industrial revolution are seen as a key lever to sustain supremacy and achieve sustain-
able competitive advantage. AI technologies have experienced enormous technological ad-
vancements since they were first introduced in the 1950s and went through two AI winters
with serious setbacks and diminishing hope to achieve the high promises that were made.
The major contributing factors were improvements in computing power, the development of
sophisticated AI methods and models as well as the increase in big data. Furthermore, the
internet led to the formation of knowledge groups that spread the interest in AI development
and education on a global scale. AI has thus become an integral part of contemporary com-
puter science and is on top of the agenda of both research and industry. As a consequence,
the maturity of AI has significantly surged and led to more complex and enhanced applica-
tions of AI in the automotive industry, from autonomous driving to automotive manufactur-
ing. Nevertheless, AI does not lend itself to be regarded as the single component that is
required to increase both the effectiveness and efficiency of automotive operations. The true
potential is highly dependent on an integrated system approach in the context of Industry
4.0, in order to establish a so-called cyber-physical production system capable of transform-
ing businesses into real-time enterprises. The latter is based on a synchronised implementa-
tion of major Industry 4.0 components, such as the internet of things, machine to machine
communication or big data. The results of the systematic literature review that was con-
ducted as part of this thesis confirmed the rising interest in the topic of AI in automotive
manufacturing and the potential advantages. The quantitative analysis revealed that the

103
number of papers increased tenfold in the year 2021 compared with the year 2015. AI models
have been implemented in various aspects of automotive manufacturing, including assem-
bly, quality, production, supply chain or business intelligence. A wide range of use cases
confirmed the industrial applicability of AI models by achieving high industrial robustness
and accuracy rates of more than 90 percent. Examples of identified automotive AI use cases
include the detection of objects, surfaces and errors to enhance quality inspection, production
process analysis and predictive machine maintenance, robot assembly and human-robot col-
laboration, prediction of material demand or AI-enabled manufacturing decision making.
The majority of use cases applied machine learning methods with mostly supervised learning
approaches where proper training datasets were available, only a few relied on unsupervised
approaches. The analysis also revealed that deep learning was largely used to solve quality
inspection problems, whereas reinforcement learning was adopted for vehicle assembly
problems. More than 70 AI models and algorithms were applied in use cases of automotive
manufacturing, with neural networks as the most used models. Further models that were
commonly applied include support vector machine, random forest, decision tree, multi-layer
perceptron and k-nearest-neighbor. A multi-method approach was adopted by almost half of
the identified papers as a means to determine the most advanced model based on trial-and-
error or to avoid any trade-offs and disadvantages entailed by the use of a single-method
approach. The latter proved to be inappropriate to cope with complex processes, dynamic
environments and unpredictable change. Problems occurred where a sufficient availability
and quality of data could not be ensured for proper model training and development. Also, a
lack of computing power and challenging environmental conditions can lead to adverse ef-
fects. In order to ensure a high degree of industrial applicability, much emphasis needs to be
placed on data integrity, model selection and development, system integration and ongoing
monitoring and optimisation. Expert knowledge and experiences are required for both AI
development and automotive manufacturing to make AI systems work as intended. In this
respect, the importance of employee engagement that was stressed by Toyota to successfully
implement the principles of its production system still holds true for the age of Industry 4.0
and the digital transformation. Automotive players that are capable of adopting AI-enabled
technologies across their value chain while developing a learning culture that embraces
change and innovation will arguably succeed and prosper from a long-term perspective. The
knowledge gained through the results of this thesis can thus aid to select appropriate models
and methods according to the respective AI problem and automotive manufacturing process.
It is recommended to use the findings for creating a supporting framework for the model
selection and development process to ensure a high degree of system robustness and accu-
racy. Further research is needed to evaluate the industrial applicability of AI and the corre-
sponding data requirements for different manufacturing problems. In addition, future studies
could address the problem of wider system integration of AI, covering more interrelated
processes, in order to promote understanding and outline the prerequisites for successful
industrial applications.

104
References

ABAGIU, MARIAN/COJOCARU, DORIAN/MANTA, FLORIN/MARINIUC,


ALEXANDRU: Detection of a Surface Defect on an Engine Block Using Com-
puter Vision, in: 22nd International Carpathian Control Conference (ICCC), Velke
Karlovice 2021, p. 1-5.

AI HLEG: A definition of AI: Main capabilities and scientific disciplines, 2019,


https://www.aepd.es/sites/default/files/2019-12/ai-definition.pdf, [25.07.2021].

ALLEN, ROBERT: Why was the Industrial Revolution British?, 2009, https://voxeu.org/arti-
cle/why-was-industrial-revolution-british, [10.5.2021].

ALSOP, THOMAS: Number of transistors in Graphcore’s Colossus MK2 vs. Nvidia’s


A100 processor, 2020, https://www.statista.com/statistics/1134502/graphcore-co-
lossus-mk2-vs-nvidia-a100-transistor-count/, [25.07.2021].

ALUKAL, GEORGE/MANOS, ANTHONY: Lean Kaizen: a simplified approach to pro-


cess improvements, Quality Press, Milwaukee.

AMANATH, BEENA/JARVIS, DAVID/HUPFER, SUSANNE: Thriving in the era of per-


vasive AI, 2020, https://www2.deloitte.com/us/en/insights/focus/cognitive-technol-
ogies/state-of-ai-and-intelligent-automation-in-business-sur-
vey.html?id=us:2pm:3ad:myil-
afy19:eng:greendot:em:tech:cn:stateAI:1x1:hbr:021519:57554769, [22.07.2021].

ARUNOZHI, MANIMUTHU/VENKATESH, V.G./SHI, PETER: Design and develop-


ment of automobile assembly model using federated AI with smart contract, in: In-
ternational Journal of Production Research, 10.1080/00207543.2021.1988750,
2021.

AVL: Simultaneous Engineering, 2021, https://www.avl.com/-/simultaneous-engineering,


[26.09.2021].

AVRAM, MARICELA-GEORGIANA: Advantages and challenges of adopting cloud


computing from an enterprise perspective, in: Procedia Technology, 2014, issue 12,
p. 529-534.

AWAD, MARIETTE/KHANNA, RAHUL: Machine Learning, 2015,


https://link.springer.com/chapter/10.1007/978-1-4302-5990-9_1, [28.10.2021].

VIII
AZIZI, AYDAN: Applications of AI Techniques in Industry 4.0, Springer, Singapore
2018.

BAI, YUN/SUN, ZHENZHONG/DENG, JUN: Manufacturing Quality Prediction Based


on Two-Step Feature Learning Approach, in: International Conference on Sensing,
Diagnostics, Prognostics, and Control (SDPC), Shanghai 2017, p. 260-263.

BATRA, GAURAV/JACOBSON, ZACH/MADHAV, SIDDARTH/QUEIROLO, AN-


DREA/SANTHANAM, NICK: Artificial-intelligence hardware: New opportunities
for semiconductor companies, 2018, https://www.mckinsey.com/~/media/McKin-
sey/Industries/Semiconductors/Our%20Insights/Artificial%20intelligence%20hard-
ware%20New%20opportunities%20for%20semiconductor%20companies/Artifi-
cial-intelligence-hardware.ashx, [25.07.2021].

BERGES, CORINNE/BIRD, JIM/SHROFF, MEHUL/RONGEN, RENE/SMITH, CHRIS:


Data analytics and machine learning: root-cause problem-solving approach to pre-
vent yield loss and quality issues in semiconductor industry for automotive applica-
tions, in: International Symposium on the Physical and Failure Analysis of Inte-
grated Circuits (IPFA), Singapore 2021, p. 1-10.

BHATT, MALAY/BUCH, SANJAY: Prediction of formability for sheet metal component


using artificial intelligent technique, in: 2nd International Conference on Signal
Processing and Integrated Networks (SPIN), Noida 2015, p. 388-393.

BIDDLE, EDD: Select and develop an AI and data science model, 2021,
https://www.ibm.com/garage/method/practices/reason/model-selection-develop-
ment-ai-data-science/, [28.10.2021].

BINDER, ALAN: Automotive industry, https://www.britannica.com/technology/automo-


tive-industry/The-modern-industry, [22.09.2021].

