AI AS AN ENABLER FOR LONG-TERM

RESILIENCE IN MANUFACTURING

Group Leader
Erwin Rauch
Assistant Professor in Manufacturing Systems, Head Smart Mini
Factory Lab for Industry 4.0, Free University of Bolzano

October 2021
CONTRIBUTORS

Tunç Acarkan Gisela Lanza

Technology Management Director, MESS Turkish Full Professor Institute of Production Science,
Employers’ Association of Metal Industries Karlsruhe Institute for Technology (KIT)

Jesús Alonso Oscar Lazaro

Senior R&D Consultant, Innovalia Association Managing Director, Innovalia Association

Fazel Ansari Irene Sterian

Assistant Professor and Head of Research Group President & CEO, REMAP Network Canada: Director
Smart and Knowledge-Based Maintenance, TU Wien of Technology and Innovation, Celestica Inc.

Ragu Athinarayanan Giacomo Tavola

Professor of Engineering Technology, Purdue Adjunct Professor, I4.0Lab@SOM Politecnico di
University, USA Milano Technology Counselor

James Balzary Simon Thevenin

Co-Founder and CEO, TilliT Assistant Professor, Department of Automation,
Production and Computer Sciences, IMT Atlantique
Giuseppe Biffi
Business Development, Siemens Digital Enterprise Raphael Vallazza
Founder and CEO, Endian, Secure Digital Platform
Michele Ermidoro for IoT
Co-Founder and Senior Partner, AISent
Aoife Doyle
Niclas Eschner Commercialisation Researcher,
Group Leader at Institute of Production Science, I-Form Advanced Manufacturing Research Centre
Karlsruhe Institute for Technology (KIT) Member, Young Manufacturing Leaders;

Emmanuel Francalanza Xin Shen

Head, Department of Industrial and Manufacturing Student RWTH Aachen
Engineering, University of Malta Member, Young Manufacturing Leaders;
This whitepaper, published in October 2021, is part of the “Back to the Future: Emerging Topics for Long-Term Re-
silience in Manufacturing” initiative, promoted by the World Manufacturing Foundation, a non-profit organisation
with a mission to spread industrial culture worldwide. The initiative involved global focus groups, each exploring a
relevant theme for building a resilient manufacturing sector. Each focus group developed a whitepaper identifying
key propositions to enable the manufacturing community to thrive in the long term.

The views and opinions expressed by whitepaper contributors are given in their personal capacity and do not nec-
essarily reflect the views of the organisations for which they work or committees of which they are members.

For more information on the project and to read other topic-focused whitepapers that are part of the initiative,
please visit https://worldmanufacturing.org/report/back-to-the-future-emerging-topics-for-long-term-resil-
ience-in-manufacturing/
INTRODUCTION CONTEXT

