This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2927793, IEEE Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.Doi Number

Internet of Things based Monitoring of

Large Rotor Vibration with a
Microelectromechanical Systems
Accelerometer
Ivar Koene1, Raine Viitala1, Petri Kuosmanen1
1
Department of Mechanical Engineering, Aalto University, 02150 Espoo, Finland
Corresponding author: Ivar Koene (e-mail: [email protected]).

This work was supported by Academy of Finland (Digital Twin of Rotor System, grant number 313675).

ABSTRACT In rotating machinery, excessive vibration can affect the lifetime of the machine and, for
example, in paper machines it can directly affect production quality. Hence it is important to monitor
vibrations. Typically the vibrations are measured with piezo-based accelerometers attached to the bearing
housing. Installation of the accelerometers and especially cabling is laborious and expensive in the case of
existing machines that have several components to monitor. In continuous process maintenance, a break
may be required as well. Therefore, typically only the critical rotors will be monitored, if at all. This
research focuses on applying wireless microelectromechanical systems (MEMS) accelerometers to the
measurement of large rotor vibration. The results indicated that MEMS accelerometers combined with
wireless communication can offer a viable alternative to more expensive piezo-based accelerometers with
traditional wire-based communication. The combination provides a flexible and cost-effective method for
the collection of vibration data from large rotors and rotor systems.

INDEX TERMS Internet of Things (IoT), microelectromechanical systems (MEMS), accelerometer,

subcritical vibration, large rotor, wireless condition monitoring

I. INTRODUCTION speed, vibration can occur. These are called harmonic

Large rotating machines are widely used, such as at factories, components. Fig. 1 presents a schematic figure of subcritical
ships and power plants. Each rotating machine vibrates. If the vibrations. Although the speed of the rotor remains below the
vibrations are excessive, they can pose problems to the usage critical speed, the vibration appears at the natural frequency
of the machine and reduce its lifetime. To measure the because the rotor can have several excitations per revolution.
vibrations and monitor the machine behavior, integrated If the rotor is rotating at one third of the critical speed and is
electronics piezo-electric (IEPE) accelerometers are typically subjected to three excitations per revolution, the vibration
adopted. However, they need to be physically connected to occurs at the natural frequency and a resonance is observed.
the measurement module by cabling, which can make the
setup of the sensors with existing machines expensive,
especially if the machine has several rotating components. A
paper machine is an example of a machine which can have
more than a hundred rotating rolls. Due to the high cost, only
the critical rotating components are typically equipped with
sensors.
Large rotors in industrial sized rotating machines can
weigh several hundred or thousand kilograms. The rotors are
typically operated below the critical speed of the rotor [1]. In
those cases, vibrations are subcritical harmonic vibrations,
which occur when the rotor is rotating at a fraction of the
critical speed. If the rotational frequency of the rotor is, for
instance, one half, one third or one fourth of the critical