BLAGOJ, DELIPETREV/TSINARAKI, CHRISA/KOSTIC, UROS: AI Watch: Historic Evo-


lution of AI, 2020, https://publications.jrc.ec.europa.eu/repository/bit-
stream/JRC120469/jrc120469_historical_evolution_of_ai-v1.1.pdf, [23.07.2021].

BRACE, IAN: Questionnaire Design, Kogan Page, London: 2004.

BRATZEL, STEFAN: AutomotiveINNOVATIONS: Ranking of the most innovative auto-


motive OEMs and premium brands, 2021, https://auto-

IX
institut.de/automotiveinnovations/automotiveinnovations-ranking-of-the-most-in-
novative-automotive-oems-and-premium-brands-2021/, [30.11.2021].

BROOKS, BRIAN: The Biggest Milestones in the History of Automotive Manufacturing,


https://www.qad.com/blog/2019/12/the-biggest-milestones-in-the-history-of-auto-
motive-manufacturing, [23.09.2021].

BRUNNER, HELMUT/ROSSBACHER, PATRICK/HIRZ, MARIO: Sustainable Product


Development: Provision of Information in Early Automotive Engineering Phases,
in: International Conference Management of Technology, 2017, https://www.re-
searchgate.net/publication/316350906_Sustainable_product_development_provi-
sion_of_information_in_early_automotive_engineering_phases, [04.10.2021].

BUKKAPATNAM, SATISH/AFRIN, KAHKASHAN/DAVE, DARPIT/KUMARA,


SOUNDAR: Machine learning and AI for long-term fault prognosis in complex
manufacturing systems, in: CIRP Annals, Vol. 68, Issue 1, 2019, p. 459-462.

BULLINARIA, JOHN: The Roots, The Goals and Sub-fields of AI,


https://www.cs.bham.ac.uk/~jxb/IAI/w2.pdf, [05.08.2021].

CAM, ARIF/CHUI, MICHAEL/HALL, BRYCE: Global AI Survey: AI proves its worth, but
few scale impact, 2019, https://www.mckinsey.com/featured-insights/artificial-intelli-
gence/global-ai-survey-ai-proves-its-worth-but-few-scale-impact, [10.10.2021].

CARINO, JESUS/DELGADO-PRIETO, MIGUEL/IGLESIAS, JOSE/SANCHIS, ARA-


CELI/ZURITA, DANIEL/MILLAN, MARTA/REDONDO, JUAN/ROMERO-TRON-
COSO, RENE: Fault Detection and Identification Methodology Under an Incremental
Learning Framework Applied to Industrial Machinery, in: IEEE Access, Vol. 6, 2018,
p. 49755-49766.

CARPART: Who are Toyota’s car parts suppliers?, 2021, https://carpart.com.au/blog/manufac-


turers/who-are-toyota-car-parts-suppliers, [24.09.2021]

CARVALHO, CLEGINALDO/BITTENCOURT, PRISCILA: Industry 4.0 Machine Learning


to Monitor the Life Span of Cutting Tools in an Automotive Production Line, in: Inter-
national Journal of Advanced Engineering Research and Science, Vol-8, Issue-5, 2021,
p. 220-228.

X
CAREY, NICK/SUN, YILEI: China’s electric carmakers make their move on Europe, 2021,
https://www.reuters.com/business/autos-transportation/chinas-electric-carmakers-
make-their-move-europe-2021-09-30/, [17.12.2021].

CDU: Merkel: Digitaler Wandel wie Industrielle Revolution, https://archiv.cdu.de/artikel/mer-


kel-digitaler-wandel-wie-industrielle-revolution, [20.05.2021]

CHEN, RONG/YICHENG, LI/JIA-MING, CAI/KUN, CAO: Atomic level deposition to


extend Moore’s law and beyond, 2020, https://www.researchgate.net/publica-
tion/342436758_Atomic_level_deposition_to_extend_Moore%27s_law_and_be-
yond, [24.07.2021].

CHEN, YI/WANG, WEITIAN/KROVI, VENKAT/JIA, YUNYI: Enabling Robot to Assist


Human in Collaborative Assembly using convolutional neural network, in: RSJ Inter-
national Conference on Intelligent Robots and Systems (IROS), Las Vegas 2020, p.
11167-11114.

CHENG, FANGZHOU/RAGHAVAN, AJAY/JUNG, DEOKWOO/SASAKI, YUKI-


NORI/TAJIKA, YOSUKE: High-Accuracy Unsupervised Fault Detection of Industrial
Robots Using Current Signal Analysis, in: International Conference on Prognostics and
Health Management (ICPHM), San Francisco 2019, p. 1-8.

CHISNALL, PETER: Marketing research, McGraw-Hill Education, Maidenhead: 2005.

CHM: 1948: The European Transistor Invention, https://www.computerhistory.org/siliconen-


gine/the-european-transistor-invention/, [15.03.2021].

CHURCHILL, GILBERT/IACOBUCCI, DAWN: Marketing Research: Methodological


Foundations, South-Western, Mason: 2005.

CERF, VINTON/LEINER, BARRY/CLARK, DAVID/KAHN, ROBERT/KLEINROCK,


LEONARD/LYNCH, DANIEL/POSTEL, JON/ROBERTS, LARRY/WOLFF,
STEPHEN: A Brief History of the Internet, 1997, https://www.internetsoci-
ety.org/wp-content/uploads/2017/09/ISOC-History-of-the-Internet_1997.pdf,
[15.07.2021].

CIAMPAGLIA, ALBERTO/MASTROPIETRO, ANTONIO/DE GREGORIO, ALES-


SANDRO/BELINGARDI, GIOVANNI: AI for damage detection in automotive
composite parts: a use case, in: SAE Int. J. Adv. & Curr. Prac. in Mobility,
3(6):2936-2945, 2021.

XI
CLIFFORD, JOE: Power to the people – Toyota’s suggestion system, 2014,
https://mag.toyota.co.uk/toyota-and-the-power-of-suggestion/, [20.11.2021].

CONTRERAS, IVAN/VEHI, JOSEP: AI for Diabetes Management and Decision Support:


Literature Review, in: Journal of Medical Internet Research, Vol. 20, 2018.

COPELAND, MICHAEL: What’s the Difference Between AI, Machine Learning and
Deep Learning?, 2016, https://blogs.nvidia.com/blog/2016/07/29/whats-difference-
artificial-intelligence-machine-learning-deep-learning-ai/, [24.07.2021].

CORNET, ANDREAS/DEUBENER, HARALD/MÖLLER, TIMO/SCHAUFUSS, PAT-


RICK/TSCHIESNER, ANDREAS: A long-term vision for the European automo-
tive industry, 2019, https://www.mckinsey.com/industries/automotive-and-assem-
bly/our-insights/a-long-term-vision-for-the-european-automotive-industry#,
[26.10.2020].

CUI, JIACHENG/LIU, WIE, ZHANG, YANG/GAO, CHANGYONG/LU, ZHE/LI,


MING/WANG, FUJI: A novel method for predicting delamination of carbon fiber
reinforced plastic (CFRP) based on multi-sensor data, in: Mechanical Systems and
Signal Processing, Vol. 157, 107708, 2021.

CULOT, GIOVANNA/NASSIMBENI, GUIDO/ORZES, GUIDO/SARTOR, MARCO:


Behind the definition of Industry 4.0: Analysis and open questions, 2020,
https://www.sciencedirect.com/science/article/abs/pii/S0925527320300050,
[20.03.2021].

CUTHBERTSON, ANTHONY: Elon Musk claims AI will overtake humans in less than
five years, 2020, https://www.independent.co.uk/life-style/gadgets-and-
tech/news/elon-musk-artificial-intelligence-ai-singularity-a9640196.html,
[23.07.2021].

DACAL-NIETO, ANGEL/AREAL, JUAN/GARCIA-FERNANDEZ, MIGUEL/LLUCH,


MARCOS: Use cases and success stories of a data analytics system in an automo-
tive Paint Shop, in: Eighth International Symposium on Computing and Network-
ing (CANDAR), Naha 2020, p. 95-100.

DAIMLER: 125 years of innovation: "The love of invention never ceases" – from the Benz
patent motor car to the second invention of the automobile, https://media.daim-
ler.com/marsMediaSite/en/instance/ko/125-years-of-innovation-The-love-of-

XII
invention-never-ceases--from-the-Benz-patent-motor-car-to-the-second-invention-
of-the-automobile.xhtml?oid=9362022, [22.09.2021].