Artificial Intelligence (AI) will increase the level of To significantly increase the long-term resilience of
intelligence in the manufacturing industry by pro- manufacturing companies, the following eight chal-
moting, inter alia, the matching of production and lenges have been identified and must be overcome.
demand, improving quality inspection, increasing
product yield, reducing product failure rates, and im- Complementing employees’ skills in manufacturing
proving production efficiency.1 While the last decade processes
of Industry 4.0 was determined by technology-driven The implementation of AI tools and Industrial IoT so-
innovation, the coming years will focus on data- and lutions on the manufacturing shop floor are aimed at
intelligence-driven innovation. In this perspective, AI automating tasks which are currently predominantly
is an enabler for the transition from smart factories carried out by human operators. Data analysis arising
towards intelligent factories with self-optimising and from quality control, visual inspection, and maintenan-
self-healing characteristics. While smart factories are ce operations are examples of tasks that are increa-
capable of applying previously acquired knowledge, singly being automated. Machine learning and data
intelligent factories will be able to autonomously ac- availability is now allowing for the possibility of analy-
quire new knowledge and apply it for self-optimisa- sing and correlating data, which was previously not
tion purposes.2 possible due to limitations of processing and human
cognition. New methods of human-robot interaction
The 2020 World Manufacturing Report: Manufactu- through AI-powered assistants are also becoming
ring in the Age of Artificial Intelligence has already a reality. As a result, optimisation of processes and
reported on the potential of AI in manufacturing. AI manufacturing systems is increasingly possible. This
applications impact the resilience, efficiency, and leads to an increased efficiency and effectiveness in
scalability of manufacturing operations and have be- manufacturing, making the industry more resilient.
come increasingly significant during recent disrup-
ting events like the current COVID-19 pandemic3 or Increasing corporate agility and reducing procure-
the blockage of the Suez Canal4. Resilience is defined ment bottlenecks
as the ability of a system to withstand potentially hi- Companies adopting agile supply chains and reconfi-
gh-impact disruptions and is characterised by the abi- gurable manufacturing systems develop a substantial
lity of a system to proactively mitigate or absorb the competitive advantage. However, these production
impact of disruptions, and quickly recover to normal systems are hard to manage because of the small ba-
conditions.5 According to several scholars, Industry tch sizes, numerous shop floor configurations, signi-
4.0 and AI play a major role in achieving more resilient ficant variability, and inaccurate predictability of para-
factories and circular value chains.6,7 meters (demands, lead times, capacities, etc). In this
context, planning and scheduling based on simple ru-
The aim of this whitepaper is to develop recommen- les and human intuition lead to suboptimal decisions,
dations for the adoption of AI in factories based on and do not provide the reactivity required to deal with
identified challenges, for long-term resilience in cogni- uncertain events. Intelligent planning algorithms may
tive manufacturing. automatically prescribe optimised production plans
and schedules. These tools can also provide robust
and flexible plans by taking advantage of the massive
amount of data generated on the shop floor.
Reducing time to market major emerging challenges are:
Product life cycles are becoming shorter, while the de-
mand for customised products is increasing. This in- 1. multi-dimensional, non-linearity modes in deci-
creases both product-side and production-side com- sion models that are influenced by multi-structu-
plexity.8 In particular, the so-called time-to-market is red, multi-channel data sources
an essential factor for the market success of a new 2. unanticipated variability of decision factors and
product in order to be able to withstand the strong preferences over time
competitive pressure. In today’s company practice, 3. undiscovered interdependencies and uninterpre-
knowledge in the form of existing CAD design mo- table semantics among predefined decision pre-
dels or established design procedures is often not yet ferences and external influential factors
used systematically.9 Project knowledge is thereby at- 4. a lack of industry-related methodologies for the
tached to specific employees.10 The question of how incorporation and explication of human-specific,
the implicit knowledge of existing solution principles experiential knowledge to feed KPIs, and
can be formalised in order to use it systematically and 5. a lack of trustworthy AI due to sophisticated, but
pass it on to new generations of engineers and pro- for humans inexplicable, models and algorithms.
ducts in an automated way, leads to a new type of
feedback interaction between humans and AI. Intelligent risk assessment and mitigation
Today’s manufacturing enterprises employ producti-
Strengthening factory resilience and robustness vity-related KPIs to shape the foundations of strategic
The ability to produce the same quality and quantity and operational decisions. The specific, non-integrati-
of product when faced with unexpected disruption is ve analytical models used for feeding these KPIs do
a requirement for improved business agility and ra- not allow for a comprehensive, anticipative identifica-
pidly changing needs. Resilient manufacturing, with tion of risks, nor do they allow for a definition of actio-
the ability to quickly adapt to disruption, has become nable measures to ensure stability and profitability in
increasingly important after challenges experienced production, which would ensure resistance towards
during the COVID-19 pandemic. According to a survey market volatility during unexpected societal and eco-
conducted by the Capgemini Research Institute, 68% nomic crises. Data-driven risk assessment and mi-
of the surveyed manufacturing enterprises repor- tigation approaches should be introduced which i)
ted that it had taken 3 months to recover from sup- predict the economic impact of productivity KPIs on
ply chain changes that occurred due to COVID-19.11 business resilience, and ii) prescribe plausible measu-
Reducing machine downtime through AI can provide res for quick adaptation to disruptions, while maintai-
flexible resource management and keep the losses ning continuous operations and safeguarding human
at a minimum, demonstrating the potential of AI to resources, assets and brand quality.
strengthen factory resilience and robustness.
Reducing cybersecurity risks
Reducing complexity in data-based With the increasing local and global connectivity of
decision-making factories, the danger of cyberattacks poses major ri-
Currently, the success of AI-enhanced decision-ma- sks, such as operational downtime of production lines
king approaches largely depends on such data pro- or the theft/loss of strategic corporate information.
perties as: the availability, quality, consistency, validity, AI can be used for intrusion detection by passively
up-to-dateness, completeness and comprehensive- monitoring the entire network and identifying ano-
ness of industrial data spaces. In future, the integra- malies or threats. Similarly, the AI-based Deep Packet
tion of AI technologies in manufacturing will result in Inspection method examines data packets passing
additional complexities due to model properties. The through firewalls, searching for potential non-com-
pliant traffic, viruses or spam to determine whether a facturing, factory maintenance, big data and cyberse-
data packet may pass or not. It is neither sensible nor curity, while also addressing ethical aspects.
economically advantageous to send all the data obtai-
ned at edge-computing level to a higher-level platform AI in product design
for analysis purposes. AI is therefore increasingly re- AI design assistants are able to identify patterns in
quired at the edge to select and filter the relevant data, the human design process and contribute towards
transmitting fewer corporate data over the Internet the prediction of necessary design steps, while ensu-
and at the same time, saving bandwidth and reducing ring that the duplication of previous design develop-
latency times. The correctness of the data is a funda- ments is avoided by identifying similar components in
mental prerequisite for correct AI decisions. For this a database to the one under construction.13
reason, both the machines at the edge and the tran-
smitted data must be effectively protected against all AI assistant systems for product design rely on a lar-
types of cyberattack. IoT security gateways, VPN and ge number of CAD design data-sets being accumula-
secure digital platforms involving federated learning ted, and used as training data. Using state-of-the-art
or differential privacy12 are needed to meet this chal- machine learning algorithms, design assistants are
lenge. able to detect patterns in the human design process
and can predict the next construction step.13 These
Addressing ethics when using AI in manufacturing kinds of AI assistants see humans as the ultimate de-
As the manufacturing industry is pushing the boun- cision maker. Augmented intelligence is created when
daries of efficiency and productivity, the use of AI is the adaptive system is able to continuously improve
becoming more pervasive, and investment in this te- the model output based on the inclusion of a human’s
chnology is bringing new value alongside new challen- acceptance or rejection of a forecast.14
ges. As AI is becoming more mainstream, negligent
applications of this technology could lead to proble- The aspect of AI assistant systems functionally as-
matic outcomes, including the reluctance to accept sisting humans by compensating for human disad-
or use it. Principles and values of trustworthy AI, that vantages, such as being unable to deal with unmana-
guide the ethical development and deployment of AI geable amounts of historical design data, should be
within the specific context of its application to manu- emphasised at this point. AI systems aim to resolve
facturing, should be developed - not only to the end the discrepancy between continuously developing
product or services, but rather to cover the entire va- design requirements and the individual employee’s
lue chain from design and engineering, planning, sup- skill level.15 Furthermore, such AI assistance systems
ply chain management, factory automation, IoT and used in design should be based on understandable AI
the workforce. concepts to maximise user confidence, as most ma-
chine-learning methods are described as black-box
models.16

