Ho Chi Minh City University of Technology and Education

Faculity of Mechanical Engineering – Machine Manufacturing

Technology

MECHANICAL ENGINEERING ENGLISH

Full name: Nguyen Nhat Nam

ID student: 21143187
Class: ENME320124-nhóm 04
Instructor: Associate Professor Ph. D Tran Ngoc Dam
FINAL EXERCISE
PORTFOLIO
I/ free paragraph
First of all, My full name is Nguyen Nhat Nam. I’m 20 years old and come from Long An
province. My family has four members. My parent work as a shop manager at my house.
Everyday we get up at 4.30 am . They don’t earn a lot of money but I repect my parent job.
My hometown near HCM city about 10km, This place famous is deligent people and the
food at there is delicous. I graduated from HCM University of Technology and Education.
I have 2 years of experience working in mechanical product design in design company
such as Bosch, Intel, …..I can use fluen a lot of software such as Autocad, inventor,…
Beside I can fluent foreign language which is English.
Thank you for listening, I hope to get the best result.
II/ Title and Abstract
The Advancements and Applications of 3D Printer Machines
Abtract: The emergence of 3D printer machines has revolutionized various industries,
from manufacturing to healthcare. This technology has enabled the production of complex
designs with precision and speed, allowing for faster prototyping and reduced costs. This
article explores the advancements and applications of 3D printer machines, including their
use in aerospace, automotive, and medical industries. Additionally, we discuss the
environmental benefits of 3D printing and how it can contribute to a sustainable future.
Overall, this article highlights the importance of 3D printing technology in shaping the
future of various industries and its potential to drive innovation and progress.
 Making a Fruit Cutter adopts the Principle of Quick Cutting in Food
Industry:
Abtract: This project involves the development of a fruit cutter that adopts the principle
of quick cutting used in the food industry. The fruit cutter will be designed to cut fruits
quickly and accurately, improving the efficiency of food preparation in homes and
restaurants. The project will involve the use of Computer-Aided Design (CAD) software
to develop the cutter's design, and it will be manufactured using stainless steel, which is
food-grade and durable. The cutter will be tested for its cutting speed, accuracy, and
durability.
Design and Development of an Arm Robot for Student Learning and
Exploration in university
Abstract: Robotics is an exciting field that offers endless opportunities for students to learn
about engineering, programming, and technology. To enhance student learning, we have
designed and developed an arm robot that is accessible and easy to use for students. Our
robot features a user-friendly interface and a variety of sensors and actuators that allow
students to explore and experiment with different robotic applications. In this paper, we
will describe the design and development process of our arm robot, highlight its unique
features, and discuss its potential for enhancing student engagement and learning in
robotics education.
INTERNET
Design And Fabrication Of Automatic Onion
Peeling And Cutting Machine
Abstract:
Automation plays an important role in industrial and domestic applications. The purpose
of automation is to get higher production rates, efficient utilization of materials, better
quality, reduce time consumption and to improve safety. In present days, automation is not
only implemented in industrial problems but also in domestic applications. This paper
describes the application of automation in one of the work in kitchen called onion peeling
and cutting. Onion peeling and cutting is one of the fatigues and time consuming job for
preparation of food and it is everyday activity in all food processing system. In hotels,
hostel mess and restaurants high man power is required for the peeling and cutting the
onion and also it consumes more time. On the other hand there is a shortage of workers for
these types of works. Hence, there is a good opportunity for the implementation of
automation in onion peeling and cutting machine. The objective of this project is to
implement the mechatronics approach for automating the onion peeling and
cutting operation.
 Comment :The introduction of this abstract is incomplete. It is missing the general
sentence and statement. The others are fully and good. This mechanical redesign
is a positive step forward for the industry and is sure to bring significant benefits
to those who use these machines.
III/ Self introduction
Hello Mr. Dam!
Have a good day , I’m very happy to be here!
Thank you for giving me the chance to have an interview today.
First of all, My full name is Nguyen Nhat Nam. I’m 20 years old and come from Long An
province. I graduated from HCM University of Technology and Education. I have 2 years
of experience working in mechanical product design in design company such as Bosch,
Intel, …..I can use fluen a lot of software such as Autocad, inventor,… Beside I can fluent
foreign language which is English. I have several strengths such á creativity, leadership
and experience in teamwork. Today Im there , I want apply for a design position in
designing mechanical products.
After becoming a part of the company, I will do my best to contribute to the company
My strengths are confidence, positive attitude and progress at work.
So, that’s all about me!
Thank you for listening! And I’m very thankful if I can have a job in your
company

IV/ Research paper

 Available online at www.sciencedirect.com

ScienceDirect
Procedia CIRP 93 (2020) 868–872
www.elsevier.com/locate/procedia

53rd CIRP Conference on Manufacturing Systems

A sensor framework for combined data streams and in-situ characterization

of machining processes
Darya Botkinaa,∗, Bernd Peukertb , Tilman Gebhardb , Benedikt Schwarzb , Andreas Archentib ,
Thomas Lundholma
a KTH Royal Institute of Technology, PMH Application Lab, Brinellvägen 68, Stockholm 10044, Sweden
b KTH Royal Institute of Technology, Department of Production Engineering, Brinellvägen 68, Stockholm 10044, Sweden

Abstract
Machining vibrations is a critical phenomenon in the industry as they negatively affect the quality and tool-life. One common avoidance strategy for
machining vibrations is the fine-tuning of process parameters, thus leading to longer production time. Our research addresses this challenge and
uses different streams of data to classify problematic processes. Data streams of machining parameters, tool position, loads, vibration sensors,
together with process plan data and cutting tool usage information, are visualized. Experiments are performed to derive classification criteria. These
results are then used to observe vibrations in a five-axis machining center for further process adjustment.
© 2020 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the scientific committee of the 53rd CIRP Conference on Manufacturing Systems
Keywords: sensor integration; data fusion; manufacturing data analytics; cyber-physical manufacturing;

chronous streams for the accurate in-situ processing is a key for

1. Introduction a deterministic performance of the implementation. The second
challenge is the classification of machining vibrations with the
Due to their reliability and simple installation, sensors and goal to identify machining signatures which are not in alignment
transducers play and important role in modern machine tools. with the targets of the process engineer. A common example is
An ideally placed sensing system detects crucial process vari- distinguishing stable and unstable cutting conditions.
ables, analyzes their influence and visualizes the data for fast
detection of faults. In perspective, the correlation of the sensor
data with other data streams provides an opportunity to analyze 2. State of the art
the influence of vibration and machining parameters.
There are several parameters which directly influence the sig- 2.1. Sensor integration and monitoring
nature of the process, e.g. tool wear, temperature profile, material
hardness, lubrication, and the clamping of tool and workpiece. The recent advent of high-speed networks and inexpensive
Collected and analysed data can lead to longer tool life, faster computational power encouraged research and development of
and higher qualitative processes, smart maintenance, real time monitoring systems for machine tools [18], especially in-situ.
monitoring, and facilitate the shift towards smart manufacturing. Downey et al. [7] combine acoustic, force and vibrations sensors
Accordingly, the acquisition of data is fundamental for research to assess the tool wear during turning operations. They also
to get an in-situ assessment of the current situation. present an imaging system for the real-time wear detection.
One of the main challenges for an accurate and reliable in-situ Archenti [2] and Österlind [15] use microphone recordings of the
process characterization is the realization of a robust fusion of cutting process for the detection of critical vibration conditions.
the various incoming data streams. Synchronizing those asyn- Industrial solution are the DAS+ vibration suppression sys-
tem by the company Bimatec Soraluce Machining Technology
∗ Corresponding author. Tel.: +46 8 790 6339 GmbH, Lahn, Germany, [3], and MVC (Machine Vibration Con-
E-mail address: [email protected] (Darya Botkina). trol) by the company DMG Mori AG, Bielefeld, Germany, [6],
with its capability to suppress chatter tendencies based on real-
time acceleration data from the spindle.

2212-8271 © 2020 The Authors. Published by Elsevier B.V.

This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the scientific committee of the 53rd CIRP Conference on Manufacturing Systems
10.1016/j.procir.2020.05.155

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 Darya Botkina et al. / Procedia CIRP 93 (2020) 868–872 869

One drawback of these industrial solutions is the lack of 3. Digital infrastructure

insight into the underlying architectures and algorithms. Never-
theless, manufacturers of new machine tools nowadays usually 3.1. MTConnect
equip their products with a multitude of sensors, e.g. power and
current sensors as well as accelerometers. This way, the systems MTConnect is a technical standard (ANSI/MTC1.4-2018) for
remain adaptable to future findings and leverage the research on communication between computer numerical control and other
monitoring solutions with built-in unprocessed technology [18]. systems. In our experimental setup MTConnect is implemented
At the same time one of the challenges remains a brownfield as a web socket running on the HMI, to download machine tool
approach and the connectivity of machine tools already used on internal data [19]. These data provide us with information regard-
the production line but not well-equipped with modern sensors ing tool positions and machining parameters, as, e.g. feed rate
and systems. The key is not only to connect the machine but to and spindle speed. The sample rate of the MTConnect adapter
understand the context of the data and turn data into knowledge. of 2 Hz limits the matching of data and needs to be improved
In addition, the collected data can be used for statistical analysis, to further investigate the change of positioning respective to
as well as for further process improvement. vibrations. Each sample consists of timestamp and data.

3.2. LISA - the line information system architecture

2.2. Machining process characterization
LISA or the line information system architecture has been de-
Stability lobe diagrams (SLDs) are representations of ma-
veloped as an event-driven service-oriented architecture, which
chine tool stability, typically showing the limit of stable cutting
allows avoiding point-to-point communication [1]. This develop-
over various spindle speeds or depth of cuts. Depending on the re-
ment has been initiated by KTH Royal Institute of Technology,
quired model order, the diagram follows from single or multiple
and aimed for the brownfield connectivity of existed resources
degrees of freedom (DOFs) oscillator equations [17]. Friedrich
and applications.
et al. [10] used an accelerometer close to the spindle bearing and
The following development is called ”the tweeting machine”
forwarded the results into a machine learning algorithm, which
[8]. The machine controller is connected to an enterprise service
continuously improves the accuracy of the diagram.
bus, carrying out messages transfer. Each message consists of a
Goyal et al. [12] present a contactless monitoring approach
timestamp, message topic for describing a type of variables, and
via laser scanning of bearings. An inexpensive micro-controller
message body with data.
is used to process the laser signal in frequency domain and de-
The tweeting machine is based on the international standard
tect short-termed pulses indicating irregularities in the rotational
ISO 10303-AP238 [13] for describing a context of a process.
motion. Shrivastava and Singh [16] decompose the signal of
Among the defined variables are information about product,
acoustic emission sensors and use a Hilbert-Huang transforma-
working step, feature, cutting tool, and many others.
tion (HHT) allowing for the in-situ detection of chatter. They
The described infrastructure provides an access to the needed
later present a more advanced procedure with potentials also in
data for tracking manufacturing process, events and their charac-
highly non-linear cutting conditions [20, 9].
teristics.
Zhang et al. [22] present an in-situ monitoring approach
for predicting the tool wear using accelerometers, current sen-
sors, and piezoelectric disk sensors for cutting force measure- 3.3. Vibration sensors
ments. They forwarded the data of a milling process to a micro-
controller connected to a database managed by Apache Spark As outlined previously, a modern machine tool manufacturer
for further analysis. equips new machine tools with a variety of standard sensors. In
Yeh and Lai [21] measure vibrations and cutting forces dur- the Hermle C 50 U, two tri-axial accelerometers of the company
ing a turning operation of slender workpieces. Their algorithm Dytran Instruments, Inc., Chatsworth, United States of Amer-
takes filtered samples of a fixed size and observes the change of ica, are bolted to the spindle housing. Although both sensors are
the standard deviation σ. Gandhi [11] analyzes the measured cut- located close to the spindle bearings, the upper sensors distance
ting force’s frequency spectrum to conclude assumptions about makes it more suitable to monitor the spindle housing vibrations
the time-dependent system behaviour for different machining than the actual spindle. It is noted, that the sensors’ coordinate
parameters. systems is rotated in reference to the machine tool coordinate
Cuka et al. [5] used two accelerometers at the spindle hous- system.
ing to assess the vibrations of a slot-milling process of inclined
surfaces of aluminium parts. Besides performing standard fre-
quency analysis based on the fast Fourier transform (FFT) [4], 4. Framework implementation
the team investigated the use of the Teager–Kaiser energy op-
erator (TEO) and the Gabor filter. TEO is a computationally A framework schema is presented in Fig. 1 showing how data
efficient algorithm which is very sensitive to fast frequency and from different sources in a CNC machine were gathered and
amplitude changes [14] while the Gabor filter can be used for classified. One can essentially identify 3 stages in the whole
signal decomposition and feature extraction. process:

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 870 Darya Botkina et al. / Procedia CIRP 93 (2020) 868–872

Fig. 1: Framework for the data transmission

4.1. Collection of data 4.3. Classification criterion

Different data streams are combined using LabVIEW Virtual In our classification we consider the frequencies of the tool
Instrument to control the accelerometers and convert analogue and the machine tool using FFT. For this application the clas-
voltages to digital values. These values are stored in CSV-files sification criterion distinguishes between stable and unstable
and sent to a Python application via a TCP protocol. Addition- signals. The stable signal has a tight concentration of a few fre-
ally, data from the CNC machine’s PLC is sent to the same quencies with peak magnitudes below 0.01 m/s2 . The dominant
application via the tweeting machine and the MTConnect agent. frequencies are the harmonic ones, and the sums of harmonic
For each data source a thread of live data is created in the Python and non-harmonic with peaks that do not exceed 0.0025 m/s2 .
application. Data are matched by timestamps. An unstable process exhibits more harmonic frequencies (more
Timestamps of data streams from MTConnect and LabVIEW than 7) and dominant non-harmonic ones, see Fig. 3.
are formatted identically to enable a comparison. The data from
the two streams are saved in a new variable whenever the times-
tamps match. The LabVIEW data stream may have a rather fine
sampling rate for which more than one timestamp per millisec-
ond may occur. In this case, if the application detects several
identical timestamps, it selects the first one for matching. This
procedure is repeated throughout the measurement cycle. The
logic is presented in Fig. 2.

Fig. 3: Example classification approach

Fig. 2: Data matching logic based on timestamps

5. Results

The manufacturing of the so-called Boxy test piece covers a

variety of machining operations and is selected as a demonstrator
4.2. Processing step for the developed data fusion framework. It is the result of
a standardization effort for STEP-NC data and resembles the
The vibration data are analysed and compared against refer- manufacturing of a gearbox housing. The performed operations
ence data. The reference data are created using an experimental include roughing and finishing operations, i.e. surface milling,
methodology which is not restricted to the particular CNC ma- pocketing, and drilling, see Fig. 4. The piece is clamped in
chine, but can be applied for other manufacturing processes. various orientations in a machine tool vice.

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 Darya Botkina et al. / Procedia CIRP 93 (2020) 868–872 871

Fig. 4: Manufacturing of the Boxy test piece with various manufacturing operations

To establish a sufficiently large database, 25 Boxys were Lastly, the data were analysed for consistently appearing
manufactured and the fused data streams were recorded as con- frequencies that are present whether an operation is ongoing or
tinuous time series. It is beyond the scope of this paper to identify not. This was done to find out if the machining center naturally
performance criteria for all the different manufacturing opera- produces certain frequencies simply from running in idle mode.
tions. However, for demonstration purposes, the authors present
the experiments on slot milling which is an integral part of the
pocketing operation. 6. Outlook
The vibration measurements were performed with a sampling
rate of 2000 Hz in the machining center, using an end mill. Two The provided framework presents an approach to manufac-
extra similar end mills were provided to investigate vibrations turing data acquisition, processing and analysis with data from
due to a used and a new tool. various data sources. This approach allows for a better descrip-
The end mill was gauged for an initial depth of 3 mm for the tion and deeper understanding of results and behavior, as well
first seven out of nine pieces. The tool cut a straight line towards as to provides the possibility to track the process conditions.
the higher cutting depth. To avoid different cutting conditions The solution gives a method to identify critical machining
and to engage both sides of the cutting tool, a 2mm separation conditions. At the same time there challenges and limitations
was left between the tracks. Different cutting conditions were remain, among them the following:
created by varying the spindle speed and the feed rate.
The machining data of several Boxys was analysed as follows.
1) The current sampling rate of the MTConnect of 2HZ is
Raw data from the CSV files were searched for peaks in the
not sufficient for high machining speeds.
amplitudes of every of the six axes measured. Those peaks were
2) Our criterion is not applicable to a classification of chatter,
further compared with other measurements of the same setup but
such classification would require further modelling.
for different Boxys. If a peak appears consistently, a correlation
3) The dynamics between housing and the spindle could be
can be made with the tool penetrating the material for the first
valuable for the vibration analysis. This would definitely
time. A peak or another anomaly might be a sign of changing
influence a formulation of any universal criterion.
forces within the cutting process leading to chatter.
A second method used were waterfall plots. Waterfall plots
Our analysis should be included as a first step for any follow-
consist of FFTs of the raw data stacked up according to the time
ing work. Accuracy can be improved by measuring matching
of the data. A Hanning window with 0.5 overlap was applied,
parameters and tool positions at higher sample rate.
enclosing a defined number of values to create conditions more
Using such system on an industrial PC continuously can
suited for the FFT. This approach allows for an identification of
improve immediate response to the wrong machining param-
shifting frequencies over time. Such shifts were correlated with
eters. In addition to that, the application of machine learning
setups, their regularity within the data was analyzed.
algorithms could give new insights into data. Another technical

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 872 Darya Botkina et al. / Procedia CIRP 93 (2020) 868–872

improvement can be the Hilber-Huang transformation, which [20] Wu, Z., Huang, N.E., 2009. Ensemble empirical mode decomposition: A
has been used in a related research. noise-assisted data analysis method. Advances in Adaptive Data Analysis
doi:10.1142/S1793536909000047.
[21] Yeh, L.J., Lai, G.J., 1995. A Study of the Monitoring and Suppression
Acknowledgements System for Turning Slender Workpieces. Proceedings of the Institution
of Mechanical Engineers, Part B: Journal of Engineering Manufacture
doi:10.1243/PIME_PROC_1995_209_077_02.
This work is supported by VINNOVA project 2017-05208 [22] Zhang, X.Y., Lu, X., Wang, S., Wang, W., Li, W.D., 2018. A multi-sensor
Nationell Testbädd Smart Produktion. based online tool condition monitoring system for milling process, in: Pro-
cedia CIRP. doi:10.1016/j.procir.2018.03.092.

