Machine Learning. A Guideline for its Usability in Production Systems


Bachelor Thesis, 2017
59 Pages, Grade: 1,3

Excerpt

List of Content

List of Content

List of Figures

1 Introduction
1.1 Objective
1.2 Structure of Thesis
1.3 Research Method and Study Design

2 Impact of Machine Learning in Industrie 4.0
2.1 Market Pull is Changing the World of Manufacturing
2.2 Key Challenges for Production Systems in an Evolving Business World
2.3 Industrie
2.4 Ubiquitous Computing and Visualization
2.5 Impact on Human Employment
2.5.1 Computerisation in Non-Routine Manual Tasks
2.5.2 Computerisation in Non-Routine Cognitive Tasks
2.5.3 Implications for Employment

3 Paradigm Shift from Abstract Models to Real World Data
3.1 What Machine Learning is and Why it is a Promising Approach
3.2 Machine Learning Techniques
3.2.1 Regression
3.2.2 Classification and Clustering
3.2.3 Dimensionality reduction
3.2.4 Association rule mining
3.3 Learning Types
3.3.1 Supervised Learning
3.3.2 Unsupervised Learning
3.3.3 Reinforcement Learning
3.4 Algorithm Selection: Implicit vs. Explicit Knowledge Representation

4 Applications of Machine Learning in Production
4.1 Descriptive Analytics
4.2 Diagnostic Analytics
4.3 Predictive Analytics
4.4 Prescriptive Analytics

5 Guidelines for the Usage of Machine Learning in Production
5.1 Domain Maturity: Machine Learning
5.2 Domain Maturity: Production
5.3 Infrastructure: Connection Task
5.4 Data: Capturing Task
5.5 Security: Cyber Security and Accountability Task
5.6 People: Knowledge and Acceptance Task
5.7 Strategy: Cooperate Design Task

6 References

List of Figures

Figure 1: Illustration of the structure of this thesis [Author]

Figure 2: Machine learning implementation in production roadmap [Author]

Figure 3: Reinforcement learning concept [53] (modified)

Figure 4: Business Analytics [57] (modified)

Figure 5: Two different types of machine learning publications [61]

Figure 6: Machine learning focuses on benchmark data sets [61]

Figure 7: Ability of machine learning to satisfy the requirements of production systems [12]

Figure 8: Two motives of companies for the use of machine learning in production systems [12]

Figure 9: Phase of machine learning usage [62] (modified)

Figure 10: Top inhibitors of machine learning initiatives in production systems [12]

1 Introduction

After the revolutionary change caused by the introduction of the steam engine, the production line, electronics and IT, into the manufacturing industry, a new disrupting change is expected [1]. Nowadays the rapidly increasing digitalization of the economy leads to the fourth industrial revolution. This global phenomenon is called ‘Industrie 4.0’ (GER) or ‘Smart Factory’ (US) [2], and it combines production technology with information and communication technology [1]. Especially, data based optimization in production is one of the predominant goals of Industrie 4.0 [3]. For the automatized analysis of large amounts of data, machine learning is an effective instrument and therefore a central element in Industrie 4.0 [3].

Recent progress in machine learning has been driven by the development of new learning algorithms and by the increasing availability of data and low-cost computation power [4]. For many applications - from computer vision to adaptive robots – it was very difficult to devise deterministic rules [5]. However, for these applications, it is possible to collect data, and now the idea is to use algorithms that learn from data, instead of being manually programmed [5]. Thus, machine learning has the potential to transform data into valuable knowledge for decision making, while making improvements possible to the production system, with approaches such as predictive maintenance [3]. The transfer of machine learning from the lab to the ‘real world’ leads to an increased interest in learning techniques, demanding further effort in explaining, on how machine learning works, and what it can be used for in other disciplines [6]. However, the entry barrier to the diverse field of machine learning is high. With many different algorithms, theories and methods, it is hard to oversee, and therefore its influence remains limited [2]. In addition, a recent study states that about 47% of jobs in the US are at high risk of computerization within the next decades [7]. Therefore, employees feel insecure [8], and demand answers on what effect machine learning will have on their future role in the factory.