VOLUME XX, 2017 1

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2927793, IEEE Access

of one of the MEMS accelerometers was poor for unknown

reasons. The two well-performing accelerometers proved that
MEMS accelerometers could be feasible for machine
condition monitoring. The fundamental speed and line
frequencies were captured as well as their multiplies.
The communication between a microcontroller and
MEMS accelerometer and the way the data are saved or
transmitted have significant effects on the sampling rate.
Chan and Huang [3] investigated the performance of a
MEMS accelerometer and a microcontroller. They applied a
universal asynchronous receiver transmitter (UART) to
transfer the data from the microcontroller to a computer in
real-time. The bandwidth of the UART was the limiting
factor for the data rate. However, when they used a batch
transfer method, they observed a sampling time limitation as
the microcontroller could only save a limited amount of data
to the static random access memory (SRAM). In addition,
when they tested jitter in their system, they reported that it
can cause sidebands in the measurement data [3]. For this
FIGURE 1. Schematic figure of the subcritical vibrations. ωn is natural reason, it is important to stabilize the triggering of the
frequency of the system and 1X, 2X, 3X and 4X are harmonic measurement.
components.
To investigate different methods for vibration
Microelectromechanical systems (MEMS) technology has measurement, accelerometers have been mounted to the
been advanced by the growing demand for cellular phones surface of the rotor. Elnady et al. [7], [8] explored this
and tablets. This technology enables cost-effective method with wireless MEMS accelerometers to see whether
accelerometer production. Prices are decreasing while it would be possible to identify the critical speed of a rotor
accuracy is improving, which has provided new applications [7]. They were able to measure the critical speed of the rotor,
for MEMS accelerometers. The typical maximum sampling which, however, did not show as a frequency peak in the
rate of a MEMS accelerometer is a few kilohertz and large plot. The critical speed was detected as a mean value of two
rotors are typically operated below critical speeds. Hence, the peaks present in the frequency plot. These peaks occurred
operating frequencies are low enough for the MEMS when the velocity was close to a harmonic component. In
accelerometers to measure rotor vibrations. In addition, addition, they compared a finite element model and an
MEMS sensors are compact and they have low power analytical model of an on-shaft accelerometer and employed
consumption. These features are useful in the internet of the measurement to validate the models [8].
things (IoT) applications because typically a small size and Feng et al. [4] investigated the condition monitoring of a
battery powered operation are required of IoT sensors. Piezo- reciprocating compressor by attaching a MEMS
based accelerometers are widely adopted in industry but new accelerometer on the shaft of the flywheel. They examined
MEMS accelerometers can replace them in some different leakage fault cases and found that when monitoring
applications. The MEMS accelerometers have been the pressure together with the fundamental speed and the
investigated for condition monitoring and compared to IEPE third harmonic components of the vibration of the rotor, it is
accelerometers [2]–[9]. possible to detect faults in the compressor.
Albarbar et al. [6], [9] tested the performance of MEMS In the previous studies, the MEMS accelerometer was
accelerometers and compared the results with an IEPE placed on the surface of the rotor [4], [7], [8] but Jiménez et
accelerometer with a particular interest in sinusoidal, random al. [2] inserted the MEMS accelerometer inside the rotor.
and impulse excitations. They compared one capacitive They compared three different imbalance conditions and
MEMS accelerometer with an IEPE accelerometer and used eddy current sensors to verify the accelerometer results
observed that the MEMS accelerometer performed well with [2]. They measured vibrations during an impulse where they
different excitations [9]. However, some phase shift was observed the same phenomenon as Elnady et al. [7], i.e., the
observed when the MEMS accelerometer was exposed to data from the accelerometer showed frequency peaks at both
sinusoidal or random excitations. They conducted similar sides of the resonance frequency.
tests using three different capacitive MEMS accelerometers MEMS accelerometers have also been analyzed when
[6]. In addition, they tested them in a CNC machine to see monitoring the condition of a wind turbine blade [5].
how they operated in real industrial conditions. In laboratory Typically, the blade condition has been monitored with
tests, two of the three MEMS accelerometers received results piezoelectric accelerometers, but they are bulkier and more
similar to those observed in [9]. However, the performance expensive than the MEMS accelerometers. The small size of