OPPORTUNITIES AI design assistants significantly influence the know-

AND RECOMMENDATIONS ledge-intensive work of design engineers. Here, both
quantitative and qualitative advantages arise. Machi-
To overcome the identified challenges and to fully ne- learning models are able to detect design patterns
exploit the opportunities of AI as an enabler for long- from the data of existing product designs and thus
term resilience in manufacturing, we have formulated contribute to efficient knowledge management. This
the following key recommendations for AI in the are- will reduce the number of design duplicates and ul-
as of product design, planning and scheduling, manu- timately the time for product development. Further-
more, design assistants enable humans to focus on manufacturing. Novel AI-predictive approaches em-
the creative part of the product development process. power the industry to use past observations to build
Humans and AI systems can work together in fun- a resilient manufacturing system. Such approaches
ction-enhancing interaction. reduce uncertainty and risk.

Generative design algorithms based on AI techniques While predictive analytics aims to forecast what will
have constrained optimisation capabilities within a happen, prescriptive analytics considers various fu-
CAD design environment such that they bring signifi- ture scenarios and prescribes the optimal production
cant improvements and speed to the product design schedule/production plan to the production manager.
life cycle. This approach can be made more general, by Suggestions are made based on mathematical re-
applying it to any type of model (not exclusively CAD), presentations and simulation of the manufacturing
i.e. any digital twin based on a non-geometric repre- systems. To provide robust and flexible plans, these
sentation of reality. Generative design algorithms al- prescriptive analytics models may account for the un-
low designers to work much more quickly than would certainty in the system by using scenario samples or
otherwise be possible. Generative design produces uncertainty sets. Some decisions are selected to be
a large amount of design configurations, which tra- robust and applicable to all scenarios, whereas ‘wait-
ditionally require long development times and a high and-see’ recourse actions are made for each scenario
level of personnel resources. Consequently, these individually to allow flexibility when reacting to proba-
design configurations require less testing, as they are bilistic events. Recourse actions may include express
already in accordance with safety and general usabi- deliveries of components from local suppliers, sub-
lity standards. Moreover, AI can be of use for sorting contracting, temporary labour, or the reallocation of
and selecting the most promising/worthwhile design reserved components to end-items that can no longer
configurations for further development. The use of be manufactured.
generative design in the product design process can
significantly reduce the costs associated with the tra- AI can also help to implement the production sche-
ditional design process, while also enabling a faster dule on the shop floor. Schedule conformance re-
launch of products to market. mains elusive in most manufacturing environments,
and the shift to more responsive supply chain ope-
AI in Planning and Scheduling rating models increases the likelihood of adherence
Production planning and scheduling aims to allocate issues. Challenges surrounding scheduling and exe-
the capacity of equipment and personnel by balan- cution can be overcome by improving upon current,
cing operational efficiency and cost. The use of AI in traditional methods of data synchronisation. Tools
planning and scheduling includes predictive analytics, to dynamically adapt to changes leverage an expli-
prescriptive analytics, and real-time rescheduling. cit real-time rescheduling and feedback connection
with shop floor processes. They can highlight areas
Predictive analysis in production planning concentra- of risk, increase efficiency, and ultimately repair the
tes on the pattern extraction from data-sets to fore- schedule. AI-enabled prediction of disturbances from
cast planning and scheduling parameters. Various pa- process monitoring can be used to amend the sche-
rameters can be estimated using statistical models dule domain model. These may include altering equi-
and forecasting methods, e.g., demand, lead time, pment process rates, varying operator availability or
process duration, capacity, yield, machine downtime, constraints, and triggering a continuous optimisation
etc. Based on the predicted outcomes, managers can process running at scale in near real-time that crea-
easily steer their planning process to enhance flexi- tes a range of alternate schedules. These potential
bility, responsiveness, robustness and productivity in countermeasures can then either be suggested to de-
cision-makers or automatically applied to adjust the developed for robust learning which allow closed-loop
schedule for a range of objective outcomes. digital twin control and monitoring systems to correct
the problem in real time. There is also a need for the
While planning and scheduling applications of tru- democratisation of AI tools which are independent of
stworthy AI do not impact fundamental human particular platforms, as well as the development of
autonomy or organisational democracy, ensuring data standards, which will allow for industry to form a
understandable AI can foster the adoption of these shared understanding.
techniques in planning and scheduling. By providing
human-readable decision criteria and outcomes, Digital Twin technology supports product or machine
users and management can increase trust levels to designers by providing valuable insights into how the
transition to automated decision-making. Particular product or machine will look and perform in the real
attention should be paid to technical robustness, data world. The development of the IPC-2551 standard
governance, privacy, and accountability. establishes the IPC Digital Twin, which comprises the
Digital Twin Product, Digital Twin Manufacturing, and
AI in Manufacturing Digital Twin Lifecycle frameworks.17 This standard
Quality control is required to detect specific anomaly enables any manufacturer, design organisation or so-
events and drifting processes, resulting in higher lution provider to initiate application interoperability to
yields, reduced costs, and thus, increased efficiency. create smart value chains, as well as the mechanism