This thesis provides an especially designed overview for the needs of decision makers in the production industry on the field of machine learning. By concerning economic and technological factors, as well as the individual challenges for companies, the goal of this thesis is to serve as a guideline for the usage of machine learning in production systems.

1.1 Objective

Regarding this thesis, we understand production systems to be comprised of both the technological elements (e.g. machines and tools) and organizational elements (e.g. labor and information) [9]. While the amount of data captured in production systems increases [2], among other reasons, due to a growing number of sensors fitted to machinery, the value of data has yet to be attained [10]. Machine learning seems to be a promising technology for data handling [5]. However, its implementation into production systems can be regarded as a highly interdisciplinary project [11].

The collaboration of different disciplines, especially data science and production is required in implementation projects [2]. Whereas machine learning expert knowledge is needed to decide upon the appropriate techniques for the data analysis, manufacturing domain expert knowledge is required for the interpretation of the results of the data analytics methods [12]. Although the interaction between the two disciplines is vital for the implementation of machine learning into production systems, the cooperation could be challenging, due to the distinct diversity of these disciplines, which we discuss in the following paragraphs.

Many companies in the production industry in Germany were established in the so-called founding years, between 1870-1914 [13]. Looking back on an over a century lasting company history, they usually have a conservative organizational culture and plenty of experience in manufacturing. Since they have gradually developed to consistently produce better and better goods and machinery. This attitude made many companies, especially from the machinery industry, become a world market leader in highly profitable niche markets. Moreover, the mindset of making ever-better products made German car brands very well known for their outstanding quality.

In contrast, many corporations with leading machine learning abilities, such as the members of the ‘AI Partnership’, which aim is to share the best practices in artificial intelligence research, for example Amazon, Apple, Facebook, Google, Microsoft, and IBM, were founded in the late 20th or early 21st century [14]. Just a few years ago most of them were startups [15]. In Silicon Valley, where most of these firms are based the gross domestic product is almost six times higher compared to the industry sector in Germany [16,17]. Instead of gradually evolving over time, the aim of many of these firms is to revolutionize their entire industry, as for example, the San Francisco-based company Uber changed the taxi industry [18]. As a researcher in the field of cooperative robotics, from the Institute for Management Cybernetics e.V., Philipp Ennen puts it: “In the field of machine learning work enormously passionate and intelligent people, who are almost all workaholics, aiming to change the world [12].” As the cultures of the two domains differ, so do the products alike. Whereas the data-driven companies largely offer services to their customers, in production many business models are still very focused on physical products. To demonstrate the observation of the different focus of the two domains towards services and products, an illustrative example provides a comparison between cars and smartphones by scrutinizing whether service or physical product is considered as standard or luxury in this context.

Whereas with cars, additional services, such as navigation applications or entertainment systems come at a very high extra cost, hence the price for additional service features can easily account for more than the basic configuration of a car. Moreover, also the number of possible extra services increases, with the base price of the car. Consequently, the service in this product driven domain can be considered as luxury.

In contrast, the services provided by smartphones, such as navigation or entertainment functions are several orders of magnitude cheaper, or even free, while the amount of services provided does not correlate much with the initial smartphone price. Hence, in this domain, services can be considered as standard, whereas the physical product, for example, the Pixel Phone comes at a high price and is considered as a luxury in this domain.

Due to their history, the attitude towards the handling of data and intellectual property of the two domains is very different. Even for research purposes, it is tough to acquire data from production systems, due to security concerns [2]. In contrast, in December 2015 Google made its neuronal network toolbox TensorFlow freely available to the public. As a reaction, Facebook and Microsoft also made their competitive product open source just a few weeks later [10]. The author assesses this trend as positive because it could make machine learning methods widely accessible, allowing it to be used by many people from all over the world. Especially in combination with free massive open online courses that teach machine learning abilities, the accessibility of machine learning tools may lead to an innovative, non-exclusive and rapidly growing worldwide community, which leverages the advantages of machine learning in real world applications.

