Portfolio NguyenTrongPhuc 20146522 First

BỘ GIÁO DỤC VÀ ĐÀO TẠO
TRƯỜNG ĐẠI HỌC SƯ PHẠM KỸ THUẬT TPHCM

KHOA CƠ KHÍ CHẾ TẠO MÁY
BÁO CÁO
MÔN HỌC: ANH VĂN CHUYÊN NGÀNH CƠ KHÍ
BÁO CÁO CUỐI KỲ
GVHD: PGS.TS Trần Ngọc Đảm

Mã lớp học phần: ENME320124
SVTH : Nguyễn Trọng Phúc_20146522
TP. Hồ Chí Minh – 2022

1. Free paragraph
(Free writing a project, which has done)
My name is Nguyen Trong Phuc, i come from Faculty of Mechanical Engineering of
HCM University of technology and education. For two years, since I have been studying
at University as a mechatronics student. I have fisnished many projects at many subjects.
Most of them were interesting to me, but some others were very hard and took me lot of
time to do it. But after all, that give me many experience and knowledge to be a best
engineer in the future. I enjoy to study, research and learn about the operating and
principles of machines, devices and many equipments. One of many projects I have done,
which i like most is create automatic banhbeo making machine. The project requires a
team of 3 members to design a mechine, which applied pneumatic technology of
hydraulic and pneumatic technology subject. After the project, I have improved my soft
skill about work in groups, understand more about the operating structure and design
method of a pneumatic machine.
2. 5 Titles and abstracts

5 titles:
- The Research of Ho Chi Minh University of Technology and Education’s Student about
Find out a New Composite Materials.
- The Development of Artificial Intelligence in The Field of Biology.
- The Effective Application of Science and Technology in Education.
-The Role of Earth Worm in Ecological and Agricultural Development in VietNam.
- The Contribute of Tesla’s Roket Technology to Aerospace Engineering.
5 abstracts:
1 - Designing an Automatic Banh Beo Making Machine Using Pneumatic System.
2 - Designing an Automatic Playing Caro Robot’s Arm with Two Degree
Using Step.
3 - Developing 3600 Welding Cutting Rotary Turn Table Welding Positioner.
4 - Alternating Pressure Control System for Hydraulic Robots.
5 – Designing an Four Wheel Control System Using Steering Mechanism.
Abstract and comment:
Abstract
Cam Shaft Mechanism DIY Ventilator
COVID19 shocked humankind with a wave of unexpected health crisis in
2019 2020 and 2021 and still continues to plague humankind in some region or the
other. As COVID family of viruses majorly affect the lungs while limiting lung
capacity and causing breathing problems, ventilator development and research has
gained particular importance.
Human lungs use the reverse pressure generated by contraction motion of
the diaphragm to suck in air for breathing. A contradictory motion is used by a
ventilator to inflate the lungs by pumping type motion.
A ventilator mechanism must be able to deliver in the range of 10 – 30
breaths per minute, with the ability to adjust rising increments in sets of 2. Along
with this the ventilator must have the ability to adjust the air volume pushed into
lungs in each breath. The last but now the least is the setting to adjust the time
duration for inhalation to exhalation ratio. Apart from this the ventilator must be
able to monitor the patients blood oxygen level and exhaled lung pressure to avoid
over/under air pressure simultaneously. The ventilator we here design and develop
using a controller that encompasses all these requirements to develop a reliable yet
affordable DIY ventilator to help in times of pandemic.
We here use a silicon ventilator bag coupled driven by Stepper motor with
single side push mechanism to push the ventilator bag. For this we hereby make
use of cam Shaft mechanism to achieve inflation and deflation. We make use of a
stepper motor to drive the mechanism. We connect the stepper motor shaft to a
cam in order to convert rotary motion into linear motion here. The cam is oval on
shape designed to push one end of the pressing arm upwards. The pressing arm is
mounted on top of the ventilator bag and connected to joint mechanism in
between. This creates a seesaw like mechanism.
Now as stepper motor rotates and pushes the arm upwards on one site, it
presses against the bag on other end. Now the rate of inflation and deflation
depends on the RPM of motor. As per settings provided we vary the motor RPM
to achieve desired BPM rate. We use toggle switch for switching and a variable
pot to adjust the breath length and the BPM value for the patient.
Our system makes use of blood oxygen sensor along with sensitive pressure
sensor to monitor the necessary vitals of the patient and display on a mini screen.
Also an emergency buzzer alert is fitted in the system to sound an alert as soon as
any anomaly is detected. The entire system is driven by a controller circuitry to
achieve desired results and to assist patients in COVID pandemic and other
emergency situations.
Comment
This abstract is almost very detailed and complete, with clear goals, methods and
results, consistent with the current situation of the world.
However, abstract only talks too much about the principle, but the conclusion
of the article doesn’t mention the actual results achieved, and the development
direction of the product to apply in practice.
3. 4 Paragraph
1- Robot in Our Life
A robot is a machine, an automaton which can carry out a variety of tasks to
replace human effort. Robots are present around us, and they don’t look like the
humanoid robots as are made popular in movies and literature. In our daily life
robots are commonly used for vacuum cleaning. It’s a small machine which can
clean the floor by itself and can do it even in the absence of a human being.
Similarly, a robotic lawnmower will mow the grass while the owner tends to other
chores. In factories use robots to assemble products on assembly lines. This is
efficient since most of the assembly work is mechanical and repetitive, and can
save human time and energy. And in many benifits The robots can go far down
into the obscure spots where the people would be squashed. The robots can play
out the assignments quicker than the people and substantially more reliably and
precisely. The majority of robots are programmed in such a way that they can
move with no human impedance. The robots can be modified to achieve the
Earth’s nadirs, for example, to borrow for the fuels.
2- What’s Artificial Intelligence
Artificial Intelligence is the wide branch of computer science. It deals with
developing smart machines that can operate without human interaction. Today we
can think of a machine that can work according to our needs without providing
instructions. Artificial Intelligence can be used in video games, mobile phones,
cars, surveillance, etc. Many famous social platforms like Google, Facebook and
many more use Artificial Intelligence. The recommendation in Google or any
other site uses AI. Google assistance and Siri in iPhone is a virtual assistance
program which is an example of Artificial Intelligence. Another famous example
of Artificial Intelligence is Amazon’s Alexa and Amazon Echo developed by
Amazon. We have heard of self-controlling cars. They do work on the principle of
Artificial Intelligence.
3- Brain’s Architecture Inspired CNNs
A typical neural network will have an input layer, hidden layers, and an output
layer. CNNs are inspired by the architecture of the brain. Just like a neuron in the
brain processes and transmits information throughout the body, artificial neurons
or nodes in CNNs take inputs, processes them and sends the result as output. The
image is fed as input. The input layer accepts the image pixels as input in the form
of arrays. In CNNs, there could be multiple hidden layers, which perform feature
extraction from the image by doing calculations. This could include convolution,
pooling, rectified linear units, and fully connected layers. Convolution is the first
layer that does feature extraction from an input image. The fully connected layer
classifies the object and identifies it in the output layer. “CNNs are feedforward
networks in that information flow takes place in one direction only, from their
inputs to their outputs. Just as artificial neural networks (ANN) are biologically
inspired, so are CNNs. The visual cortex in the brain, which consists of alternating
layers of simple and complex cells, motivates their architecture. CNN
architectures come in several variations; however, in general, they consist of
convolutional and pooling (or subsampling) layers, which are grouped into
modules. Either one or more fully connected layers, as in a standard feedforward
neural network, follow these modules,” describes a paper on “Deep Convolutional
Neural Networks for Image Classification: A Comprehensive Review”, in4. the
journal ‘Neural Computation’, which is published by MIT Press.
4- Application of artificial intelligence in medicine
Artificial intelligence (AI) is tending to be strongly applied in medicine such as:
disease diagnosis; drug research and development; optimized for individual
treatment; gene editing.
In the ability to diagnose diseases: Recently, AI has made great progress in
automatically diagnosing diseases, making diagnosis cheaper, faster and more
accessible. The huge data about normal images, pathology, body indexes... will be
"labeled", loaded into the computer, sorted, processed..., from which the computer
can receive identify, classify, and then make diagnoses when they come in contact
with certain patient data.
Optimizing for Individual Treatment: AI can automate this highly complex
statistical work and help uncover characteristics that indicate a patient will
respond specifically to a particular treatment. from which to come up with the
optimal treatment regimen for each individual patient.
Research and development of drugs: Currently, AI has been successfully applied
in all 4 main stages of drug research and development process (evaluating action
purposes; finding the right drug molecules/drugs). capable of binding to the
intended purpose; testing the new compound in the laboratory and in the clinical
setting for safety and effectiveness; gaining consent and getting the new drug into
the hands of physicians and patients ).
6 Research papers
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 11 | Nov 2022 www.irjet.net p-ISSN: 2395-0072
IOT-BASED HAZARDOUS GAS LEAKAGE DETECTION

Ruchika Chakraborty1, Ayan Prakash Gogoi Boro2, Bapon Sarkar3 , Asst. Prof. Dr. B. Siranthini 4
1Student, Department of Electrical and Electronics Engineering, SRM Institute of Science and Technology,
Ramapuram, Chennai, India


3Assistant Professor, Department of Electrical and Electronics Engineering, SRM Institute of Science and
Technology, Ramapuram, Chennai, India

---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Human beings and all the living organisms in this signal. As soon as the gas is detected by the MQ5 sensor it
world needs fresh and healthy air to survive. In recent years, displays alert messages on the LCD screen. The user will get
with the increase in gas consumption, gas leakage has become alerted via messages on the LCD screen. This is an efficient
a problem. Leaked gas causes gas waste, especially because of way to automatically detect and control gas leaks. It also
its combustible property, it can harm living organisms and prevents accidents. The idea of gas detection and control can
various industries. All components are controlled by an be implemented on a large scale in various industries. The
Arduino which acts as the central processing unit of the system can be installed in kitchens, hostel cafeterias, etc.
setup. When the sensor detects a leak of the many This helps reduce accidents caused by gas leaks in homes
combustible gases, it sounds an alarm with a buzzer. This and commercial facilities. This system is low cost, so it is
alarm supports a small LCD that displays the amount affordable, prevents many accidents, saves many accidents,
of the gas leak and alerts personnel to turn on the exhaust fan saves many properties and lives. The MQ5 sensor is used to
or stop the incoming gas at a particular section. This will detect H2 (molecular hydrogen), LPG, CH4 (methane), CO
notify to take immediate security measures. The capabilities of (carbon monoxide), and alcohol. This system is not only
this gas detection system are not only to monitor the various capable of detecting a gas leakage as well as alerting the user
surroundings but also help to prevent the gas leak and hence of the gas leakage through a buzzer alarm.
minimize the chances of fire and leakage damages.
2. LITERATURE SURVEY
Keywords- Gas detection, Arduino, Buzzer, LCD, Alert,
etc Belkacem Khadi et al [1] has proposed in an
endeavor on the concept of multi-robot intellect referred to
1. INTRODUCTION as SR(swarm robotics), inspired by the nature and observing
groups such as groups of ants, flock of birds, schools of fish
The Internet of Things strives to make life easier and faster and groups of bees. In SR, if a group of robots performs a
by automating all the small tasks involved in human life. task in an intellectual approach then they are referred to as
Today technological advances like IOT are making SI (swarm intelligence), it is a passive networking system in
everything smart. Since IoT is very beneficial in automating which every distinct bot of the group interacts with one
tasks, the advantage of IoT can also help improve convenient another and with the outside ambiance. The field deals with
security methods. the design of a large number of simple robots, their physical
properties, and their controlling behavior. There are various
Security performs a significant role at the same time as simulation platforms that are used to test the structure and
constructing homes, buildings, and industries in addition to algorithm of swarm robots. R. Imtiaz, et al [2] has proposed a
towns. The enlarged focus of certain gases within the work on implementing two different kinds of robots, which
surroundings are maybe exceptionally unsafe. Nowadays, included an explorer robot and a carrier robot. In this
everyone wishes for a facility that reduces effort and time as configuration there are four robots, one master robot or
well as expects their work to be as easy as possible. The explorer robot which works as the leader and three slave
prime aim of the paper is to alert people with the help of a robots or carrier robots. The explorer robot travels the
buzzer and detect the hazardous gases present in the air. entire path set towards target and gives the instruction
This is done using an MQ5 sensor. The MQ5 sensor is used to about the path to the carrier robots. Carrier robots follow the
detect gas leakages for various applications. The MQ5 sensor path according to master robots command. Thus, any work
detects the concentration levels of the gases and outputs an can be completed in less time and much efficiently. Zigbee
analog value which can be later converted into a digital module is used as the communication device for the
© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 343
intercommunication between robots, and is used in each Relay

robots.
A relay is an electrical switch that is used to control all other
Mohd. Daneel Khan, et al [3] has proposed work on electronic devices by using the electromagnetic mechanical
the application of S-bots in disaster management. Swarm toggle
intelligence provides a collective work to perform a certain
task which is much accurate and efficient. During
earthquakes when large number of buildings fall down and
lots of people are trapped under the building swarm it is
impossible by humans to reach that places and rescue
people. In this situation swarm of learning, birds can be used
to provide the information about the areas and give data to
make accurate decisions. It is also used for locating a safe
location and to give the exact path of the location. Abhishek,
P. Bharath, et al [4] gas leakage can be detected by using the Figure.2
gas sensor, when a small amount of gas is brought near the
sensor it starts alerting the user about the leakage of the gas LCD display
using iot. P. M. Vidya, S. Abinaya, G. G. Rajeswari, and N.
Guna, et al [5] has suggested that the leakage of the gas can The LCD is used to display the message "Gas in Zone
be detected by using various gas sensors Detected" on the first coded display in the program to
indicate hazards. All messages, data, and
Kavitha B et al [6] has designed an alarming system commands displayed on the LCD are LCD registers. its
for the industry based on the huge amount of leak of gas operating power supply ranges from +5.0 V or +3.0 V.
from the container. Kalpesh Gupta et al [7] has designed an
automatic window opening system when the gas is leaked at
a certain amount of volume continuously.
3. SYSTEM OVERVIEW
Hardware Requirements
Arduino Uno
Figure.3
The Arduino UNO is an open-source microcontroller board
based on Microchip's ATmega328P microcontroller and ESP8266 WIFI Module
developed by Arduino. cc. its main aim is to make electronics
ESP8266 is a low-cost WIFI microchip with a
to be as easy as possible. It uses different microcontrollers,
complete TCP/IP stack and microcontroller functionality
containing several input and output pins. It provides an
from EspressifSystems.
integrated development environment (IDE). Arduino
contains several numbers of parts and integrated interfaces
in a particular circuit board.
Figure.4
MQ-5 Sensor
Figure.1 The Grove Gas Sensor Module (MQ5) is useful for

gas leak detection (home and industrial). Suitable for the
detection of H2, LPG, CH4, CO and alcohol. High sensitivity
and fast response time allows the measurements to be made
quickly as possible. The sensitivity of the sensor can be
adjusted with a potentiometer.
Figure 7 depicts the architectural block diagram of the

proposed gas leakage detection system
Figure.5
Buzzer
A buzzer or buzzer is a mechanical, electromechanical or

piezoelectric (piezo for short) audible signalling device.
Common uses of buzzers and beeps include alarm devices,
timers, and user acknowledgment of user input such as
mouse clicks and keys.
Figure 7. Architecture of gas leakage detection system.
Figure.6
Software Requirements
A) Arduino IDE
B) Language C++
4. SYSTEM OPERATION
In this proposed system, the gas leakage is detected by MQ 5
sensor which is interfaced by Arduino uno, when the gas
leakage is detected through the MQ 5 sensor the lights gets
off with buzzer alarm and display alert message in LCD
display. Further, working of this proposed system is as
below
This gas sensor is used to detect hazardous gases in industry.

This sensor can detect the following gases:
• Carbon Monoxide CO 1 - 1000ppm,
• Ethanol C2H6OH 10 - 500ppm,
• Hydrogen H2 1 - 1000ppm,
• Ammonia NH3 1 - 500ppm,
• Methane CH4 >1000ppm
When detected, it will automatically turn on. The detected

values are sent as inputs to the main component, and the
buzzer will make a sound. Users can view sensor status and
take necessary actions accordingly.
Figure 8. Flow diagram of the activities of the gas leakage
detection system
5. EXPERIMENTAL RESULTS Analog value of the gas sensor
The prototype of the project system is shown in Figures 9

and 10. This proposed system is developed to detect and
monitor gas leakage. when a small amount of gas is brought
near the MQ5 sensor, it displays the message in LCD i.e “GAS
LEAKAGE DETECTED” and the gas value is also shown at the
time of the detection of the leakage of the gas and the system
monitors the LPG level and displays the message with the
gas value. The developed system will be helpful to the
customer. It helps them to upgrade their safety norms and
aids to prevent the major disaster. It also helps in protecting
life and property from apparent accidents. The primary aim
of this proposed system is therefore to detect the gas leakage
which is detected though gas sensor and user should be
notified to prevent injure or outburst via buzzer and alert When small amount of gas is brought closer to the MQ5
the user. sensor, the analog value spikes above 1000.
It helps them to upgrade their safety norms and aids to 6. CONCLUSION

prevent the major disaster. It also helps in protecting life and
property from apparent accidents. The primary aim of this This paper aims to ensure the monitoring and detection of
proposed system is therefore to detect the gas leakage which hazardous gases to meet safety standards. The system
is detected though gas sensor and user should be notified to detects gas in the atmosphere and continuously updates and
prevent injure or outburst via buzzer and alert the user. displays the gas values which can be easily viewed by the
user via a LCD screen. The system is quite responsive and
can stop crisis situations more quickly than manual methods.
The system alerts and responds to leakage by alerting the
user. In the future, this system will be packed with advanced
features to provide users with greater security and
relaxation. The popularity of handheld devices has advanced
the field of smart gas sensors, greatly expanding the range.
The need to ensure workplace safety is expected to be the
main driver of the market in the upcoming years.
REFERENCES
[1] Belkacem Khaldi, Foudil Cherif, “An Overview of Swarm
Robotics , Swarm Intelligence Applied to Multi-robotics,”
Figure.9 Prototype model of the gas leakage detection
International Journal of Computer Applications (0975 –
system
8887) Vol 126 – No.2, September 2015, India.
[2] R. Imtiaz, B. Ashokkumar, M. Danny Frazer

“Implementation Of Load Sharing Using Swarm Robotics,”
International Research Journal of Engineering and
Technology, Volume: 03 Issue: 03, (pages: 1855 – 1862) |
March-2016 , India.
[3] Mohd. Daneel Khan, Krantee Jamdaade; ”Application of

Swarm Intelligence in Disaster Management”, International
Journal on Future Revolution in Computer Science &
Communication Engineering Volume: 4 Issue: 6 | pp 77-84
Figure.10 Gas value is detected by the sensor and the June - 2018 , India
buzzer is switched on
[4] Abhishek, P. Bharath“Automation of lpg cylinder booking
and leakage monitoring system-IJCRD”
[5]P. M. Vidya, S. Abinaya, G. G. Rajeswari, and N.

Guna“Automatic lpg leakage detection and hazard
prevention for homesecurity-National Conference on VLSI,
Embedded and Communication & Networks”
[6] Kavitha B C,VallikannuR“IOT Based Intelligent Industry

MonitoringSystem –SPIN 2019”
[7] Kalpesh Gupta, Gokul Krishna G and Anjali T ”An IoT

Based System for Domestic AirQuality Monitoring and
Cooking Gas Leak Detection for A Safer Home- International
Conference on Communication and Signal Processing, 2020”
Volume: 04 Issue: 09 | Sep -2017 www.irjet.net p-ISSN: 2395-0072
Smart Home for Paralyzed Aid

Debojyoti Seth1, Debashis Chakraborty2, Debosruti Ghosh3
1,2 Department of Electronics and Communication Engineering, Future Institute of Engineering and Management,
Kolkata - 700150, West Bengal, India
3 Department of Computer Science and Engineering, University of Engineering and Management,
Kolkata - 700160, West Bengal, India

---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - The paper proposes a scope of aiding a paralyzed measured from the lobes explicitly [3]. Beta is generally
person left alone at home. Primarily focusing on obtained from deep sleep and often if the patient is suffering
Electroencephalogram data of 25 paralyzed people, the study fatigue and has not taken any food for last few hours. Alpha
of changes in brain signals while they feel hungry, thirsty, and beta are closely related to change of emotions apart
sleepy, mentally excited or stressed is conducted. On from stress and relaxed mind states. A stronger correlation
continuously monitoring and systematically analyzing the is observed in case of neural activities, whereas a weaker
brain signals for a physically challenged person, multiple one when asleep.
modules to meet the basic necessities are developed and tested Implications of IOT on healthcare was first observed in
upon. The Electroencephalogram data is preprocessed and 2009 where for a P300 based BCI system [4], EEG
classified based on neuro-fuzzy hybrids. The logical decision predictions are justified by IOT realizations in hardware
making is performed by an Internet of Things based platform. modules. But the activities were solely dependent on flash
All the control modules are fully automated and hardware is incidence, accuracy drops down to 54% from 79%, on
driven by a self-learning fuzzy control machine. Results of reducing the number of flashes from 8 to 2. Many recent
performed experimentations proved to be really promising literatures worked on designing intelligent wheelchairs,
and reached an overall accuracy of 89.73% for automating the some based on visual stimuli [5] or on overall perception
aiding units to fulfill basic needs of a paralyzed person and we through various sensory organs [6]; but an absolute need-
are looking forward to extend the research to design a smart based automated module where classified input from EEG
hospital paradigm. signals triggers multimodule IOT paradigm for various
everyday functions, was never discussed before.
Key Words: Electroencephalogram, kNN classifiers, Self- Furthermore, the disadvantages of fuzzy learning can be
learning Fuzzy, Internet of Things, paralysed, Basic need ruled off by a sliding approach for the best possible
automation optimized output with a return loop in order to learn from
previous state errors. Informally, the model designed is
1. INTRODUCTION named as Self Learning Fuzzy Sliding Mode Control
(SLFSMC) and is used for driving few hardware units for two
In the last century, engineering advances modified the support modules.
notion of healthcare. Life expectancies have increased The rest of the paper is organized as follows. Firstly, the
approximately by 34% in the last few decades [1]. The major experimentations and EEG Signal Processing is described.
problem encountered when studies related to Brain- Next the research methodology uncovers the theories behind
Computer Interfaces (BCI) are emerging is the complexity of driving the support modules for a paralysed person which is
nature of Electroencephalogram (EEG) signals. Several followed by the implications of each and every module.
predictive models are developing in the last few decades, but
mostly they are domain specific or developed for a single 2. EXPERIMENTAIONS AND SIGNAL PROCESSING
objective [2]. For example, most implications of Internet of
Things (IOT) are only for communication protocols or Fuzzy The EEG signal is acquired, pre-amplified and the artifacts
logic found a major application in domains of speech are removed by filtering. The spectral analysis derives the
recognitions for the last few decades. peak power and thus provides an overall estimate of
For not involving any pain of the test-subject and higher dominant band of activity. Next kNN based classifiers
dynamicity for research, non-invasive processes like EEG provide a scaled band activities and label frequency
(with the sole aim of studying electrical activities in brain) is parameters (alpha, beta, delta and so on) for transferring the
much more preferred than invasive ones. Previous studies possible priority for analysis in the IOT platform. Fig. 1
described our cerebral cavity divided into four lobes, namely describes the basic principles of EEG Signal Processing.
frontal, parietal, temporal and occipital; and spectral sub
band frequencies generally includes ((alpha (8-13 Hz), beta
(14-30 Hz), delta (0.5-4 Hz) and theta (4-8 Hz)) which are
Preamplific Filtering infer. Similarly, if one electrode is placed a lot away from
Acquisition
another, the loss of medical data may result in some
ation
necessary peak powers and resultant symptoms
undiagnosed. Considering all these complexities, electrodes
Symptomati Classificati Spectral are placed in the scalp obeying 10-20 International System
c Inference on Analysis for a 21-electrode setup as shown in Fig.2.
2.3 Signal Preprocessing