References

[1] Alfred Theorin, Kristofer Bengtsson, J.P.M.L.C.J.T.L.B.L., . International

Journal of Production Research doi:10.1080/00207543.2016.1201604.
[2] Archenti, A., 2011. A Computational Framework for Control of Machining
System Capability : From Formulation to Implementation. Ph.D. thesis.
KTH, Machine and Process Technology.
[3] Bimatec Soraluce Machining Technology GmbH, 2020. DAS+ Dynam-
ics Active Stabilizer. URL: https://www.bimatec-soraluce.com/
innovation/the-intelligent-machine/das/.
[4] Cooley, J.W., Lewis, P.A., Welch, P.D., 1969. The Fast Fourier Transform
and its Applications. IEEE Transactions on Education doi:10.1109/TE.
1969.4320436.
[5] Cuka, B., Cho, M., Kim, D.W., 2018. Timely and efficient chatter onset
detection for end-milling, in: Procedia Manufacturing. doi:10.1016/j.
promfg.2018.10.029.
[6] DMG Mori AG, 2020. MVC(Machine Vibration Control).
URL: https://www.bimatec-soraluce.com/innovation/
the-intelligent-machine/das/.
[7] Downey, J., Bombiński, S., Nejman, M., Jemielniak, K., 2015. Automatic
multiple sensor data acquisition system in a real-time production environ-
ment, in: Procedia CIRP. doi:10.1016/j.procir.2015.06.039.
[8] Eliwi Roa, E.S., 2017. The tweeting machine – CNC-PLC interface.
[9] Flandrin, P., Rilling, G., Gonçalvés, P., 2004. Empirical mode decomposition
as a filter bank. doi:10.1109/LSP.2003.821662.
[10] Friedrich, J., Hinze, C., Renner, A., Verl, A., Lechler, A., 2017. Estimation
of stability lobe diagrams in milling with continuous learning algorithms.
Robotics and Computer-Integrated Manufacturing doi:10.1016/j.rcim.
2015.10.003.
[11] Gandhi, N., 2012. Fft based evaluation of cutting forces and chatter vibra-
tions in turning by varying speed, feed, depth of cut and rake angle, pp.
1–4.
[12] Goyal, D., Vanraj, Pabla, B.S., Dhami, S.S., 2019. Non-contact sen-
sor placement strategy for condition monitoring of rotating machine-
elements. Engineering Science and Technology, an International Journal
doi:10.1016/j.jestch.2018.12.006.
[13] ISO, 2012. ISO 10303-238:2007 Industrial automation systems and integra-
tion — Product data representation and exchange — Part 238: Application
protocol: Application interpreted model for computerized numerical con-
trollers.
[14] Kaiser, J.F., 1990. On a simple algorithm to calculate the ’energy’ of a signal,
in: ICASSP, IEEE International Conference on Acoustics, Speech and Signal
Processing - Proceedings. doi:10.1109/icassp.1990.115702.
[15] Österlind, T., 2013. An Analysis of Machining System Capability and
Its Link with Machined Component Quality. Ph.D. thesis. KTH, Machine
and Process Technology.
[16] Shrivastava, Y., Singh, B., 2019. A comparative study of EMD and EEMD
approaches for identifying chatter frequency in CNC turning. European
Journal of Mechanics, A/Solids doi:10.1016/j.euromechsol.2018.10.
004.
[17] Siddhpura, M., Paurobally, R., 2012. A review of chatter vibration research
in turning. doi:10.1016/j.ijmachtools.2012.05.007.
[18] Teti, R., Jemielniak, K., O’Donnell, G., Dornfeld, D., 2010. Advanced mon-
itoring of machining operations. CIRP Annals - Manufacturing Technology
doi:10.1016/j.cirp.2010.05.010.
[19] Wiese, A., 2017. Development of a web-based machine-tool monitoring
system.

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 Available online at www.sciencedirect.com

ScienceDirect
Procedia CIRP 93 (2020) 891–896
www.elsevier.com/locate/procedia

53rd CIRP Conference on Manufacturing Systems

Performance assessment methodology for AI-supported decision-making in

production management
Peter Burggräfa, Johannes Wagnera, Benjamin Kokea*, Milan Bamberga
a
Chair for International Production Engineering and Management, University of Siegen, Siegen, Germany
* Corresponding author. Tel.: +40 271 740-2631; E-mail address: [email protected]

Abstract

Artificial intelligence (AI) gains importance in many domains and may soon be transferred decision-making responsibilities in production
management from production managers. For the future, it will be vital to identify each entity’s domain of decision-making superiority. Therefore,
this paper proposes and applies a model to assess AI performance in contrast to human decision-making. Relying on reinforcement learning and
item response theory, the approach describes a minimum viable setup for AI systems to identify opportunities for AI systems in manufacturing.
The model is based on operative production management decisions (job-shop scheduling) and validated through a series of academic scheduling
instances.
© 2020 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the scientific committee of the 53rd CIRP Conference on Manufacturing Systems
Keywords: Production; Artificial intelligence; Scheduling, Performance; Reinforcement Learning; Item Response Theory

1. Introduction considerations of possible implications of introducing AI into

PM [2].
During the last years, artificial intelligence (AI) has been When AI systems are considered to take over PM tasks, their
resuscitated to become a trending topic. A prominent number use should be evaluated in comparison to human performance
of achievements in AI research have foreshadowed what instead of merely on their performance metrics like accuracy or
technical advances we can expect. AI systems have scored precision. In our previous work, we put forward the notion of
super-human results in domains which experts believed would Cyber Production Management Systems (CPMS) [2]. A model
remain human-dominated for years to come. describing a CPMS comprises of three sub-models namely
AI can have a tremendous impact on re-organizing future Trust (design principles for trustworthy AI systems), Control
production management (PM). It is being said that the merger (ensuring ethical and explainable AI systems) and Perform
of the physical and digital world will shift work in production (assigning the appropriate decision-making responsibility to AI
environments towards decision-focused and creative tasks, as systems).
human flexibility and creativity is assumed to be irreplaceable This paper is attributed to our research activities within the
by machines as of now [1]. Today, research focusses on using CPMS-Perform model. Its purpose is to propose a performance
AI to enhance existing operational processes, yet we assert that assessment methodology to enable performance comparisons
true efficiency and productivity gains will be achieved by between human and AI decision-making. The approach
introducing AI into higher-level PM functions. However, while presented herein has been tested in an experimental production
much work deals with designing better algorithms to increase scheduling setting involving student participants.
AI performance, research appears to lack high-level

2212-8271 © 2020 The Authors. Published by Elsevier B.V.

This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the scientific committee of the 53rd CIRP Conference on Manufacturing Systems
10.1016/j.procir.2020.03.047

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 892 Peter Burggräf et al. / Procedia CIRP 93 (2020) 891–896

Nomenclature instance-wise performance of multivariate and binary

classifiers on different datasets, respectively. They show that
i index of problem case (𝑖𝑖 ∈ 1,2) by using IRT, training and test datasets can be improved as this
j index of examinee in sample (𝑗𝑗 ∈ 1, … ,22) method identifies outliers. These occur as items with negative
JSSP Job-Shop-Scheduling problem discrimination, potentially caused by incorrectly labelled data.
n number of jobs in JSSP In a similar vein, Lalor et al. [17] find that, similar to human
m number of machines in JSSP learning, deep neural networks perform better on simple
k index of round in reinforcement learning experiment Natural Language Inference tasks, as easy items are learned
K number of rounds matching human sample (K=22) more quickly than hard items if more data is introduced in
s current state in state space S training. While this may not be surprising, as human expertise
a current action in action space A grows with learning effort as well, this has not been
R reward for performing action a on state s demonstrated to be equally true for deep neural networks.
Rk result of algorithm at step k (𝑅𝑅𝑘𝑘 ∈ 0,1) These examples show that common machine learning metrics
such as accuracy are insufficient to report performance,
because varying problem difficulty plays a vital role in
2. Background achieving these results. Lastly, IRT was used to assess the
performance of RL algorithms in playing Atari games [3,18].
AI systems have scored super-human results in domains Similar to the work in [15,16], Martınez-Plumed & Hernández-
which experts believed would remain human-dominated for Orallo [3] find that not all games are useful to assess algorithm
years to come [3–5]. The efforts to compare the performance performance. As they have a negative discrimination factor,
of intelligent machines to humans have grown ever since the they are unfit to detect improvements of techniques over the
defeat of chess master Garri Kasparov. Chess players are benchmark, because difficult games were most often solved by
evaluated according to the Elo rating, which represents the low scoring AIs. Building on their previous work, Martínez-
expected performance of a player based on their previous Plumed & Hernández-Orallo [18] introduce the parameter
results [6]. Chess computers surpassed the grandmaster Elo generality to evaluate whether an agent is consistently good at
rating of 2,500 in the early 2000s. Today, the best chess easy problems and bad at difficult ones. This problem is well-
computer scores 3,500 Elo while the chess master comes in at known in RL algorithms as they are keen on finding loopholes
2900 [7]. that allow them to continuously exploit small rewards while
In Go, an Asian board game similar to chess, while the ostracizing long-range gains.
current leading human player scores 3,652 Elo, the first AI that The above-mentioned examples demonstrate the wide range
beat a former Go champion (AlphaGo) in 2016 scores 3,739 of assessment approaches for comparing AI to human
Elo and its successor, AlphaGo Zero, scores 5,185 Elo [8]. performance. The purpose of this paper requires an approach to
Image recognition is another cognitive domain where measurements that is flexible for use on several problems.
algorithm performance is often assessed against humans. For Based on the above research, we deem IRT a proper solution to
image recognition, AI performance is typically benchmarked apply to our model. We describe its full design in the following
in a competition using the ImageNet database [9]. The human section.
error rate in this dataset is 5%, which over 70% of AI
competition participants undercut in 2017 [10]. 3. Methodology
In reading comprehension, the Stanford Question
Answering Dataset is used to evaluate AI performance. The Decision theory in general is a complex realm, as decisions
dataset consists of over 100,000 crowdsourced question- are influenced by a number of factors that can make the
answer pairs. The human error in this test is 13.2% based on decision process arbitrarily complex [19]: information
inter-annotator agreement [11] which has recently been richness, single-agent vs. multi-agent decisions,
undercut by two AIs [12]. intertemporality, frequency and longevity of decisions, or
A research area that evaluates AI performance without the complexity (e.g. infinite time and intelligence for decision-
need for large training data sets is arcade computer games, making vs. bounded rationality; number of alternatives to be
where super-human performance is said to have been reached considered).
by now [3,13]. The performance of deep Q-learning networks To put forward our assessment methodology, we selected a
(DQN) that often excel at old Atari games (which have by now low-complexity production problem with well-known
become the benchmark problem for reinforcement learning solutions and heuristics so as to obtain unambiguous
(RL) evaluation [14]) is calculated with a human game score experiment results to feed into the assessment. In a production
that is achieved by a professional games tester. This calculation setting, this setup requires three components:
is essentially comparable to the calculation of Elo, as is the (1) a production system affected by sequential decision-
source of benchmarked scores. making,
Recently, the psychometric assessment method ‘item (2) an AI as decision-maker, and
response theory’ (IRT), traditionally used in educational (3) a methodology to compare human decisions to AI-made
testing, has been applied a number of times to assess the ones.
performance of different machine learning techniques. The following subsections discuss details about the design
Martınez-Plumed et al. [15] and Lalor et al. [16] analyzed the of these components.

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 Peter Burggräf et al. / Procedia CIRP 93 (2020) 891–896 893

3.1. Design of the production system 𝑞𝑞(𝑠𝑠𝑡𝑡 , 𝑎𝑎𝑡𝑡 ) ← 𝑞𝑞(𝑠𝑠𝑡𝑡 , 𝑎𝑎𝑡𝑡 ) + 𝛼𝛼 (𝑟𝑟𝑡𝑡+1 + 𝛾𝛾 max 𝑞𝑞 (𝑠𝑠𝑡𝑡+1 , 𝑎𝑎𝑡𝑡 ) − 𝑞𝑞(𝑠𝑠𝑡𝑡 , 𝑎𝑎𝑡𝑡 )) ȋͳȌ
𝑎𝑎

To put forward our assessment methodology, for this study Herein rt+1 is the reward incurred from executing the action
we selected low-complexity production tasks, namely static, in the environment, 𝛼𝛼 and 𝛾𝛾 are control parameters: 𝛼𝛼 (𝛼𝛼 <
non-reactive, single-agent job-shop scheduling problems 0 ≤ 1) is the learning rate. 𝛼𝛼 → 0 indicates no learning while
(JSSP) with perfect information and a relatively small solution 𝛼𝛼 → 1 gives higher precedence to future rewards. Similarly, 𝛾𝛾
space. The planning horizon is short so as to minimize controls the far-sightedness of the agent: for 𝛾𝛾 → 0, the agent
intertemporality issues, and sufficient time and (human) only considers current rewards (rt), whereas 𝛾𝛾 → 1 leads to
intelligence should be available. maximization of long-term gains [21]. As our implementation
We choose to regard a JSSP for the following reasons: they of the scheduling instances interprets the problems as
represent well-studied and important PM decision-making continuous tasks without terminal state, 𝛾𝛾 should not be 1 in
problems (e.g. see [20]), static look-ahead schedules can be order to discount future rewards to prevent them from
considered a perfect–information problem, and the problem accumulating infinitely.
size is scalable. As JSSPs can as well be modelled using an MDP approach
[22] and RL is known to be a good tool for solving JSSPs, see
3.2. Design of decision-making AI [20,22–24], we regard RL as a relevant AI technique for our
purpose.
This approach seeks to lay the groundwork for a generic
model that allows one to measure an AI’s performance on a 3.3. Item response theory
given task. Many AI and machine learning techniques, i.e.
supervised and unsupervised learning, rely on large amounts of As we seek a design that is relevant for different sets of
data from which to generate models. Another subset of AI tools tasks, a methodology that requires large background databases
is reinforcement learning (RL)[21]. RL can help to mimic and with human scores is unfit for this purpose. Furthermore,
automate the sequential human-decision making process. The benchmark results from arcade games seem unfit as their tests
goal is to solve a problem by finding the best policy of action, require uniformity across problems for comparability. The
irrespective of the deeper structure of the experience gathered performance assessment methodology has to cope with
during interaction with the environment. Thus, RL does not different scales of complexity in PM environments. IRT
require performing an intermediate system identification measures problem difficulty and the ability required for
process, enabling good learning results without the need for decision-making, which is why we have selected IRT as our
much prior knowledge of the problem [21]. method to compare human decisions to AI-made ones. IRT is
RL problems are usually modelled via a Markov Decision a psychometric test set originally designed to measure
Process (MDP) as depicted in Fig. 1. MDPs consist of a tuple examinees’ ability through a test with several questions (i.e.
(S,A,P,R) of which A represents all possible actions a that the items). It is a set of mathematical models that describe the
algorithm can use to change its state [21]. Every achievable relationship between a latent trait of interest, which is not
state for the RL problem is found within the state space S. For directly measurable, and an examinee’s answers to individual
every action a taken in state s a specific reward 𝑟𝑟(𝑠𝑠, 𝑎𝑎, 𝑠𝑠 ′ ) ∈ 𝑅𝑅 items, where the probability of a response for an item is a
arises which leads to an overall increase of return R. function of the examinee’s ability [25]. The basic 3-parameter
𝑃𝑃(𝑠𝑠 ′ , 𝑟𝑟 | 𝑠𝑠, 𝑎𝑎) describes the probability to reach a state s’ with IRT model, calculates the probability P of a correct response
reward r given current state s and taking action a. Additionally, Ui = 1 on an item i given the examinee’s ability θj as a logistic
a value function Q : S × A → ℝ indicates the expected value function:
of a state or state-action pair [21].
1−𝑐𝑐𝑖𝑖
𝑃𝑃(𝑈𝑈𝑖𝑖 = 1|𝜃𝜃𝑗𝑗 ) = 𝑐𝑐𝑖𝑖 + (−𝑎𝑎 (𝜃𝜃 −𝑏𝑏 ))
(2)
1+𝑒𝑒 𝑖𝑖 𝑗𝑗 𝑖𝑖
Agent
This model provides the item characteristics curve (ICC) for
each item tested and is described by the following parameters:
State Reward Action
St Rt
• Discrimination a: slope of the function at location
At
parameter b. Higher values indicate higher changes in item
Rt+1 response for small changes in ability;
St+1 Environment • Difficulty b (also called location parameter): describes the
point where an examinee has a 50% probability of
Fig. 1: The agent–environment interaction in a Markov decision process [21]
answering correctly (assuming no guessing);
• Guessing c: probability for correct answers for examinees
One distinct technique in RL is Q-learning. It is an off- with low ability.
policy learning algorithm, i.e. the policy being followed (often
epsilon-greedy) is different than the one being learned as q 4. Experiments
values q(st,at). Here, at is the action the agent takes on state st.
For each training epoch, the q-value is updated according to the To test the assessment model, we conducted a brief
following rule: experiment in which students competed against a simple AI.
We selected two academic scheduling problems as experiment

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 894 Peter Burggräf et al. / Procedia CIRP 93 (2020) 891–896

cases that are used in the chair’s exercises and lectures to train The answers were coded dichotomously, i.e. 0 or 1. The data
students in the field of operations research. In general, JSSPs were processed in R using the library ‘ltm’ [26]. Table 2 shows
consist of n jobs to be processed on m machines. For our study the results of the parameter calculations.
we regard the following two problems: the first (denoted as
‘Case A’) consists of atomic jobs (i.e. no sequence of Table 2. Overview over human cohort results from IRT parameter estimation.
operations) and lowest combinatorial complexity which, in an Parameter Case A Case B
n x m JSSP, has n=5 jobs and m = 1 machine. The problem has Discrimination a 5.3612 6.5651
to be solved by finding a sequence that has the least cumulative Difficulty b 1.0057 -0.0260
delay (minimum tardiness) regarding a target due date for each Guessing c 0.0679 0.3989
job. The second problem (‘Case B’) is a 5x3 Flow Shop
Scheduling Problem. The operations sequence is fixed for each Based on the participants’ answers, the results show that
job (M1→M2→M3). In contrast to Case A, the criterion for Case A (5 atomic jobs) was significantly harder to solve than
sequence optimality is makespan minimization. Each job can Case B. Notably, there is an almost 40% chance of scoring
only be passed on to the next machine after finishing operation correctly on Case A by guessing. The discrimination parameter
on the previous one. also scores lower in Case A than in case B, which indicates a
The key to solving these problems without exhaustive narrower spread in correct answers across the participants. The
computing resources is the knowledge of different heuristics mean ability of the respondents is computed to 𝜃𝜃𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 =
which had been taught to the students in lectures and exercises. 0.1363. The probabilities for correct answers in turn are
For the 5x1 problem, a suitable heuristic can be shortest process PA(𝜃𝜃𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 ) = 0.4045 for Case A and PB(𝜃𝜃𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 ) = 0.7613 for
time (SPT), while the 5x3 problem can be solved via Johnson’s Case B.
algorithm. For both cases, the optimal sequence was known to Fig. 2 displays the plots of the ICCs that correspond to the
the examiners. above parameters. The black curve represents Case A, and the
The overall logic of the experiment is as follows: in a red curve represents Case B. The vertical blue line represents
uniform sample of examinees each one answers two operations the mean ability. The horizontal dotted lines indicate the
research tasks with varying difficulty. The examinees are success probability for each case with respect to the computed
assumed to have different knowledge levels regarding possible participants’ mean ability 𝜃𝜃𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 .
heuristics to solve the tasks, meaning they all have a different
ability θ. Given different difficulties for Case A and Case B,
the mean ability over the entire sample will yield different
probabilities P(Ui=1) of finding the optimal sequence.
Subsequently, we will run the RL experiment to estimate the
number of training epochs required to match the human
probability scores. Details of the setup and results for both
cohorts are presented in the following two sub-sections.