To overcome the difference in the two disciplines may prove to be one of the toughest challenges of Industrie 4.0 [12]. Thus, the objective of this thesis is to set a basis for the mutual understanding of the impact of machine learning on production systems and to stress potential challenges, while providing a set of guidelines to overcome these barriers. This way the author hopes to contribute to improving the capabilities of production, enabling it to make better products, while using fewer resources.

1.2 Structure of Thesis

Abbildung in dieser Leseprobe nicht enthalten

Figure 1: Illustration of the structure of this thesis [Author]

This thesis outlines the economic and technological challenges of the production industry with respect to the utilization of machine learning. It presents the reader with a guideline of how to handle typical problems in the implementation of machine learning and is structured into five chapters:

(1) The present chapter describes the main motive behind this thesis, of making machine learning easier accessible to production systems, and discusses the methods which were used in the creation of the thesis.
(2) The second chapter describes the market trends, which force companies to act in order to stay competitive. Based on this observation, eight key challenges are outlined that businesses in the manufacturing industry should overcome, in order to remain successful in the future. Moreover, it is presented how the recent development of digital technologies made all the means available, that are necessary for Industrie 4.0. After the positive impact of Industrie 4.0 for production is outlined, in particular, the role of machine learning in the Industrie 4.0 and its impact on human employment is discussed.
(3) The third chapter introduces machine learning according to the needs of experts in the production domain. After outlining the significant impact of machine learning, the thesis discusses an observation that gives several hints for an ongoing paradigm shift from a deterministic approach towards a data-driven approach in factory automatization. It is explained, how the machine learning techniques, e.g. Classification or Association Rule Mining, can be used to make machines more intelligent. Subsequently, it is considered, how to teach these intelligent machines with e.g. Supervised or Unsupervised learning types. At the end of the chapter, advantages and disadvantages of several algorithms for its applicability in production systems are discussed.
(4) The fourth chapter covers four types of data analytics methods. These analytics methods are used to structure the applications of machine learning in production and show its value. In this chapter, we take a look at the technology from a user’s perspective. By going back to the eight key challenges from chapter two, it is demonstrated how the four types of analytics methods can contribute to handling the eight key challenges of production. Simultaneously, it is outlined, how the theory of machine learning from chapter three can be used in factory applications while showing that machine learning is an incremental technology, that is required to make advanced data analytics possible.
(5) The fifth chapter finally takes a look at the maturity of the domains of computer science and production to bring machine learning into production applications. Thereafter, based on the current state of the industry, a guideline for individual businesses is formulated, concerning tasks in infrastructure, data, strategy, people, and security. The guideline provides actionable tasks for decision makers on what companies need to do to be successful with machine learning in the future.

1.3 Research Method and Study Design

In the preparation of this thesis, two research methods were utilized. The methods included (1) a literature research, as well as, (2) 16 expert interviews based on a specifically developed scheme, asking participants from the production and the machine learning domain. The specific characteristics of the sample will be outlined at the end of this chapter.

(1) As the primary source for the content of this thesis, the literature research was chosen. As, the topic of this thesis requires an understanding of the perspective of two domains, machine learning and production, literature was evaluated from both disciplines. Due to this multiperspective view on the literature, it was possible to connect both perspectives on machine learning to one big picture. In the research, content was considered from scientific monographs, journal articles, lecture notes, massive open online courses, company research reports, newspaper articles and web content. With this broad range of sources, a multi-perspective view was gained in the fields of machine learning and production, allowing to secure the objective overview that is required to write this thesis.

Whereas the literature is rich on scientific papers and especially business reports, presenting the implementation of a specific application of machine learning in production, the discussion lacks the examination of implementation difficulties of machine learning into the entire production systems. The validation of these successes is tough because the performance data of these applications is typically not publicly available [2]. The purpose of the interviews was to narrow this gap, by approaching the issue of the usability of machine learning from the perspective of a challenges and difficulties point of view. In addition, the interviews were used to validate the results from the literature research.