Fig -1: EEG Signal Processing
One of the basic reasons of the randomness of the
neurophysiological signals is the noise content in it. Firstly,
2.1 Subject Selection Criteria when the signals in microvolts are amplified, the artifacts get
amplified as well [9]. The reasons of artifacts are many and
Twenty-five paralyzed people (14 male and 12 female) of include, though not limited to: all physical constraints, room
age limit between 55 and 70 years were selected for the temperature, other signal interferences, thought process,
study with their formal consent. They were all right-handed calorie content, last night sleep quality, muscular stress,
individuals and past pathologies assured normal working of psycho-physiological activity status of subjects etc. [10] The
their brain. Subjects were included only if they did not have amplified signals are hence to be vividly processed before
any past neurological histories, had normal eyesight and no subjecting to spectral analysis. Next, it is passed through
hearing ailment and not under any medications for repetitive filtration using band pass Chebyshev filters. If the
hypertension or diabetes. Standardization of subject is order of filter is kept low for a simpler calculation, the
treated as an important criterion for EEG analysis, as past number of iterations increase. Or if the filter order is
researches revealed remarkable change of perceptions and increased along with appropriate windowing expected
effects on brain signals based on changing emotional indices accuracy is obtained in a few cycles (the count stays typically
(often triggered by age or certain diseases). around or below five). Order and rate of filtration solely
depends on the nature of the raw signal acquired, and hence
2.2 Experimental Basics and Setup treated as a dynamic quantity. Type I Chebyshev filters are
used with the sole objective of quicker role-off than Type II
[11]. Parametric method for spectral density calculation
could be beneficial for such signal due to certain advantages
over non-parametric methods like avoiding side-lobe
leakages [12]. Based on past researches the Burg method
and the Yule-Walker method are found to be the most
efficient for neurophysiological signals’ spectra analysis [13].
Both are adapted for the smart home module; one may get
switched to another in case of change in specific need for a
particular module of aid.
2.4 Filtered Data Classification
The classification can be performed based on any

standard neural classifier. Support vector machine or
Fig -2: Position of electrode on scalp based on
extreme vector machine based classifications, promise a
International 10-20 System higher accuracy in cost of time [14]. Gradient descent
approach is often accompanied with standard back-
For EEG testing, signals are collected mainly in microvolts propagation paradigms [15], which further slows down the
range, though it may shoot up to a few millivolts scale for rate of computation. Fuzzy system based classification
shocked or surprised mental states [7] of a paralyzed person. gained popularity in the past few decades; though it is often
To discuss the scalp placement, as it is well known, the most found to mess up with the border overlap. For example, an
complex part of a human body is the brain itself. If the EEG left frontal alpha may at some frequency points merge
cathodes are placed too close to each other, there are with beta or delta contributions, resulting scarcity of
chances of least amplitude production due to minimum
potential difference [8]; which in turn results a massive drop
on spectral analysis and makes the study too convergent to
information extraction [16]. An emerging domain of

researchers found neuro-fuzzy hybrids quite useful in the
context of classifying EEG signals. It helps in a discrete
analysis of shorter time lapses. The kNN classifier is found
to work good for our classification purpose. In kNN
classification, the output is a class member. An object is
classified by a majority vote of its neighbours, and
weightage of voting increases with decrease in distance
between a given neighbour and target. Finally, the object is
assigned to the class, most common among k nearest
neighbours. k is a positive integer, typically of smaller
values. Given the set of classified data, kNN determines the
classification of input based on the class labels of k closest
neighbour(s) in the classified set. Major two steps of the
algorithm include: Fig -3: Fuzzy kNN classifier output on analyzing EEG for a
i) Find k nearest neighbours of the input x . 65 years old man eager to take an evening ride
ii) Calculate the class membership of x using Eq.1.
Fig.3. describes the state of a mind for a sixty-five years
k
old man who is eager to take an evening ride in the
 cix j d j countryside after spending four hours after his lunch. The
j1 boxes denote the alpha contributes, circles the beta
 ci ( x)  (1) attributes and crosses the delta attributes in Fig.2. For
k
simpler visualization three lobes are taken under
 dj
consideration for the above read: frontal (in red), parietal (in
j1
Particularly, the parameter m determines how heavily the blue) and occipital (in green). The classifier here indicates
distance is weighted while calculating the class membership. higher occurrences of beta over alpha and delta in frontal
As m decreases towards infinity, the term 1 / d approaches (68%) and parietal lobes (78%) which indicates a wish. The
alpha-beta pair in all the three lobes are normally deviated
unity regardless of distance. The term mentioned as d j is which indicates normal health status and no specific priority
defined as: task is generated. The task of identifying the expected
strands of frontal activities in occipital lobe is on the
1 intelligent module driven by IOT platform, as it is found to be
d j  2
(2) beyond the scope of kNN classifier.
x xj m1
3. RESEARCH METHODOLOGY
Interestingly, x1 , x 2 ,..., x k denotes k nearest neighbors of
x . Consequently, the neighbors x j are more evenly

weighted. As m decreases towards 1, however, the closer
neighbors are weighted far more heavily than those further
away. This also has the effect of reducing the number of
neighbors that contribute to the membership value of the
input data point. The objective of band priority deduction is
often attained solely by kNN classifiers for single priority
from user end [17]. For example, on waking up after a 6
hours’ sleep the person on study seeks assistance to drink
water. But most of the time, in a real scenario there are
multiple needs in conscious or subconscious mind, and the
objective of the classifier is to determine the prioritized and
really necessary needs and generate queries for IOT platform Fig -4: Three staged Task Objective Realization
to work on them.
Three distinct stages are followed for the research as

depicted in Fig. 4. At the early stage EEG data is collected,
processed and some initial predictions are obtained from Input Devices Output Devices
Sensors, Meters, Control
kNN classifiers. Next the IOT platform is triggered by the EEG out Modules
classified data. They have the sole responsibility of
maintaining database, communicating each layer and
decision making for identifying the best possible necessity of
the paralyzed person. The EEG data updates the buffer at an
interval of three milliseconds (peak-hours) and fifteen Analog to Digital Analog to Digital
seconds (for sleep mode). The self-learning fuzzy sliding Convertor Convertor
mode control (SLFSMC) unit is designed mainly for end-user Communication Layer (Wireless/Wired/LAN/Internet/Open Interfacing)
device controls associated with temperature control and
analysis on emotional indices. The IOT units update the
SLFSMC to operate upon and the SLFSMC acknowledges the
activity and share device status to the upper layer when Memory
Buffer Analyzer Decision
asked for routine checks or updating new orders. Memory
Maker Buffer
Control
3.1 IOT Based Decision Making Logic
The basic principle of IOT is the integration of multiple

technologies shared by a common internet. IOT is basically
one of the recent emerging technologies which is assumed to Fig -5: IOT Module for Aiding Paralyzed Person
be the pioneer in the next generation of internet network
[18]. The three effective layers of control on IOT are Based on functionality, the IOT modules can be broadly
Perception, Network and Communication layer. The major classified into three layers, Perception Layer, Network Layer
key technologies of IOT primarily include: and Application Layer. The data acquisition units of
Perception Layer play the role of identifying inputs/outputs,
a) Internet technology acquiring/releasing and updating buffer at a regular interval.
b) Sensor Network technology Another functionality of Perception Layer is to provide
access by transmitting the data from sublayer to global
c) Wireless Communication technology
object-conjunction network. Network Layer supports
d) Embedded technology
communication by best possible unit; selected based upon
The model developed can also be classified based on the priority and privacy of the data passing by. The same also aid
task to be performed: Information Unit, Control Unit, partially to information integration: a method of clustering
Decision Unit and Physical Unit. The physical unit consists of similar data by generating a label for each cluster for easy
the input layer composed of all measuring units and the identification at user end. Finally, the application layer
output layer which operates on the aiding units to the performs some selection, optimization and sorting
hardware level. The Information unit consists of all the algorithms to execute the most required action for aiding the
buffers and can sequence clusters of data over time. The paralyzed person.
Control unit in the processor has a task quite similar to the
communication layer but the only difference is, the earlier 4. SUPPORT MODULES
has the intelligence to choose priority query over another
and can also compare data with the buffer to update Some absolute need-based modules are developed for
analyzer. The analyzer activates the decision unit. The aiding a paralysed person to feed on time, letting have a nice
sequencing and/or sorting of current input with recent past sleep, offering a soothing room temperature or controlling
states is performed in the decision unit in order to identify emotional thrushes. All the support modules are completely
the major contributive necessity of the paralyzed person left automated and associated with a tertiary emergency module
alone at home. Lastly, control unit marks the output from a updating status of unexpected conditions like rapid health
cluster of possible necessities in a reverse path from the deterioration, natural calamities etc. to selected family
decision unit, once the latter completes operation. members, nearest police station, hospital for ambulance
support or fire station (triggered only for temperature
control).
4.1 Feeding Module 4.2 Sleep Control Module

The hunger and thirst are one of the most basic needs.
Not only detecting hunger or thirst psycho-physiologically, Broadly sleeps can be classified of two types, Rapid Eye
but measuring required calorie content and holding a control Movement (REM) and Non-REM sleep [20]. REM sleep
on protein, carbohydrate and vitamin supplies are some of generally implies stress, anxiety or more adversely, some
the basic needs. It also becomes necessary to differentiate latent physical disorders yet to show symptoms [21]. REM
between water-requirement or bulk roughage carbohydrate can simply occur as well due to over-scheduled last twenty-
requirement for person under control [19]. Feeding to be four hours accompanied with different streams of multiple
performed artificially for critical patients, mainly based on dreams. For a normal eight hours sleep, in early few hours
three types: REM is quite natural though, before diving in deep sleep.
i) Ryle’s tube feeding and saline support Non-REM ones are deep sleeps mainly accompanied by
ii) Normal feeding with saline support dreams.
iii) Assistive table; no feeding, drinking support Whether sleep is REM or Non-REM can easily be
For normal feeding in cases ii) and iii) a rotational assistive determined by the singularity of alpha in all the electrodes,
table is designed with preassigned diet but flexible even if not, by synchronizing left and right parietal lobes
quantities for intake; typically based upon last meal-time alpha. For continuous REM, based on other health status,
and rate of digestion. The patient when mounted to the chair certain controlled sedatives might get injected in. Whereas,
attached to the table, the table rotates to the required bowl, for continuous Non-REM (typically more than twelve hours,
lifts it up to a given height (based on height of paralysed without any sedatives or anesthetics in two days’ history) on
person) and thus aids digestion. fearing slow approach to coma certain spikes (external
activation) might be provided to sharpen pulsed delta
The major advantage obtained for this module is, effect of
concern (external stimulation to facial activity vortex) and
digestive enzymes reaches much faster to the central
wake patient up to rule out other possibilities.
nervous system to get detected by EEG for paralyzed people
than normal ones. For example, fasting for a paralyzed one,
results in wider pulse rate with sudden beta domination and 4.3 Temperature Control Module
alpha power loss, if three or more hours are exceeded than
regular meal time; but this effect gains visual approximately The realization of feeling heated or cold varies from one
fourteen to sixteen hours gap for a normal person. Observing person to another based-on immunity and several other
thirsty nature out of hunger is quite a big challenge, as both factors [22]. A typical temperature sensor is hence
returns exactly similar neuronal functions through accompanied with EEG band analysis. For temperature
noninvasive testing like EEG. The method of sorting out the control, alpha-beta pair is compared between actual
need is depicted in Fig.6; it is a visualization of a forty years temperature and temperature realization. Assuming the
old female volunteers’ data with non-functional left limbs
after a short afternoon nap. brain contribute is  and  is the real temperature, the
effective realization can be computed as:
BUFFER Patient Awake; Normal Thought Rate
Left frontal active; beta taking over alpha
k
   i  2   2 ∀ i  1,2,...k (3)
i 1
LOGIC Hunger/Thirst constant (T) activated
COMMAND: Analyzer update T to execute A fan, an air-conditioner and a heater are the three
targets device to be controlled by the fuzzy module. The
discrete time temperature control system is described by:
ANALYZER COMPUTE: Calorie/Minerals/Fluid count
INFERENCE: Meal…3.32 hours back
 b(Ts ) 
Water Intake: 6.41 Hour back
y(k  1)  a(Ts ) y(k )    (k )
0.5 y(k )
1 e (4)
DECISION Meal Necessity: 13.52%  [1  a(T )] y
Water Intake Necessity: 86.48% s 0
EXECUTE COMMAND: 
b(Ts ) 
Initiate assistance to drink water a(Ts )  e where b(T s )  ( )(1  e Ts ) (5)

Fig -6: IOT Based Priority Selection of Feeding Module
5. CONCLUSION
where index for discrete time sequencing is k, Ts the
sampling period, and y0 is the initial temperature state. The β Based on a continuous 96 hours study, an accuracy of
and α constraints are adapted from surrounding contributes 89.73% is attained in power saver mode during night-sleep
that impact weather and the system input and output are time, which further increase by about 2% if all units are kept
represented by  (k ) and y(k) respectively. A self-learning operative together. Multiple researches are going on
nowadays, from modernizing EEG devices to generating new
fuzzy algorithm helps in scaling the constraints of Eq. 5. classifiers. But the concept discussed here for automating
Firstly,  r (k ) is defined as the reference threshold
the basic needs of a paralyzed person by an IOT platform for
temperature whereas  (k ) is the output temperature from decision making and SLFSMC for device controlling, is quite a
Eq. 3. Assuming the temperature tracking error to be  , the new one. Fuzzy techniques proposed here has the ability of
change of error is given by: correcting own previous state-errors for a faster
convergence. Further researches can be extended to
automate wheelchair controls and finally implementing IOT
(6) for controlling HIS to design a Smart Hospital paradigm.
  (k )  (k )  (k  1)
As an input of the fuzzy inference rules, a sliding surface is REFERENCES

chosen as: [1] Samek, Wojciech, Frank C. Meinecke, and Klaus-Robert
Müller. "Transferring subspaces between subjects in brain--
s(k )    (k )    (k ) , σ > 0 (7)
computer interfacing." IEEE Transactions on
Biomedical Engineering 60.8 (2013): 2289-2298.
The self-learning fuzzy sliding-mode control rule base can
be defined as: [2] Sanei, Saeid, and Jonathon A. Chambers. EEG signal
processing. John Wiley & Sons, 2013.
i
Rule i :IF s  Fs , THEN u  ri, i  1(1)n (8)
[3] Borghini, Gianluca, et al. "Measuring neurophysiological
signals in aircraft pilots and car drivers for the assessment of
In Eq. (14) F denotes the fuzzy set of s and ri is a singleton mental workload, fatigue and drowsiness." Neuroscience &
function. The defuzzification is performed by Center of Biobehavioral Reviews 44 (2014): 58-75.
Gravity method.
[4] Grandy, Thomas H., et al. "On the estimation of brain
4.4 Emotion Control Module signal entropy from sparse neuroimaging data." Scientific
reports 6 (2016).
Emotion Control is obtained partly by SLFSMC logic and
partly by kNN classifier’s interpretation from various EEG [5] Tseng, Yuan-Yun, et al. "Concurrent delivery of
montages. The nature of EEG parameters, related inference carmustine, irinotecan, and cisplatin to the cerebral cavity
and possible music playing is depicted in below Table 1. A
using biodegradable nanofibers: in vitro and in vivo
common example is, when one is stressed, switching on a
meditative music can offer a sound sleep. studies." Colloids and Surfaces B: Biointerfaces 134: 254-
261 (2015).
Table -1: Fuzzy Analysis of Emotional Indices for Music
[6] Babiloni, Claudio, et al. "Alpha, beta and gamma
IF THEN EXAMPLE electrocorticographic rhythms in somatosensory, motor,
Inference Types of Music on Play
Alpha Beta Theta Delta (Feeling) premotor and prefrontal cortical areas differ in movement
0.49 0.68 0.12 0.02 Pleasant Happiness yielder execution and observation in humans." Clinical
Neurophysiology 127.1: 641-654 (2016).
0.25 0.01 Unpleasant Sharp beats; negative
0.33 0.70
emotions yielder
[7] Grønli, Janne, et al. "Beta EEG reflects sensory processing
0.31 0.18
0.32 0.04
Soothing;
Slow, classical; in active wakefulness and homeostatic sleep drive in quiet
monotonic
Meditative
instrumental wakefulness." Journal of sleep research (2016).
0.84 0.02
0.06 0.05
Energetic;
Massive positivity [8] Herrmann, Christoph S., et al. "EEG oscillations: from
yielder; generally
Boost of
Happiness relates past good feels correlation to causality." International Journal of
Psychophysiology 103: 12-21 (2016).
[9] Paris, Arnaud, et al. "Using Hidden Semi-Markov Model [20] Singla, Rajesh, Arun Khosla, and Rameshwar Jha.
for learning behavior in smarthomes." Automation "Influence of stimuli colour in SSVEP-based BCI wheelchair
Science and Engineering (CASE), 2015 IEEE International control using support vector machines." Journal of medical
Conference on. IEEE, 2015. engineering & technology 38.3 125-134 (2014).
[21] Müller, S. M. T., et al. "Robotic wheelchair
[10] Park, Seok-Hwan, et al. "Fronthaul compression for commanded by people with disabilities using low/high-
cloud radio access networks: Signal processing advances frequency ssvep- based BCI." World Congress on
inspired by network information theory." IEEE Signal Medical Physics and Biomedical Engineering, June 7-
Processing Magazine 31.6: 69-79 (2014). 12, 2015, Toronto, Canada. Springer International
Publishing, 2015.
[11] Lee, Seunghyun, Rui Zhang, and Kaibin Huang.
"Opportunistic wireless energy harvesting in cognitive radio [22] Khalghani, Mohammad Reza, et al. "A self-tuning
networks." IEEE Transactions on Wireless load frequency control strategy for microgrids: Human
Communications 12.9: 4788-4799 (2013). brain emotional learning." International Journal of
Electrical Power & Energy Systems 75 311-319 (2016).
[12] Cooper, Jesse, et al. "Efficacy of an automated ultraviolet
C device in a shared hospital bathroom." American journal
BIOGRAPHIES
of infection control 44.12: 1692-1694 (2016). Debojyoti Seth has completed his B.
Tech degree from Future Institute of
[13] Blasco, Rubén, et al. "A smart kitchen for ambient Engineering and Management in
assisted living." Sensors 14.1: 1629-1653 (2014). 2017. His research interests include
but not limited to Bioinformatics,
[14] Ghayvat, Hemant, et al. "WSN-and IOT-based smart Probabilistic Modelling, Man
homes and their extension to smart buildings." Sensors Machine Interfaces, and Medical
15.5: Automations.
10350-10379 (2015).
Debashis Chakraborty is an
[15] Robinson, John, and Henry Amirtharaj EC. "MAGDM Associate Professor of Future
problems with correlation coefficient of Triangular Fuzzy Institute of Engineering and
Management. His research interests
IFS." Theoretical and Practical Advancements for
are Statistical Signal Processing,
Fuzzy System Integration. IGI Global, 154-192 (2017). Wireless and Satellite Comm.,
Spectrum analysis, telemedicine etc.
[16] Li, Ni, et al. "A new methodology to support group
decision-making for IoT-based emergency response Debosruti Ghosh is currently a
systems." Information systems frontiers 16.5: 953- second-year B. Tech student at
977 (2014). University of Engineering and
Management. Her research interests
[17] Chi, Nai-Ching, and George Demiris. "A systematic primarily include Surgical Robotics,
review of telehealth tools and interventions to support Bioinformatics, DNA Computing,
Brain Computer Interfaces,
family caregivers." Journal of telemedicine and telecare
computational biology.
21.1:
37-44 (2015).
[18] Guger, Christoph, et al. "How many people are able to

control a P300-based brain–computer interface
(BCI)?." Neuroscience letters 462.1: 94-98 (2009).
[19] Lee, Po-Lei, et al. "Brain Computer Interface Based on

the Flash Onset and Offset Visual Evoked Potentials." Recent
Advances in Brain-Computer Interface Systems. InTech,
2011.
Application of Data Analytics to Improve Patient Care: A Systematic

Review
Mehmood Ali Mohammed1, Murtuza Ali Mohammed2, Vazeer Ali Mohammed3.
1Dept. of Computer and Information Systems, University of the Cumberlands, Williamsburg, KY USA
2Dept. of Computer and Information Systems, Lewis University, Romeoville, IL USA
3Dept. Of Computer and Information Systems, Lewis University, Romeoville, IL USA
---------------------------------------------------------------------------***--------------------------------------------------------------------
Abstract
Data analytics continuous to drive positive change by enhancing internal operations and external relationships
between organizations and communities. Data analytics is the use of scientific and mathematical techniques to derive meaning
from factual information thus gain better insights. Patient care encompasses the healthcare services that are rendered for the
benefit of patients. It is important to note that patient care systems around the world have focused on improving the health
and experience of the individual by leveraging various inputs such as modern technology. This study sought to conduct a
systematic review the application of data analytics towards improving patient care.
The specific objectives were to investigate the concept of data analytics and its applications, understand the
application of data analytics in healthcare, determine the implications of data analytics in improving patient care, and establish
the challenges and opportunities of data analytics towards enhancing patient care. The study employed a number of
theoretical frameworks as foundation for understanding the relationship between data analytics and patient care. These
theories were the Magical Thinking Theory, the Lightweight Theory, and the Classical Mathematics Theory. The study
employed systematic reviews that collate the outcomes of research studies thus obtain a collective estimate of the intervention
effect.
The eligibility criteria for the research studies include the study population, time, type of intervention, study variables,
quality of the research methodology and linguistic and cultural range. The findings from the systematic review were that data
analytics had a significant impact on the healthcare sector especially when it comes improving patient care. The study rejected
the null hypothesis and thus accepted the alternative hypothesis that stated the importance of data analytics in improving
patient care.
Keywords: Data Analytics, Patient Health, Prisma, Data Cleaning, Interpretation of Data, Digital Health.
1.0. Introduction
Data analytics is regarded as a technological revolution that continuous to drive positive change across organizations
by enhancing both internal operations and external relationships with individuals and communities. Data can be defined as
information in terms of statistics or measurements that can be used as a basis and foundations for calculations, reasoning and
discussion (Shekarian, Ramirez, & Khuntia, 2020). There are two types of data that include qualitative data that is usually
expressed in a narrative or verbal form and quantitative data that is presented in numerical form. Data analytics refers to the
utilization of various scientific and mathematical methods to derive meaning from factual information thus gain better insights
about objects, people, processes or organizational units (IFAC, 2018). It is important to incorporate data analytics as it reduces
costs, enables faster decision making and minimizes risks especially within healthcare. Below is an image illustrating the
process of data analytics:
Problem Intepretation of
Data Analysis
Identification Data
Identification of Transforming the Evaluation of

Data Sources Data Findings
Presentation of
Data Selection Data Cleaning
Findings
Fig 1: Process Map for Data Analytics
On the other hand, patient care refers to the services that are rendered by healthcare professionals including doctors,
nurses, and management personnel for the benefit of patients (Yorke, 2017). The concept has been widely discussed in the
field of healthcare and research with the aim of highlighting the most appropriate definition. In some instances, the provision
of patient-centered care has been fronted as a more elaborate explanation of services rendered to patients with the
terminology describing patient care as the provision of health services that improves the health and experiences of the
individual. Patient care systems around the world have focused on improving the health and experience of the individual by
leveraging various inputs such as modern technology (Lewis, 2019). Therefore, it is important to understand the different
technologies that can be applied in the field of healthcare thus improve patient care and thus achieve the most sought-after
patient centered healthcare system.
2.0. Background
According to Shekarian, Ramirez and Khuntia (2020), data analytics as a process of transforming, managing, and
modeling data towards deriving meaningful information plays a critical role in improving hospital performance. The study
revealed the sparse nature of evidence related to the interplay between data analytics and operations in healthcare. However,
the research sought to close the gap by evaluating the impact of data analytics in health and clinical operations. The findings
were that the relationship between data analytics and hospital performance tend to be highly complex hence the lack of
similar evidence. A study by Wang and Alexander (2019) revealed that big data analytics has the potential to improve patient
outcomes while advancing and personalizing care. The research also highlighted the capability of data analytics enhance the
performance of healthcare systems by improving relationships between healthcare providers and patients as well as reducing
medical spending.
Furthermore, Dash et al., (2019) investigated the conceptualization of big data in healthcare with regards to overall
management, analysis, and the future possibilities. The study sought to understand the impact of big data and its analytics in
the field health and clinical medicine by assessing the role played by the concept in both healthcare operations and research.
The authors revealed that big data analytics leveraged the perceived gap that exists within structured and unstructured
sources of data thus enabling a shift to the problematic integrated data system that affects the provision of quality healthcare.
Gemson and Durga (2015) conducted a study to understand the background and techniques of big data analytics that have
been applied in healthcare. The research reviewed the application of big data analytics in health systems as well as the
platforms, algorithms, pros and cons of the technological concept while providing a discussion of the future areas of interest.
2.1. Research Objectives
The study comprised of both general and specific objectives. The general objective was to review the application of
data analytics towards improving patient care.
The specific objectives were:
i. To investigate the concept of data analytics and its applications.

ii. To understand the application of data analytics in healthcare.
iii. To determine the implications of data analytics in improving patient care.
iv. To establish the challenges and opportunities of data analytics towards enhancing patient care.
2.2. Hypothesis
2.3. The study stated both a null and alternative hypothesis that guided the implantation of the systematic review. The
hypothesis for the study were:
H0 – There is no significant impact of data analytics towards improving patient care.