4.1 Human cohort results

We appointed 22 students to answer the cases. For both

cases, the task was to find the optimal job sequence. As IRT
demands that all examinees must answer the same task, we
drafted the students from the same cohort. The experiment was Fig. 2: Item characteristics curve plots for both items
run under exam conditions, meaning that both problems had to
be solved under the time constraint of three minutes. As 4.2. AI solution with reinforcement learning
additional auxiliary equipment, only calculators were allowed
In the next step, we determined the amount of training
for computations, if required. Participants were asked to
needed by the RL algorithm to match the human probability
provide an answer by the end of the time and to indicate
ratings. The algorithm was implemented in the programming
whether they had guessed the answer in order to estimate the
language R using the ‘ReinforcementLearning’ library
IRT parameter c. Table 1 shows the cases presented to the
[27]. The environment design for the agent to interact with is
examinees.
described in the following paragraphs.
The state space S for these problems is finite, meaning that
Table 1. Validation cases process times (in min for Case A and h for Case B).
all possible states are fully observable and known a priori. For
Case A Case B
both problems we used five distinct jobs A through E. The state
Job Duration Due Date M1 M2 M3
space is thus defined as all possible combinations that these
A 1,200 1,600 8 5 10
jobs can take, yielding a total complexity of 5! = 120
B 2550 5,100 16 3 7
combinations.
C 400 600 4 5 15
For the agent to change between states, the action space A
D 1,250 6,250 12 4 8
contains eleven possible actions for the agent to choose from
E 850 2,500 5 6 11
which can be summarized as two different choices. The agent
can either pick one of ten options a(d,e) with d, e ϵ {1;2;3;4;5}

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 Peter Burggräf et al. / Procedia CIRP 93 (2020) 891–896 895

and d < e to swap jobs corresponding to the stated numbers P(U1/2=1) =1. From the ICCs in Fig. 2 we see that these match
within the action. For instance, applying the action a(1,3) to state an ability of 𝜃𝜃 → ∞. A 100% chance of finding the correct
ABCDE will result in state CBADE. Alternatively, the agent sequence is achieved at ability scores around 2.
can choose use the (d, d) action, referring to an idle action,
since the one job would be swapped with itself, resulting in the
same sequence. While the (d, e) actions are used to search for
a better sequence of jobs, the ‘idle’ action is meant for high-
reward states, in which the agent is uncapable of improving the
sequence, hence yielding a better reward for keeping the
current job sequence. This way, the agent will learn the optimal
policy of swaps in order to reach the state corresponding to the
idle action where the highest reward is incurred by doing
nothing, which is described as q* (“Sopt”, “11”).
Rewards for taking an action on state s are attributed as
follows: for Case A, the cumulative delay on all jobs is fed back
as (negative) reward. The optimal sequence incurs 0 delay and
thus a reward of 0. Earlier finishes did not incur any positive Fig. 3: Results of RL training (circles), logarithmic regression (dashed lines),
reward. As zero rewards are to be avoided in order to make the and no. of training samples at human probability score (solid lines)
optimal sequence stick out from an otherwise sparsely
populated matrix, a small positive reward of 10 is given for 5. Discussion
delays of zero at the end of the episode instead. The rewards
for Case B are shaped similarly: The optimal sequence The results presented in the previous section show that very
minimizes makespan of all jobs, which is a number known a low training effort is required to match human ability for this –
priori. (Negative) rewards are computed as difference between purposefully simple – production case. As the RL ability scores
optimal and actual makespan, with the optimal sequence again surpass the human scores, we infer that the specific PM
incurring 0 reward which is being increased to 10. decision presented herein can securely be submitted to an AI.
For the experiments, we increased the amount of training At the core of this paper was a demonstration for Q-learning to
samples from 10<n<1,000 in 10 step increments. For each be a viable option for low-complexity production decisions with
round, the algorithm computed K=22 epochs to match the factorial alternatives, low intertemporality and perfect
amount of human participants. After each epoch, the state information. We induce that for similar production
corresponding to the maximum q-value for action a(1,1) was management decisions for which the same criteria apply, AI
output. Training was done with the control parameters 𝛼𝛼 = 0.2, will be equally able to outperform human decision-makers.
𝛾𝛾 = 0.8 and 𝜖𝜖 = 0.3. Furthermore, by expanding our setup with more complex
As this experiment did not require any expensive cases as well as different algorithms in the future, we can
computations and could be performed on a regular office compose a solution space that identifies decision-algorithm
laptop, the following comparisons of results refer to the outputs pairs with par-human, high-human or even greater ability
only. We omitted a cost analysis on the input side while scores, meaning that these PM tasks are automatable.
acknowledging that this is an important aspect for future work.
For the comparison of human to AI performance, a uniform Limitations and future research needs
assessment metric is required. IRT results are computed as
probability of correct response. Thus, we express the While we assert that our performance assessment
probability of correct answers as share of the right job methodology can play a vital role in the design of future Cyber
sequences found by the AI. The computation rule for the RL Production Management Systems with respect to the crucial
algorithms probability rating is given in Eq. 3. decision, which PM decisions should be automated with AI
support, we acknowledge that the approach and results
∑𝐾𝐾
𝑖𝑖 𝑅𝑅
presented herein are only a first step. Our research was subject
𝑃𝑃(𝑈𝑈𝑖𝑖 = 1|𝑛𝑛𝑘𝑘 ) = 𝑘𝑘
(3) to the following methodological limitations that require further
𝐾𝐾
consideration in the future:
Eq. 3 calculates the probability of finding a correct sequence First, following previous approaches in AI performance
for Case i given the number of training samples n in participant research [15–18], we employed item response theory to model
k as the sum of correct responses Rk per episode divided by 22. the mean ability of respondents. In this case, IRT only uses
Rk was assigned the value 1 if response of k corresponded to the dichotomous data (correct or incorrect answers) which we
known optimal sequence and 0 otherwise. expect to not be applicable at higher order problems. However,
Fig. 3 shows the training results. The Q-learning algorithm IRT also provides models for polytomous items such as Likert
exhibits next to equal performance on both cases, indicated by scales or normalized data. Future research should thus
the dashed logarithmic fit lines. Human performance for Case incorporate grading or assessment schemes that allow to
A (black, ~40% chance of answering correctly) is met after distinguish answers on a more delicate scale, i.e. normalized to
about 40 samples. For Case B (red, ~76% chance), 220 samples the best answer given or as fuzzy sets.
are needed. After a full training cycle (1,000 epochs), the Second, our sample of human participants consisted of
probability of correct answers both cases converges to production management students exclusively. The latent trait

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 896 Peter Burggräf et al. / Procedia CIRP 93 (2020) 891–896

we estimated from the responses described the knowledge of ȏʹȐ —”‰‰”¡ˆǡƒ‰‡” ǡ‘‡ȋʹͲͳͺȌ”–‹ˆ‹ ‹ƒŽ‹–‡ŽŽ‹‰‡ ‡‹’”‘†— –‹‘
ƒƒ‰‡‡–ǣ”‡˜‹‡™‘ˆ–Š‡ —””‡–•–ƒ–‡‘ˆƒˆˆƒ‹”•ƒ†”‡•‡ƒ” Š–”‡†•‹
heuristics to solve the problem. It remains to be investigated ƒ ƒ†‡‹ƒǤ2018 International Conference on Information Management and
how a sample of manufacturing experts would perform on the Processing (ICIMP)ǡ’’Ǥͺʹ–ͺͺǤ
ȏ͵Ȑ MartınezǦŽ—‡† ǡ ‡”ž†‡œǦ”ƒŽŽ‘ ȋʹͲͳ͸Ȍ ”‡•—Ž–•ˆ‘”–Š‡–ƒ”‹ʹ͸ͲͲ
same problems and whether their performance would yield ‰ƒ‡•ǣ†‹ˆˆ‹ —Ž–›ƒ††‹• ”‹‹ƒ–‹‘—•‹‰ ǤEGPAI, Evaluating General-
substantially different outcomes in terms of training samples Purpose Artificial Intelligence ͵͵Ǥ
ȏͶȐ ‹Ž˜‡”ǡ —ƒ‰ǡƒ††‹•‘ ǡ —‡œǡ‹ˆ”‡ǡ˜ƒ†‡”‹‡•• Š‡ ǡ
required for the Q-learning algorithm.  Š”‹––™‹‡•‡” ǡ–‘‘‰Ž‘— ǡƒ‡‡”•Š‡Ž˜ƒǡƒ –‘–ȋʹͲͳ͸Ȍƒ•–‡”‹‰
Third, we stress once more that demonstrating AI usefulness –Š‡‰ƒ‡‘ˆ ‘™‹–Š†‡‡’‡—”ƒŽ‡–™‘”•ƒ†–”‡‡•‡ƒ” ŠǤnature
ͷʹͻȋ͹ͷͺ͹ȌǣͶͺͶǤ
for production research is not the purpose of this paper, but ȏͷȐ ”‘™ǡƒ†Š‘ŽȋʹͲͳͻȌ—’‡”Š—ƒ ˆ‘”—Ž–‹’Žƒ›‡”’‘‡”ǤScience
demonstrating AI power in relation to human problem-solving ͵͸ͷȋ͸Ͷͷ͸Ȍǣͺͺͷ–ͻͲǤ
ȏ͸Ȑ ‘—Ž—ȋʹͲͲͺȌŠ‘Ž‡ǦŠ‹•–‘”›”ƒ–‹‰ǣƒ›‡•‹ƒ”ƒ–‹‰•›•–‡ˆ‘”’Žƒ›‡”•
skills. In fact, the problem cases used herein are too simple to ‘ˆ–‹‡Ǧ˜ƒ”›‹‰•–”‡‰–ŠǤComputers and GamesǤ’”‹‰‡”Ǥ‡”Ž‹ǡ ‡‹†‡Ž„‡”‰Ǥ
warrant the use of an AI. Further research on this topic will ȏ͹Ȑ ™‡†‹•ŠŠ‡••‘’—–‡”••‘ ‹ƒ–‹‘Ǥȋƒ ‡••‡†‘ͳʹǤͲͳǤʹͲͳͻȌǤThe SSDF
rating listǤŠ––’ǣȀȀ••†ˆǤ„‘•Œ‘Ǥ‡–Ȁȋƒ ‡••‡†‘ʹͺǤͲͳǤʹͲʹͲȌǤ
therefore use more elaborated RL techniques as well as more ȏͺȐ ‹Ž˜‡”ǡ Š”‹––™‹‡•‡” ǡ‹‘›ƒǡ–‘‘‰Ž‘— ǡ —ƒ‰ǡ —‡œǡ —„‡”–
sophisticated state spaces to test generalization ability of the ǡƒ‡”ǡƒ‹ǡ‘Ž–‘ȋʹͲͳ͹Ȍƒ•–‡”‹‰–Š‡‰ƒ‡‘ˆ‰‘™‹–Š‘—–Š—ƒ
algorithms as well. ‘™Ž‡†‰‡Ǥnature ͷͷͲȋ͹͸͹͸Ȍǣ͵ͷͶ–ͻǤ
ȏͻȐ ‡‰ ǡ‘‰ǡ‘ Š‡”ǡ‹Ǧ ǡ‹ǡ ‡‹Ǧ ‡‹ȋʹͲͲͻȌ ƒ‰‡‡–ǣŽƒ”‰‡Ǧ• ƒŽ‡
Lastly, we used only one single parametrization. Different Š‹‡”ƒ” Š‹ ƒŽ‹ƒ‰‡†ƒ–ƒ„ƒ•‡ǤIEEE Conference on Computer Vision and Pattern
parameter settings in one algorithm may lead to different Recognition, 2009: CVPR 2009 ; 20 - 25 June 2009, Miami [Beach], FL, USAǤ Ǥ
‹• ƒ–ƒ™ƒ›ǡ ǡ’’ǤʹͶͺ–ʹͷͷǤ
results, possibly even diverging drastically in between cases. In ȏͳͲȐ ‡”•Š‰‘”ǤThe Quartz guide to artificial intelligence: What is it, why is it
order to produce a more generalizing statement of AI-vs- important, and should we be afraid? Š––’•ǣȀȀ“œǤ ‘ȀͳͲͶ͸͵ͷͲȀ–Š‡Ǧ“—ƒ”–œǦ
‰—‹†‡Ǧ–‘Ǧƒ”–‹ˆ‹ ‹ƒŽǦ‹–‡ŽŽ‹‰‡ ‡Ǧ™Šƒ–Ǧ‹•Ǧ‹–Ǧ™Š›Ǧ‹•Ǧ‹–Ǧ‹’‘”–ƒ–Ǧƒ†Ǧ•Š‘—Ž†Ǧ™‡Ǧ
human performance, a sensitivity analysis of algorithm „‡Ǧƒˆ”ƒ‹†Ȁȋƒ ‡••‡†‘ʹͺǤͲͳǤʹͲʹͲȌǤ
parameters could be used in the future. ȏͳͳȐ ƒŒ’—”ƒ”ǡŠƒ‰ ǡ‘’›”‡˜ǡ‹ƒ‰ȋʹͲͳ͸Ȍ“—ƒ†ǣͳͲͲǡͲͲͲΪ“—‡•–‹‘•
ˆ‘”ƒ Š‹‡ ‘’”‡Š‡•‹‘‘ˆ–‡š–ǤarXiv preprint arXiv:1606.05250Ǥ
To overcome these limitations, the authors are creating an ȏͳʹȐ ‘†Ž‹ˆˆ‡ ǤAI Beats Humans at Reading Comprehension, but It Still Doesn’t
actor-critic based RL algorithm for reactive job-shop Truly Comprehend Language. Š––’•ǣȀȀ™™™Ǥ–‡ Š‘Ž‘‰›”‡˜‹‡™Ǥ ‘Ȁ–Š‡Ǧ
†‘™Ž‘ƒ†Ȁ͸Ͳͻͻͺ͸Ȁƒ‹Ǧ„‡ƒ–•ǦŠ—ƒ•Ǧƒ–Ǧ”‡ƒ†‹‰Ǧ ‘’”‡Š‡•‹‘Ǧ„—–Ǧ‹–Ǧ•–‹ŽŽǦ
scheduling that is able to handle more complex cases. †‘‡•–Ǧ–”—Ž›Ǧ ‘’”‡Š‡†Ȁȋƒ ‡••‡†‘ʹͺǤͲͳǤʹͲʹͲȌǤ
Simultaneously, we are designing a user-friendly interface to ȏͳ͵Ȑ ‹Šǡƒ˜— —‘‰Ž—ǡ‹Ž˜‡”ǡ—•—ǡ‡‡•• ǡ‡ŽŽ‡ƒ”‡ ǡ ”ƒ˜‡•ǡ
‹‡†‹ŽŽ‡”ǡ ‹†Œ‡Žƒ†ǡ•–”‘˜•‹ ȋʹͲͳͷȌ —ƒǦŽ‡˜‡Ž ‘–”‘Ž–Š”‘—‰Š
deploy online and in lectures to make the problem more †‡‡’”‡‹ˆ‘” ‡‡–Ž‡ƒ”‹‰Ǥnature ͷͳͺȋ͹ͷͶͲȌǣͷʹͻ–͵͵Ǥ
accessible to a wider audience. ȏͳͶȐ ‡ŽŽ‡ƒ”‡ ǡƒ††ƒˆǡ‡‡•• ǡ‘™Ž‹‰ȋʹͲͳ͵ȌŠ‡ƒ” ƒ†‡Ž‡ƒ”‹‰
‡˜‹”‘‡–ǣ‡˜ƒŽ—ƒ–‹‘’Žƒ–ˆ‘”ˆ‘”‰‡‡”ƒŽƒ‰‡–•ǤJournal of Artificial
Intelligence Research Ͷ͹ǣʹͷ͵–͹ͻǤ
6. Conclusion ȏͳͷȐ MartınezǦŽ—‡† ǡ”—†² ‹‘ǡƒ”–À‡œǦ•‘ǡ ‡”ž†‡œǦ”ƒŽŽ‘ ȋʹͲͳ͸Ȍ
ƒ‹‰•‡•‡‘ˆ‹–‡”‡•’‘•‡–Š‡‘”›‹ƒ Š‹‡Ž‡ƒ”‹‰Ǥ‹ ƒŽǤƒ‹ƒǡ
ƒ”‹ƒ ‘šǡƒ‘Ž‘‘—“—‡–ǡȋ†•ǤȌǤECAI'16: Proceedings of the Twenty-second
AI vs. human performance assessment is not entirely new. European Conference on Artificial Intelligenceǡ’’ǤͳͳͶͲ–ͳͳͶͺǤ
Most assessments rely on benchmark results from known, large ȏͳ͸Ȑ ƒŽ‘” ǡ— ǡ— ȋʹͲͳ͸Ȍ—‹Ž†‹‰ƒ˜ƒŽ—ƒ–‹‘ ƒŽ‡—•‹‰ –‡‡•’‘•‡
Š‡‘”›ǤProceedings of the Conference on Empirical Methods in Natural
databases to classify an AI’s performance in relation to human Language Processing. Conference on Empirical Methods in Natural Language
participants. As these benchmarks need labelled training data, Processing ʹͲͳ͸ǣ͸Ͷͺ–ͷ͹Ǥ
ȏͳ͹Ȑ ƒŽ‘” ǡ— ǡ—Š†ƒŽƒ‹ǡ— ȋʹͲͳ͹ȌUnderstanding Deep Learning
the tested domain is usually very narrow. Assessments that Performance through an Examination of Test Set Difficulty: A Psychometric Case
compare AI and human abilities on more general sets of tasks StudyǤ
ȏͳͺȐ ƒ”–À‡œǦŽ—‡† ǡ ‡”ž†‡œǦ”ƒŽŽ‘ ȋʹͲͳͺȌƒŽ›•‹‰‡•—Ž–•ˆ”‘ 
are scarce. Some work exists on using IRT to introduce ‡ Šƒ”•ǣ‡› †‹ ƒ–‘”•ƒ† ‘™–‘„–ƒ‹Š‡ǤarXiv preprint
uniformity and a representation of levels of difficulty in image arXiv:1811.08186Ǥ
ȏͳͻȐ ‹–œ• Š˜ȋʹͲͳͷȌEntscheidungslehre: Wie Menschen entscheiden und wie sie
recognition and games scoring. entscheiden sollten. ͳͲ–Š‡†Ǥ‡”Žƒ‰•Šƒ—•ƒ‹œ „Šǡƒ Š‡Ǥ
The novelty aspect of this paper is to adopt the IRT approach ȏʹͲȐ –”‹ ‡”ǡ—ŠŽ‡ǡ–—”ǡ ”‹‡••ȋʹͲͳͺȌ‡‹ˆ‘” ‡‡–Ž‡ƒ”‹‰ˆ‘”
to compare an RL algorithm against human performance on ƒ†ƒ’–‹˜‡‘”†‡”†‹•’ƒ– Š‹‰‹–Š‡•‡‹ ‘†— –‘”‹†—•–”›ǤCIRP Annals
͸͹ȋͳȌǣͷͳͳ–ͶǤ
two simple job-shop scheduling examples which represent low- ȏʹͳȐ —––‘ǡƒ”–‘ ȋʹͲͳͺȌReinforcement learning: An introductionǤ 
complexity decisions in PM. We demonstrated that an AI’s ’”‡••Ǥ
ȏʹʹȐ ƒ„‡Žǡ‹‡†‹ŽŽ‡”ȋʹͲͲͺȌ†ƒ’–‹˜‡”‡ƒ –‹˜‡Œ‘„Ǧ•Š‘’• Š‡†—Ž‹‰™‹–Š
performance on these tasks can be approximated based on ”‡‹ˆ‘” ‡‡–Ž‡ƒ”‹‰ƒ‰‡–•ǤInternational Journal of Information
human responses to the same case. Technology and Intelligent Computing ʹͶȋͶȌǣͳͶ–ͺǤ
ȏʹ͵Ȑ ƒ”–À‡œǡ‘™±ǡ—ž”‡œ ǡ‡ŽŽ‘ȋʹͲͳͳȌ‡‹ˆ‘” ‡‡–‡ƒ”‹‰
Our AI, based on RL, has reached par-human scores after a ’’”‘ƒ Šˆ‘”–Š‡ Ž‡š‹„Ž‡ ‘„Š‘’ Š‡†—Ž‹‰”‘„Ž‡Ǥ‹‘‡ŽŽ‘‘‡ŽŽ‘ǡ
small number of training epochs. From the findings we induce ȋ†ǤȌǤLearning and intelligent optimization: 5th international conference, LION
5, Rome, Italy, January 17 - 21, 2011 ; selected papersǤ’”‹‰‡”Ǥ‡”Ž‹ǡ’’Ǥʹͷ͵–
that these results will be replicable for similar PM decisions. ʹ͸ʹǤ
However, future research will have to operationalize human ȏʹͶȐ Š‘ƒ•ǡ‘‘ ǡŠƒ–‡”Œ‹ǡƒ‰ Š‹ȋʹͲͳͺȌ‹‡”˜ƒǣ”‡‹ˆ‘” ‡‡–
Ž‡ƒ”‹‰Ǧ„ƒ•‡†–‡ Š‹“—‡ˆ‘”‘’–‹ƒŽ• Š‡†—Ž‹‰ƒ†„‘––Ž‡‡ †‡–‡ –‹‘‹
and AI performance assessments for PM tasks of varying †‹•–”‹„—–‡†ˆƒ –‘”›‘’‡”ƒ–‹‘•Ǥ2018 10th International Conference on
complexity. With the presented experiments, we have taken a Communication Systems & Networks (COMSNETS): 2018 10th International
Conference on Communication Systems & Networks (COMSNETS) took place 3-7
first step towards the concept of Cyber Production January 2018 in Bangalore, IndiaǤ Ǥ‹• ƒ–ƒ™ƒ›ǡ ǡ’’Ǥͳʹͻ–ͳ͵͸Ǥ
Management. ȏʹͷȐ „”‡–•‘ǡ‡‹•‡ȋʹͲͳ͵ȌItem response theoryǤ•› Š‘Ž‘‰›”‡••Ǥ
ȏʹ͸Ȑ ‹œ‘’‘—Ž‘•ȋʹͲͲ͸ȌŽ–ǣ’ƒ ƒ‰‡ˆ‘”Žƒ–‡–˜ƒ”‹ƒ„Ž‡‘†‡Ž‹‰ƒ†‹–‡
”‡•’‘•‡–Š‡‘”›ƒƒŽ›•‡•ǤJournal of statistical software ͳ͹ȋͷȌǣͳ–ʹͷǤ
References ȏʹ͹Ȑ ”ÚŽŽ‘ Š•ǡ ‡—‡””‹‡‰‡ŽȋʹͲͳͺȌ‡‹ˆ‘” ‡‡–‡ƒ”‹‰‹ǤarXiv preprint
arXiv:1810.00240Ǥ
ȏͳȐ ‡—•‡ ǡ‡‹•‡”ǡ ‡‰•–‡„‡ ǡ—• Š ȋʹͲͳͷȌ ‡•–ƒŽ–—‰˜‘
”‘†—–‹‘••›•–‡‡‹‘–‡š–˜‘ †—•–”‹‡ͶǤͲǤZukunft der Arbeit in
Industrie 4.0Ǥ’”‹‰‡”‹‡™‡‰ǡ‡”Ž‹ǡ ‡‹†‡Ž„‡”‰ǡ’’Ǥͻͻ–ͳͲͻǤ