DAVIES, D. S./STAMMERS, JUDIT: “The Effect of World War II on Industrial Sci-


ence.” Proceedings of the Royal Society of London. Series A, Mathematical and
Physical Sciences, vol. 342, no. 1631, 1975, p. 505–518. JSTOR,
www.jstor.org/stable/78750, [14.03.2021].

DEANE, PHYLLIS: The first industrial revolution, 2nd Edition, Cambridge University
Press, Cambridge 1979.

DEMLEHNER, QUIRIN/SCHÖMER, DANIEL/LAUMER, SVEN: How can aritifical in-


telligence enhance car manufacturing? A Delphi study-based identification and as-
sessment of general use cases, in: International Journal of Information Manage-
ment, Vol. 58, 2021.

DENNO, PETER/DICKERSON, CHARLES/HARDING, JENNIFER: Dynamic produc-


tion system identification for smart manufacturing systems, in: Journal of Manufac-
turing Systems, Vol. 48, Part C, 2018, p. 192-203.

DIEZ, PABLO/RODRIGUEZ, JOSE: The Strategy Clock in the car manufacturing sector
in Spain, 2011, http://www.guerrasynavas.com/pdf/CB-015-EN_Strat-
egy_clock_in_the_car_industry.pdf, [25.09.2021].

DINGLI, ALEXIEI/HADDOD, FOAAD/KLÜVER, CHRISTINA: AI in Industry 4.0: A


Collection of Innovative Research Case-studies that are Reworking the Way We
Look at Industry 4.0 Thanks to AI: Springer International Publishing, Cham 2021.

DOBROTA, DAN/DUMITRASCU, OANA/DUMITRASCU, MANUEL: Performance


Evaluation for a Sustainable Supply Chain Management System in the Automotive
Industry Using AI, in: Processes, 8(11):1384, 2020.

DUNN, ALISON: The father of invention: Dick Morley looks back on the 40th anniver-
sary of the PLC, https://www.automationmag.com/855-the-father-of-invention-
dick-morley-looks-back-on-the-40th-anniversary-of-the-plc/, 2009, [15.03.2021].

EL HACHEM, CHARBEL/PERROT, GILLES/PAINVIN, LOIC/COUTURIER, RAPH-


AEL: Automation of Quality Control in the Automotive Industry Using Deep
Learning Algorithms, in: International Conference on Computer, Control and Ro-
botics (ICCCR), Shanghai 2021, p. 123-127.

XIII
ELSEVIER: ScienceDirect: Elsevier’s premier platform of peer-reviewed literature, 2021,
https://www.elsevier.com/solutions/sciencedirect, [25.03.2021].

ENGELMANN, RYAN: The Second Industrial Revolution, 2021, https://ushisto-


ryscene.com/article/second-industrial-revolution/, [10.03.2021].

ESPINOSA, RICARDO/PONCE; HIRAM/GUTIERREZ, SEBASTIAN: Click-event


sound detection in automotive industry using machine/deep learning, in: Applied
Soft Computing, Vol. 108, 2021.

ETHzürich: ETH AI CENTER, https://ai.ethz.ch/, [05.08.2021].

EUROPEAN COMMISSION: CO2 emission performance standards for cars and vans,
2021, https://ec.europa.eu/clima/eu-action/transport-emissions/road-transport-re-
ducing-co2-emissions-vehicles/co2-emission-performance-standards-cars-and-
vans_en, [02.12.2021]

EVANGELIOU, NIKOLAOS/STAMATIS, GEORGIOS/BRAVOS, GEORGE: AI for In-


bound Logistics Optimisation in Automotive Industry, in: VERMESAN,
OVIDIU/JOHN, REINER/DE LUCA, CRISTINA/COPPOLA, MARCELLO
(Eds.), AI for Digitising Industry, River Publishers, 2021.

EWING, JACK: China’s Popular Electric Vehicles Have Put Europe’s Automakers on No-
tice, https://www.nytimes.com/2021/10/31/business/electric-cars-china-eu-
rope.html, [19.11.2021].

FAST COMPANY: Most Innovative Companies, 2008, https://www.fastcom-


pany.com/most-innovative-companies/2008, [30.11.2021].

FFG: Production of the Future – the calls, https://www.ffg.at/en/program/migriert-produc-


tion-future-calls, [28.11.2021].

FISK, DALE: Programming with Punched Cars, 2005, http://www.columbia.edu/acis/his-


tory/fisk.pdf, [15.03.2021].

FLASIŃSKI, MARIUSZ: Introduction to AI, Springer International Publishing, Switzer-


land 2016.

XIV
FRANKISH, KEITH/RAMSEY, WILLIAM: The Cambridge Handbook of AI, Cambridge
University Press, Cambridge 2014.

FRERE, ERIC/ZURECK, ALEXANDER/RÖHRIG, KATHARINA: Industry 4.0 in Ger-


many – The Obstacles Regarding Smart Production in the Manufacturing Industry,
in: SSRN Electronic Journal, 2018.

FORD: The Henry Ford story, 2021, https://www.ford.co.uk/experience-ford/history-and-


heritage, [14.03.2021].

GAENZLE, STEFFEN/WEISS, CHRISTIAN/ROEMER, MICHAEL: How automakers


can survive the self-driving era, 2016, https://www.kearney.com/automotive/arti-
cle?/a/how-automakers-can-survive-the-self-driving-era, [24.09.2021].

GANKIN, DENNIS/MAYER, SEBASTIAN/ZINN, JONAS/VOGEL-HEUSER, BIR-


GIT/ENDISCH, CHRISTIAN: Modular Production Control with Multi-Agent
Deep Q-Learning, in: 26th IEEE International Conference on Emerging Technolo-
gies and Factory Automation (ETFA), Vasteras 2021, p. 1-8.

GERMIS, CARSTEN: VW verklagt Zulieferer Prevent, 2020, https://www.faz.net/aktu-


ell/wirtschaft/unternehmen/lieferstopps-volkswagen-verklagt-zulieferer-prevent-
16569655.html, [25.09.2021].

GILABERT, SAULO/ARNAIZ, AITOR: Welding Process Quality Improvement with Ma-


chine Learning Techniques, in: IFAC-PapersOnLine, Vol. 54, Issue 1, 2021, p.
343-348.

GIRASA, ROSARIO: AI as a Disruptive Technology, Palgrave Macmillan, Cham 2020.

GONCALVES, JOAO/CORTEZ, PAULO/CARVALHO, SAMEIRO/FRAZAO, NUNO:


A multivariate approach for multi-step demand forecasting in assembly industries:
Empirical evidence from an automotive supply chain, in: Decision Support Sys-
tems, Vol. 142, 113452, 2021.

GRAPHCORE: Intelligence Processing Unit, 2021, https://www.graphcore.ai/products/ipu,


[25.07.2021].

GTAI: The Automotive Industry in Germany, https://www.gtai.de/re-


source/blob/64100/817a53ea3398a88b83173d5b800123f9/industry-overview-auto-
motive-industry-en-data.pdf, [22.09.2021].

XV
GUPTA, SHUCHITA/AMABA, BEN/MCMAHON, MIKE/GUPTA, KUNAL: The Evo-
lution of AI in the Automotive Industry, in: Annual Symposium on Reliability and
Maintainability (RAMS), Orlando 2021, p. 1-7.

GUPTON, NANCY: Benjamin Franklin and the kite experiment, 2017,


https://www.fi.edu/benjamin-franklin/kite-key-experiment, [11.03.2021].

HAGEMANN, SIMON/STARK, RAINER: Hybrid AI System for the Design of Highly-


Automated Production Systems, in: 4th International Conference on Frontiers of Ed-
ucational Technologies, Moscow 2018, p. 192-196.

HAGEMANN, SIMON/SÜNNETCIOGLU, ATAKAN/STARK, RAINER: Hybrid AI


System for the Design of Highly-Automated Production Systems, in: Procedia
Manufacturing, Vol. 28, 2019, p. 160-166.

HAGEMANN, SIMON/SÜNNETCIOGLU, ATAKAN/FAHSE, TOBIAS/STARK,


RAINER: Neural Network Hyperparameter Optimization for the Assisted Selection
of Assembly Equipment, in: 23rd International Conference on Mechatronics Tech-
nology (ICMT), Salerno 2019, p. 1-7.