HI – There is a significant impact of data analytics towards improving patient care.
2.4. Literature Review
This section discussed the theoretical frameworks that provide the basis for understanding the link between data
analytics and patient care. The study sought to review the application of data analytics towards improving patient care. The
theories identified as a foundation for understanding the relationship between data analytics and patient care included the
Magical Thinking Theory, the Lightweight Theory, and the Classical Mathematics Theory. The Magical Thinking Theory
postulates that we are inclined towards seeking and interpreting connections between the events that take place around us
along with the disinclination to review our beliefs following deeper observation (Diaconis, 2016). The theoretical framework
illustrates that in some instances, a single manifestation may be believed to be a sign and that a given ritual offer the method
or technique to understanding. Therefore, the underlying belief tends to persist despite the existence of factual circumstance.
Secondly, the Lightweight Theory of data analytics emphasizes on making predictions based on the consolidation and
acceptance of theoretical frameworks (Elragal & Klischewski , 2017). The theory is related to the domain of making
predictions where big data analytics delivers predictions that rely on the execution of sequential data processing techniques.
However, there are sentiments that propose the shift from a theory-driven prediction to a process-driven one that analyzes
the steps within the utilization of big data. The theory ought to guide the analysis of big data through acquisition, processing,
analytics and finally the interpretation of the data (Rai, 2016). Therefore, the scientific theory provides a reliable school of
thought in terms of how the interrelation of the data can be done thus offer explanation and prediction.
Thirdly, the Classical Mathematical Theory of Statistics offers a highly diverse approach in terms of the description of
what should be done when analyzing or examining data (Diaconis, 2016). The theory seeks to provide an interpretation of the
patterns as the number of chances fluctuate. We should note that the theoretical framework postulates the decision on the
models and hypothesis that can be developed before interacting with the data. Therefore, classical statistics provides the
much-needed antidote for computing estimates and conducting tests based on the scientists’ assumptions while solving the
problems that may arise within the context of magical thinking. According to Kyburg (2014), the Classical Mathematical
Theory of Statistics offers values such as the P-value and the standard error that are critical towards making interpretations
that are valid and interesting.
3.0. Framework
This section described the research design employed by the researcher to collect, analyze and interpret the
informational data required to answer the study question. Since the study was a review, the researcher employed the
systematic review which is defined as the research design that summarizes the outcomes or results of a number of primary
studies (Ranganathan & Aggarwal, 2020). Systematic reviews utilize meta-analysis that is regarded as a statistical tool that
mathematically collates the outcomes of different research studies thus obtain a collective estimate of the intervention effect.
It is important to note that a systematic review entails methodologically and comprehensively synthesizing literature that is
focused entirely on a well-formulated study question (Siddaway, Wood, & Hedges, 2019). The design usually aims at
identifying and synthesizing the available scholarly research on the study topic, conducting unbiased and reproductive search
of evidence, and involves a meta-analysis.
Fig 2: Systematic Review Process
According to Hanley and Winter (2013), systematic reviews are research methods that involve critically appraising,
summarizing, and reconciling evidence. The technique entails comprehensively reviewing literature in a methodological
manner using a pre-selected protocol thus minimizing bias thus synthesizing the retrieved informational data. This research
design is characterized by having clearly stated objectives along with pre-defined inclusion and exclusion criteria for the
studies. Also, systematic reviews have a reproducible and explicit methodologies, identify studies that meet the eligibility
criteria, Ansari (2022) assess the validity of the study findings, and present and synthesize the characteristics as well as
findings of the studies included in a systematic way (Ranganathan & Aggarwal, 2020).
Furthermore, the use of systematic reviews enabled the researcher to locate the relevant published and unpublished
research studies that addressed the research questions while providing a systematic presentation and synthesis of the findings
and characteristics of included research studies. Livinski at al. (2015) agree that the rationale for conducting systematic
reviews involves making informed decisions, planning for future research agendas, establishing policies, preventing
unnecessary studies and enabling the conducting of comparative effectiveness studies. Therefore, it is important to note the
underlying differences between a systematic review and a narrative review in terms of their respective aims, structures and
models that results in different outcomes within any given study (Hanley & Winter, 2013). Also, there is a reporting guide for
systematic reviews that includes Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA), and PRISMA
+ Health Equity Reporting (PRISMA-E).
The inclusion and exclusion criteria for the research studies to be included in the study was methodologically
considered and developed. According to Meline (2016), searching for research studies across multiple databases requires a
concrete eligibility criterion that enhances the identification, locations, and retrieval of informational data that addresses the
research problem. This process specifies the studies that shall be included or excluded from the systematic review and is
fundamental to the collection of defensible and rigorous data. Therefore, the researcher utilized an eligibility that was based
on the following categories: study population, time period, type of intervention, study variables, quality of the research
methodology and linguistic and cultural range (Patino & Ferreira , 2018). Using the titles and abstracts of the selected research
studies, the researcher was able to include and exclude studies that met or did not meet the eligibility criteria.
4.0. Review
This section includes a report of the systematic review based on the Preferred Reporting Items for Systematic
Reviews and Meta-Analyses or PRISMA statement. The PRISMA statement addresses the poor reporting of most systematic
reviews by providing a checklist comprising of twenty-seven (27) items that are recommended to ensure quality reporting
(Page, et al., 2021). The study searched from several databases including PubMed, JSTOR, ScienceDirect, Public Library of
Science, BioMed Central, Cochrane Open Access and PLOS. The search from these databases resulted in a combined 180 hits
related to the study topic and thus the researcher narrowed to 21 research studies that met the eligibility criteria. The selected
studies were then analyzed based on the PRISMA checklist and the findings were summarized below based on the study
objectives and hypothesis (Ansari 2022).
The concept of data analytics and its applications.
The objective sought to find out the conceptualization of data analytics and the major applications especially in
healthcare. From the 21 research studies, a total of 19 described the concept of data analytics and various applications. For
example, Wang and Alexander (2019) in their study “Big Data Analytics in Healthcare Systems” investigated the concept of big
data in the healthcare system by comparing the various tools that are used to analyze data and their respective features that
determine their performance. The study focused on the big data analytics with respect to the operational concept and the
applications such as disease surveillance, supporting clinical decisions, managing population health, and controlling
epidemics. However, the study was non-experimental thus relied on secondary data to answer the research questions and
address the study problem. As result, the study provided the implications for future research while revealing the significance
of data analytics in improving patient care and outcomes.
Understanding the application of data analytics in healthcare.
This objective sought to understand the various applications of data analytics in healthcare by investigating the
specific case scenarios where the technique has been utilized within health systems. The search result of this objective from
the selected studies was 10 with the articles comprising of both published and unpublished studies. Most of the studies
highlighted the distinct areas within healthcare where data analytics were used. An example was the study by Batko and
Ślęzak (2022) that revealed the move by health facilities towards the provision of data-based care that had significant benefits.
In addition, Abbo and Suchithra (2021) provided a brief idea of the value added to health information through the use of data
analytics and its processes. The study revealed the various applications of big data analytics in healthcare that included
modelling for research and development of drugs, analysis of illness patterns and the tracking of diseases, and the efficient
development of vaccines.
Determining the implications of data analytics in improving patient care.
This objective sought to understand the different implications of data analytics in healthcare. As a result, the search
based on this objective resulted in 5 published studies that revealed the implications of data analytics in healthcare. An
example was a study by Galetsi et al, (2020) that described big data analytics in the healthcare sector with regards to the
underlying theoretical frameworks, techniques or methods and future prospects. The theoretical framework adopted by the
study was the resource-based view theory that focuses on how resources related to big data can be utilized by organizations
thus creating capabilities or values. As a result, the research presented various pragmatic scenarios that highlighted the
advances that had been made possible due to the integration of data analytics in healthcare. The findings were that data
analytics techniques helped healthcare personnel and scholars to make meaningful interpretations with regards to modelling,
visualizations, and clinical analysis.
Establishing the challenges and opportunities of data analytics towards enhancing patient care.
The objective sought to establish the various challenges and opportunities that have been posited by the adoption and
use of data analytics in healthcare. As a result, the researcher searched for studies that described the underlying problems
associated with the application of data analytics within medical processes and the potential opportunities that exist with
regards to the use of data analytics techniques. The search resulted in the selection of 12 studies that provided the basis for
establishing the challenges and opportunities of data analytics in healthcare. An example was the study by Shekarian et al.
(2020) that revealed the need for the healthcare industry to harness the potential of data analytics towards improving
administrative and patient care. By closing the gap in terms of research related to the use of data analytics in healthcare, the
research provided insights into the effect of the concept in improving clinical operations (Dash et al., 2022).
Discussion, Conclusion and Future Implications
Based on the outcomes of the systematic review, it was evident that data analytics had a significant impact in the
healthcare sector with regards to improving patient care. From the searched studies, big data as a concept is relevant to health
systems as it had been found to improve patient outcomes. Using the PRISMA model to report the outcomes of the systematic
review, the study addressed the selected research studies that included investigating the concept of data analytics and its
applications, understanding the application of data analytics in healthcare, determining the implications of data analytics in
improving patient care and establishing the challenges and opportunities of data analytics towards enhancing patient care
(Dash et al., 2022).
Furthermore, the outcomes of the systematic review provided a basis for answering the research problem and declining the
null hypothesis. Therefore, the study accepted the alternative hypothesis that stated that data analytics has significant impact
towards improving patient care. This assumption was supported by findings such as various applications of big data analytics
in healthcare including analysis of illness patterns and the tracking of diseases. The study suggested deeper research in this
area thus provide an in-depth understanding of the implications.
References
Abbo, I. B., & Suchithra , R. (2021). Big Data Analytics in Healthcare. International Journal of Engineering Research & Tehcnology
(IJERT), 4(27), 22-30.
Ansari, M. F. (2022). A quantitative study of risk scores and the effectiveness of AI-based Cybersecurity Awareness Training
Programs. International Journal of Smart Sensor and Adhoc Network., 1–8.
https://doi.org/10.47893/ijssan.2022.1212
Batko, K., & Ślęzak , A. (2022). The use of Big Data Analytics in healthcare. Journal of Big Data, 9(3), 21-33.AI
Dash, B., Ansari, M.F. (2022). Self-service analytics for data-driven decision making during COVID-19 pandemic: An
organization’s best defense. Academia Letters, Article 4978.
Dash, B., Ansari, M. F., Sharma, P., & Ali, A. (2022). Threats and Opportunities with AI-based Cyber Security Intrusion
Detection: A Review. International Journal of Software Engineering & Applications, 13(5), 13–21.
https://doi.org/10.5121/ijsea.2022.13502
Dash, B., Sharma, P., & Ali, A. (2022). Federated learning for privacy-preserving: A review of PII data analysis in
Fintech. International Journal of Software Engineering & Applications, 13(4), 1-13.
Dash, S., Shakyawar, S. K., Sharma, M., & Kaushik, S. (2019). Big data in healthcare: management, analysis and future prospects.
Journal of Big Data, 6(54), 14-25.
Diaconis, P. (2016). Theories of Data Analysis: From Magical Thinking through Classical Statistics. Stanford University .
Elragal, A., & Klischewski , R. (2017). Theory-driven or process-driven prediction? Epistemological challenges of big data
analytics. Journal of Big Data , 4(19), 34-47.
Galetsi, P., Katsaliak, K., & Kumar, S. (2020). Big data analytics in health sector: Theoretical framework, techniques and
prospects. International Journal of Information Management, 50, 206-216.
Gemson, A. E., & Durga, S. (2015). Big Data Analytics in Healthcare: A Survey. APRN Journal of Engineering and Applied Sciences
, 10(8), 3645-3651.
Hanley, T., & Winter, L. A. (2013). What is a systematic review? Manchester: The British Psychological Society .
IFAC. (2018). Data Analytics: An Information Resource for IFAC Members. Retrieved from International Federation of
Accountants: https://www.ifac.org/system/files/publications/files/Data-Analytics-Resource-updated.pdf
Kyburg, H. E. (2014). Classical Statistical Theory. In The Logical Foundations of Statistical Inference (pp. 22-60). Springer.
Lewis, S. (2019). Patient-Centered Care: An Introduction to What It Is and How to Achieve It. British Columbia: Oxford Press.
Livinski, A., Joubert, D., & Terry, N. (2015, October ). Undertaking a Systematic Review: What You Need to Know. Retrieved from
National Institutes of Health : https://www.nihlibrary.nih.gov/sites/default/files/SR_Training_oct2015.pdf
Meline, T. (2016). Selecting Studies for Systematic Review: Inclusion and Exclusion Criteria. Contemporary Issues in
Communication Science and Disorders , 33(1), 21-27. Retrieved from The University of Texas—Pan American,
Edinburg, TX: https://pubs.asha.org/doi/pdf/10.1044/cicsd_33_S_21
Page, M. J., McKenzie, J. E., Bossuyt, P. M., Boutron, I., Hoffmann, T. C., Mulrow, C. D., . . . Moher, D. (2021). The PRISMA 2020
statement: an updated guideline for reporting systematic reviews. Systematic Reviews , 10(89), 43-52.
Patino, C. M., & Ferreira , J. C. (2018). Inclusion and exclusion criteria in research studies: definitions and why they matter.
Brazilian Journal of Pneumology, 44(2): 84-95.
Rai, A. (2016). Synergies between big data and theory. MIS Q, 40(2), 3-9.
Ranganathan, P., & Aggarwal, R. (2020). Study designs: Part 7 – Systematic reviews. Perspectives in Clinical Research , 11(2):
97–100.
Shekarian, N., Ramirez, R., & Khuntia, J. (2020). The Impact of Data Analytics on Hospital Performance. Data Analytics and
Operational Performance, 1-6.
Siddaway, A., Wood, A. M., & Hedges, L. V. (2019). How to Do a Systematic Review: A Best Practice Guide for Conducting and
Reporting Narrative Reviews, Meta-Analyses, and Meta-Syntheses. Annual Review of Psychology, 70(1), 40-53.
Wang, L., & Alexander , C. A. (2019). Big Data Analytics in Healthcare Systems. International Journal of Mathematical,
Engineering and Management Sciences , 4(1), 17-26.Yorke, D. (2017). Patient care: what it? Journal of Patient Care,
2(2), 14-27.
Sharif, Md Haris Uddin, and Mehmood Ali Mohammed. "A literature review of financial losses statistics for cyber security and
future trend." World Journal of Advanced Research and Reviews 15.1 (2022): 138-156.
Energy and AI 11 (2023) 100211
Contents lists available at ScienceDirect
Energy and AI
journal homepage: www.elsevier.com/locate/egyai
Grey box modeling of supermarket refrigeration cabinets✩

K. Leerbeck a ,∗, P. Bacher a , C. Heerup b , H. Madsen a
a
DTU Compute, Bygning 324, 2800 Kongens Lyngby, Denmark
b
Gregersensvej 2, 2630 Taastrup, Denmark
HIGHLIGHTS GRAPHICAL ABSTRACT
• Grey box modeling of supermarket CO2

refrigeration evaporators and cabinets.
• Analysis of physically interpretable pa-
rameter estimates.
• Classification of refrigeration cabinets
based on parameter estimates character-
istics.
ARTICLE INFO ABSTRACT
Keywords: Aiming to enable robust large-scale fault diagnostics and optimized control for supermarket refrigeration
Grey box modeling systems, a data-driven grey box model for an evaporator and its surrounding cooling cabinet (or room) is
CO2 refrigeration systems presented. It is a non-linear model with two states: the cabinet temperature and the refrigerant mass in the
Refrigeration cabinets and evaporators
evaporator. To demonstrate its applicability, data with one-minute sampling resolution from ten evaporators in
System identification
a supermarket in Otterup (Denmark) was used. The model parameters were estimated using a Kalman filter and
Classification
the maximum likelihood method. Since the dynamical properties of the cabinets constantly change as goods
are added and removed, the parameters were re-estimated for each night, over a period of approximately 2.5
years. The model is validated through a statistical analysis of the residuals and the importance of the ongoing
re-estimation of parameters is highlighted. Furthermore, the physical meaning of the estimated parameters is
discussed and potential applications for characterization and classification of cabinets are demonstrated, by
showing how they can be differentiated as either open- or closed cabinets or rooms, using only the estimated
heat transfer coefficients and heat capacities. For a selected case it is shown that the estimated parameter
values are close to physics derived values, and that the accuracy measured by the standard errors of the
estimates is approximately ±10% relative to the estimated values. The analysis demonstrates that the model
is robust, accurate and reliable in terms of estimating physically meaningful parameters and it is therefore
appropriate for large-scale implementation.
✩ This document is the results of the research project Digital Twins (https://digitaltwins4hprs.dk/) partly funded by the EUDP programme.
∗ Corresponding author.
E-mail addresses: kenle@dtu.dk (K. Leerbeck), pbac@dtu.dk (P. Bacher), chp@teknologisk.dk (C. Heerup), hmad@dtu.dk (H. Madsen).
https://doi.org/10.1016/j.egyai.2022.100211
Received 6 August 2022; Received in revised form 12 October 2022; Accepted 13 October 2022
Available online 20 October 2022
2666-5468/© 2022 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
K. Leerbeck et al. Energy and AI 11 (2023) 100211
Nomenclature
Parameters
𝑚̇ Mass flow (kg/s)

𝜆 Conductivity (W/(m K))
𝑃 Pressure (Bar)
𝜌 Density ( kg/m3 )√
𝐴 Valve Constant ( 10 m2 )
𝐶 Heat capacity (kWh∕K)
𝐻 Height (m)
ℎ Enthalpy (J∕kg)
𝐿 Thickness (m)
𝑀 Mass (kg)
𝑅 Thermal resistance (K/kW)
𝑇 Temperature (◦ C)
𝑊 Width (m)
UA Heat transfer coefficient (kW/K)
Super- and subscripts
a Ambient air
c Air inside the cabinet (or room) Fig. 1. Flow sheet of a supermarket CO2 refrigeration system. The three main parts are
e Evaporator marked on the figure. The modeled evaporators and cabinets are in the ‘‘Distribution
system and display cabinets’’ part. On the figure measurement points are marked with
load Surface of wall and door between the room
numbered labels — descriptions of them are found in the main text.
and ambient air
m Surface of evaporator
r Refrigerant inside the evaporator
for refrigerated goods, each with its own compressor equipment, which
rec Receiver tank
significantly complicates the models needed.
Data-driven modeling of supermarket refrigeration systems can be
done by modeling each single component; compressors, condensers and
1. Introduction
evaporators separately, see [14] for modeling of a compressor, [15] for
evaporators and condensers, and [16] for a gas-cooler example. Alter-
To better exploit the benefits of system modeling and automation,
natively, the system may be modeled as a near-complete refrigeration
the need for digital twins (digital representations of a physical systems)
system with several components integrated, see [17] for modeling a
is growing, as reported for industrial systems in general by [1] and
vapor compression plant, [18] for phase change materials and [19]
for refrigeration systems in particular by [2]. With the increasing
for a high heat flux removal system example. Detailed data-driven
amount and resolution of data being gathered from all sorts of sources,
models of a complete system were developed by [20], using sub-
the potential and applications expand, opening up for new research
space modeling – a method for parameterization in non-linear Multiple
questions to be answered. The present paper focuses specifically on
data-driven modeling of supermarket refrigeration rooms and display Inlet Multiple Outlet (MIMO) systems [21]. In the particular study the
cabinets. The presented methodology is an outcome of the continued refrigeration cabinets were not considered individually. Using single-
research from the results presented in [3], where some limitations were component models, the parameterization can be done in higher detail,
found with the applied model and numerical implementation, e.g. lack whilst having fewer parameters per model compared to single multi-
of robustness and convergence in the estimation. component model — this makes parameter estimation easier and more
Grey box modeling – characterized by being a combination physics reliable. Thus, a clever way to use both principles is to first param-
and statistics – of supermarket refrigeration systems is a research area eterize separate models of the main components (e.g. refrigeration
of great interest and potential. The models can be used for applications cabinets, compressors and the condenser) – this was done in [22], using
in optimized control [4,5] and fault detection [6]. Thus, we achieve Predictive Error Minimization for estimation. Afterwards, the complete
physical insight of the parameters of the overall and complex system system was modeled using the estimated parameters. The model was
using measured time series data, as described in more detail by [7,8]. applied to develop control strategies utilizing the high heat capacity
Early methodologies on parameter estimation in stochastic differential of the goods to enable demand flexibility of supermarket refrigeration
equations describing a physical system have been around for many systems.
years [9], but are only recently being implemented on a large-scale A problem with previously suggested models is their lack of prac-
with the new data era. The method has been well tested in many tical identifiability due to high model complexity. This is pointed out
applications, for estimating building thermal dynamics, see [10,11], in [3], where an evaporator in a refrigeration room was modeled. It
and for refrigeration system applications, see [2]. was shown that model over-parameterization led to physically unre-
A promising approach to data-driven modeling and optimization of alistic parameter estimates. The model predicts very well the cabinet
one-stage refrigeration systems is with neural networks and predictive temperature, hence it is applicable for control, however, with physi-
control as presented in [12] to increase the efficiency of the compressor cally infeasible parameter estimation the system properties and energy
in a refrigeration system. This is later followed-up in [13], where demand cannot be determined.
an energy reduction of 17% was achieved on the one-stage system. The objective with the present paper is to demonstrate the bene-
However, supermarkets’ refrigeration systems are usually two-stage fits and potential applications from simplifying the previously applied
systems with two temperature levels, one for frozen goods and one models. The key novelty of the presented results is the application of
2
Fig. 4. Refrigeration room, ref.: Arneg/Incold.
Table 1
Evaborators with their cabinets and valves installed in the system. The
last column lists the valve constants 𝐴 – derived from a previous study
[23].
√
Cabinet Type Valve 𝐴[ 10𝑚2 ]
Fig. 2. Open display cabinet, ref.: Arneg/WICA.
MT1 Room AKV 10–3 0.58765
MT2 Room AKV 10–5 1.48523
MT3 Open cabinet AKV 10–4 0.94185
MT6 Closed cabinet AKV 10–2 0.37191
LT1 Room AKV 10–3 0.58765
LT2 Closed cabinet AKV 10–2 0.37191
The study is based on data from 2012 to 2014 where the opening hours
were 8:00–21:00 every day. The system is a CO2 booster system with a
nominal cooling capacity of around 38 kW and refrigerant mass flow of
0.293 kg/s CO2 . A flow sheet of the system is presented in Fig. 1, where
numbering labels mark measurement points. After the receiver, at the
point marked by ‘‘1’’, the refrigerant is liquefied- and from there, it is
split into the medium temperature evaporator (MT) string and the low
temperature evaporator (LT) string, where expansion valves drop the
pressure to the desired saturation temperature letting refrigerant into
Fig. 3. Closed display cabinet, ref.: Arneg/WICA.
their corresponding evaporator. The valves are controlled with either
a hysteresis or a modulating method with feedback from the measured
cabinet temperature. After the evaporators, at stages ‘‘2a’’ and ‘‘2b’’,
the refrigerant is superheated to avoid any droplets from entering the
the simpler model with fewer parameters and the demonstration of how compressor. At stage ‘‘4a’’, ‘‘4b’’ and ‘‘5’’, after the low-temperature
that ensures identifiability, and thus consistently physical meaningful compressor rack, the pressure is the same, but the enthalpy varies as
parameters. The estimation results using the simpler model for each the MT string and bypass (bp) string from the receiver connects. The
night for ten individual cabinets the parameters were re-estimated each refrigerant now enters the high-pressure compressor rack and continues
night, in total 4143 independent estimations of the model are included the condenser to a sub-cooled state at stage ‘‘8’’- and the cycle repeats.
in the presented results. It is shown how the estimates can be used The present research focus on modeling each cabinet and their cor-
for various applications e.g. characterization and fault detection, which responding evaporator, hence anything between stage ‘‘6’’ and ‘‘8’’ will
would not be possible with previous presented models. not be discussed further. The cabinets included in this study are listed
In Section 2, the supermarket refrigeration system which were in Table 1 with specifications. MT refers to medium temperature (fresh
modeled is presented and explained. In Section 3, the applied grey box food) and LT refers to low temperature (frost). We distinguish between
model and method for parameter estimation are presented. In Section 4, three types of cabinets: room, open- and closed cabinet (illustrated in
the resulting model is illustrated using a five-hour prediction on both Figs. 2–4).
in-sample and out-of-sample data. Furthermore, a residual analysis is The cooling energy released in each cabinet is mostly determined
presented for validating the model. In Section 5, a potential application by the refrigerant mass flow through its corresponding evaporator.
of the model is presented and discussed. It is shown how the model This flow is determined by the opening degree and the valve constant
— the latter is a measure of the specific evaporator valve size. The
can be used for classification of the cabinet type — i.e. whether it is an
valve constants must be known in advance for the parameter estimation
open- or closed cabinet, or a room. Finally, in Section 6 the results are
suggested in this paper otherwise the applied model will be over-
discussed and in Section 7 the conclusions are drawn.
parameterized. The valve constants can also be estimated separately,
as shown in [24].
2. The refrigeration system and data
2.1. Data
The modeled cabinets are part of the supermarket refrigeration
system in the store named ‘‘Fakta’’ in the city of Otterup, Denmark. We The study uses one-minute sampling data consisting of pressures,
consider ten cabinets — six for fresh foods and four for frozen foods. temperatures and valve openings measured at the numbered points on
3
Fig. 5. Nightly averages of the mass flows entering the evaporators for each of the eight cabinets (from 00:00 AM to 4:59 AM every night in the period).
Fig. 6. The mass flow entering the evaporator of the cooling room MT2 during 29th Fig. 7. The mass flow entering the evaporator of the closed cabinet MT6 the 29th of
of October 2013. This is an example of hysteresis control, where the valve is either October 2013. This is an example of modulating control, where the opening of the
fully open or closed. valve is set as a continuous value by the controller.
Fig. 1. The mass flow entering evaporator 𝑖 is calculated by before, models applied in previous literature [3,22,26] were three-state
√ models with the states: cabinet air temperature, temperature of the
𝑚̇ in,𝑖 = (𝑃rec − 𝑃0,𝑖 )𝜌𝑖 ⋅ 𝑜d,i ⋅ 𝐴𝑖 (1) goods and refrigerant mass in the evaporator. In most cases this leads
where 𝑃rec is the pressure in the receiver, 𝜌 is the refrigerant density, to parameter estimates, which are not realistic according to physics –
and for the 𝑖’th evaporator 𝑃0,𝑖 is the suction pressure measured at point e.g. as described in [3] the heat transfer coefficient through the walls
‘‘3a’’ and ‘‘3b’’, 𝑜d,i is the valve opening degree (between 0%–100%) and doors can end up very close to zero, because the goods can act
and 𝐴𝑖 is the valve constant, which values are specified in Table 1. In as an infinite heat reservoir. Therefore, in the present study we have
simplified the model by removing the state of the goods thus lumping
Fig. 5 the nightly averages of the mass flow entering each evaporator
together the cabinet air and the goods parts into a single part. The
are shown for the full 2.5 years period. It can be see that there is a
model for a single cabinet (note, that the 𝑖 subscript on the variables is
a relative large variation in the flows for all the MT cabinets except
omitted for clarity) consists of the system equations
MT6 . Noticeably, it is seen that the refrigerant flow variation in MT6 is
( a )
very small compared to any other cabinet — this is the only cabinet 1 𝑇𝑡 − 𝑇𝑡c 𝑇 e − 𝑇𝑡c
d𝑇𝑡c = + 𝑀𝑡r 𝑡 d𝑡 + 𝜎c d𝜔c,𝑡 (2)
using modulating control rather than hysteresis control used in the 𝐶c 𝑅load 𝑅m
other cabinets. Furthermore, the LT cabinets are seen to generally have ( )
d𝑀𝑡r = 𝑚̇ in,𝑡 − 𝑚̇ out,𝑡 d𝑡 + 𝜎r d𝜔r,𝑡 (3)
less variation than the MT cabinets.
𝑀r
A plot of the one-minute values during a single night of mass 𝑡
where 𝑚̇ out,𝑡 = 𝑅 𝛥ℎ . The latter term in each equation is the diffusion
m e
flow entering the hysteresis controlled evaporator of MT2 is shown in term, they are formed by the standard Wiener processes {𝜔c,𝑡 } and
Fig. 6. It can be seen how the valve opens fully, which occurs when {𝜔r,𝑡 }, where 𝜎c and 𝜎r become the incremental standard deviations of
the upper cabinet temperature bound is reached, and closes again, the processes. The ambient temperature, 𝑇 a𝑡 , is an input variable , which
which occurs when the lower bound is reached. In Fig. 7 the mass was set constant 𝑇 a𝑡 = 20◦ C, since measurements of the air temperature
flow for the modulating controlled MT6 is shown. It is seen that the in the store were not available. The measurement equation is
flow is continuous, however around 00:30 the valve is fully opened
and afterwards closed — the reason for this event is not known. To 𝑌𝑘 = 𝑇 c𝑡 + 𝑒𝑘 (4)
𝑘
avoid disturbances e.g. customer interference, the model is fitted for where 𝑒𝑘 ∼ 𝑁(0, 𝜎obs ) and i.i.d., hence the model output is the cabi-
all evaporators during closing hours every night from 12AM to 5AM – net temperature. The enthalpy difference over the evaporator, 𝛥ℎe , is
i.e. five hours with 300 observations as shown in the two figures. calculated using CoolProp [27]. The deterministic part of the system
equations can be illustrated by an RC-diagram [28]. A diagram for the
3. Methodology applied model is shown in Fig. 8.
In the following presentation of the results UA-values rather than
In this section grey-box model and estimation method are presented. thermal resistances 𝑅 are used, because they provide a more intuitive
understanding (UA = 𝑅1 ).
3.1. Modeling
3.2. Parameter estimation
The model is a state space model, derived from thermodynamic state
equations describing the heat and mass dynamics of the cabinet and The R package, CTSM-R, is used for maximum likelihood estimation
in the evaporator as a lumped dynamic process model [25]. As noted of the parameters. For all details on how the Kalman filter is used for
4
Fig. 8. RC-diagram diagram of the model. The refrigerant mass state, 𝑀 r , is included
in the heat source illustrating the heat transfer to the cabinet.
state estimation and calculation of the likelihood function, and how it

is maximized for the parameter estimation, see [29,30].
4. Results
In this section the obtained parameter estimates are discussed and