(2) The Interviews were conducted in a one-on-one fashion throughout February and March 2017 at the working place of the experts. On average, interviews had a duration of 30 minutes. For a better evaluation, after permission of the participants, all interviews were recorded on audio. The interview scheme was divided into two sections: The first section examines through a survey the general usability of machine learning in production systems. Whereas the second section uses questions without predefined answers to focus on the personal expertise of the participants and to promote a discussion.

In the first section, the participants had to evaluate the applicability of predefined statements, using a rating scale of the range from 1 (totally agree) to 6 (not correct at all). The questions were structured in a logical order, starting with the motivation of companies to use machine learning. The survey continued with the technological usability challenges of machine learning in production systems. Finally, the obstacles, which companies are faced with in the implementation process of machine learning into their production system were addressed.

The second section was constructed of a semi-structured interview. It was based on predefined written questions. Questions were added, when the expert knowledge promised relevant information for the research of this thesis. This flexibility was utilized especially when topics were discussed contradictory by different authors. The participants were asked about their personal experience with machine learning in production systems. Starting with an assessment of their used methods, the participants were questioned about the generalizability of those methods. Then, a closer look was taken at the implementation challenges before it finished by asking for the next developments involving machine learning in their field of expertise.

In the interview 16 experts participated, who had equally distributed professional experience in either the field of production or machine learning. Moreover, about 80% of the participants have worked in projects, applying machine learning methods in a production context. Despite this similar background of the participants, the sample is heterogenic, which is an essential requisite for the generalizability of the study results. It was deliberately chosen to have no online questionnaire because the intention was to gain a picture of the observation of experts from machine learning and production – not from the population, hence the topic is complex and therefore difficult to oversee from outsiders. Due to the high response rate from invited experts of over 80%, a negative effect on the representativeness from a low response rate of the invited participants may be neglectable in this study. As Participants originated from a wide range of industries, such as logistic industry, automotive industry (supplier and car maker), steel industry, plastics industry, tooling industry, and automatization industry. Likewise, the range of professions was profound, as the sample includes managers, project managers, implementers, and scientists.

The results of the two sections are displayed in a different manner. The results of the interviews from section two of the survey are used throughout the whole thesis and are referenced as the Supplementary Material of the Interview. The gained data from the first section of the survey appears in Chapter Five: Guideline for the Usage of Machine Learning in Production. In order to make the interpretation of the results more intuitive, the author has chosen to transform the rating scale from a range between one and six to a percental scale. Whereas the best possible rating of the scale corresponds to 100% and the worst possible rating of the scale corresponds to 0%.

2 Impact of Machine Learning in Industrie 4.0

2.1 Market Pull is Changing the World of Manufacturing

The gradual shift from the mass production of Henry Ford towards the customer-centric production of individual goods named ‘Mass Customization’ began in the 1960s [19]. The reason for this trend towards the individualization of manufactured goods is a consequence of the changing competitive conditions. As customer expectations diverge and the globalization of markets increases, the competitive pressure forces businesses to act [20]. Individualized goods provide much higher value to the customer than standardized mass produced goods. Therefore, customers are willing to pay more in exchange for them [21,22]. With the individualization of the produced goods production faces a rapid increase in the diversity of variants and growing complexity of their processes.

At the same time, it is no longer sufficient to build the most advanced machines. The production industry is gradually changing to become more like the service industry [23,21]. With increasing flexibility, the production industry can offer its production technology to open markets. For example, for buffering capacity deviation or combining the production expertise of many firms into the production of making ever more complex products. Apple, for example, does not have its own production facilities but utilizes Asian production service providers for the assembly of their goods [24]. Using production as a service, allows the company to focus on their core competence of developing new products [24].

2.2 Key Challenges for Production Systems in an Evolving Business World

The trend to the production of individualized goods and towards service driven business models, in combination with increasing market dynamics in a globalized world can be seen as a very disruptive change for manufacturing companies [21]. This is difficult in particular for those companies, that are used to take material from suppliers and turn it into products which are then pushed out to their customers [21]. This demand for adaptation motivates the description of what the author regards as the eight key challenges production is facing, based on the challenges stated in the literature. The author hopes that these key challenges can help to focus actions on strategic initiatives to overcome these challenges.