HALL, MARK: Microsoft Corporation, 2021, https://www.britannica.com/topic/Mi-


crosoft-Corporation, [20.12.2021].

HASSANI, HOSSEIN/SILVA, EMMANUEL/UNGER, STEPHANE/TAJMAZINANI,


MAEDEH/MACFEELY, STEPHEN: AI (AI) or Intelligence Augmentation (IA):
What is the Future?, in: MDPI Journal of AI, 2020, 1(2), p. 143-155.

HAO, KAREN: The computing power needed to train AI is now rising seven times faster
than ever before, 2019, https://www.technolo-
gyreview.com/2019/11/11/132004/the-computing-power-needed-to-train-ai-is-now-
rising-seven-times-faster-than-ever-before/, [25.07.2021].

HIGGINS, JULIAN/GREEN, SALLY: Cochrane Handbook for Systematic Reviews of In-


terventions, 2011, https://handbook-5-1.cochrane.org/chap-
ter_1/1_2_2_what_is_a_systematic_review.htm, [25.03.2021].

HISTORY: Morse code & the telegraph, 2019, https://www.history.com/topics/inven-


tions/telegraph, [10.03.2021].

XVI
HU, YINGJIE/LI, WENWEN/WRIGHT, DAWN/AYDIN, ORHUN/WILSON, DAN-
IEL/MAHER, OMAR/RAAD, MANSOUR: AI Approaches, 2019, https://www.re-
searchgate.net/publication/335089772_Artificial_Intelligence_Approaches,
[12.08.2021].

HUANG, POTSANG/MA, CHENG-CHIEH/KUO, CHIA-HAO: A PNN self-learning tool


breakage detection system in end milling operations, in: Applied Soft Computing,
Vol. 37, 2015, p. 114-124.

HOFMANN, MARTIN/NEUKART, FLORIAN/BÄCK, THOMAS: AI and Data Science


in the Automotive Industry, 2017, https://www.researchgate.net/publica-
tion/319534479_Artificial_Intelligence_and_Data_Science_in_the_Automotive_In-
dustry, [22.12.2021].

IEEE: IEEE at a Glance, 2021, https://www.ieee.org/about/at-a-glance.html, [25.03.2021].

INTEL: Robert Noyce: Statesman of Silicon Valley, https://www.intel.la/con-


tent/www/xl/es/history/museum-robert-noyce.html, [15.03.2021].

INTEL AI: The Rise in Computing Power: Why Ubiquitous AI Is Now A Reality, 2018,
https://www.forbes.com/sites/intelai/2018/07/17/the-rise-in-computing-power-why-
ubiquitous-artificial-intelligence-is-now-a-reality/?sh=6dedcf041d3f, [25.07.2021].

JEONG, YONGKUK/AGRAWAL, TARUN/FLORES-GARCIA, ERIK/WIKTORSSON,


MAGNUS: A reinforcement learning model for material handling task assignment
and route planning in dynamic production logistics environment, in: Procedia
CIRP, Vol. 104, 2021, p. 1807-1812.

JESPERSEN, DAM/SKJOTT-LARSEN, TAGE: Supply Chain Management: In Theory


and Practice, Holbaek Amts Bogtrykkeri, Copenhagen 2005.

JOHNSON, GERRY/WHITTINGTON, RICHARD/SCHOLES, KEVAN/ANGWIN,


DUNCAN/REGNER, PATRICK: Exploring Corporate Strategy, 8th Edition, Pear-
son Education, Harlow 2008.

JOSHI, AMEET: Machine Learning and AI, Springer Verlag, Cham 2020.

JOSHI, NAVEEN: Seven Types of AI, https://www.forbes.com/sites/cognitive-


world/2019/06/19/7-types-of-artificial-intelligence/?sh=7650c316233e,
[05.08.2021].

XVII
JOVANOVIC, MILICA/DLACIC, JASMINA/OKANOVIC, MILAN: Digitalization and
society’s sustainable development – Measures and implications, 2018,
https://www.researchgate.net/publication/329973157_Digitalization_and_Socie-
ty's_Sustainable_Development_-_Measures_and_Implications, [14.08.2021].

JYLKÄS, TITTA: Shared Path Service Design and AI in Designing Human-Centred Digi-
tal Services, 2020, https://www.researchgate.net/publica-
tion/344606640_Shared_Path_-_Service_Design_and_Artificial_Intelli-
gence_in_Designing_Human-Centred_Digital_Services, [12.08.2021].

KAVANAUGH, JEFF/TARAFDAR, RAFEE: The Live Enterprise, McGrawHill, New


York 2021.

KHAKIFIROOZ, MARZIEH/CAYARD, DIMITRI/CHIEN, CHEN FU/FATHI, MAHDI:


A System Dynamic Model for Implementation of Industry 4.0, in: 2018 Interna-
tional Conference on System Science and Engineering (ICSSE), New Taipei City
2018.

KANTER, ROSABETH: The Change Masters, Unwin Hyman, London 1985.

KLARMANN, NOAH/MALMIR, MOHAMMADHOSSEIN/JOSIFOVSKI, JOSIP: Opti-


mising Trajectories in Simulations with Deep Reinforcement Learning for Indus-
trial Robots in Automotive Manufacturing, in: VERMESAN, OVIDIU/JOHN,
REINER/DE LUCA, CRISTINA/COPPOLA, MARCELLO (Eds.), AI for Digitis-
ing Industry, River Publishers, 2021.

KOHAR, CHRISTOPHER/CONNOLLY, DANIEL/INAL, KAAN: Using AI to Aid Vehi-


cle Lightweighting in Crashworthiness with Aluminium, in: 17th International Con-
ference on Aluminium Alloys, 2020.

KOUSI, NIKI/DIMOSTHENOPOLOUS, DIMOSTHENIS/MATTHAIAKIS, ALEKSAN-


DROS-STEREOS/MICHALOS, GEORGE/MAKRIS, SOTIRIS: AI based com-
bined scheduling and motion planning in flexible robotic assembly lines, in: Proce-
dia CIRP, Vol. 86, 2019, p. 74-79.

KOZLOWSKI, ADAM/WISNIEWSKI, MARCIN: 8 Examples of How AI Drives the Au-


tomotive Industry, 2021, https://grapeup.com/blog/8-examples-of-how-ai-drives-
the-automotive-industry/, [28.11.2021].

XVIII
KPMG: Global Automotive Executive Survey, 2020,
https://home.kpmg/de/de/home/themen/2020/06/global-automotive-executive-sur-
vey-2020.html, [26.10.2020].

KPMG/EGON ZEHNER: Digital Gravity Automotive Study, 2019, https://automotive-in-


stitute.kpmg.de/EgonZehnder/executive-summary-and-lessons-learned,
[12.11.2020].

KRAUSE, ANDREAS: Trustworthy AI requires interdisciplinary collaboration,


https://ethz.ch/en/news-and-events/eth-news/news/2021/07/trustworthy-ai-requires-in-
terdisciplinary-collaboration.html, [05.08.2021].

KUHNERT, FELIX/STÜRMER, CHRISTOPH/KOSTER, ALEX: Five trends transforming


the Automotive Industry, 2018, https://www.pwc.at/de/publikationen/branchen-und-
wirtschaftsstudien/eascy-five-trends-transforming-the-automotive-industry_2018.pdf,
[02.12.2021].

KUMARI, RANI/SAINI, KAVITA: Advanced Automobile Manufacturing: An Industry 4.0,


8th International Conference on Computing for Sustainable Global Development (IN-
DIACom), New Delhi 2021, p. 899-904.

LAND STEIERMARK: Smart Production & Services: Die Steiermark ist bereit!,
https://www.politik.steiermark.at/cms/beitrag/12164148/7129389/, [20.03.2021].

LAZAREVIC, NENAD: Evolution of Time Tracking: From Punch Cards to Automatic Time
Tracking Software, 2019, https://blogs.sap.com/2019/08/27/evolution-of-time-tracking-
from-punch-cards-to-automatic-time-tracking-software/, [25.07.2021].

LEITNER, LUKAS/LAGRANGE, ANTOINE/ENDISCH, CHRISTIAN: End-of-line fault de-


tection for combustion engines using one-class classification, in: International Confer-
ence on Advanced Intelligent Mechatronics (AIM), Banff 2016, p. 207-213.