the model is validated using an example from the fresh food refriger-
Fig. 9. Zero to five hour in-sample predictions covering the selected night (2012-06-09)
ation room (MT2 ). The validation is carried out for the night the 9’th
for evaporator MT2 . The upper plot is of the measured cabinet temperature state and
of July 2012, which was selected as a representative night of regular the lower is of the hidden mass state.
operation. The estimated parameters for the night are listed in Table 2.
It is found that the UAload estimate is fairly low because it is the well
insulated refrigeration room. Assuming a simple physical model of the
room the UA value have been calculated based on the room specifica-
tions and physical material properties, as explained in Appendix. The
calculated UA value is 31.34 W K
, which is within a fair range from the
estimated value 37.5 W K
.
The specific heat transfer coefficient over the evaporator, 𝑈 𝐴m ,
describes the rate at which the evaporator transfers heat to the cabinet
— hence, it can be affected by both conditions inside the evaporator as
well as on the outside e.g. obstacles such as fouling and icing.
The heat capacity of the cabinet, 𝐶 c , reflects to some extend the
level of goods stored — it represents the combination of the heat
capacity of the air, cabinet parts and outer layer of the goods, which
responds to the temperature dynamics. The center temperature of the
goods is more or less stationary. The estimated heat capacity should
therefore be interpreted as the effective heat capacity, hence the heat
capacity which was activated during the particular operation during
the observed period.
The estimates of the diffusion terms standard deviations, 𝜎c and 𝜎r ,
and the measurement standard deviation, 𝜎obs , indicate that the model
uncertainty is absorbed in the refrigerant mass state, since its standard
deviation is much higher than the other two. Fig. 10. Zero to five hour out-of-sample predictionscovering night following the
The standard error estimates (in the ‘‘Std. Error’’ column) indicate selected night (2012-06-09) for evaporator MT2 . The upper plot is of the measured
cabinet temperature state and the lower is of the hidden mass state.
the accuracy of the parameter estimates. A 95% confidence interval
is roughly plus and minus two times the Std. Error away from the
estimate. It can be seen that the parameters are reasonably accurately Table 2
Model parameters estimated for evaporator MT3 on
determined, with standard errors roughly ±10% relative to the esti- the 9th of June 2012.
mated values and thus certainly highly significantly different from
Parameter Estimate Std. Error
zero. [ ]
The model performance is further analyzed by five-hour multi-step 𝑈 𝐴load W 37.5 4.0
[ K]
W
predictions of the two states. In Fig. 9 the prediction of the same night 𝑈 𝐴m kgK 270.2 42.3
[ ]
kJ
on which the estimation was carried out is shown — hence an in-sample Cc K 356.5 34.6
prediction. Small issues can be seen in predicting the dynamics during 𝜎c 3.3e−12 –
the off-periods when the temperature is rising. In Fig. 10 a plot of 𝜎r 1.4 –
𝜎obs 5.6e−11 –
the prediction for the following night is shown — hence an out-of-
sample prediction. Here, the prediction is much worse. The reason is
that the dynamics of the room have changed. From having a frequency
of only 11 valve openings during the five hours, it now has 14 openings. 4.1. Residual analysis
The most plausible explanation is that fewer goods are stored in the
room leading to a decreased heat capacity. Thus re-estimation of the Diagnostic plots of the in-sample one-step prediction residuals are
model must be done regularly to reflect the dynamics, which change shown in Fig. 11. They do suggest that the model can be improved.
depending on the loaded goods, especially if the model is to be used From the two upper plots (time series and auto-correlation function
for temperature prediction in a controller. (ACF)), it is seen that there is significant periodic auto-correlation in
5
Fig. 12. Each dot represents the estimates of 𝑈 𝐴load and 𝐶 c for a single cabinet during
one night. It can be seen that there is a clear different pattern from the open and closed
cabinets.
Fig. 11. Diagnostic plots of the in-sample one-step prediction residuals (2012-06-09).
The upper-left plot is of the residuals, the upper-right plot is the auto-correlation
function, the lower-left is the Q-Q plot and the histogram of the residuals.
is useful insights into the refrigeration system that can help optimizing
control settings for individual evaporators, and has potential for use
in fault detection, e.g. if the parameters drift outside of their normal
the residuals, likely caused by the increasing errors during the off-
range. Applications which can be investigated in further research.
period of every cycle. From the two lower plots, the QQ-plot and the
histogram, it can be seen that both indicate slightly skewed residuals.
6. Discussion
For predictive applications previously presented models out-perform
the current simplified model e.g. the three-state model from [3,22].
The presented results show that useful information can be extracted
However, one issue demonstrated in [3] was the lack of robustness and
from data using a rather simple grey box model. Because of its robust-
more often than not the parameter estimates were not reasonable from
ness and simplicity, it can be implemented as a general model on a far
a physical point of view, thus they could not be compared and analyzed
larger scale than previous models. The applications of the model are
from day to day. With the current simpler model, we were able to get
also different, whereas previous models primarily were derived with
consistent results and thus enabling a useful comparison of parameter
the objective of control, this model can be used for more informative
estimates with some interesting potential applications – in spite of the
applications, such as classification and fault detection. Without further
inferior model predictions and diagnostics.
knowledge about a supermarket refrigeration system, we can extract
information about the cabinet type. Potentially, the methodology can
5. Parameter analysis and applications be used to extract operational information about the systems e.g.
detection of fault states (icing built-up, refrigeration leakage etc.). It is
In this section the parameter estimates for all the cabinets from the through residual analysis and tracking of the parameter estimates that
entire period, together with some suggestions of potential applications, cabinet outliers and faulty operation can be identified — enabling early
are presented and discussed. detection of cabinets which should be repaired or replaced. Further
research involving multiple supermarkets can analyze the potential for
5.1. Cabinet classification scaling up the applications. Furthermore, tracking of the heat capacity,
𝐶c , in the storage rooms can be used to track deliveries.
From the models fitted every night over the 2.5 years, a clear These are just a few examples of the potential applications for
pattern arise in the characterization of every cabinet. In Fig. 12, the the model, and as further research into the area drives deeper, more
estimated heat capacities of the cabinets, 𝐶 c , are plotted versus the applications will most likely be revealed.
estimated heat transfer coefficients, 𝑈 𝐴load – each dot represents the
parameter estimates for a single night. It can be seen that, generally, 7. Conclusion
all cabinets can be identified using these two parameters as they all
clearly have distinct distributions. Furthermore, all closed cabinets are A grey box model describing the thermal dynamics of a supermarket
clustered with low 𝐶 c and low 𝑈 𝐴load , they are the frost cabinets, LT2 , refrigeration cabinet and its evaporator was presented. It was demon-
LT3 and LT4 , along with the fresh food cabinet, MT6 . All of them are strated how it, together with a maximum likelihood estimation method,
closed cabinets that are generally smaller and better insulated than can be used to extract important physical parameter estimates with data
open cabinets. The light blue dots represent the frost room (LT1 ) – from a regular supermarket refrigeration system. The relatively simple
they are differentiated from the closed cabinets by having a larger heat model of the cabinet temperature was proved to be sufficient for the
capacity explained by its larger volume. The open cabinets and fresh purpose. It was furthermore demonstrated that the parameter estimates
storage room all have a higher variation and higher 𝑈 𝐴load . Here, too, could be used to effectively differentiate between rooms, and open
the room (red dots) has a higher heat capacity. and closed cabinets. This enables automatic classification of cabinets
This analysis shows that the model and this type of data can be used in large scale deployments. Thus further research should include data
to clearly differentiate between rooms, and closed and open cabinets. It from multiple supermarkets and focus on development of classification
6
Table 3
Room specifications: 𝑊 is the width of the section and the
area, 𝐴, is calculated from the room height, 𝐻 = 2.5 m. The
heat transfer coefficient, 𝑈 𝐴, is calculated in two parts — first,
the total area of the parts with insulation((walls
) and roof) then
𝐴
the total area of the glass doors (𝑈 𝐴 = 𝐿
).
𝑘
Parameter 𝑊 [m] 𝐴 [m2 ] 𝑈 𝐴 [W∕K]

Insulated walls, door and roof
𝑊1 2.8 m
𝑊2 1.5 m
𝑊3 3.9 m
𝑊4 6.3 m
Roof 20.37 m2
Total 55.62 m2 17.48 W∕K
Glass doors
Total 3.9 m 12.6 m2 13.86 W∕K
Complete room
Total 31.34 W∕𝐊
Fig. 13. Floor diagram of the refrigeration room. It has glass doors towards the sales
area on the left and an insulated door to the stock area on the right. The costumers
take the cold goods through the doors on the left and staff enters the room through The wall widths are listed under 𝑊 and the areas calculated under 𝐴.
the door on the right. Dimensions used in the calculations are marked.
The heat transfer coefficient, 𝑈 𝐴, is calculated in two parts — first, the
total area of the parts with insulation (walls and roof) then the total
area of the glass doors.
models using the parameter estimates as input. Possible applications,
such as tracking of abnormal cabinet operation for early fault detection,
References
should also be investigated in follow-up studies.
[1] Jiang Y, Yin S, Li K, Luo H, Kaynak O. Industrial applications of digital twins.
CRediT authorship contribution statement Phil Trans R Soc A 2021;379.
[2] Brok Niclas Brabrand, Munk-Nielsen Thomas, Madsen Henrik, Stentoft Peter A.
K. Leerbeck: Conceptualization, Methodology, Software, Writing. Flexible control of wastewater aeration for cost-efficient, sustainable treatment.
P. Bacher: Conceptualization, Methodology, Supervision, Writing. C. IFAC-PapersOnLine 2019;52(4):494–9, IFAC Workshop on Control of Smart Grid
and Renewable Energy Systems CSGRES 2019.
Heerup: Data curation, Supervision, Review.
[3] Leerbeck Kenneth, Bacher Peder, Heerup Christian. Grey box modelling of
supermarket refrigeration room. In: Proceedings of the 2021 international con-
Declaration of competing interest ference on electrical, computer and energy technologies. European Union; 2021,
2021 International Conference on Electrical, Computer and Energy Technologies,
ICECET 2021 ; Conference date: 09-12-2021 Through 10-12-2021.
One or more of the authors of this paper have disclosed potential or
[4] Glavan Miha, Gradišar Dejan, Moscariello Salvatore, Juričić Ðani, Vrančić Damir.
pertinent conflicts of interest, which may include receipt of payment,
Demand-side improvement of short-term load forecasting using a proactive load
either direct or indirect, institutional support, or association with an management – a supermarket use case. Energy Build 2019;186:186–94.
entity in the biomedical field which may be perceived to have poten- [5] Brok Niclas, Green Torben, Heerup Christian, Oren Shmuel S, Madsen Henrik.
tial conflict of interest with this work. For full disclosure statements Optimal operation of an ice-tank for a supermarket refrigeration system. Control
refer to https://doi.org/10.1016/j.egyai.2022.100211. Kenneth Leer- Eng Pract 2022;119:104973.
[6] Li Yanfei, Sun Jian, Fricke Brian, Im Piljae, Kuruganti Teja. Grey-box fault models
beck reports financial support was provided by Danish Energy Agency.
and applications for low carbon emission CO2 refrigeration system. Int J Refrig
Peder Bacher reports financial support was provided by Danish Energy 2022;141:76–89.
Agency. Christian Heerup reports financial support was provided by [7] Kristensen Niels Rode, Madsen Henrik, Jørgensen Sten Bay. A method for
Danish Energy Agency. Henrik Madsen reports financial support was systematic improvement of stochastic grey-box models. Comput Chem Eng
provided by Danish Energy Agency. 2004;28(8):1431–49.
[8] Maruta Ichiro, Sugie Toshiharu. Projection-based identification algorithm for
grey-box continuous-time models. Systems Control Lett 2013;62(11):1090–7.
Data availability [9] Nielsen Jan Nygaard, Madsen Henrik, Young Peter C. Parameter estimation in
stochastic differential equations: An overview. Annu Rev Control 2000;24:83–94.
The authors do not have permission to share data. [10] Bacher Peder, Madsen Henrik. Identifying suitable models for the heat dynamics
of buildings. Energy Build 2011;43(7):1511–22.
Acknowledgment & funding [11] Kim Donghun, Cai Jie, Braun James E, Ariyur Kartik B. System identification
for building thermal systems under the presence of unmeasured disturbances
in closed loop operation: Theoretical analysis and application. Energy Build
This document is the results of the research projects Digital twins for 2018;167:359–69.
large-scale heat pumps and refrigeration systems (EUDP 64019-0570) and [12] Kizilkan Önder. Thermodynamic analysis of variable speed refrigeration system
Flexible Energy Denmark (FED) (IFD 8090-00069B). using artificial neural networks. Expert Syst Appl 2011;38(9):11686–92.
[13] Cirera Josep, Carino Jesus A, Zurita Daniel, Ortega Juan A. Improving the
energy efficiency of industrial refrigeration systems by means of data-driven load
Appendix. UA calculation for fresh food refrigeration room.
management. Processes 2020;8(9).
[14] Pérez-Segarra CD, Rigola J, Sòria M, Oliva A. Detailed thermodynamic charac-
The refrigeration room has glass towards the sales area on the left. terization of hermetic reciprocating compressors. Int J Refrig 2005;28:579–93.
It has a thick insulated door towards the storage area to the right. The [15] Willatzen M, Pettit NBOL, Ploug-Sørensen L. A general dynamic simulation
walls and the door towards the storage room have 7 cm of insulation model for evaporators and condensers in refrigeration. Part I: moving-
boundary formulation of two-phase flows with heat exchange. Int J Refrig
in the walls and the glass doors are 2 cm thick (properties used for
1998;21:398–403.
walls: 𝜆 = 0.022 W∕(mK) & 𝐻 = 2.5 m, and for glass doors: 𝑈 = [16] Corradi M, Cecchinato L, Schiochet G, Zilio C. Modelling fin-and-tube gas-cooler
1.1 W∕K & 𝐻 = 2 m). Data about the room is listed in Table 3. The for transcritical carbon dioxide cycles. In: International refrigeration and air
walls are referred to as 𝑊 and use the same numbering as in Fig. 13. conditioning. 2006.
7
[17] Aprea C, Renno C. An experimental analysis of a thermodynamic model of a [25] Madsen Henrik. Time series analysis. Chapman & Hall/CRC - Taylor & Francis
vapor compression refrigeration plant on varying the compressor speed. Int J Group; 2007.
Energy Res 2004;28:537–49. [26] Sarabia D, Capraro F, Larsen LFS, de Prada C. Hybrid NMPC of supermarket
[18] Wang F, Maidment G, Missenden J, Tozer R. The novel use of phase change display cases. Control Eng Pract 2009;17:428–41.
materials in refrigeration plant. Part 2: Dynamic simulation model for the [27] Bell Ian H, Wronski Jorrit, Quoilin Sylvain, Lemort Vincent. Pure and
combined system. Appl Therm Eng 2007;27:2902–10. pseudo-pure fluid thermophysical property evaluation and the open-source ther-
[19] Zhou R, Zhang T, Catano J, Wen JT, Michna G J, Peles Y, et al. The steady-state mophysical property library CoolProp. Ind Eng Chem Res 2014;53(6):2498–508.
modeling and optimization of a refrigeration system for high heat flux removal. [28] Madsen H, Holst J. Estimation of continuous-time models for the heat dynamics
Appl Therm Eng 2010;30:2347–56. of a building. Energy Build 1995;22(1):67–79.
[20] Andreasen G, Stoustrup J, Izadi-Zamanabadi R, Pardiñas Á, Hafner A. Data-driven [29] Madsen Henrik. Continuous time stochastic modeling in R - user’s guide and
modeling of a CO2 refrigeration system. In: American control conference. 2019. reference manual. CTSM-R Development Team; 2003.
[21] Katayama T. Subspace methods for system identification. Springer; 2005. [30] Kristensen Niels Rode, Madsen Henrik, Jørgensen Sten Bay. Parameter estimation
[22] Shafiei E, Rasmussen H, Skoustrup J. Modeling supermarket refrigeration systems in stochastic grey-box models. Automatica 2004;40(2):225–37.
for demand-side management. Energies 2013;6:900–20.
[23] Fredslund K. Load profiles for supermarket refrigeration. Technical Report, IPU;
2013, p. 75.
[24] Leerbeck Kenneth, Bacher Peder, Heerup Christian, Madsen Henrik. Estimation
of evaporator valve sizes in supermarket refrigeration cabinets. 2022.
8
Mechatronics 68 (2020) 102369
Mechatronics
journal homepage: www.elsevier.com/locate/mechatronics
Design, fabrication, modeling and control of a fabric-based spherical

robotic arm
Matthias Hofer∗, Raffaello D’Andrea
Institute for Dynamic Systems and Control, Sonneggstrasse 3, 8092 Zurich, Switzerland
a r t i c l e i n f o a b s t r a c t
Keywords: In this paper we present a spherical soft robotic arm made from fabric. The inflatable arm has a small mass and
Soft robotics is pneumatically actuated. A configuration is employed with only three actuators controlling the two rotational
Pneumatic bellow actuator degrees of freedom of a spherical joint. This differs from the commonly employed antagonistic pairs, including
Fabrication
four actuators for two degrees of freedom. The fabrication procedure of the lightweight and compliant system
Data-Driven modeling
is discussed in detail and uses commonly available materials and tools. The capability of the robotic arm to
Learning control
adjust the joint stiffness as a function of the actuator pressures is investigated and characterized for different
deflection directions. The static mapping from the actuator pressures to the orientation of the robotic arm is
identified from data and the inverse mapping is employed in a position controller. The modeling and controller
derivation are performed for three different stiffness levels demonstrating the ability of the spherical robotic
arm to change the joint stiffness independently of controlling the position. The position tracking performance
is experimentally evaluated by tracking a square trajectory. A comparison of the tracking performance for the
different stiffness levels shows that accurate tracking is more challenging for the smallest joint stiffness. A gray-
box model capturing the interactions of the two degrees of freedom is used in a learning scheme that is applied
for the smallest stiffness level. The learning approach reduces interactions between the two degrees of freedom
and demonstrates the control performance achievable with the system developed.
1. Introduction In this work we present a spherical robotic arm that combines two
rotational DoF in a single spherical joint (see Fig. 1). A configuration
Pneumatically actuated soft robotic systems are attracting consid- with three actuators (referred to as an actuator triplet) permits control
erable attention due to their promising intrinsic properties. A detailed of the two DoF, as well as the stiffness of the joint. Therefore, the pneu-
overview of the field can be found in [1–3]. Fabric-based soft robotic matic peripherals, such as the number of valves and pressure sensors,
systems are enabling various applications that are challenging to im- can be reduced. The fabrication procedure is presented in detail to sim-
plement with traditional rigid robotic systems. Due to their low weight plify the reproduction of the robotic arm presented or similar systems.
and compliant material behavior they are safe for human interaction The joint stiffness is experimentally characterized for different actuator
[4–6] and the antagonistic arrangement of soft inflatable actuators al- pressure levels and in different deflection directions. Measurements of
lows control of the stiffness of the joint [7,8]. Furthermore, the use of the possible interaction forces exerted by the robotic arm are performed
inflatable links to provide structural stability in combination with pneu- for different actuator elongations and varying actuator pressures. The
matic actuation permits long range manipulators for inspection [9] and steady state relation between the controlled actuator pressures and the
results in a high payload to weight ratio which is interesting for space resulting angles is identified from data for three different stiffness levels
applications [10]. Compared to other soft robotic actuation principles, denoted as soft, medium and stiff. The inverse mapping is derived for all
pneumatic actuation allows for fast actuation as demonstrated in [11]. three levels and used in separate position controllers. The control perfor-
Moreover, the flexibility of the fabric employed makes these systems mance resulting from the different stiffness levels is compared for angle
suitable for wearable applications, where for example, soft inflatable trajectory tracking. A gray-box model is presented that incorporates the
actuators can support humans in lifting tasks [12]. actuator geometry and the resulting interactions arising between the
Pneumatically actuated fabric-based systems typically rely on the se- two DoF. A learning scheme based on the gray-box model is applied for
rial combination of one degree of freedom (DoF) joints to allow move- the lowest joint stiffness mode for which accurate position tracking is
ment in a higher dimensional space.
∗
Corresponding author.
E-mail addresses: hofermat@ethz.ch (M. Hofer), rdandrea@ethz.ch (R. D’Andrea).
https://doi.org/10.1016/j.mechatronics.2020.102369
Received 19 September 2019; Received in revised form 16 April 2020; Accepted 23 April 2020
Available online 18 May 2020
0957-4158/© 2020 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license. (http://creativecommons.org/licenses/by/4.0/)
M. Hofer and R. D’Andrea Mechatronics 68 (2020) 102369
outlined in [29] for a rotary elastic chamber actuator. Advanced con-

trol approaches using the piecewise constant curvature assumption are
presented in [30] for a continuum fluidic elastomer manipulator oper-
ating in two dimensional space and in [18] for a manipulator operating
in three dimensional space. An adaptive neural network controller is
discussed in [31] for the Compact Bionic Handling Arm and a reinforce-
ment learning approach is employed in [32], for a combined tendon and
pneumatic manipulator.
1.2. Outline
The design and fabrication of the spherical soft robotic arm is pre-
sented in Section 2. The joint stiffness characterization, as well as a data-
driven model for the static mapping from actuator pressures to angles
for three different joint stiffness levels are presented in Section 3. This
section also discusses a gray-box model capturing the actuator configu-
ration and an analysis of the interaction torques of the robotic arm. A
control approach exploiting the pressure to angle mapping is discussed
in Section 4, together with a learning scheme based on the gray-box
model to improve control performance. Results from angle tracking ex-
periments for different joint stiffness levels and after applying a learning
scheme are presented in Section 5 and a conclusion is drawn in Section 6.
2. Design and fabrication
In this section, we first discuss different actuator configurations and

show the motivation for the design chosen and realized in this work.
The fabrication of the robotic arm is presented in the second part of this
Fig. 1. The spherical soft robotic arm presented in this paper. The robotic arm section.
is made of fabric and consists of two inflatable links and three bellow type actu-
ators. The pressure in each actuator can be controlled independently to adjust 2.1. Actuator configurations
its elongation. A rigid ball-and-socket joint connects the two links and restricts
the actuator elongations to two rotational degrees of freedom. A motion capture
Soft robotic systems with an antagonistic actuator configuration per-
marker is attached at the tip of the movable link to provide sensory feedback.
mit simultaneous position and stiffness control. Here, we only consider
Both links have a diameter of 10cm and a length of 30cm. The movable link
without actuators weighs 126g. joints with a single center of rotation and no continuum joints (see e.g.
[21]). Three different configurations are depicted in Fig. 2. The config-
uration on the left (C1) has a pivot joint made from soft materials that
more challenging. The learning approach is shown to improve control is controlled by two actuators and is presented in [8]. The arrangement
performance for angle trajectory tracking. in the middle (C2) consists of a ball-and-socket joint controlled by three
actuators and is realized in this work. The configuration on the right
(C3) also has a ball-and-socket joint, but uses four actuators for control.
1.1. Related work
The reachable points of configuration C1 are located on a circle. Two
DoF can be controlled with configuration C1 by serially arranging two
Considering the design of air-driven soft articulated robotic systems,
joints of this type, where one joint is rotated by 90∘ relative to the other
a one DoF system relying on soft actuators and a rigid backbone is pre-
joint (see [10]). The reachable points of configuration C2 and C3 are
sented in [13,14], while an entirely soft system is shown in [8]. Exam-
ples of higher dimensional systems with a serial arrangement of multiple
articulated single DoF joints can be found in [15–17].
The actuator configuration that we employ in this work has previ-
ously been applied to elastomer-based continuum robots (e.g. [18,19])
as well as in the articulated hand axis of Festo’s Bionic Handling Assis-
tant [20]. A six DoF system is presented in [21] using four actuators for
each two DoF continuum joint, which is made from more rigid materi-
als. A fabric-based system with multiple actuators along the length of an
inflatable link resulting in a continuum resembling system is discussed
in [22].
In terms of modeling inflatable fabric-based systems, a geometrical
approach for a high-displacement actuator is outlined in [23]. Models Fig. 2. Top view of different actuator configurations with two (C1), three (C2)
based on traditional system identification are presented in [8,24]. Non- and four actuators (C3). All configurations are displayed in the 𝑥 − 𝑦 plane. A
single actuator is indicated by the trapezoidal object with rounded corners. The
linear modeling techniques based on neural networks are discussed in
configuration C1 includes two actuators arranged in an antagonistic fashion and
[21,25]. A hybrid model consisting of continuum kinematics combined
allows for simultaneous position and stiffness control of one DoF. The config-
with a data-driven model is presented in [26] for the Bionic Handling uration C2 has three actuators arranged around a single pivot point. Two DoF
Assistant. can be controlled as well as the total joint stiffness. Configuration C3 consists of
Different approaches have been investigated for the control of pneu- four actuators arranged around a single pivot point. This permits the control of
matically actuated soft systems. Open loop control is discussed in two degrees of freedom and the independent stiffness control of both degrees of
[27,28] and an approach based on sliding mode and fuzzy control is freedom.
figuration uses only three instead of four actuators. As a consequence,

the independent stiffness control of both DoF is not possible, since there
are only three control inputs for two angular DoF and two stiffness DoF.
However, fewer pneumatic peripherals (valves, pressure sensors) are re-
quired.
2.2. Fabrication
In this section, we summarize the materials used, the fabrication pro-

cedure applied and the insights gained. The aim of this section is to sim-
plify the reproduction of the robotic arm presented and to encourage
the application of the materials and methods mentioned for the devel-
opment of similar fabric-based robotic systems. All drawings of the two
dimensional fabric and bladder pieces as well as the CAD files of the 3D
printed parts are provided in the additional files of this manuscript.
The robotic arm is made from inexpensive and commonly available
materials. The only machines required for manufacturing are a laser
cutter for preparing the different fabric pieces and a thermal heat press
Fig. 3. Explosion view of the spherical robotic arm depicting the different com- for welding them together. The single pieces can also be cut manually
ponents. The arm is mounted on the ground by means of a base plate. The actu- at the cost of increased production time.
ator triplet encloses the joint part, which includes the ball-and-socket joint, and The soft arm consists of a static link, a movable link, the actuator
is mounted between the static and movable links. The first layer of the actuator triplet, the joint part, and a base plate (see Fig. 3). The actuator triplet
triplet is clamped between the joint part and the end part of the static link and
is mounted between the two links and the elongation of the individual
the last layer between the joint part and the end part of the movable link.
actuators upon pressurization is used to control the movable link. The
static link is attached to the ground by a 3D printed base plate (PA 12
jet fusion, as all 3D printed parts). A ball-and-socket joint (IGUS AGRM-
located on a sphere. These configurations have the advantage that two 08MS, modified to increase the operating range) is integrated into the
DoF can be controlled in a single more compact joint. The drawback is 3D printed joint part. The actuator triplet laterally encloses the joint
that the complexity of the joint is increased and possible interactions be- part and the two ends of the joint part protrude from the top and bottom
tween the actuators are more pronounced. Compared to C3, the C2 con- layers of the actuator triplet (see Fig. 4 and Fig. 5). These ends fit into
Fig. 4. Images of the different components of the fabric-based soft arm (the actuator triplet is shown in detail in Fig. 5). From left to right: (A): The outer fabric
shell of the static link with Velcro straps on the side to attach the actuators and the 3D printed part on top to mount the joint part. The inflated link has a diameter
of 10cm and a length of 30cm (measured from the base plate to the pivot point). (B): The airtight inner bladder of the static link, which is slightly larger than the
outer shell to fully occupy the available volume. The top part is rounded to better match the shape of the outer shell. Tubing is connected by an angle connection
valve (Carmo, type 09-002) glued to the bladder. (C): Base plate to mount the static link on the ground. The outer shell of the static link (part A) is pulled over
the conical part and secured with the ring, which is then fixed to the conical part with four screws. Tubing to pressurize the inner bladder of the static link (part
B) is integrated. (D): 3D printed joint parts before they are attached to the fabric shells. The top and bottom triangular parts are glued to the fabric shells of both
links. The inner part houses the ball joint and fits into the top and bottom triangular parts. It is secured with three screws. The first (respectively last) fabric layer of
the actuator (referred to as the connecting triangle) is clamped between the bottom (respectively top) triangular part and the inner part housing the ball joint. The
limited range of the ball joint employed currently also limits the actuation range of the robotic arm to angular deflections of 45∘ . (E): The assembled movable link
(inflated) consisting, similarly to the static link, of the outer fabric shell, the inner bladder and two triangular parts at both ends. Similarly to the static link, Velcro
straps are attached to fix the actuators. The tubing (SMC, 4mm plastic tube) pressurizing the inner bladder can be seen on the bottom and the reflective marker for
the motion capture system is attached on the top. The movable link, when inflated, has a diameter of 10cm and a length of 30cm (measured from pivot point to
motion capture marker). The total mass of the movable link is 126g. The 3D printed end parts account for 1/5 of the mass.
Fig. 5. The actuator triplet with one actuator inflated to different elongations (increasing from left to right). The joint part housing the ball joint is enclosed and
protected by the soft actuator triplet. Each actuator consists of ten single cushions connected to each other through an inner seam. The volume of a single actuator
is approximately 0.35L. Three tubes are attached to each actuator. Two are connected to the switching valves for inflating and deflating the actuator and one to a
pressure sensor.
the 3D printed triangular end parts of the two links. The modular design
facilitates simple assembly and access to all components.
The robotic arm primarily consists of fabric and the fabrication pro-
cedure presented in [23] is applied. Two layers of poplin fabric (extrem-
textil, polyester-cotton-blend 65/35) are used in combination with one
layer of thermoplastic polyurethane (TPU) film (Perfectex, HM65 ther-
moplastic polyurethane hot melt adhesive 0.1mm) between the poplin
fabric layers to form a sandwich structure (poplin-film-poplin). The
three sheets of raw material are combined with a heat press (Insta, man-
ual heat press). The resulting layer of processed fabric is airtight, inex-
tensible, and more robust than the raw poplin fabric. It is the primary
material for the actuator and the outer shell of the links.
2.2.1. Actuator fabrication