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 Available online at www.sciencedirect.com

ScienceDirect
Procedia CIRP 93 (2020) 944–948
www.elsevier.com/locate/procedia

53rd CIRP Conference on Manufacturing Systems

The insufficiency of Lean Tradition – Lean meets Kodak Moments

Halvor Holtskoga* and Geir Ringena
Norwegian University of Science and Technology, PO box 22, 2802 Gjovik, Norway
* Corresponding author. Tel.: +47 97 78 43 20; E-mail address: [email protected]

Abstract

Little attention in the Lean literature has focused on disruptive events and change. However, there is emerging an awareness of the insufficiencies
built into this line of thinking and the way it is conducted in practice. The purpose of this paper is to highlight how underdeveloped Lean thinking
is to meet disruptive events or Kodak moments. This case study illustrates how strong commitment to Lean thinking not necessarily helps in
critical situations. Following a path turning lean from the initiative top-down to a bottom-up approach, including softer components as knowledge,
reflection, culture and involvement. While compensated the situation to some extent, but in general, the lean tool-box was insufficient to align
the organization to change. Missing in the change initiative was time for reflection and discussion about what to reflect upon, central part of
Hensai. The paper gives insights into how a company with traditional Lean implementation struggles with aligning change with market
development. The solar panel industry serve as case study. 17 in-depth interviews have been conducted with people directly involved in the cases.
© 2020 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the scientific committee of the 53rd CIRP Conference on Manufacturing Systems
Keywords: Lean; Disruptive change; continuous improvement; Hansei; Kaizen

1. Introduction discussions have been regarding Lean, where Lean is used as a

way for massive lay-offs. The most known is the struggle of the
Thinking and preparing for the future is hard and difficult. British Auto Union against British Leyland management [2]. In
In order to minimize risk, business leaders trust various proven addition, the famous book of Harrison made this as one of the
theoretical frameworks, like Lean, Total Quality Management, major arguments [3].
Business Process Reengineering (BPR), etc. Lots of articles Both of these situations describe the misuse or
and books tell stories about successful implementations of misconception of the theoretical framework. It also illustrate
these concepts giving impression of them providing a that theoretical frameworks, like Lean, are not suited to deal
seemingly secure way to success. However, over time some with every type of situation. This paper will focus on disruptive
publications emerge with aspects that tell that there is no such events where Lean do not provide an answer.
thing as one secure way to success by following the Understanding the limitation of Lean is important, specially,
frameworks. since it has become somewhat of a de-facto standard for
BPR was originally just a way big companies tried to learning, improvement and efficiency in the manufacturing
implement new technology and link it to processes for better industry. Reinforcing the importance is accelerating changes in
efficiency. After some implementations with lay-offs as a markets and global competition. These forces limit time to
result, BPR got a bad reputation. Associating BPR with major reflect upon practice, which is essential to the Lean thinking.
lay-offs was the common perception of the framework. One of
the founders, Davenport [1] wrote a defense article, trying to
explain how the framework was misused and misunderstood by
companies in need for downsizing their staff. Similar

2212-8271 © 2020 The Authors. Published by Elsevier B.V.

This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the scientific committee of the 53rd CIRP Conference on Manufacturing Systems
10.1016/j.procir.2020.04.082

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 Halvor Holtskog et al. / Procedia CIRP 93 (2020) 944–948 945

2. Theory In addition, Hansei reflects the problem solving starts with

people closes to the problem. Pushing the problem solving
Little attention in the Lean literature has focused on down in the hierarchy resulting in informal and flatter
disruptive events and step-changes. Lean thinking is celebrated organizations.
as the standard for effective and efficient production, and
spread to many other fields of business. It seems to cover what 3. Methodology
is good practice and the way to profitability [4]. Recently,
authors have claimed that Lean thinking can be used with This study reports on a research project, Lean Operations,
success in almost every industry [5]. conducted in Norway from 2012-15. The sub-projects in this
However, there is emerging awareness of the insufficiencies program were organized as case-studies, and this article
built into this line of thinking, and the way it is conducted in documents upon a study of one of the four case companies. This
practice. The critique falls into four categories; human cost [6], case study relies upon 17 in-depth interview. A group of 5
autonomy syndrome in product development [7, 8], tool researchers participated. Most of the interviews were recorded
dependency [9, 10], and union busting [11, 12]. This paper will and transcribed. The informants read and approved the
use disruptive events and step-changes, called Kodak moments, transcriptions. Further work with the documents was
to illustrate some other built in insufficiencies of Lean thinking. discussions among the researchers inspired by grounded theory
The phenomena, Kodak moments, was historically used as a approach [17]. After the discussions, two researchers went
metaphor to mark great spots to take a picture. Resulting in back to the company and interviewed more people. The latter
increased sales of films, where Kodak was the market leader. resulted in a Norwegian report, translated “Lean in the sunset”
The term resurfaced, as a phenomenon, when Kodak went [18]. This paper builds on these initial interviews and the
bankrupt due to strong belief in the superior quality of the Norwegian report. The authors were among the researchers
camera film versus digital camera. It illustrates a disruptive conduction the initial interviews. These documents have been
event and is reinforced by the fact that Kodak invented the revisited and recoded for this paper. Resulting in new
digital camera [13]. discoveries presented in this paper.
Disruption is defined as a process whereby a company is able
to successfully challenge established and/or larger businesses. 4. Case
Specifically, as incumbents focus on improving their products
and services for their most demanding customers, they exceed The case company, here called TEC, established its first
the needs of some segments and ignore the needs of others. silicon wafer factory in 2003 to serve the growing solar energy
Entrants that prove disruptive begin by successfully targeting market. Relatively good access to competence, at all levels,
those overlooked segments, gaining a foothold by delivering inexpensive energy, Lean and tempered cooling water, and
more-suitable functionality, frequently at a lower price. investment subsidiaries were all contributing factors for the
One of the essential parts of Lean thinking is continuous decision to establish the first factory in Norway. Main
improvement [4]. The idea is to engage the workforce into processes for producing silicon wafers are ingot casting
discussing and initiating improvements in operations. (heating the material to liquid state, then holding it at elevated
Motivated by lower cost, better quality and more efficient and temperature for about ten hours), cleaning, cutting ingots into
effective operations [14, 15]. Liker describe this by two wafer blocks with cross-sectional dimensions, grinding into
Japanese words, Hansei and Kaizen [4]. The latter referees to dimensional accuracy, sawing blocks to thin wafers (0.12 mm),
problem-solving process, exemplified by Liker in the “5-why” and finally cleaning and ultrasound quality control. The market
analysis. The former, Hansei, is about responsibility, self- for wafers boomed from 2002-2008, and TEC expanded their
reflection and organizational learning. According to Japanese capacity by factory number two in operation in 2007 and
culture and tradition, Hansei can be translated to reflection. The factories number three and four respectively opened in 2008
point is that Hansei and Kaizen, or reflection and continuous and 2009. The two latter factories were installed with updated
improvement, goes hand-in-hand. When a company introduces technology for sawing wafers and more automated material
Kaizen, they must make sure that the organization are able to handling between processes. In parallel to the extensively
reflect and think. When you do something wrong, it is important growth focus in that period, TEC built their production
to reflect upon the wrongdoing and figure out how to do it better philosophy on Lean principles from day one. Lean production
and in a more correct manner. In some sense, Hansei in was in 2002 seen as mandatory when operating in global
combination with Kaizen can be translated to organizational markets, and the largest owners of TEC at that time had been
learning (OL). One of the key component in OL is time to early movers in terms of adopting and implementing Lean
reflect in order to solve the core problem not just symptoms principles. However, the operating mode of Lean was said to
[16]. Organizational learning can also be distinguished into OL be rule based, top-down oriented and 5S focused, giving less
1 and OL 2, where the former relates to how the organization room for participation, cross-disciplinarily and common sense.
performs according to standard operational procedures, The market started to change in 2008-2009, partly due to the
whereas the latter describes how organizations are able to adapt global financial crises but also seeing a substantial and rapid
to its context. capacity expansion globally, particularly in China. TEC had
foreseen changes, but not such momentous one. Increased