Technological innovation is leading the way towards to assess their current IPC Digital Twin readiness le-
100% inspections being fully online. This is enabled by vel. The power and flexibility of the Digital Twin ex-
advances in process data acquisition capabilities, vi- tends further than just a monitoring solution towards
sion systems, and new machine-learning algorithms. real-time, closed-loop control during production. This
One of the biggest challenges in this field is switching is enabled by tools such as the Live Twin technology
from defect detection to anomaly detection. Quality and Virtual Sensors, which complement real, physical
inspection systems must be able to detect defects ne- sensors for optimal process control. For many years,
ver seen before. These events may be less represen- we have been using data to model processes, ma-
ted in training data-sets collected from the field. New chines and manufacturing plants. Now, the latest AI
approaches are therefore required that are capable of methods allow us to combine machine learning with
generalising and applying machine-learning models analytical models. This works very well in applications
learned from similar equipment or utilising conditions in which an initial problem is divided into an AI task
which can complement existing data with synthetised and a simulation task that, together, deliver one result.
data or similar data sources. Furthermore, these ap- The reliability matches that of a simulation, but with
proaches need to be linked to a root cause analysis, improved speed. The Digital Twin should also be a
which correlates anomalies to production process gateway for the consumer to gain information on the
changes, to identify which process variable is the one materials used within a product design, and on pro-
negatively affecting the output quality. Generating la- duct sustainability. This will open up a whole body of
belled data is essential in this regard. Efficient data evidence and trust that helps consumer confidence,
collection needs to be integrated with human expe- but which cannot be navigated without the support of
rience, and natural language processing capabilities machine learning and AI.
(e.g. reports or even operator experience), as well as
knowledge discovery by information extraction from Advances in the field of collaborative robotics are
unstructured data (e.g. existing documentation, blue- enabling new and augmented manufacturing work-
prints, pictures or written information). Sophisticated, spaces. Workers are operating in the same environ-
but lightweight, deep- learning methods need to be ment as robots, without restrictions or safeguards.
This requires developments from a safety and secu- and prediction, to reduce the likelihood or frequency
rity standpoint, as well as in the development of hu- of failures, thus increasing availability in production
man-robot interaction systems. New manufacturing systems and gaining benefits from multi-channel,
paradigms, which harness the increased flexibility multi-structured data sources.18 According to a recent
of pairing humans and robots in a coworking envi- report from Intel focusing on creating lasting value
ronment, need to be developed. This will allow for in- in the age of AI and IoT20, 80% of the manufacturing
dustry to take advantage of having the flexibility and companies surveyed were planning to invest in smart
cognitive ability of human workers, as well as the re- technologies and solutions to keep up with their com-
peatability and strength of robotic manipulators. petition in the following two to three years. The top
three most popular investments for the participating
New techniques for AI-based worker assistance are companies were in AI (51% of respondents planned to
required. These include computer vision and AI-ba- invest), predictive/big data analytics (44% planned to
sed systems which can, for example, accurately and invest), and smart machinery (35% planned to invest).
repeatedly assess and predict a human’s movement/ This acceptance of companies to embrace AI and
behaviour within the cell for increasing safety and ef- sensorised equipment will allow for the implementa-
ficiency in collaborating workspaces. New methods tion of KBM, where the data monitoring of machine/
of human-machine interaction and the development process conditions can be automatically analysed to
of AI-assistance systems, manufacturing chatbots or pick up any rules or patterns that indicate a possible
digital coaches, are also required in order to retrieve fault, through data-mining systems being applied to
and pass on information between the human, machi- accelerometric and sensor measurements, e.g., acou-
nes and advanced IT systems. These include mixed stic, temperature and pressure readings, and further
reality and voice-operated systems, which utilise na- extend the scope towards prescriptive analytics.
tural language processing in order to pass on com- These analyses can be used for wear forecasting, re-
mands, instructions and other useful information. maining useful life prediction, process recommenda-
tions, etc.21 In industrial contexts, however, the imple-
AI in factory maintenance mentation of KBM is faced with several challenges,
Factory maintenance is an integral part of production surrounding, inter alia; 1) the proper use of multiple
planning, which contributes to increased productivity, data sources, 2) the suboptimal use of multi-structu-
reduced product life cycle costs, and maximised cu- red data, 3) the multi-modality of data, i.e. the mis-
stomer satisfaction. For several decades, well-foun- sing semantic correlation of information, 4) multiple
ded industrial maintenance and reliability manage- and overlapping reliability-centered and maintenance
ment strategies have confronted persistent dilemmas strategies and approaches, 5) the multi-dimensiona-
such as a decreased ability to react, increasing main- lity of maintenance organisation/actors/teams, pro-
tenance costs, the reduced availability of machines, cesses and IT-systems, and 6) the economic and
low- to mid-level staff qualification, and a lack of insi- technical plausibility of KBM’s business cases.18 In
ght into the impact of maintenance on profitability. In the future, the scope of today’s factory maintenance
the era of Industry 4.0, AI-driven and computational should be extended towards the industrial implemen-
technologies aim to digitalise and intelligentise main- tation of KBM and taking advantage of novel, scalable