VOLUME XX, 2017 2

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2927793, IEEE Access

the MEMS accelerometers allows for the fixing of the sensor The sensor unit adopted wireless communication to transfer
on the blade during manufacturing or retrofitting it the data.
afterwards. These studies have stated that it is possible to
utilize MEMS accelerometers for monitoring the condition of 1) COMPONENTS AND STRUCTURES
a wind turbine blade. The sensor unit included the following components:
In recent years, 3G, 4G and wireless local area networks • Pycom’s WiPy 3.0 (IoT development platform)
(WLAN) have become a part of everyday life, allowing easy • Pycom’s Expansion board 2.0 (expansion board for the
access to the internet. Wireless sensor networks (WSN) have WiPy 3.0)
been tested by exploiting wireless communication • ADXL355 (MEMS accelerometer)
technologies [10], [11]. MEMS sensors are typically applied • battery (2500 mAh, 3.7 V)
in these studies because of their low energy consumption and The WiPy 3.0 (Pycom) development platform had an
low price. Typically, at least one sensor is connected to an ESP32 microcontroller chip, which contains a built-in radio
outside network, which allows it to connect to the internet module for WLAN. It had two cores: one core typically
and send information to a server, where the data are stored operated the WLAN radio and wireless communication and
and are easily accessible. However, as WSN facilitates data the other executed other tasks. Two cores helped to optimize
collection, the amount of data can increase rapidly, which the performance of the controller. Table 1 presents the
can challenge the data storing and transmission [10], [12]. specifications of ESP32.
In rotating machinery, the sensor sampling rate may need
TABLE 1
to be relatively high in some situations in order to measure
SPECIFICATION OF ESP32 MICROCONTROLLER CHIP
the vibration accurately. When measuring the bearing
vibration of an AC motor, the frequencies of interest Specifications ESP32
typically rise between 1-5 kHz [13]. The sampling rate needs Microcontroller Unit Tensilica Xtensa 32-bit LX6
to be at least twice the highest frequency of interest, as the Cores 2
Nyquist-Shannon sampling theorem presents [14]. If the Clock frequency (MHz) 240
SRAM (KiB) 520
system has several rotating components, the amount of data 802.11 b/g/n WI-FI HT40
can be large. As a way of minimizing the volume of data GPIO 36
transferred and the transfer time, data compressing has drawn SPI/I2C interfaces 4/2
research interest [15]–[18]. Compressing the data also lowers
the power consumption [16]. However, information can be WiPy 3.0 was connected to Expansion board 2.0 (Pycom),
lost from the original data when compressed [17]. which enabled the powering of the controller with a battery.
A paper machine is a representative example of an item of The microcontroller was programmed with Arduino IDE and
rotating machinery that has several rotating rolls which C/C++ language.
requires vibration measurements ensure the high quality of The MEMS accelerometer was a capacitive accelerometer
production. However, attaching sensors to existing machine ADXL355 (Analog devices). The results from the ADXL355
lines is a costly process. The cabling is time consuming and accelerometer were compared to data from a Hansford
the setup process may require an expensive maintenance sensors’ HS-1005005001 IEPE accelerometer (Fig. 2 b).
break. Therefore, a wireless sensor unit was built and tested. Table 2 presents the specifications of both accelerometers.
The sensor unit could be mounted to the machine without a
maintenance break. MEMS accelerometers are more TABLE 2
economical, consume less energy and are more compact SPECIFICATIONS OF ADXL355 AND HS-1005005001
compared to IEPE accelerometers. The MEMS sensors can
ADXL355 HS-1005005001
be installed to each rotor of the machine and not merely for
the critical rotors. Measurement range (g*) (±2, ±4, ±8) ±16g*
Sensitivity (µg*/digit) (3.9, 7.8, 15.6)** 500 (mV/g*)
This benefits a digital twin as well, which collects data Sample rate (Hz) 3.906 – 4000 2 – 10000
from the physical entity and the data can be exploited to Noise density (µg*/Hz) 25 -
optimize the operation of the product or machine [19]. In Number of axis 3 1
addition, the data can be utilized in product development to *g = gravitational acceleration
make the product more optimally suit the designed **Depends on the measurement range
application.
The cDAQ-9191 wireless data acquisition unit and NI-
II. METHODS 9234 sound and vibration input module by National
Instruments were used to acquire the data from the reference
A. SENSOR UNIT accelerometer (HS-1005005001) (Fig. 2). The cDAQ-9191
A sensor unit was built to investigate whether a MEMS wireless data acquisition unit allowed wireless
accelerometer can be used to measure large rotor vibrations. communication between the apparatus and the computer.

VOLUME XX, 2017 2

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2927793, IEEE Access

However, cabling was still required to attach the sensors to 6. The sensor unit sent a UDP packet to the server to
the NI-9234 module. indicate that the measurement was finished.
7. Server saved the data to an ASCII-file.

FIGURE 2. (a) National Instruments' cDAQ-9191 wireless data

acquisition unit and NI-9234 sound and vibration input module. (b)
Hansford sensors’ HS-1005005001 IEPE accelerometer. HS-1005005001
has a magnet attached to the bottom of the sensor.

The casing of the built sensor unit (Fig. 3) was

manufactured from aluminum with a CNC milling machine.
ADXL355 and Expansion board 2.0 was screwed to the
casing, double-sided tape kept the battery in place, and
magnets with threads kept the sensor unit in place during the
measurement.

FIGURE 3. Built sensor unit. (a) presents components of sensor unit

and (b) presents the assembled sensor unit.

FIGURE 4. The operating principle of sensor unit.

2) OPERATION PRINCIBLE
The sensor unit was controlled with a browser-based 3) POWER CONSUMPTION
graphical user interface (GUI). Fig. 4 presents the operating Power consumption can be separated into two states:
principle and the procedure is explained below. measurement state (explained in Fig. 4) and idle state. In the
1. The server sent a UDP packet to the sensor unit idle state, the sensor unit is waiting for messages from the
when the user started the measurement from the server while all other functions are turned off. In the
GUI. The packet contained information about the measurement state, the average power consumption from the
sampling rate, sampling time and which axes were battery is 490 mW (current consumption is 132 mA) and in
used. the idle state it is 160 mW (current consumption is 42 mA).
2. The sensor unit started the measurement according
to the information received from the server. B. SIGNAL PROCESSING
3. WiPy 3.0 read the measurement data from the The ADXL355 accelerometer outputs the acceleration as a
sensor and saved it to the RAM of the digital value, which is formatted as a 20-bit two’s
microcontroller. complement value. The data were converted to m/s2 using the
4. When 48 samples were saved to the RAM, the scale factor of 3.9 µg/LSB and gravitational acceleration.
sensor unit sent the data on the server using UDP.
5. Steps 3 and 4 were repeated until the sampling
time finished.