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 946 Halvor Holtskog et al. / Procedia CIRP 93 (2020) 944–948

competition was natural in a market that had seen 40-50 % contradiction between continuous improvement vs radical
yearly growth rates. Higher global volumes of silicon wafers innovation.
and lower prices was one thing, but this increased volume also
came with higher quality standards. TEC at that time had 5. Discussions
limited control of their downstream value chain and thus lacked
feedback on product quality. The mantra had been to deliver as The case illustrates a typical situation where companies
much quantity, as possible, assuming that every unit produced experience good profitability over longer periods, where the
was 100% in accordance to customer requirements. As a market pays well for any product offered by the seller. As
response TEC restructured their factories, management was noted, Lean as production philosophy was emphasized by top
replaced, and quality and performance departments were management from the early beginning. Nevertheless, not as an
revitalized. integrated and holistic system well understood and embraced
A comprehensive business system, based on Lean principles by the employees, rather a control regime failing to achieve the
and proven in other divisions of the owner's conglomerate of intentional smooth, zero defect and cost-effective one-piece-
companies, was implemented in 2010. This business system flow production. Thus, the case company went on with their
equalized the importance of both culture and structure, mass production, earning satisfactory earnings for the
meaning that all employees need understanding and knowledge shareholders, and an obvious significant improvement
on when and how to utilize methods and tools in the Lean potential.
toolbox. Thereby, making it possible to do problem solving at The top down approach is in contrast to the mainstream
the lowest level in operations. Within a year after introducing Lean literature [4, 19, 20], where support from the top is
the business system near all employees had been through a essential but it must be driven as a more bottom up approach.
learning program for continuous improvement. A team Similar argument was offered by March and Simon [16] when
structure was implemented to distribute roles and they tried to explain the difference between OL1 and OL2. If
responsibilities and to address problems at root-cause. you just attack the symptoms to the problem and use the
Identified problems and challenges were rated according to an problem solving as a control measure, it will lead to OL1 and,
improvement hierarchy, rating from the simplest state of "just sometimes, defensive actions. In other words, the
do it" to the formation of temporary and cross-discipline organizational learning suffers. At TEC, we did not find any
improvement teams aiming to solve complex problems by evidence of defensive action, but the problem solving was not
gathering facts, analyzing it by advanced statistical tools and employee driven either. On the positive side, TEC had a strong
discuss results and actions across team and value chain cultural feeling of “being in the same boat”. Employees wanted
boundaries. Successful results of improvement projects in one to learn in order to do a better job and to secure their workplace.
factory was rapidly distributed to the other factories, a This driving force was one of the main contributors for not
transformation made available by the new management and having defensive actions, which could damage their company
factory structure. The efforts gave results in the period from survival.
2010-2011 in terms of improved quality (wafer yield increased The learning and implementation of Lean tools for
from 83-95% - making TEC better than global average), cost continuous improvement and problem solving seemed to get
(cost per unit decreased by 50%), and HES (absence due to job total focus in the entire company. Battling the lowering prices,
related issues reduced by 88%). However, global prices on the attention of the leaders seemed to become more on-sided
silicon wafers continued downward at a higher phase than TEC and taking important things for granted. It meant that
was able to compensate for with the Lean business system. organizational learning was the issue to follow up, without time
This case illustrates a typical situation where companies to reflect and delegation of authority.
experience good times over longer periods, where the market Operators continuous to improve operations and saved costs
pays well for any product offered by the seller. As noted, Lean without the knowledge of the development of the market
as production philosophy was emphasized by top management prices. Leaders forgot to ask questions about is Lean, in the way
from the early beginning. However, not as an integrated and they have implemented it, the right way forward. Will
holistic system well understood and embraced by the continuous improvement and organizational learning be able to
employees, rather a control regime failing to achieve the deliver enough cost saving and efficiency to compensate the
intentional smooth, zero defect and cost-effective one-piece- down tumbling market prices? For the researchers, it seems to
flow production. Thus, TEC went on with their mass be not the case. The leaders were determined to stick with Lean
production, earning satisfactory earnings for the shareholders, and shave cost. Investments in newer technology and more
and an obvious significant improvement potential. A potential automation of operations were, of course considered, but
unleashed too late to make TEC prepared for the market having relatively new technology in place and little time and
uncertainties to come. When the crisis became present, Lean as capital to invest further this was not realistic options in the short
a holistic business philosophy seemed easy to "sell-in". The time perspective.
unsolved question is whether Lean could have been the savior, The leaders promoted a showcase for Lean thinking and,
or at least postponed the close-down by years, without a certain perhaps, felt obliged to follow this route. This impression
degree of urgency. Which leads to the dilemma or constructed manifested itself when the CEO, after the closing of the factory,
talked about how they successfully implemented Lean. In a

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 Halvor Holtskog et al. / Procedia CIRP 93 (2020) 944–948 947

presentation for other companies, he proudly presented how he 3. Harrison, B., Lean and mean: the changing
used to local media to motivate operators through positive landscape of corporate power in the age of
publicity. flexibility. 1994, New York: Basic Books. XI, 324 s.
Going back to the interviews the impression of not fully 4. Liker, J., The Toyota Way; 14 Management
understand how to manifest responsibility in the lower ranks Principles From The World's Greatest Manufacturer.
was obvious in the first stage of the lean journey. People were 2004, New York: McGraw-Hill.
waiting for orders or doing continuous improvement in the 5. Netland, T.H. and D.J. Powell, The Routledge
areas where the leadership measured. Little to no time was set Companion to Lean Management. 2017, New York:
out to self-reflection upon the direction where the company and Routledge.
each department were heading. Overall, the bigger picture was 6. Klein, J.A., The Human Cost of Manufacturing
Reform. Harvard Business Review, 1989. March-
missing.
April: p. 60 - 66.
The contribution of this paper is to reflect upon, and discuss,
7. Fujimoto, T., Capability Building and Over-
the possible shortcomings of lean in a crisis situation. The way
Adaptation: A Case of 'Fat Design' in the Japanese
Toyota describes the value of lean is that continuous Auto Industry, in Coping with Variety; Flexible
improvement of products, processes and people over time productive systems for product variety in the auto
smoothen out negative effects of unexpected events. This long industry, Y. Lung, J.-J. Chanaron, and D. Raff,
term business philosophy integrates all functions of a company Editors. 1999, Ashgate Publising Limited:
from operations, product development, technology integration Hampshire. p. 261 - 286.
and other support functions. Thus, partly and recently 8. Fujimoto, T., The Evolution of a Manufacturing
implemented lean principles will less probably withstand major System at Toyota. 1999, New York: Oxford
crisis as discussed in our case. A more holistic lean perspective University Press.
in addition to a larger degree of people involved in 9. Berggren, C., The Volvo experience: alternatives to
understanding the context, implementing and follow-up on lean lean production in the Swedish auto industry. 1993,
principles, and reflective actions upon prioritizing activities Houndmills: Macmillan. XIII, 286 s.
leading towards targets. It is hard to generalize from a one- 10. Berggren, C., NUMMI vs. Uddevalla: a Rejoinder.
case-study, but reflective actions as seem from the literature Sloan Management Review, 1994. 35(2): p. 37 - 49.
may be underestimated as a contributing factor towards 11. Green, W.C. and E.J. Yanarella, eds. North American
achieving OL 2. Auto Unions In Crisis; Lean Production as
Contested Terrain. 1996, State University of New
6. Conclusion York Press: Albany.
12. Rinehart, J., C. Huxley, and D. Robertson, Just
The two elements, time for reflection and what to reflect another car factory?: lean production and its
upon, seem to be two essential key parts in Lean or OL. Setting discontents. 1997, Ithaca: Cornell University Press.
XI, 249 s.
the boundaries for reflection and improvement to narrow can
13. Johnsen, H.C.G., H. Holtskog, and J.R. Ennals, eds.
have severe consequences in times of disruptive change.
Coping with the Future - rethinking assumptions for
Having a top-down approach enforced the misleading efforts in
society, business and work. 2018, Routledge:
Lean. Top management failed to ask the right questions and London.
truly engage the operators in Hansei. This had fatal 14. Ringen, G. and H. Holtskog, How enablers for lean
consequences when the market price dropped. The cost saving product development motivate engineers.
and increased efficiency were not enough and hardly the right International Journal of Computer Integrated
thing to do. Manufacturing, 2011: p. 1-11.
15. Ringen, G., E. Lodgaard, and C. Langeland, How to
Acknowledgements Succeed With Continuous Improvement in a Product
Development Environment, in NordPLM'09, 2nd
We wish to thank the Norwegian Research Council for Nordic Conference on Product Lifecycle
financial support of this project. In addition, the other Management. 2008.
researchers that participated in the four year project to find how 16. March, J., et al., Organizations. 1994. p. 786-787.
Norwegian companies implement Lean in their daily 17. Strauss, A. and J. Corbin, Basics of Qualitative
operations. Research: Techniques and Procedures for
Developing Grounded Theory. 2 ed. 1998, Thousand
References Oaks: Sage Publications.
18. Bygdås, A. and E. Falkum, Lean i solnedgang. AFI
1. Davenport, T.H., The Fad That Forgot People, in rapport 2012. 18.
Fastcompany. 1995. 19. Morgan, J. and J.K. Liker, The Toyota Product
2. Steward, P., et al., We sell our time no more; Development System, Integrating People, Process
Worker's struggles against Lean Production in the and Technology. 2006, New York: Productivity
British Car Industry. 2009, London: Pluto Press. Press.
20. Womack, J.P., D.T. Jones, and D. Roos, The

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 948 Halvor Holtskog et al. / Procedia CIRP 93 (2020) 944–948

machine that changed the world : based on the 1990, New York: Rawson Associates. viii, 323 p.
Massachusetts Institute of Technology 5-million
dollar 5-year study on the future of the automobile.

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 Available online at www.sciencedirect.com

ScienceDirect
Procedia CIRP 93 (2020) 1085–1090
www.elsevier.com/locate/procedia

53rd CIRP Conference on Manufacturing Systems

A Human-Robot Collaboration System towards High Accuracy

Xi Vincent Wang*, Xuetao Zhang, Ying Yang, Lihui Wang
Department of Production Engineering, KTH Royal Institute of Technology, Sweden

* * Corresponding author. Tel.: +46-736-441-433; fax: +46-8-210-851. E-mail address: [email protected]

Abstract

The next generation of manufacturing system calls for feasible solutions with high efficiency and flexibility. Thus in recent years, Human-
Robot Collaboration (HRC) research attracts many attentions worldwide since it unites the repeatability and accuracy of robots and the
adaptivity and intelligence of human operators. In this paper, the system architecture of an HRC solution is presented, together with the novel
method of vision sensing and controlling towards a smart solution with high accuracy and fast speed. The proposed method is demonstrated
through implementations in the real robot cell, and evaluated by quantifiable measurements. Future research outlooks are addressed at the end
of the paper.
© 2020 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the scientific committee of the 53rd CIRP Conference on Manufacturing Systems
Keywords: Human-Robot Collaboration; System Design; Smart Manufacturing; Cyber-Physical System; CPS

1. Introduction with the industrial robots at the same time in the same space.
The vision-based approach is popular to identify the operation
Nowadays, different initiatives are proposed worldwide objects as the first step. A bin-picking system was proposed
towards the next generation of the manufacturing industry. based on the point cloud generated by the RealSense camera
Despite the diverse terminologies and focuses, it is a consensus [2]. The 3D objects are segmented from to guide the industrial
that human should not be replaced by automation and computer robot towards accurate object-picking tasks. Similarly, the 3D
systems. Instead, the new technologies shall work as assistance sensor is utilised to guide the robot grinding operation [3]. The
and supports for the human to provide a more efficient, safe depth sensor is utilised for localisation and the 2D camera for
and friendly working environment. Among the latest profile scanning. Eventually, the grinding process can be
manufacturing technologies, Human-Robot Collaboration guided by the vision-based solution automatically.
(HRC) offers the opportunity of integrating the accuracy of To monitor the robotic environment, a framework was
robot and flexibility of human [1]. Thus in this paper, the recent presented which realised separation distance monitoring
HRC research is reviewed and discussed. Then a novel HRC between a robot and a human operator in a transparent and
system is proposed towards higher accuracy and faster tuneable fashion [4]. The separation distance is assessed
response speed. The proposed work is then validated through pairwise for key points on the robot and the human body and as
implementations and evaluated via quantifiable measurements. such can be selectively modified to account for specific
conditions. Aalerud et al. [5] proposed system architecture for
2. Literature Review mapping and real-time monitoring of a relatively large
industrial robotic environment. Multiple static sensor nodes are
In recent years, the HRC research is taken globally as it placed on the ceiling to generate the point clouds reflecting the
offers a promising approach allowing the operators to work dynamics in the robot cell. In an even bigger scale, an

2212-8271 © 2020 The Authors. Published by Elsevier B.V.

This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the scientific committee of the 53rd CIRP Conference on Manufacturing Systems
10.1016/j.procir.2020.04.085

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 1086 Xi Vincent Wang et al. / Procedia CIRP 93 (2020) 1085–1090

autonomous scanning approach was presented which allowed framework for real-time safe human-robot collaboration, using
multiple robots to perform collaborative scanning for dense 3D static hand gestures and 3D skeleton extraction. OpenPose
reconstruction of unknown indoor scenes [6]. The method can library is integrated with Microsoft Kinect V2, to obtain a 3D
conduct paths for several robots, allowing them to efficiently estimation of the human skeleton.
coordinate with each other such that the collective scanning At the collaboration phase, A dynamic robot task selection
coverage and reconstruction quality is maximised while the framework was proposed in human-robot collaborative
overall scanning effort is minimised. contexts through a voxel-based collision avoidance system
After the robotic environment is recognised, the robot needs [15]. In this research, the shared workspace is monitored by 3D
to be controlled according to the plan. Jeppesen et al. [7] point-cloud sensors. Wang et al. [16] introduced real-time
developed a lightweight HRC structure based on the RealSense active collision avoidance in an augmented environment. In this
Robotic Development Kit. The field-programmable gate array work, virtual 3D models of robots and real camera images of
is utilised for robot control and real-time tasks, e.g. following operators are integrated for monitoring and collision detection
human’s hand. A similar approach is also introduced to drive [17]. The industrial robot controllers are further linked for
the standard industrial robot, e.g. ABB IRB140 [8]. The pre- adaptive robot control, without the need for programming by
defined robot model is then matched to each robot link, based the operators [18]. An HRC system for teaching assembly was
on the depth image captured by the depth sensor. In this way, developed by Haage et al. [19]. The human demonstrates the
the robot posture can be monitored and estimated without any assembly process, and the RealSense RGBD sensor captures
markers or sensors on the robot. Instead of controlling the robot the key frames, which are converted into semantic graphs and
locally, a 3D augmented reality navigation system using assembly program later to drive the industrial robot.
stereoscopic images was developed based on a remote robot During the practice of the HRC system, safety is one of the
operating system [9]. The accurate matching between the most critical aspects. Maurtua et al. [20] measured the trust of
simulated model and the video image of the actual robot can be workers on fenceless human-robot collaboration in industrial
realised, which helps the operator to accomplish the remote robotic applications as well as to gauge the acceptance of
control task correctly and reliably. different interaction mechanisms between robots and human
During the robot manipulation, gesture recognition is an beings. In their finding, most of the participants declare that in
efficient method to communicate with the robot. Vysoky et al. the future the collaboration between robots and workers will be
[10] proposed a shared operator-robot workspace. The position possible and that they will accept the collaborative tasks.
of hands and gestures are detected, which leads to instant Besides, gesture-based interactions and hand-guiding
reactions of the robot on the presence of the operator to interaction mechanisms are also rated as promising approaches
eventually control the robot. An intuitive robot teaching method in the future.
was developed based on the hand-guided demonstration [11]. A
depth camera is used to recognise various hand gestures which 3. Proposed System
are used as the commands to control the robot gripper. Liu and
Wang [12, 13] proposed an overall model of gesture recognition Despite significant development in the vision-based HRC
for human-robot collaboration. Four essential technical system, there is still a lack of an efficient approach to perform
components in the model of gesture recognition are identified, accurate and fast interaction with the robot. Thus in this
i.e. sensor technologies, gesture identification, gesture tracking research, a novel system architecture is proposed, as illustrated
and gesture classification. Mazhar et al. [14] designed a in Fig.1. A vision sensor is utilised, which is equipped with an

Depth Sensor Image Processing Search Module Collision ROS Industrial Robot
Module Avoidance Module
‘„‘–
‡’‘”ƒŽ ‘ƒ†ȀŒ‘‹– ‘–”‘ŽŽ‡”
 ƒ† ‹†‡‘ ˆ‹Ž–‡”‹‰ ‡’–Š ‘”‡‰”‘—†
†‡’–Š ƒ‡”ƒ ‡š–”ƒ –‹‘ ‘ŽŽ‹•‹‘•–ƒ–‡ •–ƒ–‡•
ˆ”ƒ‡• ˆ”ƒ‡•
ƒ ‰”‘—†
”ƒ‡ ‡––‹‰ ‹”–—ƒŽ”‘„‘– ƒ‹’—Žƒ–‘”
‡š–”ƒ –‹‘ ‡ƒ”‡•– ‘ŽŽ‹•‹‘
‡‹‰Š„‘”•‡ƒ” Š †‡–‡ –‹‘
Ž‹‰‡– ”ƒŒ‡ –‘”›
‘•‹–‹‘‹ˆ‘”ƒ–‹‘ ’Žƒ‹‰
ƒ’–—”‡ Visualisation
Module ‘–‹—‘—•
Object ‘ŽŽ‹•‹‘
‡’–Š‹ƒ‰‡ƒ† ‘Ž‘”
ˆ”ƒ‡˜‹•—ƒŽ‹œƒ–‹‘ ƒ˜‘‹†ƒ ‡

Ž‹‰‡† ‡’–Š‹ƒ‰‡ƒ† ‘Ž‘”

ˆ”ƒ‡• ˆ”ƒ‡˜‹•—ƒŽ‹œƒ–‹‘

‘•‹–‹‘‹ˆ‘”ƒ–‹‘
˜‹•—ƒŽ‹œƒ–‹‘

Fig. 1. System Architecture for Accurate and Fast HRC.

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 Xi Vincent Wang et al. / Procedia CIRP 93 (2020) 1085–1090 1087

RGB camera and a depth camera. The RGB camera captures the different parts of the data sequentially until the visualisation
the colour video frames like normal video cameras, while the is completed.
depth camera has a built-in scanner which senses the depth of From computing’s perspective, the system consists of an
the shooting scene. As a result, the video frames are sent to the independent initialisation process and three threads that can
Image Processing Module as the output. realise video pre-processing, nearest point identification and
data visualisation at the same time. The system starts when the
3.1. Image Processing Module background frames are captured by the depth sensor and
transmitted to the processing module. Then the background
The Image Processing Module is responsible for pre- point cloud is continuously extracted and processed to apply the
processing the video frames from the cameras. More detailed temporal filter until the background is successfully set. The
workflow of the Image Processing Module can be found in Fig. initialisation stage is finished. While the background data is
2. To be more specific, the colour and the depth video of the sent to the Search Module, the Image Processing Module keep
robot cell are first captured by the two cameras and divided into acquiring the new video frames that have the target object in the
colour and depth frames that are transmitted to the Image experiment scene. The video processing thread in the Image
Processing Module. In the Processing Module, the point clouds Processing Module fetches the frames and spatially aligns all
are extracted from the video frames and smoothed by the the streams in the frames to the depth viewport to get the
temporal filter. The background point cloud is processed to set aligned frames, then smooths the depth frames by applying
the background of the experiment environment. Then the temporal filtering. The temporal filter is one kind of domain-
processed background point cloud and other depth frames are transform filters that can reduce the depth noise by calculating
delivered to the Search Module for further analysis. At the same multiple frames and the one-dimensional exponential moving
time, an alignment method is developed to merge and align the average.
colour and the depth frames into one frame. Subsequently, the
aligned frames are sent to the visualisation module to generate 3.2. Search Module
visible images.
After the depth frames reach the Search Module, the The Search Module is in charge of identifying the nearest
foreground point cloud is extracted by the foreground points between the robot and the obstacle object. After the
extraction algorithm, and processed by the nearest neighbour completion of the Image Processing Module, the filtered depth
search algorithms to compute the position information of the frames are delivered into the Search Module. More detailed
nearest point. Then the module also sends the result to the last workflow of the Search Module is shown in Fig. 3. The closest-
software module, visualisation module. After all the data is point thread controls how the nearest point is identified in the
acquired, the visualisation module starts to render and visualise frame. First, the thread fetches the depth frames and extracts
their point cloud. Then it differentiates the extracted point cloud
with the background point cloud to acquire the foreground point
–ƒ”– cloud. It needs to be highlighted that in the proposed work,
different processing algorithms, i.e. K-D tree, traversal, and
Image
ƒ’–—”‡–Š‡ Processing enumeration algorithms, are deployed in the Search Module to
„ƒ ‰”‘—†ˆ”ƒ‡ Module identify the proper method in the manufacturing scenario.