LI, CHANGMING/LI, DONGNIAN/CHEN, CHENGJUN/ZHAO, ZHENGXU/WANG,


YANG: Part Recognition and Pose Estimation Based on Convolutional Neural Net-
work, in: 2nd International Conference on Machine Learning, Big Data and Business
Intelligence (MLBDBI), Taiyuan 2020, p. 561-567.

LIKER, JEFFREY: The Toyota Way: 14 management principles from the world’s greatest
manufacturer, McGraw-Hill, New York 2004.

XIX
LIRA, CARL: Brief History of the Steam Engine, 2013,
https://www.egr.msu.edu/~lira/supp/steam/, [10.03.2021].

LIU, CHAO/VENGAYIL, HRISHIKESH/LU, YUQIAN/XU, XUN: A Cyber-Physical


Machine Tools Platform using OPC UA and MTConnect in: Journal of Manufac-
turing Systems, Vol. 51, 2019, p. 61-74.

LUCKOW, ANDRE/COOK, MATTHEW/ASHCRAFT, NATHAN/WEILL, EDWIN/DJERE-


KAROV, EMIL/VORSTER, BENNIE: Deep learning in the automotive industry: Ap-
plications and tools, in: International Conference on Big Data (Big Data), Washington
2016, p. 3759-3768.

LUCKOW, ANDRE/KENNEDY, KEN/ZIOLKOWSKI, MARCIN/DJEREKAROV,


EMIL/COOK, MATTHEW/DUFFY, EDWARD/SCHLEISS, MICHAEL/VORSTER,
BENNIE/WEILL, EDWIN/KULSHRESTHA, ANKIT/SMITH, MELISSA: AI and
Deep Learning Applications for Automotive Manufacturing, in: International Confer-
ence on Big Data (Big Data), Seattle 2018, p. 3144-3152.

LUO, REN/WANG, HAO: Diagnostic and Prediction of Machines Health Status as Exemplary
Best Practice for Vehicle Production System, in: 88th Vehicular Technology Confer-
ence (VTC-Fall), Chicago 2018, p. 1-5.

MAGNA: Complete Vehicle Manufacturing, 2021, https://www.magna.com/products/com-


plete-vehicles/complete-vehicle-manufacturing, [25.09.2021].

MAGRO, T./GHIOTTI, A./BRUSCHI, S./FERRAIULO, A: An AI approach for the in-line


evaluation of steels mechanical properties in rolling, in: Procedia CIRP, Vol. 100,
2021, p. 193-198.

MAKANTASIS, KONSTANTINOS/DOULAMIS, ANASTASIOS/DOULAMIS, NIKO-


LAOS/PSYCHAS, KONSTANTINOS: Deep learning based human behavior recogni-
tion in industrial workflows, in: International Conference on Image Processing (ICIP),
Phoenix 2016, p. 1609-1613.

MARTIN, ROGER: The High Price of Efficiency, 2019, https://hbr.org/2019/01/the-high-


price-of-efficiency, [14.03.2021].

MATA, JAVIER/DE MIGUEL, IGNACIO/DURAN, RAMON/MERAYO, NOEMI/SINGH,


SANDEPP/JUKAN, ADMELA/CHAMANIA, MOHIT: AI (AI) methods in optical
networks: A comprehensive survey, in: Optical Switching and Networking, Vol. 28,
2018, p. 43-57.

XX
MATZKA, STEPHAN: Using Process Quality Prediction to Increase Resource Efficiency in
Manufacturing Processes, First International Conference on AI for Industries, Laguna
Hills 2018, p. 110-111.

MAYR, ANDREAS/RÖLL, PHILIPP/WINKLE, DANIEL/ENZMANN, MAXIMIL-


IAN/BICKEL, BENJAMIN/FRANKE, JÖRG: Data-driven quality monitoring of
bending processes in hairpin stator production using machine learning techniques, in:
Procedia CIRP, Vol. 103, 2021, p. 256-261.

MAZZETTO, MURIEL/SOUTHIER, LUIZ/TEIXEIRA, MARCELO/CASANOVA, DALCI-


MAR: Automatic Classification of Multiple Objects in Automotive Assembly Line, in:
24th IEEE International Conference on Emerging Technologies and Factory Automa-
tion (ETFA), Zaragoza 2019, p. 363-369.

MCGIVERN, YVONNE: The Practice of Market Research, Pearson Education, 3rd Edition,
Essex 2009.

MEDEIROS, JOAO: Stephen Hawking: I Fear AI may replace humans altogether, 2017,
https://www.wired.co.uk/article/stephen-hawking-interview-alien-life-climate-change-
donald-trump, [12.08.2021].

MEINER, MORITZ/MAYR, ANDREAS/KUHN, MARLENE/RAAB, BERN-


HARD/FRANKE, JÖRG: Towards an Inline Quality Monitoring for Crimping Pro-
cesses Utilizing Machine Learning Techniques, in: 10th International Electric Drives
Production Conference (EDPC), Ludwigsburg 2020, p. 1-6.

MERCEDES-BENZ: Mercedes me, 2021, https://www.mercedes-benz.com/en/mercedes-me/,


[22.03.2021].

MERY, DOMINGO/ARTETA, CARLOS: Automatic Defect Recognition in X-Ray Testing


Using Computer Vision, in: Winter Conference on Applications of Computer Vision
(WACV), Santa Rosa 2017, p. 1026-1035.

MEYER, MICHAEL/TAN, MICHAEL/VOHRA, ROHIT/MCADOO, MICHAEL: How


ASEAN Can Move Up the Manufacturing Value Chain, 2021,
https://www.bcg.com/publications/2021/asean-manufacturing, [15.07.2021].

MHLANGA, DAVID: AI in the Industry 4.0, and Its Impact on Poverty, Innovation, Infra-
structure Development, and the Sustainable Development Goals: Lessons from

XXI
Emerging Economies?, 2021, https://res.mdpi.com/d_attachment/sustainability/sustain-
ability-13-05788/article_deploy/sustainability-13-05788.pdf, [21.03.2021].

MIAILHE, NICOLAS/HODES, CYRUS: The Third Age of AI, in: Field Action Science
Reports, Issue 17, 2017, p. 4-5.

MIT: Computer Science and AI Laboratory, 2021, http://catalog.mit.edu/mit/research/com-


puter-science-artificial-intelligence-laboratory/, [05.08.2021].

MONDEN, YASUHIRO: Toyota Production System: An integrated approach to Just-in-


time, 4th Edition, CRC Press, Boca Raton, 2012.

MOHAN, ROOSEFERT/ROSELYN, PREETHA/UTHRA, ANNIE/D., DEVARAJ/K.,


UMACHANDRAN: Intelligent machine learning based total productive mainte-
nance approach for achieving zero downtime in industrial machinery, in: Comput-
ers & Industrial Engineering, Vol. 157, 107267, 2021.

MOSNIK, LISBETH/MURHAMMER-SAS, DANIELA/PROSKE, MARKUS: AI for


Green, 2021, https://fdoc.ffg.at/s/vdb/public/node/con-
tent/anlo2weASWixNmJI14QJgA/1.0?a=true, [28.11.2021].

MURESAN, MIRCEA/CIREAP, DRAGOS/GIOSAN, ION: Automatic Vision Inspection


Solution for the Manufacturing Process of Automotive Components Through Plas-
tic Injection Molding, in: 16th International Conference on Intelligent Computer
Communication and Processing (ICCP), Cluj-Napoca 2020, p. 423-430.

MURPHY, MIKE: From dial-up to 5G: a complete guide to logging on to the internet,
2019, https://qz.com/1705375/a-complete-guide-to-the-evolution-of-the-internet/,
[22.07.2021].

NATARAJAN, RAMACHANDRAN: Total Quality Management, in: SWAMIDASS,


PAUL (Publ.): Innovations in Competitive Manufacturing, Springer, Boston 2000,
p. 69-81.

NGUYEN, HUONG/MEINERS, MORITZ/SCHMIDT, LORENZ/FRANKE, JÖRG: Deep


learning-based automated optical inspection system for crimp connections, in: 10th
International Electric Drives Production Conference (EDPC), Ludwigsburg 2020,
p. 1-7.

XXII
OECD: Recommendations of the council on AI, 2019, https://www.fsmb.org/siteassets/ar-
tificial-intelligence/pdfs/oecd-recommendation-on-ai-en.pdf, [25.07.2021].

