12th IFAC Symposium on Dynamics and Control of

12th IFAC Symposium on Dynamics and Available

Controlonline
of at www.sciencedirect.com
Process
12th IFACSystems,
Symposiumincluding Biosystems
on Dynamics and Control of
Process
12th IFACSystems,
Symposiumincluding
on Biosystems
Dynamics and Control of
Florianópolis
Process - SC,including
Systems, Brazil, April 23-26, 2019
Biosystems
Florianópolis - SC,including
Process Systems, Brazil, April 23-26, 2019
Biosystems
Florianópolis - SC, Brazil, April 23-26, 2019
Florianópolis - SC, Brazil, April 23-26, 2019 ScienceDirect
IFAC PapersOnLine 52-1 (2019) 562–567

An
An Internet-of-things
Internet-of-things Enabled Smart
Enabled Smart Manufacturing
Smart Manufacturing Testbed
Manufacturing Testbed
Testbed
An Internet-of-things Enabled
An Internet-of-things Enabled Smart Manufacturing Testbed
+
Devarshi
Devarshi Shah*, Jin Wang*,
Shah*, Jin Wang*, Q.
Q. Peter
Peter He*
He*++
Devarshi
Devarshi Shah*, Jin Wang*, Q. Peter He*
Shah*, Jin Wang*, Q. Peter He*+
*Department of Chemical Engineering, Auburn University, Auburn, AL 36849 USA
*Department of Chemical Engineering, Auburn University,
+QPH: Auburn, AL 36849 USA
(DS:dks0013@auburn.edu;
*Department JW: wang@auburn.edu; +QPH: qhe@auburn.edu, 334-844-7602)
*Department of
of Chemical
(DS:dks0013@auburn.edu; JW:
Chemical Engineering, Auburn
Auburn University,
wang@auburn.edu;
Engineering, Auburn,
Auburn, AL
AL 36849
qhe@auburn.edu,
University,
+
USA
334-844-7602)
36849 USA
(DS:dks0013@auburn.edu; JW: wang@auburn.edu; +QPH: qhe@auburn.edu, 334-844-7602)
(DS:dks0013@auburn.edu; JW: wang@auburn.edu; QPH: qhe@auburn.edu, 334-844-7602)
Abstract:
Abstract: The
The emergence
emergence of of the
the industrial
industrial Internet
Internet of
of Things
Things (IoT)
(IoT) and
and ever
ever advancing
advancing computing
computing and and
Abstract: The
communication emergence
technologies
Abstract: The emergence of the
have
of the industrial
fueled a Internet
new of Things
industrial (IoT)
revolution and ever
which isadvancing
happening computing
worldwide and
to
communication technologies haveindustrial
fueled aInternet of Things
new industrial (IoT) and
revolution everisadvancing
which happeningcomputing
worldwideand to
communication
make current
communication technologies
manufacturing have
technologies have fueled
systems a new
smarter,
fueledsmarter, industrial
safer,
a new industrial revolution
and more
revolution which
efficient. is happening
Although
which is Although worldwide
many
happeningmany general
worldwide to
to
make current manufacturing systems safer, and more efficient. general
make current
frameworks have
make currenthave manufacturing
been proposed
manufacturing systems
for
systems smarter,
IoT enabled
smarter, safer, and
systems more
for efficient.
industrial Although
application, many
there is general
limited
frameworks been proposed for IoT enabledsafer, and for
systems more efficient.
industrial Althoughthere
application, manyis general
limited
frameworks
literature on have
have been
been proposed for
for IoT enabled systems for
for industrial application, there
there is limited
frameworks
literature on demonstrations
demonstrations or
or testbeds
proposed
testbeds of
of such
IoT systems.
enabled
such In
In addition,
systems
systems. there
there is
industrial
addition, aa lack
lack of
of systematic
is application,systematic isstudy on
limited
study on
literature
the
literature on
on demonstrations
characteristics of IoT
demonstrations or
or testbeds
sensors and
testbeds of
data
of such
such systems.
analytics In addition,
challenges
systems. In there
associated
addition, there is
with
is aa lack
IoT
lack of systematic
sensor
of data.
systematic study
This on
study
study on
the characteristics of IoT sensors and data analytics challenges associated with IoT sensor data. This study
the
is
theancharacteristics
attempt to
characteristics of
help
of IoT
fill
IoT sensors
this gap
sensors and
by
and data analytics
exploring
data the
analytics challenges associated
characteristics
challenges of IoT
associated with IoT
vibration
with IoT sensor
sensors
sensor data.
and
data. This
show
This study
how
study
is an attempt to help fill this gap by exploring the characteristics of IoT vibration sensors and show how
is
is an
IoT attempt
ansensors to
to help
attemptand big fill this
this gap
gap by exploring the characteristics of IoT
IoT vibration sensors
sensors and
and show
show how
IoT sensors and big data
help fill analytics
data analytics can
by be
be used
used to
canexploring thedevelop
to real
real time
timeofmonitoring
characteristics
develop frameworks.
vibration
monitoring frameworks. how
IoT sensors and big data analytics can be used to develop real time monitoring frameworks.
© 2019, IFAC (International Federation
IoT sensors
Keywords: and big data analytics can be used to develop
of Automatic Control) real time
Hosting monitoring frameworks.
by Elsevier Ltd. Allanalysis,
rights reserved.