The actuator triplet consists of three single soft bellow actuators ar-
ranged symmetrically around a pivot point (see Fig. 2). A single bellow
actuator consists of ten cushions that are connected by an internal seam.
The geometry of the triplet is shown in Fig. 6. The actuators are arranged
symmetrically around a connecting triangle in the middle, which is used
Fig. 6. A drawing of the actuator triplet defining its geometry. All numbers
for clamping the actuator triplet between the joint part and the links.
(except angles) are denoted in mm. Three actuators are arranged symmetrically
The fabric and TPU film pieces are prepared with a laser cutter (Uni-
around a central triangle. Each actuator has three holes to connect the tubing
versal Laser Systems, VersaLaser). Two pieces of TPU film are used for and a strap with two gaps to attach the actuators with Velcro straps to the links.
a single seam. The reason for using two layers of TPU film is that the The dashed lines in the actuator on the right indicate the parts beneath the top
fabric absorbs some of the TPU film and therefore, a single layer of film layer. The rounded square in the middle specifies the location of the internal
does not provide enough adhesive strength. In a first step the TPU film seam connecting two cushions. The location of the TPU film forming the seam
pieces are welded to the fabric pieces, which are to be combined (see is indicated by the second dashed line. The dashed line close to the outer con-
Fig. 7 for a detailed explanation of the process). This step is repeated for tour specifies the location of the TPU film used to combine two layers of fabric
all pieces of the actuator. In a second step the fabric pieces that already to form a single cushion. All hot melt adhesive seams have a width of 5mm.
include TPU film at the seam are welded together. The second step is re- Placing the opening between two cushions off-center towards the middle of the
peated for each actuator separately in a bottom up process, where new triplet increases the elongation range of a single actuator and consequently the
operating range of the movable link.
fabric pieces are added to the already welded stack on top (assuming
that the heat press applies the heat from the top plate).
The three actuators are connected at the bottom with a piece of opening for the valve. Tubing is connected by an angle connection valve
fabric denoted as the connecting triangle and depicted in Fig. 6. On glued to the bladder in a second step.
the top, the actuators are completed with separate pieces of fabric (not An advantage of a two shell design is that the requirements for the
yet connected). The connecting triangle is glued on top in a later step, material can be divided. While the inner bladder is required to be air-
after the joint part is integrated into the actuator triplet. Finally, tub- tight, the outer shell is not, but only needs to constrain the expanding
ing (SMC, 6mm plastic tube) is glued (Loctite 406) over plastic tubing inner bladder when inflated. Therefore, material selection is simplified
flanges (Carmo, type 09–219) to the circular openings in the actuators. and a broad range of materials can be used for the outer shell. Note that
The final actuator triplet is shown in Fig. 5 with one actuator inflated we use the same processed fabric for the outer shell as for the actuator
to different elongation lengths. for the sake of simplicity. The first step for fabricating the outer shell is
to weld Velcro straps (Fasttech, 20mm) to fix the actuators to the pro-
2.2.2. Link fabrication cessed fabric layers of both links. Then the outer shell is glued (Loctite
The links consist of a two shell design. The outer fabric shell is 435) to the triangular 3D printed end pieces. The longitudinal seam is
made from the processed fabric (poplin-film-poplin) and the inner air glued for both links. For the movable link the inner bladder has to be
tight bladder is made from two layers of polyurethane film (Rivertex, inserted first. The different components are illustrated in Fig. 4.
0.3 mm polyurethane film) welded together with the heat press. A sheet
of parchment paper shaped to fit the inner contour is put inbetween the 2.2.3. Pneumatic peripherals
film layers to avoid welding over the full area and to restrict it to the The pressure in the actuators is controlled with fast switching valves
seam. The parchment paper can be removed after welding through the (Festo, MHJ10). These valves only have a binary state, namely fully open
Fig. 7. Exemplary section of the layered actuator composition. The TPU films
are denoted by S1 and S2, while the fabric pieces are denoted by F1 and F2.
Two layers of TPU film are used to combine two layers of fabric material. The
location of the film on the fabric material is indicated by the dashed line. The Fig. 9. The left plot shows the orientation parametrization of the spherical
lower half of the figure (F1-S2-S2-F2) shows an outer seam of a single cushion, robotic arm. The pivot point is indicated by the black circle between the two
whereas the upper half (S1-S1-F1) shows a part of the inner seam connecting two links and the tip of the robotic arm by an empty circle at the end of the movable
cushions. The first step of the fabrication procedure includes the welding of two link. A positive rotation around the inertial x-axis is denoted by 𝛼 and a posi-
pieces of TPU film (S1 and S2) to both sides of a piece of fabric material (F1). In tive rotation around the inertial y-axis is denoted by 𝛽. The maximum angular
a second step, the prepared fabric piece with TPU film on both sides is welded deflection is approximately ± 45∘ when the other angle is zero. The inertial
in a bottom up process to the prepared fabric piece underneath (S2-F2-S1). z-axis is aligned with the static link of the robotic arm. The right plot shows the
actuator configuration in the inertial coordinate frame. The axis of symmetry of
actuator A is aligned with the inertial x-axis, while the orientations of the other
two actuators are obtained by rotating actuator A by ± 120∘ around the z-axis.
3.1. Orientation parametrization
The orientation of the spherical robotic arm is parametrized with two

extrinsic Euler angles 𝛼 and 𝛽. The reachable points of the tip are located
on a spherical cap, as depicted in Fig. 9. Rotations around the longitudi-
nal axis of the movable link are not considered due to the torsional stiff-
ness of the actuator triplet. Compared to other parametrizations such
as spherical coordinates or intrinsic Euler angles, the parametrization
employed simplifies the controller derivation presented in Section 4.
3.2. Stiffness characterization for different pressure levels

Fig. 8. Pneumatic diagram of a single actuator, showing the four valves used per
actuator to control its pressure. Valves 1 and 2 are used to inflate the actuator In this section we demonstrate that the stiffness of the robotic arm
and hence, are connected to source pressure, while valves 3 and 4 are used can be adjusted by changing the actuator pressures. To this end, the pres-
for deflating the actuator and are connected to the environment. The pressure sures in all three actuators are fixed to a certain level and the movable
sensor is indicated by the gauge with arrow. The same setup is employed for the
link is manually deflected along different directions in the 𝛼 − 𝛽-plane.
other two actuators.
The required force to deflect the movable link from the upright position
is measured by means of a load cell (ATI, Mini40) over the deflection
or fully closed. However, applying Pulse Width Modulation (PWM) per- range. A lateral stop is used to guide the deflection either along the
mits continuous control of the mass flow and consequently the pressure negative 𝛼-direction or the positive 𝛽-direction (compare Fig. 9). The
in the actuator. Due to the small flow capacity of the valve type em- measured force is converted into the deflection torque by multiplying
ployed, four valves in total are used per actuator. Two valves are used by the length of the movable link. The results are presented in Fig. 10.
for inflating the soft actuator and two for deflating it (see Fig. 8). The It can be seen that the stiffness is anisotropic in the 𝛼 − 𝛽-plane.
same duty cycle is set to both inflating and both deflating valves. The Using an affine approximation of the presented data and comparing
motivation for using switching valves and not commonly employed pro- the inclination of the resulting fit, we can conclude that increasing the
portional valves is that they are less expensive. A price comparison and pressure level in the actuators from ambient pressure to 1.2bar per-
an approach to address the limitations of binary valves are presented mits an increase in stiffness from 7.67𝑚𝑁 ⋅𝑚⋅𝑑𝑒𝑔 −1 to 38.75𝑚𝑁 ⋅𝑚⋅𝑑𝑒𝑔 −1
in [24]. Pressure transducers (Bürkert, 8230) provide the pressure mea- in negative 𝛼-direction (factor of 5.05) and from 2.61𝑚𝑁 ⋅𝑚⋅𝑑𝑒𝑔 −1 to
surements in each actuator and at the source. 53.14𝑚𝑁 ⋅𝑚⋅𝑑𝑒𝑔 −1 in the positive 𝛽-direction (factor of 20.34). The stiff-
ness could in principle be further increased if actuators with a higher
3. Modeling burst pressure are utilized (see [24]).
The parametrization describing the orientation of the movable link 3.3. Steady state pressure to angle mapping
is introduced in the first part of this section. Subsequently, a charac-
terization of the joint stiffness as a function of the actuator pressures Each of the three actuators expands if a pressure higher than am-
is presented. Then, a kinematic mapping from the actuator pressures to bient pressure is applied. The exact relation between internal pressure
the arm orientation is identified from data for three different joint stiff- and elongation of the actuator is not investigated, but the focus lies on
ness levels and a simple model capturing the geometry of the actuator the robotic arm orientation as a function of the three actuator pressures.
configuration is derived. Finally, the ability of the robotic arm to exert The two DoF (𝛼, 𝛽) ∈ ℝ2 are controlled by adjusting the three actuator
interaction torques is discussed. pressures (𝑝A , 𝑝B , 𝑝C ) ∈ ℝ3 . If the actuator pressures are considered as the
Fig. 10. The measured deflection torque along two directions in the 𝛼 − 𝛽-plane for a total of four different pressure levels. Deflections along the negative 𝛼-direction
are shown in the left plot and deflections along the positive 𝛽-direction are shown in the right plot, respectively. The actuators are inflated to four different pressure
levels p, (1.00, 1.05, 1.10, 1.20bar), corresponding to the four curves depicted in each plot, and the movable link is deflected from the upright position to a maximum
deflection angle of 45∘ . It can be seen in both plots that the stiffness increases with an increasing pressure level in the actuators. The deflection torque is approximately
a linear function of the deflection angle. A deflection in the positive 𝛽-direction corresponds to a deflection in the direction of actuator A, resulting in higher stiffness
values compared to a deflection in the negative 𝛼-direction, which is not directly pointing in the direction of an actuator.
inputs to the arm dynamics, the system is over-actuated. The additional C, with the pressure in actuator A set to the lower pressure level pmin ,
pressure permits control of the overall stiffness in addition to the orien- which is in this case ambient pressure. The plots for pressures B and C
tation (see Section 2.1). Different pressure combinations that yield the set to ambient pressure as well as the pressure maps for higher stiffness
same arm orientation can differ in the resulting joint stiffness. As shown levels (medium and stiff cases) look similar and are not shown. The
in the previous section, the joint stiffness increases for a higher pressure mappings presented are used in Section 4.2 to control the orientation of
level in the three actuators. Formally, this means that the mapping be- the robotic arm.
tween the pressure space and the angle space is in general not injective.
In order to derive an injective mapping from the pressure space to the 3.4. Gray-box model
angle space that can easily be inverted for control (e.g. computing the
required pressures for given desired angles) and at the same time cover While the previously presented model only captures the static behav-
different joint stiffness levels, the following procedure is applied: ior, a first order discrete-time model is derived in this section to capture
A lower pressure level denoted by pmin is chosen in the interval [1.01, the actuator configuration and possible interactions between the actua-
1.10bar] (all pressures in this work are denoted as absolute pressures). tors. The two angles of the robotic arm, (𝛼, 𝛽), form the state and changes
Different pressure combinations are applied to the actuators and the in the three pressures, (uA , uB , uC ), form the control inputs. Both angles
resulting angles are recorded. One of the three pressures is set to pmin are assumed to be directly measurable,
and the other two pressures cover all combinations of different setpoints [ ] [ ]
larger than or equal to pmin . The procedure is repeated for multiple lower 1 0 0 cos(𝜋∕6) −cos(𝜋∕6)
𝑥(𝑘 + 1) = 𝑥(𝑘) + 𝛾 𝑢(𝑘),
pressure levels pmin . Changing the minimum pressure level allows for 0 1 −1 sin(𝜋∕6) sin(𝜋∕6)
adjustment of the stiffness level of the joint. In this work, we explore ⏟⏞⏞⏟⏞⏞⏟ ⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟
three different stiffness levels, which are referred to as soft, medium =𝐴d =𝐵d
and stiff. The lower pressure level pmin is set to ambient pressure for the [ ]
𝛼(𝑘) ⎡𝑢A (𝑘)⎤
soft setting, to 1.05bar for the medium setting and to 1.10bar for the 𝑥(𝑘) = , 𝑢(𝑘) = ⎢ 𝑢B (𝑘) ⎥ (2)
stiff setting. Furthermore, only steady state conditions are considered 𝛽(𝑘) ⎢ ⎥
⎣ 𝑢C (𝑘) ⎦
for the derivation of the mapping. The presented procedure can be ex-
[ ]
tended to an arbitrary number of stiffness levels. The resulting coverage 1 0
of the pressure and angle space is depicted in Fig. 11 for the soft and 𝑦(𝑘) = 𝑥(𝑘). (3)
0 1
stiff settings, respectively. The results for the medium stiffness level are ⏟⏞⏞⏟⏞⏞⏟
similar and not shown for the sake of better visibility. =𝐶d
For each stiffness case considered, both angles 𝛼 and 𝛽 are individ-
ually fitted to the pressures by a third order polynomial fit. The static The Bd matrix incorporates the geometry of the actuator configu-
mapping can be expressed as, ration premultiplied by the unknown parameter 𝛾. The entries of the
Bd matrix represent how each actuator affects each of the two degrees
(𝛼, 𝛽) = 𝑃 𝑀(𝑝A , 𝑝B , 𝑝C ), (1) of freedom (compare Fig. 9). The Bd matrix is derived for the movable
with PM denoting the third order polynomial model. An example for a link being in an upright orientation. Hence, the model is only valid in a
resulting fit is presented in Fig. 12 for the soft stiffness level. Both angles neighborhood of 𝛼 = 0 and 𝛽 = 0. The parameter 𝛾 is identified for each
are depicted as a function of the varying pressures in actuators B and stiffness level from transient data over a setpoint jump. The data used for
Fig. 11. The left plot shows the coverage of

the pressure state space. The pressure triples
covered for the soft setting (lowest stiffness)
are shown in blue and the pressure triples cov-
ered for the stiff setting (highest stiffness) are
shown in red. For both cases the triples cov-
ered during data collection lie on three orthog-
onal planes (indicated by the light blue and red
planes with black intersection lines) with one
pressure corresponding to pmin . The data for the
medium stiffness case is omitted for better visi-
bility and all data shown is down-sampled by a
factor of ten to give better visibility. The right
plot shows the resulting angles. The envelope
of the attainable angles is approximately circu-
lar. The area of attainable angles is significantly
smaller for the stiff case (in red), because the
pressures are limited by the maximum possible
pressure of the actuators. Note that the 𝛼 and
𝛽-axes are inverted to match the coordinate sys-
tem used throughout this work. (For interpre-
tation of the references to colour in this figure
legend, the reader is referred to the web version of this article.)
Fig. 12. The figure illustrates the pressure to angle map, PM, for the soft stiffness setting. The steady state angles 𝛼 (left plot) and 𝛽 (right plot) are shown as a
function of different pressure combinations for pB and pC . For the data shown in these plots, the pressure pA is set to ambient pressure. The measured data is shown
in black and the colored surface displays the third order polynomial fit identified. A low pressure in actuator B and a high pressure in actuator C results in a large
negative value for 𝛼, while a high pressure in actuator B and a low pressure in C results in a large positive value for 𝛼. The angle 𝛽 has largely negative values
when the pressures in both actuators B and C are close to ambient pressure. Increasing pressures in actuators B and C causes the angle 𝛽 to saturate at large positive
values (caused by the rigid joint). When pressures B and C are close to ambient pressure (as is the case for pressure A), the arm shows a negative deflection in the
𝛽-direction. This behavior at ambient conditions is caused by fabrication imperfections. Note that a slight increase in pressures B and C moves the arm to an upright
orientation.
the identification is the same as displayed in Fig. 16. The resulting val- The force exerted at the tip of the movable link is measured in its nor-
ues are 𝛾soft = 1.55𝑟𝑎𝑑∕𝑏𝑎𝑟, 𝛾med = 0.70𝑟𝑎𝑑∕𝑏𝑎𝑟 and 𝛾stiff = 0.35𝑟𝑎𝑑∕𝑏𝑎𝑟. mal direction for different orientations and pressures. Without claiming
The resulting fit for parameter 𝛾 is smaller for higher stiffness values, completeness, we investigate the applied force in one direction in the
which is due to the higher pressures required for a certain angular de- 𝛼 − 𝛽-plane, namely in direction of actuator B (compare Fig. 9). The in-
flection. The first order model does not include any spring forces of the teraction force is measured at three different positions in the actuation
actuators, damping or gravitational forces. The model will be used in direction of actuator B. The orientation in this direction is denoted by
the learning scheme presented in Section 4.3. the angle 𝜑B , with the positive direction corresponding to an expansion
of actuator B. The forces are measured for 𝜑B ∈ {−40◦ , 0◦ , 40◦ } and five
3.5. Interaction torque analysis different pressures in the range of [1.01, 1.20bar]. The orientations cor-
responding to 𝜑B = ±40◦ are chosen to be close to the limits of the actu-
In the final section of this chapter, we investigate the capability of the ation range (45∘ ) while avoiding boundary effects. A load cell is placed
robotic arm to apply forces during static interactions with the environ- at the corresponding location and the force in the normal direction of
ment. The investigation presented in this section should be understood the movable link is measured in the positive 𝜑B -direction (expanding
as a characterization of the system developed. The control approach pre- actuator). The measured force is converted into a torque by multiplying
sented in the following chapter does not build on the results presented with the known length of the movable link. The results are shown in
here. Fig. 13.
pressure controller is presented independently of the actuator and im-

plemented identically for all three actuators. The switching valves can
only be controlled to be fully open or fully closed. Continuous control
of the mass flow, and correspondingly the pressure, is achieved by ap-
plying pulse width modulation. The static mass flow of the fully opened
valve, 𝑚̇ fo , is described by the ISO 6358 standard model [33],
√
⎧ (𝑝 )2
√
⎪𝑝 𝑆𝜌 𝑇 ∕𝑇 𝑝u
d −𝑏
𝑝
𝑚̇ fo (𝑝u , 𝑝d ) = ⎨ u 0 0 u 1 − (1−𝑏)2
if 𝑝d > 𝑏 (4)
√
⎪𝑝 𝑆𝜌 𝑇 ∕𝑇 𝑝d
u
⎩ u 0 0 u if 𝑝
≤ 𝑏,
u
where pu and pd stand for up and downstream pressure, the parameters

𝜌0 , T0 denote density and temperature at the ISO 6358 reference state,
b the critical pressure ratio determining a choked flow condition and S
the sonic conductance. Both parameters, b and S are provided by the
valve manufacturer.
When applying a certain duty cycle, dc, the actual mass flow through
a single valve, 𝑚̇ , is computed as the product of the duty cycle and the
mass flow of the fully opened valve,
Fig. 13. Measured torques exerted by the robotic arm as a function of the pres-
𝑚̇ = 𝑑𝑐 ⋅ 𝑚̇ fo . (5)
sure in actuator B and the orientation of the arm denoted by 𝜑B . The three
curves denote different expansion lengths of actuator B. The highest torques are Since two different valves are used for inflation and deflation, the above
achieved if the actuator expansion is minimal (green, solid curve, 𝜑B = −40◦ ) model is applied to both valves, with the up- and downstream pressure
followed by the movable link being in the upright position (blue, dashed curve, being source pressure and actuator pressure for the inflating valve, and
𝜑B = 0◦ ) and the smallest torques are measured for actuator B expanded maxi-
actuator pressure and ambient pressure for the deflating valve. A posi-
mally (red, dash-dotted curve, 𝜑B = 40◦ ). The maximum torque of each curve is
tive 𝑚̇ corresponds to a mass flow of the inflating valve and a negative
achieved for a pressure of 1.2bar and corresponds to normal forces at the tip of
the movable link of 9.8N,6.3N and 2.7N for the three orientations considered.
mass flow to the deflating valve.
From the current pressure p and its setpoint pSP , the desired mass
flow is computed as
It can be seen that the torque approximately follows a linear trend ( )
in the actuator pressure. The torque is considerably higher when the 𝑚̇ = 𝑉 ∕(𝑅0 𝑇0 ) ⋅ 𝑘P (𝑝SP − 𝑝) + 𝑘I (𝑝SP − 𝑝) 𝑑𝑡 , (6)
∫
actuator is expanded the least compared to the case when the actuator
is almost fully expanded. where V denotes the average volume of an actuator, R0 the ideal gas
constant and T0 the temperature at the ISO 6358 reference state. The
constant term in (6) is multiplied by the sum of a proportional and an
4. Control integral term in the pressure error. The proportional and integral con-
troller gains are referred to as 𝑘P [𝑠−1 ] and 𝑘I [𝑠−2 ], respectively. The
In this section we present the control approach applied to the spher- equation is derived by taking the time derivative of the ideal gas law and
ical robotic arm. The cascaded control architecture is decomposed into neglecting changes in the volume and temperature. The time derivative
three inner loops to control the faster pressure dynamics and an outer of the pressure is then replaced by the proportional-integral term (note
loop for the slower angular dynamics (see Fig. 14). The pressure control the units of the proportional and integral gains).
is presented in the first part of the section and the angle control in the A gain scheduling approach is employed for the proportional gain kP
second part. Finally, a learning control scheme is presented to reduce as a function of the absolute pressure error,
unwanted interactions of the actuators and further improve the control ( )
performance. 𝑘P = 1 − exp(−|𝑝SP − 𝑝|∕𝑒63 ) ⋅ 𝑘̄ P (7)
where 𝑘̄ P denotes the nominal proportional gain reached for infinite

4.1. Pressure control pressure errors and e63 the error when kP takes 63.2% of its maxi-
mum value. The gain scheduling approach ensures good tracking per-
The pressure controller used in this work is closely related to the formance, while reducing valve chatter for small pressure deviations. A
approach presented in [24], with several aspects being improved. The dead band occurs for small duty cycles, because the opening time of the
Fig. 14. The cascaded control architecture