OHNO TAIICHI: Toyota Production System: Beyond Large-Scale Production, Productiv-


ity Press, Portland 1988.

O’REILLY: Waste by Using Software as a Solution, https://www.oreilly.com/li-


brary/view/eliminating-waste-in/9781430260882/9781430260882_Ch06.xhtml,
[15.10.2021].

ORTEGA-ARANDA, DAVID/LOPEZ-JUAREZ, ISMAEL/NATH-SAHA,


BAIDYA/OSORIO-COMPARAN, ROMAN/PENA-CABRERA,
MARIO/LEFRANC, GASTON: Towards learning contact states during peg-in-hole
assembly with a dual-arm robot, in: CHILEAN Conference on Electrical, Electron-
ics Engineering, Information and Communication Technologies (CHILECON),
Pucon 2017, p. 1-6.

OVERBECK, LEONARD/HUGUES, ADRIEN/CARL MAY, MARVIN/KUHNLE, AN-


DREAS/LANZA, GISELA: Reinforcement Learning Based Production Control of
Semi-automated Manufacturing Systems, in: Procedia CIRP, Vol. 103, 2021, p.
170-175.

OXFORD LEARNER’S DICTIONARIES: Definition of rational adjective, 2021,


https://www.oxfordlearnersdictionaries.com/definition/english/rational,
[25.07.2021].

PARK, HONG/PHUONG, DANG/KUMAR, SAURABH: AI Based Injection Molding


Process for Consistent Product Quality, in: Procedia Manufacturing, Vol. 28, 2019,
p. 102-106.

PASTOR-LOPEZ, IKER/DE LA PENA SORDO, JORGE/SANTOS, IGOR/BRINGAS,


PABLO: Surface defect categorization of imperfections in high precision automo-
tive iron foundries using best crossing line profile, in: 10th Conference on Indus-
trial Electronics and Applications (ICIEA), Auckland 2015, p. 339-344.

PATALAS-MALISZEWSKA, JUSTYNA/PAJAK, IWONA/SKRZESZEWSKA, MAL-


GORZATA: AI-based Decision-making Model for the Development of a Manufac-
turing Company in the context of Industry 4.0, in: International Conference on
Fuzzy Systems (FUZZ-IEEE), Glasgow 2020, p. 1-7.

XXIII
PATIL, ADITI/GAIKWAD, JITENDRA/KULKARNI, JAYANT: Bearing fault diagnosis
using discrete Wavelet Transform and Artificial neural network, in: 2nd Interna-
tional Conference on Applied and Theoretical Computing and Communication
Technology (iCATccT), p. 248-253.

PERES, RICARDO/BARATA, J./LEITAO, PAULO/GARCIA, GISELA: Detection and


prevention of assembly defects, by machine learning algorithms, in semiconductor
industry for automotive, in: IEEE Access, Vol. 7, 2019, p. 2169-3536.

PERES, RICARDO/GUEDES, MAGNO/MIRANDA, FABIO/BARATA, JOSE: Simula-


tion-Based Data Augmentation for the Quality Inspection of Structural Adhesive
With Deep Learning, in: IEEE Access, Vol. 9, 2021, p. 76532-76541.

PICCAROZZI, MICHELA/AQUILANI, BARBARA/GATTI, CORRADO: Industry 4.0 in


Management Studies: A Systematic Literature Review, 2018,
https://www.mdpi.com/2071-1050/10/10/3821/htm, [21.03.2021].

PLATTFORM INDUSTRIE 4.0: What is Industrie 4.0?, 2021, https://www.plattform-


i40.de/PI40/Navigation/EN/Industrie40/WhatIsIndustrie40/what-is-indus-
trie40.html, [20.03.2021].

PLATTFORM INDUSTRIE 4.0: Technology Scenario: ‘AI in Industrie 4.0’, 2019,


https://www.plattform-i40.de/IP/Redaktion/EN/Downloads/Publikation/AI-in-In-
dustrie4.0.pdf?__blob=publicationFile&v=5, [12.08.2021].

POLLARD, TIM: The double-bubble bursts: only 100 Peugeot RCZ coupes left in UK,
2016, https://www.carmagazine.co.uk/car-news/industry-news/peugeot/farewell-
peugeot-rcz-final-100-coupes-sell-out-in-uk/, [26.09.2021].

PORSCHE: Marriages at Porsche, 2014, https://newsroom.porsche.com/en/company/mar-


riage-10780.html, [20.11.2021].

PORTER, MICHAEL: Competitive Advantage, Free Press, New York 2004.

POSS, CHRISTIAN/MLOUKA, ONS/IRRENHAUSER, THOMAS/PRUEGLMEIER,


MARCO/GOEHRING, DANIEL/ZOGHLAMI, FIRAS/SALEHI, VAHID: Robust
Framework for intelligent Gripping Point Detection, in: 45th Annual Conference of
the IEEE Industrial Electronics, Lisbon 2019, p. 717-723.

XXIV
PRAUSE, GUNNAR/ATARI, SINA: On sustainable production networks for Industry 4.0,
2018, https://hal.archives-ouvertes.fr/hal-01860909/document, [21.03.2021].

PURSSELL, EDWARD/MCCRAE, NIALL: How to Perform a Systematic Literature Re-


view, Springer, Cham 2020.

PWC: Why Most Organizations’ Investments in AI Fall Flat, 2020, https://hbr.org/spon-


sored/2020/10/why-most-organizations-investments-in-ai-fall-flat, [02.11.2021].

QIAN, YU/ARINEZ, JORGE/XIAO, GUOXIAN/CHANG, QING: Improved Production


Performance Through Manufacturing System Learning, in: 15th International Con-
ference on Automation Science and Engineering (CASE), Vancouver 2019, p. 517-
522.

QIAN, YANJUN/KIM, JONGMUN/KWON, HYOCK-JU: Development of Hybrid AI


model for Car Steering Shaft Assembly by Combining Gaussian Process Regres-
sion and Artificial neural network, in: Canadian Conference on Electrical and Com-
puter Engineering (CCECE), 2021, p. 1-5.

QUEIROZ, JONAS/LEITAO, PAOLO/BARBOSA, JOSE/OLIVEIRA, EU-


GENIO/GISELA, GARCIA: Agent-Based Distributed Data Analysis in Industrial
Cyber-Physical Systems, in: Journal of Emerging and Selected Topics in Industrial
Electronics, Vol. 3, Issue 1, 2021, p. 5-12.

QUINTON, SARAH/SMALLBONE, THERESA: Postgraduate Research In Business: A


Critical Guide, SAGE Publications Ltd., London 2006.

RAE, JOHN: Ford and the assembly line, 2021, https://www.britannica.com/technol-


ogy/automotive-industry/Ford-and-the-assembly-line, [29.09.2021].

RAHIMI, AFSHIN/ANVARIPOUR, MOHAMMAD/HAYAT, KHIZER: Object Detec-


tion using Deep Learning in a Manufacturing Plant to Improve Manual Inspection,
in: International Conference on Prognostics and Health Management (ICPHM), De-
troit 2021, p. 1-7.

REN, JING/REN, RUI/GREEN, MARK/HUANG, XISHI: Defect Detection from X-Ray


Images Using A Three-Stage Deep Learning Algorithm, in: Canadian Conference
of Electrical and Computer Engineering (CCECE), Edmonton 2019, p. 1-4.

XXV
ResearchGate: About ResearchGate, 2021, https://www.researchgate.net/about,
[25.03.2021].

RIO-TORTO, ISABEL/CAMPANICO, ANA/PEREIRA, ANTONIO/TEIXEIRA,


LUIS/FILIPE, VITOR: Automatic quality inspection in the automotive industry: a
hierarchical approach using simulated data, in: 8th International Conference on In-
dustrial Engineering and Applications (ICIEA), Chengdu 2021, p. 342-347.

ROCKWELL, ANYOHA: The History of AI, 2017, https://sitn.hms.har-


vard.edu/flash/2017/history-artificial-intelligence/, [23.07.2021].

RUDDICK, GRAHAM: VW supplier clash stops output at six plants, 2016,


https://www.theguardian.com/business/2016/aug/22/volkswagen-supplier-clash-
stops-manufacturing-at-six-plants, [29.09.2021].