(1) Productivity: To satisfy the financial requirements for given inputs, the production system must maximize the production output [25], in order to be able to offer low prices to its customers.

(2) Reliability: To maintain stable and reliable processes, creating high-quality products consistently [25].

(3) Transparency: Factories need to be transparent, to stay in control despite increasing complexity and to continue leveraging scale effects [25].

(4) Ability to Innovate: Shorter development time in order to reduce the Time-to-Market to adapt to changing customer needs, are becoming essential for maintaining a competitive advantage [26].

(5) Individualization of Production: For decades a shift from a seller's market towards a buyers’ market strengthens the influence of customers to the trading exchange conditions. This change of power leads to a shift to personalized products, often referred to as ‘Lot Size-1’ [26].

(6) Flexibility: Due to growing dynamics in the surrounding conditions, production systems need to become more flexible in the volume and the products they produce. Therefore, also an increased flexibility in the use of human labor is required [26,25].

(7) Decentralization: In order to meet the challenges above, faster decision making processes becomes a mandatory requirement. To accelerate decision making, organizational hierarchies must be reduced [26].

(8) Efficient use of Resources: The increasing scarcity of resources, the resulting increase in price, and the societal pressure from a public opinion concerning the environmental impact of production, lead to an intensified need to focus on sustainability. Therefore, efficiency improvements have to be regarded in a two-dimensional way, economically and ecologically [26].

2.3 Industrie 4.0

The key challenges for production systems lead to the necessity, to use all means available to adapt to the changing circumstances [2]. Dramatic change has accompanied production since the beginning of the industrialization. Hence mandatory change is nothing extraordinary in the business world. Many times, businesses had to act accordingly in order to stay competitive. Yet, the velocity of change has increased over time, making the agility and flexibility of businesses to adapt increasingly important in order to prevent being outpaced by the development.

Around 1750, during the first industrial revolution [27] the fabrication of goods moved from mostly manual labor in manufactories towards factories [28], that utilized mechanical means to create goods, powered by water and steam energy [27]. With the beginning of the next century, the second industrial revolution arrived [27]. Electrical energy took over as the dominant source of power [27]. Before the introduction of electricity, machines with the highest demand for power had to be placed closest to the energy source [29]. The advent of electricity led to the decentralization of the energy supply, thus allowing to layout machines in production according to the product flow [29]. This layout was an indispensable requisite for the production line [29]. Mass production became widely spread and with it, the division of labor, which allowed first production optimizations with ‘scientific management’ [30]. The third industrial revolution is also called the digital revolution and began in the 1970s [27]. Computers and better communication technology made machines more sophisticated and flexible [27]. During this time, Lean Production played a major role in helping factories to become more productive [31]. Today, the fourth revolution is soon expected.

The next revolution is expected, because the combination of four requirements, connectivity, computer power, large amounts of data and the ability to interpret the data, are quickly developing and are therefore becoming available at reasonable costs [5,4,2]. The manufacturing industry is witnessing a never seen increase in available data for interpretation [2]. This growth in data is a result of rapidly decreasing prices for sensors, leading to intensified fitting of sensors to e.g. products, machinery, and handheld systems. Cloud-based data storage solutions improve the connectivity, by allowing to combine all the data from different areas of the factory to one location, where their interdependencies and correlation can be extracted from the observed data. As the volume of data increases, the potential of machine learning does so too [32]. Moreover, this development leads to a need for algorithms that perform well on huge amounts of data, making machine learning a fundamental necessity for Industrie 4.0 [10]. An auspicious development in recent years is deep learning, which is a machine learning model based on a brain-inspired neural network [2]. These are large networks of threshold units arranged in multiple processing layers [2], of which each threshold unit has a simple parameterized function as input [4], and its output serves as input for the lower layer [5]. Gradient-based optimization algorithms are used to adjust the parameters [4] allowing to recognize complex and non-linear patterns [2]. This approach leads to significant improvements in e.g. computer vision and language recognition but also holds a massive potential for the manufacturing industry [2].