Keywords: Internet-of-things,
Internet-of-things, smart
smart manufacturing,
manufacturing, bigbig data,
data, data
data analytics,
analytics, statistical
statistical analysis, vibration,
vibration,
Keywords:
soft sensor,
Keywords: Internet-of-things,
process monitoring. smart manufacturing, big data, data analytics, statistical analysis, vibration,
soft sensor,Internet-of-things,
process monitoring. smart manufacturing, big data, data analytics, statistical analysis, vibration,
soft sensor, process monitoring.
soft sensor, process monitoring.

11 INTRODUCTION
INTRODUCTION 1.3
1.3 “Big“Big Data”
Data” from from Smart
Smart Manufacturing
Manufacturing
11 INTRODUCTION
INTRODUCTION 1.3
1.3 “Big
“Big Data”
Data” from
from Smart
Smart Manufacturing
Manufacturing
With
With the the emergence
emergence of of the
the industrial
industrial Internet
Internet of of Things
Things (IoT)
(IoT) Manufacturing process operation databases
With the Manufacturing process operation databases are are massive
massive
and
With ever
and the emergence
ever advancing
emergence of the
advancingofcomputing the industrial
industrial
computing
Internet
power
Internet
power and of Things
Things (IoT)
andofexpansion
expansion of
(IoT) Manufacturing
of because
Manufacturing
because of
of the
the use
use
process
of
process
of process
process
operation
operation
operation
operation
databases
and
databases
and control
control
are
are massive
computers
massive
computers
and
wireless ever advancing
and ever networking
advancing computingcomputing
technologies, power
poweraa new and expansion
and generation
expansion of andof because of the
information use of
systems.processWith operation and
ever-accelerating control computers
advancement
wireless networking technologies, new generation because
and of the use
information of process
systems. Withoperation and control
ever-accelerating computers
advancement
wireless
networked, networking
information-based technologies, a
technologies, new generation
data analytics,of
of and
of information systems. With ever-accelerating advancement
wireless
networked,networking
networked,
and
information-based
information-based
predictive modeling
technologies,
are
a new data
technologies,
technologies,
providing
generation
data
analytics,
analytics,
unprecedented of IoT
and IoT devices
devices and
information and other
systems.
other Withcommunication
ever-accelerating
communication and
and sensing devices
advancement
sensing devices
networked, information-based technologies,
and predictive modeling are providing unprecedented of data analytics, of
and IoT devices
technologies, and itother
is communication
expected that the and
data sensing
generated devices
from
and
embedded predictive
computing modeling are
capabilities providing
as well unprecedented
as access to andIoT devices anditother
technologies, communication
is expected that the data and sensing
generated devices
from
and predictive
embedded computing modeling are providing
capabilities as well unprecedented
as access to smart and technologies,
manufacturing
and technologies, it is expected
systems
it issystems
expectedwill that
will the
grow
thatgrow data generated
exponentially
the data generated (Qin, from
(Qin,
from
embedded
previously computing
unimagined capabilities
potential uses as
of well
data andas access
information to smart manufacturing exponentially
embedded unimagined
previously computing potential capabilities uses as welland
of data as information smart
access to 2014).
smart manufacturing
Four V’s
manufacturing are systems
often
systems used will
to
will grow exponentially
characterize
grow the
exponentially essence(Qin,
(Qin,of
previously 2014). Four V’s are often used to characterize the essence of
previously unimagined
(Steering
(Steering Committee
unimagined
Committee potential
potential
of
of the
uses
the uses
of
of data
Advanced
Advanced data andand information
Manufacturing
information big
Manufacturing 2014).
2014).
big
Four
data
data
V’s
(Zikopoulos
Four V’s
(Zikopoulos
are
are often
oftenand
and
used to
Eaton,
used to
Eaton,
characterize
2011;
characterize
2011;
the
Zikopoulos
the
Zikopoulos
essence
et
essence
et
of
al.,
of
al.,
(Steering
(Steering Committee
Partnership 2.0, of the Advanced Manufacturing big data (Zikopoulos and Eaton, 2011; Zikopoulos et al.,
Partnership 2.0, 2014).
Committee
2014). of the Advanced Manufacturing 2012): big
2012): data Volume
(Zikopoulos
Volume (from
(from and terabytes
Eaton,
terabytes (~1012)
2011;
(~1012) to
Zikopoulos
to zettabytes
et
zettabytes al.,
Partnership
Partnership 2.0, 2.0, 2014). 2012):
1.1 Smart
2014).
Manufacturing 2012): Volume
(~1021)),
(~1021)), Variety
Variety (from
Volume (from
(from
terabytes
(from structured
terabytes to
structured
(~1012)
unstructured),
to(~1012)
unstructured),
to
to zettabytes
Velocity
zettabytes
Velocity
1.1 Smart Manufacturing (~1021)),
(from
(~1021)), batchVariety
to
Variety (from
online structured
streaming)
(fromstreaming)
structuredand to
and unstructured),
Veracity
to unstructured), (data Velocity
quality
1.1 Smart Manufacturing (from batch to online Veracity (dataVelocity quality
1.1 Smart Manufacturing (from
variations
(from batchor to online
uncertainty). streaming)
Big Data and
is Veracity
arguably a(data
major quality
focus
Although there are different names used to describe next variations or uncertainty). Big Data is arguably a majorquality
batch to online streaming) and Veracity (data focus
Although there are different names used to describe next variations variations
for the or
next uncertainty).