RUIZ, FRANCISCO/AGELL, NURIA/ANGULO, CECILIO/SANCHEZ, MONICA: A


learning system for adjustment processes based on human sensory perceptions, in:
Cognitive Systems Research, Vol. 52, 2018, p. 58-66.

RÜSSMANN, MICHAEL/LORENZ, MARKUS/GERBERT, PHILIPP/WALDNER, MA-


NUELA/ENGEL, PASCAL/HARNISCH, MICHAEL/, JUSTUS, JAN: Industry
4.0: The Future of Productivity and Growth in Manufacturing Industries,
https://www.bcg.com/publications/2015/engineered_products_project_business_in-
dustry_4_future_productivity_growth_manufacturing_industries, [20.03.2021].

RUSSELL, STUART/NORVIG, PETER: Artificial Intelligence: A Modern Approach, 3rd


edn., Prentice Hall, Harlow 2016.

RUSSELL, STUART/NORVIG, PETER: AI: A Modern Approach, 4th edn., Prentice Hall,
Harlow 2021.

RWTH Aachen: Center für Künstliche Intelligenz, https://www.ai.rwth-


aachen.de/cms/KI/~fsfai/Das-KI-Center/, [05.08.2021].

SALAZAR, LUIS ALBERTO/RYASHENTSEVA, DARIA/LÜDER, ARNDT/VOGEL-


HEUSER, BIRGIT: Cyber-physical production systems architecture based on
multi-agent’s design pattern – comparison of selected approaches mapping four
agent patterns, in: The International Journal of Advanced Manufacturing Technol-
ogy, Vol. 105, 2019, p. 4005-4034.

XXVI
SALCEDO-HERNANDEZ, JAVIER/GARCIA-BARRUETABENA, JON/PASTOR-
LOPEZ, IKER/SANZ-URQUIJO, BORJA: Predicting Enamel Layer Defects in an
Automotive Paint Shop, in: IEEE Access, Vol. 8, 2020, p. 22748-22757.

SAUNDERS, MARK/LEWIS, PHILIP/THORNHILL, ADRIAN: Research methods for


business students, 5th Edition, Pearson Education Limited, Harlow 2009.

SCHUH, G./GÜTZLAFF, A./THOMAS, K./WELSING, M.: Machine learning based de-


fect detection in a low automated assembly environment, in: Procedia CIRP, Vol.
104, 2021, p. 265-270.

SCHWICKERT, AXEL/OSTHEIMER, BERNHARD/EROGLU, MURAT: Kaizen, TPM,


Lean – Grundlagen, Abgrenzung, Zusammenhänge, 2011, https://ilias.uni-gies-
sen.de/ilias/goto.php?target=file_183959_download&client_id=JLUG,
[21.11.2021].

SEGEV, ARIE/PORRA, JAANA/ROLDAN, MALU: Internet-based EDI strategy, in: De-


cision Support Systems, Vol. 21, Issue 3, 1997, p. 157-170.

SEMENIUTA; OLEKSANDR/DRANSFELD, SEBASTIAN/FALKMAN, PETTER: Vi-


sion-based robotic system for picking and inspection of small automotive compo-
nents, in: International Conference on Automation Science and Engineering
(CASE), Fort Worth 2016, p. 549-554.

SENGE, PETER: The fifth discipline: the art and practice of the learning organization.,
Random House Business, London 2006.

SERIN, GOKBERK/SENER, BATIHAN/GUDELEK, UGUR/OZBAYOGLU, MU-


RAT/UNVER, HAKKI: Deep Multi-Layer Perceptron based Prediction of Energy
Efficiency and Surface Quality for Milling in The Era of Sustainability and Big
Data, Procedia Manufacturing, Vol. 51, 2020, p. 1166-1177.

SHANI, OR: From Science Fiction to Reality: The Evolution of AI, 2015,
https://www.wired.com/insights/2015/01/the-evolution-of-artificial-intelligence/,
[23.07.2021].

SIEGEL, ETHAN: What was it like when the first humans arose on earth, 2019,
https://www.forbes.com/sites/startswithabang/2019/05/15/what-was-it-like-when-the-
first-humans-arose-on-earth/?sh=6f9e2f926997, [23.03.2021].

XXVII
SIEPMANN, DAVID: Industrie 4.0 – Struktur und Historie, in: ROTH, Armin (Publ.):
Einführung und Umsetzung von Industrie 4.0, 1st Edition, Springer-Verlag, Berlin
Heidelberg 2016, p. 17-35.

SIEPMANN, DAVID: Industrie 4.0 – Fünf zentrale Paradigmen, in: ROTH, Armin
(Publ.): Einführung und Umsetzung von Industrie 4.0, 1st Edition, Springer-Verlag,
Berlin Heidelberg 2016, p. 35-47.

SIEPMANN, DAVID: Industrie 4.0 – Technologische Komponenten, in: ROTH, Armin


(Publ.): Einführung und Umsetzung von Industrie 4.0, 1st Edition, Springer-Verlag,
Berlin Heidelberg 2016, p. 47-73.

SIMONITE, TOM: Moore’s Law is Dead: Now What?, 2016, https://www.technolo-


gyreview.com/2016/05/13/245938/moores-law-is-dead-now-what/, [24.07.2021].

SINHA, SUMIT/FRANCIOSA, PASQUALE/CEGLAREK, DARIUSZ: Building a Scala-


ble and Interpretable Bayesian Deep Learning Framework for Quality Control of
Free Form Surfaces, in: IEEE Access, Vol. 9, 2021, p. 50188-5028.

SINHA, SUMIT/FRANCIOSA, PASQUALE/CEGLAREK, DARIUSZ: Object Shape Er-


ror Correction using Deep Reinforcement Learning for Multi-Station Assembly
Systems, in: 19th International Conference on Industrial Informatics (INDIN),
Palma de Mallorca 2021, p. 1-8.

SOCHA, GEORGE: What Is An AI Model?, 2021, https://resource.reveal-


data.com/en/blog/what-is-an-ai-model, [28.10.2021].

STATISTA: Revenue of leading automakers worldwide in 2020, 2021, https://www.sta-


tista.com/statistics/232958/revenue-of-the-leading-car-manufacturers-worldwide/,
[24.09.2021].

STATISTA: Number of cars sold worldwide between 2010 and 2021, 2021,
https://www.statista.com/statistics/200002/international-car-sales-since-1990/,
[26.09.2021].

STANFORD UNIVERSITY: Robotics: A Brief History, 1998, https://cs.stanford.edu/peo-


ple/eroberts/courses/soco/projects/1998-99/robotics/history.html, [15.03.2021].

STONE, PETER/ BROOKS, RODNEY/BRYNJOLFSSON, ERIK/CALO, RYAN/


ETZIONI, OREN/HAGER, GREG/HIRSCHBERG,

XXVIII
JULIA/KALYANAKRISHNAN, SHIVARAM/KAMAR, ECE/KRAUS, SA-
RIT/LEYTON-BROWN, KEVIN/PARKES, DAVID/PRESS, WIL-
LIAM/SAXENIAN, ANNALEE/SHAH, JULIE/TAMBE, MILIND/TELLER, AS-
TRO: AI and Life in 2030." One Hundred Year Study on AI: Report of the 2015-
2016 Study Panel, Stanford University, Stanford, https://ai100.stan-
ford.edu/sites/g/files/sbiybj18871/files/media/file/ai100report10032016fnl_sin-
gles.pdf, 2016, [05.08.2021].

SUNDARAVARADAN, NAREN/MARWAH MANISH/SHAH, AMIP/RAMAKRISH-


NAN, NAREN: Data Mining Approaches for Life Cycle Assessment, 2011,
https://www.researchgate.net/publication/224244459_Data_mining_ap-
proaches_for_life_cycle_assessment, [04.10.2021].

SWAIM, ROBERT: The Strategic Drucker, John Wiley & Sons, Singapore 2010.

SWAINE, MICHAEL: ENIAC computer, 2021, https://www.britannica.com/technol-


ogy/ENIAC, [16.03.2021].

SYSTEM DYNAMICS SOCIETY: What is System Dynamics?, 2021, https://systemdy-


namics.org/what-is-system-dynamics/, [15.12.2021].