VOLUME XX, 2017 2

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2927793, IEEE Access

1) LOW-PASS FILTER C. ROTOR SETUP

ADXL355 has its own built-in low-pass filter set to 1/4 of the The measurements were made with a paper machine roll
sampling rate. The accelerometer also has a built-in high- (Fig. 5). The roll was mounted to a CNC grinding machine,
pass filter as well, but it was deactivated. The test rotor was which controlled the acceleration and speed of the roll. The
sampled at 1000 Hz, indicating that the cut-off frequency of right side of the roll was the driving side and the left side was
the low-pass filter was 250 Hz. The first bending mode the tending side. Fig. 6 presents the dimensions of the paper
natural frequencies (around 25 Hz in x direction and 29 Hz in machine roll. The roll weighed approximately 720 kg.
y direction) of the rotor denote that frequencies below 50 Hz The sensor units and reference accelerometers were fixed
are the interesting frequencies when investigating the to the bearing housings with magnets (Fig. 7). The reference
subcritical vibration behaviour of a large rotor. accelerometers were located on the top and side of the
bearing housing, since one sensor can only measure one
2) FOURIER ANALYSIS direction. The sensor unit can measure accelerations in three
Fourier transform converts a time domain signal into the
frequency domain. The principle of the Fourier transform is
that any signal can be divided into an infinite amount of
sinusoidal functions, which reveals the frequencies and
amplitudes of the signal components [20]. In rotating
machinery, vibrations are typically monitored to prevent
damage on the machine or a reduction on the production
quality. The vibration frequency and amplitude are the
interesting components of the data. Hence, Fourier transform
is applied. However, Fourier analysis may need ample FIGURE 6. Dimensions of the paper machine roll.

computing power as the amount of data increases. To directions, thus one sensor unit per bearing housing sufficed.
expedite the transform calculations, Cooley and Tukey [21] Both bearing housings were equipped with sensors.
developed a faster method to perform numerical Fourier
transfer called the fast Fourier transform (FFT). FFT applies
complex numbers to reduce the required computing power.
For the measured acceleration data, FFT was applied in the
form of a short-time Fourier transform (STFT). STFT makes
possible the observation of frequency change over time [22].
It divides the original time domain signal into smaller
segments and takes the FFT measurement from each segment
separately. The method allows the monitoring of the
frequencies in one selected time window and represents all of
the segments as a 3D plot, which has time (s), frequency (Hz)
and amplitude (m/s2) as axes. These plots are called
spectrograms. Spectrograms are widely exploited in sound
and speech analysis [23]–[25] but can also be applied to the
analysis of how vibrations change over time.
When the vibrations are measured during a rotor
acceleration, the measured signal is not periodic. Hence, the
FIGURE 7. The measurement setup of the tending side. Two reference
signal is manipulated to mimic a periodic signal using accelerometers and one sensor unit were installed for both bearing
windowing function, thus making FFT feasible. In the housings.
present study, Hanning window was applied.

VOLUME XX, 2017 2

FIGURE
This work is licensed under 5. Paper
a Creative machine
Commons roll used
Attribution during
3.0 License. Forthis
morestudy in the see
information, CNC grinding machine.
http://creativecommons.org/licenses/by/3.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2927793, IEEE Access