used to control the spherical robotic arm. The
actuator pressures are controlled in three in-
dependent inner loops. Based on the current
angles and angular setpoints, the position con-
troller computes the pressure setpoints (the an-
gular setpoints directly used in the position
controller for feed forward control action are
not shown for the sake of clarity). An itera-
tive learning scheme is applied in parallel with
the position controller to account for repetitive
disturbances and improve the tracking perfor-
mance. The signals 𝑢𝑗A,B,C are defined over a full
iteration j. At each time step k, the correspond-
ing element is added to the pressure setpoints
of the position controller.
Fig. 15. The inverse mapping from desired angles to required pressures for the soft stiffness case. The left plot shows pressure A required to attain a certain (𝛼, 𝛽)
combination. The middle and right plots show the same for pressures B and C. The inverse mapping is only shown for the feasible values of 𝛼 and 𝛽 that are in the
actuation range of the robotic arm. Note that the surfaces of the three actuator pressures reflect the symmetry of the actuator configuration employed. A rotation of
the surface for pressure A by 120∘ along the positive vertical direction approximately results in the surface for pressure B and a rotation along the negative vertical
direction approximately results in the surface for pressure C. This corresponds to the actuator configuration shown in Fig. 9. The inverse mappings for the medium
and stiff case look similar and are not shown for the sake of brevity. The main difference is that the smallest attainable pressures are higher than ambient pressure
as is the case for the soft mode.
valve nozzle is too short for a steady mass flow to build up. This dead namely,
band was identified in [24] and is slightly refined for this work. A duty
cycle below 𝑑𝑐min = 0.05 is set to zero to avoid operating the valves when 𝑢𝛼 = 𝑘P𝛼 (𝛼SP − 𝛼) + 𝑘I𝛼 (𝛼SP − 𝛼) 𝑑𝑡 (9)
∫
no flow is possible and the valves are prone to overheating. Finally, the
duty cycles of the inflating and deflating valves are computed by divid-
ing the desired mass flow in (6) by 𝑚̇ fo of the inflating valve, if 𝑚̇ is 𝑢𝛽 = 𝑘P𝛽 (𝛽SP − 𝛽) + 𝑘I𝛽 (𝛽SP − 𝛽) 𝑑𝑡, (10)
∫
positive or by 𝑚̇ fo of the deflating valve if 𝑚̇ is negative. For a positive
mass flow 𝑚̇ , the duty cycle of the deflating valve is set to zero and vice where 𝛼 SP , 𝛽 SP denote the angular setpoints, u𝛼 , u𝛽 the setpoint adjust-
versa. ments and 𝑘P𝛼 [−], 𝑘I𝛼 [𝑠−1 ], 𝑘P𝛽 [−] and 𝑘I𝛽 [𝑠−1 ] the proportional and
integral gains for 𝛼 and 𝛽, respectively. Different controller gains are
4.2. Position control used in 𝛼 and 𝛽-direction to account for the differing control authority
in the two directions. Adding the outputs of the PI controllers to the
The position controller exploits the steady state pressure to angle original angular setpoints and applying the inverse mapping yields the
mapping presented in Section 3.3 by using the numerically inverted pressure setpoints for the three actuators,
mapping from desired angles to required pressures. The approach pre- (𝑝ASP , 𝑝BSP , 𝑝CSP ) = 𝑃 𝑀 −1 (𝛼SP + 𝑢𝛼 , 𝛽SP + 𝑢𝛽 ). (11)
sented in the following is applied for all three stiffness settings (soft,
medium, stiff). First, the fitted third order polynomial model is evalu- Note that the inverse mapping in principle decouples the angles in the
ated for an evenly spaced set of pressures to obtain the corresponding sense that changing 𝛼, while keeping 𝛽 constant yields a new set of pres-
model angles. Thereby, one of the three pressures is again set to pmin , as sure setpoints that do not change 𝛽. However, this is only valid in steady
was the case during the data collection of the identified mapping. For a state conditions and up to the modeling accuracy of the pressure to angle
grid of evenly spaced angles, the closest model angle is determined and mapping. A learning approach to compensate for possible interactions
the corresponding pressures are stored in a look-up table. The reason between the angles is presented in the next section.
why we use the third order polynomial fit to compute the model angles
and not the raw angular data directly is to obtain a smooth mapping 4.3. Learning control
from the desired angles to the required pressures. The inverse mapping
obtained for the soft case is illustrated in Fig. 15, and used to compute An Iterative Learning Control (ILC) approach is applied in parallel
the pressure setpoints for certain angular setpoints, with the previously presented cascaded control approach. This is to com-
pensate on the one hand for interactions between the two angles and on
(𝑝ASP , 𝑝BSP , 𝑝CSP ) = 𝑃 𝑀 −1 (𝛼SP , 𝛽SP ), (8)
the other hand for unmodeled effects such as hysteresis behavior due to
with 𝑃 𝑀 −1 denoting the inverse mapping. It consists of a separate look- the material employed as well as joint friction. ILC is applied to improve
up table for each pressure as a function of the two desired angles. Bilin- tracking performance for repetitive tasks. The general idea is to record
ear interpolation is applied for the pressures given the desired angles to the tracking error in one iteration and use it to improve the performance
improve accuracy. in the following iteration.
Relying purely on the inverse mapping in a feed forward sense leads The linear model of the system, presented in Section 3.4, is used to
to poor control performance due to the static nature of the mapping predict the behavior of the system in the next iteration. The correction
and neglected dynamical effects such as hysteresis and joint friction. signal to compensate for the repetitive disturbance is computed as the
Therefore, the current angles and angular setpoints are first fed into solution of a quadratic optimization problem. This approach is referred
two decoupled PI controllers that compute adjusted angular setpoints, to as Norm Optimal Iterative Learning Control (NOILC; see [34]). The
NOILC scheme has previously been applied to a SISO system (see [13]) iteration is defined by the time duration of a single trajectory and the
and is extended in this work to the MIMO case. sampling time of the learning scheme, which is denoted by TILC . Based
While ILC can learn to compensate for repetitive disturbances, it is on the model introduced in (3), the lifted system representation is used,
not able to compensate for non-repetitive disturbances. Therefore, the with
learning approach is applied in parallel with the feedback controller
presented in Section 4.2 to account for non-repetitive disturbances (as ⎡ 𝐶d 𝐵d 0 0 0 ⎤
outlined in [35]). ⎢ ⎥
𝑃 =⎢ ⋮ 𝐶d 𝐵d 0 ⎥ ∈ ℝ2𝑁×3𝑁 , (17)
We introduce the following notation, where the index j denotes the
⎢𝐶 𝐴𝑁−1 𝐵 ⋯ 𝐶d 𝐴d 𝐵d 𝐶d 𝐵d ⎥⎦
iteration number and N the number of time steps of a single iteration, ⎣ d d d
( )
𝑦𝑗 ∶= 𝛼 𝑗 (0), 𝛽 𝑗 (0), … , 𝛼 𝑗 (𝑁 − 1), 𝛽 𝑗 (𝑁 − 1) ∈ ℝ2𝑁 (12)
( ) to write the linear dynamics over a single iteration as,
𝑗 𝑗 𝑗 𝑗
𝑦SP ∶= 𝛼SP (0), 𝛽SP (0), … , 𝛼SP (𝑁 − 1), 𝛽SP (𝑁 − 1) ∈ ℝ2𝑁 (13)
( )
𝑗 𝑗 𝑗 𝑗
𝑒𝑗 ∶= 𝑒𝛼 (0), 𝑒𝛽 (0), … , 𝑒𝛼 (𝑁 − 1), 𝑒𝛽 (𝑁 − 1) ∈ ℝ2𝑁 (14) 𝑦𝑗 = 𝑃 𝑢𝑗 + 𝑑
( )
𝑗 𝑗 𝑗 𝑗
𝑑 ∶= 𝑑𝛼 (0), 𝑑𝛽 (0), … , 𝑑𝛼 (𝑁 − 1), 𝑑𝛽 (𝑁) ∈ ℝ 2 𝑁
(15) 𝑒𝑗 = 𝑦SP − 𝑦𝑗 = 𝑦SP − 𝑃 𝑢𝑗 − 𝑑
𝑗+1 (18)
𝑒 = 𝑦SP − 𝑃 𝑢𝑗+1 − 𝑑
( ) = 𝑒𝑗 − 𝑃 (𝑢𝑗+1 − 𝑢𝑗 ).
𝑢𝑗 ∶= 𝑢𝑗A (0), 𝑢𝑗B (0), 𝑢𝑗C (0), … , 𝑢𝑗A (𝑁 − 1), 𝑢𝑗B (𝑁 − 1), 𝑢𝑗C (𝑁 − 1) ∈ ℝ3𝑁 ,
(16)
The initial state of the system is treated as a repetitive disturbance. The
𝑗 correction input is computed by minimizing a quadratic cost function
where 𝛼 j (k) and 𝛼SP (𝑘) denote the angle and its setpoint in iteration j and
in the predicted tracking error of the next iteration and the current and
at time step k, 𝑒𝛼 (𝑘) the error in 𝛼 and 𝑑𝛼𝑗 (𝑘) the disturbance in 𝛼 during
𝑗
next correction input,
iteration j and at time step k. The definitions apply accordingly for angle
𝛽. The pressure compensation signals are denoted by (𝑢𝑗A (𝑘), 𝑢𝑗B (𝑘), 𝑢𝑗C (𝑘))
for iteration j and time step k. The angular setpoints ySP are the same 1 𝑗+1𝖳 𝖳
𝐽 (𝑢𝑗+1 ) = [𝑒 𝑀𝑒𝑗+1 + (𝑢𝑗+1 − 𝑢𝑗 )𝖳 𝑆(𝑢𝑗+1 − 𝑢𝑗 ) + 𝑢𝑗+1 𝐷𝖳 𝑊 𝐷𝑢𝑗+1 ],
for all iterations and the disturbances in both angles are assumed to 2
be repetitive between iterations. The number of time steps for a single (19)
Fig. 16. Comparison of the tracking performance for three different joint stiffness values. The top two plots show the angles 𝛼 and 𝛽 and the bottom three plots
show the pressures in actuators A,B and C. The angles and pressures for the smallest stiffness level (soft) are shown in blue (dotted), for the middle stiffness level
(medium) in green (dash-dotted) and for the highest stiffness value (stiff) in red (solid). The minimum pressure levels corresponding to the three stiffness levels are
indicated by the dashed line in the pressure plots for each stiffness level. The tracking performances for the low and medium stiffness modes are comparable. The
RMS tracking error of the stiff mode is reduced by roughly 30% with respect to the one from the soft mode. The position controller for each stiffness setting is tuned
for best individual performance. Note that the required pressure change for a certain change in orientation is larger for a higher joint stiffness.
where M, S, W are positive semi-definite cost matrices of appropriate + (𝑃 𝖳 𝑀𝑃 + 𝑆 + 𝐷𝖳 𝑊 𝐷)−1 𝑃 𝖳 𝑀𝑒𝑗 . (21)
dimensions and
The inverse in (21) is guaranteed to exist if S is chosen to be positive
⎡−𝐼3 𝐼3 ⎤ definite. The correction inputs 𝑢𝑗A (𝑘), 𝑢𝑗B (𝑘), 𝑢𝑗C (𝑘) are added to the pres-
⎢ ⎥
1 ⎢ ⋱ ⎥ ∈ (ℝ)3𝑁×3𝑁 sure setpoints computed by the inverse mapping in (11) during iteration
𝐷=
𝑇ILC ⎢ −𝐼3 𝐼3 ⎥ j for every time step k.
⎢ 03 03 ⎥⎦
⎣
5. Results
⎡1 0 0⎤ ⎡0 0 0⎤
⎢ ⎥ ⎢ ⎥
𝐼3 = ⎢0 1 0⎥ , 03 = ⎢0 0 0⎥ (20) Results from the experimental evaluation of the spherical robotic
⎢ ⎥ ⎢ ⎥ arm and the proposed controller are presented in this section. The
⎣0 0 1⎦ ⎣0 0 0⎦
control performance for different stiffness levels is discussed in
is an approximation of the derivative of the correction signal, by the Section 5.1 and the results of the position tracking experiment for the
first order forward numerical differentiation scheme. The first term in soft mode after applying a learning scheme are presented in Section 5.2.
(19) penalizes the predicted tracking error in the next iteration, the sec- The control algorithms are implemented in C++ and are executed
ond term adds a cost to changes between the correction signal of the on a laptop computer (Intel Core i7 CPU, 2.8GHz). The valves and pres-
current and the next iteration and the last term penalizes the deriva- sure sensors are interfaced over a Labjack T7 Pro-device. The pressure
tive of the correction signal in the next iteration. Adding the last two controllers are executed at 200Hz, the position controller at 50Hz and
terms improves the transient learning behavior by avoiding excessive the NOILC scheme has a sampling time 𝑇ILC = 1∕50𝑠. The communica-
changes between the iterations (second term) and by disabling learning tion to the hardware, data logging and the inner and outer control loops
at high frequencies (last term), which is a common practice in ILC and are implemented in separate threads for better modularity and perfor-
can also be addressed by using a so-called Q filter (see [35]). The op- mance. The source pressure is set to 2bar and the actuators are oper-
timization problem (19) does not include constraints on the correction ated in a pressure range of [1.0, 1.25bar]. Both links are inflated to a
inputs, such that the optimal solution for the control inputs can be com- pressure of 1.25bar. The switching valves are operated at a PWM fre-
puted in closed form. The feasibility of the resulting correction inputs quency of 200Hz. The angles of the movable link, (𝛼, 𝛽), are retrieved
is ensured by limiting them to the feasible range in a second step. The from position data provided by an infrared motion capture system run-
optimal solution is computed as, ning at 200Hz and providing sub-millimeter accuracy. The parameters
of the pressure controller are chosen as 𝑘̄ P = 40𝑠−1 for the soft mode and
⋆
𝑢𝑗+1 = argmin 𝐽 (𝑢𝑗+1 )
𝑢𝑗+1 𝑘̄ P = 20𝑠−1 for the medium and stiff modes (see (7)). The gain scheduling
= (𝑃 𝖳 𝑀𝑃 + 𝑆 + 𝐷𝖳 𝑊 𝐷)−1 (𝑃 𝖳 𝑀𝑃 + 𝑆)𝑢𝑗 parameter is set to 𝑒63 = 0.005𝑏𝑎𝑟 and the integral gain to 𝑘I = 120𝑠−2 for
Fig. 17. Results of the spherical robotic arm tracking a square trajectory. The top plot shows the angle 𝛼, the second plot the angle 𝛽 and the bottom three plots show
the pressures in actuators A, B and C. Two iterations of the learning scheme are shown. The green (dash-dotted) curves show the angles and pressures during iteration
zero (no learning) and the blue (solid) curves show the variables after applying the learning scheme for 48 iteration. For the top two plots, the red (dashed) curves
indicate the angular setpoints (same for all iterations). During iterations zero, the tracking performance in 𝛽 is clearly affected by a setpoint jump in 𝛼. The learning
scheme can compensate for the interaction and improve tracking performance. In the bottom three plots, the pressure setpoint is depicted by the red (dashed) curve
for iteration 48. It can be seen in the plots that the setpoints are in general well-tracked, except during the deflation phases, when the setpoint is close to ambient
pressure. Slight valve chatter occurs and is noticeable by the spikes in the measured pressures.
Table 1
Tuning parameters of the position controller for different setpoint jumps
and stiffness modes. A setpoint jump of 20∘ corresponds to the experimen-
tal results presented in Section 5.1 and a setpoint jump of 40∘ to the results
presented in Section 5.2.
setpoint jump [∘ ] stiffness 𝑘P𝛼 [ − ] 𝑘I𝛼 [𝑠−1 ] 𝑘P𝛽 [ − ] 𝑘I𝛽 [𝑠−1 ]
20 soft 0.4 0.05 0.8 0.4

20 medium 0.05 0.05 0.1 0.05
20 stiff 0.05 0.05 0.03 0.05
40 soft 0.3 0.05 0.1 0.05
all three modes. The PI gains of the position controller are chosen for
best individual performance and are summarized in Table 1. The tuning
matrices of the learning control approach are chosen as 𝑀 = 0.01 ⋅ 𝐼,
𝑆 = 5 ⋅ 𝐼 and 𝑊 = 0.002 ⋅ 𝐼, where I denotes the identity matrix of ap-
propriate dimensions.
5.1. Controller evaluation for different stiffness modes
In this section we present the results from a position tracking exper-

iment for different joint stiffness modes. Three controllers are designed
as outlined in Section 4.2 for three different stiffness modes denoted as
soft, medium and stiff. The resulting tracking performances for a square- Fig. 18. Results of the square tracking experiment depicted in the angular
shaped trajectory with setpoint jumps of 20∘ ( ± 10∘ ) are shown for two space. The trajectory consists of setpoint jumps of 40∘ and is tracked in a counter-
consecutive iterations in Fig. 16. clockwise direction. The tracking performance of iteration zero (no learning,
pure feedback control) is shown by the green (dash-dotted) line and the im-
The tracking performance increases for higher joint stiffness levels.
proved tracking performance of iteration 48 by the blue (solid) line. The set-
The behavior for the stiff mode is generally more repetitive. The higher
point is depicted by the red (dashed) lines. The interactions between the two
joint stiffness might dominate other less repetitive dynamical behaviors angles can be clearly seen in the top left corner of the square. When the arm
such as joint friction and nonlinear material behavior. An approach to moves in a positive 𝛼-direction, it is pushed into actuator C, causing the arm to
improve the tracking performance for the soft mode is presented in the deflect into a positive 𝛽-direction. The learning scheme can compensate for this
next section. effect and leads to a better overall tracking performance. Note that the 𝛼 and
𝛽-axes are inverted to match the coordinate system used throughout this work.
5.2. Learning control
In this section, we investigate the applicability of the learning

scheme to improve the tracking performance for the soft mode. The mo- C that causes the arm to deflect into the positive 𝛽-direction. The learn-
tivation for considering this stiffness mode is twofold. Firstly, the system ing scheme adjusts for this error by lowering the setpoint for pressure
presented is designed to have high joint compliance. Therefore, we ar- C and thereby increasing the speed of deflation. At the same time the
gue that it is beneficial to employ control approaches which preserve pressure in actuator A is slightly increased compared to iteration zero
the inherent compliance and do not simply alter the system dynamics and is increased in actuator B more distinctly to push into the positive
to improve tracking performance. A detailed discussion of the trade off 𝛼-direction and compensate for the deflection in the 𝛽-direction. This
between compliance and control performance is beyond the scope of transient behavior is clearly not captured by the feedback controller,
this work and the reader is referred to [36]. Secondly, from the results which is based on steady state information of the system. However, the
presented in Section 5.1, the soft mode is the most challenging mode to learning scheme relying on model knowledge of the actuator configura-
achieve a high tracking performance. If the applicability of the learning tion is able to improve the tracking performance.
scheme can be demonstrated for this mode, it is likely to also improve However, perfect tracking performance is not achieved in the final
the higher joint stiffness modes. iteration that can be seen in Fig. 18 by the non-zero tracking errors in
The tracking performance of the spherical robotic arm is evaluated both 𝛼 and 𝛽 during the last iteration. This behavior can be ascribed to
on a square-shaped trajectory including smooth setpoint jumps of 40∘ in two reasons. Firstly, non-repetitive disturbances probably occur due to
both angles (with continuous angular velocity profiles) and a duration of the soft materials employed and the interplay of the multi-threaded, cas-
10s. The results of the tracking experiment are shown in Fig. 17 as a time caded control loops. Since, the learning scheme can only address repeti-
series and in Fig. 18 in the angular space. The reader is referred to the tive disturbances, it can not compensate for such non-repetitive effects.
video attachment to gain an impression of the experiments conducted. Secondly, the objective not only includes the tracking error as a cost, but
From Fig. 17 it can be seen that in iteration zero of the learning also two terms to improve the learning behavior. The cost on the deriva-
scheme (no learning applied), an angle jump in 𝛼 causes a significant tive of the correction signal (third term in (19)) leads to a suppression
error in 𝛽. The opposite is not the case, which can be explained by the of high frequency content in the correction input. The effect can be seen
actuator arrangement shown in Fig. 9. An angle change in 𝛽 occurs along in Fig. 17 by the slightly smoothed angular trajectories relative to the
the symmetry line of the actuator configuration (namely the x-axis). setpoint trajectories.
Therefore, the angle 𝛼 is less affected by a change in 𝛽. However, a Simpler learning schemes such as PD-type ILC (see [35]) have also
setpoint change in the 𝛼-direction does not occur along a symmetry line been investigated. No model knowledge was incorporated with these
of the actuator configuration and is therefore more likely to impair the methods and the scheme has been applied in a serial configuration with
tracking performance in the 𝛽-direction. the position controller adjusting the angular setpoints. Compared to the
A positive jump in the 𝛼-direction is commanded at 𝑡 = 6𝑠. There- NOILC approach, the improvement of the tracking performance with the
fore, the pressures in actuators A and C are reduced and the pressure in PD-type ILC was very limited. A correction in 𝛼 typically impaired the
B is increased. The robotic arm is pushed against the deflating actuator performance in 𝛽 and vice versa due to the lack of model knowledge.
The inferior pressure tracking performance during deflation close different directions. This information could then be incorporated in the
to ambient pressure can be explained by the small pressure difference position controller for the realization of interactive applications.
over the valve and the correspondingly small mass flow through the Finally, the system proposed in this work seems promising for the
deflating valve (compare (4)). In addition, the limited flow capacity of investigation and realization of specific applications (e.g. pick and place
the switching valves employed further limits the mass flow. applications), which will be the subject of future work.
6. Conclusion Declaration of Competing Interest
This paper presented the design of an inflatable spherical robotic The authors declare that they have no known competing financial
arm, relying on a compact joint configuration. Only three inflatable ac- interests or personal relationships that could have appeared to influence
tuators that are arranged symmetrically around a spherical joint are re- the work reported in this paper.
quired to control the two DoF of the robotic arm. A detailed fabrication
procedure has been presented with the hope of encouraging the design CRediT authorship contribution statement
and development of similar fabric-based soft robotic systems. The sys-
tem kinematics and dynamics are identified by two models. The static Matthias Hofer: Conceptualization, Methodology, Software, Valida-
mapping from the actuator pressures to the orientation of the arm is tion, Formal analysis, Investigation, Resources, Data curation, Writing
identified for three different stiffness levels corresponding to three dif- - original draft, Writing - review & editing, Visualization, Project ad-
ferent lower actuator pressure levels. The inverse mapping is employed ministration, Funding acquisition. Raffaello D’Andrea: Conceptualiza-
in three separate position controllers to determine the actuator pres- tion, Methodology, Resources, Writing - review & editing, Supervision,
sures for a certain arm orientation. The tracking performance for each Project administration, Funding acquisition.
of the three stiffness levels is evaluated. The best tracking performance is
achieved for the highest joint stiffness level. A dynamic gray-box model
Acknowlgedgments
capturing the interactions of the two DoF for the given actuator con-
figuration is developed and used in a learning control approach. The
The authors would like to thank Michael Egli, Daniel Wagner and
learning scheme is applied to the lowest stiffness mode and is able to
Matthias Müller for their contribution to the development of the proto-
compensate for the interactions between the two DoF of the lightweight
type.
and compliant system. The results show that satisfactory control perfor-
mance can be demonstrated on a square trajectory serving as a proof of
concept for the actuator configuration proposed. Supplementary material
While the design presented is modular due to the rigid connections
between the links and the actuator triplet, it can be optimized with re- Supplementary material associated with this article can be found, in
spect to the joint part and the maximum actuator burst pressure. The the online version, at doi:10.1016/j.mechatronics.2020.102369.
rigid ball-and-socket joint induces joint friction and has a limited range
that constrains the actuation range of the robotic arm. Using soft materi- References
als for the joint and also the rigid end parts of the links could increase the
[1] Gaiser I, Wiegand R, Ivlev O, Andrés A, Breitwieser H, Schulz S, et al. Compliant
actuation range of the robotic arm and would further reduce its weight. robotics and automation with flexible fluidic actuators and inflatable structures. In-
A higher maximum burst pressure of the actuators would allow to inves- Tech; 2012. doi:10.5772/51866.
tigate higher stiffness levels of the joint and also increase the maximum [2] Polygerinos P, Correll N, Morin SA, Mosadegh B, Onal CD, Petersen K, et al.
Soft robotics: review of fluid-driven intrinsically soft devices; manufacturing,
torque the robotic arm can exert on its environment. However, it is im- sensing, control, and applications in human-robot interaction. Adv Eng Mater
portant to keep in mind that the limited burst pressure of the actuators 2017;19(12):1700016. doi:10.1002/adem.201700016.
employed in this work is a consequence of the simple manufacturing [3] Gorissen B, Reynaerts D, Konishi S, Yoshida K, Kim J-W, De Volder M. Elastic in-
flatable actuators for soft robotic applications. Adv Mater 2017;29(43):1604977.
method. Using more elaborate methods (e.g. [24]) could compromise doi:10.1002/adma.201604977.
the simplicity and availability of the manufacturing method. [4] Sanan S, Moidel JB, Atkeson CG. Robots with inflatable links. In: IEEE/RSJ
The number of rotational degrees of freedom of the spherical robotic International Conference on Intelligent Robots and Systems; 2009. p. 4331–6.
doi:10.1109/IROS.2009.5354151.
arm is currently limited to two. However, multiple joints of the type pre-
[5] Sanan S, Ornstein MH, Atkeson CG. Physical human interaction for an inflatable ma-
sented could be serially connected to form a higher dimensional system. nipulator. In: Annual International Conference of the IEEE Engineering in Medicine
The ability of a single joint to adjust its stiffness is a promising prerequi- and Biology Society; 2011. p. 7401–4. doi:10.1109/IEMBS.2011.6091723.
site to support a payload and also limit possible, undesirable interactions [6] Qi R, Khajepour A, Melek WW, Lam TL, Xu Y. Design, kinematics, and control
of a multijoint soft inflatable arm for human-safe interaction. IEEE Trans Rob
of multiple, serially connected joints. 2017;33(3):594–609. doi:10.1109/TRO.2016.2647231.
The current control approach relies on the application of a non- [7] Althoefer K. Antagonistic actuation and stiffness control in soft inflatable robots. Nat
causal learning scheme to improve control performance. Future work Rev Mater 2018;3(6):76–7. doi:10.1038/s41578-018-0004-0.
[8] Gillespie MT, Best CM, Killpack MD. Simultaneous position and stiffness control for
should address the identification of more accurate dynamic models an inflatable soft robot. In: IEEE International Conference on Robotics and Automa-
for the design of causal model-based control approaches. The gray-box tion (ICRA); 2016. p. 1095–101. doi:10.1109/ICRA.2016.7487240.
model used in this work shows promising results in combination with [9] Voisembert S, Mechbal N, Riwan A, Aoussat A. Design of a novel long-range inflat-
able robotic arm: manufacturing and numerical evaluation of the joints and actua-
the learning scheme. However, the model should be extended to be pres- tion. J Mech Robot 2013;5(4):45001. doi:10.1115/1.4025025.
sure dependent and incorporate spring forces in the actuators, damping [10] Hyatt P, Kraus D, Sherrod V, Rupert L, Day N, Killpack MD. Configuration estimation
and also account for gravity. Furthermore, the mapping identified from for accurate position control of large-scale soft robots. IEEE/ASME Trans Mechatron
2019;24(1):88–99. doi:10.1109/TMECH.2018.2878228.
actuator pressures to arm angles is only valid for a fixed lower pressure
[11] Mosadegh B, Polygerinos P, Keplinger C, Wennstedt S, Shepherd R, Gupta U,
level corresponding to a fixed stiffness level. Accordingly, the control et al. Pneumatic networks for soft robotics that actuate rapidly. Adv Funct Mater
approach presented has been demonstrated to be applicable for multi- 2014;24(15):2163–70. doi:10.1002/adfm.201303288.
[12] Thalman CM, Lam QP, Nguyen PH, Sridar S, Polygerinos P. A novel soft el-
ple, discrete stiffness levels. However, it should be investigated how the
bow exosuit to supplement bicep lifting capacity. In: 2018 IEEE/RSJ Interna-
joint stiffness can be continuously controlled. tional Conference on Intelligent Robots and Systems (IROS); 2018. p. 6965–71.
The capability of the robotic arm to apply forces during static in- doi:10.1109/IROS.2018.8594403.
teractions with the environment has been investigated along a single [13] Hofer M, Spannagl L, D’Andrea R. Iterative learning control for fast and accurate
position tracking with an articulated soft robotic arm. In: IEEE International Confer-
direction. An interesting extension would be a complete characteriza- ence on Intelligent Robots and Systems (IROS); 2019. In press. Preprint available:
tion of the possible interaction forces at different orientations and in http://arxiv.org/abs/1901.10187
[14] Ivlev O. Soft fluidic actuators of rotary type for safe physical human-machine inter- [30] Marchese AD, Katzschmann RK, Rus D. Whole arm planning for a soft and highly
action. In: IEEE International Conference on Rehabilitation Robotics; 2009. p. 1–5. compliant 2D robotic manipulator. In: IEEE/RSJ International Conference on Intel-
doi:10.1109/ICORR.2009.5209571. ligent Robots and Systems; 2014. p. 554–60. doi:10.1109/IROS.2014.6942614.
[15] Best CM, Wilson JP, Killpack MD. Control of a pneumatically actuated, fully [31] Melingui A, Lakhal O, Daachi B, Mbede JB, Merzouki R. Adaptive neural net-
inflatable, fabric-based, humanoid robot. In: IEEE-RAS International Confer- work control of a compact bionic handling arm. IEEE/ASME Trans Mechatron
ence on Humanoid Robots (Humanoids); 2015. p. 1133–40. doi:10.1109/HU- 2015;20(6):2862–75. doi:10.1109/TMECH.2015.2396114.
MANOIDS.2015.7363495. [32] Ansari Y, Manti M, Falotico E, Cianchetti M, Laschi C. Multiobjective optimization
[16] Bodily DM, Allen TF, Killpack MD. Multi-objective design optimization of a soft, for stiffness and position control in a soft robot arm module. IEEE Rob Autom Lett
pneumatic robot. In: 2017 IEEE International Conference on Robotics and Automa- 2018;3(1):108–15. doi:10.1109/LRA.2017.2734247.
tion (ICRA); 2017. p. 1864–71. doi:10.1109/ICRA.2017.7989218. [33] . Pneumatic fluid power. Determination of flow-rate characteristics of components
[17] Liang X, Cheong H, Sun Y, Guo J, Chui CK, Yeow C. Design, characterization, and using compressible fluids. Standard. International Organization for Standardization;
implementation of a two-dof fabric-based soft robotic arm. IEEE Rob Autom Lett 2013.
2018;3(3):2702–9. doi:10.1109/LRA.2018.2831723. [34] Owens D. Iterative learning control : an optimization paradigm. Springer Publishing
[18] Marchese AD, Rus D. Design, kinematics, and control of a soft spa- Company; 2016.
tial fluidic elastomer manipulator. Int J Rob Res 2016;35(7):840–69. [35] Bristow DA, Tharayil M, Alleyne AG. A survey of iterative learning control. IEEE
doi:10.1177/0278364915587925. Control Syst Mag 2006;26(3):96–114. doi:10.1109/MCS.2006.1636313.
[19] Nguyen PH, Sparks C, Nuthi SG, Vale NM, Polygerinos P. Soft poly-limbs: toward a [36] Della Santina C, Bianchi M, Grioli G, Angelini F, Catalano M, Garabini M, et al.
new paradigm of mobile manipulation for daily living tasks. Soft Rob 2019;6 1:38– Controlling soft robots: balancing feedback and feedforward elements. IEEE Robotics
53. doi:10.1089/soro.2018.0065. Automation Magazine 2017;24(3):75–83. doi:10.1109/MRA.2016.2636360.
[20] Grzesiak A, Becker R, Verl A. The bionic handling assistant: a success
story of additive manufacturing. Assembly Automation 2011;31(4):329–33. Matthias Hofer received the B.Sc. degree in Mechanical En-
doi:10.1108/01445151111172907. gineering from ETH Zurich, Zurich, Switzerland, in 2013 and
[21] Hyatt P, Wingate D, Killpack MD. Model-based control of soft actuators using the M.Sc. degree in Robotics, Systems and Control from ETH
learned non-linear discrete-time models. Frontiers in Robotics and AI 2019;6:22. Zurich, Zurich, Switzerland in 2015. He received the ETH
doi:10.3389/frobt.2019.00022. medal for his master’s thesis on approximate model predictive
[22] Liang X, Yap HK, Guo J, Yeow RCH, Sun Y, Chui CK. Design and characterization control of an aerial vehicle. During his studies, he completed
of a novel fabric-based robotic arm for future wearable robot application. In: IEEE an exchange at KTH Stockholm, Stockholm, Sweden, where
International Conference on Robotics and Biomimetics (ROBIO); 2017. p. 367–72. he focused on mathematical optimization and linear system
doi:10.1109/ROBIO.2017.8324445. theory. He is currently pursuing a doctorate degree at the In-
[23] Yang HD, Asbeck AT. A New Manufacturing Process for Soft Robots and Soft/Rigid stitute for Dynamic Systems and Control at ETH Zurich. His
Hybrid Robots. In: IEEE International Conference on Intelligent Robots and Systems main research interests include the design and development
(IROS); 2018. p. 8039–46. ISBN 9781538680940. doi:10.1109/IROS.2018.8593688. of soft robotic systems, black-box system identification, learn-
[24] Hofer M, D’Andrea R. Design, Modeling and Control of a Soft Robotic Arm. In: IEEE ing control and vision-based state estimation.
International Conference on Intelligent Robots and Systems (IROS); 2018. p. 1456–
63. ISBN 9781538680940. doi:10.1109/IROS.2018.8594221. Raffaello D’Andrea received the B.Sc. degree in Engineering
[25] Gillespie MT, Best CM, Townsend EC, Wingate D, Killpack MD. Learning nonlin- Science from the University of Toronto in 1991, and the M.S.
ear dynamic models of soft robots for model predictive control with neural net- and Ph.D. degrees in Electrical Engineering from the Califor-
works. In: IEEE International Conference on Soft Robotics (RoboSoft); 2018. p. 39– nia Institute of Technology in 1992 and 1997. He was an as-
45. doi:10.1109/ROBOSOFT.2018.8404894. sistant, and then an associate, professor at Cornell University
[26] Reinhart RF, Steil JJ. Hybrid mechanical and data-driven modeling improves from 1997 to 2007. While on leave from Cornell, from 2003
inverse kinematic control of a soft robot. Procedia Technol 2016;26:12–19. to 2007, he co-founded Kiva Systems (now Amazon Robotics),
doi:10.1016/j.protcy.2016.08.003. where he led the systems architecture, robot design, robot
[27] Shepherd RF, Ilievski F, Choi W, Morin SA, Stokes AA, Mazzeo AD, et al. Multigait navigation and coordination, and control algorithms develop-
soft robot. Proceedings of the National Academy of Sciences 2011;108(51):20400–3. ment. He is currently professor of Dynamic Systems and Con-
doi:10.1073/pnas.1116564108. trol at ETH Zurich, and chairman of the board at Verity Studios
[28] Tolley MT, Shepherd RF, Mosadegh B, Galloway KC, Wehner M, Karpel- AG.
son M, et al. A resilient, untethered soft robot. Soft Rob 2014;1(3):213–23.
doi:10.1089/soro.2014.0008.
[29] Jordan M, Pietrusky D, Mihajlov M, Ivlev O. Precise position and trajectory con-
trol of pneumatic soft-actuators for assistance robots and motion therapy devices.
In: IEEE International Conference on Rehabilitation Robotics; 2009. p. 663–8.
doi:10.1109/ICORR.2009.5209572.
Mechatronics 85 (2022) 102822
Mechatronics
journal homepage: www.elsevier.com/locate/mechatronics
Alternating pressure control system for hydraulic robots✩