round
or uncertainty). of Big
the Data is arguably
transformation aofmajor focus
advanced
Although
generation there are
manufacturing different names
systems, used
such to
as describe
industrial next
4.0, for the next round of Big the Data is arguably aofmajor
transformation focus
advanced
Although there are different names used to
generation manufacturing systems, such as industrial 4.0, manufacturing, describe next for
for the
manufacturing,
the next
next round
and
round the of
of the
analysis
the transformation
of large
transformationdata sets of
will
of advanced
become
advanced
generation manufacturing systems, such as and the analysis of large data sets will become
smart
smart manufacturing
generation manufacturing
manufacturing and intelligent
and systems,
intelligent as industrial
industrial 4.0,
suchmanufacturing,
manufacturing, the
the amanufacturing,
4.0, key
key basis
basis of
amanufacturing,
and
ofand
the
the analysis
analysis of
competitiveness,
competitiveness, of large data
data sets
productivity
large
productivity
will
will become
sets growth,
growth,becomeand
and
smart
essence manufacturing
of these
smart manufacturing is the and
application intelligent
of
and intelligent manufacturing,
increasingly powerful
manufacturing, the
and aa key
the innovation basis of
(Manyika competitiveness,
et al., 2011). productivity growth, and
essence of these is the application of increasingly powerful and key basis
innovation of
(Manyika competitiveness,
et al., 2011). productivity growth, and
essence
essence of
low-cost these is
is the
of computation the application
and of
of increasingly powerful
powerful and
low-cost these
computation and networked
application increasingly
networked information-based
information-based and innovation
innovation (Manyika
(Manyika et
et al.,
al., 2011).
2011).
low-cost
low-cost computation
technologies in and
and networked information-based
technologies in manufacturing
computation
manufacturing enterprises.
networked There
There is is aa general
enterprises. information-based general 1.4 1.4 Smart
Smart Manufacturing
Manufacturing Testbed Testbed
technologies
consensus
technologies thatin manufacturing
factories
in factories
manufacturing and enterprises.
plants connected
enterprises. There
There to is
is aa Internet
the general
general 1.4 Smart Manufacturing Testbed
consensus that and plants connected to the Internet 1.4 Smart Manufacturing Testbed
consensus
are more that
that factories and
and plants connected to
to the Internet
are more efficient,
consensus factories
efficient, productive
productive and
plants smarter
smarter than
and connected than their
the
their non-
Internet
non- Simulation
Simulation is
is a
a powerful
powerful tool
tool but
but thethe fidelity
fidelity of of the
the simulated
simulated
are more
connected efficient,
counterparts productive
(Davis and smarter than their non-
are more efficient,
connected counterparts (Davis et
productive al.,
al., 2015,
etand smarter
2015, 2012).
than their non- system
2012). Simulation
is
Simulation
system
is
limiteda powerful
by
is a powerful
is limited the tool but
understanding the fidelity
tool but the fidelity
by the understanding on theof the
the simulated
system.
on theofsystem. In
In this
simulatedthis
connected counterparts (Davis et
connected counterparts (Davis et al., 2015, 2012). al., 2015, 2012). system
project, is limited
with by
industrial the understanding
IoT still in its on the
infancy, system.
there In
is this
not
1.2 Industrial Internet-of-things system
project, with industrial IoT still in its infancy, there In
is limited by the understanding on the system. is this
not
1.2 Industrial Internet-of-things project,
sufficient
project, with industrial
understanding IoT on still
the in its infancy,
property, there
capacity is not
and
1.2 Industrial Internet-of-things sufficientwith industrial IoTonstillthein property,
understanding its infancy,capacity there is and not
1.2 Industrial Internet-of-things sufficient
performance understanding
of IoT
sufficient understanding sensors on
to
on the
enable property,
accurate
the property, capacity
simulation.
capacity This and
This
and
IoT performance of IoT sensors to enable accurate simulation.
IoT devices
devices are
are sensors,
sensors, actuators
actuators andand computers
computers with with wireless
wireless is performance
is the
the major
of
of IoT
IoT sensors
major motivation
performance motivation sensors for to
for tous enable
to
to build
usenable build
accurate
aa smart
accurate simulation.
manufacturing
smartsimulation.
manufacturing
This
This
IoT devices
networks, are
and, sensors,
most actuators
importantly, and computers
systems that with
are wireless
small and
IoT devicesand,
networks, are sensors, actuators and
most importantly, computers
systems that arewith and is
wireless
small is the
testbed
the
testbed
major
major motivation
equipped with
motivation
equipped with
for
IoT
for
IoT
us
us to
vibration
to
vibration
build
build a
a smart
sensors.
smart
sensors. In
In
manufacturing
this work
manufacturing
this work we
we
networks,
easy to and,
and, most importantly, systems that are
are small and
networks,
easy to embed.
embed. Although
most
Although the
the use
importantly, of
of IoT
usesystems sensors
IoT that
sensors has
has been
small been testbed
and investigate
testbed equipped
the
equipped
investigate the
with
capabilities
with
capabilities
IoT
IoT vibration
of these
vibration
of these
sensors.
IoT
sensors.
IoT
In
vibration
In
vibration
this
this work
sensors
work
sensors
we
for
we
for
easy
easy to to embed.
increasing Although
exponentially
Although in the use
use ofof IoT
theindustries sensors
such as has
has been
increasing embed.
exponentially in industries IoT such as retail
sensors retail and
been investigate
and process
investigate
process
the
monitoring
the
monitoring
capabilities
by
capabilities
by
of
building
of
building
these
these IoT
soft
IoT
soft
vibration
sensors
vibration
sensors to
to
sensors
predict
sensors
predict
for
key
for
key
increasing
services,
increasingtheir exponentially
use in
in industries such as
as retail and
services, use inin manufacturing
exponentially
their manufacturing has
has been
industries suchlimited.
been limited. Because
retail and process
Because monitoring
process variables.