tenance processes and systems, and thus introduce AI-driven approaches. Two major areas of untapped
new business cases. This aim can be achieved throu- potential for trustworthy AI as an enabler in factory
gh predictive and Knowledge-Based Maintenance maintenance are:
(KBM) strategies, models, and solutions.19 KBM is
about employing AI methodologies, methods, algori-
thms, and tools/technologies for analysis, modelling,
1. virtual assistance (e.g. voice-guided defect iso- processes can be potentially disrupted. Companies
lation , real-time recommender systems , auto-
22 23
must, therefore, include the concept of cybersecurity
mated profitability assessment24), and and threat intelligence as an inherent component in
2. multichannel sensorisation and intelligenti- any AI digitisation project. A secure digital platform
sation of maintenance processes (e.g. use of helps to meet these cybersecurity-related goals and
Augmented Reality25, text mining26, automated will help to avoid unplanned downtime and the enor-
scheduling27), maintenance ontologies , digital mous costs connected to a data breach. The inherent
twins29). dynamism of AI solutions calls for active, rather than
reactive, digital infrastructure protection solutions.
AI, big data and cybersecurity
To unveil the potential of AI, the manufacturing indu- Computing capabilities and connectivity of the har-
stry has to consolidate a more comprehensive data dware available must also be considered when imple-
culture in terms of industrial cybersecurity, data reuse menting AI for smart manufacturing. Due to the ‘Big
and data sharing. data’ nature of the sensorised equipment being used
in manufacturing, it is important to find a balance
The manufacturing industry has to move beyond data between using Cloud Computing resources, which are
integration at a factory level, and to be able to deal highly scalable and have the capabilities for complex
with multi-tenant scenarios with data generated and analytics to be executed on remote powerful super-
managed across the value chain (see DFA30 Zero-X visory platforms, and using Edge Computing, which
initiative). The development of new manufacturing is a more localised use of computing/storage resour-
services, products or processes relies on access to ces at the location where data is produced. Edge
common data which calls for new technologies to en- computing allows for the implementation of efficient
sure a multi-level digital continuity that goes beyond algorithms that are not computationally expensive.
the factory shop floor and operates at the level of con- Conversely, limitations of Cloud computing include
nected factories and digital value chains. the network bandwidth, data transmission speed, se-
curity, privacy, reliability, and robustness due to the
B2B connected factories’ common data processes transmission of raw sensor data from the machine to
will not be based on direct access to the raw data the cloud. There are opportunities for smart factories
entities themselves. Instead, future data value chains to adopt a combination of both approaches, in sense
will more likely be developed based on sovereign of mixed Edge-Cloud computing and Fog computing,
access to well-curated high-quality data endpoints. which consist of both cloud and edge resources that
Industrial agreements will be instrumental in provi- reduce latency and network congestion.
ding trust in the quality of the data presented in such
data end-points and its reusability. Ethics in AI
The unintended consequences of AI technology are
Data protection with encryption will be necessary that it presents new challenges for the future of work
within end-points relating to data processing and and raises legal and ethical discussions. Some con-
storage, as well as during transmission, to avoid ma- cerns are directly related to the way algorithms, and
nipulation or leakage. The increased connectivity of the data used to train them, may introduce biases or
value chains exposes each additional end-point rela- perpetuate and institutionalise existing social stere-
ting to digital infrastructure data and (micro)service, otypes and procedural biases, which may also gene-
to potential attack vectors. This creates risks at both rate misinformation. Concerns also relate to ethical
the factory and network level that should not be un- decision- making by machines on the operation of
derestimated. Business continuity and automation systems/processes having potentially harmful or ne-
gative consequences on humans. Others include data for bringing a product from prototype to its final form,
privacy or use of personal information, and Europe AI development in terms of data sharing and algori-
has led the way in addressing these challenges. thmic development can span multiple organisations,
and perhaps even multiple continents across the
In April 2019, the European Commission published a globe. In the absence of any international guidelines,
report31 establishing a European AI strategy for posi- engagement with stakeholders is necessary to come
tioning humans at the centre of all AI development to a consensus on what principles and values should
efforts. It provides seven key requirements that AI govern the development and use of AI in manufactu-
applications should adhere to in order to achieve the ring. Given their pioneering work in this domain, the
“trustworthy AI” status. These seven requirements are European Commission could provide valuable insi-
human agency and oversight; technical robustness ghts into these discussions and engage with interna-
and safety; privacy and data governance; transparen- tional stakeholders to develop AI ethics guidelines for
cy; diversity, non-discrimination; societal and environ- manufacturers globally.
mental well-being; and accountability. While many
of these requirements are broadly applicable across
a variety of different organisations, the specific con-
text in relation to how they should be applied to the
manufacturing sector has to be developed. The prin-
ciples and values that guide the ethical development
and deployment of AI in manufacturing should not be
limited to the end product or services, it should cover
the entire digital value chain from design and engine-
ering, planning, supply chain management, factory
automation, and workforce to IoT. These principles
and values should not be about meeting some rigid
checklist, rather it should be based on a framework to
ensure the deployment of AI is fair and equal, accoun-
table, safe, reliable, secure and addresses all privacy
aspects.