Sarin Kittisares ∗, Yosiharu Hirota, Hiroyuki Nabae, Gen Endo, Koichi Suzumori
Department of Mechanical Engineering, Tokyo Institute of Technology, 2 Chome-12-1 Ookayama, Meguro, 152-8550, Tokyo, Japan
ARTICLE INFO ABSTRACT
Keywords: Hydraulics is a promising technology for robots. However, traditional hydraulic infrastructures are often large
Hydraulics and power-inefficient, with large power sources that hinder mobility. In contrast, electro-hydrostatic actuators
Robotics are relatively power efficient, but their cost and weight can be excessive in systems with a higher number of
Alternating pressure
degrees of freedom. In this paper, we propose a new alternating pressure control system for hydraulic systems
with a higher number of degrees of freedom based on an alternating pressure source system. In this system,
the valves open and close in synchronization with a pump with sensor feedback, allowing either pressure or
position in each actuator to be controlled independently. With the proposed system, a centralized pump can
be used with simplified tubing and simple on–off valves. Moreover, we developed a dynamic duty ratio system
that improves performance and reduces pump utilization time. The experimental results confirmed that both
the position and pressure of each actuator can be controlled in parallel on a multi-degree-of-freedom system.
1. Introduction In contrast, the electro-hydrostatic actuator (EHA) only needs to be

powered when actuated, leading to a higher power efficiency compared
Hydraulic actuators have been employed in heavy machinery for to traditional systems [10,12–14]. Originally developed for applications
a long time owing to their large force output, high force and power in aircrafts, the EHA combines electrical motors, pumps, accumulators,
density, and good impact resistance. These advantages make them valves, hydraulic actuators, and electrical circuits into a single unit. The
suitable for tough robotic applications [1–4], mobile robots [5–8], and EHA is a pump-controlled system, and pressurized fluid is only used for
wearable robots [9,10]. Hydraulic actuators also have a potential to be transmission between the motor and actuator. In robotic applications,
adopted in other multi-degree-of-freedom mechanical systems, which EHAs are often utilized when smaller sizes and weight-to-power ratios
are also considered as robots in the context of this paper. In addition, are essential, such as in prostheses or exoskeletons [9,14].
soft hydraulic actuators that often operate using fluid power are utilized In systems with a higher number of degrees of freedom, all EHAs
for applications such as hydraulic artificial muscles. In our earlier work, operate in parallel with each other. This is an advantage for aircraft
we developed a sit-to-stand support device based on hydraulic artificial applications when each actuator is positioned in a different part of
muscles [11]. While the device had a satisfactory force output, the hy- the aircraft since piping from a centralized supply unit is then elim-
draulic supply system and its power consumption remained a challenge. inated [12,13]. On the other hand, in robots, where actuators are
In mobile robots and wearable robots, reduced power consumption will positioned close to each other, the weight of the piping is negligible.
improve operational range and duration, which are crucial in actual However, the redundant hydraulic components of each EHA, particu-
application. larly pumps and motors, result in an excessive cost and weight of the
Traditional hydraulic infrastructure consists of a centralized pump entire system.
and tank, and the hydraulic pressure is delivered to the actuators Another alternative to the conventional centralized hydraulic sys-
through piping. The supply pressure is controlled via a pressure relief tem is the alternating pressure source with a synchronized valve sys-
valve, and the actuators are controlled with flow control valves or tem, which has been previously proposed for microactuators [15,16]. In
pressure control valves. However, it has been mentioned in the liter- this system, an alternating pressure is supplied to all actuators simulta-
ature that in conventional valve-controlled hydraulic systems, where neously. Control valves, which are directly connected to the actuators,
hydraulic pressure is supplied from a central hydraulic unit and con- open and close in synchronization with the pump. For example, to
trolled through a valve, the hydraulic unit runs continuously, even elongate the actuator, the valve opens during the pressurized period
when not active, leading to a lower efficiency [10,12–14]. and closes during the suction period.
✩ This paper was recommended for publication by Associate Editor Gianluca Palli.
∗ Corresponding author.
E-mail address: kittisares.s.aa@m.titech.ac.jp (S. Kittisares).
https://doi.org/10.1016/j.mechatronics.2022.102822
Received 2 August 2021; Received in revised form 5 March 2022; Accepted 2 May 2022
Available online 18 May 2022
0957-4158/© 2022 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
S. Kittisares et al. Mechatronics 85 (2022) 102822
With the alternating pressure system, the piping is simplified and

only a simple on–off valve is required, rather than the complex 3/3
flow control valve required to achieve the same functionality as the
traditional single-acting cylinder system. In both studies, an electro-
rheological fluid, controlled with flexible electro-rheological
microvalves, was utilized as the hydraulic fluid. They achieved inde-
pendent motion of the two actuators. However, as the system was
designed for a microactuator, its functionality as a hydraulic supply
system is very limited: neither pressure nor volume can be controlled,
and the input peak-to-peak pressure is only 110 kPa. Lastly, the pump
in this system also needs to run continuously, resulting in low power
efficiency.
Another important component of hydraulic systems is hydraulic
fluid. While oil-based hydraulic liquids are often used in traditional
systems, water-based hydraulic liquids have also been proposed as al-
ternatives because they are more user-friendly, accessible, and environ-
mentally friendly [17,18]. These advantages make water an appealing
choice for wearable devices that are in close contact with the user.
In this paper, we propose a new alternating pressure control system
(APCS) based on an alternating pressure source system, wherein each
hydraulic actuator is connected to a centralized alternating pressure
source and an on–off valve. However, instead of simply synchronizing
the valves to the pump in an open-looped fashion, the pressure and Fig. 1. Outline of the proposed system, consisting of a motor, a pump, and a
position sensors allow a more complex algorithm to be utilized, thereby theoretically unlimited number of sets of on–off valves and actuators. The on/off valves
improving functionality. We propose a control algorithm utilizing on– open and close in sync with the alternating pressure from the pump.
off valves, feedback signals, and an alternating pressure source in a
multi-degree-of-freedom system. In the proposed system, the pressure
in each actuator is ‘‘updated’’ at a semi-fixed period, allowing both with water as the hydraulic fluid. The pump operated in a pressure-
pressure and volume in each actuator to be controlled simultaneously. controlled mode and had a maximum operating pressure of 5 MPa.
Furthermore, we also propose a dynamic duty ratio algorithm to reduce To utilize the double-acting cylinder as a single-acting cylinder, the
total pump utilization time and improve power consumption. The hydraulic pressure is only connected to one side of the cylinder, while
performance and effectiveness of the proposed system are validated a constant pneumatic pressure is applied to the other side as a return
through simulations and experiments. The APCS reduces the weight and mechanism. Further, instead of simply synchronizing the valves to the
cost of EHAs while enabling lower power consumption and requiring alternating pressure in an open-looped fashion, pressure sensors and
fewer components than conventional centralized systems. We believe linear potentiometers were installed for feedback, which allows a more
this system will be beneficial multi-degree-of-freedom systems where complex control algorithm to be utilized for pressure and position
high power efficiency is needed, such as mobile robots. control.
2. Proposal of the alternating pressure control system 2.1. Alternating pressure source
A general APCS consists of one centralized servo pump and one The alternating pressure source is an integral part of this study.
on–off valve per actuator, similar to the alternating pressure source Higher frequencies result in quicker system response and less delay;
system. However, in APCS, pressure sensors or position sensors can also however, this response is restricted by the limitations of the pump and
be installed for feedback according to the desired operating mode, as servo motor and the operating speed of the valve. Moreover, different
shown in Fig. 1. The inclusion of feedback sensors allows the controller alternating pressure waveforms can also lead to variable performance
to control either pressure or position of each actuator, as well as stop depending on the type of application. An APCS with a rectangular
the motor when not required to reduce power consumption. wave, associated with abrupt changes of pressure in the alternating
The control algorithm of APCS can be described as follows. First, pressure source, would enable a faster response with a higher alternat-
the controller compares the setpoint and output of the system and ing frequency. This makes a rectangular wave source pressure suitable
determines the operating mode required. If all actuators are at the if the actuator only requires the pressure to be on or off. In contrast, a
setpoint, the valves are closed, and the motor stops. However, if the sinusoidal wave alternating pressure source exhibited a more gradual
pressure or volume in one or more actuators needs to be increased or change in pressure. This allows the valves to open or close at any point
decreased, the motor turns forward or backward, respectively. Simul- between the maximum and minimum pressures and additionally allows
taneously, the valves at the corresponding actuator open, allowing the any pressure value between the minimum and maximum values to be
motor to power the actuator as a motor-controlled hydraulic system. used as a setpoint.
If the pressure or volume in certain actuators must be increased and For a rectangular wave, it is also possible to manipulate the duty
decreased simultaneously, then the motor applies alternating pressure ratio to improve the overall performance of all the cylinders. Herein,
to the system. Subsequently, the controller compares the supply and the dynamic duty ratio function for a system with n actuators is
actuator pressures, and then opens or closes the valve according to expressed as Eq. (1)
the algorithm. In addition, a dynamic duty ratio was employed to ( ∑𝑛 )
1 𝑖=1 sgn(𝑥𝑟,𝑖 − 𝑥𝑖 )
maintain a balanced performance between all the actuators during the DR = ∑𝑛 + 1 . (1)
𝑖=1 | sgn(𝑥𝑟,𝑖 − 𝑥𝑖 )|
2
alternating pressure mode, i.e., the APCS operates as a hybrid of a
pump-controlled system and a valve-controlled system. where
The system in this study consists of one positive displacement pump 𝑖 is the number of the actuator;
powered by a servo motor, four double-acting cylinders operating as 𝑥𝑟,𝑖 is the setpoint of actuator 𝑖 for the relevant mode (force or
single-acting cylinders, four on–off valves, and five pressure sensors, displacement); and
2
𝑥𝑖 is the current output of actuator 𝑖.

Function sgn denotes the signum function, which is defined as
follows:
⎧−1 if 𝑥 < 0,
⎪
sgn(𝑥) ∶= ⎨0 if 𝑥 = 0, (2)
⎪
⎩1 if 𝑥 > 0.
However, a duty ratio value that is too low results in a short pulse,
which does not provide sufficient time for the pump to reach the
desired value. Similarly, a very high duty ratio value results in the
pressure not falling to the desired value. To prevent this, the rise time
(𝑡rise ), fall time (𝑡fall ), and alternating pressure frequency (𝑓 ) of the
pump can be used to calculate the applicable range of the dynamic duty
ratio value, as shown in the equation below:
Fig. 2. Implementation of the valve controller logic in a pressure-controlled pump for
⎧0 if 𝐷𝑅 = 0,
a force-controlled mode in Simulink where 𝑒 is the value of dead band.
⎪
⎪𝑡rise 𝑓 if 0 < 𝐷𝑅 ≤ 𝑡rise 𝑓 ,
⎪ Table 1
𝐷𝑅applicable = ⎨𝐷𝑅 if 𝑡rise 𝑓 < 𝐷𝑅 < 1 − 𝑡fall 𝑓 (3) Parameters used in the simulation.
⎪ Parameter Value
⎪1 − 𝑡fall 𝑓 if 1 − 𝑡fall 𝑓 ≤ 𝐷𝑅 < 1
⎪1 if 𝐷𝑅 = 1. System temperature 293.15 K
⎩ Atmospheric pressure 0.101325 MPa
Pump rise time 0.1 s
The value of 𝑡rise and 𝑡fall in the context of this paper are obtained
Pump fall time 0.1 s
experimentally as the time taken by the pump to reach maximum Cylinder piston diameter 20 mm
pressure or minimum pressure after receiving a step input signal. Cylinder piston stroke 0.2 m
Valve delay 20 ms
While only the rectangular wave and the sinusoidal wave were used
Valve orifice diameter 1.5 mm
in this study, other waveforms such as sawtooth or triangular waves, Valve leakage area 1 × 10−10 m2
which also have gradual pressure changes, can also be considered in a
future study.
3. Simulation
2.2. Valve controller
The system was simulated using MATLAB 2021a Simulink and
2.2.1. Pressure control Simscape software. The pumps in the pressure- and volume-controlled
The valve controllers compare the source pressure, the pressure modes were modeled as perfect pressure sources. The transient response
inside the actuators, and setpoint, and then open or close the valves of the pump was modeled as an underdamped unity-gain second-order
depending on the situation. The basic logic of the valve controller is system with the transfer function shown in Eq. (6). Further, four sets
given by Eq. (4). The logical operator AND and OR are denoted as ∧ of hydraulic cylinders and valves were connected to the pump. The
components are modeled after the experimental equipment, which is
and ∨ respectively.
explained in detail in Section 4.1. We assume that liquid temperature
((𝑃𝑠 > 𝑃𝑎 ) ∧ (𝑃𝑟 > 𝑃𝑎 )) ∨ ((𝑃𝑠 < 𝑃𝑎 ) ∧ (𝑃𝑟 < 𝑃𝑎 )) ⟹ 𝑉open . (4) was constant and there was no deformation in the components. Other
conditions including cylinder friction, valve leakage, fluid compress-
where ibility and viscosity, volume of tubing were already considered. The
𝑃𝑠 is the source pressure; parameters in the simulations are obtained experimentally and are
𝑃𝑟 is the reference pressure; and given in Table 1.
𝑃𝑎 is the actuator pressure. 1002
Moreover, to prevent overshoot, which would increase the error 𝐺(𝑠) = (6)
𝑠2 + 160𝑠 + 1002
instead of reducing it, the valve should only open when 𝑃𝑟 is closer
to 𝑃𝑠 than to 𝑃𝑎 , i.e.,
3.1. Pressure control
|𝑃𝑟 − 𝑃𝑠 | < |𝑃𝑟 − 𝑃𝑎 |. (5)
In the pressure control mode, both rectangular and sinusoidal waves
The stability of the system can be improved by using a dead zone were used for the source pressure and setpoint. This results in four
around the target pressure. The implementation of the valve controller combinations:
in MATLAB Simulink is shown in Fig. 2
1. Rectangular source, rectangular setpoint (R–R)
2. Sinusoidal source, rectangular setpoint (S–R)
2.2.2. Position control
3. Rectangular source, sinusoidal setpoint (R–S)
Operation in this mode is relatively simple: a rectangular wave with
4. Sinusoidal source, sinusoidal setpoint (S–S).
a dynamic duty ratio expressed in Eq. (1) is provided to the pump. The
valve then opens or closes according to the direction of the pump and The hydraulic cylinders in the pressure control mode were simu-
the desired cylinder direction. lated to be fixed with no external force acting on the cylinder. The
The dynamic duty ratio allows all the actuators to be equally valves were simulated as gate valves with a transport delay of 50 ms to
prioritized while improving the performance when only a number of emulate the actuation time of the real valve. The models of the valve
actuators are active. and cylinder are shown in Fig. 3.
3
Fig. 3. The model of the valve and cylinder in pressure control mode simulations. The
cylinder rod was fixed, and the volume within the cylinder was constant.
Fig. 5. Simulation results in the S-R mode with a 1 Hz source frequency and a 0.2
MPa deadband.
Fig. 4. Simulation results in the R-R mode with a 1 Hz source frequency and a 0.2
MPa deadband.
The performance of the simulated system is calculated using the root

mean square error (RMSE) defined as shown in Eq. (7)
√
√ 𝑛
√1 ∑
RMSE = √ (𝑃 − 𝑃𝑖 )2 . (7)
𝑛 𝑖=1 r,𝑖
Fig. 6. Simulation results in the R-R mode with dynamic duty ratio with a 1 Hz source
where: frequency and a 0.2 MPa deadband.
𝑛 is the number of sampled points;
𝑃𝑟,𝑖 reference pressure; and Table 2
𝑃𝑖 is the cylinder pressure. RMSE and delay in S-S mode from 1 Hz to 5 Hz. The dead band was 0.2 MPa.
In addition, for sinusoidal setpoint modes, the output delay was Frequency (Hz) 1 2 3 4 5
also calculated. A discrete Fourier transform is used to obtain the RMSE (MPa) 0.69 0.52 0.41 0.40 0.46
phase difference between the setpoint and pressure output, which is Delay (s) 0.30 0.16 0.13 0.10 0.10
then used to calculate the time delay. The theoretical maximum delay
corresponds to the period of the alternating pressure.
The setpoints in the rectangular setpoint modes were rectangular cylinders were at a fully extended position with a virtually fixed vol-
waves with different frequencies and pulse widths. The APCS could ume. The current implementation of the dynamic duty ratio increases
control the pressures in all cylinders simultaneously with both rect- the theoretical maximum delay to two periods of the source pressure,
angular and sinusoidal pressure sources, as shown in Figs. 4 and 5. which also results in an increased RMSE. The simulation results are
The sinusoidal pressure source has a more gradual change in pressure, shown in Fig. 6.
resulting in a larger delay compared with the rectangular pressure Simulations with sinusoid setpoint modes yielded contrary results;
source mode. However, all delays were less than the period of the using a sinusoid wave as the source pressure resulted in lower delay and
alternating pressure source. Further, the RMSE values for the R-R and error, as shown in Figs. 7 and 8. This is because sinusoidal waves have a
S-R modes were 1.21 and 1.56 MPa, respectively. more gradual change in pressure compared to rectangular waves, which
The application of the dynamic duty ratio reduces the total pump allows the valves more time to shut at the reference pressure. As a
utilization time in the R-R mode from 50% to 41.22% and increases result, the R-S mode has a larger delay (0.61 s) and RMSE (1.22 MPa)
the RMSE to 1.62 MPa. The dynamic duty ratio decreases energy con- compared to the corresponding values for the S-S mode (0.30 s, 0.69
sumption by reducing the pump utilization time when all actuators are MPa).
already in place; in addition, it improves the performance by increasing Increasing the source pressure frequency improves both the RMSE
the pump utilization time when activating an actuator. However, the and the delay of the system. The delays and RMSEs of the S-S mode
improved performance was not apparent in this simulation, where all between the 1 and 5 Hz source frequencies are shown in Table 2.
4
Fig. 7. Simulation results in the R-S mode with a 1 Hz source frequency and a 0.2 Fig. 10. The model of the valve and cylinder in the position control mode in Simulink
MPa deadband. Simscape.
Fig. 11. Experiment setup. The controller box and compressor are not shown in this
figure.
Fig. 8. Simulation results in the S-S mode with a 1 Hz source frequency and a 0.2 with a 5 mm dead band are shown in Fig. 10. The RMSE of all four
MPa deadband. cylinders is 3.17 mm. The pump was activated for a total of 16.30 s,
which is 32.60% of the operating time.
While setting the dead band to 0 mm offers a very slight reduction
of RMSE to 3.10 mm, it increases the pump utilization time by 75%
to 28.14 s or 56.28% of the operating time. In contrast, increasing the
dead band to 10 mm increases the RMSE to 5.05 mm and reduces the
pump utilization time to 11.63 s, or 23.26% of the operating time.
4. Experiments
4.1. Experiment setup