monitoring
variables. We by building
Webyexpectbuilding
expect that
that the
soft
the
sensors
softfindings
sensorsfrom
findings
to
to predict
from this key
this project
predict key
project
services,
of the
services, their
small use
size
their use inandmanufacturing
cheap has
price, been
IoT limited.
sensors Because
offer the process
can be variables.
extended toWe expect
actual that the
manufacturing findings from
systems. this
The project
rest of
of the small sizeinand manufacturing
cheap price, hasIoT beensensors
limited.offer
Becausethe process
can be variables.
extended toWe expect
actual that the
manufacturing findings from
systems. this
The project
rest of
of the
opportunity small to size and
instrument cheap
systems price,
in IoT
massive sensors
numbers. offer the
With can be extended to actual manufacturing systems. The rest
of the smallto size
opportunity
opportunity
the huge to
amount
and cheap
instrument
instrument
of data
systems
systems
and
price,
thein
IoT sensors
in massive
massive
programmability
numbers.
numbers.
offerWith
ofWith
IoT
the
the can paper
the paper is
be extended organized
to actual
is organized as follows.
as manufacturing Sec.
follows. Sec. 2systems.2 details
details the the rest of
testbed
Thetestbed of
opportunity
the huge amount to instrument
of datasystems and theinprogrammability
massive numbers.ofWith the
IoT setup.
the paper
Sec.
paper is
is organized
33 investigates
organized as
as follows.
the
follows. Sec.
Sec. 2
2 details
characteristics
details of the
the testbed
the data
testbed
the huge amount of data setup. Sec. investigates the characteristics of the data
devices,
the
devices,hugecomes
amount
comes the
the of data and
opportunity
opportunity and to the
to
programmability
theshape the
the data
data received,
programmability
shape
of
of IoT
received, to
IoT setup.
setup. Sec.
to collected
collected from
Sec.
from
3 investigates
3 the IoT
IoT vibration
investigates
the vibration
the characteristics
the sensors. Sec. of
of the
Sec. 44 presents
characteristics
sensors. the data
presents data
devices,
address
devices,localcomes
local
comes the opportunity
redundancy
the opportunity of to shape
information,
to shape the the
and data
to received,
improve to
both collected
to representation,from the IoT
visualization vibration
and sensors.
soft sensor Sec. 4 presents
development data
and
address redundancy of information, anddata received,
to improve both collected from
representation, the IoT
visualization vibration
and sensors.
soft sensor Sec. 4 presents
development data
and
address
the local
accuracy
localand redundancy
precision of information, and to
to improve both
address
the accuracy and precisionofof
redundancy measurements
ofinformation,
measurements andlocally
locally and
and across
improve across representation,
both validation.
representation,
validation. Sec.
Sec.
visualization
5 summaries
visualization
5 summaries
and
and soft
the
soft
the
sensor
work
sensor
work and
and
development
discuss
development
discuss
and
future
and
future
the
aathe accuracy
accuracy and
distributed precision
parameter of
of measurements
system such locally
locally andand across
distributed and precision
parameter system such asas aa reactor.
measurements reactor. validation.
across directions.
validation.
directions.
Sec.
Sec. 5
5 summaries
summaries the
the work
work and
and discuss
discuss future
future
aa distributed
distributed parameter
parameter system system suchsuch asas aa reactor.
reactor. directions.
2405-8963 © 2019, IFAC (International Federation of Automatic Control) Hosting directions.
by Elsevier Ltd. All rights reserved.
Peer review under responsibility of International Federation of Automatic Control.
Copyright © 2019 IFAC 562
10.1016/j.ifacol.2019.06.122
Copyright © 2019 IFAC 562
Copyright © 2019 IFAC 562
Copyright © 2019 IFAC 562
 2019 IFAC DYCOPS
Florianópolis - SC, Brazil, April 23-26, 2019Devarshi Shah et al. / IFAC PapersOnLine 52-1 (2019) 562–567 563
2 TESTBED SETUP
2.1 Process Selection
A manufacturing plant consists of different unit operations.
Each unit operation consists of several different equipment.
Health of any given unit operation is dependent on healthy
functioning of these equipment. This makes proper
functioning of each equipment important and also
interdependent. One of the most versatile equipment in
manufacturing industry are centrifugal pump and compressors
and the associated piping system that move gas or liquid from
one location to another. Therefore, we choose a simple pipe
flow system of water that is driven and controlled by a
centrifugal pump and valves as shown in Fig. 1. The motor
revolutions per minute (rpm) and water flow rate gallon per
minute (gpm) were measured and displayed on a computer
screen with updates every second. The goal is to predict these
two key process variables through IoT sensor measurements.
Although the testbed is simple, the principles developed based
on it can be generalized to more complex real systems.
Fig. 1 The testbed of a simple pipe flow system with a
centrifugal pump and valves
2.2 Sensor Selection
One of the most commonly utilized data in industrial or
mechanical equipment is vibration. The application of
vibrational data for condition monitoring of machinery or
structure has been well documented, such as the detection of
faults or defects in gears, rotors, shafts, bearings and couplings
(Adams, 2009; Carden and Fanning, 2004; Lyon, 2013;
Tandon and Choudhury, 1999). However, their applications
for inferring process information, such as rotor speed and fluid
flow rate inside a pipe that we set to explore in this work, have
not been reported. Therefore, in this work we choose to equip
the testbed with IoT vibration sensors. Another consideration
of choosing vibration sensors is that they can be installed non-
invasively, which is an important feature if we were to equip
legacy processes with advanced sensors. The following two
major types of vibration sensors were considered and tested in
this work: (1) Piezo type vibration sensors. This type of
sensors contain a material that produces voltage when moved
or touched, which is then measured to identify vibration or
motion of the surface where it is placed. (2) Accelerometers.