D. MEASUREMENT PROCEDURE 2. The rotor was accelerated to 240 rpm before the
The measurement was made during an acceleration from 0 measurement was started.
to 1060 rpm with an approximate acceleration of 0.62 3. The first measurement was taken at that speed.
rad/s2 (5.9 rpm/s2). The sampling rate of the sensor unit The measurement lasted 5.45 seconds.
was set to 1000 Hz. STFT was exploited to observe the 4. The rotor was accelerated by 6 rpm and a new
results from the measurement. measurement was taken when the speed was
The sweep measurement proceeded as follows: stabilized.
1. One sensor unit and two reference sensors were 5. Step 4 was repeated until the speed of 1060 rpm
mounted to both the drive and the tending side was achieved. Every measurement lasted
bearing housings (Fig. 7). approximately 5.45 seconds.
2. The rotor was accelerated from 0 to 1060 rpm 6. Data processing and plots were realized with
while measuring the acceleration of the bearing MATLAB.
housing in x (horizontal) and y (vertical)
directions. The sensor units had a sampling rate of III. RESULTS
1000 Hz and the reference accelerometers had an Figures from 8 to 11 present the vibrations during the sweep
approximate sampling rate of 1650 Hz. measurement, as the rotor was constantly accelerated. 1X,
3. Data processing and plots were realized with 2X, 3X, 4x and 5X point to the harmonic components of the
MATLAB. vibrations. 1X presents the first harmonic component, which
To compare the amplitudes of the vibration received from is equal to the rotational velocity of the rotor, 2X presents the
the sweep acceleration measurements, a step-by-step sweep second harmonic component, 3X presents the third harmonic
measurement was made. However, only the reference component etc. The natural frequency marked in the picture
accelerometers were applied during this measurement. The presents the natural frequency of the first bending mode.
measurement was conducted by accelerating the rotor from Fig. 8 and Fig. 9 present vibrations at the tending side of
240 to 1060 rpm in 6 rpm steps. After each step, a the rotor and Fig. 10 and Fig. 11 present the vibrations on the
measurement was taken when the rotor was rotating at a driving side of the rotor. The vibrations occurring in the
constant speed. vertical direction (Fig. 9 and Fig. 11) have lower amplitudes
The step-by-step sweep measurement proceeded as compared to horizontal direction (Fig. 8 and Fig. 10). The
follows: MEMS and reference sensors measured a horizontal natural
1. Reference accelerometers were mounted to the frequency between 25 Hz and 25.3 Hz and the vertical
bearing housing as presented in Fig. 7. The natural frequency between 29.3 Hz and 29.5 Hz according to
approximate sampling rate of the sensors was 1650 the MEMS sensors, and between 29.3 Hz and 29.7 Hz
Hz. according to the reference accelerometers.

FIGURE 8. Horizontal vibration of the tending side of the roll. 1X, 2X, 3X, 4X and 5X represent harmonic components of vibration. Natural frequencies
marked in the picture are the natural frequency of the first bending mode. (a) presents data from the MEMS accelerometer and (b) presents data from the
reference accelerometer.

VOLUME XX, 2017 2

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2927793, IEEE Access

FIGURE 9. Vertical vibration of the tending side of the roll. 1X, 2X, 3X, 4X and 5X represent harmonic components of vibration. Natural frequencies
marked in the picture are the natural frequency of the first bending mode. (a) presents data from the MEMS accelerometer and (b) presents data from
the reference accelerometer.

FIGURE 10. Horizontal vibration of the driving side of the roll. 1X, 2X, 3X, 4X and 5X represent harmonic components of vibration. Natural frequencies
marked in the picture are the natural frequency of the first bending mode. (a) presents data from the MEMS accelerometer and (b) presents data from
the reference accelerometer.

FIGURE 11. Vertical vibration of the driving side of the roll. 1X, 2X, 3X, 4X and 5X represent harmonic components of vibration. Natural frequencies
marked in the picture are the natural frequency of the first bending mode. (a) presents data from the MEMS accelerometer and (b) presents data from the
reference accelerometer.