Such development in the field of machine learning and the availability of low-cost, but powerful computation power promises to combine all means necessary to explore the value of data [4]. The technological base of Industrie 4.0 are smart, digital networks which will eventually make a largely self-organizing production possible [33]. In the Industrie 4.0, people, machines, equipment, logistics and products communicate and cooperate directly with each other. In this way, smart value chains can evolve, which include all phases of the product life-cycle. This way, from product design to service, customer demand can be better satisfied [33]. For companies, it becomes easier to customize products according to individual customer requirements. The individualized production and maintenance of the products could become the new standard. On the other hand, the hope is to reduce the cost of manufacturing, despite the individualized production [33]. Whereas currently, individualized production for example in the machinery industry leads to considerably higher costs.

In total, it is assumed, that Industrie 4.0 can increase the efficiency of production, strengthening the competitiveness of the German industry and increase the flexibility of production as a whole [33]. The ability of machine learning based business analytics, to solve the key challenges of production systems in the context of Industrie 4.0 will be discussed further in Chapter Four: Applications of Machine Learning in Production.

2.4 Ubiquitous Computing and Visualization

When computers and humans start working together in hybrid working environments, the future challenge will be to make the cooperation between humans, machines, robots and virtual assistants, as smooth as possible [23]. However, in critical applications, the computer will remain limited to serve human operators with recommendations [7]. Advances in speech and gesture recognition technology, make the interaction with computers increasingly natural. These sophisticated interfaces are called ‘natural interface,’ and will gradually substitute graphical user interfaces [34]. Due to the development of natural interfaces and the increasing number of cyber-physical systems, ubiquitous computing becomes an increasingly popular term. The idea of ubiquitous computing is to use computers, without consciously noticing it [5]. A possible future scenario could be, instead of giving the task, to pull out all the relevant information about the performance of a certain product in a specific market, to a team member, a computer could be assigned with this task, speeding up decision making and improving its quality [34].

2.5 Impact on Human Employment

The concern of technological unemployment is not a recent phenomenon. Historically, technological advances have produced tremendous wealth, however also caused creative destruction which has lead to undesired disruptions [7]. Since the beginning of the industrialization, massive creative disruption has taken place. In France, between 1800 and 2000 the proportion of labor working in agriculture, has decreased by 95% [35]. However, it is still among the 30 wealthiest countries [36]. John Maynard Keynes, one of the most influential economists of the 20th century argued in 1933 that “due to the discovery of means of economizing the use of labor outrunning the pace at which we can find new uses for labor,” and this concern has remained until today [7]. In the past, the industrialization has lead to an increase in automatization of rule-based routine tasks [37]. However, recently the range of tasks computers can perform expanded with the use of machine learning and will continue to do so [38]. Hence, automatization will now enter areas of manual and cognitive non-routine tasks, which are currently carried out by humans.

[...]

Excerpt out of 59 pages

Details

Title
Machine Learning. A Guideline for its Usability in Production Systems
College
RWTH Aachen University
Grade
1,3
Author
Year
2017
Pages
59
Catalog Number
V489418
ISBN (eBook)
9783668968431
ISBN (Book)
9783668968448
Language
English
Series
Aus der Reihe: e-fellows.net stipendiaten-wissen
Tags
Machine Learning, Production, I4.0, Industrie 4.0, AI, Maschinelles Lernnen, Industrie, KI, Künstliche Intelligenz
Quote paper
Alexander Volz (Author), 2017, Machine Learning. A Guideline for its Usability in Production Systems, Munich, GRIN Verlag, https://www.grin.com/document/489418

Comments

  • No comments yet.
Read the ebook
Title: Machine Learning. A Guideline for its Usability in Production Systems


Upload papers

Your term paper / thesis:

- Publication as eBook and book
- High royalties for the sales
- Completely free - with ISBN
- It only takes five minutes
- Every paper finds readers

Publish now - it's free