These sensors are electromagnetic devices that measure
acceleration force on the sensors which in turn can be used to
563
sense vibration or movements. Fig. 2 shows the pictures of
some the Piezo type and accelerometers we tested. Piezo type
vibration sensors are in general the cheapest, require minimal
connections and do not require complex breakout. However,
initial tests found that these sensors were not sensitive enough
for the testbed system, and most likely not suitable for
industrial processes if we were to infer process information
buried in the vibration signals. On the other hand, although IoT
accelerometers are slightly more expensive, they are still cheap
– usually less than 10 US dollars as the ones shown in Fig. 2.
These accelerometers are most responsive vibration sensors,
easy to connect either using standard I²C (Inter-Integrated
Circuit) or SPI (Serial Peripheral Interface) protocols.
Therefore, we chose these digital accelerometers in this study.
Piezo type sensors Accelerometers
Fig. 2 Some Piezo type and accelerometer sensors tested
2.3 Testbed setup
The sensors themselves do not have built-in controllers that
can directly communicate with local computers or cloud
servers. Therefore, Raspberry Pi’s were chosen to serve as the
“third-party” computing device or micro-controller for the
sensors. Totally five accelerometers were equipped to the
testbed and their schematic locations are outlined in Fig. 3.
Due to page limit, the protocols for data transmission,
processing and storage is omitted.
3 DATA AND SENSOR CHARACTERIZATION
The testbed pump can run from 1500 rpm to 2500 rpm. At
1500 rpm, the minimum flow rate that can be measured
reliably by the flow meter is 5 gpm. The maximum flowrate
that can be achieved at 2500 rpm and maximum discharge
valve opening is around 16 gpm. The accelerometers are triple
axis accelerometers and thus measure vibration signals in x, y
and z directions of Cartesian coordinate system. Table 1 shows
the experiments performed on the testbed. Totally 85
conditions (i.e., difference combinations of motor speed and
flow rate by adjusting motor speed and discharge valve) were
tested. Data were collected for 10 min for each condition. Note
that the motor speeds and flow rates listed in Table 1 are
approximate as they do drift during the course of the
experiments and the real-time readings from the computer
screen are used as the actual values. The accelerometer
Fig. 3 Schematic of sensor locations
2019 IFAC DYCOPS
Florianópolis - SC, Brazil, April 23-26, 2019Devarshi Shah et al. / IFAC PapersOnLine 52-1 (2019) 562–567
564

sensitivity is adjustable (±2g, ±4g, or ±8g) and ±8g is used in

this study. The accelerometer sampling rate is also adjustable
(800 Hz, 1600 Hz or 3200 Hz). After some tests, it was found
that for this particular testbed, 1600 Hz is sufficient to capture
the testbed vibration characteristics.
Table 1 Experiments performed on the testbed
Conditions Motor speed (rpm) Flow rate (gpm)
3 1500 5, 7, 9
3 1600 5, 7, 9
4 1700 5, 7, 9, 11
4 1750 5, 7, 9, 11 (a) original signal
4 1800 5, 7, 9, 11
4 1850 5, 7, 9, 11
4 1900 6, 8, 10, 12
4 1950 6, 8, 10, 12
5 2000 5, 7, 9, 11, 13
5 2050 5, 7, 9, 11, 13
5 2100 6, 8, 10, 12, 14
5 2150 6, 8, 10, 12, 14
5 2200 6, 8, 10, 12, 14
5 2250 6, 8, 10, 12, 14
5 2300 7, 9, 11, 13, 15
5 2350 7, 9, 11, 13, 15 (b) zoomed-in view
5 2400 7, 9, 11, 13, 15
5 2450 7, 9, 11, 13, 15 Fig. 4 High frequency noisy measurements
5 2500 8, 10, 12, 14, 16 To

3.1 Unequal Sampling Interval

Although the sensor sampling rate on each accelerometer is set

to be 1600 Hz, the sampling rate is ultimately determined by
the code running on the Raspberry Pi that queries the sensor
data. Therefore, the sampling rate varies from cycle to cycle
due to CPU time variations. This raw data can be
downsampled to get uniform sampling rate or interval but the
downsampling effect on the signal would need to be studied.
Therefore, the raw data collected with variable sample rate are
used in this work and more details are provided in Sec. 4.1. Fig. 5 Chunks of missing values in the raw signal

3.2 High Noise Levels
Because of the high sensitivity of the accelerometer used in
this work, the signals obtained are very noisy as shown in Fig.
4. There are denoising methods such as various filtering
techniques. However, the effect of denoising on signal
distortion and information loss can vary depending on the
method and associated parameters. Therefore, in this work we
opt to use the raw signal.
3.3 Missing Values in Measurements
From time to time, missing measurements have been observed
as shown in Fig. 5, where the segments with missing values
are highlighted by the ellipses. This might be due to occasional
connection failures or communication delays between the
micro-controller and the sensor. There are techniques to
impute the missing values such as interpolation and signal
binning. But there are potential issues associated with data
imputation, especially for our case where missing values are
usually cluttered into chunks. Therefore, we opt to use the raw
signal.
summarize, the raw signals are unequally sampled with
significant noise and some missing values. There are
techniques available that can be used to pre-process the data to
get equal sampling intervals with reduced noise and imputed
missing values. However, we opt to not use any data pre-
processing or cleaning techniques with the consideration of
preserving whatever signatures or features in the raw signal
from being distorted by those data pre-processing techniques.
As a result, we need to develop methods that are robust to these
data imperfections, which is presented in the next section.