VOLUME XX, 2017 2

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2927793, IEEE Access

The frequency peaks received from the step-by-step sweep IV. DISCUSSION
measurement were compared to the peaks received from the The sensor unit was able to measure the required vibrations.
step-by-step sweep measurements. Table 3 and Table 4 The main harmonic components (the first, second, third,
present the amplitudes and frequencies of the frequency fourth and fifth harmonic components) and the natural
peaks measured from the driving and of the rotor with the frequency of the first bending mode could be observed from
MEMS and reference accelerometers during the step-by-step the data. There were other non-integer harmonic components
sweep measurement, and reference accelerometers during present in the spectrograms as well, which were produced by
sweep measurement. Fig. 12 presents the spectrogram of the the bearings. The outer ring of the bearing is the cause of
step-by-step sweep measurement and sweep measurement most of these vibrations. The equations are presented e.g., by
side by side. [26]. Fig. 13 identifies some of the vibration components
TABLE 3 caused by the outer rings of the bearings.
NATURAL FREQUENCIES (F) AND THEIR AMPLITUDES (A) ON THE DRIVING Natural frequencies were presented clearly in the sensor
SIDE OF THE ROTOR IN VERTICAL DIRECTION
unit data and they were similar to the natural frequencies
Reference sensor Reference sensor captured with reference sensors. However, as observed in
MEMS sensor (sweep (step-by-step Table 3 and 4, the amplitudes of the frequency peaks were
Harmonic
measurement) sweep
subcritical
measurement) not similar when the measurement was taken during the rotor
resonances
A f A f A f constant acceleration (sweep measurement) and during step-
(m/s2) (Hz) (m/s2) (Hz) (m/s2) (Hz) by-step acceleration. The signal is not periodic and the
2 0.027 29.5 0.023 29.7 0.044 29.4 vibrations are changing constantly during the measurement.
3 0.016 29.3 0.013 29.3 0.016 29.4 Hence, the amplitudes depend on how long segments are
4 0.010 29.5 0.010 29.7 0.014 29.6 taken from the data when the spectrograms are made. If the
5 0.012 29.5 0.016 29.3 0.020 29.6
segment is long, there are less peaks at a certain frequency
TABLE 4 compared to the length of the segment, which indicates lower
NATURAL FREQUENCIES (F) AND THEIR AMPLITUDES (A) ON THE DRIVING
SIDE OF THE ROTOR IN HORIZONTAL DIRECTION
amplitudes for the peaks. In contrast, if the segment is short,
there are more peaks compared to the length of the segment,
Reference sensor Reference sensor which indicates that the amplitudes are higher.
MEMS sensor (sweep (step-by-step
Harmonic There was a difference between the magnitudes of the
subcritical measurement) sweep
measurement) vertical and horizontal frequency peaks due to the foundation
resonances
A f A f A f stiffness of the rotor system. The foundation was more rigid
(m/s2) (Hz) (m/s2) (Hz) (m/s2) (Hz) in the vertical than in horizontal direction, and thus the
2 0.055 25.0 0.052 25.0 0.050 25.3 amplitudes of the vertical vibrations were lower and the first
3 0.087 25.0 0.072 25.0 0.076 25.5 bending mode natural frequency was higher.
4 0.078 25.3 0.067 25.0 0.056 25.5
5 0.11 25.0 0.12 25.0 0.21 25.5

FIGURE 12. Vertical vibration of the driving side of the roll. 1X, 2X, 3X, 4X and 5X represent harmonic components of vibration. Natural frequencies
marked in the picture are the natural frequency of the first bending mode. (a) presents data from step-by-step sweep measurement with the reference
sensor and (b) presents data from sweep measurement with the reference sensor.

VOLUME XX, 2017 2

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2927793, IEEE Access

FIGURE 13. Vibrations caused by the outer bearing ring in horizontal direction of the driving side. The seven first harmonic components of the outer
bearing ring are marked with OR1 to OR7. (a) presents data from the MEMS accelerometer and (b) presents data from the reference accelerometer.

The results show that the MEMS accelerometers were In this way the sensor would rotate with the rotor and reveal
suitable for measuring the subcritical vibrations of a large more information about the rotor behavior. These types of
rotor from the bearing housing. This will open new studies have been conducted for smaller-scale rotors [2], [4],
opportunities to collect more data from the large rotating [7], [8] but not for large-scale ones.
machineries. The installation needs minimal effort and the
sensors are cost effective, which makes possible the V. CONCLUSION
measurement of several components of the machine instead The present study investigated a more cost-effective and
of merely the critical components. More data can be easier way to measure vibrations from large rotating
measured, and the usage of the machine can be better machines. It is time consuming and expensive to mount a
optimized. Furthermore, the need for maintenance can be wired sensor to already existing machines that include
estimated more accurately. several rotating components. This problem was addressed
The digital twin of a rotor can also benefit from the sensor with a wireless sensor unit, which contained a MEMS
units. They can be connected to the internet, which makes it accelerometer and a microcontroller with a WLAN module
possible to send the data easily to a digital twin. These to transfer data. The sensor unit was powered by battery
sensors are controlled over WLAN and by opening the without a need for cabling, which allowed implementation
network to the internet, they could be controlled from a with minimal effort. The measurement was made from a
distance. This is beneficial for digital twins. In addition, with bearing housing and one sensor unit could measure
the sensor units, digital twin can receive more data from the acceleration in three directions. Therefore, only one unit was
physical machine because more components of the machine required per bearing housing.
can be monitored. The results verified that the sensor unit could measure
This paper suggests that the MEMS accelerometers are vibrations from large rotors. The harmonic components and
suitable for monitoring the vibration of a large rotor. natural frequencies of the rotor could be observed from the
However, wireless communication still needs more research results. In addition, bearing vibrations could also be detected.
in order to establish more reliable data transfer. The data The amplitudes of the harmonic vibration components varied
were sent during the measurement, which connoted that there from the real amplitudes (i.e., amplitudes measured with
was a limited time window in which the data could be constant speed at that rotation speed) because the
transferred. To establish sufficiently fast communication, the measurement was taken during a constant acceleration of a
user diagram protocol (UDP) was utilized. However, its rotor and the vibration was constantly changing. In such as
downside was that it did not check if the data reached its case the signal is not periodic. Hence the incorrect
destination. Occasionally some of the data were subsequently amplitudes were not induced by a sensor unit defect, but they
lost during the transfer. This could be prevented by saving were inflicted by the measurement procedure.
the data to a memory during the measurement and by sending
it forward when the measurement is finished. This way, a ACKNOWLEDGMENT
slower but more reliable communication protocol could be The authors would like to thank Mr. Ville Klar for his
adopted, since it will not interfere with the measurement contribution to the development of the sensor unit.
process.
REFERENCES
Another interesting approach to the vibration measurement [1] R. Viitala, T. Widmaier, and P. Kuosmanen, “Subcritical vibrations
of a large rotor could be to attach the sensor unit to the rotor. of a large flexible rotor efficiently reduced by modifying the bearing