The EU already has a proposal for regulatory fra-

mework to establish standards for ethics in AI32 and
applicable privacy laws such as the General Data Pro-
tection Regulation (GDPR)33 that can be adapted to
the manufacturing industry. While this has a broad re-
ach within the EU, there is a lack of global standards
that govern the development and use of AI, other than
those developed by individual organisations. For ma-
nufacturing, there is a desire for a consistent regula-
tory framework on how data is shared and used, both
internally and across the world, given the increasingly
interconnected global supply chain. Given the linka-
ges that exist between manufacturers to support in-
terconnected functions, operations, and transactions
CONCLUSION ANNEX
The core message of this whitepaper is that AI offers
enormous potential for robust and resilient manu- SUCCESS STORY 1 - Siemens
facturing in many respects. A total of 19 key recom-
mendations in 6 application fields were identified Blackout prevention out of the cloud
that enhance long-term resilience in manufacturing.
Pole-mounted transformers are a common sight in
many countries. In India, for example, well over ten
million of these transformers are in use. Every year,
15 to 20 percent of the transformers fail, causing
blackouts and leading to huge repair costs and lost
production.

Oil loss leads to fire risk

The loss of transformer oil is one of the most com-
mon causes of the high failure rate. This oil is used to
cool the coils inside the transformers and keep them
insulated from each other. In poorly maintained devi-
ces, where the housings are badly rusted or cracked,
the oil can leak out. If the oil falls below a critical level,
there is the risk of overheating or of voltage flashover.
In these cases, it is not uncommon for the transfor-
mers to burst into flames.

The digital twins make it possible to achieve a re-

alistic oil-level simulation in the cloud using fast
computers, so now grid operators can be warned in
advance if one of their transformers reaches a criti-
cal state. That is a twin that represents the essential
principles of a pole-mounted transformer but is not
customised down to the level of a specific transfor-
mer. It only becomes an individual twin when we feed
in the measured temperature values from the actual
transformer. To obtain these temperature values, the
transformers simply need to be retrofitted with four
sensors and a router that sends the measured values The requirements on the maximum acceptable defect
to the cloud. These values can be used first to deter- are very strict and even small spots can cause the
mine the shape, size, and rated power of the retrofit- bottle to be discarded. Since extrusion blow moulding
ted transformer compared with an intact transformer, is a complex process prone to a lot of uncertainties,
and then to create an individual twin for it. the shape and type of defects cannot be known in ad-
vance. In order to meet the needs, AISent developed
AI as a computing partner a quality inspection machine that does not follow the
Once the individual twin is in place, the measured traditional “defect detection” model but, instead, fol-
temperature values can be used to simulate the oil le- lows the “anomaly detection” paradigm. The anomaly
vel and indirectly determine how full the unit is. That detection approach has several advantages:
will not work with a traditional simulation, though, be-
cause the algorithms in those cases are based solely 1. It allows the creation of a model for the status of
on physical laws. The amount of time and effort ne- health without spelling out every possible defect,
eded for the calculations would be very great. That is one by one. The quality inspection machine may
why we combined these algorithms with statistically be able to detect new types of defects without
tractable techniques and methods from the world of defining them.
AI. This is the only way to bring the calculation time 2. It allows a reduction in the cost of data-set gathe-
and effort for the simulation down to the point where ring. There is no need to acquire images for every
the oil level in the transformer can be simulated as type of defect, it is just necessary to acquire he-
often and as regularly as necessary. The ease of re- althy bottles. For this specific quality inspection
trofitting is not the only benefit offered by this oil-level machine, a synthetic data-set was used.
monitoring method. Compared with other monitoring 3. It allows a measure of the product quality to be
solutions using sound or weight sensors, for exam- obtained. The output is not binary (good/bad) but
ple, it is also relatively inexpensive. Where other so- is a number between 0 (perfect) and 1 (defective).
lutions can cost almost half as much as a new tran- Therefore, the rejection threshold can be adju-
sformer, the solution using four temperature sensors, sted according to the needs of the customers
a router, and cloud-based simulation is available for (end users, type of product, type of market), span-
about one-tenth of the cost of a new unit. ning from a very strict control to a more flexible
Credits: Hubertus Breuer and Frank Krull approach.
4. The measured quality is used not just for discar-
SUCCESS STORY 2 - AISent ding pieces, but it is also used as a feedback on
the process. A product with low quality can be
AI for quality inspection of bottles seen as a symptom of incorrect process variables
or low-quality input materials. Using the quality
Serioplast is a leading company in the rigid plastic measurements, AISent is building a self-configu-
packaging industry. Serioplast supplies the FMCG ring machine to identify the optimal parameters,
companies (Unilever, P&G, Henkel, L’Oréal etc,) in the which leads to high-quality output.
home care, personal care, food & beverage and au-
tomotive markets. AISent (https://aisent.io/), a deep
tech start-up specialised in AI, Machine Learning and
Computer Vision supported Serioplast in realising an
automated quality inspection.
The quality inspection solution is installed in Dalmine, They highlighted an approach that tackled high- value
Bergamo (IT) and can analyse up to 10,000 bottles areas, starting with the application of AI-driven plan-
per hour. It is composed of 3 cameras and 3 front-li- ning and scheduling and the simultaneous adoption
ghts. The images are elaborated at the edge on an in- of IIoT cloud-connected equipment and personnel.
dustrial PC and the results are computed in real time,
discarding the abnormal pieces. The computer vision Low-cost IIoT Sensors were retrofitted to machines to
algorithm is a proprietary AI, based on Deep Learning. monitor asset performance and electronically execu-
te production processes. The knowledge base on how
machines performed, allowed accurate downtime, se-
tup, changeover and throughput rate information to
be built up, identifying areas for improvement. Paper
processes were replaced with cloud-enabled digital
workflow tools for production tasks, maintenance
and quality checks, bringing the operator closer to
real-time connectivity with equipment and providing
managers with a deeper understanding of factory
behaviour.

These performance metrics were fed into a pro-

duction scheduling system which utilised predictive
and prescriptive AI algorithms to resolve the com-
plexity inherent in their processes. Planning personnel
could see the status of any process in real time and
SUCCESS STORY 3 - TilliT
adjust the schedule as required. Operators spent far
less time transcribing tedious notes onto paper and
AI-based IIoT for mature SME manufacturers
audit records became automatic. Digital orchestra-
tion of material, personnel, equipment and processes
TilliT (https://gotillit.com/) offers tangible, high-value
became possible for the first time in the company’s
applications of AI and IIoT that are available to even
history.
the smallest manufacturers which improve resilience,
and ultimately enable them to thrive.
The result was an improvement in throughput rate, re-
duction in changeover time, lower raw material levels
Oliveri Sinkware, an Australian manufacturer of ki-
and reallocation of resources to areas of higher-value
tchen sinks and associated components represents
need within the facility.
the vast majority of SMEs worldwide. The facility is
small on a global scale, with a mixture of modern
Importantly, the ongoing improvements did not re-
equipment and machines that are decades old, ope-
quire complex industrial automation infrastructure,
rated by a skilled long-standing workforce not fami-
expensive consulting or the high capital allocation
liar with digitalised operations. Steel presses, welding
often associated with achieving digital manufactu-
and grinding machines lacked advanced logic control,
ring operations. This has set the baseline for Oliveri to
and production processes were managed on paper.
build on with more advanced technology adoption in
other areas of their business.
Management started with a view that Industry 4.0 is
not a specific project, but a strategy and mindset ap-
plication driven by a need to be digitally connected.
 ups. This demands regular and inline verification of
the status of the MT precision and real-time AI-powe-
red error compensation, which has a positive influen-
ce on high accuracy and cognitive machining process
planning and scheduling, helping to reduce expenses
on condition-based MT calibration and energy, and
material and productivity losses from defective ma-
chining processes.