4 SOFT SENSOR DEVELOPMENT AND VALIDATION
4.1 Vibration Signal Representation
Many analysis techniques have been developed for vibration
signal analysis, which can be classified into simple magnitude
analysis, time domain analysis and frequency domain analysis
(Norton and Karczub, 2003). Although most of these
techniques are developed for condition monitoring such as
fault detection, the approaches of representing vibration
signals in different domains are the same. Because of the high

564
 2019 IFAC DYCOPS
Florianópolis - SC, Brazil, April 23-26, 2019Devarshi Shah et al. / IFAC PapersOnLine 52-1 (2019) 562–567 565
frequency sampling, it was decided to use frequency domain
representation of the signals. In general, for equally spaced
signals, frequency domain representation can be obtained
through mathematical Fourier transform such as fast Fourier
transform (FFT). However, the vibration signals obtained from
the testbed are more complicated than the traditional ones,
with data imperfections (including unequal sampling)
discussed in the previous section. Therefore, we decided to use
Lomb’s algorithm to obtain power spectral density (PSD),
which does not require signal to be equally spaced.
4.2 Lomb’s Algorithm
The output of the Lomb’s algorithm is PSD of the signal under
consideration. It does not require samples to be equally spaced,
and when they are equally spaced, the mathematics of the
algorithm reduces to Fourier transform (Lomb, 1976; Scargle,
1982). Lomb’s algorithm is a subset of least-squares spectral
analysis and has been widely used in astronomy community.
Lomb’s algorithm assumes a signal as a function of cosine and
sine:
𝑃𝑃(𝑎𝑎, 𝑏𝑏, 𝑓𝑓, 𝑡𝑡) = 𝑎𝑎 cos(2𝜋𝜋𝜋𝜋𝜋𝜋) + 𝑏𝑏sin⁡(2𝜋𝜋𝜋𝜋𝜋𝜋)
(1)
where 𝑎𝑎 and 𝑏𝑏 are amplitudes, 𝑓𝑓 frequency and 𝑡𝑡 time of
sampling.
𝑃𝑃 is then fitted to the signal using a least square approach.
The sine and cosine terms are made orthogonal by shifting
the signal in time and identifying delay shift τ (offset). The
following two equations identify 𝜏𝜏 and power spectrum 𝑆𝑆 =
∑𝑁𝑁𝑖𝑖=1 𝑃𝑃 [39], [40]:
2 flow rates, although not as obvious as the first PC.
1 𝑖𝑖=1 sin(2𝜔𝜔𝑡𝑡𝑖𝑖 )
∑𝑁𝑁
𝜏𝜏 = (𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎⁡( 𝑁𝑁 )⁡
2𝜔𝜔 ∑𝑖𝑖=1 cos(2𝜔𝜔𝑡𝑡𝑖𝑖 )
1 [∑𝑁𝑁𝑖𝑖=1 𝑋𝑋𝑖𝑖 cos(𝜔𝜔(𝑡𝑡𝑖𝑖 − 𝜏𝜏))]
2 [∑𝑁𝑁
𝑖𝑖=1 𝑋𝑋𝑖𝑖 sin(𝜔𝜔(𝑡𝑡𝑖𝑖 − 𝜏𝜏))]
2
𝑆𝑆(𝜔𝜔) = ( +
2𝜎𝜎 2 ∑𝑁𝑁𝑖𝑖=1 𝑐𝑐𝑐𝑐𝑐𝑐 2 (𝜔𝜔 (𝑡𝑡 − 𝜏𝜏))
𝑖𝑖 ∑ 𝑁𝑁
𝑖𝑖=1 𝑠𝑠𝑠𝑠𝑠𝑠 2 (𝜔𝜔 (𝑡𝑡 − 𝜏𝜏))
𝑖𝑖
where 𝜎𝜎 is the variance of the signal, 𝑁𝑁 the number of
observations, 𝜔𝜔 = 2𝜋𝜋𝜋𝜋, and 𝑋𝑋 signal values after mean
centering.
4.3 Construction of Independent and Dependent Variable
Matrices
In this work, the sampling rate of vibration signals is much
higher than that of motor speed (in rpm) and water flow rate
(in gpm), which were measured and displayed on a computer
screen with update every second. But there was no mechanism
to record these measurements to the local computer. Therefore,
videos were taken during the experiments and image
processing techniques were used to extract rpm and gpm every
0.33 second. This sampling frequency is sufficient as the
screen displayed values were updated every second. Because
of the sampling rate difference, it was decided to obtain
corresponding PSD for each rpm and gpm measurement. In
total 801 data points were used for getting PSD: 400 data
points each before and after the time when rpm/gpm value is
captured. Thus for each measurement of rpm and gpm, we
have corresponding segment of data points from which PSD
can be obtained using Lomb’s algorithm. In order to reduce
spectral leakage and obtain smoother spectrum, mean-centered
signal is passed through a window function (Cerna and
Harvey, 2000; Harris, 1978; Lyon, 2009; Nuttall, 1981). In this
study Hann window function was used. PSD for frequencies
from 1 to 800 Hz with resolution of 0.2 Hz is obtained. Thus
the columns of the independent variable matrix 𝑋𝑋 consists of
the amplitude of PSD at each frequency. The columns of the
dependent variable matrix 𝑌𝑌 consists of rpm and gpm. The
rows of 𝑋𝑋 and 𝑌𝑌 correspond to the samples computed or
measured every 0.33 second.