VOLUME XX, 2017 2

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2927793, IEEE Access

inner ring roundness profile,” Mechanical Systems and Signal [22] E. Sejdić, I. Djurović, and J. Jiang, “Time-frequency feature
Processing, vol. 110, pp. 42–58, Sep. 2018. representation using energy concentration: An overview of recent
[2] S. Jiménez, M. O. T. Cole, and P. S. Keogh, “Vibration sensing in advances,” Digital Signal Processing: A Review Journal, vol. 19, no.
smart machine rotors using internal MEMS accelerometers,” Journal 1, pp. 153–183, 2009.
of Sound and Vibration, vol. 377, pp. 58–75, Sep. 2016. [23] P. K. Ajmera, D. V. Jadhav, and R. S. Holambe, “Text-independent
[3] Y. J. Chan and J.-W. Huang, “Multiple-point vibration testing with speaker identification using Radon and discrete cosine transforms
micro-electromechanical accelerometers and micro-controller unit,” based features from speech spectrogram,” Pattern Recognition, vol.
Mechatronics, vol. 44, pp. 84–93, 2017. 44, no. 10–11, pp. 2749–2759, 2011.
[4] G. Feng, N. Hu, Z. Mones, F. Gu, and A. D. Ball, “An investigation [24] B. Pinkowski, “Principal component analysis of speech spectrogram
of the orthogonal outputs from an on-rotor MEMS accelerometer for images,” Pattern Recognition, vol. 30, no. 5, pp. 777–787, 1997.
reciprocating compressor condition monitoring,” Mechanical [25] B. E. . Kingsbury, N. Morgan, and S. Greenberg, “Robust speech
Systems and Signal Processing, vol. 76–77, pp. 228–241, Aug. 2016. recognition using the modulation spectrogram,” Speech
[5] O. O. Esu, J. a Flint, and S. J. Watson, “Condition Monitoring of Communication, vol. 25, no. 1, pp. 117–132, 1998.
Wind Turbine Blades Using MEMS Accelerometers,” Renewable [26] A. H. Slocum, Precision Machine Design. New Jersey: Prentice Hall,
Energy World Europe, pp. 1–12, 2013. 1992.
[6] A. Albarbar, S. Mekid, A. Starr, and R. Pietruszkiewicz, “Suitability
of MEMS Accelerometers for Condition Monitoring: An
experimental study,” Sensors, vol. 8, pp. 784–799, 2008.
[7] M. E. Elnady, J. K. Sinha, and S. O. Oyadiji, “Identification of
Critical Speeds of Rotating Machines Using On-Shaft Wireless IVAR KOENE was born in 1993. He received his
Vibration Measurement,” Journal of Physics: Conference Series, vol. B.Sc. and M.Sc. degrees from Aalto University in
364, p. 012142, 2012. 2016 and 2018, respectively. During his studies he
[8] M. E. Elnady, A. Abdelbary, J. K. Sinha, and S. O. Oyadiji, “FE and worked as a research assistant in a paper machine
Experimental Modeling of On-shaft Vibration Measurement,” in laboratory and now he is pursuing his Ph.D.
Proceedings of the 15th International Conference on Aerospace degree at Aalto University. He is focusing on how
Sciences & Aviation Technology, 2013, no. May, pp. 1–18. cost-effective IoT sensors could be utilized in the
[9] A. Albarbar, A. Badri, J. K. Sinha, and A. Starr, “Performance condition monitoring of a large rotating
evaluation of MEMS accelerometers,” Measurement: Journal of the machinery. He is also involved in a digital twin
International Measurement Confederation, vol. 42, no. 5, pp. 790– related research, where one of the goals is to
795, 2009. utilize digital twin for predictive maintenance.
[10] S. Lu, P. Zhou, X. Wang, Y. Liu, F. Liu, and J. Zhao, “Condition
monitoring and fault diagnosis of motor bearings using undersampled
vibration signals from a wireless sensor network,” Journal of Sound RAINE VIITALA was born in 1992. He received
and Vibration, vol. 414, pp. 81–96, 2018. his M.Sc. and D.Sc. degrees in mechanical
[11] R. Simon Carbajo, E. Simon Carbajo, B. Basu, and C. Mc Goldrick, engineering from Aalto University in 2017 and