‘ Search Module
ƒ ‰”‘—† –ƒ”–
ƒ„‡•‡–ǫ ƒ ‰”‘—†
’‘‹– Ž‘—†
‡• ‡– Š‡•–Š‡†‡’–Šˆ”ƒ‡
š–”ƒ –ƒ†’”‡Ǧ ƒ†–Š‡„ƒ ‰”‘—†’‘‹–
’”‘ ‡••–Š‡’‘‹– Ž‘—† Ž‘—†

‘–‹—‡ ƒ’–—”‹‰ ‡‡”ƒ–‡–Š‡’‘‹– Ž‘—†

˜‹†‡‘ˆ”ƒ‡• ‘—–‘ˆ–Š‡„ƒ ‰”‘—†
‘”‡‰”‘—†’‘‹–Ž‘—†
‹†‡‘’”‘ ‡••‹‰
–Š”‡ƒ†ˆ‡– Š‡•–Š‡ —‹Ž†–Š‡ –”‡‡‘”Ǧ–”‡‡
ˆ”ƒ‡
Ž‹‰‡†ˆ”ƒ‡
Ž‹‰•–”‡ƒ•–‘†‡’–Š ‘‘–Š–Š‡‹ƒ‰‡„› ††–Š‡”‘„‘–’‘•‹–‹‘–‘ Robot’s
˜‹‡™’‘‹– –‡’‘”ƒŽˆ‹Ž–‡”‹‰ –Š‡•‡ƒ” Š–”‡‡ Œ‘‹–˜ƒŽ—‡

‡†–‘
Search —‡”›–Š‡–”‡‡ƒ†ˆ‹†–Š‡
Visualisation
‡’–Šˆ”ƒ‡ Module ‡ƒ”‡•–’‘‹–
Module

‘•‹–‹‘‹ˆ‘”ƒ–‹‘
† †

Fig. 2. Workflow of the Image Processing Module. Fig. 3. Workflow of the Search Module.

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 1088 Xi Vincent Wang et al. / Procedia CIRP 93 (2020) 1085–1090

In the K-D Tree approach, the foreground point cloud is used new trajectory plan is generated based on the latest collision
to build the search tree. After the tree is fully generated, the state from the previous modules and sent to the virtual robot in
position information of the robot arm is input into the tree. By the Robot Operating System (ROS).
the query algorithm of the tree, the nearest point to the robot In ROS, the actual industrial robot’s states are first acquired.
arm is found. The position information of the nearest point is It is both synchronised with the virtual robot which reflecting
then sent to the main thread in the visualisation module for the physical environment and sent to the Search Module to
video rendering. It needs to be noted that the proposed work is support the closed distance as mentioned before. Thus it forms
not designed to detect specific types of object. Instead, the a closed-loop control mechanism in the proposed system. The
reported mechanism aims to detect any item that is close to the trajectory planning function receives the position of the
industrial robot because in the industrial robot cell, there is high obstruction and robot, and generate a new route when
uncertainty in the moving objects and what approaches the necessary. Eventually, the new route is sent to the virtual and
industrial robot is unpredictable. For instance, an operator industrial robot. The module also monitors the dynamics of the
might hold a tool or box that is potentially colliding with the robot and collision detection results continuously, to be
robot. Hence, the proposed work presented a universal prepared for the next re-planning.
mechanism to detect any type of collision with the robot, to
maximise the safety in the HRC cell. 4. Implementation

3.3. Visualisation Module To validate and evaluate the proposed HRC structure, the
HRC system is implemented in the real robot cell. Intel Real
The main task of the visualisation module is to visualise the Sense D435 Depth Camera is utilised as the sensor to acquire
processed data. Three main visualisation tasks are undertaken depth and colour frames for further processing and recognition.
by the main thread, i.e. the aligned frames, object and robot and The source code package used in this research is Intel
their position information. Several OpenGL libraries are used RealSense SDK 2.0, which includes integrated interfaces for
for rendering. The workflow of the visualisation is more the D435 and OpenGL interfaces for visualisation. The original
straightforward (Fig. 4). The main thread firstly fetches the Image Processing, Search, and Visualisation modules are
aligned frames from the Processing Module and the position developed based on the architecture presented above. ROS is
information from the robot controller. Then the colourised utilised as the main robot control system. Moveit! and Octomap
depth image is rendered and then the colour frame. Eventually, are the fundamental software packages utilised in ROS.
the robot arm calculated closed point, and the distance is Moveit! provides the low-level planning function in fast speed,
rendered and displayed in the visualisation module as a part of and Octomap is the visualisation environment which also
the graphical user interface. More detailed results can be found responds in short intervals. During the implementation, the
in the implementation section. Universal Robot 5 (UR5) utilised as the physical robot and
connected to ROS.
3.4. Collision Avoidance and Robot Control The image processing and collision detection results are
shown in Fig. 5. The Fig.5.a. shows the original robotic
As shown in Fig. 1, the closest distance between the human environment frame, and Fig. 5.b. shows the results after the
and robot is sent to the Collision Avoidance Module as a frame is filtered by the Image Processing Module. The missing
dynamic vector. The collision detection function decides points in the frame are recovered by the algorithm and
whether it is necessary to re-plan the robot trajectory. If so, the highlighted in different colours while the frame quality is
higher. Fig. 5.c. presents the collision detection results without
–ƒ”– Visualisation the Image Processing Module and Fig. 5.d. shows the version
Module that is filtered. The boundary of the human is clearer in the
result and the system successfully detects the closed point
Š‡ƒ‹–Š”‡ƒ† ˆ‡– Š‡•
–Š‡ƒŽ‹‰‡†ˆ”ƒ‡ƒ†–Š‡
between the human and the robot continuously.
’‘•‹–‹‘‹ˆ‘”ƒ–‹‘ To evaluate the performance of the developed system,
quantifiable measurement is also taken focusing on the
‡†‡”–Š‡ ‘Ž‘”‹œ‡††‡’–Š response speed. The detailed data can be found in table 1. As
‹ƒ‰‡ for the outcome of the object recognition, the average
initialisation time (include background setting) is 5.813s. The
average refresh rate of the result is 10.69Hz when no nearest
‡†‡”–Š‡ ‘Ž‘”ˆ”ƒ‡
neighbour search algorithm (enumeration) is applied. When the
K-D Tree algorithm is applied, the performance (10.74Hz) is
‡†‡”–Š‡”‘„‘–ƒ”ǡ–Š‡ similar to the enumeration (11.06Hz) when the number of the
Ž‘•‡•–’‘‹–ƒ†–Š‡ target points is below 100. As the number of the target points
†‹•–ƒ ‡„‡–™‡‡–Š‡ increases, the refresh rate decreases to 2.46Hz when around
1000 points are included and becomes less than 0.1Hz when the
† magnitude reached to ten thousand. The result shows that K-D

Fig. 4. Workflow of the Visualisation Module.

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 Xi Vincent Wang et al. / Procedia CIRP 93 (2020) 1085–1090 1089

(a) (b)

(c) (d)

Fig. 5. (a) environment image without Image Processing Module, (b) environment image with Image Processing Module, (c) collision detection without
Image Processing Module, and (d) collision detection without Image Processing Module.
Tree algorithm has better performance when the ratio of the drops to 3.84HZ with a similar level of detection accuracy
data scale and the dimension of the data is low, but when the (12cm).
scale is much larger than the dimension, the calculation speed
is highly affected. Table 2. Performance evaluation on accuracy.
Search algorithm Refresh rate (Hz) Average distance error (m)
Table 1. System performance evaluation. Enumeration 10.88 0.11
System Response Time K-D Tree 0.08 -
h=1 11.49 0.54
System Initiation 5.813s
h=2 11.15 0.32
Based on the traversal algorithm 10.69Hz h=3 11.26 0.19
Based on the K-D Tree algorithm 10.74Hz Octree h=4 11.2 0.12
Based on the enumeration algorithm 11.06Hz h=5 11.09 0.14
h=6 9.61 0.09
Another crucial point is the performance of temporal
h=7 3.84 0.12
filtering. The average refresh rate with the temporal filtering
using enumeration is 10.72Hz when the target points are around
5. Discussions and conclusions
10000. The instability value is 15, while for the result without
the filter, the refresh rate is 12.93Hz and the instability value is
The modern manufacturing system calls for a new
28. The instability value is the times when the recognition result
generation of solutions that are smart, flexible, and efficient,
gets significantly affected by the noise in 60 seconds.
while the human still needs to be the centre of the system. The
Regarding the system accuracy, different algorithms are also
working environment needs to be friendly and safe for human
compared regarding the refresh speed and distance error, which
operators. The HRC offers the opportunity to integrate the
is the difference between the actual distance between the human
flexibility of human and accuracy of industrial robots. In this
and robot against the one detected in the system (Table 2). The
paper, a novel HRC system architecture is proposed. The
Enumeration method provides an acceptable response speed
scientific contribution includes the closed-loop system structure
and distance error, i.e. 11cm. The K-D tree method is not
towards accurate HRC processes. Moreover, multiple collision
capable of handling the big amount of points with a reasonable
detection algorithms are deployed to identify the suitable
speed. The performance of the Octree method heavily relies on
solution for specific HRC tasks. The proposed system is
the layers constructed in the method. The more layers are
validated through implementations in the real robot cell, and the
utilised, the higher detection accuracy is achieved, while the
performance is evaluated based on quantifiable assessments. In
response speed decreases at the same time due to the longer
the future, the proposed work can be analysed via industrial
time required for computing. The 6-layer tree structure provides
tasks with more complex objects and trajectories. Additionally,
the best accuracy during the experiment, i.e. 9cm. When the
the safety and ergonomic metrics can be also evaluated for
layers are added to 7, the response speed significantly reduced,
better user-friendliness of the proposed work.

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 1090 Xi Vincent Wang et al. / Procedia CIRP 93 (2020) 1085–1090

References [10] Vysocký A, Pastor R, Novák P. Interaction with Collaborative Robot

Using 2D and TOF Camera. Int. Conf. Model. Simul. Auton. Systesm.
[1] Wang L, Gao R, Váncza J, Krüger J, Wang X V, Makris S, Chryssolouris Springer, pp 477–489
[11] Zhao L, Li X, Liang P, Yang C, Li R. Intuitive robot teaching by hand
G. Symbiotic human-robot collaborative assembly. CIRP Ann; 2019.68(2):
p.701–726. guided demonstration. 2016 IEEE Int. Conf. Mechatronics Autom. IEEE,
[2] Gong X, Chen M, Yang X. Point cloud segmentation of 3D scattered parts pp 1578–1583
[12] Liu H, Wang L. Gesture recognition for human-robot collaboration: A
sampled by RealSense. 2017 IEEE Int. Conf. Inf. Autom. IEEE, pp 1–6
[3] Saukkoriipi J. Design and implementation of robot skill programming and review. Int J Ind Ergon; 2018.68: p.355–367.
control. Oulun yliopisto Saatavissa http//jultika oulu fi/files/nbnfioulu- [13] Liu H, Wang L. Human motion prediction for human-robot collaboration.
J Manuf Syst; 2017. p.1–8.
201902071166 pdf Hakupäivä; 2019.10: p.2019.
[4] Svarny P, Straka Z, Hoffmann M. Toward safe separation distance [14] Mazhar O, Navarro B, Ramdani S, Passama R, Cherubini A. A real-time
monitoring from RGB-D sensors in human-robot interaction. arXiv Prepr. human-robot interaction framework with robust background invariant hand
gesture detection. Robot Comput Integr Manuf; 2019.60: p.34–48.
arXiv1810.04953
[5] Aalerud A, Dybedal J, Ujkani E, Hovland G. Industrial Environment [15] Juelg C, Hermann A, Roennau A, Dillmann R. Efficient, Collaborative
Mapping Using Distributed Static 3D Sensor Nodes. 2018 14th Screw Assembly in a Shared Workspace. Int. Conf. Intell. Auton. Syst.
Springer, pp 837–848
IEEE/ASME Int. Conf. Mechatron. Embed. Syst. Appl. IEEE, pp 1–6
[6] Dong S, Xu K, Zhou Q, Tagliasacchi A, Xin S, Nießner M, Chen B. Multi- [16] Wang L, Schmidt B, Nee AYC. Vision-guided active collision avoidance
robot collaborative dense scene reconstruction. ACM Trans Graph; for human-robot collaborations. Manuf Lett; 2013.1(1): p.5–8.
[17] Schmidt B, Wang L. Depth camera based collision avoidance via active
2019.38(4): p.1–16.
[7] Jeppesen BP, Roy N, Moro L, Baronti F. An FPGA-based controller for robot control. J Manuf Syst; 2014.33(4): p.711–7118.
collaborative robotics. 2017 IEEE 26th Int. Symp. Ind. Electron. IEEE, pp [18] Mohammed A, Schmidt B, Wang L. Active Collision Avoidance for
Human–Robot Collaboration Driven by Vision Sensors. Int J Comput
1067–1072
[8] Varhegyi T, Melik-Merkumians M, Steinegger M, Halmetschlager-Funek Integr Manuf; 2017.30(9): p.970–980.
G, Schitter G. A Visual Servoing Approach for a Six Degrees-of-Freedom [19] Haage M, Piperagkas G, Papadopoulos C, Mariolis I, Malec J, Bekiroglu
Y, Hedelind M, Tzovaras D. Teaching assembly by demonstration using
Industrial Robot by RGB-D Sensing. Autom. Robot.
[9] Xin G, Huan H, Jia Q-X, Sun H-X, Song J-Z. 3D augmented reality advanced human robot interaction and a knowledge integration framework.
teleoperated robot system based on dual vision. J China Univ Posts Procedia Manuf; 2017.11: p.164–173.
[20] Maurtua I, Ibarguren A, Kildal J, Susperregi L, Sierra B. Human–robot
Telecommun; 2011.18(1): p.105–112.
collaboration in industrial applications: Safety, interaction and trust. Int J
Adv Robot Syst; 2017.14(4): p.1729881417716010.

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 INTERNATIONAL JOURNAL OF SCIENTIFIC & TECHNOLOGY RESEARCH VOLUME 8, ISSUE 12, DECEMBER 2019 ISSN 2277-8616

Design And Fabrication Of Automatic Onion

Peeling And Cutting Machine
P. Ravichandran, C.Anbu, S.Sathish kumar, A.Sakthivel, S.Thenralarasu

Abstract: Automation plays an important role in industrial and domestic applications. The purpose of automation is to get higher production rates,
efficient utilization of materials, better quality, reduce time consumption and to improve safety. In present days, automation is not only implemented in
industrial problems but also in domestic applications. This paper describes the application of automation in one of the work in kitchen called onion
peeling and cutting. Onion peeling and cutting is one of the fatigues and time consuming job for preparation of food and it is everyday activity in all food
processing system. In hotels, hostel mess and restaurants high man power is required for the peeling and cutting the onion and also it consumes more
time. On the other hand there is a shortage of workers for these types of works. Hence, there is a good opportunity for the implementation of automation
in onion peeling and cutting machine. The objective of this project is to implement the mechatronics approach for automating the onion peeling and
cutting operation.