4.4 Data Visualization and Initial Analysis through PCA
First, principal component analysis (PCA) is used to
qualitatively examine if the vibration signals contain sufficient
information that can be used to predict rpm and/or gpm and
such information is not overshadowed by the data
imperfections inherent to the IoT sensors used for the testbed,
i.e., unequal sampling, significant noise, and missing chunks
of values as observed in Figs. 4-6. For this initial analysis, it
was decided to carry out PCA for fixed rpm conditions, i.e.,
samples/spectrums included in matrix 𝑋𝑋 have fixed rpm value
but can have different gpm. Fig. 6 (a) and (b) show the score
plots comparison for different flow conditions for 2400 rpm
and the first 500 points from each condition were used to
construct 𝑋𝑋. PCA was carried out on mean centered 𝑋𝑋. The
variance was not scaled as all the variables have the same unit.
Fig. 6 (a) clearly shows that the first PC scores are very
different for different flow rates, Fig. 6 (b) also indicates the
presence of difference between signals captured for different
)
(a) (b)
Fig. 6 PCA score plots for the first PC (a) and second PC (b).
Fig. 6 (a) can be interpreted using the basic idea of wave
formation in a mechanical system. Whenever the load is
applied on a freely vibrating object, the amplitude of vibration
may increase or decrease depending on the relation with load
and natural vibrating frequency. As flowrate increases, load on
pump assembly increases and so does load on several pump
components, therefore when pump was running at 2400 rpm
with load increasing (i.e., flow rate increasing), overall
vibration amplitude of several rotating and vibrating parts of
the pump changes. PCA captures the changes in frequencies
and suggests that overall amplitude change is decreasing as
flowrate increases as indicated in Fig. 6 (a), but not in all
components/frequencies as indicated in Fig. 6 (b). Fig. 6 also
suggests that there is information contained in the vibration
signals that are associated with flow rate that is not
overshadowed by the data imperfections discussed previously.
4.5 Inferring rpm from Vibration Signal
For this initial analysis, signals collected from sensor #4 (see
565
2019 IFAC DYCOPS
Florianópolis - SC, Brazil, April 23-26, 2019Devarshi Shah et al. / IFAC PapersOnLine 52-1 (2019) 562–567
566
Fig. 3) placed on the coupling is used to infer rpm as coupling
is the connection between impeller and motor and therefore
will be directly affected by rpm change. Inspections of spectra
at different rpm's indicate that there is a linear relationship
between rpm and the frequency of the highest PSD peak. Also
it appears that the amplitude of the peak does not matter and it
is affected by the load on the pump (i.e., flow rate). Therefore,
a vector of zeros with length 1000 is generated corresponding
to frequencies with increment of 0.2 Hz. Then the frequency
of the highest PSD peak is identified and the zero in the vector
at that frequency is replaced with one. To robustify the method
considering the data veracity, four zeros corresponding to two
adjacent frequencies on each side of the identified frequency
are replaced with one’s as well. The above procedure is
performed for all samples to generate a matrix consisting of
zero’s and one’s. For this initial analysis, a partial least squares
(PLS) model was built using rpm cases of 1500, 1800, 2000,
2300 and 2400 with different flow rates for each rpm as shown
in Table 1, which combine to 22 conditions (i.e., different
combination of rpm and gpm). The calibration set consists of
200 samples from each condition. Thus the calibration matrix
𝑋𝑋𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 is a 44001000 logical matrix. The PLS model was
tested for the 22 conditions using 200 new samples from each
condition. Thus the test matrix 𝑋𝑋𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 has the same dimension
as 𝑋𝑋𝑇𝑇𝑇𝑇𝑎𝑎𝑖𝑖𝑖𝑖 . The 𝑌𝑌 matrices for training and testing were consists
of the rpm for each of the corresponding samples. The number
of PC was chosen as two. Fig. 7 demonstrates excellent
agreement between the measured and predicted rpm’s on the
test samples, indicating good performance of the PLS model.
Fig. 7 Predicted vs. measured rpm’s.
4.6 Inferring Flow Rate from Vibration Signal
As shown in the previous section, the rpm inference is quite
simple and reliable as the frequency of the highest PSD peak
is not affected by the flow rate change when vibration signals
collected from the coupling were used. The inference of flow
rate from vibration signals is much harder as the flow rate is
affected by both rpm and the discharge valve opening. As
discussed in the previous section, the flow rate affects the
amplitude of PSD peaks. Therefore, the amplitudes of PSD
peaks over the frequency of 1-800 Hz were used to build a PLS
model. We first investigate the possibility of predicting flow
566
rate under different rpm conditions using all flow rate
conditions for 1500, 1800, 2000, 2300 and 2400 rpm, which
include both low and high rpm’s, and totally 22 conditions.
Again, the signals collected from sensor #4 placed on the
coupling is used to infer flow rate. The vibration signal from
both x and z directions are used. The first 500 samples (i.e.,
PSD spectrum obtained from a vibration signal) from each of
the 22 conditions were stacked row wise (i.e., totally 11000
samples) for calibrating the PLS model. Another 250 samples
from each of the 22 conditions (i.e., totally 5500 samples) were
used for testing. Thus the final 𝑋𝑋𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 and 𝑋𝑋𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 have the
dimension of 110007992 and 55007992, respectively.
𝑦𝑦𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 and 𝑦𝑦𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 have the dimension of 110001 and 55001,
respectively. The comparison of the predicted and measured
flow rates of the test data for all 22 conditions is shown in Fig.
8, which indicates that the PLS model was able to predict the
flow rates in the vicinity of the true values but the predictions
are not very accurate. The root mean squared error (RMSE) of
the prediction is 0.60.
Fig. 8 Predictions vs. measurements for multiple flowrates at
multiple rpm’s.