Fig. 9. The model of the valve and cylinder in the position control mode simulations.
The double-acting cylinder acts as a single-acting cylinder with a constant pressure as The experimental setup consisted of a water pump (ASP035-T110,
the return mechanism. LEVEX Corporation, Kyoto, Japan) connected to a 3-phase servo mo-
tor (HG-SR51B, Mitsubishi Electric Corporation, Tokyo, Japan). The
servo motor was powered by a servo amplifier (MR-J4a, Mitsubishi
3.2. Position control Electric Corporation, Tokyo, Japan) and connected to a controller
board (MicroLabBox, dSPACE Inc., MI, USA) operating at 1 kHz. Fur-
In the position control mode, a constant pneumatic pressure of 0.9 ther, two solenoid flow control valves (HTJ262G002, ASCO Japan Co.
MPa is applied to the cylinder as a return mechanism instead of a Ltd., Hyogo, Japan) were connected in opposite directions to operate
return spring. This was simulated as a constant force of an equivalent as a single on/off valve. The double-acting hydraulic cylinders (KS-
magnitude acting on the rod. Further, the valves were simulated as gate 5LB20 × 20, JPN Co. Ltd., Tokyo, Japan) have an inner diameter
valves with a transport delay of 50 ms, similar to the pressure control of 20 mm and a stroke of 200 mm. In addition, the pressures were
mode. The models of the valve and cylinder are shown in Fig. 9. measured using pressure sensors (PSE577, SMC Corporation, Tokyo,
The simulated system could operate all four hydraulic cylinders Japan). The constant air pressure of 0.9 MPa, which acts as a return
independently. The resulting positions of the hydraulic cylinder rods mechanism in the position-controlled modes, was regulated with a
5
Fig. 12. Schematic of the experiment setup.
Fig. 14. Experiment results in the R-R mode including the dynamic duty ratio, with a
1 Hz source frequency and a 0.2 MPa deadband. The dynamic duty ratio reduces the
pump utilization.
Fig. 13. Experiment results in the R-R mode with a 1 Hz source frequency and a 0.2
MPa deadband. This is an example of a suitable configuration.
filter regulator (FRF300-02-MD, Nihon Pisco Co. Ltd., Nagano, Japan)

and a positive pressure sensor (GPD-01, Nihon Pisco Co. Ltd., Nagano,
Japan). In the experiments in a loaded condition, the external load was
simulated with a pneumatic pressure regulated by an electro pneumatic
regulator (MEVT500-0C4-T11R-4U3, CKD Corporation, Aichi, Japan).
In total, 5 pressure sensors, 4 linear potentiometers, and 8 solenoid Fig. 15. Experiment results in the S-R mode with a 1 Hz source frequency and a 0.2
valves were used. The experimental setup is illustrated in Fig. 11, and MPa deadband. This is an unsuitable configuration because sinusoidal source pressure
the schematic of each actuator is shown in Fig. 12. has a slower pressure change, resulting in a larger error.
4.2. Pressure control

sinusoidal source pressure offers a larger error and slower response
In the pressure control mode, all cylinders are in a fully-extended compared to the rectangular source pressure. Furthermore, utilization
position, and the pneumatic pressure was set to 0 MPa. of the dynamic duty ratio is not possible with a sinusoidal source
pressure (see Fig. 15).
4.2.1. Rectangular setpoint To demonstrate the performance of the APCS at a higher source
The experimental results at 1 Hz source pressure frequency were frequency, experiments were also conducted between 1 Hz and 5 Hz.
consistent with the simulation results. In the R-R mode with a constant Increasing the source frequency to 5 Hz reduced the delay significantly
duty ratio of 0.5, the RMSE was 0.61 MPa, which was lower than the as shown in Fig. 16. The RMSE at 5 Hz source frequency was 0.65 MPa.
simulated result of 1.21 MPa (see Fig. 13). Applying the dynamic duty ratio at a higher source frequency results
Applying the dynamic duty ratio to the alternating pressure source in a significantly reduced pump utilization time to 21.96% at 3 Hz
reduced the total pump utilization time to 40.78% and increased the and 22.50% at 5 Hz source frequency as shown in Fig. 17. Moreover,
RMSE to 1.55 MPa. The increase in RMSE was due to the controller the RMSE was also substantially reduced to 0.66 MPa. Utilizing the
prioritizing the activation or deactivation of the actuator. For example, dynamic duty ratio at a higher source frequency results in a signifi-
the error in Cylinder 1 in Fig. 14 between 12 and 13 s was due to the cantly reduced pump utilization time with an insignificant impact on
controller prioritizing the activation of cylinder 2 before updating the RMSE. The pump utilization percentage at different source frequencies
dynamic duty ratio at the 13 s mark. While this results in a larger error is summarized in Table 3.
in this experimental setup where the cylinders are fixed, the dynamic
duty ratio may improve performance in situations where cylinder rod 4.2.2. Sinusoidal setpoint
movement is required. Operation in the R-S mode also returned a similar result for the
In the S-R mode, the RMSE was 1.13 MPa, which was lower than simulation. The valves could not react to the sudden pressure change,
1.56 MPa in the simulations. For the rectangular setpoint modes, the resulting in a larger error compared to the S-S mode. The delay and
6
Table 3
Pump utilization time at different source frequency with the dynamic duty ratio.
Frequency (Hz) 1 2 3 4 5
Pump utilization (%) 40.78 28.62 21.96 21.89 22.50
Fig. 18. Experiment results in the R-S mode with a 1 Hz source frequency and a
0.2 MPa deadband. Operation in this mode results in ‘‘overshoot’’ errors because the
valves are not fast enough to close at the setpoint pressure. This is an example of an
unsuitable configuration.
Fig. 16. Experiment results in the R-R mode with a 5 Hz source frequency and a 0.5
MPa deadband.
Fig. 19. Experiment results in the S-S mode with a 1 Hz source frequency and a 0.2
MPa deadband. This is an example of a suitable configuration.
Fig. 17. Experiment results in the R-R mode dynamic duty ratio with a 5 Hz source
frequency and a 0.2 MPa deadband.
RMSE in the R-S mode for an alternating source frequency of 1 Hz and

a 0.2 MPa dead band were 0.29 s and 1.17 MPa, respectively.
Finally, in the S-S mode with a frequency of 1 Hz and a 0.2 MPa
dead band, the average RMSE and delay were 0.64 MPa and 0.32 s,
respectively, as shown in Fig. 19. This is similar to the corresponding
simulation results of 0.69 MPa and 0.30 s, respectively.
Similar to the simulations, increasing the source frequency reduced
the delay. Unlike the simulation, however, the RMSE increased at a
higher frequency. We believe this was caused by the valves being
slower in the experiments, resulting in a larger ‘‘overshoot’’ error,
as shown in Fig. 20. Moreover, the RMSE value also depends on
the frequency of the setpoint signal. The RMSE of Cylinder 3, which
has the slowest setpoint frequency, is significantly lower than that of
Cylinder 4 which has the fastest setpoint frequency within the same Fig. 20. Experiment results in the S-S mode with a 5 Hz source frequency and a 0.2
configurations. The RMSE and delay values between 1 and 5 Hz with MPa deadband. The ‘‘overshoot errors’’ occurred due to the faster pressure change,
similar to the R-S mode. This is another example of an unsuitable configuration.
a 0.2 MPa dead band are summarized in Table 4.
7
Table 4 Table 7
RMSE and delay in S-S mode from 1 Hz to 5 Hz. The dead band was 0.2 MPa. RMSE (MPa) in pressure control mode with external disturbance.
Frequency (Hz) 1 2 3 4 5 Mode Disturbance
Overall RMSE (MPa) 0.64 0.67 0.80 0.82 0.85 None Constant Dynamic
Cylinder 1 RMSE (MPa) 0.58 0.60 0.79 0.78 0.94
R-R 0.61 1.02 1.18
S-R 1.13 1.17 1.27
R-S 1.17 1.22 1.33
S-S 0.64 0.86 0.80
Overall delay (s) 0.32 0.17 0.14 0.11 0.13
Table 5
RMSE (mm) in position control mode at different alternating source
frequencies and dead band values.
Dead Band (mm) Frequency (Hz)
1 2 3
0 4.45 3.03 4.99
5 4.90 3.54 3.29
10 5.51 4.53 4.48
Table 6
Pump utilization time (%) in the position control mode at different
alternating source frequencies and dead band values.
Dead Band (mm) Frequency (Hz)
1 2 3
0 57.66% 55.82% 49.62%
5 47.81% 38.87% 34.51%
10 43.26% 25.72% 26.07%
Fig. 21. Experiment results in the position control mode with a 1 Hz source frequency
and 5 mm dead band.
Increasing the source frequency beyond 5 Hz was not possible with
our current experimental equipment.
4.3. Position control
The experiment in the position control mode with a 5 mm dead

band (see Fig. 21) also showed a similar result when compared to the
simulations, with an RMSE of 4.90 mm, compared to the simulated
value of 3.17 mm. Moreover, changing the frequency or dead band gap
showed no significant reduction in error in the experiment, as shown
in Table 5.
In contrast, increasing both the frequency and the dead band gap
resulted in a substantial reduction in the pump utilization time. At a
source frequency of 3 Hz and a dead band gap of 10 mm, the pump
utilization time was reduced to 26.07% as shown in Table 6.
4.4. Load disturbance
In this section, we verify the performance of APCS with load distur-

bance in both the pressure control mode and position control mode.
Fig. 22. Experiment results in the R-R mode with dynamic disturbances. The effects
of disturbance can be seen most clearly in Cylinder 4.
4.4.1. Pressure control
In the pressure control mode, the external force acting on the cylin-
der in the return direction was simulated by applying air pressure to
the hydraulic cylinder. For the constant loading experiment, a constant 4.4.2. Position control mode
pneumatic pressure of 0.75 MPa was applied to the cylinder, which In the position control mode, an external force acting on the rod
is the equivalent of 172.5 N. For the dynamic loading experiment, a in the extension direction was simulated by reducing the pneumatic
pressure to a sine wave with an amplitude of 0.125 MPa, bias of
sinusoidal pneumatic pressure with an amplitude of 0.25 MPa, bias
0.375 MPa, and angular velocity of 1.5 𝜋rad/s. This is equivalent to
of 0.25 MPa, and angular velocity of 1.5 𝜋rad/s was applied to the
an external force between 92 N to 149.5 N.
cylinder to simulate an external force between 0 to 115 N acting on
The error in position control mode increased due to the return mech-
the rod. The source pressure frequency and dead band in this section
anism having less force and thus became slower than without external
are 1 Hz and 0.2 MPa, respectively.
force. The experiment results are shown in Fig. 23 and summarized in
Adding disturbance to the system result in a slight increase in
Table 8.
RMSE, with a larger error with dynamic disturbance as shown in
Table 7. The errors in R-R mode with disturbance are much higher than 5. Discussion
without disturbance but still lower than the simulated result without
disturbance of 1.21 MPa. Experiment results in R-R mode with dynamic In the pressure control mode, we proposed the APCS with two
disturbance are shown in Fig. 22. source pressure waveforms, namely a rectangular wave and a sinusoidal
8
response. With a faster valve, both faster response and lower error
can be achieved. Another issue that was present on real hardware was
valve leakage. This was solved by connecting two solenoid valves in
an opposite direction. The computational complexity of the proposed
controller is O(𝑛), where 𝑛 is the number of actuators.
The main limitation of this study is that both the pressure and
position are updated at semi-regular intervals instead of continuously.
This may make APCS not suitable for certain applications where a
smooth motion, a high-frequency oscillating setpoint, or an instanta-
neous response is required. However, APCS may be able to reduce
power consumption and part complexity for multi-degree-of-freedom
systems where a slower response is acceptable such as a jack-up system
for structural adjustment and construction in civil engineering [19,20],
or positioning of caisson foundation [21]. With faster equipment, the
APCS will be able to provide a faster response and may be able to be
adopted in mobile robots.
6. Conclusions
Fig. 23. Experiment results in the position control mode with a 1 Hz source frequency, In this study, we developed the APCS based on an alternating
5 mm dead band and dynamic disturbances.
supply system with feedback and a dynamic duty ratio. This system
allowed the pressure and volume in each individual actuator to be
Table 8 controlled simultaneously and independently, and required only one
RMSE (mm) in position control mode with external disturbance.
centralized pump, and one on–off valve and feedback sensor for each
Dead Band (mm) Without disturbance With disturbance actuator. Furthermore, we proposed a dynamic duty ratio equation that
0 4.45 5.11 improves the energy usage in the pressure control mode and balances
5 4.90 5.81 the performance of all actuators in the position control mode.
10 5.51 10.59
We simulated the system and found that in the pressure control
mode, a rectangular wave source pressure is preferred for a rectangular
wave setpoint while a sinusoidal wave source pressure is preferred for
wave with different advantages and disadvantages. The rectangular the sinusoidal wave setpoint. In addition, we also confirmed that the
wave, which has an abrupt change in pressure, excels when the pressure application of the dynamic duty ratio reduced the pump utilization
setpoints are also rectangular waves as demonstrated in the R-R mode. time from 50% to 41.22% in the R-R mode. Further, in the position
The abrupt change allows a faster change in pressure, leading to a faster control mode, the system could follow the position command of each
response. Conversely, the abrupt change in pressure means that the actuator independently, and increasing the alternating source pressure
time window that the valve needs to operate is very short when the frequency reduced the delay and error in all operation modes.
setpoints are continuous waveforms. The delayed response of the valves The experimental results were consistent with the simulated results.
led to a larger error in R-S mode as shown in Fig. 18. We could operate four hydraulic cylinders in parallel with a similar er-
In contrast, a sinusoidal source pressure has a gradual change in ror, which were delayed compared to the simulations. In the S-S mode,
pressure, giving more time for the valves to operate. This results in a increasing the source frequency reduced the delay and increased the
lower error than rectangular source pressure with a sinusoidal setpoint error, whereas both delay and error are reduced in the simulations. In
as shown in the S-S mode. However, a gradual pressure change also re- the R-R mode with dynamic duty ratio, increasing the source pressure
sults in a slower response when a faster response is required, as shown frequency reduces pump utilization to 21.96% at 3 Hz. Increasing the
in the S-R mode. Lastly, increasing the frequency in Sinusoidal mode source frequency beyond 3 Hz produced a negligible change in pump
shortens the time window for the valves and negates its advantage. This utilization. In contrast, in the position control mode, increasing the
results in an error similar to that of the rectangular source pressure alternating source frequency or dead band gap did not show a reduced
mode as shown in Fig. 20. A suitable configuration is required for the error, but reduced the pump utilization considerably from 57.66% to
26.07%. The proposed system was able to operate under a loaded
APCS to operate optimally, namely a rectangular source pressure for a
condition in both pressure control mode and position control mode,
step reference signal, and sinusoidal source pressure for a continuous
although with an increased error.
reference signal.
Future directions include conducting experiments in other load con-
In the position control mode, the APCS is designed to operate with
ditions, improving the algorithm to operate in an open-looped fashion
single-acting cylinders as actuators. In this paper, we utilized double-
with an implementation of an observer such as those proposed in [22],
acting cylinders as single-acting cylinders by supplying a constant air
and the application of the system on a real robot.
pressure on one side instead of a return spring which is more commonly
used. However, we believe that a single-acting cylinder with a return CRediT authorship contribution statement
spring as the return mechanism can also be used to a similar effect.
The specific experiment setup used in this paper cannot hold an Sarin Kittisares: Conceptualization, Methodology, Software, Inves-
overrunning load beyond the pneumatic pressure due to the limitations tigation, Writing – original draft. Yosiharu Hirota: Resources. Hi-
of a single-acting hydraulic cylinder. If the load is mainly in the royuki Nabae: Conceptualization, Methodology, Writing – review.
extension direction, switching the pneumatic and hydraulic port will Gen Endo: Conceptualization. Koichi Suzumori: Conceptualization,
allow the load to be held. If both resistive load and overrunning load Methodology, Writing – review & editing, Supervision.
are required, changing the actuator to a double-acting cylinder with a
4/3 directional control valve will allow load in both directions to be Declaration of competing interest
help at the cost of a more complex valve and piping.
The challenges of implementation of APCS on the hardware are the The authors declare that they have no known competing finan-
limitations of the pump and operating speed of the valve which limit cial interests or personal relationships that could have appeared to
the maximum source frequency resulting in a larger error and slower influence the work reported in this paper.
9
Funding [20] Seo J, Yoo WS, Lee UK, Kim C, Kang KI, Cho H. Case study of a synchronous
hydraulic jack-up system for constructing high-rise residential buildings. Can J
Civil Eng 2010;37(6):922–6. http://dx.doi.org/10.1139/L10-026.
This work was supported by the JSPS KAKENHI Grant-in-Aid for
[21] Houlsby GT, Kelly RB, Huxtable J, Byrne BW. Field trials of suction caissons
Scientific Research(A) grant number JP18H03760. in sand for offshore wind turbine foundations. Geotechnique 2006;56(1):3–10.
http://dx.doi.org/10.1680/geot.2006.56.1.3.
References [22] Razmjooei H, Shafiei MH, Palli G, Ibeas A. Chattering-free robust finite-time
output feedback control scheme for a class of uncertain non-linear systems.
[1] Zang X, Liu Y, Zhu Y. Structure design of a mobile jack robot. In: 2013 IEEE IET Control Theory Appl 2020;14(19):3168–78. http://dx.doi.org/10.1049/iet-
international conference on information and automation. 2013, p. 1218–23. cta.2020.0910.
http://dx.doi.org/10.1109/ICINFA.2013.6720480.
[2] Kaminaga H, Ko T, Masumura R, Komagata M, Sato S, Yorita S, et al. Mechanism
and control of whole-body electro-hydrostatic actuator driven humanoid robot
Sarin Kittisares received his B.E. degree in Mechatronics
hydra. In: Springer proceedings in advanced robotics, vol. 1, 2016, p. 656–65.
from King Mongkut’s University of Technology Thonburi,
http://dx.doi.org/10.1007/978-3-319-50115-4_57.
Bangkok, Thailand in 2017. He received the M.E. degree in
[3] Suzumori K, Faudzi AA. Trends in hydraulic actuators and components in legged
Mechanical Engineering from Tokyo Institute of Technology,
and tough robots: a review. Adv Robot 2018;32(9):458–76. http://dx.doi.org/10.
Tokyo, Japan in 2020, where he is currently pursuing his
1080/01691864.2018.1455606.
doctoral degree.
[4] Hemmi M, Morita R, Hirota Y, Inoue K, Nabae H, Endo G, et al. Development of
hydraulic tough motors with high power density and their application to a 7-axis
robotic arm. In: Proceedings of the 2019 IEEE/SICE international symposium on
system integration, SII 2019. Institute of Electrical and Electronics Engineers
Inc. 2019, p. 264–9. http://dx.doi.org/10.1109/SII.2019.8700419.
[5] Raibert M, Blankespoor K, Nelson G, Playter R. Bigdog, the rough-terrain Yosiharu Hirota received his bachelor degree from the
quadruped robot. IFAC Proc Vol 2008;41(2):10822–5. http://dx.doi.org/10. Department of Science and Engineering, Waseda University,
3182/20080706-5-KR-1001.01833. Tokyo, Japan in 1968. From 1968, he worked for Yuken
[6] Semini C, Tsagarakis NG, Guglielmino E, Focchi M, Cannella F, Caldwell DG. Kogyo Co. Ltd. In 2015, he joined the school of Engineering,
Design of HyQ – a hydraulically and electrically actuated quadruped robot. Vol. Tokyo Institute of Technology as a researcher. He is a
225. No. 6. 2011, p. 831–49. http://dx.doi.org/10.1177/0959651811402275. member of The Japan Fluid Power System Society.
[7] Kohlbrecher S, Romay A, Stumpf A, Gupta A, von Stryk O, Bacim F, et al.
Human–robot teaming for rescue missions: Team vigir’s approach to the 2013
DARPA robotics challenge trials. J Field Robotics 2015;32(3):352–77. http://dx.
doi.org/10.1002/rob.21558.
[8] Li X, Zhou H, Feng H, Zhang S, Fu Y. Design and experiments of a novel
hydraulic wheel-legged robot (WLR). In: IEEE international conference on Hiroyuki Nabae received his B.E., M.E., and Ph.D. degrees
intelligent robots and systems. 2018, p. 3292–7. http://dx.doi.org/10.1109/IROS. from the University of Tokyo, Tokyo, Japan in 2010, 2012,
2018.8594484. and 2015, respectively. He worked as a JSPS research fellow
[9] Kaminaga H, Amari T, Niwa Y, Nakamura Y. Electro-hydrostatic actuators with from April 2013 to March 2015. He is currently an assistant
series dissipative property and their application to power assist devices. In: professor at Tokyo Institute of Technology. His research
2010 3rd IEEE RAS and EMBS international conference on biomedical robotics interests include mechatronics, robotics, and actuators.
and biomechatronics, BioRob 2010. 2010, p. 76–81. http://dx.doi.org/10.1109/
BIOROB.2010.5626341.
[10] Song B, Lee D, Park SY, Baek YS. Design and performance of nonlinear control for
an electro-hydraulic actuator considering a wearable robot. In: Processes 2019.
Vol. 7. p. 389. http://dx.doi.org/10.3390/PR7060389, 7 (6) (2019) 389. Gen Endo received his B.E., M.E., and Ph.D. degree from
[11] Kittisares S, Nabae H, Endo G, Suzumori K, Sakurai R. Design of knee support Tokyo Institute of Technology, Japan in 1996, 1998, and
device based on four-bar linkage and hydraulic artificial muscle. ROBOMECH J 2000, respectively. From 2000 to 2007, he worked for
2020;7(1). http://dx.doi.org/10.1186/s40648-020-00165-2. Sony Corp. and was also a visiting researcher as the Ad-
[12] Alle N, Hiremath SS, Makaram S, Subramaniam K, Talukdar A. Review on electro vanced Telecommunication Research Institute International.
hydrostatic actuator for flight control. Int J Fluid Power 2016;17(2):125–45. In 2008, he joined the faculty of Tokyo Institute of Technol-
http://dx.doi.org/10.1080/14399776.2016.1169743. ogy, as an assistant professor. In 2014, he was an associate
[13] Chao Q, Zhang J, Xu B, Huang H, Pan M. A review of high-speed electro- professor of Institute of Biomaterials and Bioengineering in
hydrostatic actuator pumps in aerospace applications: Challenges and solutions. Tokyo Medical and Dental University. He is currently a
Trans ASME, J Mech Des 141(5). http://dx.doi.org/10.1115/1.4041582, (may professor of Tokyo Institute of technology. He is a member
2019). of the Japan Society of Mechanical Engineers, the Robotics
[14] Karanovic V, Jocanovic M, Jovanovic V. Review of development stages in Society of Japan, IEEE and The Society of Instrument and
the conceptual design of an electro hydrostatic actuator for robotics. Acta Control Engineers.
Polytechnica Hungarica 2014;11(5):59–79.
[15] Miyoshi T, Yoshida K, Eom SI, Yokota S. Proposal of a multiple ER mi-
croactuator system using an alternating pressure source. Sensors Actuators A
2015;222:167–75. http://dx.doi.org/10.1016/J.SNA.2014.12.002.
[16] Sudhawiyangkul T, Yoshida K, Eom SI, wan Kim J. A novel bending mi- Koichi Suzumori was born in 1959. He received the BS,
croactuator with integrated flexible electro-rheological microvalves using an MS, and Ph.D. degrees in Mechanical Engineering from
alternating pressure source for multi-actuator systems. Microsyst Technol Yokohama National University, Japan, in 1982, 1984, and
2020;26(5):1507–19. http://dx.doi.org/10.1007/s00542-019-04685-9. 1990, respectively. He had worked for Toshiba R&D Center
[17] Takahashi T, Yamashina C, Miyakawa S. Development of water hydraulic propor- from 1984 to 2001, and also worked for Micromachine
tional control valve. In: Proceedings of the JFPS international symposium on fluid Center, Tokyo, from 1999 to 2001. He had been a Professor
power 1999. No. 4. 1999, p. 549–54. http://dx.doi.org/10.5739/isfp.1999.549. of Division of Industrial Innovation Sciences, Okayama
[18] Wang X, Yamaguchi A. Characteristics of hydrostatic bearing/seal parts for University, Japan since 2001. He has been a Professor
water hydraulic pumps and motors. Part 1: Experiment and theory. Tribol Int of Department of Mechanical and Aerospace Engineering,
2002;35(7):425–33. http://dx.doi.org/10.1016/S0301-679X(02)00023-3. Tokyo Institute of Technology since 2014. He is mainly
[19] Akai K, Tanaka Y. Ex-post-facto estimate of performance at the offshore reclama- engaging in the research fields of new actuators and their
tion of airport osaka/KIA. In: Proceedings of the 16th international conference applications. He received many awards such as JSME Medal
on soil mechanics and geotechnical engineering. IOS Press; 2005, p. 1011–4. for Outstanding Paper, in 1999, RSJ Best Paper Award in
http://dx.doi.org/10.3233/978-1-61499-656-9-1011. 2000, and JSAEM Best Book Award in 2006. He is a fellow
member of the Japan Society of Mechanical Engineers.
10
HO CHI MINH CITY UNIVERSITY OF TECHNOLOGY AND EDUCATION
FACULTY OF MECHANICAL ENGINEERING
TOPIC 3:
Research paper:
IOT-BASED HAZARDOUS GAS LEAKAGE DETECTION
Instructor: PGS.TS Tran Ngoc Dam

Presenter: Nguyen Trong Phuc
Student’s ID: 20146522
Author: Ruchika Chakraborty, Ayan Prakash Gogoi Boro, Bapon Sarkar
Ho Chi Minh city, November 2022

CONTENT
1 INTRODUCTION
2 METHODS
3 RESULTS
4 CONCLUSION
1. INTRODUCTION
❑ The Internet of Things
❑ Protect people by detect and warning hazardous gases
❑ The gas (H2, LPG, CH4, CO, alcohol) is detected by the MQ5 sensor
❑ Iot-based Hazardous Gas Leakage Detection

2. METHODS
❑ 2.1. Literature Survey
• SR(swarm robotics)
• SI (swarm intelligence)
• Zigbee module
• Application of S-bots in disaster management
• Detect gas leakage by using the gas sensor

2. METHODS
❑ 2.2. System Overview
➢ Hardware Requirements
Arduino Uno Relay LCD display
ESP8266 WIFI Module MQ-5 Sensor Buzzer

2. METHODS
❑ 2.2. System Overview
➢ Software Requirements
Arduino IDE Language C++

2. METHODS
❑ 2.3. System Operation
➢ WORKING
• The gas leakage is detected by MQ 5 sensor
• Interfaced by Arduino uno
• And display alert message in LCD display.
Flow diagram of the activities of the gas leakage

detection system
2. METHODS
❑ 2.3. System Operation
BLOCK DIAGRAM CIRCUIT DIAGRAM

2. METHODS
❑ 2.4. Detection Range
• Carbon Monoxide CO 1 - 1000ppm,

• Ethanol C2H6OH 10 - 500ppm,
• Hydrogen H2 1 - 1000ppm,
• Ammonia NH3 1 - 500ppm,
• Methane CH4 >1000ppm
3. RESULTS
❑ 3.1. Experimental Results
Prototype model of the gas leakage detection system
3. RESULTS
❑ 3.1. Experimental Results
Graph of gas leakage detection
3. RESULTS
❑ 3.2. Future Scope and Development
✓ Packed with advanced features
✓ The popularity of handheld devices has advanced
✓ Workplace safety
4. CONCLUSION
✓ The monitoring and detection of hazardous gases to meet safety standards
✓ The system is quite responsive
✓ Stop crisis situations more quickly than manual methods.

Portfolio NguyenTrongPhuc 20146522 First

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Portfolio NguyenTrongPhuc 20146522 First

Uploaded by

Copyright:

Available Formats

BỘ GIÁO DỤC VÀ ĐÀO TẠO

TRƯỜNG ĐẠI HỌC SƯ PHẠM KỸ THUẬT TPHCM

BÁO CÁO CUỐI KỲ

GVHD: PGS.TS Trần Ngọc Đảm

TP. Hồ Chí Minh – 2022

2. 5 Titles and abstracts

IOT-BASED HAZARDOUS GAS LEAKAGE DETECTION

Ramapuram, Chennai, India

Ramapuram, Chennai, India

Technology, Ramapuram, Chennai, India

intercommunication between robots, and is used in each Relay

Figure.1 The Grove Gas Sensor Module (MQ5) is useful for

Figure 7 depicts the architectural block diagram of the

A buzzer or buzzer is a mechanical, electromechanical or

Figure 7. Architecture of gas leakage detection system.

This gas sensor is used to detect hazardous gases in industry.

• Carbon Monoxide CO 1 - 1000ppm,

• Ethanol C2H6OH 10 - 500ppm,

• Ammonia NH3 1 - 500ppm,

• Methane CH4 >1000ppm

When detected, it will automatically turn on. The detected

5. EXPERIMENTAL RESULTS Analog value of the gas sensor

The prototype of the project system is shown in Figures 9

It helps them to upgrade their safety norms and aids to 6. CONCLUSION

[2] R. Imtiaz, B. Ashokkumar, M. Danny Frazer

[3] Mohd. Daneel Khan, Krantee Jamdaade; ”Application of

[5]P. M. Vidya, S. Abinaya, G. G. Rajeswari, and N.

[6] Kavitha B C,VallikannuR“IOT Based Intelligent Industry

[7] Kalpesh Gupta, Gokul Krishna G and Anjali T ”An IoT

Smart Home for Paralyzed Aid

Kolkata - 700160, West Bengal, India

2.3 Signal Preprocessing

2.4 Filtered Data Classification

The classification can be performed based on any

information extraction [16]. An emerging domain of

x . Consequently, the neighbors x j are more evenly

Three distinct stages are followed for the research as

The basic principle of IOT is the integration of multiple

4.1 Feeding Module 4.2 Sleep Control Module

Fig -6: IOT Based Priority Selection of Feeding Module

As an input of the fuzzy inference rules, a sliding surface is REFERENCES

[18] Guger, Christoph, et al. "How many people are able to

[19] Lee, Po-Lei, et al. "Brain Computer Interface Based on

Application of Data Analytics to Improve Patient Care: A Systematic

Identification of Transforming the Evaluation of

Fig 1: Process Map for Data Analytics

2.1. Research Objectives

i. To investigate the concept of data analytics and its applications.

H0 – There is no significant impact of data analytics towards improving patient care.

2.4. Literature Review

Fig 2: Systematic Review Process

The concept of data analytics and its applications.

Understanding the application of data analytics in healthcare.

Determining the implications of data analytics in improving patient care.

Discussion, Conclusion and Future Implications

Contents lists available at ScienceDirect

Grey box modeling of supermarket refrigeration cabinets✩

HIGHLIGHTS GRAPHICAL ABSTRACT

• Grey box modeling of supermarket CO2

ARTICLE INFO ABSTRACT

𝑚̇ Mass flow (kg/s)

Super- and subscripts

Fig. 4. Refrigeration room, ref.: Arneg/Incold.