Index terms : Automation, Onion peeling, Circular drum, AC motor, Punched metal sheet, Cutter, Operating time
——————————  ——————————

1. INTRODUCTION discussed about the French fries making machine. In their

Emerging automatic and semiautomatic machines like prototype model, potato was cut by horizontal square cut
extruding, dispensing, cutting, slicers, mixers, depositors are blades for French fries. Reuben and Fernandez (2015),
introduced in food processing industries. The automatic discussed about green mango peeling machine. In this study,
equipment has witnessed higher demand in recent years as a mango peeling machine was designed, fabricated and tested
compared to semiautomatic machinery. Industry 4.0 has which used a different peeling approach – peeling along the
revolution in the manufacturing sector favors automation in the direction of the longitudinal axis of the fruit against the more
food processing industry. Food processing refers to an activity common peeling around the longitudinal axis. And it was
which converts food materials to final consumable products. implemented successfully and the mango skin is peeled up-to
Onion is one of the majorly used vegetables for cooking, it is 77%. Mohan (2015) discussed the garlic peeling machine. In
used after the skin is peeled and cut down into pieces. In this model the used cutter blades which are arranged
home for cooking, onion skin is easily peeled and sliced by horizontally with 90 degree angular difference with the help of
knife, but for large quantity of onion it takes much time in vertical shaft. The blower is used to blow off the garlic skin.
restaurant and hostel mess. Time required for cooking is less
but the time taken for peeling and cutting of onion is more. So 2.2 Summary of literature review
by integrating the both peeling and cutting process, a After going through a lot of literatures and models in the
semiautomatic machine is fabricated to reach the demand. market, the peeling mechanism is inferred from the poultry de-
Generally, the available machines are heavy and their feathering machine, garlic peeling machine and peanut
marketed price refrain people from buying them. crushing machine and the cutting mechanism is inferred from
the vegetable cutting machine.
2 LITERATURE REVIEW
3 PROPOSED SYSTEM
2.1 Literature survey
Maurya and Lami (2019) discussed in detail about peanut 3.1 Existing system and its problems
peeling machine. The machine has crushing, roller and Onion skin is removed manually and knife is used to slice the
separation mechanism. It peels the dry peanut and separates onion. In villages, people put onion in the closed vessel and
the waste. Narayanan and Jagadeesh (2017) have developed they sake it to remove onion skin. The machine which is
automatic vegetable cutting machine. They used the available in the market peels excess fleshy and the removed
slidercrank mechanism to cut the vegetables. The system is skin is separated by manually from the peeled onion. The
powered by the AC motor. The rotary is converted into machine available for cutting is not compatible. Existing
reciprocating motion and it presses the vegetable and cuts the machine doesn’t have integration of both peeling and cutting
vegetables by blades. Balavignesh and Karthikeyan (2016), mechanism together. It does not collect the waste separately.
Present technology is no longer provides precise operation.
————————————————
 P.Ravichandran ME.,Ph.D., is currently working as Associate 3.3 Proposed system and its merits
Professor in Kongu Engineering College, Tamilnadu, India, PH- Integrating both onion peeling and cutting parts together to
+919965998989, E-mail: [email protected]
have a process flow. The system has two major parts that is
 C.Anbu is currently pursuing master degree program in
mechatronics engineering in Kongu Engineering College, Tamilnadu, peeling and cutting parts. Circular drum with rough punched
India. sheet covered inside the drum for peeling the onion. The
 S.Sathishkumar, A.Sakthivel and S.Thenralarasu are currently circular plate with three blades fixed 120o apart is used to cut
pursuing bachelor degree program in mechatronics engineering in the onion. From the circular drum, pathway is created to cutter
Kongu Engineering College, Tamilnadu, India to have a process flow of peeled onion to cutting. Both peeling
part and the cutter use the single motor using a belt drive. The
removed onion skin is collected separately by pipe from
bottom of the drum for using it as a fertilizer. Low budget
2067
IJSTR©2019
www.ijstr.org
 INTERNATIONAL JOURNAL OF SCIENTIFIC & TECHNOLOGY RESEARCH VOLUME 8, ISSUE 12, DECEMBER 2019 ISSN 2277-8616

machine have both onion peeling and cutting operation. Onion

skin is separately collected to use as a fertilizer. Onion peeling TABLE 1
and cutting time is effectively reduced.
LIST OF COMPONENTS
4 FEASIBILITY STUDY
S.No. Materials Specifications
4.1 Economical feasibility
The important parameter of any product which determines its
1 Single phase motor 1350 rpm, 0.25 HP
Success rate is the cost of the device. The cost of peeling
machine available in the market range from ₹20,000 to
2 Submergible pump 1φ, 0.023 HP
₹40,000 depends upon the capacity. Hence cost analysis of
our proposed model should be feasible in order to penetrate
3 Ball bearings 5 mm diameter
the market. The components of the system are easily
available.so the cost of installation will be cheaper and in
terms of mass production, it will be more economical. Our 4 Mild steel drum 40 mm diameter
model peels the onion up to 50kg per hour and it cost around
₹6000. 5 Punched Sheet As required

4.2 Operational feasibility 6 Mild steel Flat plate 36 mm diameter

The model doesn’t use blade or hard friction for peeling, so the
outcomes will whole shaped fresh onion. The operation and 7 Steel frames As required
construction of the project is very simple which leads to easy
identification, rectification and replacement of the components.
The operation is simple, so the chef can easily operate the Based on the different process the conceptual design is
machine. made and is shown in Figure.1. All electrical and mechanical
components are used to finish the required job. From this it
4.3 Technical feasibility can be inferred that setup divided into four parts Onion feeder,
As mentioned earlier, this system involves simple process. Peeler, cutter, Motor.
Since it mostly involves basic components, the project is
technically feasible with low maintenance cost. The Onion
maintenance in the setup is to periodically changing the onion Feeder
cutting blade and peeling metal sheet.

5 WORKING OF THE MODEL

Circular drum with rough punched sheet is covered inside the
drum for peeling the onion and is fixed inside by means of nut
and bold. The circular disc is fixed at the bottom of drum and a
shaft which is extended downwards from the flat plate .The
shaft which is connected to the motor by means of pulley. On
the other end (i.e.) in the cutting section there is a circular Onion
Onion
plate with three blades fixed 120o apart from each other is Peeler
Cutter
used to cut the onion. The AC motor is connected to both the
by means of belt drive. Initially the motor is switched ON, the
cutter blade and the flat plate inside the drum also rotates at
1350 rpm. When the motor reaches the rated speed onion is
feed to circular drum. It also starts rotating along with the plate Onion
and in addition to that water is sprayed to the drum by the Collector AC
Motor
submersible mini pump which is attached at the top of the
drum. The rotation of circular disc at bottom of the drum
creates centrifugal force, due to this force onion circulates
over the punched rough sheets and the onion skin peels off
slowly and it combines with the water and reaches the bottom
and comes out from the hole created at the bottom. After the
skin is peeled the onion needs to be taken out for that hole
created at the side from that hole the onion comes out when
Figure. 1 Conceptual Design
the door is kept open. Onion is guided by the hopper to the
cutting blade, it is a horizontal blade, it slices the onion into
small pieces the sliced onion pieces is collected by the 6. DESIGN CALCULATION
container placed at the bottom of the cutter blade.
6.1 Specification of the flat belt drive
5.1 Components used Driving pulley diameter (D) =40 mm
The Various components used in the project and their Driven pulley diameter (d) =30 mm
specifications are shown in the Table 1. Speed of the motor (N1) =1350 rpm
2068
IJSTR©2019
www.ijstr.org
 INTERNATIONAL JOURNAL OF SCIENTIFIC & TECHNOLOGY RESEARCH VOLUME 8, ISSUE 12, DECEMBER 2019 ISSN 2277-8616

(driving pulley speed) Circular drum = 52cm height, 32cm

Center distance between =300 mm diameter
pulley Circular disc = 30cm diameter
Cutter plate = 20cm diameter
6.2 Speed of the driven pulley Motor
N1/N2 =D/d …… Speed = 1350 rpm
(5.1) Hp = 0.25
N2 = (30×1350)/50 Supply =220 V single phase supply
N2 = 810 rpm Submersible pump
Discharge = 400 lph
6.3 Length of the flat belt drive Supply = 220v single phase supply
L = 2c + + (((D-d)^2)/4 ) ×c …… Power = 18 w
(5.2)
L =0.68 + 2.61+3.5×10^-5
L = 425mm

6.4 Belt velocity

V = ×D×N1)/60 …… (5.3)
= 3.14×0.05×1350
V = 3.5 m/s

6.5 Torque Calculation-Disc

Mass of the Onion =2.0 Kg
Mass of the Plate =0.3 Kg
Mass of the Shaft =0.150 Kg
Total mass =2.45 Kg
Acceleration due gravity = 9.81 m/s2
Force (F) =Total mass x Acceleration due gravity ……
(5.4) Figure. 2 3D Model of the Project
F =2.45 x 9.81
=24.0345 N
Torque = Force x Radius of the Disc …….
(5.5)
=24.0345x 0.15
=3.6051 N-m
Hence the torque needed to rotate the disc containing 2 kg of
onion is 3.6051 N-m

6.6 Torque Calculation-Cutter Blade

Mass of the cutter Plate =0.3 Kg
Mass of the shaft =0.150 Kg
Mass of the onion =0.5 Kg
Total mass =1.05 Kg
Acceleration due gravity = 9.81 m/s2
Force =Total Mass x Acceleration due gravity …….
(5.6)
=9.31 N
Torque = Force x Radius of the Cutter Plate ……. Figure. 3 Prepared Model of the Project
(5.7)
=9.31 x 0.1 m
=0.93 N-m 7 SEQUENCE OF OPERATION
Hence the torque required for Cutting onion is 0.93 N-m
7.1 Feeding Onion and Spraying Water
6.7 Total Torque Required
2 kg of onion is feed into peeling drum which is covered by
Total torque required =Torque required for peeling +
rough punched sheets and water is sprayed from the pump to
Torque required for cutting
the drum. After that top door was closed.
= 3.60 + 0.93
T = 4.531 N-m
7.2 Onion Peeling
Hence the total torque required for whole process is 4.6351N-
The motor is switched on and the circular disc will rotates. Due
m
to centrifugal force, the onion is circulates around rough
punched sheet and the skin is peeled.
6.8 SPECIFICATION
7.3 Separately Collecting the Peeled Skin
2069
IJSTR©2019
www.ijstr.org
 INTERNATIONAL JOURNAL OF SCIENTIFIC & TECHNOLOGY RESEARCH VOLUME 8, ISSUE 12, DECEMBER 2019 ISSN 2277-8616

The peeled skin is separated by the water and fed to the side journal of innovative research in science engineering and
gap of the circular disc. From the bottom of drum, there is a technology.
hole which is connected by a pipe. By this extent the peeled [11] Muthukrishnan and nanthagobal , 2014, Automated
skin with water is comes out and collected to use as a vegetable cutting machine, 5th International conference
fertilizer. Even without using the water the onion skin is [12] Reuben Donado and Don Martin Fernandez , 2015,
removed by using blower setup to separate the onion skin. Design and Fabrication and Testing of a Semi-automatic
Green Mango Peeling Machine, 8th IEEE International
7.4 Onion Cutting Conference Humanoid, Nanotechnology, Information
The peeled onion is passed to the hopper by the pathway. The Technology Communication and Control, Environment and
cutter rotates by the belt drive from the motor pulley, which is Management.
used for peeling. And the onion is sliced and collected. When
the onion falls on the hopper, hopper guides the onion to
reach the cutter. The cutter is powered by the motor used for
peeling operation. When the cutter blades present in the setup
slices the onion when it falls on the cutter. It is rotating around
800 revolutions per minute. Various sizes of blades are used
to cut the onion of different sizes for various preparations of
food items. This mechanism slices the onion rather than
chopping it.

8 CONCLUSIONS
The prototype model of Onion peeling and cutting machine is
designed (Figure 2) and fabricated (Figure 3) to reduce
manpower and time. The motor with torque capacity of 5 Nm is
enough to run the circular disc and the cutter for peeling and
cutting the onion respectively. The designed machine will peal
the 2kg of onion in 3 minutes and for both peeling and cutting,
it takes 8 minutes to complete the process. This model
successfully peeled the onion and while cutting, it has 75%
efficiency, but it can be overcome by the changing the cutter
blade to have a desired output. The peeled skin is stored
separately and makes a use as a fertilizer to agriculture. By
this technic of onion peeling and cutting will effectively reduce
the preparation time for onion than the cooking. Overall the
proposed model was able to satisfy the proposed advantage.

REFERENCES
[1] Agbettoye, C.A.S. and Balogun, Design and Performance
evolution of a multicrop slicing machine, United States
Patent,188,240
[2] Adetola, S.O and Darmola, G.A., Development of poultry
de-feathering machine with better efficiency, 2014, ISRO
journal of Mechanical and civil engineering
[3] Balavignesh, J. and Karthikeyan, S., 2016, Automatic
French fries making machine, International conference on
control instrumentation, communication and computational
technology.
[4] El-Ghobashy and Adel Bahnasawy, H., 2012,
Development and evaluation of onion peeling machine,
Research gate , Misr J. Ag. Eng., 29 (2): 663 –682
[5] Hari narayanan and Jagadeesh, 2019, Automatic
Vegetable cutting system,International journal of
innovative and emerging research in engineering.
[6] James Berube, A., 1988, Fruit peeling machine,
P.N.4,738,195,A.N.15,686.
[7] Javier and Aitor Aguirreta, 2012,Cutting Grid, US
2012/0125172 A1
[8] Louis Lazzarini, P.,1974, Peeling machine and method,
United States Patent,188,240.
[9] Mohan, S., Garlic peeling machine, 2015, international
journal of engineering and general science.
[10] Maurya Mohit and Cadmi meet, 2019, Design and
Development Peanut peeler machine , International
2070
IJSTR©2019
www.ijstr.org
 Special Issue - 2020 International Journal of Engineering Research & Technology (IJERT)
ISSN: 2278-0181
NCETESFT - 2020 Conference Proceedings

Design and Development of an Automated Floor

Cleaning Robot for Domestic Application
Raveena Ishalavath M Sandhyashree N
Students, Dept. of Telecommunication Engineering, Students, Dept. of Telecommunication Engineering,
GSSSIETW, Mysuru GSSSIETW, Mysuru

Rakshitha R Arpitha Shankar

Students, Dept. of Telecommunication Engineering, Assistant Professor, Dept. of Telecommunication
GSSSIETW, Mysuru Engineering, GSSSIETW, Mysuru

Abstract - With the advancement of technology, robots have automated to perform services helpful within the well-being of
gotten more attention of researchers to create lifetime of mankind humans and equipment. Robots of types including medical
easy. The actual project presents the look and development of robots, underwater robots, surveillance robots, demolition
Floor Cleaning Robot using IEEE Standards. Developing robots and other styles of robots those do a multiple jobs. they'll
Bluetooth controlled mobile robot. Scan the obstacles ahead of the clean floors, mow lawns and guard homes and can also help in
robot and to avoid collision when the robot is in motion. Raspberry assisting old and disabled people, perform some surgeries,
Pi3 is that the main component won’t to control the cleaning robot. checking pipes and sites that are highly dangerous to people,
An ultrasonic sensor which transmits the ultrasonic waves from its fight fires and defuse bombs.
sensor head and again receives the echo waves and sends its output
to the Raspberry Pi3. The ultrasonic sensor is connected with the II. LITRATURE SURVEY
servomotor, which helps within the rotation of ultrasonic sensor.
The ultrasonic sensor measures the space between the robot and Floor cleaning robot may be a trending concept in these
therefore the obstacle ahead of it. The Pi3 model will stop the robot recent days. By reviewing different paperwork and techniques
immediately and also the buzzer are actuated. The moping of used several cleaning robots, we've started acting on our
operation will be started or stopped at any point of your time as design of floor cleaning robot which is predicated on Raspberry
per the need. The moping brush is actuated by the DC motor fixed Pi 3 model. The papers surveyed for literature review are as
thereto. Signal to the present motor is fed by the controller. An follows:
LCD displays each and each operation applied by the robot.
Buzzer is an audio signalling device that provides the indication of
Aishwarya Pardeshi et. al, [1] This paper presents the look,
operating status of the robot. developed and fabricated model of programmed cleaner robot.
this type of robot performs automated function with extra
features like choose and place mechanism and dirt container
Keywords - Raspberry Pi 3; Sensors; Buzzer; LCD; DC Motor. with air vacuum mechanism. this type of labour is
straightforward and helpful in betterment of life variety of a
I. INTRODUCTION mankind.
These day’s humans lead a sophisticated life. People within Ajith Thomas et. al, [2] proposed an autonomous robotic for
the cities don't have regular and have long working times. In floor cleaning program. it's able to perform sucking and
such a situation someone will choose time saving methods. Thus cleaning, detection of obstacles, and water spraying.
robots have taken the manual works. For career oriented and job Furthermore, it's also able to add manual method. All hardware
going women it becomes hectic to handle home and office and software functions are manipulated by Raspberry pi3 model.
together. Traditionally floor is cleaned with the assistance of
mop or wet mop using the hand as a possible tool. they need to Vaibhavi Rewatkar and Sachin T. Bagde [3] provided a
clean hard on the surface. The cleaning includes cleaning of comprehensive overview of the technological advantages helped
varied surfaces basically cement floors, highly polished wooden within the real world for the convenience of just about all of the
or marble floors. Among these floors the rough surface floor like people that are extremely busy. Consequently, this has led to
cement floor, mostly present in semi urban areas are covered arriving up with a goal of constructing an automatic home
with such a lot dust which needs longer for cleaning. For saving appliance. The review includes computerized cleaner having
the time the necessity was of House Cleaning Robot, which is components to DC motor operated wheels, the dustbin,
an automatic system that works and cleans on its own without cleansing brush, mop cleansing and obstruction avoiding sensor.
human control/intervention. Autonomous robot for floor A 12V battery is employed for supplying power. Special
cleaning application reduces much time in lifestyle. It performs technique of ULTRAVIOLET germicidal cleaning technology.
sweeping and mopping tasks at a time, it also does obstacle The study has been done keeping in mind economical expense
detection, and also has automatic water spray. Service robots are of product.
getting popular recently these robots operate semi- or fully

Volume 8, Issue 14 Published by, www.ijert.org 67

 Special Issue - 2020 International Journal of Engineering Research & Technology (IJERT)
ISSN: 2278-0181
NCETESFT - 2020 Conference Proceedings