The EU’s Qu4lity digital manufacturing flagship

project, https://qu4lity-project.eu/, has trialled adap-
tive manufacturing processes integrating Innovalia
SUCCESS STORY 4 - Innovalia Metrology M3MH open cognitive 5-axis metrology MT set-up so-
lution with +GF+ smart machining control processes.
High performance machining autonomy Autonomous in terms of machine verification, it im-
proves daily operations by leveraging greener and
High precision machining is a key manufacturing pro- faster production (reducing up to 50% of machining
cess used to produce critical and geometrically com- time), avoiding process failures, variability in quality
plex parts in aeronautics. To meet such demands, output and bolstering continuous improvement sim-
each year, the aeronautical sector has to deal with ply by gaining better access to and smarter insights
high scrap rates and high numbers of unproductive from real-time data. The integration of data-driven di-
machining hours and accumulated costs primarily in mensional metrological intelligence and autonomous
extensive tests run to ensure process availability, qua- compensation modules in machining processes gives
lity and performance. operators and MTs the possibility of delivering at any
point in time the best possible part based on machine
Green & Sustainable Machining conditions and part geometry. AI-powered MT set-
Machining of critical components in aerospace is ex- up and measurement, integrated with high precision
tremely sensitive, and variances in performance are machining, reverts to increased productivity, reduced
mainly due to changes in the machine environment operational costs and more sustainable operations
(temperature, humidity, etc), material behaviour, with a marginal digital investment cost and low wor-
equipment/tool ageing and error propagation in mul- kforce re- and upskilling demands.
tistage processes. Highly skilled human expertise
and continuous process optimisation are required
for ensuring process repeatability and for tuning and
maintaining accuracy in applications at different con-
ditions. AI can play a critical role in delivering the
most awaited autonomy that will reduce scrap rates,
optimise machining performance and energy consu-
mption.

Credits: Jesus Alonso and Oscar Lazaro35

AI to realise machining autonomy
Autonomous machining performance can be levera-
ged by equally ensuring Machine Tool (MT) accuracy
status and equipment availability and optimum set-
REFERENCES