TAULLI, TOM: How To Create an AI (AI) Model, 2020,


https://www.forbes.com/sites/tomtaulli/2020/07/11/how-to-create-an-ai-artificial-
intelligence-model/?sh=4dd3e4fb7a97, [28.10.2021].

TEKINERDOGAN, BEDIR: Engineering Connected Intelligence – A Socio-Technical


Perspective, 2017, https://edepot.wur.nl/401115, [30.10.2021].

TELUKTARIE, ARNESH/SISHI, M. N.: Enterprise Definition for Industry 4.0, 2018,


https://ieeexplore.ieee.org/abstract/document/8607642/keywords#keywords,
[21.03.2021].

TIBKEN, SHARA: CES 2019: Moore’s Law is dead, says Nvidia’s CEO, 2019,
https://www.cnet.com/tech/computing/moores-law-is-dead-nvidias-ceo-jensen-
huang-says-at-ces-2019/, [24.07.2021].

TIDD, JOE/BESSANT, JOHN/PAVITT, KEITH: Managing Innovation, 3rd Edition., John


Wiley & Sons, West Sussex 2005.

XXIX
TOYOTA: History of Toyota, 2021, https://global.toyota/en/company/trajectory-of-
toyota/history/, [24.09.2021].

TOYOTA: Toyota Production System, 2021, https://global.toyota/en/company/vision-and-


philosophy/production-system/, [25.09.2021].

TOYOTA: Production Engineering, 2021, https://www.toyota-global.com/company/his-


tory_of_toyota/75years/data/automotive_business/production/production_engineer-
ing/prize_receiving/index.html, [24.09.2021].

TOYOTA: How many parts is each car made of?, 2021,


https://www.toyota.co.jp/en/kids/faq/d/01/04/, [24.09.2021].

TURING, ALAN: Computing Machinery and Intelligence, 1950,


https://www.csee.umbc.edu/courses/471/papers/turing.pdf, [23.07.2021].

VAN DUIN, STEFAN/BAKSHI, NASER: AI Defined: The most used terminology around
it, 2017, https://www2.deloitte.com/se/sv/pages/technology/articles/part1-artificial-
intelligence-defined.html, [12.08.2021].

VATER, JOHANNES/HARSCHEIDT, LARS/KNOLL, ALOIS: Smart Manufacturing


with Prescriptive Analytics, in: 8th International Conference on Industrial Technol-
ogy and Management (ICITM), Cambridge 2019, p. 224-228.

VATER, JOHANNES/HARSCHEIDT, LARS/KNOLL, ALOIS: Closing the loop: Real-


time Error Detection and Correction in automotive production using Edge-/Cloud-
Architecture and a CNN, in: International Conference on Omni-layer Intelligent
Systems (COINS), Barcelona 2020, p. 1-7.

VDI Nachrichten: Industrie 4.0: Mit dem Internet der Dinge auf dem Weg zur 4. Industriel-
len Revolution, 2011, http://www.vdi-nachrichten.com/artikel/Industrie-4-0-Mit-
dem-Internet-der-Dinge-auf-dem-Weg-zur-4-industriellen-Revolution/52570/1,
[21.03.2021].

VLASKOVITS, PATRICK: Henry Ford, innovation and that faster horse quote, 2011,
https://hbr.org/2011/08/henry-ford-never-said-the-fast, [14.03.2021].

WADHWA, VIVEK: The amazing AI we were promised is coming, finally, 2016,


https://www.washingtonpost.com/news/innovations/wp/2016/06/17/the-amazing-
artificial-intelligence-we-were-promised-is-coming-finally/, [23.07.2021].

XXX
WAGNER, RAPHAEL/FISCHER, JAKOB/GAUDER, DANIEL/HAEFNER, BENJA-
MIN/LANZA, GISELA: Virtual In-line Inspection for Function Verification in Se-
rial Production by means of AI, in: Procedia CIRP, Vol. 92, 2020, p. 63-68.

WAN, JIAFU/DAI, HONG-NING/KUSIAK, ANDREW/MARINEZ-GARCIA, MI-


GUEL/LI, DI: Artificial-Intelligence-Driven Customized Manufacturing Factory:
Key Technologies, Applications and Challenges, in: Proceedings of the IEEE, Vol.
109, Issue: 4, 2021, p. 377-398.

WANG, WEITIAN/LI, RUI/DIEKEL, MAX/JIA, YUNYI: Facilitating Human–Robot


Collaborative Tasks by Teaching-Learning-Collaboration From Human Demonstra-
tions, in: Transactions on Automation Science and Engineering, Vol. 16, Issue 2,
2019.

WAYLAND, MICHAEL: Chip shortage expected to cost auto industry $210 billion in rev-
enue in 2021, https://www.cnbc.com/2021/09/23/chip-shortage-expected-to-cost-
auto-industry-210-billion-in-2021.html, [25.09.2021].

WEDENIWSKI, SEBASTIAN: The Data-defined Vehicle Architecture, 2016,


http://mpsoc-forum.org/archive/2016/slides/Sebastian_Wedeniwski.pdf,
[30.09.2021].

WESCOAT, ETHAN/KRUGH, MATTHEW/HENDERSON, ANDREW/GOODNOUGH,


JOSH/MEARS, LAINE: Vibration Analysis Utilizing Unsupervised Learning, in:
Procedia Manufacturing, Vol. 34, 2019, p. 876-884.

WIDMER, TOBIAS/KLEIN, ACHIM/WACHTER, PHILIPP/MEYL, SEBASTIAN: Pre-


dicting Material Requirements in the Automotive Industry Using Data Mining,
22nd International Conference on Business Information Systems, SPRINGER-
VERLAG, Seville 2019, p. 147-161.

WILSON, DAVID: Managing Information, 2nd edition, Butterworth-Heinemann, Oxford


2000.

WINGETT; STEVEN: Magna Steyr wins big with Peugeot RZT, 2011, https://europe.au-
tonews.com/article/20110223/CUTAWAY01/110229974/magna-steyr-wins-big-
with-peugeot-rcz, [26.09.2021].

XXXI
XUE-BO, JIN/TING-LI, SU/JIAN-LEI, KONG/YU-TING, BAI/BEI-BEI, MAIO/DOU,
CHAO: State-of-the-Art Mobile Intelligence: Enabling Robots to Move Like Hu-
mans by Estimating Mobility with AI, in: Applied Sciences, Vol. 8, Issue 3, 2018.

ZEISER, ALEXANDER/VAN STEIN, BAS/BÄCK, THOMAS: Requirements towards


optimizing analytics in industrial processes, Procedia Computer Science, Vol. 184,
2021, p. 597-605.

ZHANG, DARUI/XU, BIN/WOOD, JASMINE: Predict failures in production lines: A


two-stage approach with clustering and supervised learning, International Confer-
ence on Big Data (Big Data), Washington DC 2016, p. 2070-2074.

ZHANG, JIAXU/HU, SHAOLIN/SHI, HAOQIANG: Visual Detection System of Auto-


motive Parts Attitude Based on Deep Learning, 8th Data Driven Control and Learn-
ing Systems Conference (DDCLS), Dali 2019, p. p. 918-922.

ZHANG, ANYI/WANG, JIANWEN: A method for analyzing abnormality of automobile


sunroof manufacturing process by using Bayesian method, International Confer-
ence on Information Technology, Big Data and AI (ICIBA), Chongqing 2020, p.
233-237.

ZHOU, JIAN/WANG, DALI/CHEN, JIAN/FENG, ZHILI/CLARSON,


BLAIR/BASELHUHN, AMBERLEE: Autonomous non-destructive evaluation of
resistance spot welded joints, in: Robotics and Computer-Integrated Manufacturing,
Vol. 72, 102183, 2021.

ZIELINSKI, KALLIL/HENDGES, LUCAS/FLORINDO, JAO/LOPES, YURI/RIBEIRO,


RICHARDSON/TEIXEIRA, MARCELO, CASANOVA, DALCIMAR: Flexible
control of Discrete Event Systems using environment simulation and Reinforce-
ment Learning, in: Applied Soft Computing, Vol. 111, 107714, 2021.

ZÜHLKE, DETLEF/OLLINGER, LISA: Advances in Automation and Robotics: Agile


Automation Systems Based on Cyber-Physical Systems and Service-Oriented Ar-
chitectures, Springer-Verlag, Berlin Heidelberg 2011.

XXXII

Das könnte Ihnen auch gefallen