“Routing in wireless sensor networks for wind turbine monitoring,” 2018. His doctoral dissertation was
Pervasive and Mobile Computing, vol. 39, pp. 1–35, Aug. 2017. acknowledged with the Aalto University School
[12] M. Xia, T. Li, Y. Zhang, and C. W. De Silva, “Closed-loop design of Engineering dissertation award. In 2018,
evolution of engineering system using condition monitoring through Viitala visited Technische Universität Darmstadt
internet of things and cloud computing,” Computer Networks, vol. researching rotating kinetic energy storages, i.e.,
101, pp. 5–18, 2016. flywheels. Viitala has a solid background in
[13] A. Vogl et al., “Design, process and characterisation of a high- experimental large rotor research, including
performance vibration sensor for wireless condition monitoring,” vibration analysis and subcritical vibration, bearing excitations, roundness
Sensors and Actuators, A: Physical, vol. 153, no. 2, pp. 155–161, measurements and manufacturing for operating conditions. His work can
2009. be widely applied in several industrial applications, such as electric motors
[14] H. Nyquist, “Certain topics in telegraph transmission theory,” Proc. and generators, turbines and paper machines. Viitala is involved in
IEEE, vol. 90, no. 2, pp. 280–305, 1928. developing an AI and simulation enhanced digital twin of a rotor system.
[15] W. Guo and P. W. Tse, “A novel signal compression method based He is instructing four doctoral candidates, three master’s thesis students
on optimal ensemble empirical mode decomposition for bearing and two bachelor’s thesis students. In addition, he lectures on a bachelor
vibration signals,” Journal of Sound and Vibration, vol. 332, no. 2, level course in Mechatronics Basics.
pp. 423–441, 2013.
[16] Q. Huang, B. Tang, L. Deng, and J. Wang, “A divide-and-compress
PETRI KUOSMANEN was born in 1963 and
lossless compression scheme for bearing vibration signals in wireless
he is the head of Engineering Design at Aalto
sensor networks,” Measurement: Journal of the International
University. Kuosmanen was appointed as a full
Measurement Confederation, vol. 67, pp. 51–60, 2015.
professor at Helsinki University of Technology
[17] W. Bao, W. Wang, R. Zhou, N. Li, J. Yang, and D. Yu, “Application
(currently Aalto University) in 1999. He
of a two-dimensional lifting wavelet transform to rotating mechanical
founded the paper machinery research group in
vibration data compression,” Proceedings of the Institution of
1990. He also established the 3D grinding
Mechanical Engineers, Part C: Journal of Mechanical Engineering
technology of large rotors, and innovated
Science, vol. 223, no. 10, pp. 2443–2449, 2009.
optimizing rotors for their operating conditions
[18] M. Oltean, J. Picheral, E. Lahalle, H. Hamdan, and J. Griffaton,
in industrial environments. Kuosmanen has
“Compression methods for mechanical vibration signals: Application
strongly promoted entrepreneurship in the
to the plane engines,” Mechanical Systems and Signal Processing,
development of new technology, both in industrial and medical
vol. 41, no. 1–2, pp. 313–327, 2013.
applications.
[19] S. Haag and R. Anderl, “Digital twin – Proof of concept,”
Manufacturing Letters, pp. 10–12, 2018.
[20] T. Widmaier, “Optimisation of the roll geometry for production
conditions,” Aalto University, 2012.
[21] J. W. Cooley and J. W. Tukey, “An Algorithm for the Machine
Calculation of Complex Fourier Series,” Mathematics of
Computation, vol. 19, no. 90, p. 297, 1965.

VOLUME XX, 2017 2

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/.