Vinod J Thomas et. al, [4] designed a cleaner robot for Zelun L, Zhicheng Huang, [9] designed a cleaning automatic
domestic application. The robotic contains a cleaning module robot predicated on the ultrasonic basics. With the sole chip
which may be used for cleaning. The Robot was created in order microcomputer AT89C52 and ultrasonic detectors the robotic
that it may well be capable of reach almost every space and can do the characteristic of practical impediment avoidance,
corner of any room that it must be as compact as possible. The programmed manage and programmed sweeping. Within the
working robot is handled using an Android phone using cleaning automatic robot, a revolving cylindrical brush is used
Wireless Bluetooth Technology. The robot was created with an before the automated robot and it sweeps garbage in to the
Arduino microcontroller at its core. The microcontroller is dustbin along the way of motion, and a mop is used behind the
complemented with communications modules like Wireless automated robot, and it can sweep the ground when the
Bluetooth motors and dirt Suction System to work accordingly. automated robot is functioning.
Manya Jain et. al, [5] discussed the event of Automatic Floor Rupinder Kaur [10] designed a swabbing automatic robot
Cleaner. The project is often used for domestic and professional which is extremely good for cleaning jobs especially in homes,
purpose to scrub the surface automatically and manually. When Office buildings, Industries where sanitation could be a
it's turned ON, it gulps within the dust particles by moving significant matter. Many research organizations are active in
everywhere the surface (floor or the other area) because it moves locating the most effective results through the unreal
over it. the driving force control mechanism are often wont to intelligence. Certainly, artificial intellect could be a branch of
drive the motors where robot having the ability to manoeuvre technology which makes computers sounds like mind. This
and also the also few sensors are accustomed detect and avoid product will sweep, and mop the bottom area with clean and
the obstacles. this can be often useful in making the approach to other wiping components; and yes it collects the dust particles
life better for humankind. and other small parts in it. Mapping is wont to instruct this small
device. These devices is just too simple to use, very affordable
Abhishek Pandey et. al, [6] reviewed the requirement of a and cleans every nook of the region. Being autonomous, it could
residence Cleaning Automatic robot. For keeping time there’s a add one’s absence.
requirement of programmed system that cleans alone without
person interventions. Also, they considered how precisely to S Monika, K Aruna Manjusha et. Al [11] This research paper
help those that have physical disabilities. Because that they had presents that floor cleaning is worn out a neater way and
to induce this done, they needed a cleaning system that may add efficiently by robot utilizing wireless system. This proposed
accordance from what we are saying, thus supporting a robot saves the time and economy of labour. within the previous
physically someone. research papers like robot household appliance and automatic
floor cleaner robot had some drawbacks like colliding with
Karthick.T et. al, [7] is intended to create up an autonomous objects before of it and this vacuum couldn’t reach to small areas
automatic robot which will move itself without constant human and left those areas unclean and therefore the automatic floor
instruction. The autonomous cleanser robot involves low power cleaner robot collects the dirt but the downside up here is that it
consuming electric components and it can operate at very low doesn't clean the wet floor. Few of the drawbacks during this
power. Electric parts are the controller board Atmega 2560, project paper are overcome. Amit Sharma,
Ultrasonic detectors, transformer IC and motor driver circuit.
Mechanized part is motor unit with gearbox founded. Ultrasonic Akash Choudhary et. al [12] the target of this project is to
detectors will identify obstructions in line with the program form a totally automated hybrid home cleaning robot. Which is
being executed. A 12V, 4.5Ah rechargeable lead acid electrical fully automated and may perform tasks like mopping and
device is that the energy source for this proposed cleaning cleaning of floor. After the testing we discover that it can
automatic robot. perform all tasks fine with none hurdle. We tested our robot on
various parameters like path following, obstacle avoidance,
Manreet Kaur and Preeti Abrol, [8] came up with the navigation, mopping and vacuum mechanism.
working of automatic robot Floor cleaning. This automatic robot
can add any of two methods. All hardware and software III. LITERATURE REVIEW AT GLANCE
functions are handled by AT89S52 microcontroller. This
automatic robot is in a position to perform sweeping and
Sr. Title of Paper Further extension Major
mopping job. RF modules is getting used for cordless No. Contribution
communication between remote (manual method) and automatic “Automatic Pick and Brings flexibility
robot has range of 50m. This robot is given with IR sensor for I. floor Cleaner” place mechanism to do work
obstacle recognition and automates water sprayer pump. Four
motors are being employed, two for cleaning purpose, one for II. “An Advanced Environmental friendly Less time
pump and one for tires. Dual relay circuit is employed to work Mobile Robot for Floor consuming
Cleaning”
the motors one for the pump and another for the cleaner. In III. ”floor cleaning robot” Auto Helps
previous works, there's no use of automated water sprayer and disposal mechanism physically
works only in programmed mode. With the automated mode disabled people
automatic robot controls all the functions itself and alter the road IV. “Automatic work automatically. able to
if just in case there's hurdle detection and moves back again. Floor Cleaner” cover large
With the manual method, the keypad will be accustomed execute floor areas.
V. A Technological Dealing with Saves time,
the expected job and operate automatic robot. In manual method, Survey on some small pieces Helps
RF component is employed to transfer the knowledge between Autonomous of garbage, physically
remote and automatic robot and display the data associated with Home Cleaning Robots such as paper chips, disabled people
the hurdle detection on LCD. the entire circuitry is associated paper & soil block
with 12V electrical device pack.

Volume 8, Issue 14 Published by, www.ijert.org 68

 Special Issue - 2020 International Journal of Engineering Research & Technology (IJERT)
ISSN: 2278-0181
NCETESFT - 2020 Conference Proceedings

VI. Simple Consumes less power Saves time

Autonomous
cleaner Robot”

IV. PROBLEM STATEMENT

Nowadays, people lead a busy life. People in urbans have
abnormal and long working hours. In such a situation an
individual will always find ways of saving time.
 For career oriented and dealing women it's hard to handle
home together with job work.
 Normally floor is cleaned with the utilization of dry mopped
or wet mopped using the hand as a base tool. they need to be
scrubbed hard on the surface.
 The cleaning module includes cleaning of varied surfaces
like cement floors, highly polished wooden or marble floors.
 The rough surface areas like cement floor, are covered with
heavy dust which consumes longer in cleaning. According to the programming instructions in the Raspberry
Pi 3, as the robot starts moving the brush fixed in the front part
V. SOLUTION STRATERGY of the robot starts sweeping. The moping operation can be
 For time saving purpose the House Cleaning Robot is started or stopped at any point of time as per the requirement.
important, which is an The moping brush is actuated by the DC motor fixed to it. Signal
to this motor is fed by the controller.
 Autonomous robot for floor cleaning application reduces
much time in existence. It does sweeping and mopping tasks An LCD is a display device that receives signal from the
at a time, it also detects obstacles, and has automatic water Raspberry Pi 3 and displays each and every operation carried out
sprayer. by the robot. Buzzer is an audio signalling device that gives the
indication of operating status of the robot.
 Automated Floor Cleaner are designed for cleaning offices,
homes also in collages. In one amongst the mode this robot According to the programming instructions in the Raspberry
is making decisions on the premise of humans or various Pi 3, as the robot starts moving the brush fixed in the front part
sensors which are employed in this robot. of the robot starts sweeping. The moping operation can be
started or stopped at any point of time as per the requirement.
 Manual works will be replaced by the robot technology and The moping brush is actuated by the DC motor fixed to it. Signal
lots of the related robot system applications are used. to this motor is fed by the controller.
VI. METHODOLOGY An LCD is a display device that receives signal from the
Raspberry Pi 3 and displays each and every operation carried out
Battery operated cleaning robot cleans and mops at the same
by the robot. Buzzer is an audio signalling device that gives the
time using the smartphone connection. Raspberry Pi 3 is the
indication of operating status of the robot.
main controller used to control the cleaning robot. It is a
Raspberry Pi 3 model based on Raspbian buster OS. Raspberry VII. OBJECTIVES
Pi 3 is open source software in which hardware can be easily
used. Raspberry Pi 3 is energized by 12V, DC battery. Bluetooth  Developing Bluetooth controlled mobile robot.
electronics app controls cleaning robot with an android device.  To scan the obstacles earlier than the robot and to avoid
This app communicates using Bluetooth to an HC-05 Bluetooth collision when the robot is in motion.
module in the robot. An ultrasonic sensor is used for obstacle
detection which transmits the ultrasonic waves from its sensor  To test the trail status using IR sensor.
head and again receives the echo waves and sends its output
signal to the Raspberry Pi 3 will stop the robot immediately and  To create a robot this carries out sweeping operation
the buzzer will be actuated. The ultrasonic sensor is connected efficiently.
with the servomotor, which helps in the rotation of ultrasonic  To try to moping operation whenever needed along with
sensor. The ultrasonic sensor measures the distance between the sweeping.
robot and the obstacle in front of it. If any obstacle is present in
front of the robot, IR sensor gives signal to an ultrasonic sensor VIII. CONCLUSION
is used for which it transmits the ultrasonic waves. L298N driver The robot is specially built on the use of modernised
circuit is used to drive the DC motors simultaneously in all of technology. it's all the features that are required for a floor
the direction. Raspberry Pi 3 sends the signal to the motor driver cleaner. It works automatically and manually. it's auto drain
circuit that controls and drives the wheel. feature. this could be locally manufactured smart floor cleaner
robot. Meanwhile it's scheduling features which can be operated
by computing only, android and windows app can make it little
more user friendly.

Volume 8, Issue 14 Published by, www.ijert.org 69

 Special Issue - 2020 International Journal of Engineering Research & Technology (IJERT)
ISSN: 2278-0181
NCETESFT - 2020 Conference Proceedings

ACKNOLEDGEMENT
We gratefully thank for the help and co-operation offered by
Dr. Parameshachari BD, professor and head, Dept of
Telecommunication Engineering, and our project guide, and
Management of GSSSIETW, Mysore for providing help and
support to carry out the project.
REFERENCES
[1] P. Aishwarya, S. More, D. Kadam, V.A. Patil, “Automatic Floor
Cleaner”, IJECT vol. 8, 2017.
[2] T. Ajith, M. S. Rohith, J. Febin, J. Cheriyan, R, Mary George, “An
Advanced Mobile Robot for Floor Cleaning”, International Journal of
Advanced Research in Electrical, Electronics and Instrumentation
Engineering, vol. 5, no. 3, 2016.
[3] R. Vaibhavi and S. T. Bagde, “A Review on Design of Automated Floor
Cleaning System”, International Journal on Recent and Innovation Trends
in Computing and Communication, vol. 3, no. 2.
[4] V. J Thomas, B. Xaviour, J. K George, “Cleaner Robot”, International
Journal of Emerging Technology and Advanced Engineering, ISSN 2250-
2459, ISO 9001:2008 Certified Journal, vol. 5, no. 12, 2015.
[5] M. Jain, P. S. Rawat, J. Morbale, “Automatic Floor Cleaner”,
International Research Journal of Engineering and Technology (IRJET),
vol. 4, no. 4 , 2017.
[6] A. Pandey, A. Kaushik, A. K. Jha, G. Kapse, “A Technological Survey on
Autonomous Home Cleaning Robots”, International Journal of Scientific
and Research Publications, vol. 4, no. 4, 2014.
[7] T. Karthick, A. Ravikumar, L. Selvakumar, T. Viknesh, B. Parthiban. and
A. Gopinath, “Simple Autonomous cleaner Robot”, International Journal
of Science, Engineering and Technology Research (IJSETR), vol. 5, no.
3, 2016.
[8] R. Vaibhavi and S. T. Bagde, “A Review on Design of Automated Floor
Cleaning System”, International Journal on Recent and Innovation Trends
in Computing and Communication, vol. 3 no. 2.
[9] J. T. Vinod, B. Xaviour, J. K. George, “Cleaner Robot”, International
Journal of Emerging Technology and Advanced Engineering, vol. 5, no.
12, 2015.
[10] M. Jain, P. S. Rawat, J. Morbale, “Automatic Floor Cleaner”,
International Research Journal of Engineering and Technology (IRJET),
vol. 04, no. 4 , 2017.
[11] A.Pandey, A. Kaushik, A, K. Jha, G. Kapse, “A Technological Survey on
Autonomous Home Cleaning Robots”, International Journal of Scientific
and Research Publications, vol. 4, no. 4, 2014.
[12] T. Karthick, A. Ravikumar, L. Selvakumar, T. Viknesh, B Parthiban. and
A. Gopinath, “Simple Autonomous cleaner Robot”, International Journal
of Science, Engineering and Technology Research (IJSETR), vol. 4, no.
5, 2015.
[13] M. Kaur, P. Abrol, “Design and Development of Floor Cleaner Robot
(Automatic and Manual)”, International Journal of Computer
Applications (0975 – 8887), vol. 97, no. 19, 2014.
[14] [12] Zelun L, Zhicheng Huang, “Design of a type of cleaning robot with
ultrasonic”, Journal of Theoretical and Applied Information Technology
31st January 2013. Vol. 47 No.3, ISSN: 1992-8645
[15] K. Rupinder, “An extremely cost efficient Swabbing Robot”,
International Journal of Engineering and Computer Science ISSN: 2319-
7242, vol. 6, no. 1, 2017.
[16] S Monika, K Aruna Manjusha, S V S Prasad, B.Naresh “Design and
Implementation of Smart Floor Cleaning Robot using Android App”
International Journal of Innovative Technology and Exploring
Engineering (IJITEE) ISSN: 2278-3075, Volume-8 Issue-4S2 March,
2019.
[17] S. Amit, A. Choudhary, A. Gaur, and A. Rajpurohit, “Fully automated
hybrid home cleaning.” International Journal of Engineering
Technologies and Management Research, vol. 5, no. 3, pp. 219-225, 2018.
[18] N. Fathima, A. Ahammed, R. Banu, B.D. Parameshachari, and N.M Naik,
“Optimized neighbor discovery in Internet of Things (IoT),” In Proc. of
International Conference on Electrical, Electronics, Communication,
Computer, and Optimization Techniques (ICEECCOT), pp. 1-5, 2017.

Volume 8, Issue 14 Published by, www.ijert.org 70

Ho Chi Minh City University of Technology and Education
Faculty of Mechanical Engineering

RESEARCH PAPERS
Automatic Drilling Machine Based on PLC

Author:Jaiprakash Kapse , Paras Choudhari, Bhushan

Khobragade, Lokesh Gaidhane, Himanshu
Barapatre, Dr.A.C.Waghmare
Origin: India
Instructor: Associate Professor Ph. D Tran Ngoc Dam.
Presenter: Nguyễn Nhất Nam
ID: 21143187.

Ho Chi Minh City, March 2023

CONTENTS
1. Introduction
1a. System Components and Hardwar Structure
2. Overview
3. Litreature Review
4. Conclusions
1. Introduction
 Conventionally drilling machine are hand operated
 This engages a Person with time consuming
 Automatic Drilling Machine Based on PLC
=> The system will reduce the Human strength
1. Introduction
 The PLC use perform by reading data
 increase reproducibility with improved accuracy
 the quality of drilling process is optimized.
1. Introduction

 The sensors and switches (input devices).

 The controller (PLC)
 The relay (output devices)
 The drilling machine which is operate
automatically.
1a. System Components and Hardware Structure

 Inductive proximity sensor and capacitive proximity sensor

 Switches are electrical or electronic contact devices
 A programmable logic controller or programmable controller
 PLC program generally loops executes repeatedly
1a. System Components and Hardware Structure

There are 5 main steps in a scan cycle :-

o 1. Reading inputs
o 2. Executing the program
o 3. Processing communication requests
o 4. Executing CPU diagnostics
o 5. Writing outputs
3. Overview

 PLCs can range from small modular devices

with tens of inputs and outputs (I/O)
 They can be designed for multiple arrangements of
digital and analog I/O
 Programs to control machine operation are typically
stored in battery-backed-up, nonvolatile memory
 4.LITREATURE REVIEW

 Automation of drilling machine with the help

of Pneumatic Devices
 The project “Automated drilling machine
with depth controllability
 Drill two holes at a time
 4.LITREATURE REVIEW

 The existing drill machine was

successfully automated using PLC control system
 The tool bit automatically
withdraws from work piece by reversible speed
=> The efficiency of this machine is very high
5.CONCLUSIONS

 The automatic drilling machine was successfully designed

 Perform smoothly and reliably
 Increase the efficiency of drilling machines instead human work
 5.CONCLUSIONS

 The system become more

Advantageous
 Reduces all the drawbacks of existing
manual drilling machine
 Increase the productivity in terms of both quality and
quantity
VI/ Paragraph
Paragraph 1:
Mechanical engineering is an important field of study that deals with the design,
manufacturing, and maintenance of mechanical systems. Mechanical engineers use their
knowledge of physics, mathematics, and materials science to create new machines,
engines, and mechanisms that can solve complex problems. They work in a wide range of
industries, from automotive and aerospace to biomedical and energy. Mechanical
engineering also plays a critical role in the development of renewable energy
technologies that can help reduce our reliance on fossil fuels and combat climate change.
Paragraph 2 :
One of the key areas of focus in mechanical engineering is automation. Automation
involves the use of machines and systems to perform tasks that would otherwise
require human intervention. Automated systems can improve efficiency, reduce costs,
and increase safety in manufacturing and other industries. Mechanical engineers
design and develop automated systems that are reliable, efficient, and cost-effective.
Paragraph 3 :
Advancements in technology have led to new areas of focus in mechanical
engineering, such as mechatronics. Mechatronics is an interdisciplinary field that
combines principles of mechanical engineering, electrical engineering, and computer
science to design and develop advanced systems. These systems can include robots,
automated vehicles, and intelligent machines that can learn and adapt to their
environment. As technology continues to evolve, the role of mechanical engineers in
creating innovative solutions will become increasingly important.
Paragraph About processing
The process of assembling a robot arm involves several steps. First, the individual
components such as the motors, gears, and sensors need to be carefully inspected and
tested for any defects or malfunctions. Once the components are deemed suitable, they
are then assembled together in the correct order using specialized tools and techniques.
This may involve soldering, fastening screws, or connecting wires. After the arm is fully
assembled, it undergoes rigorous testing to ensure that it can perform the intended
movements accurately and consistently. Once the arm has passed all the necessary tests,
it is then ready to be integrated with the rest of the robot and put into operation. The
process of assembling a robot arm requires precision, attention to detail, and a thorough
understanding of the functionality of each component.
 NAM
NGUYEN NHAT
Mechanical design engineer

CONTACT PROFILE INFO

0908925008 Today's mechanical design engineer has become an

indispensable part of 4.0 technology. I am to be a
[email protected] global engineer with extensive experience as well as
good communication skills.
4D Tran Thi Vung, LA

MY SKILLS EDUCATION

Autocad Manufacturing Engineering

UTE University
Inventor 2021-2025

Solidworks
WORK EXPERIENCE
English

Japanese Mechanical design engineer

EINTEK COMPANY 2025-2027
Marketing
Design And developing Arm Robot for Company
Design Drilling machine
Design automatic cutter machine
Employee of the month
REFERENCE

UTE university
EINTEK COMPANY

Graduated with Distinction 2021-2025

M : [email protected]
P : 0944244468 Praize for student of 5 merits
Blood Donation Every years

184 Duong Dinh Hoi TP Thu Duc