To improve model prediction accuracy, we investigate the case
where rpm is known or has been predicted independently as
shown in the previous section. Under such a case, separate
models can be built for the same or similar rpm condition. In
this work, separate models for different rpm’s were calibrated
and tested. For each model the first 500 samples from each
condition were used to construct 𝑋𝑋𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 and 𝑦𝑦𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 for
calibration and another 250 samples from the same condition
were used to construct 𝑋𝑋𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 and 𝑦𝑦𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 for testing. Due to
limited space, only one low rpm of 1800 and one high rpm of
2400 cases are presented here. Same as the mixed rpm
modeling, vibration signals from x and z directions of sensor
#4 were used. Fig. 9 (a) and (b) compare the predicted flow
rates to the measured ones for 1800 rpm and 2400 rpm,
respectively. Fig. 9 shows that the prediction performance of
models built based on separate rpm’s perform better than a
single model built including all rpm’s. The RMSE’s for 1800
rpm and 2400 rpm are 0.35 and 0.43, respectively.
5 CONCLUSIONS AND DISCUSSIONS
In this work, we introduced the design of an IoT testbed using
multi-stage centrifugal pumping system equipped with non-
invasive IoT vibration sensors. We studied the characteristics
of the data collected from these IoT sensors, focusing on data
veracity (i.e., unequal sampling intervals, significant noise and
 2019 IFAC DYCOPS
Florianópolis - SC, Brazil, April 23-26, 2019Devarshi Shah et al. / IFAC PapersOnLine 52-1 (2019) 562–567 567
missing values) and its challenges for data analytics. We
demonstrated that some robust methods such as Lomb’s
algorithm can properly handle these data veracity
characteristics. We also discussed data volume and velocity
resulted from high frequency sampling and showed that proper
data representation (e.g., in frequency domain) can help
overcome this challenge. Finally, we developed data-driven
predictive models using the frequency domain representation
of the vibration signals to infer key process variables for
process monitoring, namely the flow rate inside the pipe and
rpm of the pump motor. The model predictions were validated
with experimental measurements. Altogether, this study serves
as a demonstration of how IoT sensors and big data analytics
can be integrated and utilized for real-time process monitoring.
(a)
(b)
Fig. 9 Predicted vs. measured flowrates based on individual
rpm models
It is worth noting that this work presents some initial results of
this project, which is far from comprehensive and there are
many areas that are worth further investigation, such as
exploring other modeling approaches such as neural networks,
utilizing data collected from all sensors instead of just one as
in this work, improving soft sensor performance through
variable selection, etc.
6 ACKNOWLEDGEMENT
This work was supported by National Science Foundation
567
(CBET-1547124, CBET-1547163, and CBET-1805950).
REFERENCES
Adams, M.L., 2009. Rotating machinery vibration: from
analysis to troubleshooting. CRC Press.
Carden, E.P., Fanning, P., 2004. Vibration based condition
monitoring: a review. Struct. Health Monit. 3,
355–377.
Cerna, M., Harvey, A.F., 2000. The fundamentals of FFT-
based signal analysis and measurement. Natl.
Instruments, Junho.
Davis, J., Edgar, T., Graybill, R., Korambath, P., Schott,
B., Swink, D., Wang, J., Wetzel, J., 2015. Smart
Manufacturing. Annu. Rev. Chem. Biomol. Eng.
6, 141–160.
Davis, J., Edgar, T., Porter, J., Bernaden, J., Sarli, M.,
2012. Smart manufacturing, manufacturing
intelligence and demand-dynamic performance.
Comput. & Chem. Eng. 47, 145–156.
Harris, F.J., 1978. On the use of windows for harmonic
analysis with the discrete Fourier transform. Proc.
IEEE 66, 51–83.
Lomb, N.R., 1976. Least-squares frequency analysis of
unequally spaced data. Astrophys. Space Sci. 39,
447–462.
Lyon, D.A., 2009. The discrete fourier transform, part 4:
spectral leakage. J. Object Technol. 8.
Lyon, R.H., 2013. Machinery noise and diagnostics.
Butterworth-Heinemann.
Manyika, J., Chui, M., Brown, B., Bughin, J., Dobbs, R.,
Roxburgh, C., Byers, A.H., Institute, M.G., 2011.
Big data: The next frontier for innovation,
competition, and productivity.
Norton, M.P., Karczub, D.G., 2003. Fundamentals of noise
and vibration analysis for engineers. Cambridge
university press.
Nuttall, A., 1981. Some windows with very good sidelobe
behavior. IEEE Trans. Acoust. Speech, Signal
Process. 29, 84–91.
Qin, S.J., 2014. Process data analytics in the era of big
data. AIChE J. 60, 3092–3100.
Scargle, J.D., 1982. Studies in astronomical time series
analysis. II-Statistical aspects of spectral analysis
of unevenly spaced data. Astrophys. J. 263, 835–
853.
Steering Committee of the Advanced Manufacturing
Partnership 2.0, 2014. Report to the president:
accelerating U.S. advanced manufacturing.
Tandon, N., Choudhury, A., 1999. A review of vibration
and acoustic measurement methods for the
detection of defects in rolling element bearings.
Tribol. Int. 32, 469–480.
Zikopoulos, P., Eaton, C., 2011. Understanding big data:
Analytics for enterprise class hadoop and
streaming data. McGraw-Hill Osborne Media.
Zikopoulos, P., Parasuraman, K., Deutsch, T., Giles, J.,
Corrigan, D., 2012. Harness the Power of Big
Data The IBM Big Data Platform. McGraw Hill
Professional.