1
Industry 4.0: Artificial Intelligence in Manufacturing (2021, June 29). Retrieved from: https://www.
futuremanageralliance.com/industry-4-0-artifical-intelligence-in-manufacturing/
2
Rauch, E. (2020). Industry 4.0+: The Next Level of Intelligent and Self-optimizing Factories. In: Ivanov V.,
Trojanowska J., Pavlenko I., Zajac J., Peraković D. (eds) Advances in Design, Simulation and Manufacturing
III. DSMIE 2020. Lecture Notes in Mechanical Engineering. Springer, Cham. https://doi.org/10.1007/978-3-
030-50794-7_18
3
WMF report 2020: Manufacturing in the Age of Artificial Intelligence (2020). Retrieved from https://
worldmanufacturing.org/report/report-2020/
4
How Blockchain, IIoT and AI Make the Global Supply Chain More Resilient (2021, May 20). Retrieved from:
https://stellarfoodforthought.net/how-blockchain-iiot-and-ai-make-the-global-supply-chain-more-resilient/
5
Youn, B. D., Hu, C., Wang, P. (2011). Resilience-driven system design of complex engineered systems.
Journal of Mechanical Design, 133(10), 101011.
6
Belhadi, A., Mani, V., Kamble, S. S., Khan, S. A. R., & Verma, S. (2021). Artificial intelligence-driven innovation
for enhancing supply chain resilience and performance under the effect of supply chain dynamism: an
empirical investigation. Annals of Operations Research, 1-26.
7
Modgil, S., Gupta, S., Stekelorum, R., & Laguir, I. (2021). AI technologies and their impact on supply chain
resilience during COVID-19. International Journal of Physical Distribution & Logistics Management.
8
Bauernhansl, T. (2014). Die Vierte Industrielle Revolution – Der Weg in ein wertschaffendes
Produktionsparadigma. In: Bauernhansl T., ten Hompel M, Vogel-Heuser B (eds.) Industrie 4.0 in Produktion,
Automatisierung und Logistik. Springer Fachmedien Wiesbaden, Wiesbaden, pp. 5–35.
9
Humpa, M. (2016). CAD-Methodik zur Produktivitätssteigerung in der Prozesskette Konstruktion-
Fertigung. Dissertation, Faculty of Engineering. Retrieved from: https://duepublico2.uni-due.de/servlets/
MCRFileNodeServlet/duepublico_derivate_00042999/Humpa_Diss.pdf.
10
Fleischer, B. (2019). Methodisches Konstruieren in Ausbildung und Beruf. Praxisorientierte
Konstruktionsentwicklung und rechnergestützte Optimierung. Springer Fachmedien Wiesbaden, Wiesbaden.
ISBN: 978-3-658-27690-4.
11
Capgemini (2020). Fast Forward - Rethinking supply chain resilience for a post-COVID-19 world. Retrieved
from: https://www.capgemini.com/wp-content/uploads/2020/11/Fast-forward_Report.pdf
12
Kohl, L., Ansari, F., Sihn, W. (2021). A Modular Federated Learning Architecture for Integration of AI-
enhanced Assistance in Industrial Maintenance. Academic Society for Work and Industrial Organization, Gito
Verlag Berlin, 2021 (in press).
13
Krahe, C., Iberl, M., Jacob, A., Lanza, G. (2019). AI-based Computer Aided Engineering for automated
product design-A first approach with a Multi-View based classification. Procedia CIRP, 86, 104-109.
14
Kirschniak, C. (2018). Auswirkungen der Nutzung von künstlicher Intelligenz in Deutschland. Retrieved
from: https://www.pwc.de/de/business-analytics/sizing-the-price-final-juni-2018.pdf.
15
Apt, W., Bovenschulte, M., Priesack, K., Weiss, C., Hartmann, E. (2018). Einsatz von digitalen
Assistenzsystemen im Betrieb. Bundesministerium für Arbeit und Soziales. Retrieved from: https://www.iit-
berlin.de/publikation/einsatz-von-digitalen-assistenzsystemen-im-betrieb/.
16
Hagras, H. (2018). Toward human-understandable, explainable AI. Computer, 51(9), 28-36.
17
IPC-2551: International Standard for Digital Twins. Retrieved from: https://shop.ipc.org/2551-0-0-english
18
Ansari, F., Glawar, R., Nemeth, T. (2019). PriMa: A Prescriptive Maintenance Model for Cyber-Physical
Production Systems, International Journal of Computer Integrated Manufacturing, 32(4-5), pp. 482-503.
19
Ansari, F., Glawar, R. (2018). Knowledge-Based Maintenance. In: Matyas K (Ed.) Maintenance Logistics, 7th
Edition, Carl Hanser Verlag, pp. 318-342.
20
Petrick, I., McCreary, F. (2019). Creating lasting value in the age of AI + IoT: Futureproofing your business,
Intel Newsroom, December 2019. Retrieved from: https://newsroom.intel.com/wp-content/uploads/
sites/11/2019/12/futureproofing-your-business.pdf
21
Ansari, F., Kohl, L., Giner, J., Meier, H. (2021). Text mining for AI enhanced failure detection and availability
optimization in production systems, CIRP Annals - Manufacturing Technology, 70(1), 2021, pp. 373-376.
22
Wellsandta, S., Rusak, Z., Ruiz Arenas, S., Aschenbrenner, D., Hribernik, K. A., Thoben, K. D. (2020). Concept
of a Voice-Enabled Digital Assistant for Predictive Maintenance in Manufacturing. Available at SSRN
3718008.
23
Pech, M., Vrchota, J., Bednář, J. (2021). Predictive Maintenance and Intelligent Sensors in Smart Factory.
Sensors, 21(4), 1470.
24
Schenkelberg, K., Seidenberg, U., Ansari, F. (2020). Supervised Machine Learning for Knowledge-Based
Analysis of Maintenance Impact on Profitability. IFAC-PapersOnLine, 53(2), 10651-10657.
25
Kostoláni, M., Murín, J., Kozák, Š. (2019, June). Intelligent predictive maintenance control using augmented
reality. In 2019 22nd International Conference on Process Control (PC19) (pp. 131-135). IEEE.
Ansari, F. (2020). Cost-based text understanding to improve maintenance knowledge intelligence in
26

manufacturing enterprises. Computers & Industrial Engineering, 141, 106319.

Glawar, R., Ansari, F., Viharos, Z. J., Matyas, K., Sihn, W. (2021). Integrating Maintenance Strategies in
27

Autonomous Production Control using a Cost-based Model, Acta IMEKO - The e-Journal of the International
Measurement Confederation (IMEKO).
28
Ansari, F., Khobreh, M., Seidenberg, U., Sihn, W. (2018). A problem-solving ontology for human-centered
cyber physical production systems. CIRP Journal of Manufacturing Science and Technology, 22, 91-106.
Passath, T., Huber, C., Kohl, L., Ansari, F. (2021). A Knowledge-Based Digital Lifecycle-Oriented Asset
29

Optimisation. Tehnički glasnik, 15(2), 226-234.

30
Digital Factory Alliance (DFA) – https://digitalfactoryalliance.eu
31
European Commission - Building Trust in Human-Centric Artificial Intelligence. Retrieved from: https://
ec.europa.eu/jrc/communities/en/community/digitranscope/document/building-trust-human-centric-
artificial-intelligence
32
Proposal for a Regulation of the European Parliament and of the Council - Laying Down Harmonised Rules
on Artificial Intelligence (Artificial Intelligence Act) and Amending Certain Union Legislative Acts. Retrieved
from: https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:52021PC0206&from=EN
33
General Data Protection Regulation (GDPR). Retrieved from: https://gdpr-info.eu/
34
Blackout prevention out of the cloud. Simulations supported by artificial intelligence can protect millions of
pole mounted transformers from sudden, total failure. Retrieved from: https://new.siemens.com/global/en/
company/stories/research-technologies/digitaltwin/oil-level-monitoring-for-pole-mounted-transformers.html
35
Innovalia Metrology M3MH – Machining without defects. Retrieved from: https://www.innovalia-metrology.
com/m3mh-machining-without-defects/
World Manufacturing Foundation
Via Pantano, 9 - 20122 Milano, Italy

worldmanufacturing.org